iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[MM][Perf][CG] Support ViT full CUDA graph for Qwen3.5 (#42151)

Browse Source 
Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2026-05-13 16:00:32 +08:00
committed by GitHub
parent 85b2fecab7
commit 92def124bc
4 changed files with 112 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/design/cuda_graphs_multimodal.md
+2 -1
View File
@@ -85,8 +85,9 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
| Architecture | Models | CG for Image | CG for Video |
| ------------ | ------ | ------------ | ------------ |
| `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ |
| `Qwen2_5_VLForConditionalGeneration` | `Qwen2.5-VL` | ✅︎ | ✅︎ |
| `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ |
| `Qwen3_5ForConditionalGeneration` | `Qwen3.5` | ✅︎ | ✅︎ |
!!! note
Encoder CUDA Graphs have currently been tested with `--mm-encoder-attn-backend=FLASH_ATTN` and `--mm-encoder-attn-backend=FLASHINFER` on Blackwell GPUs.
examples/generate/multimodal/vision_language_offline.py
+93 -1
View File
@@ -2179,6 +2179,92 @@ def run_qwen3_vl_moe(questions: list[str], modality: str) -> ModelRequestData:
)
# Qwen3.5-Dense
def run_qwen3_5(questions: list[str], modality: str) -> ModelRequestData:
model_name = "Qwen/Qwen3.5-4B"
mm_limit = {"image": 1, "video": 1} if modality == "image+video" else {modality: 1}
engine_args = EngineArgs(
model=model_name,
max_model_len=4096,
max_num_seqs=5,
mm_processor_kwargs={
"min_pixels": 28 * 28,
"max_pixels": 1280 * 28 * 28,
"fps": 1,
},
limit_mm_per_prompt=mm_limit,
)
image_placeholder = "<|vision_start|><|image_pad|><|vision_end|>"
video_placeholder = "<|vision_start|><|video_pad|><|vision_end|>"
if modality == "image":
placeholder = image_placeholder
elif modality == "video":
placeholder = video_placeholder
elif modality == "image+video":
placeholder = image_placeholder + video_placeholder
prompts = [
(
"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
f"<|im_start|>user\n{placeholder}"
f"{question}<|im_end|>\n"
"<|im_start|>assistant\n"
)
for question in questions
]
return ModelRequestData(
engine_args=engine_args,
prompts=prompts,
)
# Qwen3.5-MoE
def run_qwen3_5_moe(questions: list[str], modality: str) -> ModelRequestData:
model_name = "Qwen/Qwen3.5-35B-A3B"
mm_limit = {"image": 1, "video": 1} if modality == "image+video" else {modality: 1}
engine_args = EngineArgs(
model=model_name,
max_model_len=4096,
max_num_seqs=5,
mm_processor_kwargs={
"min_pixels": 28 * 28,
"max_pixels": 1280 * 28 * 28,
"fps": 1,
},
limit_mm_per_prompt=mm_limit,
)
image_placeholder = "<|vision_start|><|image_pad|><|vision_end|>"
video_placeholder = "<|vision_start|><|video_pad|><|vision_end|>"
if modality == "image":
placeholder = image_placeholder
elif modality == "video":
placeholder = video_placeholder
elif modality == "image+video":
placeholder = image_placeholder + video_placeholder
prompts = [
(
"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
f"<|im_start|>user\n{placeholder}"
f"{question}<|im_end|>\n"
"<|im_start|>assistant\n"
)
for question in questions
]
return ModelRequestData(
engine_args=engine_args,
prompts=prompts,
)
# R-4B
def run_r_vl(questions: list[str], modality: str) -> ModelRequestData:
assert modality == "image"
@@ -2442,6 +2528,8 @@ model_example_map = {
"qwen2_5_omni": run_qwen2_5_omni,
"qwen3_vl": run_qwen3_vl,
"qwen3_vl_moe": run_qwen3_vl_moe,
"qwen3_5": run_qwen3_5,
"qwen3_5_moe": run_qwen3_5_moe,
"rvl": run_r_vl,
"skywork_chat": run_skyworkr1v,
"smolvlm": run_smolvlm,
@@ -2460,13 +2548,17 @@ MODELS_NEED_VIDEO_METADATA = [
"molmo2",
"qwen3_vl",
"qwen3_vl_moe",
"qwen3_5",
"qwen3_5_moe",
]
MODELS_SUPPORT_VIT_CUDA_GRAPH = [
"qwen2_5_vl",
"qwen3_vl",
"qwen3_vl_moe",
"qwen2_5_vl",
"qwen3_5",
"qwen3_5_moe",
]
tests/models/multimodal/generation/test_vit_cudagraph.py
+15 -3
View File
@@ -42,6 +42,18 @@ def qwen_vl_chat_template(content: str) -> str:
MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
"qwen2_5_vl": VitCudagraphTestConfig(
model="Qwen/Qwen2.5-VL-3B-Instruct",
image_prompt=qwen_vl_chat_template(
"<|vision_start|><|image_pad|><|vision_end|>What is in this image?"
),
video_prompt=qwen_vl_chat_template(
"<|vision_start|><|video_pad|><|vision_end|>"
"Describe this video in one sentence."
),
needs_video_metadata=False,
marks=[pytest.mark.core_model],
),
"qwen3_vl": VitCudagraphTestConfig(
model="Qwen/Qwen3-VL-2B-Instruct",
image_prompt=qwen_vl_chat_template(
@@ -54,8 +66,8 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
needs_video_metadata=True,
marks=[pytest.mark.core_model],
),
"qwen2_5_vl": VitCudagraphTestConfig(
model="Qwen/Qwen2.5-VL-3B-Instruct",
"qwen3_5": VitCudagraphTestConfig(
model="Qwen/Qwen3.5-0.8B",
image_prompt=qwen_vl_chat_template(
"<|vision_start|><|image_pad|><|vision_end|>What is in this image?"
),
@@ -63,7 +75,7 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
"<|vision_start|><|video_pad|><|vision_end|>"
"Describe this video in one sentence."
),
needs_video_metadata=False,
needs_video_metadata=True,
marks=[pytest.mark.core_model],
),
}
vllm/model_executor/models/qwen3_5.py
+2
View File
@@ -565,6 +565,7 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid)
multimodal_config = vllm_config.model_config.multimodal_config
self.config = config
self.model_config = vllm_config.model_config
self.multimodal_config = multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
# Qwen3.5 does not support multimodal pruning (EVS).
@@ -778,6 +779,7 @@ class Qwen3_5MoeForConditionalGeneration(
multimodal_config = vllm_config.model_config.multimodal_config
self.config = config
self.model_config = vllm_config.model_config
self.multimodal_config = multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
# Qwen3.5 does not support multimodal pruning (EVS).