Compare commits

..
Author SHA1 Message Date
Gao Han 69fd466fba Merge branch 'main' into feat/diffusion-prefetch-protection 2026-06-02 23:22:08 +08:00
tzhouam cb08af976b fix: correct logging message format in hub_prefetch.py for better clarity during prefetch attempts 2026-06-02 09:01:47 +00:00
tzhouam 617c6ca55f feat: enhance model loading with prefetching and retry logic
- Introduced `from_pretrained_with_prefetch` to handle racy cache scenarios by re-prefetching and retrying on failures.
- Updated various model pipelines to utilize the new prefetching mechanism, ensuring robust loading of model components.
- Prefetch logic added to multiple models, including HunyuanVideo, LongCatImage, and QwenImage, to mitigate issues with incomplete cache states.
2026-06-02 09:01:47 +00:00
240 changed files with 3489 additions and 26382 deletions

No files matched your search

-23
View File
@@ -389,29 +389,6 @@ steps:
volumes: volumes:
- "/fsx/hf_cache:/fsx/hf_cache" - "/fsx/hf_cache:/fsx/hf_cache"
- label: "TTS · VoxCPM2 Online Test"
timeout_in_minutes: 30
commands:
- |
timeout 30m bash -c "
export VLLM_LOGGING_LEVEL=DEBUG
export VLLM_WORKER_MULTIPROC_METHOD=spawn
pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts.py -m 'advanced_model and cuda' --run-level 'advanced_model'
"
agents:
queue: "gpu_1_queue"
plugins:
- docker#v5.2.0:
image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
always-pull: true
propagate-environment: true
shm-size: "8gb"
environment:
- "HF_HOME=/fsx/hf_cache"
- "HF_TOKEN"
volumes:
- "/fsx/hf_cache:/fsx/hf_cache"
- label: "Omni · Higgs-Audio V2 Test" - label: "Omni · Higgs-Audio V2 Test"
timeout_in_minutes: 30 timeout_in_minutes: 30
commands: commands:
+4 -188
View File
@@ -394,11 +394,11 @@ steps:
depends_on: upload-nightly-pipeline depends_on: upload-nightly-pipeline
if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test" if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "tts-test"
steps: steps:
- label: ":full_moon: TTS · Function Test with L4" - label: ":full_moon: TTS · Function Test"
timeout_in_minutes: 90 timeout_in_minutes: 90
commands: commands:
- export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1" - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
- pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_voxcpm2_tts_expansion.py - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
agents: agents:
queue: "gpu_1_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU queue: "gpu_1_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
plugins: plugins:
@@ -413,69 +413,6 @@ steps:
volumes: volumes:
- "/fsx/hf_cache:/fsx/hf_cache" - "/fsx/hf_cache:/fsx/hf_cache"
# - label: ":full_moon: TTS · Function Test with H100"
# timeout_in_minutes: 90
# commands:
# - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
# - pytest -s -v tests/e2e/ -m "full_model and H100 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
# agents:
# queue: "mithril-h100-pool"
# plugins:
# - kubernetes:
# podSpec:
# containers:
# - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
# resources:
# limits:
# nvidia.com/gpu: 1
# volumeMounts:
# - name: devshm
# mountPath: /dev/shm
# - name: hf-cache
# mountPath: /root/.cache/huggingface
# env:
# - name: HF_HOME
# value: /root/.cache/huggingface
# - name: HF_TOKEN
# valueFrom:
# secretKeyRef:
# name: hf-token-secret
# key: token
# nodeSelector:
# node.kubernetes.io/instance-type: gpu-h100-sxm
# volumes:
# - name: devshm
# emptyDir:
# medium: Memory
# - name: hf-cache
# hostPath:
# path: /mnt/hf-cache
# type: DirectoryOrCreate
- label: ":full_moon: TTS · Accuracy Test · VoxCPM2"
timeout_in_minutes: 30
commands:
- |
timeout 30m bash -c "
export VLLM_LOGGING_LEVEL=DEBUG
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts_expansion.py -m 'full_model and cuda and tts' --run-level 'full_model'
"
agents:
queue: "gpu_1_queue"
plugins:
- docker#v5.2.0:
image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
always-pull: true
propagate-environment: true
shm-size: "8gb"
environment:
- "HF_HOME=/fsx/hf_cache"
- "HF_TOKEN"
volumes:
- "/fsx/hf_cache:/fsx/hf_cache"
- label: ":full_moon: TTS · Perf Test" - label: ":full_moon: TTS · Perf Test"
key: nightly-tts-performance key: nightly-tts-performance
timeout_in_minutes: 180 timeout_in_minutes: 180
@@ -522,8 +459,8 @@ steps:
path: /mnt/hf-cache path: /mnt/hf-cache
type: DirectoryOrCreate type: DirectoryOrCreate
- label: ":full_moon: TTS · VoxCPM2 · Perf Test" - label: ":full_moon: VoxCPM2 · Perf Test"
key: nightly-tts-performance-voxcpm2 key: nightly-voxcpm2-performance
timeout_in_minutes: 180 timeout_in_minutes: 180
commands: commands:
- export BENCHMARK_DIR=tests/dfx/perf/results - export BENCHMARK_DIR=tests/dfx/perf/results
@@ -758,87 +695,6 @@ steps:
path: /mnt/hf-cache path: /mnt/hf-cache
type: DirectoryOrCreate type: DirectoryOrCreate
- label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU DreamZero"
timeout_in_minutes: 120
commands:
- >-
pytest -sv
tests/e2e/online_serving/test_dreamzero_expansion.py
-m "full_model and diffusion and H100 and distributed_cuda"
--run-level "full_model"
agents:
queue: "mithril-h100-pool"
plugins:
- kubernetes:
podSpec:
containers:
- image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
resources:
limits:
nvidia.com/gpu: 2
volumeMounts:
- name: devshm
mountPath: /dev/shm
- name: hf-cache
mountPath: /root/.cache/huggingface
env:
- name: HF_HOME
value: /root/.cache/huggingface
- name: HF_TOKEN
valueFrom:
secretKeyRef:
name: hf-token-secret
key: token
nodeSelector:
node.kubernetes.io/instance-type: gpu-h100-sxm
volumes:
- name: devshm
emptyDir:
medium: Memory
- name: hf-cache
hostPath:
path: /mnt/hf-cache
type: DirectoryOrCreate
- label: ":full_moon: Diffusion X2I(&A&T) · GLM-Image Function Test with H100"
timeout_in_minutes: 120
commands:
- pytest -sv tests/e2e/online_serving/test_glm_image_expansion.py -m "full_model and diffusion and H100" --run-level "full_model"
agents:
queue: "mithril-h100-pool"
plugins:
- kubernetes:
podSpec:
containers:
- image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
resources:
limits:
nvidia.com/gpu: 2
volumeMounts:
- name: devshm
mountPath: /dev/shm
- name: hf-cache
mountPath: /root/.cache/huggingface
env:
- name: HF_HOME
value: /root/.cache/huggingface
- name: HF_TOKEN
valueFrom:
secretKeyRef:
name: hf-token-secret
key: token
nodeSelector:
node.kubernetes.io/instance-type: gpu-h100-sxm
volumes:
- name: devshm
emptyDir:
medium: Memory
- name: hf-cache
hostPath:
path: /mnt/hf-cache
type: DirectoryOrCreate
- label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4" - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
timeout_in_minutes: 60 timeout_in_minutes: 60
@@ -937,46 +793,6 @@ steps:
path: /mnt/hf-cache path: /mnt/hf-cache
type: DirectoryOrCreate type: DirectoryOrCreate
- label: ":full_moon: HunyuanImage3-DIT · Accuracy Test"
timeout_in_minutes: 180
commands:
- export HUNYUAN_IMAGE3_MODEL="tencent/HunyuanImage-3.0-Instruct"
- export HUNYUAN_IMAGE3_DEVICES="0,1,2,3"
- pytest -s -v tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py -m full_model --run-level full_model
agents:
queue: "mithril-h100-pool"
plugins:
- kubernetes:
podSpec:
containers:
- image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
resources:
limits:
nvidia.com/gpu: 4
volumeMounts:
- name: devshm
mountPath: /dev/shm
- name: hf-cache
mountPath: /root/.cache/huggingface
env:
- name: HF_HOME
value: /root/.cache/huggingface
- name: HF_TOKEN
valueFrom:
secretKeyRef:
name: hf-token-secret
key: token
nodeSelector:
node.kubernetes.io/instance-type: gpu-h100-sxm
volumes:
- name: devshm
emptyDir:
medium: Memory
- name: hf-cache
hostPath:
path: /mnt/hf-cache
type: DirectoryOrCreate
- label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image" - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image"
key: nightly-diffusion-x2iat-performance-qwen-image key: nightly-diffusion-x2iat-performance-qwen-image
timeout_in_minutes: 180 timeout_in_minutes: 180
-82
View File
@@ -273,42 +273,6 @@ steps:
path: /mnt/hf-cache path: /mnt/hf-cache
type: DirectoryOrCreate type: DirectoryOrCreate
- label: "Omni · Qwen3-Omni Realtime WebSocket"
commands:
- |
timeout 30m bash -c "
export VLLM_WORKER_MULTIPROC_METHOD=spawn
pytest -s -v tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py::TestQwen3OmniRealtimeWebSocket::test_streaming_audio_input_pcm_output_async_chunk_smoke -m 'core_model' --run-level 'core_model'
"
agents:
queue: "mithril-h100-pool"
plugins:
- kubernetes:
podSpec:
containers:
- image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
resources:
limits:
nvidia.com/gpu: 2
volumeMounts:
- name: devshm
mountPath: /dev/shm
- name: hf-cache
mountPath: /root/.cache/huggingface
env:
- name: HF_HOME
value: /root/.cache/huggingface
nodeSelector:
node.kubernetes.io/instance-type: gpu-h100-sxm
volumes:
- name: devshm
emptyDir:
medium: Memory
- name: hf-cache
hostPath:
path: /mnt/hf-cache
type: DirectoryOrCreate
- label: "Omni · VoxCPM2 Test" - label: "Omni · VoxCPM2 Test"
timeout_in_minutes: 20 timeout_in_minutes: 20
commands: commands:
@@ -332,29 +296,6 @@ steps:
volumes: volumes:
- "/fsx/hf_cache:/fsx/hf_cache" - "/fsx/hf_cache:/fsx/hf_cache"
- label: "TTS · VoxCPM2 Online Test"
timeout_in_minutes: 30
commands:
- |
timeout 30m bash -c "
export VLLM_LOGGING_LEVEL=DEBUG
export VLLM_WORKER_MULTIPROC_METHOD=spawn
pytest -s -v tests/e2e/online_serving/test_voxcpm2_tts.py -m 'core_model and tts' --run-level 'core_model'
"
agents:
queue: "gpu_1_queue"
plugins:
- docker#v5.2.0:
image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
always-pull: true
propagate-environment: true
shm-size: "8gb"
environment:
- "HF_HOME=/fsx/hf_cache"
- "HF_TOKEN"
volumes:
- "/fsx/hf_cache:/fsx/hf_cache"
- label: "TTS · Qwen3-TTS CustomVoice Test" - label: "TTS · Qwen3-TTS CustomVoice Test"
commands: commands:
- | - |
@@ -399,29 +340,6 @@ steps:
volumes: volumes:
- "/fsx/hf_cache:/fsx/hf_cache" - "/fsx/hf_cache:/fsx/hf_cache"
- label: "TTS · Higgs-Audio-V3 Test"
commands:
- |
timeout 20m bash -c "
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export VLLM_USE_DEEP_GEMM=0
export VLLM_MOE_USE_DEEP_GEMM=0
pytest -s -v tests/e2e/online_serving/test_higgs_audio_v3.py -m 'core_model' --run-level 'core_model'
"
agents:
queue: "gpu_4_queue"
plugins:
- docker#v5.2.0:
image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
always-pull: true
propagate-environment: true
shm-size: "8gb"
environment:
- "HF_HOME=/fsx/hf_cache"
- "HF_TOKEN"
volumes:
- "/fsx/hf_cache:/fsx/hf_cache"
- label: "Diffusion · Qwen Image Edit Test" - label: "Diffusion · Qwen Image Edit Test"
commands: commands:
- | - |
-97
View File
@@ -1,97 +0,0 @@
---
name: precheck-pr
description: Self-check your branch before creating a PR — catch dead code, verify accuracy/perf claims, validate PR title format, and confirm merge readiness. Use when the user says "precheck", "self review", "pre-submit check", or "check my PR before I open it." Never posts to GitHub.
---
# PR Pre-Check
Self-review your branch before creating a PR against `vllm-project/vllm-omni`. Two modes: **quick** catches showstoppers, **full** does a thorough maintainer-grade review. Never posts to GitHub; the report is for the contributor's terminal only.
## Mode Selection
| Mode | When | Time |
|------|------|------|
| **Quick** | About to push, final sanity check | ~3 min |
| **Full** | Ready for review, want maintainer-level scan | ~10 min |
Default to quick if unsure. Run full before marking a PR "ready for review."
## Workflow
### Step 1: Detect Base Branch
```bash
BASE_SHA=$(git merge-base HEAD origin/main 2>/dev/null \
|| git merge-base HEAD main 2>/dev/null \
|| echo origin/main)
echo "diffing against ${BASE_SHA}"
git diff --name-only ${BASE_SHA}...HEAD
```
### Step 2: Validate PR Title
Check the most recent commit message (or branch name if no commit yet) against the project convention. Valid prefixes:
| Prefix | Applies to |
|--------|-----------|
| `[Bugfix]` | Bug fixes |
| `[CI/Build]` | Build or CI improvements |
| `[Doc]` | Documentation changes |
| `[Model]` | New/improved models (include model name) |
| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
| `[Kernel]` | CUDA/kernel changes |
| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
| `[Misc]` | Other changes (use sparingly) |
✗ if: missing prefix, wrong case (`[bugfix]`), or WIP/Draft in title.
⚠ if: `[Model]` prefix without the model identifier (e.g., `[Model] Add new model` — should be `[Model] Add <ModelName> ...`).
### Step 3: Categorize the PR
| Diff contains | PR type |
|---------------|---------|
| New files under `vllm_omni/model_executor/models/<name>/` | **New Model** |
| Changes to `vllm_omni/diffusion/` | **Diffusion Model** |
| `[Bugfix]` prefix or single-file fix | **Bug Fix** |
| Perf/benchmark/throughput claims in commit msg or diff | **Performance** |
| Everything else | **General** |
If multiple rows apply (e.g., a diffusion model is also a new model), union the checklists.
### Step 4: Run Checklist
Ask: "Quick mode or full mode?" Then walk the checklist for the detected PR type from [references/checklists.md](references/checklists.md). Each item produces ✓, ✗, or ⚠.
### Step 5: Print Report
```
Pre-check report for <branch>
Mode: quick | full
Type: <new-model | diffusion-model | bug-fix | perf | general>
Dimension Result
───────────────── ──────
PR title format ✓
PR desc integrity ✓
Registry/config ✓
Dead code ⚠ 2 warnings
Accuracy ✓
Benchmark ✗ missing software versions
Verdict: 1 blocking | 2 warnings | recommend fixing ✗ before PR
```
**Severity:**
| Mark | Meaning |
|------|---------|
| ✗ | Blocking — fix before opening PR |
| ⚠ | Warning — consider fixing |
| ✓ | Pass |
| — | Skipped (not applicable) |
### Stop Here
Do not post comments, open PRs, or modify files. The report is for the contributor's terminal only.
@@ -1,135 +0,0 @@
# Pre-Check Checklists
Each item is a concrete check to run against the diff. ✗ = fix before PR. ⚠ = consider fixing.
---
## All PRs: Title Format (check first)
PR title must follow the project convention documented in `docs/contributing/README.md`:
| Prefix | Applies to |
|--------|-----------|
| `[Bugfix]` | Bug fixes |
| `[CI/Build]` | Build or CI improvements |
| `[Doc]` | Documentation changes |
| `[Model]` | New/improved models (include model name) |
| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
| `[Kernel]` | CUDA/kernel changes |
| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
| `[Misc]` | Other changes (use sparingly) |
**Examples:**
```
✓ [Model] Add MiniCPM-o 4.5 support
✓ [Bugfix] Fix Ovis image text encoder dtype
✓ [Core] Qwen3-Omni performance optimization
✓ [Hardware][Ascend] Fix VoxCPM2 on Ascend
✗ Fix Ovis image text encoder dtype ← missing prefix
✗ [bugfix] Fix Ovis image text encoder ← wrong case
✗ [WIP] Add model ← WIP in title
```
- [ ] **Title has valid prefix** — ✗ if missing prefix, wrong case, or WIP/Draft in title. For `[Model]` PRs, the title MUST also include the model identifier (e.g., `[Model] Add <ModelName> ...`). ⚠ if missing the model name.
---
## Bug Fix PRs
### Quick
- [ ] **PR body has 4 sections:** what broke (error message), repro steps, root cause, fix description
- [ ] **Fix is minimal:** diff touches only files directly related to the bug. No unrelated refactors or style changes.
- [ ] **Repro is runnable:** the PR body includes a copy-paste command that reproduces the bug
- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
### Full (adds)
- [ ] **Regression test exists:** `git diff --name-only` includes at least one `tests/` file
- [ ] **Regression test reproduces the original error:** the test would fail on main, pass on this branch
- [ ] **Fix matches root cause exactly:** no "fix the symptom + something else" — if the root cause is dtype, the fix is dtype, not dtype + reformatting
- [ ] **No silent failure risk:** no bare `except: pass`, no `try: ... except: return None`, no empty fallback added
- [ ] **Upstream pattern match:** the fix follows the same pattern as existing code for similar cases (grep for analogous `from_pretrained` calls, etc.)
- [ ] **Environment documented:** torch/transformers/vllm versions listed if the bug is version-dependent
---
## Performance PRs
### Quick
- [ ] **Before/after numbers in PR body:** at minimum, one metric (RTF, RPS, latency, VRAM) with main vs PR values
- [ ] **Hardware specified:** GPU model, count, VRAM. "L4" or "H20" is sufficient. "tested on GPU" is not.
- [ ] **Concurrency scaling:** if the optimization claims better throughput, numbers at 2+ concurrency levels
- [ ] **No unexplained regressions:** if any metric goes down at any concurrency level, explain why
- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
### Full (adds)
- [ ] **Software versions:** torch, CUDA, vllm, vllm-omni versions in PR body
- [ ] **Warmup stated:** "1 warmup + 3 measured" or equivalent
- [ ] **Stddev or range:** not just a single best run. "average of 2 rounds" is minimum.
- [ ] **VRAM measured:** peak GPU memory before and after
- [ ] **Benchmark script checked in:** under `tests/` or `examples/`, runnable by reviewer
- [ ] **Exact command line in PR:** not pseudocode, not "run the benchmark"
- [ ] **Pytest summary included:** `36 passed, 3 skipped, 17 warnings in 2897.66s` — proves the whole suite passes
- [ ] **Config changes documented:** if `max_num_seqs`, `gpu_memory_utilization`, or deploy YAML changed, old and new values stated with reasoning
---
## New Model PRs
### Quick
- [ ] **PR body file list matches `git diff --name-only`:** every file path in the description exists in the diff
- [ ] **Registry entries match claimed count:** count `_OMNI_MODELS` or `_DIFFUSION_MODELS` entries added vs claimed in body
- [ ] **Pipeline config `custom_process_input_func` paths resolve:** each referenced module.function exists
- [ ] **`__init__.py` exports match registry:** each registered class appears in `__init__.py` and `__all__`
- [ ] **Shell script modes match README:** `run_curl.sh` supports every mode the README documents
- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
### Full (adds)
- [ ] **Dead code scan (5 patterns):**
1. Dead `forward()` — defined but never called through the pipeline path
2. Dead factory/builder functions — instantiated nowhere
3. Dead wrappers — thin wrappers with no callers
4. Dead branch guards — `if version > X` where X is always satisfied
5. Unused parameters — `__init__` args never accessed
- [ ] **Copy-paste detection:** no string constant defined in 3+ files, no cross-module validation duplication, no near-identical shape coercion fns
- [ ] **Import hygiene:** no re-export-only imports, no module-level side effects without docs, no `import os` in function body when already at top. Function-body imports are acceptable when the import is only reachable from a closure or factory return value (i.e., `import` would otherwise execute at module load even though the user may never trigger that path). Either lift to top-of-file or annotate with `# noqa: PLC0415`.
- [ ] **Accuracy:** at least one test compares output against reference; audio = valid WAV at correct sample rate; determinism verified if seed claimed. ⚠ if missing and not explicitly deferred to a follow-up PR or RFC section; — if explicitly deferred (the deferral target SHOULD be linked in the PR body).
- [ ] **Performance:** hardware, software versions, warmup, and metrics (RTF, VRAM, RPS, latency) stated. ⚠ if missing and not explicitly deferred; — if explicitly deferred.
- [ ] **Benchmark settings:** model config (TP, PP, max_model_len, enforce_eager, quant), runtime config (batch, input/output spec), environment versions
- [ ] **Benchmark script checked in + exact command line + pytest summary**
---
## Diffusion Model PRs
Diffusion models live under `vllm_omni/diffusion/`. In addition to the New Model checklist above, verify:
- [ ] **Transformer adapter follows contract:** inherits from `nn.Module`, implements `load_weights()`, uses `vllm_omni.diffusion.attention.layer.Attention`
- [ ] **Pipeline has both offline and online paths:** `forward()` for batch generation, streaming path for online serving
- [ ] **Cache-DiT integration:** at least one acceleration (TeaCache, FBCache, etc.) is wired up. ✗ for new **top-level** pipelines. For **subclass** pipelines that inherit `__call__` from a parent where Cache-DiT is already absent, downgrade to ⚠ and recommend filing a follow-up issue against the parent pipeline family.
- [ ] **Parallelism config:** TP/SP/USP/CFG-Parallel options exposed in pipeline config
- [ ] **Docs table updated:** model added to `docs/models/supported_models.md` with correct metadata
- [ ] **E2E test exists:** at least one test exercises the full generate path
---
## General PRs
### Quick
- [ ] **PR body matches diff:** description doesn't claim changes not in the diff
- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
- [ ] **CI gates passing:** DCO, pre-commit, build — check `gh pr view --json statusCheckRollup` if a PR already exists
### Full (adds)
- [ ] **No new dead code in changed files:** scan for unreachable paths, unused imports
- [ ] **Test coverage:** if core code changed (`engine/`, `stages/`, `connectors/`), tests are added or existing tests cover the path
- [ ] **Import hygiene:** no redundant imports added
- [ ] **No unrelated changes:** diff scope matches PR description
-5
View File
@@ -25,12 +25,7 @@ include:
- `add-tts-model`: covers integration of new TTS models and related serving - `add-tts-model`: covers integration of new TTS models and related serving
workflows workflows
- `generate-release-note`: helps prepare release notes for repository changes - `generate-release-note`: helps prepare release notes for repository changes
- `precheck-pr`: self-check a branch before creating a PR — validates PR title
format, catches dead code, verifies accuracy/perf claims, and confirms merge
readiness
- `review-pr`: provides a structured workflow for reviewing pull requests - `review-pr`: provides a structured workflow for reviewing pull requests
- `vllm-omni-npu-model-runner-upgrade`: upgrades NPU model runners to align with the
latest vllm-ascend NPUModelRunner
## Maintenance Guidelines ## Maintenance Guidelines
+1 -4
View File
@@ -1,6 +1,4 @@
<!-- markdownlint-disable --> <!-- markdownlint-disable -->
**Before submitting:** run the precheck-pr skill with code agent for a self-check against project conventions — catches dead code, missing benchmarks, and title format issues.
PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED. PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
## Purpose ## Purpose
@@ -20,5 +18,4 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTT
- [ ] (Optional) Release notes update. If your change is user-facing, please update the release notes draft. - [ ] (Optional) Release notes update. If your change is user-facing, please update the release notes draft.
</details> </details>
**BEFORE SUBMITTING,** please read <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md> **BEFORE SUBMITTING, PLEASE READ <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md>** (anything written below this line will be removed by GitHub Actions)
(anything written below this line will be removed by GitHub Actions)
-1
View File
@@ -16,7 +16,6 @@ Easy, fast, and cheap omni-modality model serving for everyone
--- ---
*Latest News* 🔥 *Latest News* 🔥
- [2026/06] We released [0.22.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.22.0) - an **omnimodal world-model** release aligned with vLLM 0.22, featuring [Nvidia Cosmos3](recipes/cosmos3/Cosmos3-Nano.md)/DreamZero world model support, expanded quantization coverage across Blackwell/NPU/XPU, TTS production improvements, new models including MiniCPM-o 4.5, MOSS-TTS, and Lance, plus RL integration with [VeRL-Omni](https://github.com/verl-project/verl-omni).
- [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends. - [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends.
- [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments. - [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
- [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup! - [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup!
-9
View File
@@ -151,9 +151,6 @@ class VllmOmniImageClient:
seed: int | None = None, seed: int | None = None,
negative_prompt: str | None = None, negative_prompt: str | None = None,
output_compression: int | None = None, output_compression: int | None = None,
bot_task: str | None = None,
sys_type: str | None = None,
system_prompt: str | None = None,
) -> Image.Image: ) -> Image.Image:
if not isinstance(images, list): if not isinstance(images, list):
images = [images] images = [images]
@@ -173,12 +170,6 @@ class VllmOmniImageClient:
data["negative_prompt"] = negative_prompt data["negative_prompt"] = negative_prompt
if output_compression is not None: if output_compression is not None:
data["output_compression"] = str(output_compression) data["output_compression"] = str(output_compression)
if bot_task is not None:
data["bot_task"] = bot_task
if sys_type is not None:
data["sys_type"] = sys_type
if system_prompt is not None:
data["system_prompt"] = system_prompt
files = [ files = [
( (
@@ -64,47 +64,6 @@ def resolve_model_name(*, model_name: str | None, model: str | None = None, outp
raise ValueError("model-name is required when it cannot be inferred from model or output-root") raise ValueError("model-name is required when it cannot be inferred from model or output-root")
def _coerce_score_int(value: Any) -> int:
if isinstance(value, bool):
raise TypeError(f"Expected numeric score, got bool: {value!r}")
if isinstance(value, (int, float)):
return int(value)
if isinstance(value, str):
stripped = value.strip()
if stripped.lstrip("-").isdigit():
return int(stripped)
raise TypeError(f"Expected numeric score, got {type(value).__name__}: {value!r}")
def _expand_score_entry(entry: Any) -> list[int]:
if isinstance(entry, dict):
if "score" in entry:
return [_coerce_score_int(entry["score"])]
if "value" in entry:
return [_coerce_score_int(entry["value"])]
numeric_values = [_coerce_score_int(value) for value in entry.values()]
if numeric_values:
return numeric_values
raise TypeError(f"Score dict has no numeric values: {entry!r}")
return [_coerce_score_int(entry)]
def _normalize_score_list(score: Any) -> list[int]:
if isinstance(score, dict):
entries: list[Any] = [score]
elif isinstance(score, list):
entries = score
else:
entries = [score]
normalized: list[int] = []
for entry in entries:
normalized.extend(_expand_score_entry(entry))
if not normalized:
raise ValueError(f"Empty score list after normalization: {score!r}")
return normalized
def parse_score_payload(raw_text: str) -> dict[str, Any]: def parse_score_payload(raw_text: str) -> dict[str, Any]:
try: try:
parsed = extract_json_object(raw_text) parsed = extract_json_object(raw_text)
@@ -114,7 +73,10 @@ def parse_score_payload(raw_text: str) -> dict[str, Any]:
parsed = {"score": json.loads(stripped), "reasoning": ""} parsed = {"score": json.loads(stripped), "reasoning": ""}
else: else:
raise raise
parsed["score"] = _normalize_score_list(parsed.get("score", [])) score = parsed.get("score", [])
if not isinstance(score, list):
score = [score]
parsed["score"] = [int(value) for value in score]
return parsed return parsed
@@ -429,9 +391,6 @@ class GEditBenchRunner:
num_inference_steps: int = 20, num_inference_steps: int = 20,
guidance_scale: float | None = None, guidance_scale: float | None = None,
seed: int | None = 42, seed: int | None = 42,
bot_task: str | None = None,
sys_type: str | None = None,
system_prompt: str | None = None,
): ):
self.dataset_ref = dataset_ref self.dataset_ref = dataset_ref
self.output_root = output_root self.output_root = output_root
@@ -441,9 +400,6 @@ class GEditBenchRunner:
self.num_inference_steps = num_inference_steps self.num_inference_steps = num_inference_steps
self.guidance_scale = guidance_scale self.guidance_scale = guidance_scale
self.seed = seed self.seed = seed
self.bot_task = bot_task
self.sys_type = sys_type
self.system_prompt = system_prompt
self.client = VllmOmniImageClient(base_url=base_url, api_key=api_key) self.client = VllmOmniImageClient(base_url=base_url, api_key=api_key)
def generate( def generate(
@@ -522,9 +478,6 @@ class GEditBenchRunner:
num_inference_steps=self.num_inference_steps, num_inference_steps=self.num_inference_steps,
guidance_scale=self.guidance_scale, guidance_scale=self.guidance_scale,
seed=self.seed, seed=self.seed,
bot_task=self.bot_task,
sys_type=self.sys_type,
system_prompt=self.system_prompt,
) )
edited_image.save(output_path) edited_image.save(output_path)
return { return {
@@ -680,25 +633,6 @@ def build_parser() -> argparse.ArgumentParser:
generate.add_argument("--num-inference-steps", type=int, default=20) generate.add_argument("--num-inference-steps", type=int, default=20)
generate.add_argument("--guidance-scale", type=float, default=None) generate.add_argument("--guidance-scale", type=float, default=None)
generate.add_argument("--seed", type=int, default=42) generate.add_argument("--seed", type=int, default=42)
generate.add_argument(
"--bot-task",
type=str,
default=None,
choices=["think", "recaption", "think_recaption", "vanilla"],
help="HunyuanImage-3 prompting mode for /v1/images/edits (optional).",
)
generate.add_argument(
"--sys-type",
type=str,
default=None,
help="Hunyuan system-prompt type (e.g. en_unified). Defaults from --bot-task when omitted.",
)
generate.add_argument(
"--system-prompt",
type=str,
default=None,
help="Custom Hunyuan system prompt body (sys_type=custom).",
)
generate.add_argument("--workers", type=int, default=1) generate.add_argument("--workers", type=int, default=1)
generate.add_argument("--max-samples", type=int, default=None) generate.add_argument("--max-samples", type=int, default=None)
generate.add_argument("--samples-per-group", type=int, default=None) generate.add_argument("--samples-per-group", type=int, default=None)
@@ -741,9 +675,6 @@ def main(argv: list[str] | None = None) -> int:
num_inference_steps=args.num_inference_steps, num_inference_steps=args.num_inference_steps,
guidance_scale=args.guidance_scale, guidance_scale=args.guidance_scale,
seed=args.seed, seed=args.seed,
bot_task=args.bot_task,
sys_type=args.sys_type,
system_prompt=args.system_prompt,
) )
records = runner.generate( records = runner.generate(
model_name=model_name, model_name=model_name,
+1
View File
@@ -92,6 +92,7 @@ Core scheduling and caching components.
Configuration classes. Configuration classes.
- [vllm_omni.config.model.OmniModelConfig][] - [vllm_omni.config.model.OmniModelConfig][]
- [vllm_omni.config.stage_config.ModelPipeline][]
- [vllm_omni.config.stage_config.StageConfig][] - [vllm_omni.config.stage_config.StageConfig][]
- [vllm_omni.config.stage_config.StageConfigFactory][] - [vllm_omni.config.stage_config.StageConfigFactory][]
- [vllm_omni.config.stage_config.StageType][] - [vllm_omni.config.stage_config.StageType][]
Binary file not shown.

Before

Width:  |  Height:  |  Size: 403 KiB

After

Width:  |  Height:  |  Size: 408 KiB

+3 -111
View File
@@ -2,7 +2,7 @@
The vllm bench command launches the vLLM-Omni benchmark to evaluate the performance of multimodal models. The vllm bench command launches the vLLM-Omni benchmark to evaluate the performance of multimodal models.
## Notes ## Notes
We currently only support using the "openai-chat-omni" and "openai-image-edits-omni" backend. We currently only support using the "openai-chat-omni" backend.
## Basic Parameter Description ## Basic Parameter Description
You can use `vllm bench serve --omni --help=all` to get descriptions of all parameters. The commonly used parameters are described below: You can use `vllm bench serve --omni --help=all` to get descriptions of all parameters. The commonly used parameters are described below:
@@ -52,11 +52,7 @@ You can use `vllm bench serve --omni --help=all` to get descriptions of all para
- `--percentile-metrics` - `--percentile-metrics`
"Comma-separated list of selected metrics to report percentiles." "Comma-separated list of selected metrics to report percentiles."
"This argument specifies the metrics to report percentiles." "This argument specifies the metrics to report percentiles."
'Allowed metric names are "ttft", "tpot", "itl", "ttfc", "tpoc", "icl", ' 'Allowed metric names are "ttft", "tpot", "itl", "e2el", "audio_ttfp", "audio_rtf", "audio_duration". '
'"tpop", "e2el", "audio_ttfp", "audio_rtf", "audio_duration". '
- `-print-stage`
Print per-stage benchmark metrics for --omni serving when stage metrics are returned by the server. Disabled by default.
- `--save-result` - `--save-result`
Specify to save benchmark results to a json file Specify to save benchmark results to a json file
@@ -280,9 +276,7 @@ Generate synthetic image、video、audio inputs alongside random text prompts to
Notes: Notes:
- Works only with online benchmark via: - Works only with online benchmark via the OpenAI backend (`--backend openai-chat-omni`) and endpoint `/v1/chat/completions`.
- the OpenAI chat backend (`--backend openai-chat-omni`) and endpoint `/v1/chat/completions`.
- the OpenAI edit image backend (`--backend openai-image-edits-omni`) and endpoint `/v1/images/edits`.
Start the server (example): Start the server (example):
@@ -363,105 +357,3 @@ How sampling works:
This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`. This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`. - The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
</details> </details>
### Multi-Stage Benchmark
<details class="admonition abstract" markdown="1">
<summary>Show more</summary>
Start the server (change the port and model):
```bash
vllm serve --omni --port 29999 --model ~/Qwen3-Omni-30B-A3B-Instruct
```
Then run the benchmarking script (remember to add the flag `--print-stage`):
```bash
vllm bench serve --omni \
--dataset-name random \
--port 29999 \
--max-concurrency 1 \
--num-warmups 2 \
--model ~/Qwen3-Omni-30B-A3B-Instruct \
--endpoint /v1/chat/completions \
--backend openai-chat-omni \
--num-prompts 3 \
--random-input-len 2500 \
--ignore-eos \
--percentile-metrics ttft,tpot,itl,e2el,audio_ttfp,audio_rtf,ttfc,tpoc,icl \
--random-output-len 900 \
--extra_body '{"modalities": ["text","audio"]}' \
--print-stage
```
Besides the "Serving Benchmark Result", there will be "Stage Benchmark Result" below:
```text
============= Stage Benchmark Result =============
=============== Stage 0 (thinker) ================
-------------------Stage Timing-------------------
Mean stage_gen_time (ms): 18243.16
Median stage_gen_time (ms): 18220.83
P99 stage_gen_time (ms): 18296.75
================== Text Result ===================
Stage generated tokens: 2700
-----------Serving Time to First Token------------
Mean Serving TTFT (ms): 816.31
Median Serving TTFT (ms): 818.94
P99 Serving TTFT (ms): 829.76
-----Time per Output Token (excl. 1st token)------
Mean TPOP (ms): 19.41
Median TPOP (ms): 19.40
P99 TPOP (ms): 19.47
---------------Inter-token Latency----------------
Mean ITL (ms): 19.41
Median ITL (ms): 19.12
P99 ITL (ms): 23.58
================ Stage 1 (talker) ================
-------------------Stage Timing-------------------
Mean stage_gen_time (ms): 18849.64
Median stage_gen_time (ms): 13641.15
P99 stage_gen_time (ms): 33655.58
============= Internal Stream Result =============
-----------Serving Time to First Chunk------------
Mean Serving TTFC (ms): 1029.14
Median Serving TTFC (ms): 1027.98
P99 Serving TTFC (ms): 1032.59
-----Time per Output Chunk (excl. 1st chunk)------
Mean TPOP (ms): 17.53
Median TPOP (ms): 18.70
P99 TPOP (ms): 18.88
---------------Inter-chunk Latency----------------
Mean ICL (ms): 16.27
Median ICL (ms): 18.19
P99 ICL (ms): 25.86
=============== Stage 2 (code2wav) ===============
-------------------Stage Timing-------------------
Mean stage_gen_time (ms): 18860.74
Median stage_gen_time (ms): 13655.84
P99 stage_gen_time (ms): 33659.79
================== Audio Result ==================
Stage audio duration generated(s): 260.11
Stage audio frames generated: 6242730
-----------Serving Time to First Packet-----------
Mean Serving AUDIO_TTFP (ms): 1366.16
Median Serving AUDIO_TTFP (ms): 1367.20
P99 Serving AUDIO_TTFP (ms): 1368.26
-----------------Real Time Factor-----------------
Mean AUDIO_RTF: 0.24
Median AUDIO_RTF: 0.26
P99 AUDIO_RTF: 0.27
```
Explanation:
- stage_gen_time: Time from submitting a request to a specific stage to that stage finishing generation.
- Serving TTFC (Time to First Chunk): Time from the HTTP request being accepted by the serving frontend to
the stage producing its first non-empty output chunk. Keep the name TTFT in text stage, TTFP in audio stage
for easier mapping.
- TPOP (Time per Output Chunk): Average time from the first output chunk to stage completion, divided by the
number of remaining output chunks. The TPOP abbreviation follows Qwen3.5-Omni Technical Report.
- ICL (Inter-Chunk Latency): Time between two consecutive output chunks produced by the same stage. Keep the
name ITL in text stage for easier mapping.
</details>
-14
View File
@@ -116,20 +116,6 @@ Only specific types of PRs will be reviewed. The PR title is prefixed appropriat
!!! note !!! note
If the PR spans more than one category, please include all relevant prefixes. If the PR spans more than one category, please include all relevant prefixes.
### Pre-Check Before Submitting
Before submitting a PR, run the precheck-pr skill with code agent for a self-review against project conventions:
```
/precheck-pr
```
The skill offers two modes:
- **Quick (~3 min):** catches showstoppers — PR title format, missing benchmark claims, rebase status
- **Full (~10 min):** thorough maintainer-grade review — dead code scan, copy-paste detection, import hygiene
The precheck covers five PR types: Bug Fix, Performance, New Model, Diffusion Model, and General. Each type has a tailored checklist that validates evidence quality (repro steps, A/B benchmarks, registry entries, etc.). See the skill in `.claude/skills/precheck-pr/` for the full checklist.
### Local Test ### Local Test
Please run the L1 and L2 test cases locally first and attach the results before contacting us to add the "ready" label. Please refer to the [test instructions](./ci/test_guide.md) for running the test cases. Please run the L1 and L2 test cases locally first and attach the results before contacting us to add the "ready" label. Please refer to the [test instructions](./ci/test_guide.md) for running the test cases.
+2 -2
View File
@@ -305,8 +305,8 @@ class GPUWorker:
) )
# Load model # Load model
model_loader = DiffusersPipelineLoader(load_config, self.od_config) model_loader = DiffusersPipelineLoader(load_config)
self.pipeline = model_loader.load_model(load_device=f"cuda:{rank}") self.pipeline = model_loader.load_model(od_config, load_device=f"cuda:{rank}")
# Setup cache backend # Setup cache backend
from vllm_omni.diffusion.cache.selector import get_cache_backend from vllm_omni.diffusion.cache.selector import get_cache_backend
@@ -45,7 +45,7 @@ If you do not need to modify source code of vLLM, you can directly install the s
uv pip install vllm==0.22.0 --torch-backend=auto uv pip install vllm==0.22.0 --torch-backend=auto
``` ```
The 0.22.0 release of vLLM ships CUDA 12.9-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead. The 0.22.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
#### Installation of vLLM-Omni #### Installation of vLLM-Omni
Since vllm-omni is rapidly evolving, it's recommended to install it from source Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -70,8 +70,8 @@ git checkout v0.22.0
``` ```
Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file. Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
```bash ```bash
#For CUDA 12.9 (the default for v0.22.0; the wheel filename has no `+cu129` suffix) #For CUDA 13.0 (the default for v0.22.0; the wheel filename has no `+cu130` suffix)
export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.22.0/vllm-0.22.0-cp38-abi3-manylinux_2_28_x86_64.whl export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.22.0/vllm-0.22.0-cp38-abi3-manylinux_2_35_x86_64.whl
``` ```
Install vllm with command below (If you have no existing PyTorch). Install vllm with command below (If you have no existing PyTorch).
```bash ```bash
+1 -1
View File
@@ -22,7 +22,7 @@ source .venv/bin/activate
uv pip install vllm==0.22.0 --torch-backend=auto uv pip install vllm==0.22.0 --torch-backend=auto
# On ROCm # On ROCm
uv pip install vllm==0.22.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.22.0/rocm722 uv pip install vllm==0.22.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.22.0/rocm721
git clone https://github.com/vllm-project/vllm-omni.git git clone https://github.com/vllm-project/vllm-omni.git
cd vllm-omni cd vllm-omni
+1 -10
View File
@@ -33,7 +33,7 @@ th {
| `ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `WanPipeline` | Wan2.1-T2V, Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`, `Wan-AI/Wan2.1-T2V-14B-Diffusers`, `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V, V2V, T2V with sound, action policy | `nvidia/Cosmos3-Nano` | ✅︎ | | | | | `Cosmos3OmniDiffusersPipeline` | Cosmos3 T2I, T2V, I2V | `nvidia/Cosmos3-Nano` | ✅︎ | | | |
| `WanSpeechToVideoPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `WanSpeechToVideoPipeline` | Wan2.2-S2V | `Wan-AI/Wan2.2-S2V-14B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `Wan22VACEPipeline` | Wan2.1-VACE | `Wan-AI/Wan2.1-VACE-1.3B-diffusers`, `Wan-AI/Wan2.1-VACE-14B-diffusers` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `LTX2Pipeline` | LTX-2-T2V | `Lightricks/LTX-2` | ✅︎ | ✅︎ | | | | `LTX2Pipeline` | LTX-2-T2V | `Lightricks/LTX-2` | ✅︎ | ✅︎ | | |
@@ -42,7 +42,6 @@ th {
| `LTX2ImageToVideoTwoStagesPipeline` | LTX-2-I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | | | `LTX2ImageToVideoTwoStagesPipeline` | LTX-2-I2V | `rootonchair/LTX-2-19b-distilled` | ✅︎ | ✅︎ | | |
| `LTX23Pipeline` | LTX-2.3-T2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | | `LTX23Pipeline` | LTX-2.3-T2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
| `LTX23ImageToVideoPipeline` | LTX-2.3-I2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | | | `LTX23ImageToVideoPipeline` | LTX-2.3-I2V | `dg845/LTX-2.3-Diffusers` | ✅︎ | ✅︎ | | |
| `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | | | |
| `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | | | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | |
| `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | | | `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | |
| `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ | | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ |
@@ -50,8 +49,6 @@ th {
| `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ | | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ |
| `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ | | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ |
| `OmniVoiceModel` | OmniVoice | `k2-fsa/OmniVoice` | ✅︎ | | | |
| `VoxCPM2TalkerForConditionalGeneration` | VoxCPM2 | `openbmb/VoxCPM2` | ✅︎ | | | |
| `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | | | `MammothModa2ForConditionalGeneration` | MammothModa2-Preview | `bytedance-research/MammothModa2-Preview` | ✅︎ | ✅︎ | | |
| `Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `FluxKontextPipeline` | FLUX.1-Kontext-dev | `black-forest-labs/FLUX.1-Kontext-dev` | ✅︎ | ✅︎ | | | | `FluxKontextPipeline` | FLUX.1-Kontext-dev | `black-forest-labs/FLUX.1-Kontext-dev` | ✅︎ | ✅︎ | | |
@@ -64,10 +61,6 @@ th {
| `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
| `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | | | `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | |
| `MossTTSNanoForCausalLM` | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✅︎ | | | |
| `MossTTSDelayModel` | MOSS-TTS, MOSS-TTSD, MOSS-SoundEffect, MOSS-VoiceGenerator | `OpenMOSS-Team/MOSS-VoiceGenerator` | ✅︎ | | | |
| `MossTTSRealtime` | MOSS-TTS-Realtime | `OpenMOSS-Team/MOSS-TTS-Realtime` | ✅︎ | | | |
| `HiggsAudioV2ForConditionalGeneration` | Higgs-Audio v2 | `bosonai/higgs-audio-v2-generation-3B-base` | ✅︎ | | | |
| `NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` | ✅︎ | ✅︎ | | ✅︎ | | `NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` | ✅︎ | ✅︎ | | ✅︎ |
| `MiMoAudioModel` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` | ✅︎ | ✅︎ | | | | `MiMoAudioModel` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` | ✅︎ | ✅︎ | | |
| `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | | | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | |
@@ -75,13 +68,11 @@ th {
| `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | | | `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | |
| `DreamIDOmniPipeline` | DreamID-Omni | `XuGuo699/DreamID-Omni` | ✅︎ | ✅︎ | | | | `DreamIDOmniPipeline` | DreamID-Omni | `XuGuo699/DreamID-Omni` | ✅︎ | ✅︎ | | |
| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | | | `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | |
| `LancePipeline` | Lance | `bytedance-research/Lance` | ✅︎ | | | |
| `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | | | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | |
| `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | | | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | |
| `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | | | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | |
| `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | | | `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | |
|`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | | |`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | |
| `MiniCPMO45OmniForConditionalGeneration` | MiniCPM-o 4.5 | `openbmb/MiniCPM-o-4_5` | ✅︎ | | ✅︎ | |
| `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
|`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | | |`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | |
+1 -1
View File
@@ -116,7 +116,7 @@ Content-Type: application/json
| `initial_codec_chunk_frames` | integer | null | Per-request initial chunk size override for TTFA tuning. When null, IC is computed dynamically based on server load. | | `initial_codec_chunk_frames` | integer | null | Per-request initial chunk size override for TTFA tuning. When null, IC is computed dynamically based on server load. |
| `stream` | bool | false | Stream raw PCM chunks as they are decoded (requires `response_format="pcm"`) | | `stream` | bool | false | Stream raw PCM chunks as they are decoded (requires `response_format="pcm"`) |
**Supported languages:** Only applicable to Qwen3-TTS. Derived from the model configuration (`talker_config.codec_language_id` in the checkpoint's `config.json`), plus `Auto`, which is always accepted. Official Qwen3-TTS checkpoints support: Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian. **Supported languages:** Auto, Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
#### Voice Clone Parameters (Base task) #### Voice Clone Parameters (Base task)
+3 -44
View File
@@ -67,9 +67,8 @@ curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o output.mp4
| Parameter | Type | Default | Description | | Parameter | Type | Default | Description |
|-----------|------|---------|-------------| |-----------|------|---------|-------------|
| `input_reference` | file | null | Uploaded reference image or video for image-to-video/video-to-video requests | | `input_reference` | file | null | Uploaded reference image for image-to-video requests |
| `image_reference` | string | null | JSON-encoded reference image payload; do not combine with `input_reference` or `video_reference` | | `image_reference` | string | null | JSON-encoded reference image payload; do not combine with `input_reference` |
| `video_reference` | string | null | JSON-encoded reference video payload; do not combine with `input_reference` or `image_reference` |
| `width` | integer | model default | Output video width | | `width` | integer | model default | Output video width |
| `height` | integer | model default | Output video height | | `height` | integer | model default | Output video height |
| `num_frames` | integer | model default | Number of generated frames | | `num_frames` | integer | model default | Number of generated frames |
@@ -81,8 +80,6 @@ curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o output.mp4
| `flow_shift` | number | null | Scheduler flow-shift value | | `flow_shift` | number | null | Scheduler flow-shift value |
| `true_cfg_scale` | number | null | True CFG scale when supported by the model | | `true_cfg_scale` | number | null | True CFG scale when supported by the model |
| `seed` | integer | null | Random seed for reproducibility | | `seed` | integer | null | Random seed for reproducibility |
| `generate_sound` | boolean | false | Request model-generated audio for video models that support sound generation |
| `sound_duration` | number | null | Duration in seconds for generated audio; defaults to generated video duration |
| `negative_prompt` | string | null | Text describing what to avoid in the generated video | | `negative_prompt` | string | null | Text describing what to avoid in the generated video |
| `enable_frame_interpolation` | boolean | null | Enable post-generation frame interpolation | | `enable_frame_interpolation` | boolean | null | Enable post-generation frame interpolation |
| `frame_interpolation_exp` | integer | null | Interpolation exponent; `1=2x`, `2=4x`, and so on | | `frame_interpolation_exp` | integer | null | Interpolation exponent; `1=2x`, `2=4x`, and so on |
@@ -126,41 +123,6 @@ curl -s http://localhost:8091/v1/videos \
-F "fps=16" -F "fps=16"
``` ```
### Video-to-Video
For models that support video conditioning, upload the reference video with
`input_reference`:
```bash
curl -s http://localhost:8091/v1/videos \
-F "prompt=continue this motion with consistent subjects and lighting" \
-F "input_reference=@input.mp4;type=video/mp4" \
-F "width=1280" \
-F "height=720" \
-F "num_frames=80" \
-F "fps=16"
```
You can also pass a JSON-safe video URL or `data:video/...;base64,...` payload
through `video_reference`. Do not send `video_reference` together with
`input_reference` or `image_reference`.
```bash
curl -s http://localhost:8091/v1/videos \
-F "prompt=continue this motion with consistent subjects and lighting" \
-F 'video_reference={"video_url":"https://example.com/input.mp4"}' \
-F "width=1280" \
-F "height=720" \
-F "num_frames=80" \
-F "fps=16"
```
JSON references currently support `image_url`/`video_url`; `file_id` references
are not implemented yet. Models may expose additional V2V controls through
`extra_params`. For example, Cosmos3 supports
`condition_frame_indexes_vision` and `condition_video_keep` to select which
decoded reference frames are used as clean conditioning.
### Synchronous Generation ### Synchronous Generation
```bash ```bash
@@ -184,10 +146,7 @@ export VLLM_OMNI_STORAGE_PATH=/var/tmp/vllm-omni-videos
## Model-Specific Examples ## Model-Specific Examples
For complete text-to-video, image-to-video, and model-specific video-to-video For complete text-to-video and image-to-video walkthroughs, see:
walkthroughs, see:
- [Text-to-Video](../user_guide/examples/online_serving/text_to_video.md) - [Text-to-Video](../user_guide/examples/online_serving/text_to_video.md)
- [Image-to-Video](../user_guide/examples/online_serving/image_to_video.md) - [Image-to-Video](../user_guide/examples/online_serving/image_to_video.md)
- [Cosmos3 recipes](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Nano.md)
for model-specific video-to-video examples and conditioning controls
+1 -2
View File
@@ -145,10 +145,9 @@ The following tables show which models support each feature:
| Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution | | Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
|------------------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-----------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:| |------------------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-----------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:|
| **Wan2.2** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (encode/decode) | ❌ | ❌ | | **Wan2.2** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (encode/decode) | ❌ | ❌ |
| **Wan2.2-S2V** | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ (encode/decode) | ❌ | ❌ |
| **Wan2.1-VACE** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ | | **Wan2.1-VACE** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
| **LTX-2** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | **LTX-2** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| **LTX-2.3** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | | **LTX-2.3** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| **Helios** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | **Helios** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
| **HunyuanVideo-1.5 T2V I2V** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ (encode/decode) | ✅ | ❌ | | **HunyuanVideo-1.5 T2V I2V** | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ (encode/decode) | ✅ | ❌ |
| **DreamID-Omni** | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | | **DreamID-Omni** | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-44
View File
@@ -28,50 +28,6 @@ in deep DiT blocks.
Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
guide. FP8 on Ampere may use a weight-only path where available. guide. FP8 on Ampere may use a weight-only path where available.
### Faster FP8 GEMM on Blackwell (quack)
On Blackwell (SM 100+), vLLM runs FP8 linears through the FlashInfer kernel, which
applies the bias as a separate kernel after the GEMM. On the small GEMMs in video
DiTs this bias add is a significant overhead. Installing the optional `quack` kernel
lets vLLM-Omni fuse `alpha * (A @ B) + bias` into a single CuteDSL GEMM, recovering
that overhead (e.g. HunyuanVideo-1.5 FP8 goes from slower-than-BF16 to faster).
```bash
# CUDA 12.9
pip install vllm-omni[quack]
# CUDA 13.x
pip install 'quack-kernels[cu13]' --extra-index-url https://download.pytorch.org/whl/cu130
```
It is enabled automatically once installed (no flag needed) and is **Blackwell-only**:
on Hopper/Ada the CUTLASS FP8 kernel already fuses bias, so quack is not used there.
Set `VLLM_OMNI_USE_QUACK_FP8=0` to force the FlashInfer path. If `quack-kernels` is
not installed, FP8 still works — it just keeps the unfused FlashInfer path.
#### Compile cache and warmup
quack JIT-compiles its kernel once per distinct GEMM shape (tens of seconds, longer
the first time across all autotuned configs). The compiled `.o` files are cached on
disk and reused on later runs, so this is a one-time cost — **not** per request.
vLLM-Omni points that cache at `~/.cache/vllm_omni/quack` (override with
`QUACK_CACHE_DIR`) instead of quack's default under `/tmp`, so it survives restarts.
In containers, set `QUACK_CACHE_DIR` to a mounted/persistent path — or bake it into
the image — so the first cold start does not recompile. The engine's startup dummy
run already exercises the kernels, so with a warm cache the first real request is fast.
To pre-warm specific shapes (e.g. at image build time):
```python
from vllm_omni.quantization.quack_fp8 import warmup_quack_fp8
# (M, K, N) per linear; M = number of tokens for your resolution/frame count
warmup_quack_fp8([(14040, 2048, 6144), (14040, 2048, 2048)])
```
> The PyPI package is `quack-kernels` (imported as `quack`); plain `pip install
> quack` is an unrelated statistics library. Requires CUDA 12.9+ and Python 3.12.
## Model Type Support ## Model Type Support
### Diffusion Model (Qwen-Image, Wan2.2) ### Diffusion Model (Qwen-Image, Wan2.2)
@@ -1,14 +0,0 @@
{
"prompt": "A robotic arm, primarily white with black joints and cables, is shown in a clean, modern indoor setting with a white tabletop. The arm, equipped with a gripper holding a small, light green pitcher, is positioned above a clear glass containing a reddish-brown liquid and a spoon. The robotic arm is in the process of pouring a transparent liquid into the glass. To the left of the pitcher, there is an opened jar with a similar reddish-brown substance visible through its transparent body. In the background, a vase with white flowers and a brown couch are partially visible, adding to the contemporary ambiance. The lighting is bright, casting soft shadows on the table. The robotic arm's movements are smooth and controlled, demonstrating precision in its task. As the video progresses, the robotic arm completes the pour, leaving the glass half-filled with the reddish-brown liquid. The jar remains untouched throughout the sequence, and the spoon inside the glass remains stationary. The other robotic arm on the right side also stays stationary throughout the video. The final frame captures the robotic arm with the pitcher finishing the pour, with the glass now filled to a higher level, while the pitcher is slightly tilted but still held securely by the gripper.",
"negative_prompt": "",
"vision_path": "https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_pouring.mp4",
"height": 720,
"width": 1280,
"num_frames": 189,
"num_inference_steps": 35,
"guidance_scale": 6.0,
"flow_shift": 10.0,
"fps": 24,
"condition_frame_indexes_vision": [0, 1],
"condition_video_keep": "first"
}
@@ -1,304 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import argparse
import uuid
from pathlib import Path
import cv2
import numpy as np
import torch
from PIL import Image
from vllm_omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
from vllm_omni.outputs import OmniRequestOutput
WORKER_EXTENSION = "vllm_omni.diffusion.models.dreamzero.video_export_worker.DreamZeroVideoExportWorkerExtension"
ASSET_REPO_ID = "YangshenDeng/vllm-omni-dreamzero-assets"
DEFAULT_SESSION_PREFIX = "dreamzero-export"
RELATIVE_OFFSETS = [-23, -16, -8, 0]
ACTION_HORIZON = 24
CAMERA_FILES = {
"observation/exterior_image_0_left": "exterior_image_1_left.mp4",
"observation/exterior_image_1_left": "exterior_image_2_left.mp4",
"observation/wrist_image_left": "wrist_image_left.mp4",
}
def _parse_args() -> argparse.Namespace:
repo_root = Path(__file__).resolve().parents[3]
parser = argparse.ArgumentParser(description="Export DreamZero prediction video from downloaded example inputs.")
parser.add_argument("--model", default="GEAR-Dreams/DreamZero-DROID")
parser.add_argument("--deploy-config", type=Path, required=True)
parser.add_argument(
"--video-dir",
type=Path,
default=repo_root / "outputs" / "dreamzero" / "assets",
help="Directory containing the three camera MP4 files.",
)
parser.add_argument(
"--output-dir", type=Path, default=repo_root / "outputs" / "dreamzero" / "generated_predictions"
)
parser.add_argument("--output-stem", default="dreamzero_prediction")
parser.add_argument(
"--prompt",
default="Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan",
)
parser.add_argument("--session-id", default=None)
parser.add_argument("--save-input-video", action="store_true")
parser.add_argument("--save-gif", action="store_true")
parser.add_argument("--save-actions", action="store_true")
parser.add_argument("--fps", type=int, default=5)
return parser.parse_args()
def _load_all_frames(video_path: Path) -> np.ndarray:
cap = cv2.VideoCapture(str(video_path))
frames = []
while True:
ok, frame = cap.read()
if not ok:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
cap.release()
if not frames:
raise RuntimeError(f"No frames loaded from {video_path}")
return np.stack(frames, axis=0)
def _load_camera_frames(video_dir: Path) -> dict[str, np.ndarray]:
camera_frames: dict[str, np.ndarray] = {}
for camera_key, file_name in CAMERA_FILES.items():
video_path = video_dir / file_name
if not video_path.exists():
raise FileNotFoundError(
f"Missing DreamZero example asset: {video_path}. "
"Download the example videos with: "
f"`hf download {ASSET_REPO_ID} --repo-type dataset --local-dir {video_dir}`"
)
camera_frames[camera_key] = _load_all_frames(video_path)
return camera_frames
def _build_frame_schedule(total_frames: int, num_chunks: int) -> list[list[int]]:
chunks: list[list[int]] = []
current_frame = 23
for _ in range(num_chunks):
indices = [max(current_frame + offset, 0) for offset in RELATIVE_OFFSETS]
if indices[-1] >= total_frames:
break
chunks.append(indices)
current_frame += ACTION_HORIZON
return chunks
def _make_obs_from_video(
camera_frames: dict[str, np.ndarray],
frame_indices: list[int],
*,
prompt: str,
session_id: str,
) -> dict:
obs: dict = {}
for camera_key, all_frames in camera_frames.items():
selected = all_frames[frame_indices]
obs[camera_key] = selected[0] if len(frame_indices) == 1 else selected
obs["observation/joint_position"] = np.zeros(7, dtype=np.float32)
obs["observation/cartesian_position"] = np.zeros(6, dtype=np.float32)
obs["observation/gripper_position"] = np.zeros(1, dtype=np.float32)
obs["prompt"] = prompt
obs["session_id"] = session_id
return obs
def _build_observations(video_dir: Path, prompt: str, session_id: str) -> tuple[dict[str, np.ndarray], list[dict]]:
camera_frames = _load_camera_frames(video_dir)
total_frames = min(frames.shape[0] for frames in camera_frames.values())
chunks = _build_frame_schedule(total_frames, 1)
observations = [
_make_obs_from_video(camera_frames, [0], prompt=prompt, session_id=session_id),
]
if chunks:
observations.append(
_make_obs_from_video(
camera_frames,
chunks[0],
prompt=prompt,
session_id=session_id,
)
)
if len(observations) < 2:
raise RuntimeError("Need at least two DreamZero example observations to export a prediction video.")
return camera_frames, observations[:2]
def _extract_latents(output: OmniRequestOutput) -> torch.Tensor:
if not isinstance(output, OmniRequestOutput):
raise TypeError(f"Expected OmniRequestOutput, got {type(output)!r}")
if not output.images:
raise RuntimeError("DreamZero output does not contain video latents in `images`.")
latents = output.images[0]
if not isinstance(latents, torch.Tensor):
raise TypeError(f"Expected tensor latents, got {type(latents)!r}")
latents = latents.detach().cpu()
if latents.dim() == 4:
latents = latents.unsqueeze(0)
if latents.dim() != 5:
raise ValueError(f"Unexpected latent shape: {tuple(latents.shape)}")
if latents.shape[1] < latents.shape[2]:
latents = latents.transpose(1, 2).contiguous()
return latents
def _write_mp4(path: Path, frames: np.ndarray, fps: int) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
height, width = frames.shape[1:3]
writer = cv2.VideoWriter(
str(path),
cv2.VideoWriter_fourcc(*"mp4v"),
float(fps),
(width, height),
)
if not writer.isOpened():
raise RuntimeError(f"Failed to open video writer for {path}")
try:
for frame in frames:
writer.write(cv2.cvtColor(frame, cv2.COLOR_RGB2BGR))
finally:
writer.release()
def _write_gif(path: Path, frames: np.ndarray, fps: int) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
images = [Image.fromarray(frame) for frame in frames]
duration_ms = max(int(round(1000 / max(fps, 1))), 1)
images[0].save(
path,
save_all=True,
append_images=images[1:],
duration=duration_ms,
loop=0,
)
def _stitch_input_frames(camera_frames: dict[str, np.ndarray]) -> np.ndarray:
total_frames = min(frames.shape[0] for frames in camera_frames.values())
stitched = []
for frame_index in range(total_frames):
left = camera_frames["observation/exterior_image_0_left"][frame_index]
right = camera_frames["observation/exterior_image_1_left"][frame_index]
wrist = camera_frames["observation/wrist_image_left"][frame_index]
pad = np.zeros((left.shape[0], left.shape[1], 3), dtype=np.uint8)
canvas = np.concatenate([left, right], axis=1)
bottom = np.concatenate([wrist, pad], axis=1)
stitched.append(np.concatenate([canvas, bottom], axis=0))
return np.stack(stitched, axis=0)
def _run_generation(
model: str, deploy_config_path: Path, observations: list[dict]
) -> tuple[Omni, list[OmniRequestOutput]]:
omni = Omni(
model=model,
deploy_config=str(deploy_config_path),
enforce_eager=True,
worker_extension_cls=WORKER_EXTENSION,
)
outputs: list[OmniRequestOutput] = []
for index, obs in enumerate(observations):
sampling_params = OmniDiffusionSamplingParams(
extra_args={
"reset": index == 0,
"session_id": obs["session_id"],
"robot_obs": obs,
}
)
result = omni.generate(obs["prompt"], sampling_params_list=[sampling_params])
if not result:
raise RuntimeError(f"No output returned for DreamZero request {index}")
outputs.append(result[0])
return omni, outputs
def _decode_with_worker(omni: Omni, full_latents: torch.Tensor) -> np.ndarray:
stage_client = omni.engine.stage_clients[0]
engine = getattr(stage_client, "_engine", None)
if engine is None:
raise RuntimeError("DreamZero export requires inline diffusion stage access.")
decoded = engine.executor.collective_rpc(
"decode_video_latents_to_uint8",
args=(full_latents,),
unique_reply_rank=0,
exec_all_ranks=True,
)
if isinstance(decoded, torch.Tensor):
decoded = decoded.numpy()
if not isinstance(decoded, np.ndarray):
raise TypeError(f"Unexpected decoded output type: {type(decoded)!r}")
return decoded
def main() -> None:
args = _parse_args()
session_id = args.session_id or f"{DEFAULT_SESSION_PREFIX}-{uuid.uuid4()}"
camera_frames, observations = _build_observations(
video_dir=args.video_dir,
prompt=args.prompt,
session_id=session_id,
)
args.output_dir.mkdir(parents=True, exist_ok=True)
if args.save_input_video:
input_frames = _stitch_input_frames(camera_frames)
_write_mp4(args.output_dir / f"{args.output_stem}_input.mp4", input_frames, fps=15)
if args.save_gif:
_write_gif(args.output_dir / f"{args.output_stem}_input.gif", input_frames[::3], fps=5)
omni = None
try:
omni, outputs = _run_generation(
model=args.model,
deploy_config_path=args.deploy_config,
observations=observations,
)
latent_steps = [_extract_latents(output) for output in outputs]
full_latents = torch.cat(latent_steps, dim=2)
frames = _decode_with_worker(omni, full_latents)
finally:
if omni is not None:
omni.close()
mp4_path = args.output_dir / f"{args.output_stem}.mp4"
_write_mp4(mp4_path, frames, fps=args.fps)
print(f"SAVED_MP4={mp4_path}")
if args.save_gif:
gif_path = args.output_dir / f"{args.output_stem}.gif"
_write_gif(gif_path, frames, fps=args.fps)
print(f"SAVED_GIF={gif_path}")
if args.save_actions:
npz_path = args.output_dir / f"{args.output_stem}_actions.npz"
np.savez(
npz_path,
step0=np.asarray(outputs[0].multimodal_output.get("actions")),
step1=np.asarray(outputs[1].multimodal_output.get("actions")),
)
print(f"SAVED_ACTIONS={npz_path}")
if __name__ == "__main__":
main()
@@ -42,7 +42,7 @@ pip install \
## 3. Hardware and VRAM Requirements ## 3. Hardware and VRAM Requirements
This example uses a 3-stage pipeline on one GPU by default This example uses a 3-stage pipeline on one GPU by default
([`dynin_omni.yaml`](../../../vllm_omni/deploy/dynin_omni.yaml)): ([`dynin_omni.yaml`](../../../vllm_omni/model_executor/stage_configs/dynin_omni.yaml)):
- Stage-0 (`token2text`): `gpu_memory_utilization: 0.5` - Stage-0 (`token2text`): `gpu_memory_utilization: 0.5`
- Stage-1 (`token2image`): `gpu_memory_utilization: 0.1` - Stage-1 (`token2image`): `gpu_memory_utilization: 0.1`
@@ -878,8 +878,8 @@ def parse_args(repo_root: Path) -> argparse.Namespace:
parser.add_argument( parser.add_argument(
"--stage-config-path", "--stage-config-path",
type=str, type=str,
default=str(repo_root / "vllm_omni/deploy/dynin_omni.yaml"), default=str(repo_root / "vllm_omni/model_executor/stage_configs/dynin_omni.yaml"),
help="Path to deploy config yaml.", help="Path to stage config yaml.",
) )
parser.add_argument( parser.add_argument(
"--dynin-config-path", "--dynin-config-path",
@@ -16,10 +16,9 @@ Usage:
""" """
import argparse import argparse
import json import os
import time import time
from pathlib import Path from pathlib import Path
from typing import Any
import torch import torch
@@ -30,16 +29,6 @@ from vllm_omni.outputs import OmniRequestOutput
from vllm_omni.platforms import current_omni_platform from vllm_omni.platforms import current_omni_platform
def parse_profiler_config(value: str) -> dict[str, Any]:
try:
config = json.loads(value)
except json.JSONDecodeError as e:
raise argparse.ArgumentTypeError(f"--profiler-config must be valid JSON: {e}") from e
if not isinstance(config, dict):
raise argparse.ArgumentTypeError("--profiler-config must be a JSON object")
return config
def parse_args() -> argparse.Namespace: def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser( parser = argparse.ArgumentParser(
description="Generate a talking-head video from a reference image and audio (Wan2.2 S2V)." description="Generate a talking-head video from a reference image and audio (Wan2.2 S2V)."
@@ -98,12 +87,6 @@ def parse_args() -> argparse.Namespace:
default=3.0, default=3.0,
help="Scheduler flow shift (default: 3.0).", help="Scheduler flow shift (default: 3.0).",
) )
parser.add_argument(
"--boundary-ratio",
type=float,
default=None,
help="Boundary split ratio for low/high DiT stages (S2V). Default varies by model.",
)
parser.add_argument( parser.add_argument(
"--output", "--output",
type=str, type=str,
@@ -195,10 +178,30 @@ def parse_args() -> argparse.Namespace:
help="Enable cache-dit summary logging after diffusion forward passes.", help="Enable cache-dit summary logging after diffusion forward passes.",
) )
parser.add_argument( parser.add_argument(
"--profiler-config", "--profile-dir",
type=parse_profiler_config, type=str,
default=None, default=None,
help='JSON profiler config for torch/cuda profiling, e.g. \'{"profiler":"torch","torch_profiler_dir":"./perf"}\'.', help="Enable torch profiler and save traces to this directory.",
)
parser.add_argument(
"--profile-record-shapes",
action="store_true",
help="Record tensor shapes in profiler (increases trace size).",
)
parser.add_argument(
"--profile-with-stack",
action="store_true",
help="Record stack traces in profiler (increases overhead).",
)
parser.add_argument(
"--profile-with-memory",
action="store_true",
help="Profile memory usage.",
)
parser.add_argument(
"--profile-with-flops",
action="store_true",
help="Estimate FLOPs for operations.",
) )
return parser.parse_args() return parser.parse_args()
@@ -236,9 +239,23 @@ def main():
ulysses_degree=args.ulysses_degree, ulysses_degree=args.ulysses_degree,
) )
profiler_enabled = args.profiler_config is not None # Check if profiling is requested via CLI args or environment variable
profile_dir = args.profile_dir or os.getenv("VLLM_TORCH_PROFILER_DIR")
profiler_enabled = bool(profile_dir)
profiler_config = None
if profiler_enabled:
from vllm.config import ProfilerConfig
omni_kwargs = dict( profiler_config = ProfilerConfig(
profiler="torch",
torch_profiler_dir=profile_dir,
torch_profiler_record_shapes=args.profile_record_shapes,
torch_profiler_with_stack=args.profile_with_stack,
torch_profiler_with_memory=args.profile_with_memory,
torch_profiler_with_flops=args.profile_with_flops,
)
omni = Omni(
model=args.model, model=args.model,
model_class_name="WanS2VPipeline", model_class_name="WanS2VPipeline",
flow_shift=args.flow_shift, flow_shift=args.flow_shift,
@@ -252,12 +269,8 @@ def main():
cache_backend=args.cache_backend, cache_backend=args.cache_backend,
cache_config=cache_config, cache_config=cache_config,
enable_cache_dit_summary=args.enable_cache_dit_summary, enable_cache_dit_summary=args.enable_cache_dit_summary,
profiler_config=args.profiler_config, profiler_config=profiler_config,
) )
if args.boundary_ratio is not None:
omni_kwargs["boundary_ratio"] = args.boundary_ratio
omni = Omni(**omni_kwargs)
# Print generation configuration # Print generation configuration
print(f"\n{'=' * 60}") print(f"\n{'=' * 60}")
@@ -312,16 +325,13 @@ def main():
if profiler_enabled: if profiler_enabled:
print("\n[Profiler] Stopping profiler and collecting results...") print("\n[Profiler] Stopping profiler and collecting results...")
profile_results = omni.stop_profile() profile_results = omni.stop_profile()
if profile_results and isinstance(profile_results, dict): if profile_results and isinstance(profile_results, list):
traces = profile_results.get("traces", [])
print("\n" + "=" * 60) print("\n" + "=" * 60)
print("PROFILING RESULTS:") print("PROFILING RESULTS:")
for rank, trace in enumerate(traces): for i, result in enumerate(profile_results):
print(f"\nRank {rank}:") print(f"\nStage {i}:")
if trace: if result:
print(f" • Trace: {trace}") print(f" {result}")
if not traces:
print(" No traces collected.")
print("=" * 60) print("=" * 60)
else: else:
print("[Profiler] No valid profiling data returned.") print("[Profiler] No valid profiling data returned.")
@@ -1,204 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Offline higgs-audio v3 inference example.
Runs Stage 0 (Qwen3 talker) + Stage 1 (HiggsAudio codec) end-to-end
through the vLLM-Omni engine without going through the HTTP server, and
saves a 24 kHz mono WAV per prompt.
Example (plain TTS):
python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \\
--texts "Hello world." \\
--output-dir results/higgs_v3_wavs
Example (voice clone):
python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \\
--texts "Hello world." \\
--ref-audio path/to/reference.wav \\
--ref-text "Transcript of the reference clip." \\
--output-dir results/higgs_v3_wavs
"""
from __future__ import annotations
import os
os.environ.setdefault("VLLM_USE_DEEP_GEMM", "0")
os.environ.setdefault("VLLM_MOE_USE_DEEP_GEMM", "0")
import re
import time
from pathlib import Path
import numpy as np
import soundfile as sf
import torch
from vllm_omni import Omni
SAMPLE_RATE = 24_000
DEFAULT_TEXTS = (
"Hello world.",
"The quick brown fox jumps over the lazy dog.",
"Today is a beautiful day for a walk in the park.",
)
def parse_args():
import argparse
parser = argparse.ArgumentParser(description="Offline higgs-audio v3 inference")
parser.add_argument(
"--model",
type=str,
default="bosonai/higgs-audio-v3-tts-4b",
help="Stage-0 talker model id or path.",
)
parser.add_argument(
"--texts",
type=str,
nargs="+",
default=list(DEFAULT_TEXTS),
help="One or more plain-text prompts to synthesize.",
)
parser.add_argument(
"--output-dir",
type=str,
default="results/higgs_v3_wavs",
help="Directory to write per-prompt WAV files.",
)
parser.add_argument(
"--deploy-config",
type=str,
default=None,
help="Override the deploy config path.",
)
parser.add_argument(
"--max-new-tokens",
type=int,
default=2048,
help="Cap on Stage-0 codec frames per request.",
)
parser.add_argument(
"--ref-audio",
type=str,
default=None,
help="Reference audio for voice clone (WAV/FLAC/MP3 path).",
)
parser.add_argument(
"--ref-text",
type=str,
default=None,
help="Transcript of the reference audio. Optional but improves fidelity.",
)
return parser.parse_args()
def _slugify(text: str) -> str:
slug = re.sub(r"\s+", "_", text.strip().lower())
slug = re.sub(r"[^a-z0-9_]+", "", slug)
return slug[:48] or "prompt"
def _extract_pcm(multimodal_output: dict) -> torch.Tensor:
audio = multimodal_output.get("model_outputs")
if audio is None:
audio = multimodal_output.get("audio")
if audio is None:
raise ValueError(f"no audio key in multimodal_output: {list(multimodal_output.keys())}")
if isinstance(audio, list):
valid = [torch.as_tensor(a).float().cpu().reshape(-1) for a in audio if a is not None]
if not valid:
raise ValueError("audio list is empty")
return torch.cat(valid, dim=0) if len(valid) > 1 else valid[0]
return torch.as_tensor(audio).float().cpu().reshape(-1)
def _pcm_to_int16(pcm: torch.Tensor) -> np.ndarray:
arr = pcm.numpy()
if arr.dtype.kind == "f":
arr = np.clip(arr, -1.0, 1.0)
arr = (arr * 32767.0).astype(np.int16)
else:
arr = arr.astype(np.int16)
return arr
def main():
args = parse_args()
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
engine = Omni(model=args.model, deploy_config=args.deploy_config, trust_remote_code=True)
from transformers import AutoTokenizer
from vllm_omni.model_executor.models.higgs_audio_v3.higgs_audio_v3_tokenizer import (
HiggsAudioV3TokenizerAdapter,
apply_delay_pattern,
encode_reference_audio,
)
tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
adapter = HiggsAudioV3TokenizerAdapter(tokenizer)
# Load and encode reference audio once (if voice cloning)
ref_codes_delayed: torch.Tensor | None = None
if args.ref_audio is not None:
ref_wav, ref_sr = sf.read(args.ref_audio, always_2d=False)
if ref_wav.ndim == 2:
ref_wav = ref_wav.mean(axis=1)
ref_codes_raw = encode_reference_audio(ref_wav, int(ref_sr))
ref_codes_delayed = apply_delay_pattern(ref_codes_raw)
print(f"Reference : {args.ref_audio}")
print(f"Ref codes : {ref_codes_raw.shape[0]} frames -> {ref_codes_delayed.shape[0]} delayed")
if args.ref_text:
print(f"Ref text : {args.ref_text}")
print(f"Model : {args.model}")
print(f"Prompts : {len(args.texts)}")
print(f"Output dir : {output_dir}")
print(f"Voice clone : {'yes' if ref_codes_delayed is not None else 'no'}")
total_elapsed = 0.0
total_dur = 0.0
for text in args.texts:
if ref_codes_delayed is not None:
prompt_ids = adapter.build_prompt(
text,
num_ref_tokens=int(ref_codes_delayed.shape[0]),
reference_text=args.ref_text,
)
prompt = {
"prompt_token_ids": prompt_ids,
"additional_information": {
"audio_input_ids": ref_codes_delayed.to(torch.long),
"audio_input_ids_mask": torch.ones(ref_codes_delayed.shape[0], dtype=torch.bool),
},
}
else:
prompt_ids = adapter.build_prompt(text)
prompt = {"prompt_token_ids": prompt_ids}
t_start = time.perf_counter()
outputs = engine.generate([prompt])
elapsed = time.perf_counter() - t_start
total_elapsed += elapsed
mm = outputs[0].outputs[0].multimodal_output
pcm = _extract_pcm(mm)
slug = _slugify(text)
out_path = output_dir / f"{slug}.wav"
sf.write(str(out_path), _pcm_to_int16(pcm), SAMPLE_RATE, format="WAV", subtype="PCM_16")
dur = pcm.numel() / SAMPLE_RATE
total_dur += dur
print(f" {slug:<50} dur={dur:5.2f}s -> {out_path}")
rtf = total_elapsed / total_dur if total_dur > 0 else float("inf")
print(f"Total infer : {total_elapsed:.2f}s total audio: {total_dur:.2f}s RTF: {rtf:.3f}")
if __name__ == "__main__":
main()
-132
View File
@@ -1,132 +0,0 @@
# DreamZero
DreamZero is the robot-policy serving example for the OpenPI-compatible `/v1/realtime/robot/openpi` endpoint.
## Supported checkpoints
| Model | HuggingFace repo |
|---|---|
| DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` |
## Quick start
### Start the server
```bash
bash examples/online_serving/dreamzero/run_server.sh
```
By default this launches:
```bash
vllm serve GEAR-Dreams/DreamZero-DROID --omni --port 8000 \
--served-model-name dreamzero-droid \
--deploy-config vllm_omni/deploy/dreamzero_tp1_cfg2.yaml \
--enforce-eager --disable-log-stats
```
Override `MODEL`, `PORT`, `HOST`, `DEPLOY_CONFIG`, or `SERVED_MODEL_NAME` through the script environment if needed.
### Install example dependencies
The core `pip install -e .` setup does not include the extra packages used by the DreamZero example scripts.
- `openpi_client.py`:
`openpi-client`, `websockets`, `opencv-python`
- `droid_sim_eval_client.py`:
`mediapy`, `websockets`, `openpi-client`
The DROID sim-eval script also needs an Isaac Lab environment that provides `isaaclab`, `isaaclab_tasks`, `sim_evals`, and `gymnasium`.
If you run the DROID client on Python < 3.12, also install `typing-extensions`.
### Configure TP and CFG parallelism
The bundled DreamZero configs intentionally keep only:
| Config | Purpose |
|---|---|
| `vllm_omni/deploy/dreamzero.yaml` | Default TP=1, CFG parallel disabled |
| `vllm_omni/deploy/dreamzero_tp1_cfg2.yaml` | TP=1, CFG parallel size=2 |
For other topologies, use CLI parallelism flags and update stage 0 `devices` with `--stage-overrides`. The number of listed devices must match `tensor_parallel_size * cfg_parallel_size`.
TP=2 with CFG parallel disabled:
```bash
vllm serve GEAR-Dreams/DreamZero-DROID --omni --port 8000 \
--served-model-name dreamzero-droid \
--deploy-config vllm_omni/deploy/dreamzero.yaml \
--tensor-parallel-size 2 \
--stage-overrides '{"0": {"devices": "0,1"}}' \
--enforce-eager --disable-log-stats
```
TP=2 with CFG parallel size=2:
```bash
vllm serve GEAR-Dreams/DreamZero-DROID --omni --port 8000 \
--served-model-name dreamzero-droid \
--deploy-config vllm_omni/deploy/dreamzero.yaml \
--tensor-parallel-size 2 \
--cfg-parallel-size 2 \
--stage-overrides '{"0": {"devices": "0,1,2,3"}}' \
--enforce-eager --disable-log-stats
```
### Download example assets
The OpenPI client and DROID sim-eval example expect the three camera MP4 files in `outputs/dreamzero/assets`.
```bash
hf download YangshenDeng/vllm-omni-dreamzero-assets --repo-type dataset --local-dir outputs/dreamzero/assets
```
### Run the OpenPI client
```bash
python examples/online_serving/dreamzero/openpi_client.py \
--host 127.0.0.1 \
--port 8000 \
--video-dir outputs/dreamzero/assets
```
This client uses downloaded example videos and talks to the OpenPI websocket server.
### Run DROID sim eval
```bash
${ISAACLAB_LAUNCHER} -p examples/online_serving/dreamzero/droid_sim_eval_client.py \
--host 127.0.0.1 \
--port 8000 \
--scene 1 \
--episodes 1 \
--headless \
--device cuda:0
```
Set `ISAACLAB_LAUNCHER=path/to/isaaclab.sh` from the vLLM-Omni repository root before running the command.
This launches Isaac Lab / sim-evals and runs the DROID benchmark loop against the same websocket endpoint.
### Export comparison videos offline
```bash
python examples/offline_inference/dreamzero/export_prediction_video.py \
--deploy-config vllm_omni/deploy/dreamzero_tp1_cfg2.yaml \
--save-input-video \
--save-gif
```
The export script uses local `Omni` inference, not the websocket server. It writes the input rollout video and the predicted output video artifacts for side-by-side inspection.
### MolmoSpace demo
```bash
python examples/online_serving/dreamzero/molmospace_dreamzero_eval_demo.py \
--host 127.0.0.1 \
--port 8000 \
--benchmark_dir /path/to/benchmark \
--output_dir /path/to/output
```
This demo adapts DreamZero to the MolmoSpace-style remote policy eval loop.
@@ -1,819 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Run one or more DROID sim-eval rollouts against the vLLM DreamZero server.
This script is the vLLM/OpenPI adaptation of the upstream DreamZero
`eval_utils/run_sim_eval.py` sim-eval client.
Behavior intentionally kept close to upstream:
- same DROID observation extraction (`external_cam`, `external_cam_2`,
`wrist_cam`, joint position, gripper position)
- same resize-with-pad preprocessing to `(180, 320)`
- same `open_loop_horizon=8`
- same gripper binarization rule (`> 0.5 -> 1`, else `0`)
- same per-scene language prompts
Unlike upstream DreamZero, vLLM serves the compatible websocket policy endpoint
at `/v1/realtime/robot/openpi`, so this script includes the path suffix in the
client URI.
Run this script through Isaac Lab's launcher from the vLLM-Omni repository
root, for example:
"${ISAACLAB_LAUNCHER}" -p \
examples/online_serving/dreamzero/droid_sim_eval_client.py \
--host 127.0.0.1 \
--port 8000 \
--scene 1 \
--episodes 1 \
--headless \
--device cuda:1
"""
from __future__ import annotations
import argparse
import json
import logging
import sys
import time
import uuid
from dataclasses import dataclass
from datetime import UTC, datetime
from pathlib import Path
from typing import Any
import numpy as np
import torch
try:
import mediapy
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires `mediapy`.") from exc
try:
from typing import override
except ImportError:
try:
from typing_extensions import override
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires `typing-extensions` on Python < 3.12.") from exc
try:
import websockets.sync.client
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires `websockets`.") from exc
# NOTE:
# This directory already contains a local file named `openpi_client.py`.
# However, what we want here is the *installed* `openpi_client` package from
# upstream OpenPI, not the sibling example file. When a script is executed
# directly, Python often puts the script directory at `sys.path[0]`, which
# would cause `import openpi_client` to resolve to the local example file and
# create a circular import.
#
# To avoid that ambiguity, temporarily remove the current example directory
# from the front of `sys.path`, import the real package, and then restore the
# path afterwards.
example_dir = str(Path(__file__).resolve().parent)
removed_path = False
if sys.path and sys.path[0] == example_dir:
sys.path.pop(0)
removed_path = True
try:
from openpi_client import image_tools, msgpack_numpy
from openpi_client.base_policy import BasePolicy
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires the `openpi-client` package.") from exc
finally:
if removed_path:
sys.path.insert(0, example_dir)
# -----------------------------------------------------------------------------
# Constants
# -----------------------------------------------------------------------------
#
# These values intentionally mirror the upstream DreamZero sim-eval client
# where possible. The important distinction is:
#
# - ACTION_HORIZON = 24
# The model returns 24 future actions per inference call.
# - DEFAULT_OPEN_LOOP_HORIZON = 8
# The sim client only executes the first 8 actions locally before asking
# the server to replan from a fresh observation.
#
# So a single server call predicts 24x8 actions, but the rollout consumes only
# 8 of them before replanning.
PING_INTERVAL_SECS = 60
PING_TIMEOUT_SECS = 600
DEFAULT_PATH = "/v1/realtime/robot/openpi"
DEFAULT_OPEN_LOOP_HORIZON = 8
ACTION_HORIZON = 24
ACTION_DIM = 8
DEFAULT_OUTPUT_ROOT = Path("runs") / "dreamzero_sim_eval"
SCENE_PROMPTS = {
1: "put the cube in the bowl",
2: "pick up the can and put it in the mug",
3: "put the banana in the bin",
}
def _decode_action_response(response: bytes | str) -> np.ndarray:
if isinstance(response, str):
raise RuntimeError(f"Error in inference server:\n{response}")
decoded = msgpack_numpy.unpackb(response)
if isinstance(decoded, dict) and decoded.get("type") == "error":
message = decoded.get("message", decoded)
raise RuntimeError(f"Error in inference server:\n{message}")
return np.asarray(decoded, dtype=np.float32)
@dataclass(frozen=True)
class StepRecord:
"""One fully materialized rollout step for later JSON export.
The `episode_00.json` artifact is intended to be human-readable and
post-process-friendly. Instead of keeping raw tensors around, each step is
flattened into plain Python types so it can be serialized directly.
"""
# Index of this control step within the episode.
step_index: int
# Whether this step triggered a fresh model call (as opposed to reusing the
# cached open-loop chunk from the previous server response).
used_server_call: bool
# End-to-end latency of the server call that produced the current chunk.
# This is `None` on steps that only reuse cached actions.
chunk_latency_s: float | None
# The concrete 8-D action sent into the simulator at this step.
action: list[float]
# Observed 7-DoF arm joint positions before the next environment step.
joint_position: list[float]
# Observed gripper scalar before the next environment step.
gripper_position: list[float]
# Reward and termination signals directly returned by the simulator.
reward: float
terminated: bool
truncated: bool
# Optional scene object positions for downstream debugging / success
# heuristics. This may be empty if the environment does not expose them.
scene_objects: dict[str, list[float]]
class OpenPIWebsocketClientPolicy(BasePolicy):
"""Minimal websocket client for the DreamZero/OpenPI policy protocol.
Protocol shape:
- connect -> server immediately sends a metadata dict
- infer -> send msgpack observation, receive action chunk
- reset -> send msgpack reset command, receive confirmation string
This class intentionally stays very small because the more interesting
DreamZero-specific behavior lives one layer above, in
`DreamZeroJointPosClient`.
"""
def __init__(
self,
host: str = "127.0.0.1",
port: int = 8000,
path: str = DEFAULT_PATH,
) -> None:
# vLLM serves the robot endpoint under `/v1/realtime/robot/openpi`.
self._uri = f"ws://{host}:{port}{path}"
# Upstream protocol uses msgpack with numpy support, not JSON.
self._packer = msgpack_numpy.Packer()
# Connect immediately and cache the server handshake metadata.
self._ws, self._server_metadata = self._wait_for_server()
def get_server_metadata(self) -> dict[str, Any]:
"""Return a copy of the server handshake metadata."""
return dict(self._server_metadata)
def _wait_for_server(self):
"""Connect to the websocket server and read the initial metadata frame."""
logging.info("Connecting to %s", self._uri)
conn = websockets.sync.client.connect(
self._uri,
compression=None,
max_size=None,
ping_interval=PING_INTERVAL_SECS,
ping_timeout=PING_TIMEOUT_SECS,
)
metadata = msgpack_numpy.unpackb(conn.recv())
if not isinstance(metadata, dict):
raise TypeError(f"Expected dict metadata from server, got {type(metadata)!r}")
return conn, metadata
@override
def infer(self, obs: dict[str, Any]) -> np.ndarray:
"""Send an inference request and return the decoded action chunk."""
# Keep the upstream DreamZero/OpenPI convention that the request itself
# tells the server which logical endpoint is being called.
payload = dict(obs)
payload["endpoint"] = "infer"
self._ws.send(self._packer.pack(payload))
response = self._ws.recv()
return _decode_action_response(response)
@override
def reset(self, reset_info: dict[str, Any] | None = None) -> str:
"""Tell the server to reset its session-side state."""
payload = dict(reset_info or {})
payload["endpoint"] = "reset"
self._ws.send(self._packer.pack(payload))
response = self._ws.recv()
if isinstance(response, str):
return response
decoded = msgpack_numpy.unpackb(response)
if not isinstance(decoded, dict) or decoded.get("status") != "reset successful":
raise RuntimeError(f"Unexpected reset response: {decoded!r}")
return str(decoded["status"])
def close(self) -> None:
"""Close the websocket connection explicitly."""
self._ws.close()
class DreamZeroJointPosClient:
"""DROID sim-eval client that talks to the vLLM OpenPI websocket server.
This is the main compatibility layer between:
- Isaac Lab DROID observations (`obs["policy"][...]`)
- DreamZero/OpenPI websocket payloads
- local open-loop action reuse across several simulator steps
In other words:
simulator obs -> websocket request -> action chunk -> one action per step
"""
def __init__(
self,
remote_host: str = "127.0.0.1",
remote_port: int = 8000,
path: str = DEFAULT_PATH,
open_loop_horizon: int = DEFAULT_OPEN_LOOP_HORIZON,
) -> None:
# Low-level transport client.
self.client = OpenPIWebsocketClientPolicy(remote_host, remote_port, path=path)
# Number of actions to execute locally before replanning.
self.open_loop_horizon = open_loop_horizon
# Cursor into the currently cached action chunk.
self.actions_from_chunk_completed = 0
# Most recent `(ACTION_HORIZON, ACTION_DIM)` server response.
self.pred_action_chunk: np.ndarray | None = None
# Session id is part of the DreamZero serving contract. Changing it
# causes the server side to treat the rollout as a fresh episode.
self.session_id = str(uuid.uuid4())
# Simple runtime stats for reporting.
self.server_calls = 0
self.last_chunk_latency_s: float | None = None
self.last_used_server_call = False
def metadata(self) -> dict[str, Any]:
"""Expose the server metadata to callers / logs."""
return self.client.get_server_metadata()
def reset(self) -> str:
"""Reset local chunk state and remote session state.
Local reset:
- drop cached action chunk
- rewind chunk cursor
- allocate a fresh session id
Remote reset:
- send a websocket `reset` message so the server can clear any
request/session-side state it associates with this client
"""
self.actions_from_chunk_completed = 0
self.pred_action_chunk = None
self.session_id = str(uuid.uuid4())
self.last_chunk_latency_s = None
self.last_used_server_call = False
return self.client.reset({})
def infer(self, obs: dict[str, Any], instruction: str) -> dict[str, Any]:
"""Turn one simulator observation into one executable 8-D action.
Key behavior:
- call the server only when the local chunk cache is empty/exhausted
- otherwise, keep consuming the cached chunk open-loop
- always return exactly one 8-D action for the current simulator step
"""
# Convert Isaac Lab observation structure into a plain numpy-friendly
# record that is easier to serialize and visualize.
curr_obs = self._extract_observation(obs)
self.last_used_server_call = False
# Replan if:
# 1. this is the first step of a rollout / chunk
# 2. we already consumed `open_loop_horizon` actions from the current chunk
# 3. no cached chunk is currently available
if (
self.actions_from_chunk_completed == 0
or self.actions_from_chunk_completed >= self.open_loop_horizon
or self.pred_action_chunk is None
):
self.actions_from_chunk_completed = 0
# Build the exact DreamZero/OpenPI payload expected by the server.
#
# Notes:
# - images are resized/padded to the serving contract's 180x320
# - proprio is cast to float64 to match upstream client behavior
# - cartesian_position is currently unused by DreamZero DROID, so
# a dummy zero vector is sent for protocol completeness
request_data = {
"observation/exterior_image_0_left": image_tools.resize_with_pad(curr_obs["right_image"], 180, 320),
"observation/exterior_image_1_left": image_tools.resize_with_pad(curr_obs["left_image"], 180, 320),
"observation/wrist_image_left": image_tools.resize_with_pad(curr_obs["wrist_image"], 180, 320),
"observation/joint_position": curr_obs["joint_position"].astype(np.float64),
"observation/cartesian_position": np.zeros((6,), dtype=np.float64),
"observation/gripper_position": curr_obs["gripper_position"].astype(np.float64),
"prompt": instruction,
"session_id": self.session_id,
}
# Measure end-to-end server latency for this chunk request.
start = time.perf_counter()
actions = self.client.infer(request_data)
self.last_chunk_latency_s = time.perf_counter() - start
self.last_used_server_call = True
self.server_calls += 1
# DreamZero DROID serving is expected to return an action chunk with
# 24 future actions, each action being 8-D.
if actions.ndim != 2:
raise AssertionError(f"Expected 2D action array, got shape {actions.shape}")
if actions.shape != (ACTION_HORIZON, ACTION_DIM):
raise AssertionError(f"Expected action shape {(ACTION_HORIZON, ACTION_DIM)}, got {actions.shape}")
self.pred_action_chunk = actions
# Consume exactly one action row from the cached chunk for this
# simulator step.
action = np.array(self.pred_action_chunk[self.actions_from_chunk_completed], copy=True)
self.actions_from_chunk_completed += 1
# Upstream DreamZero sim-eval binarizes the gripper command.
action[-1] = 1.0 if action[-1].item() > 0.5 else 0.0
# Produce a human-friendly visualization strip for videos:
# right external | wrist | left external
img1 = image_tools.resize_with_pad(curr_obs["right_image"], 224, 224)
img2 = image_tools.resize_with_pad(curr_obs["wrist_image"], 224, 224)
img3 = image_tools.resize_with_pad(curr_obs["left_image"], 224, 224)
viz = np.concatenate([img1, img2, img3], axis=1)
# Return both the executable action and auxiliary debug info.
return {
"action": action,
"viz": viz,
"joint_position": curr_obs["joint_position"],
"gripper_position": curr_obs["gripper_position"],
"used_server_call": self.last_used_server_call,
"chunk_latency_s": self.last_chunk_latency_s if self.last_used_server_call else None,
}
@staticmethod
def _extract_observation(obs_dict: dict[str, Any]) -> dict[str, np.ndarray]:
"""Extract the pieces DreamZero cares about from Isaac Lab observations.
`sim-evals` exposes camera frames and robot state inside the
`obs["policy"]` group. This helper converts those tensors into numpy
arrays so they can be fed into image preprocessing / websocket packing.
"""
policy = obs_dict["policy"]
# Isaac Lab stores camera observations as batched tensors; use env 0.
right_image = policy["external_cam"][0].clone().detach().cpu().numpy()
left_image = policy["external_cam_2"][0].clone().detach().cpu().numpy()
wrist_image = policy["wrist_cam"][0].clone().detach().cpu().numpy()
# Robot proprioception.
joint_position = policy["arm_joint_pos"].clone().detach().cpu().numpy()
gripper_position = policy["gripper_pos"].clone().detach().cpu().numpy()
return {
"right_image": right_image,
"left_image": left_image,
"wrist_image": wrist_image,
"joint_position": joint_position,
"gripper_position": gripper_position,
}
def _scene_instruction(scene: int) -> str:
"""Map a numeric scene id onto the fixed language prompt used for rollout."""
try:
return SCENE_PROMPTS[scene]
except KeyError as exc:
raise ValueError(f"Unsupported scene {scene}. Available scenes: {sorted(SCENE_PROMPTS)}") from exc
def _capture_scene_objects(env: Any) -> dict[str, list[float]]:
"""Best-effort extraction of scene object root positions.
This is only for debugging / reporting. The rollout logic itself does not
depend on these positions.
"""
objects: dict[str, list[float]] = {}
scene = getattr(env, "scene", None)
if scene is None:
return objects
# Skip non-task entities such as cameras, the robot, and lighting.
for name in scene.keys():
if name in {"robot", "external_cam", "external_cam_2", "wrist_cam", "sphere_light", "scene"}:
continue
entity = scene[name]
data = getattr(entity, "data", None)
root_pos_w = getattr(data, "root_pos_w", None)
if root_pos_w is None:
continue
# Convert tensors to plain lists for JSON serialization.
value = root_pos_w[0].detach().cpu().to(torch.float32).tolist()
objects[str(name)] = [float(x) for x in value]
return objects
def _maybe_infer_success(scene: int, final_objects: dict[str, list[float]]) -> dict[str, Any]:
"""Best-effort geometric heuristic.
The simulator itself does not expose a built-in success term; this function
provides a transparent fallback for human-readable reporting only.
"""
task_pairs = {
1: ("cube", "bowl"),
2: ("can", "mug"),
3: ("banana", "bin"),
}
source_name, target_name = task_pairs.get(scene, (None, None))
if source_name not in final_objects or target_name not in final_objects:
return {
"has_builtin_success": False,
"heuristic_success": None,
"reason": "scene object names unavailable for heuristic",
}
source = np.asarray(final_objects[source_name], dtype=np.float32)
target = np.asarray(final_objects[target_name], dtype=np.float32)
xy_distance = float(np.linalg.norm(source[:2] - target[:2]))
z_delta = float(source[2] - target[2])
heuristic_success = bool(xy_distance < 0.12 and z_delta > -0.08)
return {
"has_builtin_success": False,
"heuristic_success": heuristic_success,
"xy_distance": xy_distance,
"z_delta": z_delta,
"source_object": source_name,
"target_object": target_name,
"reason": "no built-in env success flag; using final object pose heuristic",
}
def _dump_json(path: Path, payload: dict[str, Any]) -> None:
"""Write a JSON file with stable UTF-8 formatting."""
path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
def _scalar_from_env_value(value: Any) -> float:
"""Normalize simulator scalar outputs into a plain Python float.
Isaac Lab / Gym values may come back as tensors, numpy arrays, tuples, or
direct Python scalars depending on the wrapper stack. Centralizing the
conversion here makes the rollout loop cleaner and more robust.
"""
if isinstance(value, torch.Tensor):
return float(value.reshape(-1)[0].detach().cpu().item())
if isinstance(value, np.ndarray):
return float(value.reshape(-1)[0])
if isinstance(value, (list, tuple)):
return float(np.asarray(value).reshape(-1)[0])
return float(value)
def _bool_from_env_value(value: Any) -> bool:
"""Normalize simulator boolean-like outputs into a plain Python bool."""
if isinstance(value, torch.Tensor):
return bool(value.reshape(-1)[0].detach().cpu().item())
if isinstance(value, np.ndarray):
return bool(value.reshape(-1)[0])
if isinstance(value, (list, tuple)):
return bool(np.asarray(value).reshape(-1)[0])
return bool(value)
def _make_output_dir(output_root: Path, scene: int) -> Path:
"""Create a timestamped output directory for one sim-eval run."""
timestamp = datetime.now(UTC).strftime("%Y%m%dT%H%M%SZ")
output_dir = output_root / f"scene{scene}_{timestamp}"
output_dir.mkdir(parents=True, exist_ok=True)
return output_dir
def main() -> None:
"""Entry point for one or more DROID simulation rollouts.
High-level flow:
1. parse command-line flags
2. bootstrap Isaac Lab / sim-evals imports
3. create the DROID environment
4. connect the DreamZero websocket client
5. run `episodes` rollouts
6. export videos + JSON summaries
"""
# Script-level arguments.
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--episodes", type=int, default=1, help="Number of episodes to run.")
parser.add_argument("--scene", type=int, default=1, help="DROID scene id (1/2/3).")
parser.add_argument("--host", type=str, default="127.0.0.1", help="vLLM DreamZero server host.")
parser.add_argument("--port", type=int, default=8000, help="vLLM DreamZero server port.")
parser.add_argument("--path", type=str, default=DEFAULT_PATH, help="Websocket path suffix.")
parser.add_argument(
"--open-loop-horizon",
type=int,
default=DEFAULT_OPEN_LOOP_HORIZON,
help="How many actions to consume locally before requesting the next chunk.",
)
parser.add_argument(
"--output-root",
type=Path,
default=DEFAULT_OUTPUT_ROOT,
help="Directory where videos and trajectory logs are stored.",
)
try:
from isaaclab.app import AppLauncher
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError(
"Optional DROID sim-eval client requires Isaac Lab (`isaaclab`). "
"Launch it from an Isaac Lab environment, e.g. via `isaaclab.sh -p`."
) from exc
# Let Isaac Lab inject its own runtime flags (e.g. headless, device).
AppLauncher.add_app_launcher_args(parser)
args = parser.parse_args()
# DreamZero sim-eval always needs camera observations enabled.
args.enable_cameras = True
# Boot Isaac Sim / Isaac Lab.
app_launcher = AppLauncher(args)
simulation_app = app_launcher.app
# Set defaults so the `finally` block can clean up safely even if an
# earlier step fails.
env = None
client = None
# Import simulator modules only *after* the app is launched. This matches
# Isaac Lab's required import ordering.
try:
import gymnasium as gym
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires `gymnasium`.") from exc
try:
import sim_evals.environments # noqa: F401
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError(
"Optional DROID sim-eval client requires the external `sim-evals` package or checkout to be importable."
) from exc
try:
from isaaclab_tasks.utils import parse_env_cfg
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("Optional DROID sim-eval client requires `isaaclab_tasks`.") from exc
# Resolve output location and scene prompt.
output_dir = _make_output_dir(args.output_root.expanduser().resolve(), args.scene)
instruction = _scene_instruction(args.scene)
# Build the DROID environment configuration from `sim-evals`.
env_cfg = parse_env_cfg(
"DROID",
device=args.device,
num_envs=1,
use_fabric=True,
)
# Select one of the pre-authored scenes/tasks.
env_cfg.set_scene(args.scene)
env = gym.make("DROID", cfg=env_cfg)
# Upstream sim-evals resets twice so materials / cameras are fully ready.
obs, _ = env.reset()
obs, _ = env.reset()
# Connect the websocket policy client.
client = DreamZeroJointPosClient(
remote_host=args.host,
remote_port=args.port,
path=args.path,
open_loop_horizon=args.open_loop_horizon,
)
# Aggregated per-run results.
all_episode_summaries: list[dict[str, Any]] = []
max_steps = int(env.env.max_episode_length)
logging.info("DreamZero metadata: %s", client.metadata())
logging.info("Scene %s prompt: %s", args.scene, instruction)
logging.info("Writing outputs to %s", output_dir)
try:
# No gradients are needed in inference-only rollout mode.
with torch.no_grad():
for episode_index in range(args.episodes):
# Per-episode collectors.
frames: list[np.ndarray] = []
step_records: list[StepRecord] = []
episode_start = time.perf_counter()
server_time_s = 0.0
final_reward = 0.0
terminated = False
truncated = False
for step_index in range(max_steps):
# Ask the policy for the next action. Internally this may or
# may not trigger a real server request depending on whether
# the local chunk cache has been exhausted.
logging.debug("Episode %d step %d: requesting action", episode_index, step_index)
result = client.infer(obs, instruction)
logging.debug(
"Episode %d step %d: got action (server_call=%s latency=%s)",
episode_index,
step_index,
result["used_server_call"],
result["chunk_latency_s"],
)
# Save one visualization frame per simulator step.
frames.append(result["viz"])
# Isaac Lab expects batched actions, hence `[None]`.
action_tensor = torch.tensor(result["action"], dtype=torch.float32)[None]
logging.debug("Episode %d step %d: stepping env", episode_index, step_index)
obs, reward, term, trunc, info = env.step(action_tensor)
logging.debug("Episode %d step %d: env.step returned", episode_index, step_index)
logging.debug("Episode %d step %d: parsing reward/flags", episode_index, step_index)
logging.debug(
"Episode %d step %d: raw types reward=%s term=%s trunc=%s",
episode_index,
step_index,
type(reward).__name__,
type(term).__name__,
type(trunc).__name__,
)
# Normalize environment outputs into plain Python scalars so
# the rest of the code does not depend on wrapper-specific types.
reward_value = _scalar_from_env_value(reward)
term_value = _bool_from_env_value(term)
trunc_value = _bool_from_env_value(trunc)
logging.debug(
"Episode %d step %d: parsed reward=%s term=%s trunc=%s",
episode_index,
step_index,
reward_value,
term_value,
trunc_value,
)
# Keep scene-object capture optional. It is useful for
# debugging / success heuristics, but the rollout should not
# fail if the environment does not expose object roots.
scene_objects = _capture_scene_objects(env)
# Accumulate total server-side time only on steps that
# triggered a fresh chunk inference.
if result["chunk_latency_s"] is not None:
server_time_s += float(result["chunk_latency_s"])
# Materialize one JSON-serializable trajectory record.
logging.debug("Episode %d step %d: appending trajectory", episode_index, step_index)
step_records.append(
StepRecord(
step_index=step_index,
used_server_call=bool(result["used_server_call"]),
chunk_latency_s=(
float(result["chunk_latency_s"]) if result["chunk_latency_s"] is not None else None
),
action=[float(x) for x in np.asarray(result["action"], dtype=np.float32).tolist()],
joint_position=[
float(x) for x in np.asarray(result["joint_position"], dtype=np.float32).tolist()
],
gripper_position=[
float(x) for x in np.asarray(result["gripper_position"], dtype=np.float32).tolist()
],
reward=reward_value,
terminated=term_value,
truncated=trunc_value,
scene_objects=scene_objects,
)
)
logging.debug("Episode %d step %d: trajectory appended", episode_index, step_index)
# Track final status for summary export.
final_reward = reward_value
terminated = term_value
truncated = trunc_value
if term_value or trunc_value:
# End the rollout early if the environment terminates.
break
# Episode-level timing and video export.
episode_wall_time_s = time.perf_counter() - episode_start
video_path = output_dir / f"episode_{episode_index:02d}.mp4"
logging.info("Episode %d: writing video to %s", episode_index, video_path)
mediapy.write_video(video_path, frames, fps=15)
# Reset the policy server between episodes.
logging.info("Episode %d: sending reset", episode_index)
reset_response = client.reset()
final_objects = step_records[-1].scene_objects if step_records else {}
success_report = _maybe_infer_success(args.scene, final_objects)
# Assemble the per-episode summary that is written to
# `episode_XX.json`.
episode_summary = {
"episode_index": episode_index,
"prompt": instruction,
"video_path": str(video_path),
"steps_executed": len(step_records),
"max_steps": max_steps,
"terminated": terminated,
"truncated": truncated,
"final_reward": final_reward,
"server_calls": client.server_calls,
"server_time_s": server_time_s,
"episode_wall_time_s": episode_wall_time_s,
"avg_server_time_per_call_s": (
server_time_s / client.server_calls if client.server_calls else None
),
"reset_response": reset_response,
"success_report": success_report,
"server_metadata": client.metadata(),
"trajectory": [record.__dict__ for record in step_records],
}
_dump_json(output_dir / f"episode_{episode_index:02d}.json", episode_summary)
all_episode_summaries.append(episode_summary)
logging.info(
"Episode %d done: steps=%d wall=%.2fs server_calls=%d heuristic_success=%s",
episode_index,
len(step_records),
episode_wall_time_s,
client.server_calls,
success_report.get("heuristic_success"),
)
# Reset per-episode counters while keeping the client alive.
client.server_calls = 0
# Top-level run summary across all episodes.
summary = {
"scene": args.scene,
"prompt": instruction,
"episodes": args.episodes,
"host": args.host,
"port": args.port,
"path": args.path,
"device": args.device,
"output_dir": str(output_dir),
"server_metadata": client.metadata(),
"episodes_summary": all_episode_summaries,
}
_dump_json(output_dir / "summary.json", summary)
# Also print the summary to stdout so the caller can capture it in logs.
print(json.dumps(summary, ensure_ascii=False, indent=2))
finally:
# Best-effort cleanup. Avoid masking the main error if cleanup fails.
try:
if client is not None:
client.client.close()
except Exception:
pass
if env is not None:
env.close()
simulation_app.close()
if __name__ == "__main__":
# Keep the script-level logs readable. Per-step rollout details are still
# available via `DEBUG` if needed, but websocket / asyncio internals are
# usually too noisy for normal usage.
logging.basicConfig(level=logging.INFO)
logging.getLogger("websockets").setLevel(logging.WARNING)
logging.getLogger("asyncio").setLevel(logging.WARNING)
main()
@@ -1,86 +0,0 @@
from __future__ import annotations
import argparse
import os
import sys
from pathlib import Path
os.environ.setdefault("MUJOCO_GL", "egl")
os.environ.setdefault("PYOPENGL_PLATFORM", "egl")
# Import base configs at module top level so the subclasses below are pickle-
# resolvable (worker processes import this module fresh via __main__).
from molmo_spaces.configs.policy_configs_baselines import ( # noqa: E402
DreamZeroPolicyConfig,
)
from molmo_spaces.evaluation.configs.evaluation_configs import ( # noqa: E402
DreamZeroPolicyEvalConfig,
)
# We only need to change the backend host and port to the vllm-host!
class DreamZeroVllmOmniPolicyConfig(DreamZeroPolicyConfig):
remote_config: dict = dict(host="127.0.0.1", port=8000)
class DreamZeroVllmOmniEvalConfig(DreamZeroPolicyEvalConfig):
policy_config: DreamZeroVllmOmniPolicyConfig = DreamZeroVllmOmniPolicyConfig()
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--port", type=int, default=8000)
parser.add_argument(
"--benchmark_dir",
required=True,
help=(
"Path to a MolmoSpaces benchmark directory, for example "
"$MOLMOSPACES_BENCHMARK_DIR/20260327/ithor/FrankaCloseHardBench/"
"FrankaCloseHardBench_20260206_json_benchmark"
),
)
parser.add_argument("--max_episodes", type=int, default=1)
parser.add_argument("--task_horizon_steps", type=int, default=80)
parser.add_argument(
"--output_dir",
required=True,
help="Directory to write evaluation outputs (created if missing).",
)
parser.add_argument("--episode_idx", type=int, default=None)
args = parser.parse_args()
policy_config = DreamZeroVllmOmniPolicyConfig(remote_config=dict(host=args.host, port=args.port))
DreamZeroVllmOmniPolicyConfig.model_fields["remote_config"].default = policy_config.remote_config
DreamZeroVllmOmniEvalConfig.model_fields["policy_config"].default = policy_config
# Import after env vars are set so MuJoCo picks EGL.
from molmo_spaces.evaluation import run_evaluation
cfg_cls = DreamZeroVllmOmniEvalConfig
output_dir = args.output_dir
Path(output_dir).mkdir(parents=True, exist_ok=True)
print(f"[eval] benchmark_dir={args.benchmark_dir}")
print(f"[eval] max_episodes={args.max_episodes} task_horizon_steps={args.task_horizon_steps}")
print(f"[eval] remote policy: ws://{args.host}:{args.port}/v1/realtime/robot/openpi")
results = run_evaluation(
eval_config_cls=cfg_cls,
benchmark_dir=Path(args.benchmark_dir),
max_episodes=args.max_episodes,
task_horizon_steps=args.task_horizon_steps,
num_workers=1,
use_wandb=False,
output_dir=output_dir,
episode_idx=args.episode_idx,
)
print(f"[eval] success={results.success_count}/{results.total_count} ({results.success_rate:.1%})")
print(f"[eval] output_dir={results.output_dir}")
return 0
if __name__ == "__main__":
sys.exit(main())
@@ -1,311 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import argparse
import json
import logging
import sys
import uuid
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import numpy as np
try:
import cv2
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("DreamZero OpenPI example requires `opencv-python`.") from exc
try:
example_dir = str(Path(__file__).resolve().parent)
removed_path = False
if sys.path and sys.path[0] == example_dir:
sys.path.pop(0)
removed_path = True
try:
from openpi_client.websocket_client_policy import WebsocketClientPolicy
finally:
if removed_path:
sys.path.insert(0, example_dir)
except ImportError as exc: # pragma: no cover - runtime dependency guard
raise ImportError("DreamZero OpenPI example requires `openpi-client`.") from exc
DEFAULT_HOST = "127.0.0.1"
DEFAULT_PORT = 8000
DEFAULT_PATH = "/v1/realtime/robot/openpi"
DEFAULT_PROMPT = "Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan"
ACTION_HORIZON = 24
DEFAULT_ACTION_DIM = 8
RELATIVE_OFFSETS = [-23, -16, -8, 0]
REPO_ROOT = Path(__file__).resolve().parents[3]
ASSET_REPO_ID = "YangshenDeng/vllm-omni-dreamzero-assets"
DEFAULT_VIDEO_DIR = REPO_ROOT / "outputs" / "dreamzero" / "assets"
CAMERA_FILES = {
"observation/exterior_image_0_left": "exterior_image_1_left.mp4",
"observation/exterior_image_1_left": "exterior_image_2_left.mp4",
"observation/wrist_image_left": "wrist_image_left.mp4",
}
def _as_action_array(response: Any) -> np.ndarray:
if isinstance(response, dict) and response.get("type") == "error":
message = response.get("message", response)
raise RuntimeError(f"Inference failed: {message}")
return np.asarray(response, dtype=np.float32)
@dataclass(frozen=True)
class DreamZeroServerMetadata:
image_resolution: tuple[int, int]
n_external_cameras: int
needs_wrist_camera: bool
needs_stereo_camera: bool
needs_session_id: bool
action_space: str
@classmethod
def from_dict(cls, payload: dict[str, Any]) -> DreamZeroServerMetadata:
required_keys = (
"image_resolution",
"n_external_cameras",
"needs_wrist_camera",
"needs_stereo_camera",
"needs_session_id",
"action_space",
)
missing_keys = [key for key in required_keys if key not in payload]
if missing_keys:
raise ValueError(f"Missing DreamZero metadata keys: {missing_keys}")
image_resolution = payload["image_resolution"]
if not isinstance(image_resolution, (list, tuple)) or len(image_resolution) != 2:
raise ValueError(f"Invalid image_resolution: {image_resolution!r}")
return cls(
image_resolution=(int(image_resolution[0]), int(image_resolution[1])),
n_external_cameras=int(payload["n_external_cameras"]),
needs_wrist_camera=bool(payload["needs_wrist_camera"]),
needs_stereo_camera=bool(payload["needs_stereo_camera"]),
needs_session_id=bool(payload["needs_session_id"]),
action_space=str(payload["action_space"]),
)
def _openpi_uri(host: str, port: int, path: str) -> str:
if host.startswith(("ws://", "wss://")):
return host
return f"ws://{host}:{port}{path}"
def _close_policy_client(client: WebsocketClientPolicy) -> None:
ws = getattr(client, "_ws", None)
if ws is not None:
ws.close()
def load_all_frames(video_path: Path) -> np.ndarray:
cap = cv2.VideoCapture(str(video_path))
frames = []
while True:
ok, frame = cap.read()
if not ok:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
cap.release()
if not frames:
raise RuntimeError(f"No frames loaded from {video_path}")
return np.stack(frames, axis=0)
def load_camera_frames(video_dir: Path) -> dict[str, np.ndarray]:
camera_frames: dict[str, np.ndarray] = {}
for camera_key, file_name in CAMERA_FILES.items():
video_path = video_dir / file_name
if not video_path.exists():
raise FileNotFoundError(
f"Missing DreamZero example asset: {video_path}. "
"Download the example videos with: "
f"`hf download {ASSET_REPO_ID} --repo-type dataset --local-dir {video_dir}`"
)
camera_frames[camera_key] = load_all_frames(video_path)
return camera_frames
def build_frame_schedule(total_frames: int, num_chunks: int) -> list[list[int]]:
chunks: list[list[int]] = []
current_frame = 23
for _ in range(num_chunks):
indices = [max(current_frame + offset, 0) for offset in RELATIVE_OFFSETS]
if indices[-1] >= total_frames:
break
chunks.append(indices)
current_frame += ACTION_HORIZON
return chunks
def make_obs_from_video(
camera_frames: dict[str, np.ndarray],
frame_indices: list[int],
*,
prompt: str,
session_id: str,
) -> dict[str, Any]:
obs: dict[str, Any] = {}
for camera_key, all_frames in camera_frames.items():
selected = all_frames[frame_indices]
obs[camera_key] = selected[0] if len(frame_indices) == 1 else selected
obs["observation/joint_position"] = np.zeros(7, dtype=np.float32)
obs["observation/cartesian_position"] = np.zeros(6, dtype=np.float32)
obs["observation/gripper_position"] = np.zeros(1, dtype=np.float32)
obs["prompt"] = prompt
obs["session_id"] = session_id
return obs
def build_demo_observations(
camera_frames: dict[str, np.ndarray],
*,
prompt: str,
session_id: str,
num_chunks: int = 2,
) -> list[dict[str, Any]]:
if num_chunks < 1:
raise ValueError("num_chunks must be at least 1")
total_frames = min(frames.shape[0] for frames in camera_frames.values())
observations = [
make_obs_from_video(
camera_frames,
[0],
prompt=prompt,
session_id=session_id,
)
]
for indices in build_frame_schedule(total_frames, num_chunks - 1):
observations.append(
make_obs_from_video(
camera_frames,
indices,
prompt=prompt,
session_id=session_id,
)
)
return observations
def validate_session_result(
result: dict[str, Any],
*,
expected_action_horizon: int = ACTION_HORIZON,
expected_action_dim: int = DEFAULT_ACTION_DIM,
expected_action_count: int | None = None,
) -> None:
metadata = DreamZeroServerMetadata.from_dict(result["metadata"])
if metadata.image_resolution != (180, 320):
raise AssertionError(f"Unexpected image_resolution: {metadata.image_resolution}")
if metadata.n_external_cameras != 2:
raise AssertionError(f"Unexpected n_external_cameras: {metadata.n_external_cameras}")
if not metadata.needs_wrist_camera:
raise AssertionError("DreamZero example expects wrist camera metadata")
if metadata.action_space != "joint_position":
raise AssertionError(f"Unexpected action_space: {metadata.action_space}")
actions = result["actions"]
if expected_action_count is not None and len(actions) != expected_action_count:
raise AssertionError(f"Expected {expected_action_count} action tensors, got {len(actions)}")
if not actions:
raise AssertionError("Expected at least one action tensor")
for index, action in enumerate(actions):
if action.shape != (expected_action_horizon, expected_action_dim):
raise AssertionError(
f"Action {index} shape mismatch: expected "
f"{(expected_action_horizon, expected_action_dim)}, got {action.shape}"
)
if not np.isfinite(action).all():
raise AssertionError(f"Action {index} contains non-finite values")
def run_policy_session(
*,
host: str = DEFAULT_HOST,
port: int = DEFAULT_PORT,
path: str = DEFAULT_PATH,
video_dir: Path = DEFAULT_VIDEO_DIR,
prompt: str = DEFAULT_PROMPT,
session_id: str | None = None,
num_chunks: int = 2,
) -> dict[str, Any]:
session_id = session_id or str(uuid.uuid4())
camera_frames = load_camera_frames(video_dir)
observations = build_demo_observations(
camera_frames,
prompt=prompt,
session_id=session_id,
num_chunks=num_chunks,
)
uri = _openpi_uri(host, port, path)
logging.info("Connecting to %s", uri)
client = WebsocketClientPolicy(host=uri)
try:
metadata = dict(client.get_server_metadata())
actions = [_as_action_array(client.infer(obs)) for obs in observations]
return {
"metadata": metadata,
"actions": actions,
"session_id": session_id,
}
finally:
_close_policy_client(client)
def format_action_summary(index: int, action: np.ndarray) -> str:
return (
f"Action {index}: shape={tuple(action.shape)} dtype={action.dtype} "
f"min={action.min():.6f} max={action.max():.6f}"
)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="DreamZero OpenPI client example with downloaded real videos.")
parser.add_argument("--host", default=DEFAULT_HOST)
parser.add_argument("--port", type=int, default=DEFAULT_PORT)
parser.add_argument("--path", default=DEFAULT_PATH)
parser.add_argument(
"--video-dir", type=Path, default=DEFAULT_VIDEO_DIR, help="Directory containing the three camera MP4 files."
)
parser.add_argument("--prompt", default=DEFAULT_PROMPT)
parser.add_argument("--session-id", default=None)
parser.add_argument("--num-chunks", type=int, default=2)
return parser.parse_args()
def main() -> int:
args = parse_args()
logging.basicConfig(level=logging.INFO)
result = run_policy_session(
host=args.host,
port=args.port,
path=args.path,
video_dir=args.video_dir,
prompt=args.prompt,
session_id=args.session_id,
num_chunks=args.num_chunks,
)
validate_session_result(result, expected_action_count=args.num_chunks)
print("Server metadata:", json.dumps(result["metadata"], sort_keys=True))
for index, action in enumerate(result["actions"]):
print(format_action_summary(index, action))
print("Session ID:", result["session_id"])
return 0
if __name__ == "__main__":
raise SystemExit(main())
@@ -1,24 +0,0 @@
#!/usr/bin/env bash
set -euo pipefail
MODEL="${MODEL:-GEAR-Dreams/DreamZero-DROID}"
HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-8000}"
DEPLOY_CONFIG="${DEPLOY_CONFIG:-vllm_omni/deploy/dreamzero_tp1_cfg2.yaml}"
SERVED_MODEL_NAME="${SERVED_MODEL_NAME:-dreamzero-droid}"
args=(
serve
"$MODEL"
--omni
--host "$HOST"
--port "$PORT"
--served-model-name "$SERVED_MODEL_NAME"
--deploy-config "$DEPLOY_CONFIG"
--enforce-eager
--disable-log-stats
)
ATTENTION_BACKEND="${ATTENTION_BACKEND:-torch}" \
DIFFUSION_ATTENTION_BACKEND="${DIFFUSION_ATTENTION_BACKEND:-TORCH_SDPA}" \
vllm "${args[@]}"
+5 -5
View File
@@ -22,14 +22,13 @@ export PYTHONPATH=<transformers_modules_path>:$PYTHONPATH
export HF_MODULES_CACHE=<transformers_modules_path> export HF_MODULES_CACHE=<transformers_modules_path>
``` ```
Run from repository root. The 3-stage pipeline is resolved automatically from Run from repository root:
the model type (`vllm_omni/deploy/dynin_omni.yaml`), so no config path is
required. Pass `--deploy-config` only to override the defaults:
```bash ```bash
vllm-omni serve snu-aidas/Dynin-Omni \ vllm-omni serve snu-aidas/Dynin-Omni \
--omni \ --omni \
--port 8091 --port 8091 \
--stage-configs-path "$(pwd)/vllm_omni/model_executor/stage_configs/dynin_omni.yaml"
``` ```
If `vllm-omni` is not in PATH, run: If `vllm-omni` is not in PATH, run:
@@ -37,7 +36,8 @@ If `vllm-omni` is not in PATH, run:
```bash ```bash
PYTHONPATH="$(pwd)" python -m vllm_omni.entrypoints.cli.main serve snu-aidas/Dynin-Omni \ PYTHONPATH="$(pwd)" python -m vllm_omni.entrypoints.cli.main serve snu-aidas/Dynin-Omni \
--omni \ --omni \
--port 8091 --port 8091 \
--stage-configs-path "$(pwd)/vllm_omni/model_executor/stage_configs/dynin_omni.yaml"
``` ```
Wait until the server logs show both `All stages initialized successfully` and Wait until the server logs show both `All stages initialized successfully` and
@@ -198,14 +198,7 @@ def _build_prompt_for_query_type(
return query_func(custom_prompt=custom_prompt) return query_func(custom_prompt=custom_prompt)
if query_type == "use_audio_in_video": if query_type == "use_audio_in_video":
return query_func(video_path=video_path, custom_prompt=custom_prompt) return query_func(video_path=video_path, custom_prompt=custom_prompt)
if query_type == "use_mixed_modalities": # use_mixed_modalities / use_multi_audios
return query_func(
video_path=video_path,
image_path=image_path,
audio_path=audio_path,
custom_prompt=custom_prompt,
)
# use_multi_audios
return query_func(custom_prompt=custom_prompt) return query_func(custom_prompt=custom_prompt)
@@ -51,18 +51,13 @@ def build_payload(
response_format: str = "wav", response_format: str = "wav",
stream: bool = False, stream: bool = False,
) -> dict: ) -> dict:
"""Build the /v1/audio/speech request payload. """Build the /v1/audio/speech request payload."""
Fish Speech has no built-in speakers; omit ``voice`` so the model
generates with a random timbre (controllable via text tags like
``[cool]``, ``[warm]``, etc.). When ``ref_audio`` is supplied we
enter voice-cloning mode instead.
"""
if not text or not text.strip(): if not text or not text.strip():
raise gr.Error("Please enter text to synthesize.") raise gr.Error("Please enter text to synthesize.")
payload: dict = { payload: dict = {
"input": text.strip(), "input": text.strip(),
"voice": "default",
"response_format": "pcm" if stream else response_format, "response_format": "pcm" if stream else response_format,
"stream": stream, "stream": stream,
} }
@@ -198,8 +193,7 @@ def build_interface(api_base: str, stream_chunk_seconds: float = 0.5):
with gr.Accordion("Voice Cloning (optional)", open=False): with gr.Accordion("Voice Cloning (optional)", open=False):
gr.Markdown( gr.Markdown(
"Upload or link a short reference audio (10-30s) and provide its transcript to clone the voice. " "Upload or link a short reference audio (10-30s) and provide its transcript to clone the voice."
"Leave empty for random timbre; use text tags like `[cool]`, `[warm]`, `[happy]` to steer the output."
) )
ref_audio = gr.Audio( ref_audio = gr.Audio(
label="Reference Audio", label="Reference Audio",
@@ -39,6 +39,7 @@ def run_tts(args) -> None:
payload = { payload = {
"model": args.model, "model": args.model,
"input": args.text, "input": args.text,
"voice": "default",
"response_format": args.response_format, "response_format": args.response_format,
} }
@@ -1,40 +0,0 @@
# Higgs-Audio V3 Online Serving
## Start the server
```bash
# Default: GPU 0, port 8095
./examples/online_serving/text_to_speech/higgs_audio_v3/run_server.sh
# Custom GPU / port
PORT=8096 GPUS=0,1 ./examples/online_serving/text_to_speech/higgs_audio_v3/run_server.sh
```
## Plain text TTS
```bash
python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
--base-url http://localhost:8095 \
--output-dir /tmp/higgs_v3_batch \
--prompts "Hello world." "The quick brown fox jumps over the lazy dog."
```
## Voice clone
```bash
python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
--base-url http://localhost:8095 \
--output-dir /tmp/higgs_v3_clone \
--ref-audio path/to/reference.wav \
--ref-text "transcript of the reference clip" \
--prompts "Text to synthesize in the cloned voice."
```
## curl example
```bash
curl -X POST http://localhost:8095/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{"model": "higgs_audio_v3", "input": "Hello world."}' \
--output hello.wav
```
@@ -1,115 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Batch client for the higgs-audio v3 online server.
Sends prompts to ``/v1/audio/speech`` and saves the returned WAV files.
Usage (plain text -> speech):
python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
--base-url http://localhost:8095 \
--output-dir /tmp/higgs_v3_batch \
--prompts "Hello world." "The quick brown fox jumps over the lazy dog."
Usage (voice clone):
python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
--base-url http://localhost:8095 \
--output-dir /tmp/higgs_v3_clone \
--ref-audio path/to/reference.wav \
--ref-text "the transcript of the reference clip" \
--prompts "Hello world."
"""
from __future__ import annotations
import argparse
import base64
import sys
from pathlib import Path
DEFAULT_PROMPTS = (
"Hello world.",
"The quick brown fox jumps over the lazy dog.",
"Today is a beautiful day for a walk in the park.",
"Innovation distinguishes between a leader and a follower.",
)
def _slug(text: str) -> str:
import re
s = re.sub(r"\s+", "_", text.strip().lower())
return re.sub(r"[^a-z0-9_]+", "", s)[:32] or "prompt"
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
parser.add_argument("--base-url", default="http://localhost:8095")
parser.add_argument("--model", default="higgs_audio_v3")
parser.add_argument("--prompts", nargs="+", default=list(DEFAULT_PROMPTS))
parser.add_argument("--output-dir", type=Path, default=Path("/tmp/higgs_v3_batch"))
parser.add_argument("--format", choices=("wav", "pcm"), default="wav")
parser.add_argument("--max-new-tokens", type=int, default=2048)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--timeout-s", type=float, default=120.0)
parser.add_argument(
"--ref-audio",
type=Path,
default=None,
help="Reference clip for voice clone (WAV/FLAC/MP3 path). Pair with --ref-text.",
)
parser.add_argument(
"--ref-text",
type=str,
default=None,
help="Transcript of the reference clip. Optional but improves fidelity.",
)
args = parser.parse_args()
ref_audio_data_url: str | None = None
if args.ref_audio is not None:
if not args.ref_audio.exists():
print(f"ref-audio file not found: {args.ref_audio}", file=sys.stderr)
return 2
mime = "audio/wav" if args.ref_audio.suffix.lower() == ".wav" else "audio/mpeg"
ref_b64 = base64.b64encode(args.ref_audio.read_bytes()).decode("ascii")
ref_audio_data_url = f"data:{mime};base64,{ref_b64}"
try:
import httpx
except ImportError:
print("this client needs `httpx`. Install with `pip install httpx`.", file=sys.stderr)
return 2
args.output_dir.mkdir(parents=True, exist_ok=True)
url = args.base_url.rstrip("/") + "/v1/audio/speech"
failures = 0
with httpx.Client(timeout=args.timeout_s) as client:
for prompt in args.prompts:
payload = {
"model": args.model,
"input": prompt,
"response_format": args.format,
"max_new_tokens": args.max_new_tokens,
"seed": args.seed,
}
if ref_audio_data_url is not None:
payload["ref_audio"] = ref_audio_data_url
if args.ref_text:
payload["ref_text"] = args.ref_text
resp = client.post(url, json=payload)
if resp.status_code != 200:
print(f"[FAIL] {prompt!r} -> {resp.status_code}: {resp.text[:200]}", file=sys.stderr)
failures += 1
continue
suffix = ".wav" if args.format == "wav" else ".pcm"
out = args.output_dir / f"{_slug(prompt)}{suffix}"
out.write_bytes(resp.content)
print(f"[ ok ] {prompt!r} -> {out} ({len(resp.content)} bytes)")
return 1 if failures else 0
if __name__ == "__main__":
sys.exit(main())
@@ -1,32 +0,0 @@
#!/bin/bash
# Launch vLLM-Omni server for higgs-audio v3.
#
# Supports plain text TTS and voice cloning via /v1/audio/speech.
#
# Usage:
# ./run_server.sh # default port 8095, GPU 0
# PORT=8096 GPUS=0,1 ./run_server.sh
# MODEL=/path/to/local/checkpoint ./run_server.sh
set -e
MODEL="${MODEL:-bosonai/higgs-audio-v3-tts-4b}"
PORT="${PORT:-8095}"
GPUS="${GPUS:-0}"
GPU_UTIL="${GPU_UTIL:-0.6}"
echo "Starting higgs-audio v3 server"
echo " MODEL=$MODEL"
echo " PORT=$PORT"
echo " CUDA_VISIBLE_DEVICES=$GPUS"
CUDA_VISIBLE_DEVICES="$GPUS" \
VLLM_USE_DEEP_GEMM=0 \
VLLM_MOE_USE_DEEP_GEMM=0 \
vllm-omni serve "$MODEL" \
--deploy-config vllm_omni/deploy/higgs_multimodal_qwen3.yaml \
--host 0.0.0.0 \
--port "$PORT" \
--gpu-memory-utilization "$GPU_UTIL" \
--trust-remote-code \
--omni
@@ -1,247 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Verify a ModelOpt FP8 diffusers checkpoint exported by
quantize_wan2_2_modelopt_fp8.py / quantize_hunyuanvideo_15_modelopt_fp8.py.
Three checks:
A. transformer/config.json has a sane quantization_config block.
B. transformer/*.safetensors contains FP8 (float8_e4m3fn) quantized tensors.
C. transformer disk size is materially smaller than a BF16 baseline.
Example:
python examples/quantization/check_modelopt_fp8_export.py \\
--output ./hv15-480p-modelopt-fp8
# Optional: compare disk size against a local or HF BF16 baseline.
python examples/quantization/check_modelopt_fp8_export.py \\
--output ./hv15-480p-modelopt-fp8 \\
--baseline hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v
"""
from __future__ import annotations
import argparse
import json
import sys
from collections import Counter
from pathlib import Path
SUPPORTED_ALGOS = ("FP8",)
def _check_config(transformer_dir: Path) -> tuple[int, str | None]:
"""Returns (status, quant_algo). status: 0 pass, 1 fail, 2 warn."""
cfg_path = transformer_dir / "config.json"
if not cfg_path.exists():
print(f"[FAIL] {cfg_path} missing.")
return 1, None
with cfg_path.open(encoding="utf-8") as f:
cfg = json.load(f)
qc = cfg.get("quantization_config")
if not isinstance(qc, dict):
print(f"[FAIL] No `quantization_config` block in {cfg_path}.")
return 1, None
print(f"[A] quantization_config from {cfg_path}:")
print(json.dumps(qc, indent=2))
quant_algo = qc.get("quant_algo")
issues = []
if qc.get("quant_method") != "modelopt":
issues.append(f"quant_method={qc.get('quant_method')!r} (expected 'modelopt')")
if quant_algo not in SUPPORTED_ALGOS:
issues.append(
f"quant_algo={quant_algo!r} (expected one of {SUPPORTED_ALGOS} — vllm-omni adapter may not auto-detect)"
)
if issues:
print("[A] WARN — config looks incomplete:")
for issue in issues:
print(f" - {issue}")
return 2, quant_algo
print(f"[A] PASS — config looks correct (quant_algo={quant_algo}).")
return 0, quant_algo
def _read_safetensors_header(path: Path) -> dict:
"""Read the JSON header of a safetensors file. Bypass-safe — doesn't materialize tensors.
Returns {tensor_name: {'dtype': 'F8_E4M3', 'shape': [...], 'data_offsets': [...]}}.
Header dtype strings: F8_E4M3, F8_E5M2, BF16, F16, F32, F64, I8, I16, I32, I64, BOOL, U8, ...
"""
import struct
with open(path, "rb") as f:
header_len = struct.unpack("<Q", f.read(8))[0]
header = json.loads(f.read(header_len))
header.pop("__metadata__", None)
return header
def _classify_weight_scale_granularity(weight_scale_shapes: list[list[int]]) -> str:
"""Infer per-tensor vs per-channel vs per-block from sample weight_scale shapes.
ModelOpt block-wise produces shapes like `[16, 1, 16, 1]` (broadcasting dims of 1
interleaved with block-count dims). We count "meaningful" dims — ones with size > 1 —
and classify: 0 meaningful dims = per-tensor (scalar), 1 = per-channel, 2+ = per-block.
"""
if not weight_scale_shapes:
return "no weight_scale tensors found"
def meaningful_dims(shape: list[int]) -> int:
return sum(1 for d in shape if d > 1)
per_tensor = sum(1 for s in weight_scale_shapes if meaningful_dims(s) == 0)
per_channel = sum(1 for s in weight_scale_shapes if meaningful_dims(s) == 1)
per_block = sum(1 for s in weight_scale_shapes if meaningful_dims(s) >= 2)
total = len(weight_scale_shapes)
if per_tensor == total:
return "per-tensor (all scalar scales)"
if per_channel == total:
return "per-channel (1 meaningful dim)"
if per_block == total:
return "per-block (2+ meaningful dims — e.g. [M//bm, 1, N//bn, 1] for tiles)"
return f"mixed: per-tensor={per_tensor}, per-channel={per_channel}, per-block={per_block} of {total}"
def _check_safetensors(transformer_dir: Path) -> int:
"""Returns 0 on pass, 1 on fail. Reads on-disk dtype from the safetensors header."""
files = sorted(transformer_dir.glob("*.safetensors"))
if not files:
print(f"[FAIL] No *.safetensors in {transformer_dir}.")
return 1
header_dtype_counts: Counter[str] = Counter()
sample_quant_weight_keys: list[str] = []
sample_scale_keys: list[str] = []
weight_scale_shapes: list[list[int]] = []
sample_weight_scale_entries: list[tuple[str, list[int]]] = []
for f in files:
try:
header = _read_safetensors_header(f)
except Exception as exc:
print(f"[B] WARN — could not parse header of {f}: {exc}")
continue
for k, info in header.items():
dtype = info.get("dtype", "?")
header_dtype_counts[dtype] += 1
# FP8 stores weights as F8_E4M3 directly.
if dtype.startswith("F8") and k.endswith(".weight") and len(sample_quant_weight_keys) < 5:
sample_quant_weight_keys.append(k)
if k.endswith(("_scale", ".weight_scale", ".input_scale", "_scale_inv")) and len(sample_scale_keys) < 5:
sample_scale_keys.append(k)
if k.endswith(".weight_scale"):
weight_scale_shapes.append(info.get("shape", []))
if len(sample_weight_scale_entries) < 5:
sample_weight_scale_entries.append((k, info.get("shape", [])))
print(f"\n[B] On-disk dtype counts across {len(files)} safetensors file(s) (from header, not get_tensor):")
for dtype, count in sorted(header_dtype_counts.items(), key=lambda kv: -kv[1]):
marker = " <-- FP8" if dtype.startswith("F8") else ""
print(f" {dtype:10s} {count:>6d}{marker}")
quant_count = sum(c for d, c in header_dtype_counts.items() if d.startswith("F8"))
if quant_count == 0:
print("[B] FAIL — no FP8 tensors on disk. Calibration likely did not actually quantize the weights.")
return 1
print(f"[B] PASS — {quant_count} FP8 tensors stored on disk.")
if sample_quant_weight_keys:
print(f" sample quantized weight tensors: {sample_quant_weight_keys[:3]}")
if sample_scale_keys:
print(f" sample scale tensors: {sample_scale_keys[:3]}")
print(" (Note: torch's get_tensor() may return these as bf16 views on some versions —")
print(" irrelevant; vLLM's loader uses native FP8 ops.)")
# Weight-scale granularity — per-tensor (scalar) vs per-channel (1-D) vs per-block (N-D).
print(f"\n weight_scale granularity: {_classify_weight_scale_granularity(weight_scale_shapes)}")
for key, shape in sample_weight_scale_entries[:3]:
print(f" {key}: shape {shape}")
return 0
def _disk_size_gib(p: Path) -> float:
return sum(f.stat().st_size for f in p.rglob("*") if f.is_file()) / (1024**3)
def _check_size_vs_baseline(transformer_dir: Path, baseline: str | None) -> int:
"""Returns 0 always (informational only)."""
quant_size = _disk_size_gib(transformer_dir)
print(f"\n[C] FP8 transformer disk size: {quant_size:.2f} GiB")
if baseline is None:
print("[C] SKIP — pass --baseline <path or HF id> to compare against BF16.")
return 0
baseline_path = Path(baseline)
if not baseline_path.exists():
# Try HF download.
try:
from huggingface_hub import snapshot_download
except ImportError:
print("[C] SKIP — huggingface_hub not installed and baseline not a local path.")
return 0
print(f" Downloading baseline transformer from HF: {baseline}")
baseline_path = Path(snapshot_download(baseline, allow_patterns=["transformer/*"]))
bf16_dir = baseline_path / "transformer" if (baseline_path / "transformer").exists() else baseline_path
bf16_size = _disk_size_gib(bf16_dir)
if bf16_size == 0:
print(f"[C] WARN — baseline transformer dir empty: {bf16_dir}")
return 0
# Expected reduction for FP8: ~50%.
min_reduction = 30
reduction = (1 - quant_size / bf16_size) * 100
print(f"[C] BF16 baseline transformer disk size: {bf16_size:.2f} GiB ({bf16_dir})")
print(f"[C] Disk reduction: {reduction:.1f}% (FP8 transformer is {quant_size / bf16_size:.0%} of BF16)")
if reduction < min_reduction:
print(
f"[C] WARN — FP8 should typically reduce disk by ~40-50%; <{min_reduction}% suggests partial quantization."
)
return 0
def main() -> None:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--output", required=True, help="Path to the exported ModelOpt FP8 checkpoint root.")
p.add_argument(
"--baseline",
default=None,
help="Optional BF16 baseline (local diffusers dir or HF id) for disk-size comparison.",
)
args = p.parse_args()
out_root = Path(args.output).expanduser().resolve()
transformer_dir = out_root / "transformer"
if not transformer_dir.exists():
print(f"[FAIL] {transformer_dir} does not exist.")
sys.exit(1)
print(f"Checking: {out_root}\n")
fail = 0
config_status, _quant_algo = _check_config(transformer_dir)
fail |= config_status
fail |= _check_safetensors(transformer_dir)
_check_size_vs_baseline(transformer_dir, args.baseline)
print()
if fail == 0:
print("=" * 60)
print("ALL CHECKS PASSED — checkpoint looks ready for vllm-omni serving.")
elif fail == 1:
print("=" * 60)
print("FAILURES detected — calibration may need to be re-run.")
sys.exit(1)
else:
print("=" * 60)
print("WARNINGS only — checkpoint may serve but with caveats. See [A] above.")
if __name__ == "__main__":
main()
@@ -1,448 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Quantize HunyuanVideo-1.5 (480p T2V) to a ModelOpt FP8 Hugging Face checkpoint.
Calibrates the DiT transformer using a small video prompt set and exports a
diffusers-style directory whose transformer carries ModelOpt FP8 metadata.
The exported checkpoint is consumable by vllm-omni's ModelOpt FP8 adapter
(see vllm_omni/diffusion/model_loader/checkpoint_adapters/modelopt.py).
Layers kept full precision match the #2728 / #2795 pattern: modulation,
AdaLayerNorm, entry/exit projections, embeddings, the token refiner path,
and final proj_out. MHA quantizers are off by default; HV-1.5 self-attention
empirically degrades under FP8 (see #2920 ablation).
Example:
python examples/quantization/quantize_hunyuanvideo_15_modelopt_fp8.py \\
--model hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \\
--output ./hv15-480p-modelopt-fp8 \\
--overwrite
"""
from __future__ import annotations
import argparse
import copy
import json
import re
import shutil
import sys
from pathlib import Path
from typing import Any
import torch
from diffusers import DiffusionPipeline
DEFAULT_PROMPTS = [
"A dog running across a field of golden wheat.",
"An astronaut riding a horse across the surface of Mars, red dust swirling, cinematic wide shot.",
"A hummingbird hovering in front of a vibrant red flower, slow motion, macro shot.",
"A crackling campfire at night under a starry sky, sparks rising into the dark.",
"An underwater shot of a coral reef with tropical fish swimming by, sun rays piercing the water.",
"A close-up of a blooming rose covered in morning dew, soft natural light.",
"A peaceful mountain village at dawn, mist rolling over the rooftops, cinematic establishing shot.",
"A skateboarder doing a kickflip in an urban plaza, slow motion, golden hour lighting.",
]
def _build_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--model", required=True, help="Input HV-1.5 diffusers directory or HF id.")
p.add_argument("--output", required=True, help="Output directory for the ModelOpt FP8 checkpoint.")
p.add_argument("--dtype", choices=("bfloat16", "float16"), default="bfloat16")
p.add_argument("--height", type=int, default=480)
p.add_argument("--width", type=int, default=832)
p.add_argument(
"--num-frames",
type=int,
default=33,
help="Frames per calibration sample. 33 matches the typical short benchmark.",
)
p.add_argument("--guidance-scale", type=float, default=6.0)
p.add_argument(
"--calib-steps",
type=int,
default=10,
help="Denoising steps per calibration prompt (10 is enough for amax statistics).",
)
p.add_argument("--calib-size", type=int, default=8, help="How many prompts to use for calibration.")
p.add_argument("--seed", type=int, default=42)
p.add_argument(
"--prompt",
action="append",
default=[],
help="Custom calibration prompt. Repeat to provide multiple.",
)
p.add_argument(
"--quantize-mha",
action="store_true",
help="Enable FP8 attention K/V/softmax quantizers. Off by default — empirically degrades HV-1.5 video output.",
)
p.add_argument(
"--weight-block-size",
type=str,
default=None,
help="Per-block weight quantization as 'M,N' (e.g. '128,128' for 128x128 tiles). "
"Default: per-tensor (one scale per linear). Block-wise typically gives tighter quality at "
"negligible memory cost. Static FP8 is exempt from upstream vLLM's online block-wise gate.",
)
p.add_argument("--overwrite", action="store_true", help="Replace an existing output directory.")
return p
def _parse_block_size(spec: str | None) -> list[int] | None:
if spec is None:
return None
parts = [int(x) for x in spec.split(",") if x.strip()]
if len(parts) != 2:
raise SystemExit(f"--weight-block-size must be 'M,N' (2 ints), got {spec!r}")
return parts
def _require_modelopt() -> Any:
try:
import modelopt.torch.quantization as mtq
except ModuleNotFoundError as exc:
raise SystemExit(
"NVIDIA ModelOpt is not installed. Install with:\n"
" pip install 'nvidia-modelopt[all]'\n"
f"Original error: {exc}"
) from exc
return mtq
def _ensure_paths(args: argparse.Namespace) -> tuple[str, Path]:
model_path = args.model
output_dir = Path(args.output).expanduser().resolve()
if output_dir.exists():
if not args.overwrite:
raise SystemExit(f"Output directory already exists: {output_dir}\nPass --overwrite to replace it.")
shutil.rmtree(output_dir)
return model_path, output_dir
def _select_dtype(name: str) -> torch.dtype:
return {"bfloat16": torch.bfloat16, "float16": torch.float16}[name]
def _build_prompts(args: argparse.Namespace) -> list[str]:
prompts = args.prompt or DEFAULT_PROMPTS
if args.calib_size <= 0:
raise SystemExit("--calib-size must be positive.")
if len(prompts) < args.calib_size:
repeats = (args.calib_size + len(prompts) - 1) // len(prompts)
prompts = (prompts * repeats)[: args.calib_size]
return prompts[: args.calib_size]
# Layers to KEEP at full precision (mirror of the #2920 wiring + #2728/#2795 skip pattern).
# - x_embedder, image_embedder, context_embedder*, time_embed*, cond_type_embed: entry/embedding
# - norm_out, norm1*.linear, norm1_context*.linear, norm2*, norm2_context*: AdaLayerNorm modulation
# - proj_out: final output projection
# - token_refiner*: text-encoder refinement uses diffusers raw nn.Linear
def _filter_func_hv15(name: str) -> bool:
pattern = re.compile(
r"(proj_out.*|"
r".*(x_embedder|image_embedder|context_embedder|context_embedder_2|"
r"time_embed|cond_type_embed|"
r"norm_out|norm1\.linear|norm1_context\.linear|norm2|norm2_context|"
r"token_refiner).*)"
)
return pattern.match(name) is not None
def _mha_filter_func(name: str) -> bool:
pattern = re.compile(
r".*(q_bmm_quantizer|k_bmm_quantizer|v_bmm_quantizer|softmax_quantizer|bmm2_output_quantizer).*"
)
return pattern.match(name) is not None
def _disable_known_problematic_quantizers(mtq: Any, backbone: torch.nn.Module, *, quantize_mha: bool) -> None:
if not hasattr(mtq, "disable_quantizer"):
return
mtq.disable_quantizer(backbone, _filter_func_hv15)
if not quantize_mha:
mtq.disable_quantizer(backbone, _mha_filter_func)
def _load_pipeline(model_path: str, dtype: torch.dtype) -> DiffusionPipeline:
pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype)
if hasattr(pipe, "set_progress_bar_config"):
pipe.set_progress_bar_config(disable=True)
pipe.to("cuda")
return pipe
def _build_forward_loop(pipe: DiffusionPipeline, args: argparse.Namespace, prompts: list[str]):
generator = torch.Generator(device="cuda")
# Try to set guidance on the pipeline's guider object up front (modern
# diffusers HV-1.5 uses a Guider abstraction, not a per-call kwarg). Falls
# back silently — calibration uses whatever default the pipeline ships with.
guider = getattr(pipe, "guider", None)
if guider is not None and hasattr(guider, "guidance_scale"):
try:
guider.guidance_scale = args.guidance_scale
except Exception:
pass
base_kwargs = dict(
height=args.height,
width=args.width,
num_frames=args.num_frames,
num_inference_steps=args.calib_steps,
output_type="latent",
)
def forward_loop(*_unused_args, **_unused_kwargs) -> None:
with torch.inference_mode():
for idx, prompt in enumerate(prompts):
generator.manual_seed(args.seed + idx)
# Try with guidance_scale first; fall back without on TypeError
# for pipelines (like HV-1.5) that take CFG via guider config.
try:
pipe(prompt=prompt, generator=generator, guidance_scale=args.guidance_scale, **base_kwargs)
except TypeError as exc:
if "guidance_scale" not in str(exc):
raise
pipe(prompt=prompt, generator=generator, **base_kwargs)
return forward_loop
def _summarize_export(output_dir: Path) -> None:
cfg_path = output_dir / "transformer" / "config.json"
if not cfg_path.exists():
print(f"[warn] {cfg_path} missing.", file=sys.stderr)
return
with cfg_path.open(encoding="utf-8") as f:
cfg = json.load(f)
qc = cfg.get("quantization_config")
if not isinstance(qc, dict):
print("[warn] No quantization_config in transformer/config.json.", file=sys.stderr)
return
print("Export summary:")
print(f" quant_method: {qc.get('quant_method')}")
print(f" quant_algo: {qc.get('quant_algo')}")
producer = qc.get("producer")
if isinstance(producer, dict):
print(f" producer: {producer.get('name')} {producer.get('version')}")
print(f" config path: {cfg_path}")
def _force_export_quantized_weights(backbone: torch.nn.Module, dtype: torch.dtype) -> int:
"""Convert in-memory weights of quantized modules to actual FP8 storage.
`export_hf_checkpoint` skips this step for unknown model types (HV-1.5 isn't
in ModelOpt's recognized-model registry), so we must call the per-weight
export helper ourselves. Same workaround as the HunyuanImage-3 calibration
helper.
"""
from modelopt.torch.export.quant_utils import (
QUANTIZATION_NONE,
get_quantization_format,
quantizer_attr_names,
weight_attr_names,
)
from modelopt.torch.export.unified_export_hf import _export_quantized_weight
exported = 0
for name, module in backbone.named_modules():
try:
quantization_format = get_quantization_format(module)
except Exception as exc:
print(f"[warn] Could not inspect quantization format for {name}: {exc}", file=sys.stderr)
continue
if quantization_format == QUANTIZATION_NONE:
continue
for weight_name in weight_attr_names(module):
quantizer_attrs = quantizer_attr_names(weight_name)
weight_quantizer = getattr(module, quantizer_attrs.weight_quantizer, None)
if weight_quantizer is None or not getattr(weight_quantizer, "is_enabled", False):
continue
_export_quantized_weight(module, dtype, weight_name)
exported += 1
return exported
def _hv15_quant_config_block(weight_block_size: list[int] | None = None) -> dict:
"""Mirror ModelOpt FP8 metadata expected by vllm-omni's adapter (#2913).
Same shape as the HunyuanImage-3 author's _hunyuan_quant_config(). When
`weight_block_size` is given, advertise block-wise weight quantization in
the saved metadata (so consumers know to expect multi-element scale tensors).
"""
weights_cfg: dict = {"dynamic": False, "num_bits": 8, "type": "float"}
if weight_block_size is not None:
weights_cfg["strategy"] = "block"
weights_cfg["block_structure"] = f"{weight_block_size[0]}x{weight_block_size[1]}"
return {
"config_groups": {
"group_0": {
"input_activations": {"dynamic": False, "num_bits": 8, "type": "float"},
"weights": weights_cfg,
"targets": ["Linear"],
}
},
"ignore": [
"context_embedder*",
"context_embedder_2*",
"cond_type_embed*",
"image_embedder*",
"norm1.linear*",
"norm1_context.linear*",
"norm2*",
"norm2_context*",
"norm_out*",
"proj_out*",
"time_embed*",
"token_refiner*",
"x_embedder*",
],
"producer": {"name": "modelopt"},
"quant_algo": "FP8",
"quant_method": "modelopt",
}
def _patch_quant_config(output_dir: Path, weight_block_size: list[int] | None = None) -> None:
"""Inject quant_algo: FP8 + config_groups into transformer/config.json so
vllm-omni's adapter (#2913) recognises the checkpoint as ModelOpt FP8."""
cfg_path = output_dir / "transformer" / "config.json"
with cfg_path.open(encoding="utf-8") as f:
cfg = json.load(f)
new_qc = _hv15_quant_config_block(weight_block_size=weight_block_size)
existing = cfg.get("quantization_config")
if isinstance(existing, dict):
producer = existing.get("producer")
if isinstance(producer, dict):
new_qc["producer"] = producer
cfg["quantization_config"] = new_qc
with cfg_path.open("w", encoding="utf-8") as f:
json.dump(cfg, f, indent=2)
def _save_pipeline_with_fp8_transformer(
pipe: DiffusionPipeline,
model_path: str,
output_dir: Path,
max_shard_size: str = "5GB",
) -> None:
"""Save the pipeline with the (now FP8) transformer.
Copies the source directory verbatim except for `transformer/`, then
saves the transformer with quantizers hidden so the state dict contains
only the FP8 weights + scale tensors.
"""
from modelopt.torch.export.diffusers_utils import hide_quantizers_from_state_dict
src = Path(model_path)
if not src.exists():
from huggingface_hub import snapshot_download
src = Path(snapshot_download(model_path))
if output_dir.exists():
shutil.rmtree(output_dir)
shutil.copytree(src, output_dir, ignore=shutil.ignore_patterns("transformer"))
transformer_out = output_dir / "transformer"
# `hide_quantizers_from_state_dict` walks named_modules(); pass the actual
# nn.Module (transformer), not the diffusers Pipeline wrapper.
with hide_quantizers_from_state_dict(pipe.transformer):
pipe.transformer.save_pretrained(
str(transformer_out),
safe_serialization=True,
max_shard_size=max_shard_size,
)
def main() -> None:
args = _build_parser().parse_args()
if not torch.cuda.is_available():
raise SystemExit("CUDA is required for ModelOpt FP8 quantization.")
mtq = _require_modelopt()
model_path, output_dir = _ensure_paths(args)
dtype = _select_dtype(args.dtype)
prompts = _build_prompts(args)
print("Quantization plan:")
weight_block_size = _parse_block_size(args.weight_block_size)
print(f" input: {args.model}")
print(f" output: {output_dir}")
print(f" dtype: {dtype}")
print(f" height/width: {args.height}x{args.width}")
print(f" num_frames: {args.num_frames}")
print(f" calib_size: {len(prompts)}")
print(f" calib_steps: {args.calib_steps}")
print(f" quantize_mha: {args.quantize_mha}")
print(
f" weight strategy: {'block-wise ' + str(weight_block_size) if weight_block_size else 'per-tensor (default)'}"
)
pipe = _load_pipeline(model_path, dtype)
backbone = pipe.transformer
quant_config = copy.deepcopy(mtq.FP8_DEFAULT_CFG)
if weight_block_size is not None:
# Switch from per-tensor (default) to block-wise weight quantization.
# ModelOpt's wildcard "*weight_quantizer" matches every linear's weight quantizer.
quant_config["quant_cfg"]["*weight_quantizer"] = {
"num_bits": (4, 3), # E4M3 (FP8 weights, same as default)
"block_sizes": {-1: weight_block_size[1], -2: weight_block_size[0]},
}
print(
f" -> overriding weight quantizer with block_sizes={weight_block_size} "
f"({weight_block_size[0]}x{weight_block_size[1]} tiles)"
)
forward_loop = _build_forward_loop(pipe, args, prompts)
quantized = mtq.quantize(backbone, quant_config, forward_loop)
if quantized is not None:
pipe.transformer = quantized
backbone = quantized
_disable_known_problematic_quantizers(mtq, backbone, quantize_mha=args.quantize_mha)
print("\nForcing FP8 weight serialization (HV-1.5 isn't in ModelOpt's recognized-model registry,")
print("so we have to call the per-weight export helper ourselves)...")
exported = _force_export_quantized_weights(backbone, dtype)
print(f" -> {exported} weights converted to FP8 in memory")
if exported == 0:
raise SystemExit(
"No quantized weights were exported. Calibration may have skipped every layer "
"(check the disable_quantizer regex) or `mtq.quantize` did not actually wrap any "
"weight quantizers."
)
print("\nSaving pipeline with FP8 transformer...")
_save_pipeline_with_fp8_transformer(pipe, model_path, output_dir)
_patch_quant_config(output_dir, weight_block_size=weight_block_size)
print(f"Saved to: {output_dir}")
_summarize_export(output_dir)
print("\nNext: validate the checkpoint with vllm-omni:")
print(
" python examples/offline_inference/text_to_video/text_to_video.py \\\n"
f" --model {output_dir} \\\n"
" --quantization fp8 \\\n"
" --prompt 'A dog running across a field of golden wheat.' \\\n"
f" --height {args.height} --width {args.width} --num-frames {args.num_frames} \\\n"
" --num-inference-steps 30 --guidance-scale 6.0 --seed 42 \\\n"
" --output outputs/hv15_modelopt_fp8.mp4 \\\n"
" --enforce-eager"
)
print(
"\n (--quantization fp8 is auto-upgraded to ModelOpt FP8 at runtime because the "
"checkpoint's config.json has modelopt metadata.)"
)
if __name__ == "__main__":
main()
@@ -1,471 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Quantize Wan2.2 (TI2V-5B, 704x1280 T2V) to a ModelOpt FP8 Hugging Face checkpoint.
Calibrates the DiT transformer using a small video prompt set and exports a
diffusers-style directory whose transformer carries ModelOpt FP8 metadata.
The exported checkpoint is consumable by vllm-omni's ModelOpt FP8 adapter
(see vllm_omni/diffusion/model_loader/checkpoint_adapters/modelopt.py).
Layers kept full precision match the #2728 / #2795 pattern: condition embedder
(time/text/image), patch embedding, modulation (scale_shift_table), final
norm + proj_out, and sequence-parallel helpers. All attention + FFN linears
are quantized — static calibration handles the numerics that online FP8
couldn't (see #2920 ablation).
Default target is `Wan-AI/Wan2.2-TI2V-5B-Diffusers`, the dense 5B variant that
fits 80GB BF16. The A14B MoE variants need 2+ GPUs and are out of scope here.
Example:
python examples/quantization/quantize_wan2_2_modelopt_fp8.py \\
--model Wan-AI/Wan2.2-TI2V-5B-Diffusers \\
--output ./wan22-ti2v-modelopt-fp8 \\
--overwrite
"""
from __future__ import annotations
import argparse
import copy
import json
import re
import shutil
import sys
from pathlib import Path
from typing import Any
import torch
from diffusers import DiffusionPipeline
DEFAULT_PROMPTS = [
"A dog running across a field of golden wheat.",
"An astronaut riding a horse across the surface of Mars, red dust swirling, cinematic wide shot.",
"A hummingbird hovering in front of a vibrant red flower, slow motion, macro shot.",
"A crackling campfire at night under a starry sky, sparks rising into the dark.",
"An underwater shot of a coral reef with tropical fish swimming by, sun rays piercing the water.",
"A close-up of a blooming rose covered in morning dew, soft natural light.",
"A peaceful mountain village at dawn, mist rolling over the rooftops, cinematic establishing shot.",
"A skateboarder doing a kickflip in an urban plaza, slow motion, golden hour lighting.",
]
def _build_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--model", required=True, help="Input Wan2.2 diffusers directory or HF id.")
p.add_argument("--output", required=True, help="Output directory for the ModelOpt FP8 checkpoint.")
p.add_argument("--dtype", choices=("bfloat16", "float16"), default="bfloat16")
p.add_argument("--height", type=int, default=704, help="Calibration video height (Wan2.2 TI2V-5B native: 704).")
p.add_argument("--width", type=int, default=1280, help="Calibration video width (Wan2.2 TI2V-5B native: 1280).")
p.add_argument(
"--num-frames",
type=int,
default=49,
help="Frames per calibration sample. 49 matches the typical short benchmark; "
"use 17 to reduce memory pressure during calibration.",
)
p.add_argument("--guidance-scale", type=float, default=5.0)
p.add_argument(
"--calib-steps",
type=int,
default=10,
help="Denoising steps per calibration prompt (10 is enough for amax statistics).",
)
p.add_argument("--calib-size", type=int, default=8, help="How many prompts to use for calibration.")
p.add_argument("--seed", type=int, default=42)
p.add_argument(
"--prompt",
action="append",
default=[],
help="Custom calibration prompt. Repeat to provide multiple.",
)
p.add_argument(
"--quantize-mha",
action="store_true",
help="Enable FP8 attention K/V/softmax quantizers. Off by default — Wan2.2's long attention "
"sequences amplified FP8 drift in the online ablation (see #2920).",
)
p.add_argument(
"--weight-block-size",
type=str,
default=None,
help="Per-block weight quantization as 'M,N' (e.g. '128,128'). Default per-tensor. "
"Note: vllm-omni's ModelOpt adapter may not yet dispatch block-wise scales — check #2924 "
"for the HV-1.5 investigation status before relying on this.",
)
p.add_argument("--overwrite", action="store_true", help="Replace an existing output directory.")
return p
def _parse_block_size(spec: str | None) -> list[int] | None:
if spec is None:
return None
parts = [int(x) for x in spec.split(",") if x.strip()]
if len(parts) != 2:
raise SystemExit(f"--weight-block-size must be 'M,N' (2 ints), got {spec!r}")
return parts
def _require_modelopt() -> Any:
try:
import modelopt.torch.quantization as mtq
except ModuleNotFoundError as exc:
raise SystemExit(
"NVIDIA ModelOpt is not installed. Install with:\n"
" pip install 'nvidia-modelopt[all]'\n"
f"Original error: {exc}"
) from exc
return mtq
def _ensure_paths(args: argparse.Namespace) -> tuple[str, Path]:
model_path = args.model
output_dir = Path(args.output).expanduser().resolve()
if output_dir.exists():
if not args.overwrite:
raise SystemExit(f"Output directory already exists: {output_dir}\nPass --overwrite to replace it.")
shutil.rmtree(output_dir)
return model_path, output_dir
def _select_dtype(name: str) -> torch.dtype:
return {"bfloat16": torch.bfloat16, "float16": torch.float16}[name]
def _build_prompts(args: argparse.Namespace) -> list[str]:
prompts = args.prompt or DEFAULT_PROMPTS
if args.calib_size <= 0:
raise SystemExit("--calib-size must be positive.")
if len(prompts) < args.calib_size:
repeats = (args.calib_size + len(prompts) - 1) // len(prompts)
prompts = (prompts * repeats)[: args.calib_size]
return prompts[: args.calib_size]
# Layers to KEEP at full precision. Wan2.2's module naming:
# - condition_embedder: time_embedder, time_proj, text_embedder, image_embedder (I2V)
# - patch_embedding: Conv3dLayer (already not Linear, belt-and-suspenders skip)
# - scale_shift_table: nn.Parameter modulation (not Linear, but pattern guard)
# - norm_out: AdaLayerNorm final
# - proj_out: final nn.Linear
# - timestep_proj_prepare / output_scale_shift_prepare: SP helpers
def _filter_func_wan22(name: str) -> bool:
pattern = re.compile(
r"(proj_out.*|"
r".*(condition_embedder|patch_embedding|"
r"norm_out|scale_shift_table|"
r"timestep_proj_prepare|output_scale_shift_prepare).*)"
)
return pattern.match(name) is not None
def _mha_filter_func(name: str) -> bool:
pattern = re.compile(
r".*(q_bmm_quantizer|k_bmm_quantizer|v_bmm_quantizer|softmax_quantizer|bmm2_output_quantizer).*"
)
return pattern.match(name) is not None
def _disable_known_problematic_quantizers(mtq: Any, backbone: torch.nn.Module, *, quantize_mha: bool) -> None:
if not hasattr(mtq, "disable_quantizer"):
return
mtq.disable_quantizer(backbone, _filter_func_wan22)
if not quantize_mha:
mtq.disable_quantizer(backbone, _mha_filter_func)
def _load_pipeline(model_path: str, dtype: torch.dtype) -> DiffusionPipeline:
pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype)
if hasattr(pipe, "set_progress_bar_config"):
pipe.set_progress_bar_config(disable=True)
pipe.to("cuda")
return pipe
def _build_forward_loop(pipe: DiffusionPipeline, args: argparse.Namespace, prompts: list[str]):
generator = torch.Generator(device="cuda")
# Try setting guidance on the pipeline's guider if present (newer diffusers APIs).
guider = getattr(pipe, "guider", None)
if guider is not None and hasattr(guider, "guidance_scale"):
try:
guider.guidance_scale = args.guidance_scale
except Exception:
pass
base_kwargs = dict(
height=args.height,
width=args.width,
num_frames=args.num_frames,
num_inference_steps=args.calib_steps,
output_type="latent",
)
def forward_loop(*_unused_args, **_unused_kwargs) -> None:
with torch.inference_mode():
for idx, prompt in enumerate(prompts):
generator.manual_seed(args.seed + idx)
# Try with guidance_scale first; fall back without on TypeError
# for pipelines that take CFG via guider config only.
try:
pipe(prompt=prompt, generator=generator, guidance_scale=args.guidance_scale, **base_kwargs)
except TypeError as exc:
if "guidance_scale" not in str(exc):
raise
pipe(prompt=prompt, generator=generator, **base_kwargs)
return forward_loop
def _summarize_export(output_dir: Path) -> None:
cfg_path = output_dir / "transformer" / "config.json"
if not cfg_path.exists():
print(f"[warn] {cfg_path} missing.", file=sys.stderr)
return
with cfg_path.open(encoding="utf-8") as f:
cfg = json.load(f)
qc = cfg.get("quantization_config")
if not isinstance(qc, dict):
print("[warn] No quantization_config in transformer/config.json.", file=sys.stderr)
return
print("Export summary:")
print(f" quant_method: {qc.get('quant_method')}")
print(f" quant_algo: {qc.get('quant_algo')}")
producer = qc.get("producer")
if isinstance(producer, dict):
print(f" producer: {producer.get('name')} {producer.get('version')}")
print(f" config path: {cfg_path}")
def _force_export_quantized_weights(backbone: torch.nn.Module, dtype: torch.dtype) -> int:
"""Convert in-memory weights of quantized modules to actual FP8 storage.
`export_hf_checkpoint` skips this step for unknown model types (Wan2.2 isn't
in ModelOpt's recognized-model registry), so we must call the per-weight
export helper ourselves. Same workaround as the HunyuanVideo-1.5 / HunyuanImage-3
calibration helpers.
"""
from modelopt.torch.export.quant_utils import (
QUANTIZATION_NONE,
get_quantization_format,
quantizer_attr_names,
weight_attr_names,
)
from modelopt.torch.export.unified_export_hf import _export_quantized_weight
exported = 0
for name, module in backbone.named_modules():
try:
quantization_format = get_quantization_format(module)
except Exception as exc:
print(f"[warn] Could not inspect quantization format for {name}: {exc}", file=sys.stderr)
continue
if quantization_format == QUANTIZATION_NONE:
continue
for weight_name in weight_attr_names(module):
quantizer_attrs = quantizer_attr_names(weight_name)
weight_quantizer = getattr(module, quantizer_attrs.weight_quantizer, None)
if weight_quantizer is None or not getattr(weight_quantizer, "is_enabled", False):
continue
_export_quantized_weight(module, dtype, weight_name)
exported += 1
return exported
def _wan22_quant_config_block(weight_block_size: list[int] | None = None) -> dict:
"""Mirror ModelOpt FP8 metadata expected by vllm-omni's adapter (#2913)."""
weights_cfg: dict = {"dynamic": False, "num_bits": 8, "type": "float"}
if weight_block_size is not None:
weights_cfg["strategy"] = "block"
weights_cfg["block_structure"] = f"{weight_block_size[0]}x{weight_block_size[1]}"
return {
"config_groups": {
"group_0": {
"input_activations": {"dynamic": False, "num_bits": 8, "type": "float"},
"weights": weights_cfg,
"targets": ["Linear"],
}
},
"ignore": [
"condition_embedder*",
"norm_out*",
"output_scale_shift_prepare*",
"patch_embedding*",
"proj_out*",
"scale_shift_table*",
"timestep_proj_prepare*",
],
"producer": {"name": "modelopt"},
"quant_algo": "FP8",
"quant_method": "modelopt",
}
def _patch_quant_config(output_dir: Path, weight_block_size: list[int] | None = None) -> None:
"""Inject quant_algo: FP8 + config_groups into each transformer's config.json
so vllm-omni's adapter (#2913) recognises the checkpoint as ModelOpt FP8.
Patches both ``transformer`` and ``transformer_2`` (A14B) when present."""
for sub in ("transformer", "transformer_2"):
cfg_path = output_dir / sub / "config.json"
if not cfg_path.exists():
continue
with cfg_path.open(encoding="utf-8") as f:
cfg = json.load(f)
new_qc = _wan22_quant_config_block(weight_block_size=weight_block_size)
existing = cfg.get("quantization_config")
if isinstance(existing, dict):
producer = existing.get("producer")
if isinstance(producer, dict):
new_qc["producer"] = producer
cfg["quantization_config"] = new_qc
with cfg_path.open("w", encoding="utf-8") as f:
json.dump(cfg, f, indent=2)
def _list_transformers(pipe: DiffusionPipeline) -> list[tuple[str, torch.nn.Module]]:
"""Return [(attr_name, module)] for every DiT on the pipeline.
Wan2.2 A14B has both ``transformer`` (low-noise) and ``transformer_2``
(high-noise). TI2V-5B has only ``transformer``.
"""
out = []
for attr in ("transformer", "transformer_2"):
mod = getattr(pipe, attr, None)
if mod is not None:
out.append((attr, mod))
return out
def _save_pipeline_with_fp8_transformer(
pipe: DiffusionPipeline,
model_path: str,
output_dir: Path,
max_shard_size: str = "5GB",
) -> None:
"""Copy source dir verbatim minus transformer dirs, then save each quantized DiT.
Handles A14B's dual transformer (``transformer`` + ``transformer_2``) as well
as the single-DiT TI2V-5B.
"""
from modelopt.torch.export.diffusers_utils import hide_quantizers_from_state_dict
src = Path(model_path)
if not src.exists():
from huggingface_hub import snapshot_download
src = Path(snapshot_download(model_path))
if output_dir.exists():
shutil.rmtree(output_dir)
shutil.copytree(src, output_dir, ignore=shutil.ignore_patterns("transformer", "transformer_2"))
for attr, backbone in _list_transformers(pipe):
transformer_out = output_dir / attr
# Pass the nn.Module (transformer), not the Pipeline wrapper.
with hide_quantizers_from_state_dict(backbone):
backbone.save_pretrained(
str(transformer_out),
safe_serialization=True,
max_shard_size=max_shard_size,
)
def main() -> None:
args = _build_parser().parse_args()
if not torch.cuda.is_available():
raise SystemExit("CUDA is required for ModelOpt FP8 quantization.")
mtq = _require_modelopt()
model_path, output_dir = _ensure_paths(args)
dtype = _select_dtype(args.dtype)
prompts = _build_prompts(args)
weight_block_size = _parse_block_size(args.weight_block_size)
print("Quantization plan:")
print(f" input: {args.model}")
print(f" output: {output_dir}")
print(f" dtype: {dtype}")
print(f" height/width: {args.height}x{args.width}")
print(f" num_frames: {args.num_frames}")
print(f" calib_size: {len(prompts)}")
print(f" calib_steps: {args.calib_steps}")
print(f" quantize_mha: {args.quantize_mha}")
print(
f" weight strategy: {'block-wise ' + str(weight_block_size) if weight_block_size else 'per-tensor (default)'}"
)
pipe = _load_pipeline(model_path, dtype)
transformers = _list_transformers(pipe)
if not transformers:
raise SystemExit("Pipeline has no transformer or transformer_2 attribute.")
print(f" found {len(transformers)} transformer(s): {', '.join(a for a, _ in transformers)}")
quant_config = copy.deepcopy(mtq.FP8_DEFAULT_CFG)
if weight_block_size is not None:
quant_config["quant_cfg"]["*weight_quantizer"] = {
"num_bits": (4, 3),
"block_sizes": {-1: weight_block_size[1], -2: weight_block_size[0]},
}
print(
f" -> overriding weight quantizer with block_sizes={weight_block_size} "
f"({weight_block_size[0]}x{weight_block_size[1]} tiles)"
)
forward_loop = _build_forward_loop(pipe, args, prompts)
# Quantize each DiT. The forward_loop runs the full pipeline, so quantizers
# on every attached transformer (A14B: transformer + transformer_2) get
# calibrated during the same passes.
for attr, backbone in transformers:
print(f"\nQuantizing {attr}...")
quantized = mtq.quantize(backbone, quant_config, forward_loop)
if quantized is not None:
setattr(pipe, attr, quantized)
_disable_known_problematic_quantizers(mtq, getattr(pipe, attr), quantize_mha=args.quantize_mha)
# Export weights to FP8 only AFTER all transformers are calibrated. Exporting
# inside the loop would convert transformer's weights to real FP8, and the
# next transformer's forward_loop (full pipeline) would then fake-quantize an
# already-FP8 weight -> "fake_e4m3fy not implemented for Float8_e4m3fn".
print("\nForcing FP8 weight serialization (Wan2.2 isn't in ModelOpt's recognized-model registry,")
print("so we have to call the per-weight export helper ourselves)...")
total_exported = 0
for attr, backbone in _list_transformers(pipe):
exported = _force_export_quantized_weights(backbone, dtype)
print(f" -> {exported} weights converted to FP8 in {attr}")
total_exported += exported
if total_exported == 0:
raise SystemExit(
"No quantized weights were exported. Calibration may have skipped every layer "
"(check the disable_quantizer regex) or `mtq.quantize` did not actually wrap any "
"weight quantizers."
)
print("\nSaving pipeline with FP8 transformer(s)...")
_save_pipeline_with_fp8_transformer(pipe, model_path, output_dir)
_patch_quant_config(output_dir, weight_block_size=weight_block_size)
print(f"Saved to: {output_dir}")
_summarize_export(output_dir)
print("\nNext: validate the checkpoint with vllm-omni:")
print(
" python examples/offline_inference/text_to_video/text_to_video.py \\\n"
f" --model {output_dir} \\\n"
" --quantization fp8 \\\n"
" --prompt 'A dog running across a field of golden wheat.' \\\n"
f" --height {args.height} --width {args.width} --num-frames {args.num_frames} \\\n"
" --num-inference-steps 30 --guidance-scale 5.0 --seed 42 \\\n"
" --output outputs/wan22_modelopt_fp8.mp4"
)
print(
"\n (--quantization fp8 is auto-upgraded to ModelOpt FP8 at runtime because the "
"checkpoint's config.json has modelopt metadata.)"
)
if __name__ == "__main__":
main()
+1 -9
View File
@@ -64,9 +64,7 @@ dev = [
"lpips>=0.1.4", # for quantization fp8 "lpips>=0.1.4", # for quantization fp8
"orjson>=3.11.7", # to make v1/embeddings API fast "orjson>=3.11.7", # to make v1/embeddings API fast
"voxcpm>=2.0.2", # for VoxCPM2 path-locating in tests "voxcpm>=2.0.2", # for VoxCPM2 path-locating in tests
"torchdiffeq>=0.2.5", # for covo_audio Test "torchdiffeq>=0.2.5" # for covo_audio Test
"FlagEmbedding", # for hunyuan_image3 tests
"kernels<0.15.0" # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
] ]
demo = [ demo = [
@@ -90,12 +88,6 @@ minicpmo = [
"stepaudio2-minicpmo", "stepaudio2-minicpmo",
] ]
# Optional Blackwell-only fused-bias FP8 GEMM; package is `quack-kernels` (imported
# as `quack`), not the unrelated `quack`. See docs/user_guide/quantization/fp8.md.
quack = [
"quack-kernels>=0.3.11",
]
docs = [ docs = [
"mkdocs>=1.5.0", "mkdocs>=1.5.0",
"mkdocs-api-autonav", "mkdocs-api-autonav",
-110
View File
@@ -1,110 +0,0 @@
# Higgs-Audio V3 TTS
> Multilingual text-to-speech with voice cloning on 1×H100
## Summary
- Vendor: Boson AI
- Model: `bosonai/higgs-audio-v3-tts-4b`
- Task: Text-to-speech synthesis with optional voice cloning (100+ languages)
- Mode: Online serving with the OpenAI-compatible `/v1/audio/speech` API; offline batch inference
- Maintainer: @yuekaiz
## When to use this recipe
Use this recipe to serve `bosonai/higgs-audio-v3-tts-4b` for high-quality
multilingual TTS. The model generates 24 kHz speech, supports zero-shot voice
cloning from a reference clip, and handles 100+ languages with inline control
tokens for emotion, style, and prosody. The architecture is a ~4B Qwen3
backbone with fused multi-codebook embedding/head (8 codebooks × 1026 vocab,
MusicGen-style delay pattern).
## References
- Model card: [bosonai/higgs-audio-v3-tts-4b](https://huggingface.co/bosonai/higgs-audio-v3-tts-4b)
- Offline example: [`examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py`](../../examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py)
- Online example: [`examples/online_serving/text_to_speech/higgs_audio_v3/`](../../examples/online_serving/text_to_speech/higgs_audio_v3/)
- Benchmark results: see Performance section below
## Hardware Support
## GPU
### 1×H100 80GB
#### Environment
- OS: Linux
- Python: 3.12+
- CUDA: 12.x
- vLLM version: 0.22.0
- vLLM-Omni version or commit: `36e048fd` (branch `higgs-v3`)
#### Command
**Online serving:**
```bash
vllm-omni serve bosonai/higgs-audio-v3-tts-4b \
--host 0.0.0.0 --port 8095 \
--trust-remote-code --omni
```
The default deploy config `vllm_omni/deploy/higgs_multimodal_qwen3.yaml` is
loaded automatically by model registry (HF `model_type=higgs_multimodal_qwen3`).
**Offline batch inference:**
```bash
python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \
--texts "Hello world." "The quick brown fox jumps over the lazy dog." \
--output-dir results/higgs_v3_wavs
```
**Offline voice clone:**
```bash
python examples/offline_inference/text_to_speech/higgs_audio_v3/end2end.py \
--texts "Text to synthesize in the cloned voice." \
--ref-audio path/to/reference.wav \
--ref-text "Transcript of the reference clip." \
--output-dir results/higgs_v3_clone
```
#### Verification
Basic TTS via curl:
```bash
curl -X POST http://localhost:8095/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{
"model": "bosonai/higgs-audio-v3-tts-4b",
"input": "Hello, how are you?"
}' --output hello.wav
```
Voice clone via Python client:
```bash
python examples/online_serving/text_to_speech/higgs_audio_v3/batch_speech_client.py \
--base-url http://localhost:8095 \
--model bosonai/higgs-audio-v3-tts-4b \
--ref-audio path/to/reference.wav \
--ref-text "Transcript of the reference." \
--prompts "Text to clone."
```
#### Notes
- Memory usage: Stage 0 (talker, ~4B) uses ~60% GPU memory; Stage 1 (codec decoder) uses ~25%.
- Key flags: `--trust-remote-code` and `--omni` are required.
- Output: 24 kHz mono WAV.
- Voice cloning: `ref_audio` accepts WAV/FLAC/MP3; `ref_text` is optional but improves fidelity.
- Deploy config: `vllm_omni/deploy/higgs_multimodal_qwen3.yaml` (auto-discovered from `model_type`).
- `max_num_seqs=16` for both stages.
- `enforce_eager=true` for both stages (CUDA graph available for stage 0 but no throughput gain at batch>1).
- Known limitations:
- CUDA graph for stage 0 helps batch=1 (-13% RTF) but not batch>1 due to model-owned sampler running outside graph.
- Stage 1 (code2wav) must use `enforce_eager=true` (`@torch.inference_mode` incompatible with graph capture).
- Async streaming (chunk-based) not yet implemented; sync-only pipeline.
+1 -3
View File
@@ -36,8 +36,7 @@ recipes/
| [`LTX/LTX-2.md`](./LTX/LTX-2.md) | Text-to-video and image-to-video serving | 1x H200 141GB | | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | Text-to-video and image-to-video serving | 1x H200 141GB |
| [`LTX/LTX-2.3.md`](./LTX/LTX-2.3.md) | Text-to-video with audio generation (22B) | 1x GPU (96GB VRAM) | | [`LTX/LTX-2.3.md`](./LTX/LTX-2.3.md) | Text-to-video with audio generation (22B) | 1x GPU (96GB VRAM) |
| [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB | | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
| [`cosmos3/Cosmos3-Nano.md`](./cosmos3/Cosmos3-Nano.md) | Text-to-image, text-to-video, image-to-video, video-to-video generation, text to video with sound, action policy | 1x H200 141GB / B300 | | [`nvidia/Cosmos3-Nano.md`](./nvidia/Cosmos3-Nano.md) | Text-to-image, text-to-video, and image-to-video generation | 1x H200 141GB / B300 |
| [`cosmos3/Cosmos3-Super.md`](./cosmos3/Cosmos3-Super.md) | 64B T2I / T2V / I2V / V2V generation (+ optional audio) / Action policy | 8x H200/H100/A100 / 2x H200 / B300 |
| [`OpenBMB/MiniCPM-o-4_5.md`](./OpenBMB/MiniCPM-o-4_5.md) | Online serving for omni multimodal chat (text / image / audio / video → text + 24 kHz speech) | 2x A100/H100 80GB / 3x mid-tier GPU / 8x RTX 4090 24GB | | [`OpenBMB/MiniCPM-o-4_5.md`](./OpenBMB/MiniCPM-o-4_5.md) | Online serving for omni multimodal chat (text / image / audio / video → text + 24 kHz speech) | 2x A100/H100 80GB / 3x mid-tier GPU / 8x RTX 4090 24GB |
| [`OpenBMB/VoxCPM2.md`](./OpenBMB/VoxCPM2.md) | Online + offline TTS with native AR pipeline (48 kHz, 30+ languages) | 1x RTX 4090 24GB | | [`OpenBMB/VoxCPM2.md`](./OpenBMB/VoxCPM2.md) | Online + offline TTS with native AR pipeline (48 kHz, 30+ languages) | 1x RTX 4090 24GB |
| [`Qwen/Qwen-Image.md`](./Qwen/Qwen-Image.md) | Text-to-image serving with step-wise continuous batching replay and ModelOpt mixed FP8/NVFP4 | 1x A100 80GB / 2x B200 | | [`Qwen/Qwen-Image.md`](./Qwen/Qwen-Image.md) | Text-to-image serving with step-wise continuous batching replay and ModelOpt mixed FP8/NVFP4 | 1x A100 80GB / 2x B200 |
@@ -50,7 +49,6 @@ recipes/
| [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 | | [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |
| [`Wan-AI/Wan2.2-I2V.md`](./Wan-AI/Wan2.2-I2V.md) | Image-to-video serving (Wan2.2 14B) | 8x Ascend NPU (A2/A3) | | [`Wan-AI/Wan2.2-I2V.md`](./Wan-AI/Wan2.2-I2V.md) | Image-to-video serving (Wan2.2 14B) | 8x Ascend NPU (A2/A3) |
| [`Wan-AI/Wan2.2-S2V.md`](./Wan-AI/Wan2.2-S2V.md) | Speech-to-video serving (Wan2.2 14B) | 2x A100/H100 80GB | | [`Wan-AI/Wan2.2-S2V.md`](./Wan-AI/Wan2.2-S2V.md) | Speech-to-video serving (Wan2.2 14B) | 2x A100/H100 80GB |
| [`StabilityAI/Stable-Diffusion-3.5.md`](./StabilityAI/Stable-Diffusion-3.5.md) | Text-to-image serving (SD 3.5-medium and SD 3.5-large) | 1x RTX A6000 48GB |
| [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB | | [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB |
| [`GLM/GLM-Image.md`](./GLM/GLM-Image.md) | Online serving for image generation | 1x A800 80GB / 2x A800 80GB | | [`GLM/GLM-Image.md`](./GLM/GLM-Image.md) | Online serving for image generation | 1x A800 80GB / 2x A800 80GB |
| |
-139
View File
@@ -1,139 +0,0 @@
# Stable Diffusion 3.5
> Text-to-image serving with Stability AI's SD 3.5 family
## Summary
- Vendor: Stability AI
- Model: `stabilityai/stable-diffusion-3.5-medium`, `stabilityai/stable-diffusion-3.5-large`
- Task: Text-to-image generation
- Mode: Online serving with the OpenAI-compatible images API
- Maintainer: Community
## When to use this recipe
Use this recipe when you want a lightweight, fast text-to-image model that
fits on a single consumer or workstation GPU. SD 3.5-medium requires only
~17 GiB VRAM and generates 1024×1024 images in ~4 seconds, making it one of
the most accessible diffusion models in vLLM-Omni. SD 3.5-large produces
noticeably sharper images at the cost of ~34 GiB VRAM and ~21 seconds per
image.
## References
- Model (medium): <https://huggingface.co/stabilityai/stable-diffusion-3.5-medium>
- Model (large): <https://huggingface.co/stabilityai/stable-diffusion-3.5-large>
- Related example under `examples/`:
[`examples/online_serving/text_to_image/README.md`](../../examples/online_serving/text_to_image/README.md)
- User guide:
[`docs/user_guide/examples/online_serving/text_to_image.md`](../../docs/user_guide/examples/online_serving/text_to_image.md)
## Hardware Support
This recipe documents two CUDA GPU configurations on a single RTX A6000 48 GB,
one for each model variant. Extend it with more hardware sections as community
validation lands.
## GPU
### 1x NVIDIA RTX A6000 48GB — stable-diffusion-3.5-medium
#### Environment
- OS: Ubuntu 22.04
- Python: 3.12
- Driver / runtime: NVIDIA driver 580.126.09, CUDA 13.0
- vLLM version: 0.22.0
- vLLM-Omni version or commit: 0.22.0rc2.dev8+g41a795b5d
#### Command
```bash
vllm serve stabilityai/stable-diffusion-3.5-medium --omni --port 8091
```
#### Verification
```bash
curl -X POST http://localhost:8091/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
"model": "stabilityai/stable-diffusion-3.5-medium",
"prompt": "a cute cat sitting on a windowsill, highly detailed, sharp focus",
"negative_prompt": "blurry, out of focus",
"guidance_scale": 4.5,
"n": 1,
"size": "1024x1024"
}' | python3 -c "
import json, base64, sys
data = json.load(sys.stdin)
with open('output.png', 'wb') as f:
f.write(base64.b64decode(data['data'][0]['b64_json']))
print('saved output.png')
"
```
Or use the example client:
```bash
python examples/online_serving/text_to_image/openai_chat_client.py \
--server http://localhost:8091 \
--prompt "a cute cat sitting on a windowsill" \
--output output.png
```
#### Notes
- Memory usage: Model weights ~15.6 GiB, peak VRAM ~17.0 GiB (no CFG), ~20.9 GiB with `guidance_scale > 1` (CFG enabled).
- Generation time: ~4 seconds without CFG, ~8 seconds with `guidance_scale=4.5`.
- Model loading time: ~55 seconds (including weight download on first run).
- `negative_prompt` requires `guidance_scale > 1` to take effect; SD3 defaults to 1.0 (no CFG).
### 1x NVIDIA RTX A6000 48GB — stable-diffusion-3.5-large
#### Environment
- OS: Ubuntu 22.04
- Python: 3.12
- Driver / runtime: NVIDIA driver 580.126.09, CUDA 13.0
- vLLM version: 0.22.0
- vLLM-Omni version or commit: 0.22.0rc2.dev8+g41a795b5d
#### Command
```bash
vllm serve stabilityai/stable-diffusion-3.5-large --omni --port 8091
```
> **Note**: `stable-diffusion-3.5-large` is a gated model. Accept the license
> on [Hugging Face](https://huggingface.co/stabilityai/stable-diffusion-3.5-large)
> and set `HF_TOKEN` before running.
#### Verification
```bash
curl -X POST http://localhost:8091/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
"model": "stabilityai/stable-diffusion-3.5-large",
"prompt": "a cute cat sitting on a windowsill, highly detailed, sharp focus",
"negative_prompt": "blurry, out of focus",
"guidance_scale": 4.5,
"n": 1,
"size": "1024x1024"
}' | python3 -c "
import json, base64, sys
data = json.load(sys.stdin)
with open('output.png', 'wb') as f:
f.write(base64.b64decode(data['data'][0]['b64_json']))
print('saved output.png')
"
```
#### Notes
- Memory usage: Peak VRAM ~31.6 GiB at 1024×1024, ~45.5 GiB at 2048×2048 with `guidance_scale=4.5`.
- Generation time: ~19 seconds at 1024×1024, ~102 seconds at 2048×2048.
- Image quality is noticeably better than medium, with sharper details.
- `negative_prompt` requires `guidance_scale > 1` to take effect; SD3 defaults to 1.0 (no CFG).
- 2048×2048 fits within 48 GB but leaves little headroom (~45.5 GiB used out of 49 GiB).
-353
View File
@@ -1,353 +0,0 @@
# Cosmos3-Nano
> Text-to-image, text-to-video, image-to-video, and video-to-video serving
## Summary
- Vendor: NVIDIA
- Model: `nvidia/Cosmos3-Nano`
- Task: Text-to-image (T2I), text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) generation, with optional synchronized audio (video + sound), action policy
- Mode: Online serving with the OpenAI-compatible image/video APIs, plus offline generation via the `Omni` API
- Maintainer: Community
## When to use this recipe
Use this recipe to deploy `nvidia/Cosmos3-Nano` for image and video generation.
A single pipeline class (`Cosmos3OmniDiffusersPipeline`) serves these modes; the
mode is selected per request:
- **T2I** — `POST /v1/images/generations` (or a prompt carrying `modalities=["image"]`).
- **T2V** — `POST /v1/videos/sync` with `num_frames > 1` and no reference image/video.
- **I2V** — `POST /v1/videos/sync` with a reference image (`input_reference` file
upload, or `image_reference` JSON).
- **V2V** — `POST /v1/videos/sync` with a reference video (`input_reference`
video upload, or `video_reference` JSON). Cosmos3 conditions on selected
reference-video latent frames; use `extra_params.condition_frame_indexes_vision`
and `extra_params.condition_video_keep` to choose which prefix/tail frames guide
generation.
- **T2VS / I2VS** — add `generate_sound=true` (and optional `sound_duration`) to a
T2V/I2V `/v1/videos/sync` request to also generate synchronized audio, muxed into
the mp4 as AAC 48 kHz stereo. See the official model card's "Video + Audio" examples.
- **Action** — pass `extra_params={"action_mode": ...}` to drive Physical-AI tasks:
- `forward_dynamics` — given a first frame or video **and** an action trajectory,
roll out the resulting video. Synchronous: `POST /v1/videos/sync`.
- `policy` — given a first frame or video and a language instruction,
**predict** the action trajectory (and a rollout video). Use the async
`POST /v1/videos` endpoint and read the predicted action from the top-level
`action` field.
- `inverse_dynamics` — given a video, **recover** the action trajectory. Use
the async `POST /v1/videos` endpoint and read the recovered action from
the top-level `action` field
(`{data, shape, dtype, raw_action_dim, domain_id}`).
Action requests also take `domain_name` (e.g. `av`, `bridge_orig_lerobot`,
`droid_lerobot`, `agibotworld`, …; or a numeric `domain_id`), `raw_action_dim`,
and `action_chunk_size` (must equal `num_frames` or `num_frames - 1`). For
`forward_dynamics` also pass the `action` array. The dedicated policy checkpoint
**`nvidia/Cosmos3-Nano-Policy-DROID`** is served the same way
(`domain_name=droid_lerobot`).
Action requests can use `input_reference` or `video_reference` for video input.
`policy` and `forward_dynamics` can also use an image reference; `inverse_dynamics`
requires a video reference.
## References
- Model card (authoritative usage + example assets): <https://huggingface.co/nvidia/Cosmos3-Nano>
- Example inputs/outputs live in the repo's `assets/` (`example_t2v_prompt.json`,
`example_i2v_prompt.json`, `example_i2v_input.jpg`, `negative_prompt.json`;
audio examples: `example_t2vs_prompt.json`, `example_t2vs_output.mp4`,
`example_i2vs_output.mp4`).
- Prompt upsampling (recommended for quality): the model expects JSON-upsampled
structured prompts; see NVIDIA's `cosmos-framework` prompt-upsampling docs.
- Pipeline: [`vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py`](../../vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py)
- Smoke tests (canonical request formats): [`tests/e2e/accuracy/test_cosmos3_similarity.py`](../../tests/e2e/accuracy/test_cosmos3_similarity.py)
## Hardware Support
## GPU
### 1x H200 141GB / B300 (Online serving)
#### Environment
- OS: Ubuntu 22.04+
- Python: 3.12+
- Driver / runtime: NVIDIA CUDA environment
- vLLM version: match the repository requirements from your current checkout
- vLLM-Omni version or commit: use the commit you are deploying from
#### Command
Requires the `vllm-omni` package (or the `vllm/vllm-omni:cosmos3` container),
which provides the `vllm serve … --omni` entrypoint used below.
Safety guardrails are **on by default** (NVIDIA Open Model License). They load
the **gated** `nvidia/Cosmos-1.0-Guardrail` model, so to keep them on you must:
1. `pip install cosmos-guardrail`
2. Accept the license at <https://huggingface.co/nvidia/Cosmos-1.0-Guardrail>
3. Export a token with access: `export HF_TOKEN=hf_...`
Then launch the recommended server:
```bash
vllm serve nvidia/Cosmos3-Nano \
--omni \
--host 0.0.0.0 --port 8000 \
--init-timeout 1800
```
To run **without** guardrails (you are responsible for license compliance),
add `--no-guardrails` (no token/`cosmos-guardrail` needed). For extra GPUs use
`--ulysses-degree N` (context parallel) or `--tensor-parallel-size N`;
`--enable-layerwise-offload` reduces VRAM on smaller GPUs. The pipeline
auto-resolves from `model_index.json`; pass
`--model-class-name Cosmos3OmniDiffusersPipeline` to force it explicitly.
#### Verification
Best quality uses the JSON-upsampled prompts from `assets/` (download with
`hf download nvidia/Cosmos3-Nano assets/ --local-dir Cosmos3-Nano`). Minimal
self-contained examples:
```bash
curl http://localhost:8000/v1/models
# Text-to-image -> /v1/images/generations (1024x1024, 50 steps; base64 PNG)
curl -sS -X POST http://localhost:8000/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/Cosmos3-Nano",
"prompt": "A photorealistic red sports car on a city street at golden hour, cinematic lighting.",
"negative_prompt": "blurry, distorted, low quality",
"size": "1024x1024", "n": 1, "response_format": "b64_json",
"num_inference_steps": 50, "guidance_scale": 7.0, "seed": 42
}' | python -c "import sys,json,base64; open('cosmos3_t2i.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
# Text-to-video -> /v1/videos/sync (720p, 189 frames @ 24fps; official params)
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=A robot arm is cleaning a plate in the kitchen" \
-F "negative_prompt=blurry, distorted, low quality, jittery, deformed" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=123" \
-o cosmos3_t2v.mp4
# Image-to-video -> /v1/videos/sync with an uploaded reference image
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=The scene comes to life with smooth, natural motion." \
-F "negative_prompt=blurry, distorted, low quality" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=1111" \
-F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
-o cosmos3_i2v.mp4
# Video-to-video -> /v1/videos/sync with an uploaded reference video.
# By default Cosmos3 conditions on latent indexes [0, 1]. For the default
# temporal VAE stride this decodes only the first 5 input frames.
# The model works best when the prompt describes the actual situation happening in the video.
# Generic prompts may create sub-standard generations.
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=Continue the same scene with smooth natural motion and consistent subjects." \
-F "negative_prompt=blurry, distorted, low quality, jittery, deformed" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true,"condition_frame_indexes_vision":[0,1],"condition_video_keep":"first"}' \
-F "seed=2222" \
-F "input_reference=@/path/to/reference.mp4;type=video/mp4" \
-o cosmos3_v2v.mp4
# V2V can also use a JSON-safe URL/data-URL video reference. Do not combine
# video_reference with input_reference or image_reference.
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=Continue the same scene with smooth natural motion and consistent subjects." \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"condition_frame_indexes_vision":[0,1],"condition_video_keep":"last"}' \
-F 'video_reference={"video_url":"https://example.com/reference.mp4"}' \
-o cosmos3_v2v_from_url.mp4
# Text-to-video-with-sound
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=The video opens with a view of a well-lit indoor fruit display. A robotic arm picks up a pear, an orange, and a carambola one by one, placing each into a plastic bag in a shopping cart with red handles. The video is 7.875 seconds long, 24 FPS, and 1280x720. Audio description: soft servo whirs, gentle fruit thuds, plastic bag rustling, and a faint refrigeration hum." \
-F "negative_prompt=blurry, distorted, low quality" \
-F "size=1280x720" \
-F "num_frames=189" \
-F "fps=24" \
-F "num_inference_steps=35" \
-F "guidance_scale=6.0" \
-F "max_sequence_length=4096" \
-F "flow_shift=10.0" \
-F "seed=0" \
-F "generate_sound=true" \
-F "sound_duration=7.875" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-o cosmos3_t2v_with_sound.mp4
# Action — forward dynamics (first frame + action trajectory -> rollout video).
# Synchronous; `action` is a JSON array shaped [action_chunk_size, raw_action_dim].
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
--form-string "model=nvidia/Cosmos3-Nano" \
--form-string "prompt=You are an autonomous vehicle. This video is captured from a first-person perspective." \
-F "input_reference=@first_frame.jpg;type=image/jpeg" \
-F "size=640x480" -F "num_frames=61" -F "fps=10" \
-F "num_inference_steps=30" -F "guidance_scale=1.0" -F "flow_shift=5.0" \
--form-string "extra_params={\"action_mode\":\"forward_dynamics\",\"domain_name\":\"av\",\"raw_action_dim\":9,\"action_chunk_size\":60,\"action\":$(cat action.json)}" \
-F "seed=0" \
-o cosmos3_forward_dynamics.mp4
# Action — policy (first frame + instruction -> predicted action trajectory + video).
# Asynchronous: POST returns a job id; poll, then read the predicted action from
# the top-level `action` field ({data, shape, dtype, raw_action_dim, domain_id}).
VIDEO_ID=$(curl -sS -X POST http://localhost:8000/v1/videos \
-H "Accept: application/json" \
--form-string "model=nvidia/Cosmos3-Nano" \
--form-string "prompt=Pick up the banana and place it in the bowl." \
-F "input_reference=@first_frame.jpg;type=image/jpeg" \
-F "size=640x480" -F "num_frames=17" -F "fps=5" \
-F "num_inference_steps=30" -F "guidance_scale=1.0" -F "flow_shift=5.0" \
--form-string 'extra_params={"action_mode":"policy","domain_name":"bridge_orig_lerobot","raw_action_dim":10,"action_chunk_size":16}' \
-F "seed=0" | jq -r '.id')
# poll until status == completed, then:
curl -sS "http://localhost:8000/v1/videos/$VIDEO_ID" | jq '.action | {shape, dtype, raw_action_dim, domain_id}'
curl -sS -L "http://localhost:8000/v1/videos/$VIDEO_ID/content" -o cosmos3_policy.mp4
# Action — inverse dynamics (video -> recovered action trajectory).
# Asynchronous: use the job metadata to read the recovered action.
VIDEO_ID=$(curl -sS -X POST http://localhost:8000/v1/videos \
-H "Accept: application/json" \
--form-string "model=nvidia/Cosmos3-Nano" \
--form-string "prompt=Recover the robot action trajectory from this clip." \
-F "input_reference=@motion_clip.mp4;type=video/mp4" \
-F "size=640x480" -F "num_frames=17" -F "fps=5" \
-F "num_inference_steps=30" -F "guidance_scale=1.0" -F "flow_shift=5.0" \
--form-string 'extra_params={"action_mode":"inverse_dynamics","domain_name":"bridge_orig_lerobot","raw_action_dim":10,"action_chunk_size":16}' \
-F "seed=0" | jq -r '.id')
# poll until status == completed, then:
curl -sS "http://localhost:8000/v1/videos/$VIDEO_ID" | jq '.action | {shape, dtype, raw_action_dim, domain_id}'
curl -sS -L "http://localhost:8000/v1/videos/$VIDEO_ID/content" -o cosmos3_inverse_dynamics.mp4
```
#### Notes
- **Measured latency (1x B300, bf16, guardrails off):**
- T2I 1024² — 10 / 25 / 50 steps → ~0.4 / 0.7 / **1.3 s**
- T2V 1280×720 @ 35 steps — 25 / 49 / 93 / **189** frames → ~7 / 15 / 33 / **~93 s**
- I2V 1280×720, 189 frames @ 35 steps → ~**99 s**
- Action 640×480 @ 30 steps — forward-dynamics 61f ~**4 s**, policy 17f ~**1–3 s**.
- Guardrails-on overhead: ~8% on T2I, negligible on video.
- **Memory:** transformer ~17 GiB (bf16); peak ~46 GiB for 720p video on 1 GPU;
full repo (transformer + Wan VAE + Qwen3-VL vision encoder + audio tokenizer)
~33 GB on disk.
- **Determinism:** identical seed reproduces identical output on the same
hardware; outputs are not bit-identical across different GPU types.
- **Supported sizes (per model card):** 256p / 480p / 720p at 16:9, 4:3, 1:1,
3:4, 9:16. Defaults: T2I 1024², 50 steps, guidance 7.0; T2V/I2V/V2V
1280×720, 189 frames, 35 steps, guidance 6.0, `flow_shift=10.0`.
- **Key flags / params:** `--no-guardrails` (server) or
`extra_params={"guardrails":false}` (per request) toggles safety. The
per-request flag only takes effect when the server was launched **with**
guardrails enabled (it cannot re-enable them on a `--no-guardrails` server).
`use_resolution_template` / `use_duration_template` are off by default and only
needed when not using upsampled prompts that already encode resolution/duration.
For V2V, `condition_frame_indexes_vision` selects the clean conditioned latent
frame indexes (default `[0, 1]`), and `condition_video_keep` selects whether the
API decodes the first or last needed reference frames (`"first"` by default).
- **Known limitations:**
- Guardrails-on requires `cosmos-guardrail` **and** access to the gated
`nvidia/Cosmos-1.0-Guardrail` repo (accept license + `HF_TOKEN`); otherwise
the server fails at pipeline build with a gated-repo / safety-checker error.
- A guardrail-blocked prompt currently returns HTTP 500
(`"Guardrail blocked prompt"`).
- Action `forward_dynamics`, `policy`, and `inverse_dynamics` are supported
online. Use async `POST /v1/videos` when you need the predicted/recovered
action payload under the top-level `action` field; sync `/v1/videos/sync`
returns raw MP4 bytes and does not expose action metadata in the response body.
### 1x GPU (Offline generation)
#### Environment
- OS: Ubuntu 22.04+
- Python: 3.12+
- Driver / runtime: NVIDIA CUDA environment
- vLLM-Omni version or commit: use the commit you are deploying from
#### Command
```python
# cosmos3_offline.py — run with: python cosmos3_offline.py
import torch
from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
def main():
omni = Omni(
model="nvidia/Cosmos3-Nano",
model_class_name="Cosmos3OmniDiffusersPipeline",
trust_remote_code=True,
enforce_eager=True,
# Guardrails are disabled here for a quick local run; install
# cosmos-guardrail + gated-repo access and drop this to enable them.
model_config={"guardrails": False},
)
gen = torch.Generator(device="cpu").manual_seed(42)
# Text-to-image (modalities=["image"]). For T2V use modalities=["video"]
# plus num_frames/fps; for I2V add multi_modal_data={"image": <PIL.Image>}.
outputs = omni.generate(
{
"prompt": "A photorealistic red sports car at golden hour, cinematic lighting.",
"negative_prompt": "blurry, distorted, low quality",
"modalities": ["image"],
},
OmniDiffusionSamplingParams(
height=1024, width=1024, generator=gen,
guidance_scale=7.0, num_inference_steps=50, num_outputs_per_prompt=1,
),
)
outputs[0].request_output.images[0].save("cosmos3_t2i.png")
omni.close()
if __name__ == "__main__":
main()
```
#### Verification
```bash
python cosmos3_offline.py
python -c "from PIL import Image; im=Image.open('cosmos3_t2i.png'); print('image', im.size, im.mode)"
```
#### Notes
- Same `Cosmos3OmniDiffusersPipeline` as online; mode is chosen by
`prompt["modalities"]` (`["image"]` → T2I, `["video"]` → T2V) plus
`num_frames`/`fps`, `multi_modal_data={"image": ...}` for I2V, and
`multi_modal_data={"video": [<PIL frames>]}` or a video tensor/array for V2V.
For video, frames are returned in `outputs[0].request_output.images` as an
`(B, F, H, W, 3)` array.
- The offline entry must be guarded by `if __name__ == "__main__":` — the engine
spawns workers with the `spawn` start method.
-124
View File
@@ -1,124 +0,0 @@
# Cosmos3-Super
> Frontier 64B world model: text-to-image, text-to-video, image-to-video, video-to-video (+ optional audio)
## Summary
- Vendor: NVIDIA
- Model: `nvidia/Cosmos3-Super` (64B; also `Cosmos3-Super-Text2Image`, `Cosmos3-Super-Image2Video`)
- Task: T2I, T2V, I2V, V2V generation, with optional synchronized audio (video + sound)
- Mode: Online serving with the OpenAI-compatible image/video APIs
- Maintainer: Community
## When to use this recipe
Use this recipe to deploy the 64B `nvidia/Cosmos3-Super` for the highest-quality
Cosmos3 generation. It shares the same `Cosmos3OmniDiffusersPipeline` and request
formats as [Cosmos3-Nano](./Cosmos3-Nano.md) — only the checkpoint size and the
recommended parallelism differ. Mode is selected per request (T2I →
`/v1/images/generations`; T2V/I2V/V2V → `/v1/videos/sync`; add
`generate_sound=true` for audio).
## References
- Model card (authoritative usage + example assets): <https://huggingface.co/nvidia/Cosmos3-Super>
- Nano recipe (same APIs/params): [`Cosmos3-Nano.md`](./Cosmos3-Nano.md)
- Pipeline: [`vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py`](../../vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py)
## Hardware Support
## GPU
Requires the `vllm-omni` package (or the `vllm/vllm-omni:cosmos3` container),
which provides the `vllm serve … --omni` entrypoint used below.
### 8x H200/H100/A100 (recommended, per model card)
```bash
vllm serve nvidia/Cosmos3-Super \
--omni \
--host 0.0.0.0 --port 8000 \
--cfg-parallel-size 2 \
--ulysses-degree 4 \
--use-hsdp --hsdp-shard-size 8 \
--init-timeout 1800
```
### 2x H200 / B300 (minimum)
```bash
vllm serve nvidia/Cosmos3-Super \
--omni \
--host 0.0.0.0 --port 8000 \
--cfg-parallel-size 2 \
--use-hsdp --hsdp-shard-size 2 \
--init-timeout 1800
```
Guardrails are on by default (gated `nvidia/Cosmos-1.0-Guardrail` — `pip install
cosmos-guardrail`, accept the license, set `HF_TOKEN`); add `--no-guardrails` to
disable. `--enable-layerwise-offload` reduces VRAM on smaller GPUs.
#### Verification
Requests are identical to Nano (see [`Cosmos3-Nano.md`](./Cosmos3-Nano.md) for full
T2I/T2V/I2V/V2V/T2VS curls); official params: `size=1280x720, num_frames=189,
fps=24, num_inference_steps=35, guidance_scale=6.0, flow_shift=10.0,
max_sequence_length=4096`.
```bash
curl http://localhost:8000/v1/models
# T2V (official prompt assets give best quality)
curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Super" -F "prompt=A robot arm is cleaning a plate in the kitchen" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
-F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=17" -o cosmos3_super_t2v.mp4
# I2V — add an uploaded reference image
curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Super" -F "prompt=The scene comes to life with smooth, natural motion." \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
-F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=1111" -F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
-o cosmos3_super_i2v.mp4
# V2V — add an uploaded reference video. condition_video_keep can be "first" or "last".
curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Super" -F "prompt=Continue the same scene with smooth natural motion." \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
-F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"condition_frame_indexes_vision":[0,1],"condition_video_keep":"first"}' \
-F "seed=2222" -F "input_reference=@/path/to/reference.mp4;type=video/mp4" \
-o cosmos3_super_v2v.mp4
# T2V + sound — add generate_sound/sound_duration (output muxes AAC 48 kHz stereo)
curl -sS -X POST http://localhost:8000/v1/videos/sync -H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Super" -F "prompt=A robot arm is cleaning a plate in the kitchen" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" -F "num_inference_steps=35" \
-F "guidance_scale=6.0" -F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F "generate_sound=true" -F "sound_duration=7.875" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=17" -o cosmos3_super_t2vs.mp4
```
#### Notes
- **Measured (2x B300, bf16, guardrails off, official 2-GPU config above):**
- T2I 1024², 50 steps → **~6 s**
- T2V 1280×720, 189 frames, 35 steps → **~197 s**
- I2V 1280×720, 189 frames, 35 steps → **~200 s**
- T2V + sound (189 frames, 35 steps) → **~198 s**, output muxes **AAC 48 kHz stereo**
- (NVIDIA's reference: 8×H200 @ 50 steps ≈ 55 s/video; 2×H200 @ 35 steps ≈ 3 min/video.)
- **Memory:** ~61.5 GiB per GPU when sharded across 2 GPUs (HSDP shard 2); repo ~135 GB on disk.
- Same generation defaults, supported sizes, V2V reference-video controls
(`condition_frame_indexes_vision`, `condition_video_keep`), and
`generate_sound`/`sound_duration`
semantics as Nano, including the **action** modality: `forward_dynamics`,
`policy`, and `inverse_dynamics` — see the Cosmos3-Nano recipe for the request
shapes. Use async `/v1/videos` when you need predicted/recovered action metadata
under the top-level `action` field. Verified on the 64B Super under
`--cfg-parallel-size 2`: async `policy` returns the predicted action (`[16, 10]`)
and the rollout video reliably.
+6 -9
View File
@@ -57,6 +57,12 @@ Adjust `devices` in the YAML to match your hardware.
#### Command #### Command
Thinker only (text output):
```bash
vllm serve Jonathan1909/Ming-flash-omni-2.0 --omni --port 8091
```
Thinker + talker (text and/or audio output): Thinker + talker (text and/or audio output):
```bash ```bash
@@ -66,15 +72,6 @@ vllm serve Jonathan1909/Ming-flash-omni-2.0 \
--log-stats --log-stats
``` ```
Thinker only (text-only output):
```bash
vllm serve Jonathan1909/Ming-flash-omni-2.0 \
--omni \
--deploy-config vllm_omni/deploy/ming_flash_omni_thinker_only.yaml \
--port 8091
```
`--log-stats` is optional but recommended while validating the deployment. `--log-stats` is optional but recommended while validating the deployment.
#### Verification #### Verification
+215
View File
@@ -0,0 +1,215 @@
# Cosmos3-Nano
> Text-to-image, text-to-video, and image-to-video serving
## Summary
- Vendor: NVIDIA
- Model: `nvidia/Cosmos3-Nano`
- Task: Text-to-image (T2I), text-to-video (T2V), and image-to-video (I2V) generation
- Mode: Online serving with the OpenAI-compatible image/video APIs, plus offline generation via the `Omni` API
- Maintainer: Community
## When to use this recipe
Use this recipe to deploy `nvidia/Cosmos3-Nano` for image and video generation.
A single pipeline class (`Cosmos3OmniDiffusersPipeline`) serves all three modes;
the mode is selected per request:
- **T2I** — `POST /v1/images/generations` (or a prompt carrying `modalities=["image"]`).
- **T2V** — `POST /v1/videos/sync` with `num_frames > 1` and no reference image.
- **I2V** — `POST /v1/videos/sync` with a reference image (`input_reference` file
upload, or `image_reference` JSON).
## References
- Model card (authoritative usage + example assets): <https://huggingface.co/nvidia/Cosmos3-Nano>
- Example inputs/outputs live in the repo's `assets/` (`example_t2v_prompt.json`,
`example_i2v_prompt.json`, `example_i2v_input.jpg`, `negative_prompt.json`).
- Prompt upsampling (recommended for quality): the model expects JSON-upsampled
structured prompts; see NVIDIA's `cosmos-framework` prompt-upsampling docs.
- Pipeline: [`vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py`](../../vllm_omni/diffusion/models/cosmos3/pipeline_cosmos3.py)
- Smoke tests (canonical request formats): [`tests/e2e/accuracy/test_cosmos3_similarity.py`](../../tests/e2e/accuracy/test_cosmos3_similarity.py)
## Hardware Support
## GPU
### 1x H200 141GB / B300 (Online serving)
#### Environment
- OS: Ubuntu 22.04+
- Python: 3.12+
- Driver / runtime: NVIDIA CUDA environment
- vLLM version: match the repository requirements from your current checkout
- vLLM-Omni version or commit: use the commit you are deploying from
#### Command
Safety guardrails are **on by default** (NVIDIA Open Model License). They load
the **gated** `nvidia/Cosmos-1.0-Guardrail` model, so to keep them on you must:
1. `pip install cosmos-guardrail`
2. Accept the license at <https://huggingface.co/nvidia/Cosmos-1.0-Guardrail>
3. Export a token with access: `export HF_TOKEN=hf_...`
Then launch the recommended server:
```bash
vllm serve nvidia/Cosmos3-Nano \
--omni \
--host 0.0.0.0 --port 8000 \
--init-timeout 1800
```
To run **without** guardrails (you are responsible for license compliance),
add `--no-guardrails` (no token/`cosmos-guardrail` needed). For extra GPUs use
`--ulysses-degree N` (context parallel) or `--tensor-parallel-size N`;
`--enable-layerwise-offload` reduces VRAM on smaller GPUs. The pipeline
auto-resolves from `model_index.json`; pass
`--model-class-name Cosmos3OmniDiffusersPipeline` to force it explicitly.
#### Verification
Best quality uses the JSON-upsampled prompts from `assets/` (download with
`hf download nvidia/Cosmos3-Nano assets/ --local-dir Cosmos3-Nano`). Minimal
self-contained examples:
```bash
curl http://localhost:8000/v1/models
# Text-to-image -> /v1/images/generations (1024x1024, 50 steps; base64 PNG)
curl -sS -X POST http://localhost:8000/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/Cosmos3-Nano",
"prompt": "A photorealistic red sports car on a city street at golden hour, cinematic lighting.",
"negative_prompt": "blurry, distorted, low quality",
"size": "1024x1024", "n": 1, "response_format": "b64_json",
"num_inference_steps": 50, "guidance_scale": 7.0, "seed": 42
}' | python -c "import sys,json,base64; open('cosmos3_t2i.png','wb').write(base64.b64decode(json.load(sys.stdin)['data'][0]['b64_json']))"
# Text-to-video -> /v1/videos/sync (720p, 189 frames @ 24fps; official params)
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=A robot arm is cleaning a plate in the kitchen" \
-F "negative_prompt=blurry, distorted, low quality, jittery, deformed" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=123" \
-o cosmos3_t2v.mp4
# Image-to-video -> /v1/videos/sync with an uploaded reference image
curl -sS -X POST http://localhost:8000/v1/videos/sync \
-H "Accept: video/mp4" \
-F "model=nvidia/Cosmos3-Nano" \
-F "prompt=The scene comes to life with smooth, natural motion." \
-F "negative_prompt=blurry, distorted, low quality" \
-F "size=1280x720" -F "num_frames=189" -F "fps=24" \
-F "num_inference_steps=35" -F "guidance_scale=6.0" \
-F "max_sequence_length=4096" -F "flow_shift=10.0" \
-F 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
-F "seed=1111" \
-F "input_reference=@/path/to/reference.jpg;type=image/jpeg" \
-o cosmos3_i2v.mp4
```
#### Notes
- **Measured latency (1x B300, bf16, guardrails off):**
- T2I 1024² — 10 / 25 / 50 steps → ~0.4 / 0.7 / **1.3 s**
- T2V 1280×720 @ 35 steps — 25 / 49 / 93 / **189** frames → ~7 / 15 / 33 / **~93 s**
- I2V 1280×720, 189 frames @ 35 steps → ~**99 s**
- Guardrails-on overhead: ~8% on T2I, negligible on video.
- **Memory:** transformer ~17 GiB (bf16); peak ~46 GiB for 720p video on 1 GPU;
full repo (transformer + Wan VAE + Qwen3-VL vision encoder + audio tokenizer)
~33 GB on disk.
- **Determinism:** identical seed reproduces identical output on the same
hardware; outputs are not bit-identical across different GPU types.
- **Supported sizes (per model card):** 256p / 480p / 720p at 16:9, 4:3, 1:1,
3:4, 9:16. Defaults: T2I 1024², 50 steps, guidance 7.0; T2V/I2V 1280×720,
189 frames, 35 steps, guidance 6.0, `flow_shift=10.0`.
- **Key flags / params:** `--no-guardrails` (server) or
`extra_params={"guardrails":false}` (per request) toggles safety;
`use_resolution_template` / `use_duration_template` are off by default and only
needed when not using upsampled prompts that already encode resolution/duration.
- **Known limitations:**
- Guardrails-on requires `cosmos-guardrail` **and** access to the gated
`nvidia/Cosmos-1.0-Guardrail` repo (accept license + `HF_TOKEN`); otherwise
the server fails at pipeline build with a gated-repo / safety-checker error.
- A guardrail-blocked prompt currently returns HTTP 500
(`"Guardrail blocked prompt"`).
- Video + audio, and action (policy / forward- / inverse-dynamics) modalities
are not part of this integration yet.
### 1x GPU (Offline generation)
#### Environment
- OS: Ubuntu 22.04+
- Python: 3.12+
- Driver / runtime: NVIDIA CUDA environment
- vLLM-Omni version or commit: use the commit you are deploying from
#### Command
```python
# cosmos3_offline.py — run with: python cosmos3_offline.py
import torch
from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
def main():
omni = Omni(
model="nvidia/Cosmos3-Nano",
model_class_name="Cosmos3OmniDiffusersPipeline",
trust_remote_code=True,
enforce_eager=True,
# Keep guardrails on by installing cosmos-guardrail + gated-repo access;
# this disables them for a quick local run.
model_config={"guardrails": False},
)
gen = torch.Generator(device="cpu").manual_seed(42)
# Text-to-image (modalities=["image"]). For T2V use modalities=["video"]
# plus num_frames/fps; for I2V add multi_modal_data={"image": <PIL.Image>}.
outputs = omni.generate(
{
"prompt": "A photorealistic red sports car at golden hour, cinematic lighting.",
"negative_prompt": "blurry, distorted, low quality",
"modalities": ["image"],
},
OmniDiffusionSamplingParams(
height=1024, width=1024, generator=gen,
guidance_scale=7.0, num_inference_steps=50, num_outputs_per_prompt=1,
),
)
outputs[0].request_output.images[0].save("cosmos3_t2i.png")
omni.close()
if __name__ == "__main__":
main()
```
#### Verification
```bash
python cosmos3_offline.py
python -c "from PIL import Image; im=Image.open('cosmos3_t2i.png'); print('image', im.size, im.mode)"
```
#### Notes
- Same `Cosmos3OmniDiffusersPipeline` as online; mode is chosen by
`prompt["modalities"]` (`["image"]` → T2I, `["video"]` → T2V) plus
`num_frames`/`fps`, and `multi_modal_data={"image": ...}` for I2V. For video,
frames are returned in `outputs[0].request_output.images` as an
`(B, F, H, W, 3)` array.
- The offline entry must be guarded by `if __name__ == "__main__":` — the engine
spawns workers with the `spawn` start method.
Binary file not shown.

Before

Width:  |  Height:  |  Size: 1.1 MiB

@@ -345,14 +345,6 @@ def test_parse_score_payload_handles_raw_json_and_delimited_json():
assert parse_score_payload(wrapped)["score"] == [6] assert parse_score_payload(wrapped)["score"] == [6]
def test_parse_score_payload_handles_qwen_vl_nested_score_dicts():
qwen_style = '{"score": [{"naturalness": 8, "artifact_free": 7}], "reasoning": "good quality"}'
assert parse_score_payload(qwen_style)["score"] == [8, 7]
nested_values = '{"score": [{"score": 9}, {"score": 6}], "reasoning": "ok"}'
assert parse_score_payload(nested_values)["score"] == [9, 6]
def test_summarize_gedit_generated_records_groups_by_task_and_language(): def test_summarize_gedit_generated_records_groups_by_task_and_language():
records = [] records = []
for group in GEDIT_GROUPS[:2]: for group in GEDIT_GROUPS[:2]:
@@ -61,7 +61,7 @@ def test_stage0_streaming_update_discards_outstanding_async_placeholder_token()
sched._update_request_as_session(session, _make_update([10, 20])) sched._update_request_as_session(session, _make_update([10, 20]))
assert session.async_tokens_to_discard == 1 assert session.discard_latest_async_tokens is True
assert session.num_output_placeholders == 0 assert session.num_output_placeholders == 0
assert session.spec_token_ids == [] assert session.spec_token_ids == []
# The async placeholder makes token 9 unconfirmed, so only 7 and 8 are # The async placeholder makes token 9 unconfirmed, so only 7 and 8 are
@@ -83,7 +83,7 @@ def test_stage0_streaming_update_keeps_all_computed_tokens_without_placeholder()
sched._update_request_as_session(session, _make_update([10, 20])) sched._update_request_as_session(session, _make_update([10, 20]))
assert getattr(session, "async_tokens_to_discard", 0) == 0 assert session.discard_latest_async_tokens is False
assert session.num_output_placeholders == 0 assert session.num_output_placeholders == 0
assert session.prompt_token_ids == [1, 2, 3, 7, 8, 9, 10, 20] assert session.prompt_token_ids == [1, 2, 3, 7, 8, 9, 10, 20]
assert list(session._all_token_ids) == [1, 2, 3, 7, 8, 9, 10, 20] assert list(session._all_token_ids) == [1, 2, 3, 7, 8, 9, 10, 20]
@@ -1,7 +1,7 @@
[ [
{ {
"test_name": "test_hunyuan_image3_ti2i_inline", "test_name": "test_hunyuan_image3_ti2i_tp4_tp4_inline",
"description": "AR (TP=2) + 2 * DiT (TP=2) - using inline dataset and deploy-config", "description": "AR (TP=4) + DiT (TP=4) - using inline dataset and deploy-config",
"server_type": "vllm-omni", "server_type": "vllm-omni",
"benchmark_endpoint": "/v1/images/edits", "benchmark_endpoint": "/v1/images/edits",
"server_params": { "server_params": {
@@ -21,7 +21,7 @@
"zmq_port": 50051, "zmq_port": 50051,
"protocol": "rdma", "protocol": "rdma",
"device_name": "", "device_name": "",
"memory_pool_size": 17179869184, "memory_pool_size": 4294967296,
"memory_pool_device": "cpu" "memory_pool_device": "cpu"
} }
} }
@@ -32,13 +32,13 @@
"is_comprehension": true, "is_comprehension": true,
"final_output": true, "final_output": true,
"final_output_type": "text", "final_output_type": "text",
"max_num_seqs": 16, "max_num_seqs": 4,
"gpu_memory_utilization": 0.95, "gpu_memory_utilization": 0.95,
"enforce_eager": false, "enforce_eager": false,
"enable_prefix_caching": false, "enable_prefix_caching": false,
"max_num_batched_tokens": 32768, "max_num_batched_tokens": 32768,
"devices": "0,1", "devices": "0,1,2,3",
"tensor_parallel_size": 2, "tensor_parallel_size": 4,
"hf_overrides": { "hf_overrides": {
"rope_parameters": { "rope_parameters": {
"mrope_section": [0, 32, 32], "mrope_section": [0, 32, 32],
@@ -65,14 +65,13 @@
"stage_id": 1, "stage_id": 1,
"max_num_seqs": 1, "max_num_seqs": 1,
"enforce_eager": true, "enforce_eager": true,
"devices": "2,3", "devices": "4,5,6,7",
"distributed_executor_backend": "mp", "distributed_executor_backend": "mp",
"omni_kv_config": { "omni_kv_config": {
"need_recv_cache": true "need_recv_cache": true
}, },
"parallel_config": { "parallel_config": {
"tensor_parallel_size": 2, "tensor_parallel_size": 4,
"vae_patch_parallel_size": 2,
"enable_expert_parallel": true "enable_expert_parallel": true
}, },
"input_connectors": { "input_connectors": {
@@ -115,8 +114,8 @@
}, },
"task": "ti2i", "task": "ti2i",
"bot-task": "think_recaption", "bot-task": "think_recaption",
"max-concurrency": 16, "max-concurrency": 4,
"num-prompts": 96, "num-prompts": 16,
"skip-performance-assertion": true, "skip-performance-assertion": true,
"baseline": { "baseline": {
"throughput_qps": 0.1, "throughput_qps": 0.1,
@@ -23,10 +23,10 @@
"ignore_eos": true, "ignore_eos": true,
"percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration", "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
"baseline": { "baseline": {
"mean_ttft_ms": [203, 283, 470, 869], "mean_ttft_ms": [1000, 2000, 3000, 4000],
"mean_tpot_ms": [28, 36, 41, 45], "mean_tpot_ms": [200, 200, 200, 200],
"mean_audio_ttfp_ms": [861, 1200, 1528, 2100], "mean_audio_ttfp_ms": [10000, 15000, 20000, 25000],
"mean_audio_rtf": [0.24, 0.30, 0.35, 0.39] "mean_audio_rtf": [1.5, 1.5, 1.5, 1.5]
} }
}, },
{ {
@@ -49,9 +49,9 @@
}, },
"percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration", "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
"baseline": { "baseline": {
"mean_ttft_ms": [377], "mean_ttft_ms": [2000],
"mean_audio_ttfp_ms": [684], "mean_audio_ttfp_ms": [2000],
"mean_audio_rtf": [0.14] "mean_audio_rtf": [0.25]
} }
}, },
{ {
@@ -76,9 +76,9 @@
}, },
"percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration", "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
"baseline": { "baseline": {
"mean_ttft_ms": [228], "mean_ttft_ms": [6000],
"mean_audio_ttfp_ms": [686], "mean_audio_ttfp_ms": [6000],
"mean_audio_rtf": [0.16] "mean_audio_rtf": [0.7]
} }
}, },
{ {
@@ -105,9 +105,9 @@
}, },
"percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration", "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
"baseline": { "baseline": {
"mean_ttft_ms": [522], "mean_ttft_ms": [12000],
"mean_audio_ttfp_ms": [1119], "mean_audio_ttfp_ms": [12000],
"mean_audio_rtf": [0.21] "mean_audio_rtf": [1.0]
} }
} }
] ]
@@ -31,6 +31,11 @@
], ],
"random_input_len": 2500, "random_input_len": 2500,
"random_output_len": 900, "random_output_len": 900,
"temperature": 0.4,
"top_p": 0.9,
"top_k": 1,
"seed": 42,
"repetition_penalty": 1.05,
"ignore_eos": true, "ignore_eos": true,
"extra_body": { "extra_body": {
"modalities": [ "modalities": [
@@ -69,6 +74,11 @@
"random_input_len": 100, "random_input_len": 100,
"random_output_len": 100, "random_output_len": 100,
"random_range_ratio": 0.0, "random_range_ratio": 0.0,
"temperature": 0.4,
"top_p": 0.9,
"top_k": 1,
"seed": 42,
"repetition_penalty": 1.05,
"ignore_eos": true, "ignore_eos": true,
"extra_body": { "extra_body": {
"modalities": [ "modalities": [
@@ -107,6 +117,11 @@
"random_input_len": 100, "random_input_len": 100,
"random_output_len": 100, "random_output_len": 100,
"random_range_ratio": 0.0, "random_range_ratio": 0.0,
"temperature": 0.4,
"top_p": 0.9,
"top_k": 1,
"seed": 42,
"repetition_penalty": 1.05,
"ignore_eos": true, "ignore_eos": true,
"extra_body": { "extra_body": {
"modalities": [ "modalities": [
@@ -147,6 +162,11 @@
"random_input_len": 100, "random_input_len": 100,
"random_output_len": 100, "random_output_len": 100,
"random_range_ratio": 0.0, "random_range_ratio": 0.0,
"temperature": 0.4,
"top_p": 0.9,
"top_k": 1,
"seed": 42,
"repetition_penalty": 1.05,
"ignore_eos": true, "ignore_eos": true,
"extra_body": { "extra_body": {
"modalities": [ "modalities": [
@@ -1,46 +1,19 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Tests for the DiffusersPipelineLoader.
"""
from types import SimpleNamespace from types import SimpleNamespace
import pytest import pytest
import torch import torch
import torch.nn as nn import torch.nn as nn
from huggingface_hub import snapshot_download
from vllm.config.load import LoadConfig
from vllm_omni.diffusion.config import get_current_diffusion_config, get_current_diffusion_config_or_none from vllm_omni.diffusion.config import get_current_diffusion_config, get_current_diffusion_config_or_none
from vllm_omni.diffusion.data import OmniDiffusionConfig
from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter from vllm_omni.diffusion.model_loader.gguf_adapters import get_gguf_adapter
from vllm_omni.diffusion.models.helios import HeliosPipeline
from vllm_omni.diffusion.registry import initialize_model from vllm_omni.diffusion.registry import initialize_model
pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu] pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
model_path = "hf-internal-testing/tiny-helios-modular-pipe"
@pytest.fixture(scope="module")
def prefetch_helios_model():
"""Downloads the tiny helios model prior to running a test."""
snapshot_download(model_path)
@pytest.fixture(scope="function")
def mock_tp_group(mocker):
"""Mocks the tensor parallel group; this is needed to initialize the Helios model."""
mocker.patch("vllm.model_executor.layers.linear.get_tensor_model_parallel_world_size", return_value=1)
mocker.patch("vllm.model_executor.layers.linear.get_tensor_model_parallel_rank", return_value=0)
mock_group = mocker.MagicMock()
mock_group.world_size = 1
mock_group.rank_in_group = 0
mocker.patch("vllm.distributed.parallel_state.get_tp_group", return_value=mock_group)
class _DummyPipelineModel(nn.Module): class _DummyPipelineModel(nn.Module):
def __init__(self, *, source_prefix: str): def __init__(self, *, source_prefix: str):
@@ -69,13 +42,7 @@ class _DummyPipelineModel(nn.Module):
def _make_loader_with_weights(weight_names: list[str]) -> DiffusersPipelineLoader: def _make_loader_with_weights(weight_names: list[str]) -> DiffusersPipelineLoader:
od_config = SimpleNamespace( loader = object.__new__(DiffusersPipelineLoader)
dtype=torch.float32,
parallel_config=SimpleNamespace(use_hsdp=False),
quantization_config=None,
)
loader = DiffusersPipelineLoader(LoadConfig(), od_config)
loader.counter_before_loading_weights = 0.0 loader.counter_before_loading_weights = 0.0
loader.counter_after_loading_weights = 0.0 loader.counter_after_loading_weights = 0.0
@@ -183,15 +150,16 @@ def test_load_model_custom_pipeline_sets_current_diffusion_config(monkeypatch):
quantization_config=None, quantization_config=None,
) )
loader = DiffusersPipelineLoader(LoadConfig(), od_config) loader = object.__new__(DiffusersPipelineLoader)
loader.load_weights = lambda model: None # type: ignore[assignment] loader.load_weights = lambda model: None # type: ignore[assignment]
loader._process_weights_after_loading = lambda model, target_device: None # type: ignore[assignment] loader._process_weights_after_loading = lambda model, target_device: None # type: ignore[assignment]
loader._is_gguf_quantization = lambda: False # type: ignore[assignment] loader._is_gguf_quantization = lambda _od_config: False # type: ignore[assignment]
monkeypatch.setattr(loader_mod, "resolve_obj_by_qualname", lambda _name: _ConfigAwareModel) monkeypatch.setattr(loader_mod, "resolve_obj_by_qualname", lambda _name: _ConfigAwareModel)
monkeypatch.setattr(loader_mod.torch, "device", lambda _name: _DeviceContext("cpu")) monkeypatch.setattr(loader_mod.torch, "device", lambda _name: _DeviceContext("cpu"))
model = loader.load_model( model = loader.load_model(
od_config,
load_device="cpu", load_device="cpu",
load_format="custom_pipeline", load_format="custom_pipeline",
custom_pipeline_name="tests.dummy.ConfigAwarePipeline", custom_pipeline_name="tests.dummy.ConfigAwarePipeline",
@@ -200,33 +168,3 @@ def test_load_model_custom_pipeline_sets_current_diffusion_config(monkeypatch):
assert model.captured_config is od_config assert model.captured_config is od_config
assert model.seen_config_during_init is od_config assert model.seen_config_during_init is od_config
assert get_current_diffusion_config_or_none() is None assert get_current_diffusion_config_or_none() is None
def test_get_all_weights(prefetch_helios_model, mock_tp_group):
"""Ensure that get all weights on a tiny model resolves to nonempty weights."""
od_config = OmniDiffusionConfig(
model_class_name="HeliosPipeline",
model=model_path,
)
loader = DiffusersPipelineLoader(
load_config=LoadConfig(),
od_config=od_config,
)
pipeline = HeliosPipeline(od_config=od_config)
weights = list(loader.get_all_weights(pipeline))
assert len(weights) > 0
def test_load_model(prefetch_helios_model, mock_tp_group):
"""Ensure that load model creates an instance of the expected pipeline class."""
od_config = OmniDiffusionConfig(
model_class_name="HeliosPipeline",
model=model_path,
)
loader = DiffusersPipelineLoader(
load_config=LoadConfig(),
od_config=od_config,
)
model = loader.load_model(load_device="cpu")
assert isinstance(model, HeliosPipeline)
@@ -62,7 +62,6 @@ def _make_loader() -> DiffusersPipelineLoader:
ignore_patterns=["*.tmp"], ignore_patterns=["*.tmp"],
) )
loader.od_config = None loader.od_config = None
loader.quant_config = None
loader.counter_before_loading_weights = 0.0 loader.counter_before_loading_weights = 0.0
loader.counter_after_loading_weights = 0.0 loader.counter_after_loading_weights = 0.0
return loader return loader
@@ -70,47 +69,53 @@ def _make_loader() -> DiffusersPipelineLoader:
def test_is_gguf_quantization_accepts_dict_config(): def test_is_gguf_quantization_accepts_dict_config():
loader = _make_loader() loader = _make_loader()
loader.quant_config = { od_config = SimpleNamespace(
"method": "gguf", quantization_config={
"gguf_model": "weights.gguf", "method": "gguf",
} "gguf_model": "weights.gguf",
}
)
assert loader._is_gguf_quantization() is True assert loader._is_gguf_quantization(od_config) is True
def test_is_gguf_quantization_rejects_non_gguf_dict_config(): def test_is_gguf_quantization_rejects_non_gguf_dict_config():
loader = _make_loader() loader = _make_loader()
loader.quant_config = { od_config = SimpleNamespace(
"method": "fp8", quantization_config={
"gguf_model": "weights.gguf", "method": "fp8",
} "gguf_model": "weights.gguf",
}
)
assert loader._is_gguf_quantization() is False assert loader._is_gguf_quantization(od_config) is False
def test_is_gguf_quantization_requires_gguf_model_for_dict_config(): def test_is_gguf_quantization_requires_gguf_model_for_dict_config():
loader = _make_loader() loader = _make_loader()
loader.quant_config = {"method": "gguf"} od_config = SimpleNamespace(quantization_config={"method": "gguf"})
with pytest.raises(ValueError, match="gguf_model"): with pytest.raises(ValueError, match="gguf_model"):
loader._is_gguf_quantization() loader._is_gguf_quantization(od_config)
def test_is_gguf_quantization_accepts_object_config(): def test_is_gguf_quantization_accepts_object_config():
loader = _make_loader() loader = _make_loader()
loader.quant_config = SimpleNamespace( od_config = SimpleNamespace(
get_name=lambda: "gguf", quantization_config=SimpleNamespace(
gguf_model="weights.gguf", get_name=lambda: "gguf",
gguf_model="weights.gguf",
)
) )
assert loader._is_gguf_quantization() is True assert loader._is_gguf_quantization(od_config) is True
def test_is_gguf_quantization_uses_fallback_object_without_get_name(): def test_is_gguf_quantization_uses_fallback_object_without_get_name():
loader = _make_loader() loader = _make_loader()
loader.quant_config = SimpleNamespace(gguf_model="weights.gguf") od_config = SimpleNamespace(quantization_config=SimpleNamespace(gguf_model="weights.gguf"))
assert loader._is_gguf_quantization() is True assert loader._is_gguf_quantization(od_config) is True
def test_get_model_loadable_names_collects_parameters_and_buffers(): def test_get_model_loadable_names_collects_parameters_and_buffers():
@@ -166,14 +171,16 @@ def test_resolve_gguf_model_path_downloads_by_quant_type(monkeypatch: pytest.Mon
def test_get_gguf_weights_iterator_prefixes_source_names(monkeypatch: pytest.MonkeyPatch): def test_get_gguf_weights_iterator_prefixes_source_names(monkeypatch: pytest.MonkeyPatch):
loader = _make_loader() loader = _make_loader()
loader.quant_config = SimpleNamespace(gguf_model="weights.gguf")
loader.od_config = SimpleNamespace(revision="main")
source = DiffusersPipelineLoader.ComponentSource( source = DiffusersPipelineLoader.ComponentSource(
model_or_path="dummy", model_or_path="dummy",
subfolder="transformer", subfolder="transformer",
revision=None, revision=None,
prefix="transformer.", prefix="transformer.",
) )
od_config = SimpleNamespace(
revision="main",
quantization_config=SimpleNamespace(gguf_model="weights.gguf"),
)
class _Adapter: class _Adapter:
def weights_iterator(self): def weights_iterator(self):
@@ -183,7 +190,7 @@ def test_get_gguf_weights_iterator_prefixes_source_names(monkeypatch: pytest.Mon
monkeypatch.setattr(loader, "_resolve_gguf_model_path", lambda gguf_model, revision: "resolved.gguf") monkeypatch.setattr(loader, "_resolve_gguf_model_path", lambda gguf_model, revision: "resolved.gguf")
monkeypatch.setattr(loader_module, "get_gguf_adapter", lambda gguf_file, model, source, od_config: _Adapter()) monkeypatch.setattr(loader_module, "get_gguf_adapter", lambda gguf_file, model, source, od_config: _Adapter())
weights = list(loader._get_gguf_weights_iterator(source, object())) weights = list(loader._get_gguf_weights_iterator(source, object(), od_config))
assert weights[0][0] == "transformer.weight" assert weights[0][0] == "transformer.weight"
assert torch.equal(weights[0][1], torch.ones((2, 2))) assert torch.equal(weights[0][1], torch.ones((2, 2)))
@@ -194,12 +201,13 @@ def test_get_gguf_weights_iterator_prefixes_source_names(monkeypatch: pytest.Mon
def test_load_weights_with_gguf_falls_back_only_for_missing_transformer_weights(monkeypatch: pytest.MonkeyPatch): def test_load_weights_with_gguf_falls_back_only_for_missing_transformer_weights(monkeypatch: pytest.MonkeyPatch):
loader = _make_loader() loader = _make_loader()
model = _DummyModel() model = _DummyModel()
od_config = SimpleNamespace()
monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources)) monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources))
monkeypatch.setattr( monkeypatch.setattr(
loader, loader,
"_get_gguf_weights_iterator", "_get_gguf_weights_iterator",
lambda source, model: iter([("transformer.weight", torch.ones((2, 2), dtype=torch.float32))]), lambda source, model, od_config: iter([("transformer.weight", torch.ones((2, 2), dtype=torch.float32))]),
) )
def _hf_iter(source): def _hf_iter(source):
@@ -214,7 +222,7 @@ def test_load_weights_with_gguf_falls_back_only_for_missing_transformer_weights(
lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"}, lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"},
) )
loaded = loader._load_weights_with_gguf(model) loaded = loader._load_weights_with_gguf(model, od_config)
assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"} assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"}
assert model.calls[0] == ["transformer.weight"] assert model.calls[0] == ["transformer.weight"]
@@ -225,12 +233,13 @@ def test_load_weights_with_gguf_falls_back_only_for_missing_transformer_weights(
def test_load_weights_with_gguf_skips_transformer_fallback_when_gguf_is_complete(monkeypatch: pytest.MonkeyPatch): def test_load_weights_with_gguf_skips_transformer_fallback_when_gguf_is_complete(monkeypatch: pytest.MonkeyPatch):
loader = _make_loader() loader = _make_loader()
model = _DummyModel() model = _DummyModel()
od_config = SimpleNamespace()
monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources)) monkeypatch.setattr(loader, "_get_weight_sources", lambda _model: tuple(model.weights_sources))
monkeypatch.setattr( monkeypatch.setattr(
loader, loader,
"_get_gguf_weights_iterator", "_get_gguf_weights_iterator",
lambda source, model: iter( lambda source, model, od_config: iter(
[ [
("transformer.weight", torch.ones((2, 2), dtype=torch.float32)), ("transformer.weight", torch.ones((2, 2), dtype=torch.float32)),
("transformer.bias", torch.zeros(2, dtype=torch.float32)), ("transformer.bias", torch.zeros(2, dtype=torch.float32)),
@@ -253,7 +262,7 @@ def test_load_weights_with_gguf_skips_transformer_fallback_when_gguf_is_complete
lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"}, lambda _model: {"transformer.weight", "transformer.bias", "vae.weight"},
) )
loaded = loader._load_weights_with_gguf(model) loaded = loader._load_weights_with_gguf(model, od_config)
assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"} assert loaded == {"transformer.weight", "transformer.bias", "vae.weight"}
assert hf_calls == ["vae"] assert hf_calls == ["vae"]
@@ -73,42 +73,14 @@ class StubCosmos3VAE:
return (latents,) return (latents,)
class StubCosmos3AVAE:
def __init__(self, **kwargs: Any) -> None:
self.kwargs = kwargs
self.sample_rate = int(kwargs["sample_rate"])
self.audio_channels = int(kwargs["audio_channels"])
self.latent_ch = int(kwargs["io_channels"])
self.temporal_compression_factor = int(kwargs["hop_size"])
def get_latent_num_samples(self, num_audio_samples: int) -> int:
return int(num_audio_samples) // self.temporal_compression_factor
def get_audio_num_samples(self, num_latent_samples: int) -> int:
return int(num_latent_samples) * self.temporal_compression_factor
def decode(self, latents: torch.Tensor) -> torch.Tensor:
return torch.zeros(latents.shape[0], self.audio_channels, 8)
class StubCosmos3Transformer(nn.Module): class StubCosmos3Transformer(nn.Module):
def __init__( def __init__(
self, self,
*, *,
latent_channel_size: int = 2, latent_channel_size: int = 2,
sound_gen: bool = False,
sound_dim: int = 3,
sound_latent_fps: float = 25.0,
action_gen: bool = False,
action_dim: int = 4,
) -> None: ) -> None:
super().__init__() super().__init__()
self.latent_channel_size = latent_channel_size self.latent_channel_size = latent_channel_size
self.sound_gen = sound_gen
self.sound_dim = sound_dim
self.sound_latent_fps = sound_latent_fps
self.action_gen = action_gen
self.action_dim = action_dim
self.cached_kv: Any | None = None self.cached_kv: Any | None = None
self.cached_freqs_gen: Any | None = None self.cached_freqs_gen: Any | None = None
self.calls: list[dict[str, Any]] = [] self.calls: list[dict[str, Any]] = []
@@ -127,9 +99,8 @@ class StubCosmos3Transformer(nn.Module):
text_ids: torch.Tensor, text_ids: torch.Tensor,
text_mask: torch.Tensor, text_mask: torch.Tensor,
**kwargs: Any, **kwargs: Any,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: ) -> torch.Tensor:
token = int(text_ids.reshape(-1)[0].item()) if text_ids.numel() else 0 token = int(text_ids.reshape(-1)[0].item()) if text_ids.numel() else 0
sound_latents = kwargs.get("sound_latents")
self.calls.append( self.calls.append(
{ {
"token": token, "token": token,
@@ -143,13 +114,7 @@ class StubCosmos3Transformer(nn.Module):
marker = torch.tensor([token], dtype=torch.float32) marker = torch.tensor([token], dtype=torch.float32)
self.cached_kv = [(marker, marker + 100)] self.cached_kv = [(marker, marker + 100)]
self.cached_freqs_gen = (marker + 200, marker + 300) self.cached_freqs_gen = (marker + 200, marker + 300)
action_latents = kwargs.get("action_latents") return torch.full_like(hidden_states, float(token))
outputs: list[torch.Tensor] = [torch.full_like(hidden_states, float(token))]
if action_latents is not None:
outputs.append(torch.full_like(action_latents, float(token + 20)))
if sound_latents is not None:
outputs.append(torch.full_like(sound_latents, float(token + 10)))
return outputs[0] if len(outputs) == 1 else tuple(outputs)
def passthrough_progress_bar(iterable): def passthrough_progress_bar(iterable):
@@ -190,7 +155,6 @@ def make_cosmos3_pipeline():
pipeline._guidance_scale = None pipeline._guidance_scale = None
pipeline._num_timesteps = None pipeline._num_timesteps = None
pipeline._cache_dit_requires_paired_cfg = False pipeline._cache_dit_requires_paired_cfg = False
pipeline._sound_tokenizer = None
pipeline.progress_bar = passthrough_progress_bar pipeline.progress_bar = passthrough_progress_bar
return pipeline return pipeline
@@ -249,109 +213,7 @@ def test_pipeline_registered_and_exported() -> None:
assert "Cosmos3OmniDiffusersPipeline" in cosmos3.__all__ assert "Cosmos3OmniDiffusersPipeline" in cosmos3.__all__
@pytest.fixture def test_preprocess_i2v_image_input() -> None:
def stub_real_pipeline_init(monkeypatch: pytest.MonkeyPatch):
from vllm_omni.diffusion.models.cosmos3 import pipeline_cosmos3
class _StubAutoTokenizer:
@classmethod
def from_pretrained(cls, *args, **kwargs):
return SimpleNamespace()
class _StubDiffusersVAE:
config = SimpleNamespace(scale_factor_temporal=4, scale_factor_spatial=8)
@classmethod
def from_pretrained(cls, *args, **kwargs):
return cls()
def to(self, _device):
return self
class _StubDiffusersScheduler:
config = SimpleNamespace(flow_shift=1.0)
@classmethod
def from_pretrained(cls, *args, **kwargs):
return cls()
class _StubVideoProcessor:
def __init__(self, *args, **kwargs) -> None:
pass
monkeypatch.setattr(pipeline_cosmos3, "AutoTokenizer", _StubAutoTokenizer)
monkeypatch.setattr(pipeline_cosmos3, "DistributedAutoencoderKLWan", _StubDiffusersVAE)
monkeypatch.setattr(pipeline_cosmos3, "UniPCMultistepScheduler", _StubDiffusersScheduler)
monkeypatch.setattr(pipeline_cosmos3, "VideoProcessor", _StubVideoProcessor)
monkeypatch.setattr(pipeline_cosmos3, "get_local_device", lambda: torch.device("cpu"))
def _make_od_config(*, sound_gen: bool) -> SimpleNamespace:
tf_model_config = {
"hidden_size": 8,
"num_hidden_layers": 0,
"num_attention_heads": 2,
"num_key_value_heads": 2,
"head_dim": 4,
"intermediate_size": 16,
"vocab_size": 32,
"latent_patch_size": 1,
"latent_channel": 2,
"rope_scaling": {"mrope_section": [1, 1, 0]},
}
if sound_gen:
tf_model_config["sound_gen"] = True
return SimpleNamespace(
enable_cpu_offload=False,
enable_diffusion_pipeline_profiler=False,
model="/nonexistent/model/path",
dtype=torch.float32,
flow_shift=None,
quantization_config=None,
custom_pipeline_args={},
model_config={},
tf_model_config=tf_model_config,
)
def test_pipeline_init_skips_tokenizer_when_sound_disabled(stub_real_pipeline_init) -> None:
from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import Cosmos3OmniDiffusersPipeline
pipeline = Cosmos3OmniDiffusersPipeline(od_config=_make_od_config(sound_gen=False))
assert pipeline._sound_tokenizer is None
assert pipeline.transformer.sound_gen is False
assert not hasattr(pipeline.transformer, "audio_proj_in")
assert not hasattr(pipeline.transformer, "audio_proj_out")
def test_pipeline_init_passes_tokenizer_attrs_into_transformer(
stub_real_pipeline_init,
monkeypatch: pytest.MonkeyPatch,
) -> None:
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import Cosmos3OmniDiffusersPipeline
stub_tokenizer = sound_tokenizer.Cosmos3SoundTokenizer(
StubCosmos3AVAE(sample_rate=32000, audio_channels=2, io_channels=5, hop_size=800)
)
monkeypatch.setattr(
sound_tokenizer.Cosmos3SoundTokenizer,
"from_config",
classmethod(lambda cls, od_config: stub_tokenizer),
)
pipeline = Cosmos3OmniDiffusersPipeline(od_config=_make_od_config(sound_gen=True))
assert pipeline._sound_tokenizer is stub_tokenizer
assert pipeline.transformer.sound_gen is True
assert pipeline.transformer.sound_dim == pipeline._sound_tokenizer.latent_ch == 5
assert pipeline.transformer.sound_latent_fps == pipeline._sound_tokenizer.latent_fps == 40.0
assert pipeline.transformer.audio_proj_in.in_features == 5
assert pipeline.transformer.audio_proj_out.out_features == 5
def test_preprocess_i2v_image_and_action_video_inputs() -> None:
from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import get_cosmos3_pre_process_func from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import get_cosmos3_pre_process_func
preprocess = get_cosmos3_pre_process_func(SimpleNamespace()) preprocess = get_cosmos3_pre_process_func(SimpleNamespace())
@@ -364,31 +226,8 @@ def test_preprocess_i2v_image_and_action_video_inputs() -> None:
assert (result.sampling_params.height, result.sampling_params.width) == (672, 1344) assert (result.sampling_params.height, result.sampling_params.width) == (672, 1344)
assert tuple(result.prompts[0]["additional_information"]["preprocessed_image"].shape[-2:]) == (672, 1344) assert tuple(result.prompts[0]["additional_information"]["preprocessed_image"].shape[-2:]) == (672, 1344)
frames = [Image.new("RGB", (8, 4), color) for color in ("red", "green", "blue")]
action = SimpleNamespace(
prompts=[{"prompt": "Move.", "multi_modal_data": {"video": frames}}],
sampling_params=SimpleNamespace(height=16, width=32, extra_args={"action_mode": "forward_dynamics"}),
)
additional = preprocess(action).prompts[0]["additional_information"] def test_postprocess_handles_image_video_and_validation() -> None:
assert tuple(additional["preprocessed_image"].shape) == (1, 3, 16, 32)
assert tuple(additional["preprocessed_video"].shape) == (1, 3, 3, 16, 32)
frames = [Image.new("RGB", (8, 4), color) for color in ("red", "green", "blue", "yellow", "purple", "black")]
v2v = SimpleNamespace(
prompts=[{"prompt": "Continue.", "multi_modal_data": {"video": frames}}],
sampling_params=SimpleNamespace(
height=16,
width=32,
extra_args={"condition_frame_indexes_vision": [0, 1], "condition_video_keep": "last"},
),
)
additional = preprocess(v2v).prompts[0]["additional_information"]
assert tuple(additional["preprocessed_video"].shape) == (1, 3, 5, 16, 32)
assert additional["condition_frame_indexes_vision"] == [0, 1]
def test_postprocess_handles_image_video_audio_and_validation() -> None:
from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import get_cosmos3_post_process_func from vllm_omni.diffusion.models.cosmos3.pipeline_cosmos3 import get_cosmos3_post_process_func
func = get_cosmos3_post_process_func(SimpleNamespace()) func = get_cosmos3_post_process_func(SimpleNamespace())
@@ -396,16 +235,6 @@ def test_postprocess_handles_image_video_audio_and_validation() -> None:
assert func(video, output_type="latent") is video assert func(video, output_type="latent") is video
assert func({"image": video})[0].size == (4, 4) assert func({"image": video})[0].size == (4, 4)
# Video-only postprocess returns the bare processed video (not a dict),
# matching the image/latent branches and peer audio-capable pipelines.
assert not isinstance(func({"video": video}), dict)
assert (
func(
{"video": video, "audio": torch.ones(1, 2, 16), "audio_sample_rate": 48000},
sampling_params=SimpleNamespace(extra_args={"resolved_frame_rate": 12}),
)["audio_sample_rate"]
== 48000
)
with pytest.raises(ValueError, match="text-to-image postprocess expects"): with pytest.raises(ValueError, match="text-to-image postprocess expects"):
func({"image": torch.zeros(1, 3, 2, 4, 4)}) func({"image": torch.zeros(1, 3, 2, 4, 4)})
@@ -464,7 +293,7 @@ def test_prompt_formatting_and_checkpoint_key_remap(make_cosmos3_pipeline) -> No
assert {key: Cosmos3OmniDiffusersPipeline._remap_ckpt_key(key) for key in remaps} == remaps assert {key: Cosmos3OmniDiffusersPipeline._remap_ckpt_key(key) for key in remaps} == remaps
def test_prepare_latents_for_video_image_sound_and_action(make_cosmos3_pipeline) -> None: def test_prepare_latents_for_video_and_image(make_cosmos3_pipeline) -> None:
pipeline = make_cosmos3_pipeline() pipeline = make_cosmos3_pipeline()
latents = pipeline._prepare_latents(16, 24, 5, torch.Generator(device="cpu").manual_seed(0)) latents = pipeline._prepare_latents(16, 24, 5, torch.Generator(device="cpu").manual_seed(0))
assert latents.shape == (1, 2, 2, 2, 3) assert latents.shape == (1, 2, 2, 2, 3)
@@ -477,49 +306,8 @@ def test_prepare_latents_for_video_image_sound_and_action(make_cosmos3_pipeline)
assert velocity_mask.tolist() == [[[[[0.0]], [[1.0]]]]] assert velocity_mask.tolist() == [[[[[0.0]], [[1.0]]]]]
assert image_latent.shape == (1, 2, 1, 2, 3) assert image_latent.shape == (1, 2, 1, 2, 3)
pipeline._encode_video_tensor = lambda *args, **kwargs: torch.full((1, 2, 3, 2, 3), 6.0)
v2v_latents, v2v_velocity_mask, v2v_condition = pipeline._prepare_latents_v2v(
torch.zeros(1, 3, 5, 16, 24),
16,
24,
9,
torch.Generator(device="cpu").manual_seed(0),
[0, 1],
)
torch.testing.assert_close(v2v_latents[:, :, 0:2], torch.full((1, 2, 2, 2, 3), 6.0))
assert v2v_velocity_mask.tolist() == [[[[[0.0]], [[0.0]], [[1.0]]]]]
assert v2v_condition.shape == (1, 2, 3, 2, 3)
pipeline.transformer = pipeline.transformer.__class__(latent_channel_size=2, sound_gen=True, sound_dim=3) def test_diffuse_covers_cfg_and_i2v_steps(make_cosmos3_pipeline) -> None:
pipeline._sound_tokenizer = SimpleNamespace(
sample_rate=10,
latent_ch=3,
hop_size=4,
decode=lambda x: torch.ones(x.shape[0], 2, 24),
)
assert pipeline._resolve_sound_target_samples(SimpleNamespace(extra_args={"sound_duration": 2.0}), 9, 3.0) == (
20,
2.0,
10,
)
sound_latents, latent_frames = pipeline._prepare_sound_latents(21, torch.Generator(device="cpu").manual_seed(0))
assert (sound_latents.shape, latent_frames) == (torch.Size([1, 3, 6]), 6)
assert pipeline._decode_sound_latents(torch.zeros(1, 3, 6), target_audio_samples=21).shape == (1, 2, 21)
pipeline.transformer = pipeline.transformer.__class__(action_gen=True, action_dim=4)
action, action_mask, clean, raw_dim = pipeline._prepare_action_latents(
mode="forward_dynamics",
action_chunk_size=2,
raw_action_dim=None,
generator=torch.Generator(device="cpu").manual_seed(0),
sp=SimpleNamespace(extra_args={"action": [[1.0, 2.0], [3.0, 4.0]]}),
)
assert raw_dim == 2
assert action_mask.tolist() == [[[0.0], [0.0]]]
torch.testing.assert_close(action, clean)
def test_diffuse_covers_cfg_i2v_and_multimodal_steps(make_cosmos3_pipeline) -> None:
pipeline = make_cosmos3_pipeline() pipeline = make_cosmos3_pipeline()
latents = torch.zeros(1, 2, 1, 1, 1) latents = torch.zeros(1, 2, 1, 1, 1)
@@ -551,23 +339,6 @@ def test_diffuse_covers_cfg_i2v_and_multimodal_steps(make_cosmos3_pipeline) -> N
) )
torch.testing.assert_close(i2v[:, :, 0:1], torch.full((1, 2, 1, 1, 1), 7.0)) torch.testing.assert_close(i2v[:, :, 0:1], torch.full((1, 2, 1, 1, 1), 7.0))
pipeline.transformer = pipeline.transformer.__class__(latent_channel_size=2, action_gen=True, action_dim=4)
video_result, action_result = pipeline.diffuse(
latents=latents,
action_latents=torch.zeros(1, 3, 4),
action_velocity_mask=torch.ones(1, 3, 1),
action_condition_latents=torch.zeros(1, 3, 4),
timesteps=torch.tensor([7, 3]),
cond_ids=_ids(2),
cond_mask=_mask(),
uncond_ids=_ids(1),
uncond_mask=_mask(),
guidance_scale=1.0,
shared_kwargs={"video_shape": (1, 1, 1), "fps": 24.0, "action_domain_ids": torch.tensor([0])},
)
torch.testing.assert_close(video_result, torch.full_like(latents, 4.0))
torch.testing.assert_close(action_result, torch.full((), 44.0).expand_as(action_result))
def test_diffuse_keeps_paired_cfg_when_cache_dit_active(make_cosmos3_pipeline) -> None: def test_diffuse_keeps_paired_cfg_when_cache_dit_active(make_cosmos3_pipeline) -> None:
"""With cache-dit active the uncond pass is kept even outside the guidance """With cache-dit active the uncond pass is kept even outside the guidance
@@ -624,12 +395,7 @@ class TestForwardRouting:
def fake_diffuse(**kwargs): def fake_diffuse(**kwargs):
captured["diffuse_calls"].append(kwargs) captured["diffuse_calls"].append(kwargs)
outputs = [kwargs["latents"] + len(captured["diffuse_calls"])] return kwargs["latents"] + len(captured["diffuse_calls"])
if kwargs.get("action_latents") is not None:
outputs.append(kwargs["action_latents"] + 3.0)
if kwargs.get("sound_latents") is not None:
outputs.append(kwargs["sound_latents"] + 2.0)
return outputs[0] if len(outputs) == 1 else tuple(outputs)
pipeline._format_and_tokenize_prompts = fake_format pipeline._format_and_tokenize_prompts = fake_format
pipeline._prepare_latents = fake_prepare pipeline._prepare_latents = fake_prepare
@@ -671,7 +437,7 @@ class TestForwardRouting:
assert captured["flow_shifts"] == expected["flow"] assert captured["flow_shifts"] == expected["flow"]
assert [call[0] for call in pipeline.scheduler.set_timesteps_calls] == expected["steps"] assert [call[0] for call in pipeline.scheduler.set_timesteps_calls] == expected["steps"]
def test_forward_i2v_sound_and_action_routes(self, make_cosmos3_pipeline) -> None: def test_forward_i2v_route(self, make_cosmos3_pipeline) -> None:
pipeline = make_cosmos3_pipeline() pipeline = make_cosmos3_pipeline()
captured = self._install_forward_stubs(pipeline) captured = self._install_forward_stubs(pipeline)
image_tensor = torch.zeros(1, 3, 16, 16) image_tensor = torch.zeros(1, 3, 16, 16)
@@ -696,83 +462,11 @@ class TestForwardRouting:
) )
assert captured["diffuse_calls"][-1]["shared_kwargs"]["noisy_frame_mask"] is velocity_mask assert captured["diffuse_calls"][-1]["shared_kwargs"]["noisy_frame_mask"] is velocity_mask
video_tensor = torch.zeros(1, 3, 5, 16, 16)
v2v_condition = torch.full((1, 2, 2, 1, 1), 4.0)
v2v_mask = torch.tensor([[[[[0.0]], [[1.0]]]]])
pipeline._prepare_latents_v2v = lambda *args, **kwargs: (
torch.zeros(1, 2, 2, 1, 1),
v2v_mask,
v2v_condition,
)
pipeline.forward(
SimpleNamespace(
prompts=[
{
"prompt": "continue",
"modalities": ["video"],
"additional_information": {
"preprocessed_video": video_tensor,
"condition_frame_indexes_vision": [0],
},
}
],
sampling_params=make_sampling_params(height=16, width=16, num_frames=5),
)
)
assert captured["flow_shifts"][-1] == 10.0
assert captured["format"]["negative_prompt"] == ""
assert captured["diffuse_calls"][-1]["shared_kwargs"]["noisy_frame_mask"] is v2v_mask
assert captured["diffuse_calls"][-1]["condition_latents"] is v2v_condition
pipeline.transformer = pipeline.transformer.__class__(latent_channel_size=2, sound_gen=True, sound_dim=3)
sound_latents = torch.zeros(1, 3, 4)
pipeline._resolve_sound_target_samples = lambda *args: (20, 2.0, 10)
pipeline._prepare_sound_latents = lambda *args: (sound_latents, 4)
pipeline._decode_sound_latents = lambda *args: torch.ones(1, 2, 20)
output = pipeline.forward(
SimpleNamespace(
prompts=[{"prompt": "A robot", "modalities": ["video"], "generate_sound": True}],
sampling_params=make_sampling_params(num_frames=9, frame_rate=3.0),
)
)
assert captured["diffuse_calls"][-1]["sound_latents"] is sound_latents
assert output.output["audio_sample_rate"] == 10
pipeline.transformer = pipeline.transformer.__class__(latent_channel_size=2, action_gen=True, action_dim=4)
output = pipeline.forward(
SimpleNamespace(
prompts=[
{
"prompt": "Pick the block.",
"modalities": ["video"],
"additional_information": {"preprocessed_image": image_tensor},
}
],
sampling_params=make_sampling_params(
height=16,
width=16,
extra_args={
"action_mode": "policy",
"action_chunk_size": 2,
"raw_action_dim": 2,
"domain_name": "bridge_orig_lerobot",
},
),
)
)
assert captured["diffuse_calls"][-1]["shared_kwargs"]["action_domain_ids"].tolist() == [7]
assert output.custom_output["action"].shape == (1, 2, 2)
@pytest.mark.parametrize( @pytest.mark.parametrize(
("prompt", "sampling_params", "message"), ("prompt", "sampling_params", "message"),
[ [
(["one", "two"], make_sampling_params(), "single prompt"), (["one", "two"], make_sampling_params(), "single prompt"),
([{"prompt": "one", "modalities": ["image", "video"]}], make_sampling_params(), "both image and video"), ([{"prompt": "one", "modalities": ["image", "video"]}], make_sampling_params(), "both image and video"),
(
[{"prompt": "x", "modalities": ["image"], "generate_sound": True}],
make_sampling_params(),
"only for video",
),
], ],
) )
def test_forward_rejects_invalid_public_requests( def test_forward_rejects_invalid_public_requests(
@@ -783,7 +477,6 @@ class TestForwardRouting:
message, message,
) -> None: ) -> None:
pipeline = make_cosmos3_pipeline() pipeline = make_cosmos3_pipeline()
pipeline.transformer = pipeline.transformer.__class__(latent_channel_size=2, sound_gen=True, sound_dim=3)
with pytest.raises(ValueError, match=message): with pytest.raises(ValueError, match=message):
pipeline.forward(SimpleNamespace(prompts=prompt, sampling_params=sampling_params)) pipeline.forward(SimpleNamespace(prompts=prompt, sampling_params=sampling_params))
@@ -1,236 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import json
from pathlib import Path
from types import SimpleNamespace
import pytest
import torch
pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME = "diffusion_pytorch_model.safetensors"
class _FakeAVAEAudioTokenizer:
def __init__(self, **kwargs) -> None:
self.kwargs = kwargs
self.sample_rate = int(kwargs["sample_rate"])
self.audio_channels = int(kwargs["audio_channels"])
self.latent_ch = int(kwargs["io_channels"])
self.temporal_compression_factor = int(kwargs["hop_size"])
def get_latent_num_samples(self, num_audio_samples: int) -> int:
return int(num_audio_samples) // self.temporal_compression_factor
def get_audio_num_samples(self, num_latent_samples: int) -> int:
return int(num_latent_samples) * self.temporal_compression_factor
def decode(self, latents: torch.Tensor) -> torch.Tensor:
return torch.zeros(latents.shape[0], self.audio_channels, 8)
def _write_component(root: Path, config: dict | None = None, checkpoint_name: str | None = None) -> Path:
tokenizer_dir = root / "sound_tokenizer"
tokenizer_dir.mkdir(parents=True)
if checkpoint_name:
(tokenizer_dir / checkpoint_name).write_bytes(b"stub")
(tokenizer_dir / "config.json").write_text(json.dumps(config or {}), encoding="utf-8")
return tokenizer_dir
def _patch_fake_avae(monkeypatch: pytest.MonkeyPatch, created: dict) -> None:
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
class FakeAVAE(_FakeAVAEAudioTokenizer):
def __init__(self, **kwargs) -> None:
created.update(kwargs)
super().__init__(**kwargs)
monkeypatch.setattr(sound_tokenizer, "Cosmos3AVAEAudioTokenizer", FakeAVAE)
monkeypatch.setattr(sound_tokenizer, "get_local_device", lambda: torch.device("cpu"))
def test_from_config_loads_local_diffusers_component(tmp_path, monkeypatch: pytest.MonkeyPatch) -> None:
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
model_dir = tmp_path / "model"
tokenizer_dir = _write_component(model_dir, checkpoint_name=DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
created = {}
_patch_fake_avae(monkeypatch, created)
tokenizer = sound_tokenizer.Cosmos3SoundTokenizer.from_config(
SimpleNamespace(
model=str(model_dir),
custom_pipeline_args={"sound_sample_rate": 32000, "sound_hop_size": 800, "sound_dim": 3},
dtype=torch.float32,
)
)
assert created["checkpoint_path"] == str(tokenizer_dir / DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
assert created["config_path"] == str(tokenizer_dir / "config.json")
assert (tokenizer.sample_rate, tokenizer.latent_ch, tokenizer.hop_size, tokenizer.latent_fps) == (
32000,
3,
800,
40.0,
)
def test_from_config_downloads_component_from_hf_repo(tmp_path, monkeypatch: pytest.MonkeyPatch) -> None:
import huggingface_hub
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
cache_dir = tmp_path / "hf"
_write_component(cache_dir, checkpoint_name=DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
calls = []
created = {}
_patch_fake_avae(monkeypatch, created)
def fake_snapshot_download(repo_id: str, *, revision: str | None, allow_patterns: list[str]) -> str:
calls.append((repo_id, revision, allow_patterns))
return str(cache_dir)
monkeypatch.setattr(huggingface_hub, "snapshot_download", fake_snapshot_download)
sound_tokenizer.Cosmos3SoundTokenizer.from_config(
SimpleNamespace(
model="nvidia/cosmos3",
revision="test-rev",
custom_pipeline_args={"sound_sample_rate": 32000, "sound_hop_size": 800, "sound_dim": 3},
dtype=torch.float32,
)
)
assert created["checkpoint_path"].endswith(DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
assert calls == [
(
"nvidia/cosmos3",
"test-rev",
["sound_tokenizer/config.json", f"sound_tokenizer/{DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME}"],
)
]
@pytest.mark.parametrize(
("checkpoint_name", "message"),
[
(None, "no AVAE sound tokenizer checkpoint"),
("model.safetensors", DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME),
],
)
def test_default_component_requires_diffusers_checkpoint_name(tmp_path, checkpoint_name, message) -> None:
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
model_dir = tmp_path / "model"
_write_component(model_dir, checkpoint_name=checkpoint_name)
with pytest.raises(ValueError, match=message):
sound_tokenizer.Cosmos3SoundTokenizer.from_config(
SimpleNamespace(model=str(model_dir), custom_pipeline_args={}, dtype=torch.float32)
)
def test_component_config_precedence_and_conflict_detection(tmp_path, monkeypatch: pytest.MonkeyPatch) -> None:
from vllm_omni.diffusion.models.cosmos3 import sound_tokenizer
component_config = {
"sampling_rate": 48000,
"dec_out_channels": 2,
"vocoder_input_dim": 64,
"hop_size": 1920,
}
model_dir = tmp_path / "model"
_write_component(model_dir, component_config, DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME)
created = {}
_patch_fake_avae(monkeypatch, created)
tokenizer = sound_tokenizer.Cosmos3SoundTokenizer.from_config(
SimpleNamespace(
model=str(model_dir),
custom_pipeline_args={
"sound_normalize_latents": True,
"sound_normalization_type": "tanh",
"sound_tanh_input_scale": 2.0,
},
model_config={
"sound_tokenizer": {
"sample_rate": 32000,
"audio_channels": 1,
"io_channels": 3,
"hop_size": 800,
"normalize_latents": False,
"normalization_type": "none",
}
},
dtype=torch.float32,
)
)
assert (created["sample_rate"], created["audio_channels"], created["io_channels"], created["hop_size"]) == (
48000,
2,
64,
1920,
)
assert (created["normalize_latents"], created["normalization_type"], created["tanh_input_scale"]) == (
True,
"tanh",
2.0,
)
assert (tokenizer.sample_rate, tokenizer.latent_ch, tokenizer.hop_size, tokenizer.latent_fps) == (
48000,
64,
1920,
25.0,
)
with pytest.raises(ValueError, match=r"sample_rate.*48000.*32000"):
sound_tokenizer.Cosmos3SoundTokenizer.from_config(
SimpleNamespace(
model=str(model_dir),
custom_pipeline_args={"sound_sample_rate": 32000},
dtype=torch.float32,
)
)
def test_avae_uses_diffusers_decoder_state_dict_layout(tmp_path) -> None:
from safetensors.torch import save_file
from vllm_omni.diffusion.models.cosmos3.audio_tokenizer import avae
config = {
"sampling_rate": 8000,
"hop_size": 2,
"dec_dim": 4,
"dec_c_mults": [1],
"dec_strides": [2],
"dec_out_channels": 1,
"vocoder_input_dim": 2,
"normalization_type": "none",
}
checkpoint_path = tmp_path / DIFFUSERS_SOUND_TOKENIZER_CHECKPOINT_NAME
config_path = tmp_path / "config.json"
config_path.write_text(json.dumps(config), encoding="utf-8")
decoder = avae.OobleckDecoder(4, 2, 1, [2], [1])
save_file({f"decoder.{key}": value for key, value in decoder.state_dict().items()}, str(checkpoint_path))
tokenizer = avae.Cosmos3AVAEAudioTokenizer(
checkpoint_path=checkpoint_path,
config_path=config_path,
dtype=torch.float32,
device="cpu",
)
keys = set(tokenizer.state_dict())
assert {"decoder.conv1.weight_g", "decoder.block.0.conv_t1.weight_g", "decoder.conv2.weight_g"} <= keys
assert not any(key.startswith(("decoder.layers.", "model.decoder.")) for key in keys)
assert tokenizer.decode(torch.zeros(1, 2, 3)).shape == (1, 1, 6)
with pytest.raises(NotImplementedError, match="decoder-only"):
tokenizer.encode(torch.zeros(1, 1, 6))
@@ -4,7 +4,6 @@
from __future__ import annotations from __future__ import annotations
from types import SimpleNamespace from types import SimpleNamespace
from typing import Any
import pytest import pytest
import torch import torch
@@ -30,10 +29,8 @@ def _tiny_cosmos3_config(**overrides):
return config return config
def test_mrope_position_ids_cover_text_video_sound_and_action() -> None: def test_mrope_position_ids_cover_text_and_video() -> None:
from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import ( from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import (
compute_mrope_position_ids_action,
compute_mrope_position_ids_sound,
compute_mrope_position_ids_text, compute_mrope_position_ids_text,
compute_mrope_position_ids_vision, compute_mrope_position_ids_vision,
) )
@@ -59,14 +56,6 @@ def test_mrope_position_ids_cover_text_video_sound_and_action() -> None:
torch.testing.assert_close(modulated_ids[0], torch.tensor([10.0, 12.0])) torch.testing.assert_close(modulated_ids[0], torch.tensor([10.0, 12.0]))
assert modulated_offset == 13 assert modulated_offset == 13
sound_ids, sound_offset = compute_mrope_position_ids_sound(3, temporal_offset=10, sound_latent_fps=25.0)
torch.testing.assert_close(sound_ids[0], torch.tensor([10.0, 10.96, 11.92]))
assert sound_offset == 12
action_ids, action_offset = compute_mrope_position_ids_action(3, temporal_offset=10, action_fps=None)
assert action_ids.tolist() == [[11, 12, 13], [0, 0, 0], [0, 0, 0]]
assert action_offset == 14
@pytest.mark.parametrize( @pytest.mark.parametrize(
("key", "value"), ("key", "value"),
@@ -131,142 +120,7 @@ def test_forward_returns_video_prediction(monkeypatch: pytest.MonkeyPatch) -> No
assert tuple(output.shape) == (1, 2, 1, 2, 2) assert tuple(output.shape) == (1, 2, 1, 2, 2)
def test_sound_and_action_modules_follow_config() -> None: def test_compute_rope_freqs_places_text_and_video_positions() -> None:
from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer
tiny = _tiny_cosmos3_config()
no_modal = Cosmos3VFMTransformer(SimpleNamespace(tf_model_config=tiny, dtype=torch.float32))
with_sound = Cosmos3VFMTransformer(
SimpleNamespace(tf_model_config=tiny, dtype=torch.float32),
sound_gen=True,
sound_dim=5,
sound_latent_fps=40.0,
)
with_action = Cosmos3VFMTransformer(
SimpleNamespace(
tf_model_config={**tiny, "action_gen": True, "max_action_dim": 6, "num_embodiment_domains": 9},
dtype=torch.float32,
)
)
assert no_modal.sound_gen is False
assert no_modal.action_gen is False
assert not hasattr(no_modal, "audio_proj_in")
assert not hasattr(no_modal, "action_proj_in")
assert with_sound.sound_dim == 5
assert with_sound.sound_latent_fps == 40.0
assert with_sound.audio_proj_in.in_features == 5
assert with_action.action_dim == 6
assert with_action.action_proj_in.num_domains == 9
@pytest.mark.parametrize(
"kwargs",
[
{"sound_gen": True},
{"sound_gen": True, "sound_dim": 5},
{"sound_gen": True, "sound_latent_fps": 40.0},
],
)
def test_transformer_requires_sound_dim_and_fps_when_sound_gen_true(kwargs: dict[str, Any]) -> None:
from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer
with pytest.raises(ValueError, match=r"requires an explicit sound_dim and sound_latent_fps"):
Cosmos3VFMTransformer(
SimpleNamespace(tf_model_config=_tiny_cosmos3_config(), dtype=torch.float32),
**kwargs,
)
def test_sound_and_action_pack_unpack_validate_shapes() -> None:
from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer
model = object.__new__(Cosmos3VFMTransformer)
nn.Module.__init__(model)
model.sound_dim = 3
model.action_dim = 3
sound = torch.arange(2 * 3 * 4, dtype=torch.float32).reshape(2, 3, 4)
action = torch.arange(2 * 4 * 3, dtype=torch.float32).reshape(2, 4, 3)
torch.testing.assert_close(model.unpack_sound(model.pack_sound(sound)), sound)
torch.testing.assert_close(model.unpack_action(model.pack_action(action)), action)
with pytest.raises(ValueError, match="channel mismatch"):
model.pack_sound(torch.zeros(1, 4, 2))
with pytest.raises(ValueError, match="dimension mismatch"):
model.pack_action(torch.zeros(1, 2, 4))
@pytest.mark.parametrize(
("config", "transformer_kwargs", "extra_kwargs", "expected_shapes"),
[
(
_tiny_cosmos3_config(),
{"sound_gen": True, "sound_dim": 3, "sound_latent_fps": 24.0},
{"sound_latents": torch.zeros(1, 3, 4)},
[(1, 2, 1, 2, 2), (1, 3, 4)],
),
(
_tiny_cosmos3_config(action_gen=True, max_action_dim=3, num_embodiment_domains=4),
{},
{"action_latents": torch.zeros(1, 5, 3), "action_domain_ids": torch.tensor([2])},
[(1, 2, 1, 2, 2), (1, 5, 3)],
),
],
)
def test_forward_returns_video_plus_optional_modality_predictions(
monkeypatch: pytest.MonkeyPatch,
config,
transformer_kwargs,
extra_kwargs,
expected_shapes,
) -> None:
from vllm_omni.diffusion.models.cosmos3 import transformer_cosmos3
monkeypatch.setattr(transformer_cosmos3, "_get_ulysses_state", lambda: (1, 0, None))
output = transformer_cosmos3.Cosmos3VFMTransformer(
SimpleNamespace(tf_model_config=config, dtype=torch.float32),
**transformer_kwargs,
)(
hidden_states=torch.zeros(1, 2, 1, 2, 2),
timestep=torch.tensor([1.0]),
text_ids=torch.tensor([[1, 2]], dtype=torch.long),
text_mask=torch.ones(1, 2, dtype=torch.long),
video_shape=(1, 2, 2),
fps=24.0,
action_noisy_mask=torch.ones(1, 5, 1),
**extra_kwargs,
)
assert isinstance(output, tuple)
assert [tuple(tensor.shape) for tensor in output] == expected_shapes
def test_forward_with_sound_ulysses_error_mentions_combined_sequence(monkeypatch: pytest.MonkeyPatch) -> None:
import vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 as cosmos3_module
model = cosmos3_module.Cosmos3VFMTransformer(
SimpleNamespace(tf_model_config=_tiny_cosmos3_config(), dtype=torch.float32),
sound_gen=True,
sound_dim=3,
sound_latent_fps=40.0,
)
monkeypatch.setattr(cosmos3_module, "_get_ulysses_state", lambda: (2, 0, None))
with pytest.raises(ValueError, match=r"GEN sequence length \(3 = video tokens 2 \+ sound tokens 1\)"):
model(
hidden_states=torch.zeros(1, 2, 1, 1, 2),
timestep=torch.tensor([1.0]),
text_ids=torch.tensor([[1, 2]], dtype=torch.long),
text_mask=torch.ones(1, 2, dtype=torch.long),
video_shape=(1, 1, 2),
fps=24.0,
sound_latents=torch.zeros(1, 3, 1),
)
def test_compute_rope_freqs_places_text_video_action_and_sound_positions() -> None:
from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer from vllm_omni.diffusion.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer
class FakeRotary: class FakeRotary:
@@ -286,8 +140,6 @@ def test_compute_rope_freqs_places_text_video_action_and_sound_positions() -> No
model.temporal_modality_margin = 100 model.temporal_modality_margin = 100
model.base_fps = 24.0 model.base_fps = 24.0
model.temporal_compression_factor = 4 model.temporal_compression_factor = 4
model.temporal_compression_factor_sound = 1
model.sound_latent_fps = 25.0
model.enable_fps_modulation = False model.enable_fps_modulation = False
freqs_und, freqs_gen = model._compute_rope_freqs( freqs_und, freqs_gen = model._compute_rope_freqs(
@@ -304,21 +156,3 @@ def test_compute_rope_freqs_places_text_video_action_and_sound_positions() -> No
assert vision_pos[0, 0].tolist() == [102, 103] assert vision_pos[0, 0].tolist() == [102, 103]
assert freqs_und[0].shape == (2, 3, 1, 4) assert freqs_und[0].shape == (2, 3, 1, 4)
assert freqs_gen[0].shape == (2, 2, 1, 4) assert freqs_gen[0].shape == (2, 2, 1, 4)
rotary.position_ids.clear()
model._compute_rope_freqs(
text_mask=torch.tensor([[1, 1]], dtype=torch.long),
t=2,
hp=1,
wp=1,
fps=24.0,
device=torch.device("cpu"),
dtype=torch.float32,
t_action=2,
action_start_frame_offset=1,
t_sound=1,
)
_, gen_pos = rotary.position_ids
assert gen_pos.shape == (3, 1, 5)
assert gen_pos[0, 0].tolist() == [102, 103, 103, 104, 102]
@@ -1,2 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
@@ -1,160 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for HunyuanVideo-1.5 quant_config propagation through transformer creation.
Tests cover:
- HunyuanVideo15Pipeline passes quant_config to HunyuanVideo15Transformer3DModel
- HunyuanVideo15I2VPipeline passes quant_config to HunyuanVideo15Transformer3DModel
"""
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
import pytest
import torch
import vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5 as t2v_module
import vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v as i2v_module
pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
class TestHunyuanVideoQuantConfigPropagation:
"""Verify quant_config is propagated to the transformer model in HunyuanVideo-1.5 pipelines."""
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.get_local_device",
return_value="cpu",
)
@patch("vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.prefetch_subfolders")
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.from_pretrained_with_prefetch",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.Qwen2Tokenizer.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.ByT5Tokenizer.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.AutoConfig.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.T5EncoderModel",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.FlowMatchEulerDiscreteScheduler.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5.get_transformer_config_kwargs",
return_value={},
)
def test_t2v_quant_config_passed_through(
self,
mock_get_kwargs,
mock_scheduler,
mock_t5_encoder,
mock_auto_config,
mock_byt5_tok,
mock_qwen2_tok,
mock_from_pretrained,
mock_prefetch,
mock_get_device,
):
captured_kwargs = {}
class FakeTransformer:
def __init__(self, **kwargs):
captured_kwargs.update(kwargs)
with patch.object(t2v_module, "HunyuanVideo15Transformer3DModel", FakeTransformer):
fake_qc = MagicMock()
od_config = SimpleNamespace(
model="fake-t2v-model",
dtype=torch.bfloat16,
flow_shift=None,
tf_model_config=SimpleNamespace(),
quantization_config=fake_qc,
enable_diffusion_pipeline_profiler=False,
)
t2v_module.HunyuanVideo15Pipeline(od_config=od_config)
assert captured_kwargs.get("quant_config") is fake_qc
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.get_local_device",
return_value="cpu",
)
@patch("vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.prefetch_subfolders")
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.from_pretrained_with_prefetch",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.Qwen2Tokenizer.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.ByT5Tokenizer.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.AutoConfig.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.T5EncoderModel",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.SiglipImageProcessor.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.FlowMatchEulerDiscreteScheduler.from_pretrained",
return_value=MagicMock(),
)
@patch(
"vllm_omni.diffusion.models.hunyuan_video.pipeline_hunyuan_video_1_5_i2v.get_transformer_config_kwargs",
return_value={},
)
def test_i2v_quant_config_passed_through(
self,
mock_get_kwargs,
mock_scheduler,
mock_siglip_processor,
mock_t5_encoder,
mock_auto_config,
mock_byt5_tok,
mock_qwen2_tok,
mock_from_pretrained,
mock_prefetch,
mock_get_device,
):
captured_kwargs = {}
class FakeTransformer:
def __init__(self, **kwargs):
captured_kwargs.update(kwargs)
with patch.object(i2v_module, "HunyuanVideo15Transformer3DModel", FakeTransformer):
fake_qc = MagicMock()
od_config = SimpleNamespace(
model="fake-i2v-model",
dtype=torch.bfloat16,
flow_shift=None,
tf_model_config=SimpleNamespace(),
quantization_config=fake_qc,
enable_diffusion_pipeline_profiler=False,
)
i2v_module.HunyuanVideo15I2VPipeline(od_config=od_config)
assert captured_kwargs.get("quant_config") is fake_qc
@@ -72,31 +72,6 @@ class TestPipelineIndependence:
assert issubclass(LTX23Pipeline, CFGParallelMixin) assert issubclass(LTX23Pipeline, CFGParallelMixin)
def test_ltx23_pipeline_declares_offload_components(self):
"""LTX23Pipeline must expose LTX-2.3-specific modules to offload discovery."""
from vllm_omni.diffusion.models.ltx2.pipeline_ltx2_3 import LTX23Pipeline
from vllm_omni.diffusion.offloader.module_collector import ModuleDiscovery
pipe = object.__new__(LTX23Pipeline)
torch.nn.Module.__init__(pipe)
pipe.transformer = torch.nn.Linear(1, 1)
pipe.text_encoder = torch.nn.Linear(1, 1)
pipe.connectors = torch.nn.Linear(1, 1)
pipe.vae = torch.nn.Linear(1, 1)
pipe.audio_vae = torch.nn.Linear(1, 1)
pipe.vocoder = torch.nn.Linear(1, 1)
modules = ModuleDiscovery.discover(pipe)
assert LTX23Pipeline._dit_modules == ["transformer"]
assert LTX23Pipeline._encoder_modules == ["text_encoder", "connectors"]
assert LTX23Pipeline._vae_modules == ["vae", "audio_vae"]
assert LTX23Pipeline._resident_modules == ["vocoder"]
assert modules.dit_names == ["transformer"]
assert modules.encoder_names == ["text_encoder", "connectors"]
assert modules.resident_names == ["vocoder"]
assert len(modules.vaes) == 2
class TestCFGParallelHelpers: class TestCFGParallelHelpers:
"""Test LTX-2.3 CFG helper math without loading model weights.""" """Test LTX-2.3 CFG helper math without loading model weights."""
@@ -1,22 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from torch import nn
from vllm_omni.diffusion.models.ltx2.ltx2_transformer import _make_rms_norm
def test_ltx_rms_norm_no_affine_identity_weight_is_non_persistent_buffer():
norm = _make_rms_norm(8, eps=1e-6, elementwise_affine=False)
assert "weight" not in dict(norm.named_parameters())
assert "weight" in dict(norm.named_buffers())
assert "weight" not in norm.state_dict()
def test_ltx_rms_norm_affine_weight_remains_parameter():
norm = _make_rms_norm(8, eps=1e-6, elementwise_affine=True)
assert isinstance(dict(norm.named_parameters())["weight"], nn.Parameter)
assert "weight" not in dict(norm.named_buffers())
assert "weight" in norm.state_dict()
+6 -47
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0 # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import asyncio
import queue import queue
import threading import threading
from types import SimpleNamespace from types import SimpleNamespace
@@ -446,7 +447,11 @@ class TestDiffusionEngine:
@pytest.mark.asyncio @pytest.mark.asyncio
async def test_step_raises_aborted_error(self, mocker: MockerFixture) -> None: async def test_step_raises_aborted_error(self, mocker: MockerFixture) -> None:
engine = DiffusionEngine.__new__(DiffusionEngine) engine = DiffusionEngine.__new__(DiffusionEngine)
engine._check_and_start_background_loop = mocker.AsyncMock() engine._closed = False
engine._loop_started = True
engine._init_lock = asyncio.Lock()
engine.main_loop = asyncio.get_running_loop()
engine.stop_event = threading.Event()
engine.pre_process_func = None engine.pre_process_func = None
engine.async_add_req_and_wait_for_response = mocker.AsyncMock( engine.async_add_req_and_wait_for_response = mocker.AsyncMock(
return_value=DiffusionOutput(aborted=True, abort_message="Request req-abort aborted.") return_value=DiffusionOutput(aborted=True, abort_message="Request req-abort aborted.")
@@ -521,52 +526,6 @@ class TestDiffusionEngine:
with pytest.raises(RuntimeError, match="Dummy run failed: boom"): with pytest.raises(RuntimeError, match="Dummy run failed: boom"):
engine._dummy_run() engine._dummy_run()
@pytest.mark.asyncio
async def test_step_multi_request_reuses_multimodal_slice_logic(self, mocker: MockerFixture) -> None:
engine = DiffusionEngine.__new__(DiffusionEngine)
engine.od_config = SimpleNamespace(
model_class_name="mock_model",
enable_cpu_offload=False,
)
engine.pre_process_func = None
engine.post_process_func = None
engine._check_and_start_background_loop = mocker.AsyncMock()
engine.async_add_req_and_wait_for_response = mocker.AsyncMock(
return_value=DiffusionOutput(
output={
"video": ["frame-0", "frame-1"],
"audio": ["audio-0", "audio-1"],
"actions": torch.tensor([[1.0, 2.0], [3.0, 4.0]]),
}
)
)
request = OmniDiffusionRequest(
prompts=["prompt-0", "prompt-1"],
sampling_params=OmniDiffusionSamplingParams(
num_inference_steps=1,
num_outputs_per_prompt=1,
),
request_id="req-batch",
)
mocker.patch("vllm_omni.diffusion.diffusion_engine.supports_audio_output", return_value=False)
outputs = await engine.step(request)
assert len(outputs) == 2
assert outputs[0].images == ["frame-0"]
assert outputs[1].images == ["frame-1"]
assert outputs[0].multimodal_output["audio"] == "audio-0"
assert outputs[1].multimodal_output["audio"] == "audio-1"
torch.testing.assert_close(
outputs[0].multimodal_output["actions"],
torch.tensor([1.0, 2.0]),
)
torch.testing.assert_close(
outputs[1].multimodal_output["actions"],
torch.tensor([3.0, 4.0]),
)
class TestStepScheduler: class TestStepScheduler:
def setup_method(self) -> None: def setup_method(self) -> None:
+1 -1
View File
@@ -41,7 +41,7 @@ def mock_od_config(mocker: MockerFixture):
config.cache_backend = None config.cache_backend = None
config.cache_config = None config.cache_config = None
config.model = "test-model" config.model = "test-model"
config.diffusion_load_format = "default" config.diffusion_load_format = None
config.dtype = "float32" config.dtype = "float32"
config.max_cpu_loras = 0 config.max_cpu_loras = 0
config.lora_path = None config.lora_path = None
@@ -32,8 +32,8 @@ BAGEL_MOONCAKE_CI_DEPLOY = get_deploy_config_path("ci/bagel_mooncake.yaml")
# Reference pixel data extracted from the known-good output image # Reference pixel data extracted from the known-good output image
# Each entry contains (x, y) position and expected (R, G, B) values # Each entry contains (x, y) position and expected (R, G, B) values
# "Generated with seed=52, num_inference_steps=14, # "Generated with seed=52, num_inference_steps=15,
# prompt='A cute cat'" # prompt='A futuristic city skyline at twilight, cyberpunk style'"
REFERENCE_PIXELS = [ REFERENCE_PIXELS = [
{"position": (100, 100), "rgb": (115, 113, 94)}, {"position": (100, 100), "rgb": (115, 113, 94)},
{"position": (400, 50), "rgb": (159, 160, 144)}, {"position": (400, 50), "rgb": (159, 160, 144)},
@@ -63,7 +63,7 @@ def _find_free_port() -> int:
return port return port
def _configure_sampling_params(omni: Omni, num_inference_steps: int = 14) -> list: def _configure_sampling_params(omni: Omni, num_inference_steps: int = 15) -> list:
"""Configure sampling parameters for Bagel text2img generation. """Configure sampling parameters for Bagel text2img generation.
Args: Args:
@@ -269,6 +269,7 @@ def _load_mooncake_config(host: str, rpc_port: int, http_port: int) -> str:
@pytest.mark.core_model @pytest.mark.core_model
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.diffusion @pytest.mark.diffusion
@pytest.mark.skip(reason="Skip failed CI issue 3977: https://github.com/vllm-project/vllm-omni/issues/3977")
@hardware_test(res={"cuda": "H100"}, num_cards=1) @hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_bagel_text2img_mooncake_connector(run_level): def test_bagel_text2img_mooncake_connector(run_level):
"""Test Bagel text2img with Mooncake connector for inter-stage communication.""" """Test Bagel text2img with Mooncake connector for inter-stage communication."""
@@ -24,13 +24,14 @@ from tests.helpers.mark import hardware_test
from tests.helpers.runtime import OmniRunner from tests.helpers.runtime import OmniRunner
from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
from vllm_omni.entrypoints.omni import Omni from vllm_omni.entrypoints.omni import Omni
from vllm_omni.platforms import current_omni_platform
pytestmark = [pytest.mark.usefixtures("clean_gpu_memory_between_tests")] pytestmark = [pytest.mark.usefixtures("clean_gpu_memory_between_tests")]
BAGEL_CI_DEPLOY = get_deploy_config_path("ci/bagel.yaml") BAGEL_CI_DEPLOY = get_deploy_config_path("ci/bagel.yaml")
# Reference pixel data extracted from the known-good output image # Reference pixel data extracted from the known-good output image
# Generated with seed=52, num_inference_steps=14, # Generated with seed=52, num_inference_steps=15,
# prompt='Change the grass color to red', # prompt='Change the grass color to red',
# input image: 2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg # input image: 2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg
REFERENCE_PIXELS = [ REFERENCE_PIXELS = [
@@ -46,10 +47,23 @@ REFERENCE_PIXELS = [
{"position": (256, 256), "rgb": (181, 201, 221)}, {"position": (256, 256), "rgb": (181, 201, 221)},
] ]
if current_omni_platform.is_rocm():
REFERENCE_PIXELS = [
{"position": (100, 100), "rgb": (156, 172, 217)},
{"position": (400, 50), "rgb": (105, 144, 217)},
{"position": (700, 100), "rgb": (118, 159, 232)},
{"position": (150, 400), "rgb": (180, 22, 52)},
{"position": (512, 336), "rgb": (221, 211, 194)},
{"position": (700, 400), "rgb": (192, 10, 46)},
{"position": (100, 600), "rgb": (102, 12, 22)},
{"position": (400, 600), "rgb": (161, 28, 47)},
{"position": (700, 600), "rgb": (100, 87, 94)},
{"position": (256, 256), "rgb": (181, 201, 221)},
]
# text2img reference pixels (aligned with offline `bagel/end2end.py` text2img, 15 steps) # text2img reference pixels (aligned with offline `bagel/end2end.py` text2img, 15 steps)
# "Generated with seed=52, num_inference_steps=14, # "Generated with seed=52, num_inference_steps=15,
# prompt='A cute cat'" # prompt='A futuristic city skyline at twilight, cyberpunk style'"
TEXT2IMG_REFERENCE_PIXELS = [ TEXT2IMG_REFERENCE_PIXELS = [
{"position": (100, 100), "rgb": (115, 113, 94)}, {"position": (100, 100), "rgb": (115, 113, 94)},
{"position": (400, 50), "rgb": (159, 160, 144)}, {"position": (400, 50), "rgb": (159, 160, 144)},
@@ -63,6 +77,19 @@ TEXT2IMG_REFERENCE_PIXELS = [
{"position": (256, 256), "rgb": (167, 156, 150)}, {"position": (256, 256), "rgb": (167, 156, 150)},
] ]
if current_omni_platform.is_rocm():
TEXT2IMG_REFERENCE_PIXELS = [
{"position": (100, 100), "rgb": (115, 113, 94)},
{"position": (400, 50), "rgb": (159, 160, 144)},
{"position": (700, 100), "rgb": (164, 151, 123)},
{"position": (150, 400), "rgb": (120, 121, 107)},
{"position": (512, 512), "rgb": (165, 133, 127)},
{"position": (700, 400), "rgb": (217, 130, 66)},
{"position": (100, 700), "rgb": (191, 168, 152)},
{"position": (400, 700), "rgb": (130, 96, 77)},
{"position": (700, 700), "rgb": (247, 203, 140)},
{"position": (256, 256), "rgb": (167, 156, 150)},
]
PIXEL_TOLERANCE = 10 PIXEL_TOLERANCE = 10
TEXT2IMG_PIXEL_TOLERANCE = 5 TEXT2IMG_PIXEL_TOLERANCE = 5
@@ -88,7 +115,7 @@ def _find_free_port() -> int:
return port return port
def _configure_sampling_params(omni: Omni, num_inference_steps: int = 14) -> list: def _configure_sampling_params(omni: Omni, num_inference_steps: int = 15) -> list:
"""Configure sampling parameters for Bagel img2img generation. """Configure sampling parameters for Bagel img2img generation.
Args: Args:
@@ -231,6 +258,7 @@ def _resolve_deploy_config(config_path: str, run_level: str) -> str:
@pytest.mark.core_model @pytest.mark.core_model
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.diffusion @pytest.mark.diffusion
@pytest.mark.skip(reason="Skip failed CI issue 3977: https://github.com/vllm-project/vllm-omni/issues/3977")
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}) @hardware_test(res={"cuda": "H100", "rocm": "MI325"})
def test_bagel_img2img_shared_memory_connector(run_level): def test_bagel_img2img_shared_memory_connector(run_level):
"""Test Bagel img2img with shared memory connector.""" """Test Bagel img2img with shared memory connector."""
@@ -248,6 +276,7 @@ def test_bagel_img2img_shared_memory_connector(run_level):
@pytest.mark.core_model @pytest.mark.core_model
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.diffusion @pytest.mark.diffusion
@pytest.mark.skip(reason="Skip failed CI issue 3977: https://github.com/vllm-project/vllm-omni/issues/3977")
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}) @hardware_test(res={"cuda": "H100", "rocm": "MI325"})
def test_bagel_text2img_shared_memory_connector(run_level): def test_bagel_text2img_shared_memory_connector(run_level):
"""Test Bagel text2img with shared memory connector.""" """Test Bagel text2img with shared memory connector."""
-307
View File
@@ -1,307 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import uuid
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import numpy as np
try:
import cv2
except ImportError: # pragma: no cover - optional e2e dependency
cv2 = None
try:
import websockets.sync.client as websockets_client
except ImportError: # pragma: no cover - optional e2e dependency
websockets_client = None
try:
from openpi_client import msgpack_numpy
except ImportError: # pragma: no cover - optional e2e dependency
msgpack_numpy = None
PING_INTERVAL_SECS = 300
PING_TIMEOUT_SECS = 3600
DEFAULT_HOST = "127.0.0.1"
DEFAULT_PORT = 8000
DEFAULT_PATH = "/v1/realtime/robot/openpi"
DEFAULT_PROMPT = "Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan"
ACTION_HORIZON = 24
DEFAULT_ACTION_DIM = 8
RELATIVE_OFFSETS = [-23, -16, -8, 0]
CAMERA_FILES = {
"observation/exterior_image_0_left": "exterior_image_1_left.mp4",
"observation/exterior_image_1_left": "exterior_image_2_left.mp4",
"observation/wrist_image_left": "wrist_image_left.mp4",
}
def _decode_action_response(response: bytes | str) -> np.ndarray:
if isinstance(response, str):
raise RuntimeError(f"Inference failed: {response}")
decoded = msgpack_numpy.unpackb(response)
if isinstance(decoded, dict) and decoded.get("type") == "error":
message = decoded.get("message", decoded)
raise RuntimeError(f"Inference failed: {message}")
return np.asarray(decoded, dtype=np.float32)
def require_dependencies() -> None:
missing = []
if cv2 is None:
missing.append("opencv-python")
if websockets_client is None:
missing.append("websockets")
if msgpack_numpy is None:
missing.append("openpi-client")
if missing:
raise ModuleNotFoundError(f"DreamZero OpenPI test dependencies are missing: {', '.join(missing)}")
@dataclass(frozen=True)
class DreamZeroServerMetadata:
image_resolution: tuple[int, int]
n_external_cameras: int
needs_wrist_camera: bool
needs_stereo_camera: bool
needs_session_id: bool
action_space: str
@classmethod
def from_dict(cls, payload: dict[str, Any]) -> DreamZeroServerMetadata:
required_keys = (
"image_resolution",
"n_external_cameras",
"needs_wrist_camera",
"needs_stereo_camera",
"needs_session_id",
"action_space",
)
missing_keys = [key for key in required_keys if key not in payload]
if missing_keys:
raise ValueError(f"Missing DreamZero metadata keys: {missing_keys}")
image_resolution = payload["image_resolution"]
if not isinstance(image_resolution, (list, tuple)) or len(image_resolution) != 2:
raise ValueError(f"Invalid image_resolution: {image_resolution!r}")
return cls(
image_resolution=(int(image_resolution[0]), int(image_resolution[1])),
n_external_cameras=int(payload["n_external_cameras"]),
needs_wrist_camera=bool(payload["needs_wrist_camera"]),
needs_stereo_camera=bool(payload["needs_stereo_camera"]),
needs_session_id=bool(payload["needs_session_id"]),
action_space=str(payload["action_space"]),
)
class OpenPIWebsocketClient:
def __init__(
self,
*,
host: str = DEFAULT_HOST,
port: int = DEFAULT_PORT,
path: str = DEFAULT_PATH,
) -> None:
require_dependencies()
self._uri = f"ws://{host}:{port}{path}"
self._packer = msgpack_numpy.Packer()
self._ws, self._server_metadata = self._connect()
def _connect(self):
conn = websockets_client.connect(
self._uri,
compression=None,
max_size=None,
ping_interval=PING_INTERVAL_SECS,
ping_timeout=PING_TIMEOUT_SECS,
)
metadata = msgpack_numpy.unpackb(conn.recv())
if not isinstance(metadata, dict):
raise TypeError(f"Expected dict metadata from server, got {type(metadata)!r}")
return conn, metadata
def get_server_metadata(self) -> dict[str, Any]:
return dict(self._server_metadata)
def infer(self, obs: dict[str, Any]) -> np.ndarray:
payload = dict(obs)
payload["endpoint"] = "infer"
self._ws.send(self._packer.pack(payload))
response = self._ws.recv()
return _decode_action_response(response)
def reset(self, reset_info: dict[str, Any] | None = None) -> str:
payload = dict(reset_info or {})
payload["endpoint"] = "reset"
self._ws.send(self._packer.pack(payload))
response = self._ws.recv()
if isinstance(response, str):
return response
decoded = msgpack_numpy.unpackb(response)
if not isinstance(decoded, dict) or decoded.get("status") != "reset successful":
raise RuntimeError(f"Unexpected reset response: {decoded!r}")
return str(decoded["status"])
def close(self) -> None:
self._ws.close()
def load_all_frames(video_path: Path) -> np.ndarray:
require_dependencies()
cap = cv2.VideoCapture(str(video_path))
frames = []
while True:
ok, frame = cap.read()
if not ok:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
cap.release()
if not frames:
raise RuntimeError(f"No frames loaded from {video_path}")
return np.stack(frames, axis=0)
def load_camera_frames(video_dir: Path) -> dict[str, np.ndarray]:
camera_frames: dict[str, np.ndarray] = {}
for camera_key, file_name in CAMERA_FILES.items():
video_path = video_dir / file_name
if not video_path.exists():
raise FileNotFoundError(f"Missing DreamZero test asset: {video_path}")
camera_frames[camera_key] = load_all_frames(video_path)
return camera_frames
def build_frame_schedule(total_frames: int, num_chunks: int) -> list[list[int]]:
chunks: list[list[int]] = []
current_frame = 23
for _ in range(num_chunks):
indices = [max(current_frame + offset, 0) for offset in RELATIVE_OFFSETS]
if indices[-1] >= total_frames:
break
chunks.append(indices)
current_frame += ACTION_HORIZON
return chunks
def make_obs_from_video(
camera_frames: dict[str, np.ndarray],
frame_indices: list[int],
*,
prompt: str,
session_id: str,
) -> dict[str, Any]:
obs: dict[str, Any] = {}
for camera_key, all_frames in camera_frames.items():
selected = all_frames[frame_indices]
obs[camera_key] = selected[0] if len(frame_indices) == 1 else selected
obs["observation/joint_position"] = np.zeros(7, dtype=np.float32)
obs["observation/cartesian_position"] = np.zeros(6, dtype=np.float32)
obs["observation/gripper_position"] = np.zeros(1, dtype=np.float32)
obs["prompt"] = prompt
obs["session_id"] = session_id
return obs
def build_demo_observations(
camera_frames: dict[str, np.ndarray],
*,
prompt: str,
session_id: str,
num_chunks: int = 2,
) -> list[dict[str, Any]]:
if num_chunks < 1:
raise ValueError("num_chunks must be at least 1")
total_frames = min(frames.shape[0] for frames in camera_frames.values())
observations = [
make_obs_from_video(
camera_frames,
[0],
prompt=prompt,
session_id=session_id,
)
]
for indices in build_frame_schedule(total_frames, num_chunks - 1):
observations.append(
make_obs_from_video(
camera_frames,
indices,
prompt=prompt,
session_id=session_id,
)
)
return observations
def validate_session_result(
result: dict[str, Any],
*,
expected_action_horizon: int = ACTION_HORIZON,
expected_action_dim: int = DEFAULT_ACTION_DIM,
) -> None:
metadata = DreamZeroServerMetadata.from_dict(result["metadata"])
if metadata.image_resolution != (180, 320):
raise AssertionError(f"Unexpected image_resolution: {metadata.image_resolution}")
if metadata.n_external_cameras != 2:
raise AssertionError(f"Unexpected n_external_cameras: {metadata.n_external_cameras}")
if not metadata.needs_wrist_camera:
raise AssertionError("DreamZero test expects wrist camera metadata")
if metadata.action_space != "joint_position":
raise AssertionError(f"Unexpected action_space: {metadata.action_space}")
actions = result["actions"]
if len(actions) != 3:
raise AssertionError(f"Expected 3 action tensors, got {len(actions)}")
for index, action in enumerate(actions):
if action.shape != (expected_action_horizon, expected_action_dim):
raise AssertionError(
f"Action {index} shape mismatch: expected "
f"{(expected_action_horizon, expected_action_dim)}, got {action.shape}"
)
if not np.isfinite(action).all():
raise AssertionError(f"Action {index} contains non-finite values")
if result["reset_status"] != "reset successful":
raise AssertionError(f"Unexpected reset status: {result['reset_status']!r}")
def run_policy_session(
*,
host: str = DEFAULT_HOST,
port: int = DEFAULT_PORT,
path: str = DEFAULT_PATH,
video_dir: Path,
prompt: str = DEFAULT_PROMPT,
session_id: str | None = None,
num_chunks: int = 2,
) -> dict[str, Any]:
session_id = session_id or str(uuid.uuid4())
camera_frames = load_camera_frames(video_dir)
observations = build_demo_observations(
camera_frames,
prompt=prompt,
session_id=session_id,
num_chunks=num_chunks,
)
client = OpenPIWebsocketClient(host=host, port=port, path=path)
try:
metadata = client.get_server_metadata()
actions = [client.infer(obs) for obs in observations]
reset_status = client.reset({})
actions.append(client.infer(observations[0]))
return {
"metadata": metadata,
"actions": actions,
"reset_status": reset_status,
"session_id": session_id,
}
finally:
client.close()
-24
View File
@@ -1,24 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm_omni.diffusion.models.dreamzero.utils import (
DEFAULT_CFG_SCALE,
DEFAULT_EMBODIMENT_NAME_TO_ID,
DEFAULT_NEGATIVE_PROMPT,
DEFAULT_NUM_INFERENCE_STEPS,
DEFAULT_SEED,
DEFAULT_SIGMA_SHIFT,
)
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
def test_dreamzero_default_constants_match_source_baseline():
assert DEFAULT_NUM_INFERENCE_STEPS == 16
assert DEFAULT_CFG_SCALE == 5.0
assert DEFAULT_SIGMA_SHIFT == 5.0
assert DEFAULT_SEED == 1140
assert "worst quality" in DEFAULT_NEGATIVE_PROMPT
assert DEFAULT_EMBODIMENT_NAME_TO_ID["oxe_droid"] == 17
@@ -1,46 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import numpy as np
import pytest
from tests.dreamzero import openpi_client_helper
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
class FakeMsgpackNumpy:
payloads = {}
@staticmethod
def packb(obj):
key = str(len(FakeMsgpackNumpy.payloads)).encode()
FakeMsgpackNumpy.payloads[key] = obj
return key
@staticmethod
def unpackb(data):
return FakeMsgpackNumpy.payloads[data]
def test_decode_action_response_surfaces_structured_error(monkeypatch):
monkeypatch.setattr(openpi_client_helper, "msgpack_numpy", FakeMsgpackNumpy)
payload = FakeMsgpackNumpy.packb(
{
"type": "error",
"message": "Internal inference error",
}
)
with pytest.raises(RuntimeError, match="Internal inference error"):
openpi_client_helper._decode_action_response(payload)
def test_decode_action_response_converts_action_payload_to_float32(monkeypatch):
monkeypatch.setattr(openpi_client_helper, "msgpack_numpy", FakeMsgpackNumpy)
payload = FakeMsgpackNumpy.packb(np.asarray([[1.0, 2.0]], dtype=np.float64))
actions = openpi_client_helper._decode_action_response(payload)
assert actions.dtype == np.float32
np.testing.assert_array_equal(actions, np.asarray([[1.0, 2.0]], dtype=np.float32))
-59
View File
@@ -1,59 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections import OrderedDict
import pytest
import torch
from vllm_omni.diffusion.models.dreamzero.pipeline_dreamzero import DreamZeroPipeline
from vllm_omni.diffusion.models.dreamzero.state_dreamzero import DreamZeroState
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
def _empty_pipeline() -> DreamZeroPipeline:
pipeline = DreamZeroPipeline.__new__(DreamZeroPipeline)
pipeline._states = OrderedDict()
pipeline._max_session_states = 2
return pipeline
def test_dreamzero_pipeline_state_is_session_keyed() -> None:
pipeline = _empty_pipeline()
session_a = pipeline._get_or_create_state("session-a")
session_b = pipeline._get_or_create_state("session-b")
session_a.call_count = 7
session_b.call_count = 3
assert pipeline._get_or_create_state("session-a") is session_a
assert pipeline._get_or_create_state("session-b") is session_b
assert session_a.call_count == 7
assert session_b.call_count == 3
def test_dreamzero_pipeline_state_lru_caps_retained_sessions() -> None:
pipeline = _empty_pipeline()
session_a = pipeline._get_or_create_state("session-a")
pipeline._get_or_create_state("session-b")
assert pipeline._get_or_create_state("session-a") is session_a
pipeline._get_or_create_state("session-c")
assert list(pipeline._states) == ["session-a", "session-c"]
assert "session-b" not in pipeline._states
def test_dreamzero_state_cache_access_requires_initialization() -> None:
state = DreamZeroState()
with pytest.raises(RuntimeError, match="KV caches not initialized"):
state.get_kv_caches()
with pytest.raises(RuntimeError, match="Cross-attn caches not initialized"):
state.get_crossattn_caches()
with pytest.raises(RuntimeError, match="create_kv_caches first"):
state.update_kv_cache(0, torch.empty(0))
@@ -1,381 +0,0 @@
#!/usr/bin/env python3
"""
Copied from the DreamZero repository's `test_client_AR.py`:
https://github.com/dreamzero0/dreamzero/blob/main/test_client_AR.py
Kept here for end-to-end compatibility / parity testing against the vLLM
OpenPI server.
Test client for AR_droid policy server using roboarena interface.
Sends real video frames from debug_image/ directory instead of zero dummy images.
Frame schedule (matching debug_inference.py):
- Step 0 (initial): send frame [0] (1 frame, H W 3)
- Step 1: send frames [0, 7, 15, 23] (4 frames, 4 H W 3)
- Step 2: send frames [24, 31, 39, 47] (4 frames)
- Step 3: send frames [48, 55, 63, 71] (4 frames)
- ...
Expected server configuration:
- image_resolution: (180, 320)
- n_external_cameras: 2
- needs_wrist_camera: True
- action_space: "joint_position"
Usage:
# Start server with roboarena interface:
torchrun --nproc_per_node=8 socket_test_optimized_AR.py --port 8000
# Run against the original DreamZero websocket server (default path):
python test_client_AR.py --host <server_host> --port 8000
# Run against the vLLM OpenPI server:
python test_client_AR.py --host <server_host> --port 8000 --path /v1/realtime/robot/openpi
# Use zero images instead of real video (old behavior):
python test_client_AR.py --host <server_host> --port 8000 --use-zero-images
"""
import argparse
import logging
import os
import sys
import time
import uuid
from pathlib import Path
import cv2
import numpy as np
from openpi_client import msgpack_numpy
_DREAMZERO_REPO_ENV = os.environ.get("DREAMZERO_REPO")
DREAMZERO_REPO = Path(_DREAMZERO_REPO_ENV).expanduser() if _DREAMZERO_REPO_ENV else None
def _import_upstream_policy_modules():
if DREAMZERO_REPO is None:
raise ImportError("Set DREAMZERO_REPO to an upstream DreamZero checkout before using this helper.")
if DREAMZERO_REPO.exists() and str(DREAMZERO_REPO) not in sys.path:
sys.path.insert(0, str(DREAMZERO_REPO))
import eval_utils.policy_server as policy_server
from eval_utils.policy_client import WebsocketClientPolicy
return policy_server, WebsocketClientPolicy
policy_server, WebsocketClientPolicy = _import_upstream_policy_modules()
VIDEO_DIR = os.environ.get(
"DREAMZERO_VIDEO_DIR",
str(DREAMZERO_REPO / "debug_image") if DREAMZERO_REPO is not None else "debug_image",
)
# roboarena key -> video filename
CAMERA_FILES = {
"observation/exterior_image_0_left": "exterior_image_1_left.mp4",
"observation/exterior_image_1_left": "exterior_image_2_left.mp4",
"observation/wrist_image_left": "wrist_image_left.mp4",
}
# Frame schedule constants (matching debug_inference.py)
RELATIVE_OFFSETS = [-23, -16, -8, 0]
ACTION_HORIZON = 24
DEFAULT_WEBSOCKET_PATH = ""
class OpenPIWebsocketClientPolicy(WebsocketClientPolicy):
"""DreamZero websocket client with a configurable path suffix.
The original DreamZero client connects to ``ws://host:port``.
vLLM serves the compatible robot policy endpoint at
``/v1/realtime/robot/openpi`` when ``path`` is set accordingly.
"""
def __init__(
self,
host: str = "0.0.0.0",
port: int = 8000,
path: str = DEFAULT_WEBSOCKET_PATH,
) -> None:
self._uri = f"ws://{host}:{port}{path}"
self._packer = msgpack_numpy.Packer()
self._ws, self._server_metadata = self._wait_for_server()
def load_all_frames(video_path: str) -> np.ndarray:
"""Load all frames from a video file. Returns (N, H, W, 3) uint8 array (RGB)."""
cap = cv2.VideoCapture(video_path)
frames = []
while True:
ret, frame = cap.read()
if not ret:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
cap.release()
if not frames:
raise RuntimeError(f"No frames loaded from {video_path}")
return np.stack(frames, axis=0)
def load_camera_frames() -> dict[str, np.ndarray]:
"""Load all video frames for each camera from the debug_image/ directory.
Returns:
Dict mapping roboarena camera keys to (N, H, W, 3) uint8 arrays.
"""
camera_frames: dict[str, np.ndarray] = {}
for cam_key, fname in CAMERA_FILES.items():
path = os.path.join(VIDEO_DIR, fname)
camera_frames[cam_key] = load_all_frames(path)
logging.info(f"Loaded {cam_key}: {camera_frames[cam_key].shape}")
return camera_frames
def build_frame_schedule(total_frames: int, num_chunks: int) -> list[list[int]]:
"""Build the frame index schedule for multi-frame chunks.
Returns a list of frame-index lists. Each inner list has 4 indices.
"""
chunks: list[list[int]] = []
current_frame = 23 # first anchor frame
for _ in range(num_chunks):
indices = [max(current_frame + off, 0) for off in RELATIVE_OFFSETS]
if indices[-1] >= total_frames:
logging.info(f"Frame {indices[-1]} >= {total_frames}, stopping at {len(chunks)} chunks")
break
chunks.append(indices)
current_frame += ACTION_HORIZON
return chunks
def _make_obs_from_video(
camera_frames: dict[str, np.ndarray],
frame_indices: list[int],
prompt: str,
session_id: str,
) -> dict:
"""Build an observation dict from real video frames.
For 1 frame: each image key is (H, W, 3).
For 4 frames: each image key is (4, H, W, 3).
"""
obs: dict = {}
for cam_key, all_frames in camera_frames.items():
selected = all_frames[frame_indices] # (T, H, W, 3)
if len(frame_indices) == 1:
selected = selected[0] # (H, W, 3)
obs[cam_key] = selected
obs["observation/joint_position"] = np.zeros(7, dtype=np.float32)
obs["observation/cartesian_position"] = np.zeros(6, dtype=np.float32)
obs["observation/gripper_position"] = np.zeros(1, dtype=np.float32)
obs["prompt"] = prompt
obs["session_id"] = session_id
return obs
def _make_zero_observation(
server_config: policy_server.PolicyServerConfig,
prompt: str = "pick up the object",
session_id: str | None = None,
) -> dict:
"""Create a dummy observation matching AR_droid expectations.
AR_droid expects:
- 2 external cameras (exterior_image_0_left, exterior_image_1_left)
- 1 wrist camera (wrist_image_left)
- Image resolution: 180x320 (H x W)
- joint_position: 7 DoF
- gripper_position: 1 DoF
"""
obs = {}
# Determine image resolution
if server_config.image_resolution is not None:
h, w = server_config.image_resolution
else:
# Default for AR_droid
h, w = 180, 320
# External cameras (0-indexed in roboarena)
for i in range(server_config.n_external_cameras):
obs[f"observation/exterior_image_{i}_left"] = np.zeros((h, w, 3), dtype=np.uint8)
if server_config.needs_stereo_camera:
obs[f"observation/exterior_image_{i}_right"] = np.zeros((h, w, 3), dtype=np.uint8)
# Wrist camera
if server_config.needs_wrist_camera:
obs["observation/wrist_image_left"] = np.zeros((h, w, 3), dtype=np.uint8)
if server_config.needs_stereo_camera:
obs["observation/wrist_image_right"] = np.zeros((h, w, 3), dtype=np.uint8)
# Session ID - should be passed in to ensure consistency within a session
if server_config.needs_session_id:
import uuid
# Generate unique session ID if not provided
obs["session_id"] = session_id if session_id else str(uuid.uuid4())
# State observations (AR_droid: 7 DoF arm + 1 gripper)
obs["observation/joint_position"] = np.zeros(7, dtype=np.float32)
obs["observation/cartesian_position"] = np.zeros(6, dtype=np.float32)
obs["observation/gripper_position"] = np.zeros(1, dtype=np.float32)
# Language prompt
obs["prompt"] = prompt
return obs
def test_ar_droid_policy_server(
host: str = "localhost",
port: int = 8000,
path: str = DEFAULT_WEBSOCKET_PATH,
num_chunks: int = 15,
prompt: str = "Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan",
use_zero_images: bool = False,
):
"""Test the AR_droid policy server with roboarena interface.
When use_zero_images is False (default), loads real video frames from
debug_image/ and follows the frame schedule from debug_inference.py.
"""
logging.info(f"Connecting to AR_droid server at ws://{host}:{port}{path} ...")
client = OpenPIWebsocketClientPolicy(host=host, port=port, path=path)
# Validate server metadata
metadata = client.get_server_metadata()
logging.info(f"Server metadata: {metadata}")
assert isinstance(metadata, dict), "Metadata should be a dict"
try:
server_config = policy_server.PolicyServerConfig(**metadata)
except Exception as e:
logging.error(f"Error parsing metadata: {e}")
raise e
# Validate expected AR_droid configuration
logging.info(f"Server config: {server_config}")
assert server_config.n_external_cameras == 2, f"Expected 2 external cameras, got {server_config.n_external_cameras}"
assert server_config.needs_wrist_camera, "Expected wrist camera to be enabled"
assert server_config.action_space == "joint_position", (
f"Expected joint_position action space, got {server_config.action_space}"
)
logging.info("Server configuration validated for AR_droid")
# Generate unique session ID for this test run
session_id = str(uuid.uuid4())
logging.info(f"Session ID: {session_id}")
# ── Zero-image fallback mode ──────────────────────────────────────
if use_zero_images:
logging.info("Using ZERO dummy images (legacy mode)")
for i in range(num_chunks):
obs = _make_zero_observation(server_config, prompt=prompt, session_id=session_id)
logging.info(f"Inference {i + 1}/{num_chunks}: prompt='{prompt}'")
t0 = time.time()
actions = client.infer(obs)
dt = time.time() - t0
_log_action(actions, dt)
logging.info("Sending reset...")
client.reset({})
logging.info("Done (zero-image mode).")
return
# ── Real video frame mode ─────────────────────────────────────────
logging.info("Loading real video frames from debug_image/ directory")
camera_frames = load_camera_frames()
total_frames = min(v.shape[0] for v in camera_frames.values())
logging.info(f"Total frames available: {total_frames}")
# Build frame schedule
chunks = build_frame_schedule(total_frames, num_chunks)
logging.info("Frame schedule:")
logging.info(" Initial: [0]")
for i, indices in enumerate(chunks):
logging.info(f" Chunk {i}: {indices}")
# Step 0: initial single frame
logging.info("=== Initial: frame [0] ===")
obs = _make_obs_from_video(camera_frames, [0], prompt, session_id)
t0 = time.time()
actions = client.infer(obs)
dt = time.time() - t0
_log_action(actions, dt)
# Subsequent chunks: send 4 frames at a time
for chunk_idx, frame_indices in enumerate(chunks):
logging.info(f"=== Chunk {chunk_idx}: frames {frame_indices} ===")
obs = _make_obs_from_video(camera_frames, frame_indices, prompt, session_id)
t0 = time.time()
actions = client.infer(obs)
dt = time.time() - t0
_log_action(actions, dt)
# Reset triggers video save on the server
logging.info("Sending reset to save video...")
client.reset({})
logging.info("Done.")
def _log_action(actions: np.ndarray, dt: float) -> None:
"""Pretty-print action shape, range, and timing."""
assert isinstance(actions, np.ndarray), f"Expected numpy array, got {type(actions)}"
assert actions.ndim == 2, f"Expected 2D array, got shape {actions.shape}"
assert actions.shape[-1] == 8, f"Expected 8 action dims (7 joints + 1 gripper), got {actions.shape[-1]}"
logging.info(f" Action shape: {actions.shape}, range: [{actions.min():.4f}, {actions.max():.4f}], time: {dt:.2f}s")
def main():
parser = argparse.ArgumentParser(description="Test AR_droid policy server with real video frames from debug_image/")
parser.add_argument("--host", default="localhost", help="Server hostname")
parser.add_argument("--port", type=int, default=8000, help="Server port")
parser.add_argument(
"--path",
default=DEFAULT_WEBSOCKET_PATH,
help="WebSocket path suffix (default: empty string for the original DreamZero server)",
)
parser.add_argument(
"--num-chunks",
type=int,
default=15,
help="Number of 4-frame chunks to send after the initial frame (default: 15)",
)
parser.add_argument(
"--prompt",
default="Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan",
help="Language prompt for the policy",
)
parser.add_argument(
"--use-zero-images",
action="store_true",
help="Use zero dummy images instead of real video frames (legacy mode)",
)
args = parser.parse_args()
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
)
test_ar_droid_policy_server(
host=args.host,
port=args.port,
path=args.path,
num_chunks=args.num_chunks,
prompt=args.prompt,
use_zero_images=args.use_zero_images,
)
if __name__ == "__main__":
main()
@@ -1,358 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Formal OpenPI end-to-end parity: upstream DreamZero server vs `vllm serve`.
This test uses DreamZero's own client-side observation builders from
`${DREAMZERO_REPO}/test_client_AR.py`, and client-side websocket protocol from
`${DREAMZERO_REPO}/eval_utils/policy_client.py`.
The only client-side adaptation for vLLM is the websocket path:
DreamZero's upstream server serves at `/`, while vLLM serves OpenPI at
`/v1/realtime/robot/openpi`.
Current scope for this test:
- default two-GPU run (`nproc_per_node=2` on upstream, `--cfg-parallel-size 2` on `vllm serve`)
- non-`torch.compile` (upstream launched through
`upstream_socket_server_no_compile.py`, vLLM with `--enforce-eager`)
- non-DiT-cache / non-skip-schedule (`NUM_DIT_STEPS=16`)
Serving contract locked by this test:
- upstream DreamZero still boots from the local checkpoint directory
- vLLM boots from the official DreamZero HF repo name (`GEAR-Dreams/DreamZero-DROID`)
rather than a prepared local bundle path
"""
from __future__ import annotations
import os
import shutil
import subprocess
import sys
import time
from pathlib import Path
import numpy as np
import pytest
import torch
from tests.helpers.runtime import get_open_port
msgpack_numpy = pytest.importorskip("openpi_client.msgpack_numpy")
_DREAMZERO_REPO_ENV = os.environ.get("DREAMZERO_REPO")
DREAMZERO_REPO = Path(_DREAMZERO_REPO_ENV).expanduser() if _DREAMZERO_REPO_ENV else None
if DREAMZERO_REPO is not None and str(DREAMZERO_REPO) not in sys.path:
sys.path.insert(0, str(DREAMZERO_REPO))
try:
import test_client_AR as dreamzero_client
from eval_utils.policy_client import WebsocketClientPolicy
except Exception: # pragma: no cover - guarded by pytest skip below
dreamzero_client = None
WebsocketClientPolicy = None
_BaseWebsocketClientPolicy = WebsocketClientPolicy if WebsocketClientPolicy is not None else object
CHECKPOINT_DIR = DREAMZERO_REPO / "checkpoints" / "dreamzero" if DREAMZERO_REPO is not None else None
VLLM_MODEL = os.environ.get("VLLM_DREAMZERO_MODEL", "GEAR-Dreams/DreamZero-DROID")
SERVICE_READY_TIMEOUT_S = int(os.environ.get("OPENPI_SERVICE_READY_TIMEOUT_S", "900"))
PROMPT = "Move the pan forward and use the brush in the middle of the plates to brush the inside of the pan"
SESSION_ID = "openpi-e2e-parity-session"
pytestmark = [
pytest.mark.skipif(not torch.cuda.is_available(), reason="GPU required"),
pytest.mark.skipif(
dreamzero_client is None or WebsocketClientPolicy is None,
reason="DreamZero client modules are required on PYTHONPATH",
),
pytest.mark.skipif(
DREAMZERO_REPO is None or not DREAMZERO_REPO.exists(),
reason="DreamZero source repo is required at DREAMZERO_REPO",
),
pytest.mark.skipif(
CHECKPOINT_DIR is None or not CHECKPOINT_DIR.exists(), reason="DreamZero local checkpoint is required"
),
]
class OpenPIWebsocketClientPolicy(_BaseWebsocketClientPolicy):
"""DreamZero client protocol with an OpenPI websocket path suffix."""
def __init__(
self,
host: str = "127.0.0.1",
port: int = 8000,
path: str = "/v1/realtime/robot/openpi",
) -> None:
self._uri = f"ws://{host}:{port}{path}"
self._packer = msgpack_numpy.Packer()
self._ws, self._server_metadata = self._wait_for_server()
def _vllm_executable() -> str:
fallback = Path(sys.executable).with_name("vllm")
if fallback.exists():
return str(fallback)
exe = shutil.which("vllm")
if exe:
return exe
raise FileNotFoundError("Unable to locate `vllm` executable in current environment.")
def _cfg_parallel_size() -> int:
return int(os.environ.get("OPENPI_E2E_CFG_PARALLEL_SIZE", "2"))
def _pick_test_gpus() -> list[str]:
cfg_parallel_size = _cfg_parallel_size()
override = os.environ.get("OPENPI_E2E_GPUS") or os.environ.get("OPENPI_E2E_GPU")
if override is not None:
gpus = [part.strip() for part in override.split(",") if part.strip()]
if not gpus:
raise ValueError("OPENPI_E2E_GPUS is set but empty.")
if len(gpus) < cfg_parallel_size:
raise RuntimeError(f"Need {cfg_parallel_size} GPUs, but OPENPI_E2E_GPUS only provided {gpus}.")
return gpus
query = subprocess.check_output(
[
"nvidia-smi",
"--query-gpu=index,memory.used",
"--format=csv,noheader,nounits",
],
text=True,
)
gpu_rows = []
for line in query.strip().splitlines():
gpu_index, used_mb = [part.strip() for part in line.split(",", maxsplit=1)]
gpu_rows.append((int(used_mb), gpu_index))
gpu_rows.sort()
gpus = [gpu_index for _, gpu_index in gpu_rows[: max(cfg_parallel_size, 1)]]
if len(gpus) < cfg_parallel_size:
raise RuntimeError(
f"Need {cfg_parallel_size} GPUs for cfg_parallel_size={cfg_parallel_size}, "
f"but found only {len(gpus)} candidates."
)
return gpus
def _torchrun_argv(script: str, port: int) -> list[str]:
return [
sys.executable,
"-m",
"torch.distributed.run",
"--standalone",
"--nproc_per_node",
str(_cfg_parallel_size()),
script,
"--port",
str(port),
"--model_path",
str(CHECKPOINT_DIR),
]
def _run_upstream_service(port: int, log_path: Path) -> subprocess.Popen[str]:
env = os.environ.copy()
env.setdefault("PYTHONPATH", "")
env["PYTHONPATH"] = f"{Path.cwd()}:{DREAMZERO_REPO}:{env['PYTHONPATH']}".rstrip(":")
env["CUDA_VISIBLE_DEVICES"] = ",".join(_pick_test_gpus())
env.setdefault("NO_ALBUMENTATIONS_UPDATE", "1")
env["ATTENTION_BACKEND"] = "torch"
env.setdefault("ENABLE_TENSORRT", "false")
env["ENABLE_DIT_CACHE"] = "false"
env["NUM_DIT_STEPS"] = "16"
env["DYNAMIC_CACHE_SCHEDULE"] = "false"
argv = _torchrun_argv(
str(Path("tests/dreamzero/upstream/upstream_socket_server_no_compile.py")),
port,
)
log_file = log_path.open("w")
proc = subprocess.Popen(
argv,
stdout=log_file,
stderr=subprocess.STDOUT,
text=True,
cwd=str(Path.cwd()),
env=env,
)
proc._codex_log_file = log_file # type: ignore[attr-defined]
return proc
def _run_vllm_service(port: int, log_path: Path) -> subprocess.Popen[str]:
env = os.environ.copy()
gpus = _pick_test_gpus()
cfg_parallel_size = _cfg_parallel_size()
if cfg_parallel_size > len(gpus):
raise RuntimeError(
f"cfg_parallel_size={cfg_parallel_size} requires at least {cfg_parallel_size} GPUs, but only got {gpus}."
)
env["CUDA_VISIBLE_DEVICES"] = ",".join(gpus[:cfg_parallel_size])
env.setdefault("ATTENTION_BACKEND", "torch")
env.setdefault("DIFFUSION_ATTENTION_BACKEND", "TORCH_SDPA")
env.setdefault("MASTER_PORT", str(get_open_port()))
argv = [
_vllm_executable(),
"serve",
VLLM_MODEL,
"--omni",
"--host",
"127.0.0.1",
"--port",
str(port),
"--served-model-name",
"dreamzero-droid",
"--enforce-eager",
]
if cfg_parallel_size > 1:
argv.extend(["--cfg-parallel-size", str(cfg_parallel_size)])
log_file = log_path.open("w")
proc = subprocess.Popen(
argv,
stdout=log_file,
stderr=subprocess.STDOUT,
text=True,
env=env,
cwd=str(Path.cwd()),
)
proc._codex_log_file = log_file # type: ignore[attr-defined]
return proc
def _stop_process(proc: subprocess.Popen[str]) -> None:
log_file = getattr(proc, "_codex_log_file", None)
if proc.poll() is None:
proc.terminate()
try:
proc.wait(timeout=30)
except subprocess.TimeoutExpired: # pragma: no cover - cleanup path
proc.kill()
proc.wait(timeout=10)
if log_file is not None:
log_file.close()
def _build_obs_sequence() -> tuple[dict, dict]:
camera_frames = dreamzero_client.load_camera_frames()
chunks = dreamzero_client.build_frame_schedule(
min(v.shape[0] for v in camera_frames.values()),
1,
)
obs0 = dreamzero_client._make_obs_from_video(camera_frames, [0], PROMPT, SESSION_ID)
obs1 = dreamzero_client._make_obs_from_video(camera_frames, chunks[0], PROMPT, SESSION_ID)
return obs0, obs1
def _wait_for_client_ready(client_factory, timeout_s: float, proc=None, log_path: Path | None = None):
deadline = time.time() + timeout_s
last_err: Exception | None = None
while time.time() < deadline:
if proc is not None and proc.poll() is not None:
details = ""
if log_path is not None and log_path.exists():
details = log_path.read_text(errors="replace")[-8000:]
raise RuntimeError(f"Service exited before becoming ready with code {proc.returncode}.\n{details}")
try:
return client_factory()
except Exception as exc: # pragma: no cover - retry path
last_err = exc
time.sleep(1)
raise TimeoutError(f"Timed out waiting for websocket service: {last_err}")
def _collect_outputs_with_client(client) -> tuple[dict, list[np.ndarray]]:
metadata = client.get_server_metadata()
obs0, obs1 = _build_obs_sequence()
outputs = [
client.infer(dict(obs0)),
client.infer(dict(obs1)),
]
assert _normalize_reset_response(client.reset({})) == "reset successful"
outputs.append(client.infer(dict(obs0)))
client._ws.close()
return metadata, outputs
def _normalize_reset_response(response) -> str:
if isinstance(response, str):
return response
decoded = msgpack_numpy.unpackb(response)
if isinstance(decoded, dict):
return str(decoded.get("status"))
return str(decoded)
def _normalize_metadata(metadata: dict) -> dict:
normalized = dict(metadata)
if isinstance(normalized.get("image_resolution"), tuple):
normalized["image_resolution"] = list(normalized["image_resolution"])
return normalized
def _assert_logs_clean(log_path: Path) -> None:
text = log_path.read_text(errors="replace")
if "SignalException: Process" in text and "got signal: 15" in text:
text = text.split("Traceback (most recent call last):", 1)[0]
assert "Traceback" not in text, text
assert "RuntimeError:" not in text, text
def _assert_upstream_log_matches_vllm_baseline(log_path: Path) -> None:
text = log_path.read_text(errors="replace")
assert "DIT Compute Steps 8 steps" not in text, text
assert "DIT Compute Steps 16 steps" in text, text
def test_openpi_service_matches_upstream_server_noncompile(tmp_path: Path) -> None:
expected_metadata = {
"image_resolution": [180, 320],
"n_external_cameras": 2,
"needs_wrist_camera": True,
"needs_stereo_camera": False,
"needs_session_id": True,
"action_space": "joint_position",
}
upstream_port = get_open_port()
upstream_log = tmp_path / "dreamzero_upstream.log"
upstream_proc = _run_upstream_service(upstream_port, upstream_log)
try:
upstream_client = _wait_for_client_ready(
lambda: WebsocketClientPolicy(host="127.0.0.1", port=upstream_port),
timeout_s=SERVICE_READY_TIMEOUT_S,
proc=upstream_proc,
log_path=upstream_log,
)
upstream_metadata, upstream_outputs = _collect_outputs_with_client(upstream_client)
finally:
_stop_process(upstream_proc)
_assert_logs_clean(upstream_log)
_assert_upstream_log_matches_vllm_baseline(upstream_log)
vllm_port = get_open_port()
vllm_log = tmp_path / "vllm_openpi.log"
vllm_proc = _run_vllm_service(vllm_port, vllm_log)
try:
vllm_client = _wait_for_client_ready(
lambda: OpenPIWebsocketClientPolicy(host="127.0.0.1", port=vllm_port),
timeout_s=SERVICE_READY_TIMEOUT_S,
proc=vllm_proc,
log_path=vllm_log,
)
vllm_metadata, vllm_outputs = _collect_outputs_with_client(vllm_client)
finally:
_stop_process(vllm_proc)
_assert_logs_clean(vllm_log)
assert _normalize_metadata(upstream_metadata) == expected_metadata
assert _normalize_metadata(vllm_metadata) == expected_metadata
for idx, (actual, expected) in enumerate(zip(vllm_outputs, upstream_outputs, strict=True)):
np.testing.assert_allclose(
actual,
expected,
rtol=1e-2,
atol=1e-3,
err_msg=f"OpenPI step {idx} output mismatch",
)
@@ -1,184 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Launch the upstream DreamZero websocket server with `torch.compile` disabled.
This wrapper is meant for formal parity tests against `vllm serve --omni`.
It monkeypatches `torch.compile` before importing DreamZero modules so that
all import-time decorators and `post_initialize()` compile calls become eager.
For the current DreamZero port baseline we also disable all upstream DiT cache
and step-skipping behavior so the reference server matches vLLM's current
eager/no-skip implementation:
- `ENABLE_DIT_CACHE=false`
- `NUM_DIT_STEPS=16`
- `DYNAMIC_CACHE_SCHEDULE=false`
In our CI/dev environment we also do not install Transformer Engine or
FlashAttention, so upstream's default `socket_test_optimized_AR.main()` path
(`ATTENTION_BACKEND=TE` and then fallback to `FA2`) cannot execute. To keep the
formal server-vs-server test runnable without pulling in those heavyweight
optional deps, this wrapper reproduces upstream `main()` but pins attention to
PyTorch SDPA (`ATTENTION_BACKEND=torch`) for this subprocess only.
Usage:
PYTHONPATH="${DREAMZERO_REPO}" \\
.venv/bin/python -m torch.distributed.run --standalone --nproc_per_node=2 \\
tests/dreamzero/upstream/upstream_socket_server_no_compile.py --port 18081 \\
--model_path "${DREAMZERO_REPO}/checkpoints/dreamzero"
"""
from __future__ import annotations
import os
import sys
from pathlib import Path
import torch
import torch.nn.functional as F
def _identity_compile(*args, **kwargs):
if args and callable(args[0]) and len(args) == 1 and not kwargs:
return args[0]
def deco(fn):
return fn
return deco
torch.compile = _identity_compile
os.environ.setdefault("NO_ALBUMENTATIONS_UPDATE", "1")
DREAMZERO_REPO_ENV = os.environ.get("DREAMZERO_REPO")
if not DREAMZERO_REPO_ENV:
raise RuntimeError("Set DREAMZERO_REPO to an upstream DreamZero checkout before launching this helper.")
DREAMZERO_REPO = Path(DREAMZERO_REPO_ENV).expanduser()
if str(DREAMZERO_REPO) not in sys.path:
sys.path.insert(0, str(DREAMZERO_REPO))
import socket_test_optimized_AR as upstream # noqa: E402
import tyro # noqa: E402
from groot.vla.model.dreamzero.modules import attention as upstream_attention # noqa: E402
from groot.vla.model.dreamzero.modules import wan2_1_submodule as upstream_submodule # noqa: E402
from groot.vla.model.dreamzero.modules import wan_video_dit as upstream_wan_video_dit # noqa: E402
def _torch_varlen_flash_attention(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
q_lens: torch.Tensor | None = None,
k_lens: torch.Tensor | None = None,
dropout_p: float = 0.0,
causal: bool = False,
dtype: torch.dtype = torch.bfloat16,
**_: object,
) -> torch.Tensor:
if q_lens is not None or k_lens is not None:
upstream_attention.warnings.warn(
"Padding mask is disabled in the test-only SDPA fallback.",
)
out_dtype = q.dtype
q = q.transpose(1, 2).to(dtype)
k = k.transpose(1, 2).to(dtype)
v = v.transpose(1, 2).to(dtype)
out = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=None,
is_causal=causal,
dropout_p=dropout_p,
)
return out.transpose(1, 2).contiguous().to(out_dtype)
upstream_attention.flash_attention = _torch_varlen_flash_attention
upstream_submodule.flash_attention = _torch_varlen_flash_attention
upstream_wan_video_dit.FLASH_ATTN_3_AVAILABLE = False
upstream_wan_video_dit.FLASH_ATTN_2_AVAILABLE = False
upstream_wan_video_dit.SAGE_ATTN_AVAILABLE = False
def main(args: upstream.Args) -> None:
os.environ["ENABLE_DIT_CACHE"] = "false"
os.environ["ATTENTION_BACKEND"] = "torch"
os.environ["NUM_DIT_STEPS"] = "16"
os.environ["DYNAMIC_CACHE_SCHEDULE"] = "false"
torch._dynamo.config.recompile_limit = 800
embodiment_tag = "oxe_droid"
model_path = args.model_path
device_mesh = upstream.init_mesh()
rank = upstream.dist.get_rank()
timeout_delta = upstream.datetime.timedelta(seconds=args.timeout_seconds)
signal_group = upstream.dist.new_group(backend="gloo", timeout=timeout_delta)
upstream.logger.info("Rank %s initialized signal_group (gloo)", rank)
policy = upstream.GrootSimPolicy(
embodiment_tag=upstream.EmbodimentTag(embodiment_tag),
model_path=model_path,
device="cuda" if torch.cuda.is_available() else "cpu",
device_mesh=device_mesh,
)
hostname = upstream.socket.gethostname()
local_ip = upstream.socket.gethostbyname(hostname)
if rank == 0:
upstream.logging.info("Creating server (host: %s, ip: %s)", hostname, local_ip)
parent_dir = os.path.dirname(model_path)
date_suffix = upstream.datetime.datetime.now().strftime("%Y%m%d")
checkpoint_name = os.path.basename(model_path)
output_dir = os.path.join(
parent_dir,
f"real_world_eval_gen_{date_suffix}_{args.index}",
checkpoint_name,
)
os.makedirs(output_dir, exist_ok=True)
upstream.logging.info("Videos will be saved to: %s", output_dir)
else:
output_dir = None
upstream.logging.info("Rank %s starting as worker for distributed inference...", rank)
wrapper_policy = upstream.ARDroidRoboarenaPolicy(
groot_policy=policy,
signal_group=signal_group,
output_dir=output_dir,
)
server_config = upstream.PolicyServerConfig(
image_resolution=(180, 320),
needs_wrist_camera=True,
n_external_cameras=2,
needs_stereo_camera=False,
needs_session_id=True,
action_space="joint_position",
)
if rank == 0:
upstream.logging.info("Using roboarena policy server interface")
upstream.logging.info("Server config: %s", server_config)
roboarena_server = upstream.RoboarenaServer(
policy=wrapper_policy,
server_config=server_config,
host="0.0.0.0",
port=args.port,
)
roboarena_server.serve_forever()
else:
worker = upstream.DistributedWorker(
policy=policy,
signal_group=signal_group,
)
upstream.asyncio.run(worker.run())
if __name__ == "__main__":
upstream.logging.basicConfig(level=upstream.logging.INFO, force=True)
main(tyro.cli(upstream.Args))
@@ -8,7 +8,7 @@ L4-style smoke runs against an already-running server. The pytest wrappers in
H100 / MI325) because they launch the live Omni server inside the test. H100 / MI325) because they launch the live Omni server inside the test.
1. **Daily-Omni** — MCQ accuracy fields in the saved JSON (``daily_omni_accuracy``, …); by default the 1. **Daily-Omni** — MCQ accuracy fields in the saved JSON (``daily_omni_accuracy``, …); by default the
run **fails** if accuracy is strictly below **0.69** (``--min-daily-omni-accuracy`` / ``ACC_BENCH_MIN_DAILY_OMNI_ACCURACY``). run **fails** if accuracy is strictly below **0.67** (``--min-daily-omni-accuracy`` / ``ACC_BENCH_MIN_DAILY_OMNI_ACCURACY``).
2. **Seed-TTS** — ``seed-tts-eval``-style metrics when ``--seed-tts-wer-eval`` is used 2. **Seed-TTS** — ``seed-tts-eval``-style metrics when ``--seed-tts-wer-eval`` is used
(WER / SIM / UTMOS keys from :func:`compute_seed_tts_wer_metrics`). (WER / SIM / UTMOS keys from :func:`compute_seed_tts_wer_metrics`).
@@ -308,9 +308,9 @@ def build_arg_parser() -> argparse.ArgumentParser:
p.add_argument( p.add_argument(
"--min-daily-omni-accuracy", "--min-daily-omni-accuracy",
type=float, type=float,
default=float((os.environ.get("ACC_BENCH_MIN_DAILY_OMNI_ACCURACY") or "0.69").strip() or "0.69"), default=float((os.environ.get("ACC_BENCH_MIN_DAILY_OMNI_ACCURACY") or "0.67").strip() or "0.67"),
help="Fail when daily_omni_accuracy is strictly below this threshold (0–1). " help="Fail when daily_omni_accuracy is strictly below this threshold (0–1). "
"Default baseline 0.69; override with env ACC_BENCH_MIN_DAILY_OMNI_ACCURACY or pass 0 to disable the floor.", "Default baseline 0.67; override with env ACC_BENCH_MIN_DAILY_OMNI_ACCURACY or pass 0 to disable the floor.",
) )
p.add_argument( p.add_argument(
@@ -340,7 +340,7 @@ def build_arg_parser() -> argparse.ArgumentParser:
p.add_argument( p.add_argument(
"--max-seed-tts-mean-wer", "--max-seed-tts-mean-wer",
type=float, type=float,
default=0.35, default=0.5,
help="If set, fail when seed_tts_content_error_mean is strictly above this value.", help="If set, fail when seed_tts_content_error_mean is strictly above this value.",
) )
p.add_argument( p.add_argument(
@@ -218,7 +218,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> tuple[Image.I
torch_dtype=torch.bfloat16, torch_dtype=torch.bfloat16,
trust_remote_code=True, trust_remote_code=True,
).to("cuda") ).to("cuda")
pipe.transformer.set_attention_backend("_flash_3_hub")
_diffusers_dummy_run(pipe) _diffusers_dummy_run(pipe)
generator = torch.Generator(device="cuda").manual_seed(SEED) generator = torch.Generator(device="cuda").manual_seed(SEED)
@@ -1,218 +0,0 @@
from __future__ import annotations
import base64
import copy
import io
import os
import tempfile
from pathlib import Path
import pytest
import requests
import yaml
from PIL import Image
from tests.e2e.accuracy.helpers import assert_images_pixel_close, assert_similarity, model_output_dir
from tests.helpers.mark import hardware_test
from tests.helpers.runtime import OmniServer
pytestmark = [pytest.mark.full_model, pytest.mark.diffusion]
MODEL_NAME = "tencent/HunyuanImage-3.0-Instruct"
SEED = 42
NUM_INFERENCE_STEPS = 50
GUIDANCE_SCALE = 2.5
HEIGHT = 1024
WIDTH = 1024
PROMPT = "A brown and white dog is running on the grass."
MEAN_THRESHOLD = 3e-2
P99_THRESHOLD = 3e-1
SSIM_THRESHOLD = 0.97
PSNR_THRESHOLD = 30.0
_REPO_ROOT = Path(__file__).resolve().parent.parent.parent.parent
BASELINE_PATH = _REPO_ROOT / "tests" / "assets" / "hunyuan" / "hunyuan_baseline.png"
_OFFLINE_SCRIPT = _REPO_ROOT / "examples" / "offline_inference" / "hunyuan_image3" / "end2end.py"
# DiT-only deploy config with trust_remote_code (based on hunyuan_image3_dit.yaml).
_DEPLOY_CONFIG = {
"pipeline": "hunyuan_image3_dit",
"async_chunk": False,
"trust_remote_code": True,
"stages": [
{
"stage_id": 0,
"max_num_seqs": 1,
"gpu_memory_utilization": 0.9,
"enforce_eager": True,
"trust_remote_code": True,
"devices": "0,1,2,3", # set dynamically by _write_deploy_config
"vae_use_slicing": False,
"vae_use_tiling": False,
"parallel_config": {
"pipeline_parallel_size": 1,
"data_parallel_size": 1,
"tensor_parallel_size": 4,
"enable_expert_parallel": True,
"sequence_parallel_size": 1,
"ulysses_degree": 1,
"ring_degree": 1,
"cfg_parallel_size": 1,
"vae_patch_parallel_size": 1,
"use_hsdp": False,
"hsdp_shard_size": -1,
"hsdp_replicate_size": 1,
},
"default_sampling_params": {
"seed": SEED,
},
},
],
}
def _model_name() -> str:
return os.environ.get("HUNYUAN_IMAGE3_MODEL", MODEL_NAME)
def _devices() -> str:
return os.environ.get("HUNYUAN_IMAGE3_DEVICES", "0,1,2,3")
def _write_deploy_config(path: Path) -> None:
config = copy.deepcopy(_DEPLOY_CONFIG)
devices = _devices()
config["stages"][0]["devices"] = devices
config["stages"][0]["parallel_config"]["tensor_parallel_size"] = len(devices.split(","))
path.write_text(yaml.dump(config, default_flow_style=False, sort_keys=False))
def _run_vllm_omni_hunyuan_image3_online(*, model: str, deploy_config: str, output_path: Path) -> Image.Image:
server_args = [
"--deploy-config",
deploy_config,
"--stage-init-timeout",
"300",
"--init-timeout",
"900",
"--enforce-eager",
"--trust-remote-code",
]
with OmniServer(model, server_args, use_omni=True) as omni_server:
response = requests.post(
f"http://{omni_server.host}:{omni_server.port}/v1/images/generations",
json={
"model": omni_server.model,
"prompt": PROMPT,
"size": f"{WIDTH}x{HEIGHT}",
"n": 1,
"response_format": "b64_json",
"num_inference_steps": NUM_INFERENCE_STEPS,
"guidance_scale": GUIDANCE_SCALE,
"seed": SEED,
"bot_task": "none",
"use_system_prompt": "en_unified",
},
timeout=600,
)
response.raise_for_status()
payload = response.json()
assert len(payload["data"]) == 1
image_bytes = base64.b64decode(payload["data"][0]["b64_json"])
image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
image.load()
image.save(output_path)
return image
def _run_vllm_omni_hunyuan_image3_offline(*, model: str, deploy_config: str, output_path: Path) -> Image.Image:
import subprocess
import sys
output_dir = str(output_path.parent)
subprocess.run(
[
sys.executable,
str(_OFFLINE_SCRIPT),
"--modality",
"text2img",
"--deploy-config",
deploy_config,
"--prompts",
PROMPT,
"--output",
output_dir,
"--steps",
str(NUM_INFERENCE_STEPS),
"--guidance-scale",
str(GUIDANCE_SCALE),
"--seed",
str(SEED),
"--height",
str(HEIGHT),
"--width",
str(WIDTH),
"--bot-task",
"none",
"--sys-type",
"en_unified",
"--model",
model,
"--enforce-eager",
],
check=True,
)
images = sorted(Path(output_dir).glob("output_*.png"))
assert images, f"No output image found in {output_dir}"
image = Image.open(images[0]).convert("RGB")
image.load()
image.save(output_path)
return image
def _assert_against_baseline(image: Image.Image, label: str) -> None:
assert BASELINE_PATH.exists(), f"Baseline image not found at {BASELINE_PATH}"
baseline_image = Image.open(BASELINE_PATH).convert("RGB")
assert_images_pixel_close(
model_name=f"{MODEL_NAME} ({label} vs baseline)",
vllm_image=image,
diffusers_image=baseline_image,
mean_threshold=MEAN_THRESHOLD,
p99_threshold=P99_THRESHOLD,
)
assert_similarity(
model_name=f"{MODEL_NAME} ({label} vs baseline)",
vllm_image=image,
diffusers_image=baseline_image,
ssim_threshold=SSIM_THRESHOLD,
psnr_threshold=PSNR_THRESHOLD,
)
@hardware_test(res={"cuda": "H100"}, num_cards=4)
def test_hunyuan_image3_pixel_accuracy_online(accuracy_artifact_root: Path) -> None:
model = _model_name()
output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME)
with tempfile.TemporaryDirectory() as tmpdir:
deploy_config_path = Path(tmpdir) / "deploy.yaml"
_write_deploy_config(deploy_config_path)
image = _run_vllm_omni_hunyuan_image3_online(
model=model, deploy_config=str(deploy_config_path), output_path=output_dir / "vllm_omni_online.png"
)
_assert_against_baseline(image, "online")
@hardware_test(res={"cuda": "H100"}, num_cards=4)
def test_hunyuan_image3_pixel_accuracy_offline(accuracy_artifact_root: Path) -> None:
model = _model_name()
output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME)
with tempfile.TemporaryDirectory() as tmpdir:
deploy_config_path = Path(tmpdir) / "deploy.yaml"
_write_deploy_config(deploy_config_path)
image = _run_vllm_omni_hunyuan_image3_offline(
model=model, deploy_config=str(deploy_config_path), output_path=output_dir / "vllm_omni_offline.png"
)
_assert_against_baseline(image, "offline")
-2
View File
@@ -106,7 +106,6 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> Image.Image:
trust_remote_code=True, trust_remote_code=True,
local_files_only=_local_files_only(model), local_files_only=_local_files_only(model),
).to("cuda") ).to("cuda")
pipe.transformer.set_attention_backend("_flash_3_hub")
generator = torch.Generator(device="cuda").manual_seed(SEED) generator = torch.Generator(device="cuda").manual_seed(SEED)
result = pipe( # pyright: ignore[reportCallIssue] result = pipe( # pyright: ignore[reportCallIssue]
prompt=PROMPT, prompt=PROMPT,
@@ -168,7 +167,6 @@ def _run_diffusers_qwen_image_2512(*, model: str, output_path: Path) -> Image.Im
trust_remote_code=True, trust_remote_code=True,
local_files_only=_local_files_only(model), local_files_only=_local_files_only(model),
).to("cuda") ).to("cuda")
pipe.transformer.set_attention_backend("_flash_3_hub")
generator = torch.Generator(device="cuda").manual_seed(SEED_2512) generator = torch.Generator(device="cuda").manual_seed(SEED_2512)
result = pipe( # pyright: ignore[reportCallIssue] result = pipe( # pyright: ignore[reportCallIssue]
prompt=PROMPT_2512, prompt=PROMPT_2512,
@@ -1,273 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""End-to-end ModelOpt FP8 quality accuracy for video-gen DiTs.
Modeled on ``test_hunyuan_image3.py``'s ``test_quantized_dit_matches_bf16_accuracy``:
for each calibrated FP8 checkpoint the test generates one video with BF16 and one
with the FP8 DiT under the same seed, scores prompt-faithfulness with CLIP on the
middle frame, then asserts both ``CLIP >= absolute_floor`` and
``CLIP_drop <= drop_threshold``. The test is gated by an opt-in env var (model
paths are local) so it never runs in generic CI.
To run::
VIDEOGEN_RUN_QUANT_ACCURACY=1 \\
WAN22_A14B_BF16_MODEL=Wan-AI/Wan2.2-T2V-A14B-Diffusers \\
WAN22_A14B_FP8_MODEL=/path/to/wan22-a14b-modelopt-fp8 \\
HV15_BF16_MODEL=hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v \\
HV15_FP8_MODEL=/path/to/hv15-modelopt-fp8 \\
pytest -s -v tests/e2e/accuracy/test_videogen_modelopt_quant.py
"""
from __future__ import annotations
import gc
import json
import os
import tempfile
from dataclasses import dataclass
from pathlib import Path
from typing import Any
import numpy as np
import pytest
import torch
from PIL import Image
from tests.e2e.accuracy.helpers import CLIPScorer, model_output_dir
pytestmark = [pytest.mark.local_model, pytest.mark.diffusion]
# ---------------------------------------------------------------------------
# Tunables (override via env to suit local hardware / calibration budget)
# ---------------------------------------------------------------------------
SEED = 42
HEIGHT = int(os.environ.get("VIDEOGEN_QUANT_HEIGHT", "480"))
WIDTH = int(os.environ.get("VIDEOGEN_QUANT_WIDTH", "832"))
NUM_FRAMES = int(os.environ.get("VIDEOGEN_QUANT_NUM_FRAMES", "17"))
NUM_INFERENCE_STEPS = int(os.environ.get("VIDEOGEN_QUANT_STEPS", "20"))
WAN_PROMPT = "A red fox running through a snowy forest at sunrise, cinematic."
HV_PROMPT = "A red fox running through a snowy forest at sunrise, cinematic."
# CLIP gates: absolute floor + max drop from the BF16 reference.
CLIP_ABSOLUTE_FLOOR = float(os.environ.get("VIDEOGEN_QUANT_CLIP_FLOOR", "20.0"))
CLIP_DROP_FP8 = float(os.environ.get("VIDEOGEN_QUANT_CLIP_DROP_FP8", "7.0"))
RUN_ENV = "VIDEOGEN_RUN_QUANT_ACCURACY"
_TRUE = {"1", "true", "yes", "on"}
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class _Spec:
"""A model under test (BF16 + its FP8 variant)."""
family: str
bf16_env: str
fp8_env: str
prompt: str
guidance_scale: float
guidance_scale_2: float | None = None
boundary_ratio: float | None = None
WAN_SPEC = _Spec(
family="wan22_a14b_t2v",
bf16_env="WAN22_A14B_BF16_MODEL",
fp8_env="WAN22_A14B_FP8_MODEL",
prompt=WAN_PROMPT,
guidance_scale=3.5,
guidance_scale_2=4.0,
boundary_ratio=0.875,
)
HV_SPEC = _Spec(
family="hunyuanvideo15_t2v",
bf16_env="HV15_BF16_MODEL",
fp8_env="HV15_FP8_MODEL",
prompt=HV_PROMPT,
guidance_scale=6.0,
)
@dataclass(frozen=True)
class _QuantCase:
name: str
spec: _Spec
clip_drop: float
def _cases() -> list[pytest.ParameterSet]:
cases = [
_QuantCase("wan_fp8", WAN_SPEC, CLIP_DROP_FP8),
_QuantCase("hv_fp8", HV_SPEC, CLIP_DROP_FP8),
]
params: list[pytest.ParameterSet] = []
opted_in = os.environ.get(RUN_ENV, "").lower() in _TRUE
for case in cases:
marks: list[Any] = []
if not opted_in:
marks.append(pytest.mark.skip(reason=f"Set {RUN_ENV}=1 to run video-gen quant accuracy."))
if not os.environ.get(case.spec.bf16_env):
marks.append(pytest.mark.skip(reason=f"Set {case.spec.bf16_env} to run {case.name}."))
if not os.environ.get(case.spec.fp8_env):
marks.append(pytest.mark.skip(reason=f"Set {case.spec.fp8_env} to run {case.name}."))
params.append(pytest.param(case, id=case.name, marks=marks))
return params
# ---------------------------------------------------------------------------
def _middle_frame(frames: list[Image.Image] | list[np.ndarray] | np.ndarray) -> Image.Image:
if isinstance(frames, np.ndarray):
idx = frames.shape[0] // 2
arr = frames[idx]
if arr.dtype != np.uint8:
arr = (arr * 255).clip(0, 255).astype(np.uint8)
return Image.fromarray(arr).convert("RGB")
if not frames:
raise AssertionError("Empty video output")
mid = frames[len(frames) // 2]
if isinstance(mid, np.ndarray):
if mid.dtype != np.uint8:
mid = (mid * 255).clip(0, 255).astype(np.uint8)
return Image.fromarray(mid).convert("RGB")
return mid.convert("RGB")
def _extract_frames(result) -> list[Image.Image]:
"""Pull a frame list out of an Omni text-to-video result."""
from vllm_omni.entrypoints.omni import OmniRequestOutput
if isinstance(result, list):
result = result[0] if result else None
if isinstance(result, OmniRequestOutput) and result.is_pipeline_output and result.request_output is not None:
result = result.request_output
images = getattr(result, "images", None)
if not images:
raise AssertionError("Pipeline output had no images")
if isinstance(images, list) and images and isinstance(images[0], tuple):
# (frames, fps)
images = images[0][0]
return list(images)
def _generate_video(
*,
model: str,
spec: _Spec,
quantization: str | None,
output_path: Path,
) -> tuple[Image.Image, float]:
"""Generate one video and return its middle frame + elapsed wall time."""
from vllm_omni.diffusion.data import DiffusionParallelConfig
from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
from vllm_omni.platforms import current_omni_platform
kwargs: dict[str, Any] = dict(
model=model,
parallel_config=DiffusionParallelConfig(),
enforce_eager=True,
vae_use_tiling=True,
)
if quantization:
kwargs["quantization"] = quantization
if spec.boundary_ratio is not None:
kwargs["boundary_ratio"] = spec.boundary_ratio
omni = Omni(**kwargs)
try:
gen = torch.Generator(device=current_omni_platform.device_type).manual_seed(SEED)
sampling = dict(
height=HEIGHT,
width=WIDTH,
generator=gen,
guidance_scale=spec.guidance_scale,
num_inference_steps=NUM_INFERENCE_STEPS,
num_frames=NUM_FRAMES,
seed=SEED,
)
if spec.guidance_scale_2 is not None:
sampling["guidance_scale_2"] = spec.guidance_scale_2
import time as _time
t0 = _time.perf_counter()
out = omni.generate({"prompt": spec.prompt}, OmniDiffusionSamplingParams(**sampling))
elapsed = _time.perf_counter() - t0
frames = _extract_frames(out)
mid = _middle_frame(frames)
mid.save(output_path)
return mid, elapsed
finally:
del omni
gc.collect()
if torch.cuda.is_available():
torch.accelerator.empty_cache()
@pytest.mark.parametrize("case", _cases())
@pytest.mark.skipif(
not torch.cuda.is_available() or torch.accelerator.device_count() < 1,
reason="Needs at least 1 GPU.",
)
def test_quantized_videogen_dit_matches_bf16_clip(case: _QuantCase, accuracy_artifact_root: Path) -> None:
"""FP8 DiT should preserve prompt-aligned video quality vs BF16."""
out_dir = model_output_dir(accuracy_artifact_root, case.spec.family + "-quant")
with tempfile.TemporaryDirectory():
bf16_model = os.environ[case.spec.bf16_env]
quant_model = os.environ[case.spec.fp8_env]
bf16_frame, bf16_time = _generate_video(
model=bf16_model,
spec=case.spec,
quantization=None,
output_path=out_dir / "bf16.png",
)
quant_frame, quant_time = _generate_video(
model=quant_model,
spec=case.spec,
quantization="fp8",
output_path=out_dir / f"{case.name}.png",
)
clip = CLIPScorer()
bf16_clip = clip.score(bf16_frame, case.spec.prompt)
quant_clip = clip.score(quant_frame, case.spec.prompt)
clip_drop = bf16_clip - quant_clip
speedup = (bf16_time / quant_time) if quant_time else float("nan")
metrics = {
"case": case.name,
"family": case.spec.family,
"quantization": "fp8",
"bf16_model": bf16_model,
"quant_model": quant_model,
"prompt": case.spec.prompt,
"height": HEIGHT,
"width": WIDTH,
"num_frames": NUM_FRAMES,
"num_inference_steps": NUM_INFERENCE_STEPS,
"bf16_elapsed_s": bf16_time,
"quant_elapsed_s": quant_time,
"speedup_vs_bf16": speedup,
"bf16_clip_score": bf16_clip,
"quant_clip_score": quant_clip,
"clip_score_drop": clip_drop,
}
metrics_path = out_dir / f"{case.name}_metrics.json"
metrics_path.write_text(json.dumps(metrics, indent=2, sort_keys=True), encoding="utf-8")
print(f"\n[{case.name}] BF16 {bf16_time:.2f}s | FP8 {quant_time:.2f}s (x{speedup:.2f})")
print(f" CLIP BF16={bf16_clip:.2f} FP8={quant_clip:.2f} drop={clip_drop:.2f}")
print(f" metrics={metrics_path}")
assert quant_clip >= CLIP_ABSOLUTE_FLOOR, (
f"{case.name} CLIP below floor: got {quant_clip:.2f}, expected >= {CLIP_ABSOLUTE_FLOOR:.2f}"
)
assert clip_drop <= case.clip_drop, (
f"{case.name} CLIP drop too large: got {clip_drop:.2f}, expected <= {case.clip_drop:.2f}"
)
+7 -2
View File
@@ -32,8 +32,7 @@ from vllm_omni import Omni
# Selected by the nightly "TTS · Function Test" (-m "full_model and L4 and tts"). # Selected by the nightly "TTS · Function Test" (-m "full_model and L4 and tts").
# Without the tts + run-level marks these tests are collected but never run. # Without the tts + run-level marks these tests are collected but never run.
# TODO: Fix this test pytestmark = [pytest.mark.full_model, pytest.mark.tts]
# pytestmark = [pytest.mark.full_model, pytest.mark.tts]
SAMPLE_RATE = 24_000 # All MOSS-TTS full variants output 24 kHz SAMPLE_RATE = 24_000 # All MOSS-TTS full variants output 24 kHz
@@ -129,6 +128,7 @@ def delay_engine():
return Omni(model=_DELAY_MODEL, stage_init_timeout=300) return Omni(model=_DELAY_MODEL, stage_init_timeout=300)
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_delay_english(delay_engine, ref_audio_path): def test_moss_tts_delay_english(delay_engine, ref_audio_path):
"""MossTTSDelayModel: English voice_clone produces non-empty 24 kHz audio.""" """MossTTSDelayModel: English voice_clone produces non-empty 24 kHz audio."""
@@ -140,6 +140,7 @@ def test_moss_tts_delay_english(delay_engine, ref_audio_path):
assert not torch.all(audio == 0), "Audio is silence" assert not torch.all(audio == 0), "Audio is silence"
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_delay_chinese(delay_engine, ref_audio_path): def test_moss_tts_delay_chinese(delay_engine, ref_audio_path):
"""MossTTSDelayModel: Chinese input produces non-empty audio.""" """MossTTSDelayModel: Chinese input produces non-empty audio."""
@@ -151,6 +152,7 @@ def test_moss_tts_delay_chinese(delay_engine, ref_audio_path):
assert not torch.all(audio == 0) assert not torch.all(audio == 0)
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_delay_deterministic(delay_engine, ref_audio_path): def test_moss_tts_delay_deterministic(delay_engine, ref_audio_path):
"""MossTTSDelayModel: same seed yields identical waveforms.""" """MossTTSDelayModel: same seed yields identical waveforms."""
@@ -162,6 +164,7 @@ def test_moss_tts_delay_deterministic(delay_engine, ref_audio_path):
assert torch.allclose(audio1, audio2, atol=1e-4), "Waveforms differ across identical seeds" assert torch.allclose(audio1, audio2, atol=1e-4), "Waveforms differ across identical seeds"
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_delay_batch(delay_engine, ref_audio_path): def test_moss_tts_delay_batch(delay_engine, ref_audio_path):
"""MossTTSDelayModel: batch of two requests each returns non-empty audio.""" """MossTTSDelayModel: batch of two requests each returns non-empty audio."""
@@ -203,6 +206,7 @@ def realtime_engine():
return Omni(model=_REALTIME_MODEL, stage_init_timeout=300) return Omni(model=_REALTIME_MODEL, stage_init_timeout=300)
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_realtime_english(realtime_engine, ref_audio_path): def test_moss_tts_realtime_english(realtime_engine, ref_audio_path):
"""MossTTSRealtime: English voice_clone produces non-empty 24 kHz audio.""" """MossTTSRealtime: English voice_clone produces non-empty 24 kHz audio."""
@@ -214,6 +218,7 @@ def test_moss_tts_realtime_english(realtime_engine, ref_audio_path):
assert not torch.all(audio == 0), "Audio is silence" assert not torch.all(audio == 0), "Audio is silence"
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_realtime_deterministic(realtime_engine, ref_audio_path): def test_moss_tts_realtime_deterministic(realtime_engine, ref_audio_path):
"""MossTTSRealtime: same seed yields identical waveforms.""" """MossTTSRealtime: same seed yields identical waveforms."""
@@ -120,6 +120,7 @@ def _collect_audio(omni: Omni, request: dict) -> tuple[torch.Tensor, int]:
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_nano_english(omni_runner: OmniRunner, ref_audio_path) -> None: def test_moss_tts_nano_english(omni_runner: OmniRunner, ref_audio_path) -> None:
"""English TTS produces non-empty 48 kHz stereo audio.""" """English TTS produces non-empty 48 kHz stereo audio."""
@@ -132,6 +133,7 @@ def test_moss_tts_nano_english(omni_runner: OmniRunner, ref_audio_path) -> None:
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_nano_chinese(omni_runner: OmniRunner, ref_audio_path) -> None: def test_moss_tts_nano_chinese(omni_runner: OmniRunner, ref_audio_path) -> None:
"""Chinese TTS produces non-empty audio.""" """Chinese TTS produces non-empty audio."""
@@ -144,6 +146,7 @@ def test_moss_tts_nano_chinese(omni_runner: OmniRunner, ref_audio_path) -> None:
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_nano_deterministic(omni_runner: OmniRunner, ref_audio_path) -> None: def test_moss_tts_nano_deterministic(omni_runner: OmniRunner, ref_audio_path) -> None:
"""Same seed produces identical waveforms.""" """Same seed produces identical waveforms."""
@@ -156,6 +159,7 @@ def test_moss_tts_nano_deterministic(omni_runner: OmniRunner, ref_audio_path) ->
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}) @hardware_test(res={"cuda": "L4"})
def test_moss_tts_nano_batch(omni_runner: OmniRunner, ref_audio_path) -> None: def test_moss_tts_nano_batch(omni_runner: OmniRunner, ref_audio_path) -> None:
"""Batch of two requests returns audio for each.""" """Batch of two requests returns audio for each."""
@@ -24,11 +24,10 @@ from vllm import SamplingParams
from tests.helpers.mark import hardware_test from tests.helpers.mark import hardware_test
from vllm_omni import Omni from vllm_omni import Omni
# H100-gated (8B). Picked up by the nightly "TTS · Function Test with H100" # H100-gated (8B). Picked up by the nightly H100 omni job
# job (-m "full_model and H100 and tts"); the L4 MOSS-TTS tests land in the # (-m "full_model and H100 and omni"); the tts mark keeps it consistent with
# sibling L4 tts lane instead. # the other MOSS-TTS tests and any tts-scoped sweeps.
# TODO: Fix this test pytestmark = [pytest.mark.full_model, pytest.mark.tts]
# pytestmark = [pytest.mark.full_model, pytest.mark.tts]
SAMPLE_RATE = 24_000 SAMPLE_RATE = 24_000
REF_AUDIO_URL = "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS/main/assets/audio/reference_zh_1.wav" REF_AUDIO_URL = "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS/main/assets/audio/reference_zh_1.wav"
@@ -102,6 +101,7 @@ def _collect_audio(omni: Omni, request: dict) -> tuple[torch.Tensor, int]:
raise AssertionError("No stage outputs received") raise AssertionError("No stage outputs received")
@pytest.mark.omni
@hardware_test(res={"cuda": "H100"}) @hardware_test(res={"cuda": "H100"})
def test_moss_tts_v15_voice_clone(v15_engine, ref_audio_path): def test_moss_tts_v15_voice_clone(v15_engine, ref_audio_path):
"""MOSS-TTS-v1.5 (8B): voice_clone produces non-empty 24 kHz audio.""" """MOSS-TTS-v1.5 (8B): voice_clone produces non-empty 24 kHz audio."""
@@ -1,90 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""E2E online serving test for DreamZero OpenPI websocket serving."""
from __future__ import annotations
import os
from pathlib import Path
import numpy as np
import pytest
from tests.dreamzero import openpi_client_helper as openpi_client
from tests.helpers.mark import hardware_test
from tests.helpers.runtime import OmniServerParams, get_open_port
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
MODEL = "GEAR-Dreams/DreamZero-DROID"
test_params = [
OmniServerParams(
model=MODEL,
port=8091,
server_args=[
"--deploy-config",
"vllm_omni/deploy/dreamzero_tp1_cfg2.yaml",
"--enforce-eager",
"--disable-log-stats",
],
env_dict={
"ATTENTION_BACKEND": "torch",
"DIFFUSION_ATTENTION_BACKEND": "TORCH_SDPA",
"VLLM_DISABLE_COMPILE_CACHE": "1",
"MASTER_PORT": str(get_open_port()),
},
)
]
def _write_synthetic_video(path: Path, cv2_module, *, channel: int) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
height, width, num_frames = 180, 320, 24
writer = cv2_module.VideoWriter(str(path), cv2_module.VideoWriter_fourcc(*"mp4v"), 15.0, (width, height))
if not writer.isOpened():
raise RuntimeError(f"Failed to open video writer for {path}")
try:
for frame_idx in range(num_frames):
frame = np.zeros((height, width, 3), dtype=np.uint8)
frame[..., channel] = (frame_idx * 7) % 255
frame[..., (channel + 1) % 3] = 64
writer.write(cv2_module.cvtColor(frame, cv2_module.COLOR_RGB2BGR))
finally:
writer.release()
def _write_synthetic_dreamzero_videos(client_mod, video_dir: Path) -> None:
for channel, file_name in enumerate(client_mod.CAMERA_FILES.values()):
_write_synthetic_video(video_dir / file_name, client_mod.cv2, channel=channel)
@pytest.mark.full_model
@pytest.mark.diffusion
@pytest.mark.distributed_cuda
@hardware_test(res={"cuda": "H100"}, num_cards=2)
@pytest.mark.parametrize("omni_server", test_params, indirect=True)
def test_dreamzero_openpi_online(omni_server, tmp_path: Path) -> None:
try:
openpi_client.require_dependencies()
except ModuleNotFoundError as exc:
pytest.skip(str(exc))
video_dir = tmp_path / "dreamzero_videos"
_write_synthetic_dreamzero_videos(openpi_client, video_dir)
result = openpi_client.run_policy_session(
host=omni_server.host,
port=omni_server.port,
video_dir=video_dir,
session_id="dreamzero-online-e2e",
)
openpi_client.validate_session_result(result)
metadata = result["metadata"]
assert metadata["needs_session_id"] is True
assert metadata["needs_stereo_camera"] is False
assert tuple(metadata["image_resolution"]) == (180, 320)
@@ -1,183 +0,0 @@
"""
Comprehensive tests of diffusion features that are available in online serving mode
and are supported by the following model:
- zai-org/GLM-Image (t2i and i2i)
Coverage (per mode):
- Baseline
- Cache-DiT
- Tensor-Parallel
- HSDP
- Sequence parallel
Topology for multi-GPU cases uses runtime deploy YAML (except TP)
so stage-1 parallel fields reach the diffusion worker without relying on CLI
assert_diffusion_response validates successful generation and the expected resolution.
"""
import os
import pytest
from tests.helpers.mark import hardware_marks
from tests.helpers.media import generate_synthetic_image
from tests.helpers.runtime import (
OmniServer,
OmniServerParams,
OpenAIClientHandler,
dummy_messages_from_mix_data,
)
from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
MODEL = os.environ.get("GLM_IMAGE_MODEL_PATH", "zai-org/GLM-Image")
T2I_PROMPT = "A Vincent van Gogh style impressionist painting."
I2I_PROMPT = "Transform this modern, geometric image into a Vincent van Gogh style impressionist painting."
NEGATIVE_PROMPT = "low quality, blurry, distorted, unnatural colors"
TWO_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
GLM_DEPLOY = get_deploy_config_path("glm_image.yaml")
GLM_CACHEDIT_DEPLOY = modify_stage_config(
GLM_DEPLOY,
updates={
"stages": {
1: {"cache_backend": "cache_dit"},
},
},
)
GLM_TP_2_DEPLOY = modify_stage_config(
GLM_DEPLOY,
updates={
"stages": {
0: {"devices": "0", "tensor_parallel_size": 1},
1: {"devices": "0,1", "tensor_parallel_size": 2},
},
},
)
GLM_HSDP_2_DEPLOY = modify_stage_config(
GLM_DEPLOY,
updates={
"stages": {
0: {"devices": "0"},
1: {
"devices": "0,1",
"parallel_config": {
"use_hsdp": True,
"hsdp_shard_size": 2,
},
},
},
},
)
GLM_SP_2_DEPLOY = modify_stage_config(
GLM_DEPLOY,
updates={
"stages": {
0: {"devices": "0"},
1: {
"devices": "0,1",
"parallel_config": {
"ulysses_degree": 2,
"ring_degree": 1,
},
},
},
},
)
def _get_diffusion_feature_cases(model: str):
return [
# Baseline: default deploy (stage 0 -> GPU 0, stage 1 -> GPU 1)
pytest.param(
OmniServerParams(
model=model,
stage_config_path=GLM_DEPLOY,
),
id="baseline",
marks=TWO_CARD_FEATURE_MARKS,
),
# Cache-DiT via runtime deploy patch (cache_backend -> engine_extras on stage 1)
pytest.param(
OmniServerParams(
model=model,
stage_config_path=GLM_CACHEDIT_DEPLOY,
),
id="cachedit",
marks=TWO_CARD_FEATURE_MARKS,
),
# Tensor-Parallel (2 GPUs on diffusion stage)
pytest.param(
OmniServerParams(
model=model,
stage_config_path=GLM_TP_2_DEPLOY,
),
id="tensor_parallel_2",
marks=TWO_CARD_FEATURE_MARKS,
),
# HSDP (2 GPUs on diffusion stage)
pytest.param(
OmniServerParams(
model=model,
stage_config_path=GLM_HSDP_2_DEPLOY,
),
id="hsdp_2",
marks=TWO_CARD_FEATURE_MARKS,
),
# Ulysses sequence parallel degree=2 (2 GPUs)
pytest.param(
OmniServerParams(
model=model,
stage_config_path=GLM_SP_2_DEPLOY,
),
id="sequence_parallel_2",
marks=TWO_CARD_FEATURE_MARKS,
),
]
MODES = ["t2i", "i2i"]
@pytest.mark.parametrize("mode", MODES)
@pytest.mark.parametrize(
"omni_server",
_get_diffusion_feature_cases(MODEL),
indirect=True,
)
def test_glm_image(
omni_server: OmniServer,
mode: str,
openai_client: OpenAIClientHandler,
):
"""GLM-Image diffusion feature coverage for t2i and i2i on two H100s."""
if mode == "t2i":
messages = dummy_messages_from_mix_data(content_text=T2I_PROMPT)
else:
image_size = 1024
image_data_url = f"data:image/jpeg;base64,{generate_synthetic_image(image_size, image_size)['base64']}"
messages = dummy_messages_from_mix_data(
image_data_url=image_data_url,
content_text=I2I_PROMPT,
)
request_config = {
"model": omni_server.model,
"messages": messages,
"extra_body": {
"height": 1024,
"width": 1024,
"num_inference_steps": 2,
"guidance_scale": 1.5,
"true_cfg_scale": 4.0,
"negative_prompt": NEGATIVE_PROMPT,
"seed": 42,
},
}
openai_client.send_diffusion_request(request_config)
@@ -1,388 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""End-to-end online tests for higgs-audio v3 against /v1/audio/speech.
Mirrors the higgs_audio_v2 test layout. Covers the plain-text-in / audio-out
happy path, a small concurrent burst (Stage 0 prefix caching has to coexist
with batching), the voice-clone path via ``ref_audio`` + ``ref_text``, the
upstream ``references[]`` cookbook alias added by
``normalize_references_alias`` in ``protocol/audio.py``, and the few
validator rejections that should remain 4xx.
"""
from __future__ import annotations
import os
os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
os.environ.setdefault("VLLM_USE_DEEP_GEMM", "0")
os.environ.setdefault("VLLM_MOE_USE_DEEP_GEMM", "0")
import pytest
from tests.helpers.mark import hardware_test
from tests.helpers.media import load_test_audio_data_url
from tests.helpers.runtime import OmniServerParams
from tests.helpers.stage_config import get_deploy_config_path
MODEL = "bosonai/higgs-audio-v3-tts-4b"
STAGE_CONFIG = get_deploy_config_path("higgs_multimodal_qwen3.yaml")
SERVER_ARGS = ["--trust-remote-code", "--disable-log-stats"]
SERVER_ENV = {"VLLM_USE_DEEP_GEMM": "0", "VLLM_MOE_USE_DEEP_GEMM": "0"}
TEST_PARAMS = [
pytest.param(
OmniServerParams(
model=MODEL,
stage_config_path=STAGE_CONFIG,
server_args=SERVER_ARGS,
env_dict=SERVER_ENV,
),
id="higgs_audio_v3_plain_text",
)
]
DEFAULT_SPEECH_TIMEOUT_S = 180.0
# Floor for ~0.5 s of 24 kHz mono PCM_16: 24000 * 0.5 * 2 bytes ~= 24 KiB.
# A WAV header adds 44 bytes; conservative floor catches truncated /
# silence-only outputs without flagging short legitimate clips.
_MIN_AUDIO_BYTES = 20_000
# Reuse the shared TTS reference clip (clean ~5 s 24 kHz mono human speech)
# vendored under tests/assets/qwen3_tts/. Keeps a single WAV across TTS
# tests rather than duplicating asset bytes.
_REF_AUDIO_URL = load_test_audio_data_url("qwen3_tts/clone_2.wav")
_REF_TEXT = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."
@pytest.mark.parametrize("omni_server", TEST_PARAMS, indirect=True)
class TestHiggsAudioV3OnlineHappyPath:
"""Plain-text -> audio happy paths against the live HTTP server."""
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_plain_text_wav(self, omni_server, openai_client) -> None:
"""Single non-streaming WAV request - canonical TTS happy path."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "Hello world.",
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_plain_text_with_max_new_tokens(self, omni_server, openai_client) -> None:
"""``max_new_tokens`` is one of the few extra fields the v3 validator accepts."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "Innovation distinguishes between a leader and a follower.",
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"max_new_tokens": 500,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_concurrent_plain_text(self, omni_server, openai_client) -> None:
"""Three concurrent non-streaming requests - guards per-slot audio state and Stage-0 PC under batching."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "It was the night before my birthday.",
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
},
request_num=3,
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_plain_text_pcm_streaming(self, omni_server, openai_client) -> None:
"""Streaming PCM via talker2code2wav_async_chunk + the connector's codec_streaming path.
The deploy YAML pins ``async_chunk: true`` and ``codec_streaming: true`` so chunks
flow Stage 0 -> Stage 1 per AR loop. The Stage-1 codec honors
``meta.left_context_size`` and ``meta.right_holdback_size`` so per-chunk codec
windows stitch into a coherent PCM stream. The byte-count gate is the same as
the sync paths; per-chunk audio content is verified offline against Whisper.
NOTE: ``min_hnr_db=0.0`` sits below the typical speech-noise floor so the
check still catches catastrophic codec failure (silence, white noise, sample
scramble all give HNR << 0) while allowing for the sliding-window codec's
slightly-noisier-than-sync output. The default 1.0 dB threshold has only
~0.16 dB margin over measured single-request output (1.16 dB on L4), so
flake risk is too high there.
"""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "The quick brown fox jumps over the lazy dog.",
"stream": True,
"response_format": "pcm",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
"min_hnr_db": 0.0,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_concurrent_pcm_streaming(self, omni_server, openai_client) -> None:
"""Three concurrent streaming requests - guards per-request frame cursors
in ``talker2code2wav_async_chunk`` and per-slot delay-pattern state under batched AR.
NOTE: ``min_hnr_db=-2.0`` is looser than the single-request streaming test
because batched AR adds per-request codec quality variance (DAC boundary
artifacts compounded across the 3-way batch); -2.0 dB gives ~1.6 dB margin
over the measured worst-of-3 on L4 (-0.42 dB) while still well above the
catastrophic-failure region.
"""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "She sells seashells by the seashore.",
"stream": True,
"response_format": "pcm",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
"min_hnr_db": -2.0,
},
request_num=3,
)
@pytest.mark.parametrize("omni_server", TEST_PARAMS, indirect=True)
class TestHiggsAudioV3OnlineInlineControlTokens:
"""Inline control-token surface from the bosonai cookbook.
The model exposes four categories of ``<|category:value|>`` tokens
(emotion, style, prosody, sfx). Delivery tokens (emotion / style /
prosody speed-pitch-expressive) go at the START of input; positional
tokens (pause, sfx) go inline. SFX tokens must pair with their written
onomatopoeia.
These tests check the *serving surface* - the validator accepts the
payload, the engine produces audio, and the WAV has non-trivial size.
They do not assert audio quality (out of scope for CI).
"""
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_inline_emotion_and_expressive(self, omni_server, openai_client) -> None:
"""Delivery tokens (emotion + expressive_high) at the start of input."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": ("<|emotion:amusement|><|prosody:expressive_high|>Wait, that was actually hilarious."),
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_inline_style_whispering(self, omni_server, openai_client) -> None:
"""Style token at the start - ``whispering``."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "<|style:whispering|>It is just between you and me, alright?",
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_inline_prosody_speed_and_pitch(self, omni_server, openai_client) -> None:
"""Two prosody tokens at the start - slow speed plus low pitch."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": (
"<|prosody:speed_slow|><|prosody:pitch_low|>The radar shows a storm approaching from the east."
),
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_inline_pause_mid_text(self, omni_server, openai_client) -> None:
"""Positional pause token placed inline between two clauses."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "Hold on a moment <|prosody:pause|> let me think about it.",
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_inline_sfx_with_onomatopoeia(self, omni_server, openai_client) -> None:
"""SFX token paired with its written onomatopoeia (``<|sfx:laughter|>Hehe``)."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": (
"<|emotion:amusement|>I cannot believe that just happened. "
"<|sfx:laughter|>Hehe, I am still recovering from it."
),
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.parametrize("omni_server", TEST_PARAMS, indirect=True)
class TestHiggsAudioV3OnlineVoiceClone:
"""Voice clone via the two payload shapes the model serves."""
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_voice_clone_ref_audio_ref_text(self, omni_server, openai_client) -> None:
"""Canonical vllm-omni voice clone via ``ref_audio`` + ``ref_text``."""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "Hello world.",
"ref_audio": _REF_AUDIO_URL,
"ref_text": _REF_TEXT,
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_voice_clone_references_alias(self, omni_server, openai_client) -> None:
"""BosonAI cookbook payload: ``references=[{audio_path, text}]``.
``normalize_references_alias`` (``protocol/audio.py``) is supposed to
translate the cookbook field into ``ref_audio`` / ``ref_text``; without
it the request would silently fall through to zero-shot synthesis.
"""
openai_client.send_audio_speech_request(
{
"model": omni_server.model,
"input": "Hello world.",
"references": [{"audio_path": _REF_AUDIO_URL, "text": _REF_TEXT}],
"stream": False,
"response_format": "wav",
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
)
@pytest.mark.parametrize("omni_server", TEST_PARAMS, indirect=True)
class TestHiggsAudioV3OnlineValidatorRejections:
"""Out-of-scope shapes must come back as 4xx.
These cases use ``send_audio_speech_http_request`` (raw HTTP POST against
``/v1/audio/speech``) rather than the OpenAI SDK helper because:
- We need the failure-path matcher; the SDK helper asserts success.
- The cookbook ``references`` field is not in the OpenAI SDK schema, so
the SDK strips it before the request leaves the client and the server
sees a plain TTS request that returns 200. Raw HTTP keeps the JSON body
intact so the validator actually fires.
"""
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_rejects_empty_input(self, omni_server, openai_client) -> None:
"""Empty text input - validator should reject before reaching the engine."""
openai_client.send_audio_speech_http_request(
{
"json": {
"model": omni_server.model,
"input": "",
"response_format": "wav",
},
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"err_code": (400, 422),
"err_message": ("input", "empty"),
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_rejects_multi_reference_payload(self, omni_server, openai_client) -> None:
"""``references[]`` with more than one entry - multi-shot voice clone is not supported."""
openai_client.send_audio_speech_http_request(
{
"json": {
"model": omni_server.model,
"input": "Hello world.",
"references": [
{"audio_path": _REF_AUDIO_URL, "text": _REF_TEXT},
{"audio_path": _REF_AUDIO_URL, "text": _REF_TEXT},
],
"response_format": "wav",
},
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"err_code": (400, 422),
"err_message": "references",
}
)
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "H100"}, num_cards=1)
def test_rejects_conflicting_ref_audio_and_references(self, omni_server, openai_client) -> None:
"""``ref_audio`` and ``references`` cannot both be set to different values."""
openai_client.send_audio_speech_http_request(
{
"json": {
"model": omni_server.model,
"input": "Hello world.",
"ref_audio": _REF_AUDIO_URL,
"references": [{"audio_path": "https://example.com/other.wav"}],
"response_format": "wav",
},
"timeout": DEFAULT_SPEECH_TIMEOUT_S,
"err_code": (400, 422),
"err_message": "mutually exclusive",
}
)
@@ -30,8 +30,7 @@ from tests.helpers.mark import hardware_test
from tests.helpers.runtime import OmniServerParams from tests.helpers.runtime import OmniServerParams
from tests.helpers.stage_config import get_deploy_config_path from tests.helpers.stage_config import get_deploy_config_path
# TODO: Fix this test pytestmark = [pytest.mark.full_model, pytest.mark.tts]
# pytestmark = [pytest.mark.full_model, pytest.mark.tts]
MODEL = "OpenMOSS-Team/MOSS-TTS-Realtime" MODEL = "OpenMOSS-Team/MOSS-TTS-Realtime"
REF_AUDIO_URL = "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS/HEAD/assets/audio/reference_zh_1.wav" REF_AUDIO_URL = "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS/HEAD/assets/audio/reference_zh_1.wav"
+95 -91
View File
@@ -9,40 +9,63 @@ import pytest
from tests.helpers.mark import hardware_test from tests.helpers.mark import hardware_test
from tests.helpers.media import generate_synthetic_audio, generate_synthetic_image, generate_synthetic_video from tests.helpers.media import generate_synthetic_audio, generate_synthetic_image, generate_synthetic_video
from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data from tests.helpers.runtime import OmniServerParams, dummy_messages_from_mix_data
from tests.helpers.stage_config import get_deploy_config_path from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
from vllm_omni.platforms import current_omni_platform
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]
# Set VLLM_TEST_PD_MODE=1 to test PD disaggregation (follow-up — deploy overlay not yet migrated). # Set VLLM_TEST_PD_MODE=1 to test PD disaggregation (follow-up — deploy overlay not yet migrated).
_USE_PD = os.environ.get("VLLM_TEST_PD_MODE", "0") == "1" _USE_PD = os.environ.get("VLLM_TEST_PD_MODE", "0") == "1"
_MODEL = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
_CI_DEPLOY = get_deploy_config_path("ci/qwen3_omni_moe.yaml") _CI_DEPLOY = get_deploy_config_path("ci/qwen3_omni_moe.yaml")
# For prefix caching checks against we enable it on the thinker and talker via CLI override def get_chunk_config(config_path: str | None = None):
# and enable prompt token details so that we can determine if any tokens were cached. """Load the qwen3_omni CI deploy yaml with async_chunk modifications for streaming mode."""
# We also explicitly set block size so that we can make sure the cached token counts are a if config_path is None:
# multiple of the block size. config_path = _CI_DEPLOY
BLOCK_SIZE = 16 # TODO: remove this workaround once legacy `stage_args` path is deleted.
# The pipeline (qwen3_omni/pipeline.py) already wires
# thinker2talker_async_chunk / talker2code2wav_async_chunk on stage 0/1,
# so only async_chunk needs flipping. Writing nested `engine_args:` into
# the new-schema overlay trips _parse_stage_deploy's legacy branch and
# drops flat fields (load_format, max_num_seqs, ...).
return modify_stage_config(config_path, updates={"async_chunk": True})
# Platform-specific overrides live inside the new deploy yaml's ``platforms:``
# section, so a single ``_CI_DEPLOY`` path serves CUDA, ROCm, and XPU.
# TODO: re-add VLLM_TEST_PD_MODE branch once the PD-disaggregation deploy
# overlay has been migrated to the new schema (previously used the deleted
# ``qwen3_omni_moe_pd_ci.yaml`` stage-configs file).
if current_omni_platform.is_xpu():
stage_configs = [_CI_DEPLOY]
else: # CUDA + ROCm MI325 share the same deploy config
stage_configs = [get_chunk_config()]
# Create parameter combinations for model and stage config
test_params = [ test_params = [
pytest.param( OmniServerParams(model=model, stage_config_path=stage_config) for model in models for stage_config in stage_configs
OmniServerParams( ]
model=_MODEL, # For prefix caching, we enable it on the thinker (stage 0) via CLI override
stage_config_path=_CI_DEPLOY, # and enable prompt token details so that we can determine if any tokens were cached.
use_stage_cli=True, BLOCK_SIZE = 16
server_args=[ prefix_test_params = [
"--no-async-chunk", OmniServerParams(
"--block-size", model=model,
str(BLOCK_SIZE), stage_config_path=_CI_DEPLOY,
"--stage-overrides", server_args=[
'{"0": {"enable_prefix_caching": true}, "1": {"enable_prefix_caching": true}}', "--block-size",
"--enable-prompt-tokens-details", str(BLOCK_SIZE),
], "--stage-overrides",
), '{"0": {"enable_prefix_caching": true}}',
id="default", "--enable-prompt-tokens-details",
],
) )
for model in models
] ]
@@ -143,14 +166,45 @@ def test_text_to_text_001(omni_server, openai_client) -> None:
openai_client.send_omni_request(request_config, request_num=get_max_batch_size()) openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
def _run_prefix_cache_check(openai_client, request_config: dict): @pytest.mark.advanced_model
"""Make two requests given a request config, and validate that: @pytest.mark.core_model
1. The second request actually had cached tokens @pytest.mark.omni
2. The number of cached tokens is divisible by the block size used in @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
test_params, because currently upstream vLLM does not cache partial @pytest.mark.parametrize("omni_server", prefix_test_params, indirect=True)
blocks. def test_thinker_prefix_caching(omni_server, openai_client, run_level) -> None:
""" """
openai_client.send_omni_request(request_config, request_num=1)[0] Test thinker prefix caching by sending identical requests with an image (i.e.,
a large shared prefix) and verifying that the second request uses cached tokens
& produces the same output with greedy decoding.
NOTE: The reason that we check against logprobs instead of direct text here is that
the outputs may still diverge a bit even though we set the seed and temperature.
This is mostly because the GEMM algorithm may vary based on the input tensors dims.
Because of this, we don't check the logprobs if it's a dummy load, since in that case
the top logprobs will all be very close.
"""
seed = 10
img_res = generate_synthetic_image(224, 224, seed=seed)
image_data_url = f"data:image/jpeg;base64,{img_res['base64']}"
messages = dummy_messages_from_mix_data(
system_prompt=get_system_prompt(),
image_data_url=image_data_url,
content_text=get_prompt("text_image"),
)
top_k = 10
sampling_params = {"seed": seed, "temperature": 0, "max_tokens": 8, "logprobs": top_k}
request_config = {
"model": omni_server.model,
"messages": messages,
"stream": False,
"modalities": ["text"],
"logprobs": True,
"top_logprobs": top_k,
"sampling_params_list": [sampling_params] * 3,
}
uncached_response = openai_client.send_omni_request(request_config, request_num=1)[0]
cached_response = openai_client.send_omni_request(request_config, request_num=1)[0] cached_response = openai_client.send_omni_request(request_config, request_num=1)[0]
# Ensure that we have a prefix cache hit on the second request and that only the last # Ensure that we have a prefix cache hit on the second request and that only the last
@@ -163,66 +217,16 @@ def _run_prefix_cache_check(openai_client, request_config: dict):
assert num_cached_tokens % BLOCK_SIZE == 0 assert num_cached_tokens % BLOCK_SIZE == 0
assert (num_cached_tokens + num_uncached_tokens) == num_prompt_tokens assert (num_cached_tokens + num_uncached_tokens) == num_prompt_tokens
# Ensure that we have logprobs and tokens were generated for both requests
assert uncached_response.logprobs is not None
assert cached_response.logprobs is not None
n_tokens = min(len(uncached_response.logprobs), len(cached_response.logprobs))
assert n_tokens > 0
@pytest.mark.advanced_model if run_level == "advanced_model":
@pytest.mark.core_model # For each token index where both responses have an output, ensure that the greedy token
@pytest.mark.omni # predicted in the uncached case is in the top k logprobs for the cached case
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2) for idx in range(n_tokens):
@pytest.mark.parametrize("omni_server", test_params, indirect=True) greedy_token = uncached_response.logprobs[idx].token
def test_thinker_prefix_caching_text_output(omni_server, openai_client) -> None: cached_top_k = {lp.token for lp in cached_response.logprobs[idx].top_logprobs}
""" assert greedy_token in cached_top_k
Test thinker prefix caching by sending identical requests with an image (i.e.,
a large shared prefix) and verifying that the second request uses cached tokens
& produces the same output with greedy decoding.
NOTE: Checking the output of prefix caching directly can be a bit unstable
due to slight numerical differences as a result of running different scheduled
sequence lengths. As such, for E2E tests on prefix cache, we only check the cached
token count and not the output, since the omni tensor cache has solid unit tests,
and the core prefix cache algorithm is largely tested by upstream vLLM.
"""
img_res = generate_synthetic_image(224, 224)
image_data_url = f"data:image/jpeg;base64,{img_res['base64']}"
messages = dummy_messages_from_mix_data(
system_prompt=get_system_prompt(),
image_data_url=image_data_url,
content_text=get_prompt("text_image"),
)
request_config = {
"model": omni_server.model,
"messages": messages,
"stream": False,
"modalities": ["text"],
}
_run_prefix_cache_check(openai_client, request_config)
@pytest.mark.advanced_model
@pytest.mark.core_model
@pytest.mark.omni
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
@pytest.mark.parametrize("omni_server", test_params, indirect=True)
def test_thinker_prefix_caching_audio_output(omni_server, openai_client) -> None:
"""
Verify that thinker prefix caching does not hang when the request
produces audio output (text + audio modalities). Sends two identical
requests so the second exercises the prefix-cached path through the
full thinker -> talker -> code2wav pipeline.
Regression test for https://github.com/vllm-project/vllm-omni/issues/3510
"""
messages = dummy_messages_from_mix_data(
system_prompt=get_system_prompt(),
content_text=get_prompt(),
)
request_config = {
"model": omni_server.model,
"messages": messages,
"stream": True,
"stream_options": {
"include_usage": True,
},
}
_run_prefix_cache_check(openai_client, request_config)
@@ -1,97 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
E2E Online tests for VoxCPM2 with text input and audio output.
These tests verify the /v1/audio/speech endpoint works correctly with
actual model inference, not mocks. VoxCPM2 zero-shot needs only ``input``
and a placeholder ``voice`` (``default``); no ``task_type`` / ``ref_text``.
"""
import os
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
import pytest
from tests.helpers.mark import hardware_test
from tests.helpers.runtime import OmniServerParams
from tests.helpers.stage_config import get_deploy_config_path
MODEL = "openbmb/VoxCPM2"
DEFAULT_AUDIO_SPEECH_TIMEOUT_S = 300.0
MAX_CONCURRENT = 4
# ~0.5 s of 48 kHz mono PCM_16 in WAV (~48k payload + header).
_MIN_AUDIO_BYTES = 40_000
def get_prompt(prompt_type="text"):
"""English prompt for zero-shot TTS (Whisper-friendly for advanced_model accuracy)."""
prompts = {
"text": "The weather is nice today, perfect for a walk in the park.",
}
return prompts.get(prompt_type, prompts["text"])
tts_server_params = [
pytest.param(
OmniServerParams(
model=MODEL,
stage_config_path=get_deploy_config_path("voxcpm2.yaml"),
server_args=["--trust-remote-code", "--disable-log-stats"],
),
id="voxcpm2",
)
]
@pytest.mark.advanced_model
@pytest.mark.core_model
@pytest.mark.tts
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
def test_text_to_audio_001(omni_server, openai_client) -> None:
"""
Test zero-shot text-to-audio via OpenAI API.
Deploy Setting: voxcpm2.yaml
Input Modal: text
Output Modal: audio
Input Setting: stream=False
Datasets: few requests (max_num_seqs=4)
"""
request_config = {
"model": omni_server.model,
"input": get_prompt(),
"stream": False,
"timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
"response_format": "wav",
"voice": "default",
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
openai_client.send_audio_speech_request(request_config, request_num=MAX_CONCURRENT)
@pytest.mark.advanced_model
@pytest.mark.tts
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
def test_text_to_audio_002(omni_server, openai_client) -> None:
"""
Test zero-shot streaming text-to-audio via OpenAI API.
Deploy Setting: voxcpm2.yaml
Input Modal: text
Output Modal: audio
Input Setting: stream=True
Datasets: single request
"""
request_config = {
"model": omni_server.model,
"input": get_prompt(),
"stream": True,
"timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
"response_format": "wav",
"voice": "default",
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
openai_client.send_audio_speech_request(request_config)
@@ -1,95 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
E2E Online expansion tests for VoxCPM2 (voice clone, PCM, concurrency).
Voice clone uses ``ref_audio`` only (no ``ref_text``). Reference clip is
vendored under tests/assets/qwen3_tts/clone_2.wav — see test_qwen3_tts_base.py.
"""
import os
import pytest
from tests.helpers.mark import hardware_test
from tests.helpers.media import load_test_audio_data_url
from tests.helpers.runtime import OmniServerParams
from tests.helpers.stage_config import get_deploy_config_path
pytestmark = [pytest.mark.full_model, pytest.mark.tts]
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
MODEL = "openbmb/VoxCPM2"
DEFAULT_AUDIO_SPEECH_TIMEOUT_S = 300.0
_MIN_AUDIO_BYTES = 40_000
MAX_CONCURRENT = 4
REF_AUDIO_URL = load_test_audio_data_url("qwen3_tts/clone_2.wav")
def get_prompt(prompt_type="text"):
prompts = {
"text": "The weather is nice today, perfect for a walk in the park.",
}
return prompts.get(prompt_type, prompts["text"])
tts_server_params = [
pytest.param(
OmniServerParams(
model=MODEL,
stage_config_path=get_deploy_config_path("voxcpm2.yaml"),
server_args=["--trust-remote-code", "--disable-log-stats"],
),
id="voxcpm2",
)
]
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
def test_voice_clone_streaming_001(omni_server, openai_client) -> None:
"""
Test voice-clone text-to-audio via OpenAI API.
Deploy Setting: voxcpm2.yaml
Input Modal: text + ref_audio
Output Modal: audio
Input Setting: stream=True
Datasets: few requests (max_num_seqs=4)
"""
request_config = {
"model": omni_server.model,
"input": get_prompt(),
"stream": True,
"timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
"response_format": "wav",
"voice": "default",
"ref_audio": REF_AUDIO_URL,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
openai_client.send_audio_speech_request(request_config, request_num=MAX_CONCURRENT)
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
def test_response_format_001(omni_server, openai_client) -> None:
"""
Test voice-clone non-stream PCM output via OpenAI API.
Deploy Setting: voxcpm2.yaml
Input Modal: text + ref_audio
Output Modal: audio (pcm)
Input Setting: stream=False
Datasets: single request
"""
request_config = {
"model": omni_server.model,
"input": get_prompt(),
"response_format": "pcm",
"stream": False,
"timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
"voice": "default",
"ref_audio": REF_AUDIO_URL,
"min_audio_bytes": _MIN_AUDIO_BYTES,
}
openai_client.send_audio_speech_request(request_config)
-2
View File
@@ -427,7 +427,6 @@ async def _enqueue_add_request(
sampling_params_list=sampling_params_list, sampling_params_list=sampling_params_list,
final_stage_id=final_stage_id, final_stage_id=final_stage_id,
preprocess_ms=0.0, preprocess_ms=0.0,
request_timestamp=time.time(),
enqueue_ts=time.perf_counter(), enqueue_ts=time.perf_counter(),
) )
) )
@@ -985,7 +984,6 @@ async def test_handle_streaming_update_passes_prompt_text_to_stage_pool() -> Non
sampling_params_list=[_sampling_params()], sampling_params_list=[_sampling_params()],
final_stage_id=0, final_stage_id=0,
preprocess_ms=0.0, preprocess_ms=0.0,
request_timestamp=time.time(),
enqueue_ts=time.perf_counter(), enqueue_ts=time.perf_counter(),
) )
) )
@@ -156,7 +156,7 @@ class FakeAsyncOmni:
self.captured_prompt = None self.captured_prompt = None
self._images = images or [Image.new("RGB", (64, 64), color="green")] self._images = images or [Image.new("RGB", (64, 64), color="green")]
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
if sampling_params_list is not None: if sampling_params_list is not None:
self.captured_sampling_params_list = sampling_params_list self.captured_sampling_params_list = sampling_params_list
else: else:
@@ -254,7 +254,7 @@ def async_omni_test_client():
self._images = [Image.new("RGB", (64, 64), color="green")] self._images = [Image.new("RGB", (64, 64), color="green")]
self.od_config = SimpleNamespace(supports_multimodal_inputs=True) self.od_config = SimpleNamespace(supports_multimodal_inputs=True)
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
if sampling_params_list is not None: if sampling_params_list is not None:
self.captured_sampling_params_list = sampling_params_list self.captured_sampling_params_list = sampling_params_list
else: else:
@@ -318,7 +318,7 @@ def async_omni_rgba_test_client():
self._images = [Image.new("RGBA", (64, 64), color=(0, 255, 0, 128))] self._images = [Image.new("RGBA", (64, 64), color=(0, 255, 0, 128))]
self.od_config = SimpleNamespace(supports_multimodal_inputs=True) self.od_config = SimpleNamespace(supports_multimodal_inputs=True)
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
if sampling_params_list is not None: if sampling_params_list is not None:
self.captured_sampling_params_list = sampling_params_list self.captured_sampling_params_list = sampling_params_list
else: else:
@@ -382,7 +382,7 @@ def async_omni_stage_configs_only_client():
self._images = [Image.new("RGB", (64, 64), color="green")] self._images = [Image.new("RGB", (64, 64), color="green")]
self.od_config = SimpleNamespace(supports_multimodal_inputs=True) self.od_config = SimpleNamespace(supports_multimodal_inputs=True)
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
if sampling_params_list is not None: if sampling_params_list is not None:
self.captured_sampling_params_list = sampling_params_list self.captured_sampling_params_list = sampling_params_list
else: else:
@@ -442,7 +442,7 @@ def streaming_image_edit_client():
self.captured_prompt = None self.captured_prompt = None
self.od_config = SimpleNamespace(supports_multimodal_inputs=True) self.od_config = SimpleNamespace(supports_multimodal_inputs=True)
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
self.captured_prompt = prompt self.captured_prompt = prompt
self.captured_sampling_params_list = sampling_params_list or [sampling_params] self.captured_sampling_params_list = sampling_params_list or [sampling_params]
assert self.captured_sampling_params_list[0].output_kind == RequestOutputKind.DELTA assert self.captured_sampling_params_list[0].output_kind == RequestOutputKind.DELTA
@@ -696,7 +696,7 @@ def test_generate_images_async_omni_glm_image_sets_stage0_max_tokens():
self._images = [Image.new("RGB", (64, 64), color="green")] self._images = [Image.new("RGB", (64, 64), color="green")]
self.od_config = SimpleNamespace(supports_multimodal_inputs=True) self.od_config = SimpleNamespace(supports_multimodal_inputs=True)
async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate(self, prompt, request_id, sampling_params=None, sampling_params_list=None):
self.captured_sampling_params_list = ( self.captured_sampling_params_list = (
sampling_params_list if sampling_params_list is not None else [sampling_params] sampling_params_list if sampling_params_list is not None else [sampling_params]
) )
@@ -815,9 +815,7 @@ def test_image_edits_streaming_returns_ar_delta_then_image(streaming_image_edit_
def test_image_edits_streaming_ar_delta_chunks_include_index(streaming_image_edit_client): def test_image_edits_streaming_ar_delta_chunks_include_index(streaming_image_edit_client):
async def generate_multi_output_delta( async def generate_multi_output_delta(prompt, request_id, sampling_params=None, sampling_params_list=None):
prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs
):
yield MockStageResult(stage_id=0, final_output_type="text", texts=["first", "second"]) yield MockStageResult(stage_id=0, final_output_type="text", texts=["first", "second"])
yield MockStageResult( yield MockStageResult(
stage_id=1, stage_id=1,
@@ -845,7 +843,7 @@ def test_image_edits_streaming_ar_delta_chunks_include_index(streaming_image_edi
def test_image_edits_streaming_errors_without_final_image(streaming_image_edit_client): def test_image_edits_streaming_errors_without_final_image(streaming_image_edit_client):
async def generate_without_image(prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate_without_image(prompt, request_id, sampling_params=None, sampling_params_list=None):
yield MockStageResult(stage_id=0, final_output_type="text", text="recap") yield MockStageResult(stage_id=0, final_output_type="text", text="recap")
streaming_image_edit_client.app.state.engine_client.generate = generate_without_image streaming_image_edit_client.app.state.engine_client.generate = generate_without_image
@@ -868,7 +866,7 @@ def test_image_edits_streaming_errors_without_final_image(streaming_image_edit_c
def test_image_edits_streaming_errors_on_empty_final_image(streaming_image_edit_client): def test_image_edits_streaming_errors_on_empty_final_image(streaming_image_edit_client):
async def generate_empty_image(prompt, request_id, sampling_params=None, sampling_params_list=None, **kwargs): async def generate_empty_image(prompt, request_id, sampling_params=None, sampling_params_list=None):
yield MockStageResult(stage_id=0, final_output_type="text", text="recap") yield MockStageResult(stage_id=0, final_output_type="text", text="recap")
yield MockStageResult(stage_id=1, final_output_type="image", images=[]) yield MockStageResult(stage_id=1, final_output_type="image", images=[])
@@ -2,11 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
""" """
E2E online tests for Qwen3-Omni /v1/realtime WebSocket (streaming PCM in, audio out). E2E online tests for Qwen3-Omni /v1/realtime WebSocket (streaming PCM in, audio out).
Three scenarios:
- Ready CI: async_chunk on, smoke only (no send delay, no accuracy check).
- Merge CI: async_chunk on + send delay, full accuracy check.
- Merge CI: async_chunk off, no send delay, full accuracy check.
""" """
from __future__ import annotations from __future__ import annotations
@@ -39,32 +34,18 @@ REALTIME_SYNTH_PHRASE_TEXT = (
"Translate into Chinese: Beijing is the Capital of China. It is the center of culture and politics" "Translate into Chinese: Beijing is the Capital of China. It is the center of culture and politics"
) )
# Simulate realtime upload pacing (``openai_realtime_client.py --send-delay-ms``). # The new-schema CI overlay bakes in async_chunk: False and covers CUDA/ROCm/XPU
SEND_DELAY_MS = 200 # via its ``platforms:`` section, so one path serves all three.
# CI overlay bakes in async_chunk: False and covers CUDA/ROCm/XPU via ``platforms:``.
default_stage_config = get_deploy_config_path("ci/qwen3_omni_moe.yaml") default_stage_config = get_deploy_config_path("ci/qwen3_omni_moe.yaml")
realtime_sync_server_params = [ realtime_server_params = [
pytest.param( pytest.param(
OmniServerParams( OmniServerParams(
model=MODEL, model=MODEL,
stage_config_path=default_stage_config, stage_config_path=default_stage_config,
use_stage_cli=True, use_stage_cli=True,
), ),
id="sync", id="default",
),
]
realtime_async_chunk_server_params = [
pytest.param(
OmniServerParams(
model=MODEL,
stage_config_path=default_stage_config,
use_stage_cli=True,
server_args=["--async-chunk"],
),
id="async_chunk",
), ),
] ]
@@ -99,7 +80,6 @@ async def _run_realtime_audio_roundtrip(
pcm16: bytes, pcm16: bytes,
*, *,
chunk_ms: int = 100, chunk_ms: int = 100,
send_delay_ms: int = 0,
) -> dict: ) -> dict:
uri = f"ws://{host}:{port}/v1/realtime" uri = f"ws://{host}:{port}/v1/realtime"
incremental: list[bytes] = [] incremental: list[bytes] = []
@@ -125,8 +105,6 @@ async def _run_realtime_audio_roundtrip(
} }
) )
) )
if send_delay_ms > 0:
await asyncio.sleep(send_delay_ms / 1000.0)
await ws.send(json.dumps({"type": "input_audio_buffer.commit", "final": True})) await ws.send(json.dumps({"type": "input_audio_buffer.commit", "final": True}))
@@ -178,92 +156,26 @@ async def _run_realtime_audio_roundtrip(
} }
def _synthetic_pcm16_input() -> bytes:
syn = generate_synthetic_audio(
10,
1,
sample_rate=16000,
phrase_text=REALTIME_SYNTH_PHRASE_TEXT,
)
wav_bytes = base64.b64decode(syn["base64"])
return _pcm16_mono_16k_from_wav_bytes(wav_bytes)
def _assert_realtime_smoke(result: dict) -> None:
out_pcm = result["output_pcm"]
assert result["delta_events"] >= 1
assert out_pcm, "No output PCM from response.audio.delta"
assert len(out_pcm) % 2 == 0
assert len(out_pcm) >= 4096, "Output audio unexpectedly small"
assert result["output_sample_rate"] > 0
def _assert_realtime_accuracy(result: dict) -> None:
final_text = (result["transcription_text"] or "").strip()
assert final_text, "Expected non-empty transcription (model text stream)"
wav_out = _wav_bytes_from_pcm16(result["output_pcm"], result["output_sample_rate"])
whisper_text = convert_audio_bytes_to_text(wav_out).strip()
assert whisper_text, "Whisper returned empty string for synthesized output audio"
sim = cosine_similarity_text(whisper_text.lower(), final_text.lower())
assert sim > 0.8, (
f"Output audio transcript should match model text (sim={sim:.3f}): "
f"whisper={whisper_text!r}, model_text={final_text!r}"
)
class TestQwen3OmniRealtimeWebSocket: class TestQwen3OmniRealtimeWebSocket:
@pytest.mark.core_model
@pytest.mark.omni
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
@pytest.mark.parametrize("omni_server", realtime_async_chunk_server_params, indirect=True)
def test_streaming_audio_input_pcm_output_async_chunk_smoke(self, omni_server) -> None:
"""Ready CI: async_chunk on, no send delay, smoke only."""
pcm16 = _synthetic_pcm16_input()
result = asyncio.run(
_run_realtime_audio_roundtrip(
omni_server.host,
omni_server.port,
omni_server.model,
pcm16,
chunk_ms=100,
send_delay_ms=0,
)
)
_assert_realtime_smoke(result)
@pytest.mark.advanced_model @pytest.mark.advanced_model
@pytest.mark.omni @pytest.mark.omni
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2) @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
@pytest.mark.parametrize("omni_server", realtime_async_chunk_server_params, indirect=True) @pytest.mark.parametrize("omni_server", realtime_server_params, indirect=True)
def test_streaming_audio_input_pcm_output_async_chunk(self, omni_server) -> None:
"""Merge CI: async_chunk on, paced upload, full accuracy check."""
pcm16 = _synthetic_pcm16_input()
result = asyncio.run(
_run_realtime_audio_roundtrip(
omni_server.host,
omni_server.port,
omni_server.model,
pcm16,
chunk_ms=100,
send_delay_ms=SEND_DELAY_MS,
)
)
_assert_realtime_smoke(result)
_assert_realtime_accuracy(result)
@pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
@pytest.mark.parametrize("omni_server", realtime_sync_server_params, indirect=True)
def test_streaming_audio_input_pcm_output(self, omni_server) -> None: def test_streaming_audio_input_pcm_output(self, omni_server) -> None:
"""Merge CI: async_chunk off, no send delay, full accuracy check.""" """
pcm16 = _synthetic_pcm16_input() Short streamed 16 kHz mono PCM16 input; expect streamed PCM16 audio deltas and
transcription. Verify Whisper(output audio) aligns with model text (same idea
as multimodal omni e2e). Input speech is synthesized from
``REALTIME_SYNTH_PHRASE_TEXT``.
"""
syn = generate_synthetic_audio(
10,
1,
sample_rate=16000,
phrase_text=REALTIME_SYNTH_PHRASE_TEXT,
)
wav_bytes = base64.b64decode(syn["base64"])
pcm16 = _pcm16_mono_16k_from_wav_bytes(wav_bytes)
result = asyncio.run( result = asyncio.run(
_run_realtime_audio_roundtrip( _run_realtime_audio_roundtrip(
@@ -272,9 +184,25 @@ class TestQwen3OmniRealtimeWebSocket:
omni_server.model, omni_server.model,
pcm16, pcm16,
chunk_ms=100, chunk_ms=100,
send_delay_ms=0,
) )
) )
_assert_realtime_smoke(result) out_pcm = result["output_pcm"]
_assert_realtime_accuracy(result) assert result["delta_events"] >= 1
assert out_pcm, "No output PCM from response.audio.delta"
assert len(out_pcm) % 2 == 0
assert len(out_pcm) >= 4096, "Output audio unexpectedly small"
assert result["output_sample_rate"] > 0
final_text = (result["transcription_text"] or "").strip()
assert final_text, "Expected non-empty transcription (model text stream)"
wav_out = _wav_bytes_from_pcm16(out_pcm, result["output_sample_rate"])
whisper_text = convert_audio_bytes_to_text(wav_out).strip()
assert whisper_text, "Whisper returned empty string for synthesized output audio"
sim = cosine_similarity_text(whisper_text.lower(), final_text.lower())
assert sim > 0.8, (
f"Output audio transcript should match model text (sim={sim:.3f}): "
f"whisper={whisper_text!r}, model_text={final_text!r}"
)
@@ -61,7 +61,7 @@ def test_build_multistage_generation_inputs_applies_stage_specific_overrides(ser
assert engine_prompt["prompt"] == "draw a robot" assert engine_prompt["prompt"] == "draw a robot"
assert engine_prompt["modalities"] == ["img2img"] assert engine_prompt["modalities"] == ["img2img"]
assert engine_prompt["negative_prompt"] == "blurry" assert engine_prompt["negative_prompt"] == "blurry"
assert engine_prompt["mm_processor_kwargs"] == {"target_h": 768, "target_w": 1024, "vae_generator_seed": 0} assert engine_prompt["mm_processor_kwargs"] == {"target_h": 768, "target_w": 1024}
assert engine_prompt["multi_modal_data"]["img2img"].size == (24, 24) assert engine_prompt["multi_modal_data"]["img2img"].size == (24, 24)
assert len(sampling_params_list) == 3 assert len(sampling_params_list) == 3
@@ -33,7 +33,6 @@ from vllm_omni.entrypoints.openai.protocol.audio import (
SpeechBatchItem, SpeechBatchItem,
) )
from vllm_omni.entrypoints.openai.serving_speech import ( from vllm_omni.entrypoints.openai.serving_speech import (
_TTS_LANGUAGES,
OmniOpenAIServingSpeech, OmniOpenAIServingSpeech,
_create_wav_header, _create_wav_header,
) )
@@ -1258,67 +1257,6 @@ class TestTTSMethods:
# Verify speakers are normalized to lowercase # Verify speakers are normalized to lowercase
assert server.supported_speakers == {"ryan", "vivian", "aiden"} assert server.supported_speakers == {"ryan", "vivian", "aiden"}
def test_load_supported_languages_from_config(self, speech_server):
"""Languages/dialects from codec_language_id are loaded title-cased; 'Auto' is added."""
speech_server._tts_model_type = "qwen3_tts"
speech_server.engine_client.model_config = SimpleNamespace(
hf_config=SimpleNamespace(
talker_config=SimpleNamespace(
codec_language_id={"chinese": 2055, "english": 2050, "beijing_dialect": 2074}
)
)
)
assert speech_server._load_supported_languages() == {
"Chinese",
"English",
"Beijing_Dialect",
"Auto",
}
def test_load_supported_languages_from_dict_config(self, speech_server):
"""talker_config provided as a plain dict is handled the same as an object."""
speech_server._tts_model_type = "qwen3_tts"
speech_server.engine_client.model_config = SimpleNamespace(
hf_config=SimpleNamespace(talker_config={"codec_language_id": {"chinese": 2055, "english": 2050}})
)
assert speech_server._load_supported_languages() == {"Chinese", "English", "Auto"}
def test_validate_language_custom_dialect_accepted(self, speech_server):
"""A language present in the model config passes validation, case-insensitively."""
speech_server.supported_languages = {"Chinese", "English", "Beijing_Dialect", "Auto"}
for language in ("Beijing_Dialect", "beijing_dialect", "English", "english", "Auto", "AUTO"):
req = OpenAICreateSpeechRequest(input="Hello", language=language)
result = speech_server._validate_tts_request(req)
assert result is None or "Invalid language" not in result
# Language is normalized to the title-cased config form.
assert req.language == language.title()
def test_validate_language_unknown_rejected(self, speech_server):
"""A language not in the configured set is rejected."""
speech_server.supported_languages = {"Chinese", "English", "Auto"}
for language in ("Klingon", "klingon"):
req = OpenAICreateSpeechRequest(input="Hello", language=language)
assert "Invalid language" in speech_server._validate_tts_request(req)
def test_load_supported_languages_default_when_no_config(self, speech_server):
"""Empty/missing codec_language_id on a Qwen3-TTS model falls back to the default list."""
speech_server._tts_model_type = "qwen3_tts"
speech_server.engine_client.model_config = SimpleNamespace(
hf_config=SimpleNamespace(talker_config=SimpleNamespace(codec_language_id={}))
)
assert speech_server._load_supported_languages() == _TTS_LANGUAGES
def test_load_supported_languages_default_on_config_error(self, speech_server):
"""If the model config cannot be read, fall back to the default list."""
speech_server._tts_model_type = "qwen3_tts"
speech_server.engine_client = SimpleNamespace() # no model_config -> AttributeError
assert speech_server._load_supported_languages() == _TTS_LANGUAGES
def test_load_supported_languages_default_for_non_qwen(self, speech_server):
"""Non-qwen3_tts model types get the default language set."""
speech_server._tts_model_type = None
assert speech_server._load_supported_languages() == _TTS_LANGUAGES
def test_build_tts_params_with_uploaded_voice(self, speech_server, mocker: MockerFixture): def test_build_tts_params_with_uploaded_voice(self, speech_server, mocker: MockerFixture):
"""Test _build_tts_params auto-sets ref_audio for uploaded voices (x_vector only).""" """Test _build_tts_params auto-sets ref_audio for uploaded voices (x_vector only)."""
speech_server.uploaded_speakers = { speech_server.uploaded_speakers = {
@@ -13,14 +13,12 @@ import threading
import time import time
from types import SimpleNamespace from types import SimpleNamespace
import numpy as np
import pytest import pytest
from fastapi import FastAPI from fastapi import FastAPI
from fastapi.testclient import TestClient from fastapi.testclient import TestClient
from PIL import Image from PIL import Image
from pytest_mock import MockerFixture from pytest_mock import MockerFixture
from vllm_omni.diffusion.utils.media_utils import mux_video_audio_bytes
from vllm_omni.entrypoints.openai import api_server from vllm_omni.entrypoints.openai import api_server
from vllm_omni.entrypoints.openai.api_server import router from vllm_omni.entrypoints.openai.api_server import router
from vllm_omni.entrypoints.openai.protocol.videos import ( from vllm_omni.entrypoints.openai.protocol.videos import (
@@ -86,8 +84,7 @@ class BlockingVideoHandler:
if self.stage_configs is None: if self.stage_configs is None:
self.stage_configs = stage_configs self.stage_configs = stage_configs
async def generate_video_bytes(self, request, reference_id, *, reference_image=None, reference_video=None): async def generate_video_bytes(self, request, reference_id, *, reference_image=None):
del request, reference_id, reference_image, reference_video
self.started.set() self.started.set()
try: try:
await asyncio.Future() await asyncio.Future()
@@ -133,30 +130,6 @@ def _make_test_image_data_url(size=(64, 64)) -> str:
return f"data:image/png;base64,{encoded}" return f"data:image/png;base64,{encoded}"
def _make_test_video_bytes(size=(32, 24), num_frames=3) -> bytes:
width, height = size
frames = np.zeros((num_frames, height, width, 3), dtype=np.uint8)
for idx in range(num_frames):
frames[idx, :, :, 0] = idx * 40
frames[idx, :, :, 1] = 128
frames[idx, :, :, 2] = 255 - idx * 40
return mux_video_audio_bytes(frames, fps=8, video_codec_options={"preset": "ultrafast", "threads": "0"})
def _make_test_video_data_url(size=(32, 24), num_frames=3) -> str:
encoded = base64.b64encode(_make_test_video_bytes(size, num_frames)).decode("utf-8")
return f"data:video/mp4;base64,{encoded}"
def _cosmos3_stage_configs():
return [
SimpleNamespace(
stage_type="diffusion",
engine_args=SimpleNamespace(model_class_name="Cosmos3OmniDiffusersPipeline"),
)
]
def _wait_for_status(client: TestClient, video_id: str, status: str, timeout_s: float = 2.0): def _wait_for_status(client: TestClient, video_id: str, status: str, timeout_s: float = 2.0):
deadline = time.time() + timeout_s deadline = time.time() + timeout_s
last_payload = None last_payload = None
@@ -359,104 +332,6 @@ def test_i2v_video_generation_with_image_reference_form(test_client, mocker: Moc
assert input_image.size == (40, 24) assert input_image.size == (40, 24)
def test_v2v_video_generation_form(test_client, mocker: MockerFixture):
video_bytes = _make_test_video_bytes((32, 24), num_frames=3)
mocker.patch(
"vllm_omni.entrypoints.openai.serving_video._encode_video_bytes",
return_value=b"fake-video",
)
response = test_client.post(
"/v1/videos",
data={"prompt": "Continue this motion."},
files={"input_reference": ("input.mp4", video_bytes, "video/mp4")},
)
assert response.status_code == 200
video_id = response.json()["id"]
_wait_for_status(test_client, video_id, VideoGenerationStatus.COMPLETED.value)
engine = test_client.app.state.openai_serving_video._engine_client
prompt = engine.captured_prompt
assert "multi_modal_data" in prompt
assert "video" in prompt["multi_modal_data"]
input_video = prompt["multi_modal_data"]["video"]
assert len(input_video) == 3
assert all(isinstance(frame, Image.Image) for frame in input_video)
assert input_video[0].size == (32, 24)
def test_v2v_video_generation_with_video_reference_form(test_client, mocker: MockerFixture):
mocker.patch(
"vllm_omni.entrypoints.openai.serving_video._encode_video_bytes",
return_value=b"fake-video",
)
response = test_client.post(
"/v1/videos",
data={
"prompt": "Continue this motion.",
"video_reference": json.dumps({"video_url": _make_test_video_data_url((32, 24), 2)}),
},
)
assert response.status_code == 200
video_id = response.json()["id"]
_wait_for_status(test_client, video_id, VideoGenerationStatus.COMPLETED.value)
engine = test_client.app.state.openai_serving_video._engine_client
input_video = engine.captured_prompt["multi_modal_data"]["video"]
assert len(input_video) == 2
assert input_video[0].size == (32, 24)
def test_decode_video_bytes_can_keep_last_frames():
from vllm_omni.entrypoints.openai.video_api_utils import _decode_video_bytes
frames = _decode_video_bytes(
_make_test_video_bytes((32, 24), num_frames=6),
source="input_reference",
max_frames=2,
keep="last",
)
assert len(frames) == 2
red_means = [np.asarray(frame)[:, :, 0].mean() for frame in frames]
assert red_means[0] > 100
assert red_means[1] > red_means[0]
def test_cosmos3_reference_video_limit_uses_v2v_condition_frames():
request = VideoGenerationRequest(
prompt="Continue this motion.",
num_frames=189,
extra_params={"condition_frame_indexes_vision": [0, 2]},
)
spec = api_server._reference_video_decode_spec(request, _cosmos3_stage_configs())
assert spec.max_frames == 9
assert spec.keep == "first"
def test_cosmos3_reference_video_limit_preserves_action_frames():
request = VideoGenerationRequest(
prompt="Predict the action.",
num_frames=17,
extra_params={"action_mode": "inverse_dynamics", "action_chunk_size": 16},
)
assert api_server._reference_video_decode_spec(request, _cosmos3_stage_configs()).max_frames == 17
def test_cosmos3_reference_video_limit_caps_condition_frames_to_output_frames():
request = VideoGenerationRequest(
prompt="Continue this motion.",
num_frames=5,
extra_params={"condition_frame_indexes_vision": [0, 20]},
)
assert api_server._reference_video_decode_spec(request, _cosmos3_stage_configs()).max_frames == 5
def test_seconds_defaults_fps_and_frames(test_client, mocker: MockerFixture): def test_seconds_defaults_fps_and_frames(test_client, mocker: MockerFixture):
fps_values = [] fps_values = []
@@ -524,8 +399,6 @@ def test_sampling_params_pass_through(test_client, mocker: MockerFixture):
"true_cfg_scale": "4.0", "true_cfg_scale": "4.0",
"boundary_ratio": "0.7", "boundary_ratio": "0.7",
"flow_shift": "0.25", "flow_shift": "0.25",
"generate_sound": "true",
"sound_duration": "2.5",
}, },
) )
@@ -540,8 +413,6 @@ def test_sampling_params_pass_through(test_client, mocker: MockerFixture):
assert captured.true_cfg_scale == 4.0 assert captured.true_cfg_scale == 4.0
assert captured.boundary_ratio == 0.7 assert captured.boundary_ratio == 0.7
assert captured.extra_args["flow_shift"] == 0.25 assert captured.extra_args["flow_shift"] == 0.25
assert captured.extra_args["generate_sound"] is True
assert captured.extra_args["sound_duration"] == 2.5
def test_frame_interpolation_params_pass_to_diffusion_sampling_params(test_client, mocker: MockerFixture): def test_frame_interpolation_params_pass_to_diffusion_sampling_params(test_client, mocker: MockerFixture):
@@ -752,115 +623,6 @@ def test_video_job_persists_profiler_metadata(test_client, mocker: MockerFixture
assert completed["stage_durations"] == {"diffuse": 2.5, "vae.decode": 0.3} assert completed["stage_durations"] == {"diffuse": 2.5, "vae.decode": 0.3}
assert completed["peak_memory_mb"] == 4096.5 assert completed["peak_memory_mb"] == 4096.5
assert completed["action"] is None
def test_video_generation_response_exposes_action_payload(mocker: MockerFixture):
engine = FakeAsyncOmni()
handler = OmniOpenAIServingVideo.for_diffusion(
diffusion_engine=engine,
model_name="Cosmos3-8B-UVA",
)
async def _generate(prompt, request_id, sampling_params_list):
del prompt, request_id, sampling_params_list
import numpy as np
yield MockVideoResult(
[object()],
custom_output={
"action": np.array([[[1.5, 2.5], [3.5, 4.5]]], dtype=np.float32),
"raw_action_dim": 2,
"action_mode": "policy",
"domain_id": 7,
},
)
engine.generate = _generate
mocker.patch(
"vllm_omni.entrypoints.openai.serving_video.encode_video_base64",
return_value="encoded-video",
)
response = asyncio.run(
handler.generate_videos(
VideoGenerationRequest(prompt="predict actions"),
"action-json",
)
)
action = response.data[0].action
assert action is not None
assert action.data == [[1.5, 2.5], [3.5, 4.5]]
assert action.shape == [2, 2]
assert action.dtype == "float32"
assert action.raw_action_dim == 2
assert action.action_mode == "policy"
assert action.domain_id == 7
assert response.model_dump(mode="json")["data"][0]["action"]["data"] == [[1.5, 2.5], [3.5, 4.5]]
def test_video_job_persists_action_metadata(test_client, mocker: MockerFixture):
engine = test_client.app.state.openai_serving_video._engine_client
async def _generate(prompt, request_id, sampling_params_list):
import numpy as np
engine.captured_prompt = prompt
engine.captured_sampling_params_list = sampling_params_list
yield MockVideoResult(
[object()],
custom_output={
"action": np.array([[[1.0, 2.0], [3.0, 4.0]]], dtype=np.float32),
"raw_action_dim": 2,
"action_mode": "policy",
"domain_id": 7,
},
)
engine.generate = _generate
mocker.patch(
"vllm_omni.entrypoints.openai.serving_video._encode_video_bytes",
return_value=b"fake-video",
)
response = test_client.post("/v1/videos", data={"prompt": "profile me"})
assert response.status_code == 200
video_id = response.json()["id"]
completed = _wait_for_status(test_client, video_id, VideoGenerationStatus.COMPLETED.value)
expected_action = {
"data": [[1.0, 2.0], [3.0, 4.0]],
"shape": [2, 2],
"dtype": "float32",
"raw_action_dim": 2,
"action_mode": "policy",
"domain_id": 7,
}
assert completed["action"] == expected_action
listed = test_client.get("/v1/videos").json()
assert listed["data"][0]["action"] == expected_action
def test_action_extraction_accepts_unbatched_action():
import numpy as np
result = MockVideoResult(
[object()],
custom_output={
"action": np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32),
"raw_action_dim": 2,
"action_mode": "policy",
"domain_id": 7,
},
)
actions = OmniOpenAIServingVideo._extract_action_outputs(result, expected_count=1)
assert actions[0] is not None
assert actions[0].data == [[1.0, 2.0], [3.0, 4.0]]
assert actions[0].shape == [2, 2]
def test_missing_handler_returns_503(): def test_missing_handler_returns_503():
@@ -919,20 +681,7 @@ def test_rejects_input_reference_and_image_reference_together(test_client):
files={"input_reference": ("input.png", _make_test_image_bytes(), "image/png")}, files={"input_reference": ("input.png", _make_test_image_bytes(), "image/png")},
) )
assert response.status_code == 400 assert response.status_code == 400
assert "only one of input_reference, image_reference, or video_reference" in response.json()["detail"].lower() assert "either input_reference or image_reference" in response.json()["detail"].lower()
def test_rejects_image_reference_and_video_reference_together(test_client):
response = test_client.post(
"/v1/videos",
data={
"prompt": "bad refs",
"image_reference": '{"image_url": "https://example.com/cat.png"}',
"video_reference": '{"video_url": "https://example.com/cat.mp4"}',
},
)
assert response.status_code == 400
assert "only one of input_reference, image_reference, or video_reference" in response.json()["detail"].lower()
def test_invalid_seconds_returns_422(test_client): def test_invalid_seconds_returns_422(test_client):
@@ -994,18 +743,6 @@ def test_unsupported_image_reference_file_id_returns_400(test_client):
assert response.json()["detail"] == "Invalid image_reference: file_id is not supported yet." assert response.json()["detail"] == "Invalid image_reference: file_id is not supported yet."
def test_unsupported_video_reference_file_id_returns_400(test_client):
response = test_client.post(
"/v1/videos",
data={
"prompt": "unsupported ref",
"video_reference": '{"file_id": "file-123"}',
},
)
assert response.status_code == 400
assert response.json()["detail"] == "Invalid video_reference: file_id is not supported yet."
def test_invalid_uploaded_input_reference_returns_400(test_client): def test_invalid_uploaded_input_reference_returns_400(test_client):
response = test_client.post( response = test_client.post(
"/v1/videos", "/v1/videos",
@@ -1013,15 +750,12 @@ def test_invalid_uploaded_input_reference_returns_400(test_client):
files={"input_reference": ("input.png", b"not-an-image", "image/png")}, files={"input_reference": ("input.png", b"not-an-image", "image/png")},
) )
assert response.status_code == 400 assert response.status_code == 400
assert response.json()["detail"] == "Invalid input_reference: provided content is not a valid image or video." assert response.json()["detail"] == "Invalid input_reference: provided content is not a valid image."
def test_video_request_validation(): def test_video_request_validation():
req = VideoGenerationRequest(prompt="test") req = VideoGenerationRequest(prompt="test")
assert req.prompt == "test" assert req.prompt == "test"
assert req.generate_sound is False
assert req.sound_duration is None
assert VideoGenerationRequest(prompt="test", generate_sound=True, sound_duration=1.5).generate_sound is True
with pytest.raises(ValueError): with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", size="invalid") VideoGenerationRequest(prompt="test", size="invalid")
@@ -1030,14 +764,10 @@ def test_video_request_validation():
with pytest.raises(ValueError): with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", image_reference={"file_id": "file-1", "image_url": "https://example.com"}) VideoGenerationRequest(prompt="test", image_reference={"file_id": "file-1", "image_url": "https://example.com"})
with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", video_reference={"file_id": "file-1", "video_url": "https://example.com"})
with pytest.raises(ValueError): with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", frame_interpolation_exp=0) VideoGenerationRequest(prompt="test", frame_interpolation_exp=0)
with pytest.raises(ValueError): with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", frame_interpolation_scale=0) VideoGenerationRequest(prompt="test", frame_interpolation_scale=0)
with pytest.raises(ValueError):
VideoGenerationRequest(prompt="test", sound_duration=0)
def test_list_videos_supports_order_after_and_limit(test_client, mocker: MockerFixture): def test_list_videos_supports_order_after_and_limit(test_client, mocker: MockerFixture):
@@ -1299,31 +1029,6 @@ def test_sample_solver_forwarded_via_extra_params(test_client, mocker: MockerFix
assert captured.extra_args["sample_solver"] == "euler" assert captured.extra_args["sample_solver"] == "euler"
def test_extra_params_allows_inline_action(test_client, mocker: MockerFixture):
"""Inline ``action`` data is accepted and forwarded verbatim to
``extra_args`` (the supported way to pass forward-dynamics actions)."""
mocker.patch(
"vllm_omni.entrypoints.openai.serving_video._encode_video_bytes",
return_value=b"fake-video",
)
action = [[0.1, 0.2], [0.3, 0.4]]
response = test_client.post(
"/v1/videos",
data={
"prompt": "forward dynamics inline",
"extra_params": json.dumps({"action_mode": "forward_dynamics", "action": action}),
},
)
assert response.status_code == 200
video_id = response.json()["id"]
_wait_for_status(test_client, video_id, VideoGenerationStatus.COMPLETED.value)
engine = test_client.app.state.openai_serving_video._engine_client
captured = engine.captured_sampling_params_list[0]
assert captured.extra_args["action"] == action
assert captured.extra_args["action_mode"] == "forward_dynamics"
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
# Sync endpoint tests (POST /v1/videos/sync) # Sync endpoint tests (POST /v1/videos/sync)
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
@@ -1417,23 +1122,6 @@ def test_sync_i2v_with_image_reference(test_client, mocker: MockerFixture):
assert response.content == b"ref-video" assert response.content == b"ref-video"
def test_sync_v2v_returns_video_bytes(test_client, mocker: MockerFixture):
video_bytes = _make_test_video_bytes((32, 24), num_frames=3)
_mock_encode_video_bytes(mocker, b"v2v-video-data")
response = test_client.post(
"/v1/videos/sync",
data={"prompt": "Continue this motion."},
files={"input_reference": ("input.mp4", video_bytes, "video/mp4")},
)
assert response.status_code == 200
assert response.content == b"v2v-video-data"
engine = test_client.app.state.openai_serving_video._engine_client
input_video = engine.captured_prompt["multi_modal_data"]["video"]
assert len(input_video) == 3
assert input_video[0].size == (32, 24)
def test_sync_missing_handler_returns_503(): def test_sync_missing_handler_returns_503():
app = FastAPI() app = FastAPI()
app.include_router(router) app.include_router(router)
@@ -1466,7 +1154,7 @@ def test_sync_rejects_both_references(test_client):
files={"input_reference": ("input.png", _make_test_image_bytes(), "image/png")}, files={"input_reference": ("input.png", _make_test_image_bytes(), "image/png")},
) )
assert response.status_code == 400 assert response.status_code == 400
assert "only one of input_reference, image_reference, or video_reference" in response.json()["detail"].lower() assert "either input_reference or image_reference" in response.json()["detail"].lower()
def test_sync_generation_error_returns_500(test_client, mocker: MockerFixture): def test_sync_generation_error_returns_500(test_client, mocker: MockerFixture):
+1 -2
View File
@@ -16,7 +16,7 @@ OMNI_MODEL = "Qwen/Qwen2.5-Omni-7B"
OMNI_STAGE_CONFIG = get_deploy_config_path("ci/qwen2_5_omni_thinker_only.yaml") OMNI_STAGE_CONFIG = get_deploy_config_path("ci/qwen2_5_omni_thinker_only.yaml")
async def _noop(*args, **kw): async def _noop(**kw):
pass pass
@@ -61,7 +61,6 @@ def get_async_omni_instance(fake_add_request=_noop, fake_abort_request=_noop) ->
omni._final_output_handler = lambda: None omni._final_output_handler = lambda: None
omni.resolve_sampling_params_list = lambda params, allow_delta_coercion: params omni.resolve_sampling_params_list = lambda params, allow_delta_coercion: params
omni._compute_final_stage_id = lambda output_modalities: 0 omni._compute_final_stage_id = lambda output_modalities: 0
omni._compute_final_output_stage_ids = lambda output_modalities: [0]
omni._process_orchestrator_results = fake_process_results omni._process_orchestrator_results = fake_process_results
omni._log_summary_and_cleanup = lambda request_id: omni.request_states.pop(request_id, None) omni._log_summary_and_cleanup = lambda request_id: omni.request_states.pop(request_id, None)
return omni return omni
@@ -413,19 +413,6 @@ def test_openai_serving_models_can_consume_async_omni_compat_attrs():
def test_get_diffusion_od_config_returns_diffusion_stage_config(): def test_get_diffusion_od_config_returns_diffusion_stage_config():
diffusion_od_config = object()
omni = object.__new__(AsyncOmni)
omni.engine = SimpleNamespace(
stage_clients=[
SimpleNamespace(stage_type="llm"),
SimpleNamespace(stage_type="diffusion", od_config=diffusion_od_config),
]
)
assert omni.get_diffusion_od_config() is diffusion_od_config
def test_get_diffusion_od_config_falls_back_to_inner_engine():
diffusion_od_config = object() diffusion_od_config = object()
omni = object.__new__(AsyncOmni) omni = object.__new__(AsyncOmni)
omni.engine = SimpleNamespace( omni.engine = SimpleNamespace(
Loaded 100 of 240 files, more files were not shown because too many files have changed in this diff. Show more