Commit Graph
312 Commits
Author SHA1 Message Date
tzhouamandClaude Opus 5 7869a8dc0d Add cpu_int4, an asymmetric 4-bit weight path for vLLM's CPU backend
Single-stream CPU decode reads every weight to emit one token, so bits per
weight set the token rate. This registers a W4A16 group-wise method
out-of-tree through register_quantization_config -- vLLM itself is untouched
-- using PyTorch's _weight_int4pack_mm_for_cpu tinygemm kernel.

Two paths over one set of weights, because that kernel dequantizes once per
row of activations and so is linear in the batch (12.9x slower than bf16 at
512 rows). Single-row steps take tinygemm; wider batches unpack the weight to
bf16 once and use the ordinary oneDNN/AMX GEMM. Shipping only the first
measured pp512 = 557 tok/s against bf16's 4455; with both, 3589.

The batch threshold is 2 rather than the ~8 rows where the unpack path starts
winning on speed, because running tinygemm on a multi-row batch and then
vLLM's cpu_attention_with_kv_cache in the same forward dies with SIGILL
inside the attention kernel -- an AMX instruction with no valid tile config
loaded. The same prompt is fine in bf16 and fine once every multi-row GEMM
takes the unpack path.

Positioning, measured on a Xeon 8480C with 32 pinned cores and a warm cache
(analysis/spark_cpu_int4.md): vLLM's own GPTQ W4A8 AMX path is faster --
88.7 tok/s decode and 4355 prefill against 85.5 and 3589 here -- but its
config rejects sym=False at 4 bits, so it can only be fed a symmetric grid.
This one is asymmetric and costs +12.5% perplexity against that path's
+27.2%, and needs no AMX. llama.cpp Q4_K_M, for reference, is 98.0 tok/s
decode and 1356 prefill at +7.6% and 5.15 bits per weight.

Expanding a quantized checkpoint back to bf16 offline and scoring it
reproduces the quantized run to within 0.2% perplexity, so both GEMM paths,
the nibble packing and the quantized tied embedding are faithful.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GXhbAfkSNRaeqZrus5Z2fx
2026-09-12 03:48:04 +08:00
tzhouamandClaude Opus 5 e4a565b3df [Edge][Model] Spark-X2.5 (Spark2_5ForCausalLM) support
Spark-X2.5-1.7B is an Apache-2.0 on-device model built for agentic tool
use. vLLM 0.28 has no implementation for it, so it could not be served at
all; llama.cpp already ships one, which made it the only edge engine able
to run the model.

Three traits separate it from a Llama-style decoder:

* Hybrid attention -- layer_types interleaves three 512-token sliding
  layers with one full-attention layer, which is what makes the 1M-token
  context affordable on a phone.
* Per-layer-type RoPE -- sliding layers rotate all 256 head dims with
  theta 1e4, full layers only the first quarter with theta 5e6.
* A head-wise attention output gate -- g_proj reads the same post-norm
  input as QKV and scales each head's attention output by a sigmoid.

The checkpoint also stores QKV pre-fused as a single q_k_v_proj, so
load_weights splits it before vLLM's sharded QKVParallelLinear sees it.

Verified against the official remote-code implementation on transformers
4.57.1 (the version config.json targets; it does not import on the 5.x in
our venvs). In fp32 all three scenarios match exactly -- 32/32 greedy
tokens with a max top-20 logprob difference of 2.2e-05. In bf16 the short
and the 2465-token long-context prompts still match 32/32, which is what
exercises the sliding layers; the tool-calling prompt diverges at token 23
on a near-tie, and fp32 confirms that is rounding rather than a defect.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
2026-09-11 13:47:02 +08:00
tzhouamandClaude Fable 5.1 37c0ea9a6b [Edge][WP8] Per-step overhead attribution, edge scheduling helpers, CPU thread sweep
- vllm_omni/edge/step_stats.py: process-local counters enabled by
  VLLM_OMNI_STEP_STATS_DIR (scheduler schedule/update_from_output, engine-core
  step, chunk hop enqueue->put / put->get / get, orchestrator dispatch), dumped
  per process and merged into an attribution table (share of the 80 ms frame).
  Hooks: stage_engine_core_proc (instrument after construction, dump at exit),
  chunk_transfer_adapter (put/get timing, put->get via the shm lock-file mtime),
  orchestrator dispatch loop. Bench: stream_latency_bench.py --step-stats.
- vllm_omni/edge/scheduling.py: playback slack, prefill-throttle decision and
  the no-preemption KV admission rule (exact bytes/token from the talker KV
  geometry); adapt.derive_overrides now sizes max_num_seqs with it and
  auto_profile passes the geometry read from the model config.
- Engine re-reads parallel_stage_init after apply_deploy_profile so the edge
  profile's orchestrator default takes effect (was read before the profile).
- benchmarks/tts/cpu_thread_sweep.py: CPU-only 8/16/32-thread sweep (+inductor)
  with per-step attribution.
- Tests: tests/core/test_edge_scheduling.py, tests/edge/test_step_stats.py,
  tests/benchmarks/test_cpu_thread_sweep.py, profile default test.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-09 14:55:55 +08:00
tzhouamandClaude Fable 5.1 8e912db6c8 [Edge][WP6] Slim the AR/TTS import path and add an edge dependency set
* diffusers is no longer imported by vllm_omni.diffusion.data (annotation only);
  vllm_omni.diffusion.io_support (diffusion registry + distributed state) is
  imported lazily by config_factory and async_omni_engine, so constructing a
  Qwen3-TTS pipeline config / AsyncOmni no longer loads diffusers,
  x_transformers or whisper (guarded by tests/config/test_import_guard.py).
* requirements/common.txt = core.txt (AR/speech runtime) + media.txt
  (diffusion/video/extra audio); VLLM_OMNI_EDGE_BUILD=1 makes setup.py use
  requirements/<device>-edge.txt (core only); `edge` extra adds websockets and
  sounddevice for on-device clients.

Measured on this host (CUDA venv): `import vllm` 7.0-7.9 s, `import vllm_omni`
adds ~20 s of which ~17 s is transformers' own import (transformers.utils.import_utils)
pulled in by vllm_omni.patch / vllm_omni.version; that floor is not
vLLM-Omni's and is unchanged. Guard test: no diffusion module loaded on the TTS path.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-09 11:14:50 +08:00
tzhouamandClaude Fable 5.1 073d31b9cd [Edge][WP2] Add initialization timeline and cold-start profiler
* vllm_omni/engine/init_timeline.py: InitTimeline phase collector; engine
  phases (stages_init, g3_device_lock, g2_spawn/ready) and worker phases
  (proc_entry, engine_core_init, init_device, load_model/load_weights,
  talker_mtp_init, attn_prewarm, cudagraph_capture, talker_mtp_capture,
  code2wav_cudagraph, predictor_setup_compile/warmup/cudagraph) recorded via
  worker_phase/timed_phase; workers append JSONL under
  VLLM_OMNI_INIT_TIMELINE_DIR (inherited at spawn) and the engine merges them.
  Enabled by VLLM_OMNI_INIT_TIMELINE=<json> or log_stats; table logged at INFO,
  partial table on orchestrator startup timeout, suggested timeouts logged.
* omni_base.py: named deploy profiles default to stage_init_timeout=900 /
  init_timeout=1500 unless given explicitly.
* gpu_model_runner._init_talker_mtp tolerates max_cudagraph_capture_size=None;
  gpu_ar_model_runner.shutdown skips CUDA graph teardown on CPU.
* benchmarks/tts/init_profile.py: runs stream_latency_bench --init-only per
  configuration (default, cold_cache, eager_stage1, eager_both,
  parallel_stage_init, edge, ...) and tabulates phase durations.
* stream_latency_bench: --parallel-stage-init / --deploy-profile flags; the
  resource poller no longer shadows Thread._stop and skips GPU sampling on CPU.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-09 11:01:39 +08:00
tzhouamandClaude Fable 5.1 b22dcc59f7 [Edge][WP1] Add edge deploy profile and deploy_profile selector
* vllm_omni/deploy/edge/qwen3_tts.yaml: single-device Qwen3-TTS profile with a
  ramped codec chunk schedule (2->4->8->16->25 frames) that removes the ~80 ms
  playback stall of the default 1->25 schedule, decode batch 1, max_num_seqs 4.
* deploy_profile="edge" (Omni/AsyncOmni kwarg, --deploy-profile CLI, headless
  serve) resolves vllm_omni/deploy/<profile>/<model_type>.yaml; mutually
  exclusive with deploy_config; explicit profiles never fall back silently.
* tests/helpers/stage_config.py: modify_stage_config anchors a relative
  base_config before writing the temp YAML so overlay-derived configs resolve.
* async_omni_engine.py also carries the init-timeline hooks used by WP2.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-09 10:27:33 +08:00
BruceLoveDecimal 5cd925631c [Feature][LingBot World]stepwise execution (#6844)
Signed-off-by: BruceLoveDecimal <liuqihao970610@gmail.com>
Signed-off-by: liuqihao <liuqihao970610@gmail.com>
2026-09-07 13:38:23 +08:00
f26f9c3a7f [Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)
Signed-off-by: asukaqaq-s <1311722138@qq.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: asukaqaq-s <1311722138@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-09-06 05:41:54 +08:00
SYLAR ea3cf99894 [diffusion][feature] Add component-selective offload policies (#5929)
Signed-off-by: lishunyang12 <lishunyang12@163.com>
2026-09-06 05:40:09 +08:00
Zhou TaichangandClaude Fable 5 142151fef5 Parallel stage initialization (admission + SH/EX device locks) (Worker refactor 1/N) (#5224)
Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk>
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-09-05 22:22:41 +08:00
5d7fdf9350 [Core] Graduate MiniCPM-o 4.5 and PersonaPlex full-duplex serving out of experimental (#6196)
Signed-off-by: chickeyton <ngton2014@gmail.com>
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
2026-09-05 08:20:21 +00:00
SYLARandhsliuustc0106 7493cbd3c2 [Bugfix] Fix Qwen3-Omni MoE backend selection (#7019)
Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>
Co-authored-by: hsliuustc0106 <liuhongsheng4@huawei.com>
2026-09-04 21:16:58 +08:00
Yuekai ZhangandClaude Fable 5 4179b0cae1 [Perf][NemotronVoiceChat] Fast defaults for all three deploy profiles: 2.8x offline, realtime full-duplex audio (#6354)
Signed-off-by: Yuekai Zhang <zhangyuekai@foxmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-09-03 18:49:19 -07:00
Arifuzzaman Joy 3cddda9781 [Bugfix] Prevent build_engine_args_dict from mutating stage_config.engine_args (#6783)
Signed-off-by: Arifuzzamanjoy <s1710374103@ru.ac.bd>
2026-09-01 16:45:28 +00:00
Andy Zhou 2298359f67 [Bugfix][Diffusion] Preserve stage attention backend (#6645)
Signed-off-by: AndyZhou952 <jzhoubc@connect.ust.hk>
2026-09-01 20:18:44 +08:00
Yuxuan RhoandHongsheng Liu 7992c2213e [Config] Remove diffusion_batch_size and use max_num_seqs instead. (#6484)
Signed-off-by: LyxWxj <yuxuanlou621@foxmail.com>
Signed-off-by: Yuxuan Rho <104440472+LyxWxj@users.noreply.github.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-09-01 01:02:54 +00:00
TJian ded8934626 [AMD] [CI] Fix tests for v0.28.0 (#6830)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
2026-08-31 09:10:27 +08:00
Gao Han f7f48929db [Bugfix][Diffusion] Restore inline single-stage execution (#6813)
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
2026-08-30 21:40:47 +08:00
Yueqian Lin a6b559cf54 [Bugfix][MiniMax-Music3] Resolve stage subdirs against a real snapshot (#6640)
Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
2026-08-30 06:51:24 +00:00
baonudesifeizhai 8be601ed7f [Diffusion] Add native SymmMem Fast Ulysses transport (#6340)
Signed-off-by: roG0d <baonudesifeizhai@gmail.com>
2026-08-30 04:14:21 +00:00
21dd4b33ac [Bugfix][MiniCPM-o] Stabilize native duplex streaming and Stage-1 handoffs (#6529)
Signed-off-by: Sy03 <1370724210@qq.com>
Co-authored-by: Nathan Price <nathan@abridge.com>
Co-authored-by: Eric(Qun) <186020953+0xtoward@users.noreply.github.com>
2026-08-29 13:05:24 -07:00
Sy03 3adee400a0 [Bugfix][Qwen3-TTS] Handle codec generations that exhaust their token budget (#6728)
Signed-off-by: Sy03 <1370724210@qq.com>
2026-08-29 08:59:05 -07:00
bb5bf8fd0a [Bugfix] Fix colocate-async sleep admission race and deliver AR abort tokens (#6367)
Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>
Signed-off-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Samit <285365963@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Samit <285365963@qq.com>
2026-08-29 09:11:39 +00:00
Bob Tian d934a5f7b8 [Bugfix] Report engine-queued requests as waiting in vllm_omni:num_requests gauges (#6549)
Signed-off-by: Bob Tian <bobzetian@google.com>
2026-08-29 03:58:26 +00:00
d94061201d [Bugfix][Engine] Make abort idempotent during shutdown (#6327)
Signed-off-by: Allen Wu <allenwu2795@gmail.com>
Co-authored-by: TRAE CLI <noreply@bytedance.com>
Co-authored-by: TRAE CLI <traecli@bytedance.com>
2026-08-28 19:51:25 -07:00
Sy03 009b992c5b [Bugfix][MiniCPM-o] Isolate duplex handoff failures and barge-in (#6170)
Signed-off-by: Sy03 <1370724210@qq.com>
2026-08-28 15:46:41 +08:00
b6a4baa659 [Feature] Support text encoder Disaggregation for H3 (#5885)
Signed-off-by: Canlin Guo <canlinguosdu@gmail.com>
Signed-off-by: lishunyang12 <lishunyang12@163.com>
Signed-off-by: yuanwu <yuan.wu@intel.com>
Co-authored-by: lishunyang12 <lishunyang12@163.com>
Co-authored-by: yuanwu <yuan.wu@intel.com>
2026-08-28 14:27:05 +08:00
Mu GuanLinandHongsheng Liu 5e4daaf197 [Diffusion] Add LTX-2.5 Diffusion VAE decoder support (#6189)
Signed-off-by: mglyn <1203789601@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-27 23:54:40 +08:00
NATURE 54c1fc64b7 [Bugfix] Gate async-chunk segment resume by session mode (#6680)
Signed-off-by: natureofnature <wzliu@connect.hku.hk>
2026-08-27 18:52:09 +08:00
Zhou TaichangandClaude Opus 5 651b03225b [Rebase] Rebase to vllm 0.28.0 (#6606)
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-27 06:12:55 +00:00
c2723979c0 [Bugfix] Validate per-stage device layout before spawning workers (#5003) (#5742)
Signed-off-by: ChoHee15 <cc5281@126.com>
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
2026-08-26 19:25:22 +00:00
Yueqian LinandYueqian Lin 8cfede4309 [Perf][Engine] Event-driven orchestration loop (opt-in) — S1 of #4855 (#5221)
Signed-off-by: Yueqian Lin <gao@yueqian.us>
Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
Co-authored-by: Yueqian Lin <gao@yueqian.us>
2026-08-26 22:15:25 +08:00
psv666andamy-why-3459 5ec576ab4c [Bugfix] Fix missing Realtime audio completion (#6564)
Signed-off-by: psv666 <2693925048@qq.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
2026-08-26 18:27:25 +08:00
7ae15c74c8 text and image metric (#6150)
Signed-off-by: zhuhh97 <zhuhonghan@huawei.com>
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
2026-08-26 02:54:45 +00:00
Alex Brooks 0eb47b76d0 [Bugfix] Handle DoS Overflow Cases (#6598)
Signed-off-by: Alex Brooks <albrooks@redhat.com>
2026-08-25 16:09:03 +00:00
Hongsheng Liuandhsliuustc0106 c608910d20 [Diffusion] Register HWR mmap for direct DLO H2D (#6591)
Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
2026-08-25 23:30:11 +08:00
NumberWan bc8ea489ef [BugFix] Qwen-image performance regressed - Avoid mapping diffusion_batch_size onto scheduler max_num_seqs (#6525)
Signed-off-by: NumberWan <wantszkin2003@gmail.com>
2026-08-25 03:36:11 -07:00
wy17003andHongsheng Liu b454e34966 [Refactor][2/N Scheduler]Replace _FULL_PAYLOAD_INPUT_STAGES with a resolved stage transport capability. (#6149)
Signed-off-by: wy17003 <yaow17003@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-25 15:24:59 +08:00
f0808c9793 [Bugfix] Fix remote replica membership lifecycle races (#5277)
Signed-off-by: hbhflw2000 <417911774@qq.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Samit <285365963@qq.com>
2026-08-25 10:34:59 +08:00
Hongsheng Liuandhsliuustc0106 8360346a66 [Diffusion] Integrate no-AllGather DLO with Host Weight Runtime (#6486)
Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Signed-off-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
Co-authored-by: hsliuustc0106 <hsliuustc0106@users.noreply.github.com>
2026-08-25 08:03:53 +08:00
Nguyen Kha Nhat LongandCursor d15b7ca3e1 [Feature]: Support pause / resume and sleep / wake for AR stages in AsyncOmni (#6084)
Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-24 16:21:32 +08:00
bjf-frzandHongsheng Liu 072bfc02dd [FEAT]Add FastVideo VSA backend for Wan2.2 (#4820)
Signed-off-by: bjf-frz <frz123db@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-24 14:48:31 +08:00
b39f599b09 [Core] Add UniProc diffusion executor for single-GPU (#6308)
Signed-off-by: Rahul Steiger <rsteiger@nvidia.com>
Signed-off-by: Rahul Steiger <rsteiger@aws-cmh-slurm-1-vscode-04.cm.cluster>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Co-authored-by: Rahul Steiger <rsteiger@aws-cmh-slurm-1-vscode-04.cm.cluster>
2026-08-22 21:58:21 +08:00
Yash Jain c6e38860e9 [Bugfix][Core] Fix replica device split when a stage omits devices (#5445)
Signed-off-by: Yash Jain <yashjainjain1704@gmail.com>
2026-08-21 00:44:03 +08:00
Inesh Reddy Chappidi cd3867c373 [Model] Suppress silence codec tokens for the first N Qwen3-TTS decod… (#5048)
Signed-off-by: Inesh Reddy <ineshreddy249@gmail.com>
2026-08-20 09:22:37 -07:00
amy-why-3459andCursor 9c61d5b67b [Bugfix][MiniCPM-o] Stop Talker mid-utterance truncation and load cod… (#6346)
Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-20 11:50:07 +08:00
bjf-frzandHongsheng Liu 14bd8f8772 [REFACT]Refactor diffusion parallel state (#5531)
Signed-off-by: bjf-frz <frz123db@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-19 11:23:10 +08:00
Wang fuyinandHongsheng Liu e1c5f59aab [Config] Reuse vLLM configs in VllmOmniConfig (#6050)
Signed-off-by: Acerak01-fy <wfy2003324@163.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-18 21:27:58 +08:00
Yueqian Lin b283215c28 [Model] Add MiniMax Music 3 text-to-music (#6186)
Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
2026-08-18 04:33:13 +00:00
Yueqian LinandHongsheng Liu 611c083587 [BugFix][TTS] R1.1-R1.4: turn silent async-chunk failures into visible ones (#6033)
Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
2026-08-17 10:51:17 +08:00