Isotr0py 3affc7766d Enable VAE tiling for Qwen-Image-2.1 (#5) ...

* init

Signed-off-by: Isotr0py <mozf@inferact.ai>

* fix

Signed-off-by: Isotr0py <mozf@inferact.ai>

* test(diffusion): add Qwen-Image-2.1 unit tests

Signed-off-by: Isotr0py <mozf@inferact.ai>

* fix(diffusion): Qwen-Image-2.1 VAE spatial ratio, tiling e2e verification

- AutoencoderKLQwenImage21: derive spatial_compression_ratio from the
  architecture (2 ** len(temperal_downsample) = 16x) instead of trusting
  the checkpoint's scale_factor_spatial=8, which corrupted tiled
  decode/encode geometry (output cropped to half size) and tiling gating.
- pipeline: _read_vae_scale_factor prefers the architecture-derived value
  for the same reason.
- models/qwen_image_21/__init__: lazy exports to break the circular import
  distributed VAE -> package __init__ -> pipeline -> distributed VAE.
- Add VAE unit tests: ratio derivation, tiled vs untiled decode/encode
  consistency, distributed tile_split/exec/merge == tiled_decode.
- e2e matrix (GB200, seed 42, 20 steps) all green; see e2e_out/qwen21_vae/.
- examples: qwen_image_21_edit.py gains --vae-use-tiling.
- docs: mark VAE patch-parallel verified for Qwen-Image-2.1; recipe notes
  the default-tile banding limitation at 2048px+.

* clean tests

Signed-off-by: Isotr0py <mozf@inferact.ai>

* fix cudnn determination

Signed-off-by: Isotr0py <mozf@inferact.ai>

---------

Signed-off-by: Isotr0py <mozf@inferact.ai>

2026-09-18 10:02:50 +08:00

..

attention

[Rebase] Rebase to vLLM 0.29.0 (#7230)

2026-09-10 12:08:28 +08:00

cache

[Model] Add native MAGI-2 Preview diffusion support (#5918)

2026-09-05 15:32:30 +08:00

diffusion_kv

[Rebase] Rebase to vLLM 0.29.0 (#7230)

2026-09-10 12:08:28 +08:00

distributed

Enable VAE tiling for Qwen-Image-2.1 (#5)

2026-09-18 10:02:50 +08:00

executor

[diffusion][feature] Add component-selective offload policies (#5929)

2026-09-06 05:40:09 +08:00

hooks

[Perf][Diffusion] Skip the Ulysses length all-gather when shards are padded equal (#5717)

2026-09-06 16:23:30 +08:00

interaction

[Core][Diffusion] Refactor prompt updates into a modality-ready interaction framework (#6294)

2026-09-07 11:40:04 +08:00

layers

[Rebase] Rebase to vLLM 0.29.0 (#7230)

2026-09-10 12:08:28 +08:00

lora

Initialize QMI pipeline implementation (#2)

2026-09-16 15:14:23 +08:00

model_loader

Add vLLM-Omni library info to Hugging Face Hub requests (#5381)

2026-09-08 10:08:38 -06:00

models

Enable VAE tiling for Qwen-Image-2.1 (#5)

2026-09-18 10:02:50 +08:00

offloader

[Refactor][Diffusion] Resolve offload topology through one plan resolver (#7209)

2026-09-09 08:17:38 +08:00

postprocess

[Bugfix] Fix RIFE device selection for CPU-transported videos (#2876)

2026-04-17 19:19:34 +08:00

profiler

[Bugfix] Report stepwise Qwen-Image profiler metrics (#4517)

2026-07-03 04:32:22 +00:00

quantization

[Bugfix] Fix HSDP compatibility with the new online FP8 linear method (#5677)

2026-08-14 15:45:55 +08:00

sched

[Bugfix][Diffusion] Include guidance_scale_2_provided in the request-batch key (#7078)

2026-09-07 15:58:19 +08:00

utils

[1/N] Stream Wan VAE chunks to the media consumer (#7016)

2026-09-08 07:41:13 +08:00

worker

[Feat][OmniVoice]Support Varlen Attn, Request-Batch and Step-Execution (#6408)

2026-09-08 18:58:38 -07:00

__init__.py

[Diffusion] Qwen image support (#82)

2025-11-28 14:22:11 +08:00

compile.py

fix: keep DLO hooks outside regional compilation (#6073)

2026-08-22 21:44:03 +08:00

config.py

[Diffusion][Attention] Support per-role attention backend via CLI (#2681)

2026-05-10 11:59:05 +08:00

data.py

Enable CUDA Graph for DiT decode layers (#4)

2026-09-18 10:02:39 +08:00

diffusion_engine.py

[Feature][Diffusion] Expose detailed pipeline timings (#6822)

2026-09-09 09:58:42 +00:00

envs.py

[Diffusion] Add native SymmMem Fast Ulysses transport (#6340)

2026-08-30 04:14:21 +00:00

forward_context.py

[Perf][Diffusion] Skip the Ulysses length all-gather when shards are padded equal (#5717)

2026-09-06 16:23:30 +08:00

inline_stage_diffusion_client.py

[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

2026-09-06 05:41:54 +08:00

io_support.py

[Feature] Support text encoder Disaggregation for H3 (#5885)

2026-08-28 14:27:05 +08:00

ipc.py

[CI]Remove DreamID-Omni and MagiHuman support (#6357)

2026-08-26 01:31:12 +00:00

model_metadata.py

Initialize QMI pipeline implementation (#2)

2026-09-16 15:14:23 +08:00

output_formatter.py

[Feature] Support text encoder Disaggregation for H3 (#5885)

2026-08-28 14:27:05 +08:00

registry.py

Initialize QMI pipeline implementation (#2)

2026-09-16 15:14:23 +08:00

request.py

[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

2026-09-06 05:41:54 +08:00

stage_diffusion_client.py

[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

2026-09-06 05:41:54 +08:00

stage_diffusion_proc.py

[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

2026-09-06 05:41:54 +08:00

vllm_config.py

[Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) (#6310)

2026-09-06 05:41:54 +08:00