[Bugfix][MM][CG] Enable dual-path ViT CUDA graph for Step3-VL (#46034)

Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Shanshan Shen
2026-06-28 00:06:43 -07:00
committed by GitHub
co-authored by Isotr0py
parent c7ca0bccae
commit 35e6c86caa
2 changed files with 87 additions and 125 deletions
+1 -1
View File
@@ -136,7 +136,7 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
| `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ | ❌︎ | | `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ | ❌︎ |
| `Qwen3_5ForConditionalGeneration` | `Qwen3.5`, `Qwen3.6` | ✅︎ | ✅︎ | ❌︎ | | `Qwen3_5ForConditionalGeneration` | `Qwen3.5`, `Qwen3.6` | ✅︎ | ✅︎ | ❌︎ |
| `Qwen3_5MoeForConditionalGeneration` | `Qwen3.5-MoE`, `Qwen3.6-MoE` | ✅︎ | ✅︎ | ❌︎ | | `Qwen3_5MoeForConditionalGeneration` | `Qwen3.5-MoE`, `Qwen3.6-MoE` | ✅︎ | ✅︎ | ❌︎ |
| `Step3VLForConditionalGeneration` | `Step3-VL` | ✅︎ | ❌︎ | ❌︎ | | `Step3VLForConditionalGeneration` | `Step3-VL` | ✅︎ | ❌︎ | ✅︎ |
!!! note !!! note
Encoder CUDA Graphs have currently been tested with `--mm-encoder-attn-backend=FLASH_ATTN` and `--mm-encoder-attn-backend=FLASHINFER` on Blackwell GPUs. Encoder CUDA Graphs have currently been tested with `--mm-encoder-attn-backend=FLASH_ATTN` and `--mm-encoder-attn-backend=FLASHINFER` on Blackwell GPUs.
+86 -124
View File
@@ -589,6 +589,31 @@ class Step3VLForConditionalGeneration(
h2 = (h1 - 1) // 2 + 1 h2 = (h1 - 1) // 2 + 1
return h2 * h2 return h2 * h2
@property
def img_output_tokens(self) -> int:
return self._compute_spatial_tokens(
self.config.vision_config.image_size,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
@property
def patch_output_tokens(self) -> int:
return self._compute_spatial_tokens(
504,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
def _batched_encoder_forward(
self,
pixel_values: torch.Tensor,
) -> torch.Tensor:
image_features = self._process_image_features(
self._get_vision_model_output(pixel_values)
)
return image_features.reshape(-1, image_features.shape[-1])
def _parse_and_validate_image_input( def _parse_and_validate_image_input(
self, **kwargs: object self, **kwargs: object
) -> Step3VLImageInputs | None: ) -> Step3VLImageInputs | None:
@@ -695,6 +720,8 @@ class Step3VLForConditionalGeneration(
is_multimodal=is_multimodal, is_multimodal=is_multimodal,
) )
# -- SupportsEncoderCudaGraph protocol methods --
def get_encoder_cudagraph_config(self): def get_encoder_cudagraph_config(self):
from vllm.v1.worker.encoder_cudagraph_defs import ( from vllm.v1.worker.encoder_cudagraph_defs import (
EncoderCudaGraphConfig, EncoderCudaGraphConfig,
@@ -707,18 +734,16 @@ class Step3VLForConditionalGeneration(
"patch_pixel_values", "patch_pixel_values",
], ],
out_hidden_size=self.config.hidden_size, out_hidden_size=self.config.hidden_size,
enable_dual_path_graph=True,
global_token_per_image=self.img_output_tokens,
local_token_per_patch=self.patch_output_tokens,
) )
def get_encoder_cudagraph_budget_range( def get_encoder_cudagraph_budget_range(
self, self,
vllm_config: "VllmConfig", vllm_config: "VllmConfig",
) -> tuple[int, int]: ) -> tuple[int, int]:
# An image without patches min_budget = self.img_output_tokens
min_budget = self._compute_spatial_tokens(
self.config.vision_config.image_size,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
max_budget = min( max_budget = min(
vllm_config.scheduler_config.max_num_batched_tokens, vllm_config.scheduler_config.max_num_batched_tokens,
self.model_config.max_model_len, self.model_config.max_model_len,
@@ -732,22 +757,6 @@ class Step3VLForConditionalGeneration(
from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec
num_patches = mm_kwargs.get("num_patches") num_patches = mm_kwargs.get("num_patches")
img_output_tokens = self._compute_spatial_tokens(
self.config.vision_config.image_size,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
# NOTE: 504 is the hard coded size for each patch after processing
# by the vision model, which is determined by the current architecture
# of the vision model and may need to be updated if the architecture changes.
# The number of tokens for each patch is calculated based on this
# size and the patch size.
patch_output_tokens = self._compute_spatial_tokens(
504,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
img_grid = ( img_grid = (
self.config.vision_config.image_size // self.config.vision_config.patch_size self.config.vision_config.image_size // self.config.vision_config.patch_size
@@ -759,7 +768,11 @@ class Step3VLForConditionalGeneration(
return [ return [
EncoderItemSpec( EncoderItemSpec(
input_size=(total_image_pixel + num_patch * total_patch_pixel), input_size=(total_image_pixel + num_patch * total_patch_pixel),
output_tokens=(img_output_tokens + num_patch * patch_output_tokens), output_tokens=(
self.img_output_tokens + num_patch * self.patch_output_tokens
),
global_output_tokens=self.img_output_tokens,
local_output_tokens=num_patch * self.patch_output_tokens,
) )
for num_patch in num_patches for num_patch in num_patches
] ]
@@ -810,46 +823,30 @@ class Step3VLForConditionalGeneration(
EncoderCudaGraphCaptureInputs, EncoderCudaGraphCaptureInputs,
) )
# For pixel_value, the max input size is max_batch_size assert path in ("global", "local")
img_output_tokens = self._compute_spatial_tokens( if path == "global":
self.config.vision_config.image_size, max_num_images = token_budget // self.img_output_tokens
self.config.vision_config.patch_size, max_batch_size = min(max_batch_size, max_num_images)
self.config.understand_projector_stride, dummy_pixel_values = torch.randn(
) max_batch_size,
patch_output_tokens = self._compute_spatial_tokens( 3,
504, self.config.vision_config.image_size,
self.config.vision_config.patch_size, self.config.vision_config.image_size,
self.config.understand_projector_stride, device=device,
) dtype=dtype,
dummy_pixel_values = torch.randn( )
max_batch_size, values = {"pixel_values": dummy_pixel_values}
3, else:
self.config.vision_config.image_size, max_num_patches = token_budget // self.patch_output_tokens
self.config.vision_config.image_size, dummy_patch_pixel_values = torch.randn(
device=device, max_num_patches,
dtype=dtype, 3,
) 504,
# max_num_patches is the max total patches across the whole batch. 504,
# token_budget = max_batch_size * img_out + max_num_patches * patch_out device=device,
max_num_patches = max( dtype=dtype,
0, )
(token_budget - max_batch_size * img_output_tokens) // patch_output_tokens, values = {"patch_pixel_values": dummy_patch_pixel_values}
)
dummy_patch_pixel_values = torch.randn(
max_num_patches,
3,
504,
504,
device=device,
dtype=dtype,
)
# num_patches is NOT in values -- the per-item merge is done
# CPU-side by finalize_encoder_cudagraph_output using the actual
# batch's num_patches from mm_kwargs.
values = {
"pixel_values": dummy_pixel_values,
"patch_pixel_values": dummy_patch_pixel_values,
}
return EncoderCudaGraphCaptureInputs( return EncoderCudaGraphCaptureInputs(
values=values, values=values,
@@ -860,42 +857,22 @@ class Step3VLForConditionalGeneration(
values: dict[str, torch.Tensor], values: dict[str, torch.Tensor],
path: str = "default", path: str = "default",
) -> torch.Tensor: ) -> torch.Tensor:
# Graph captures only the compute (vision model + conv projector). assert path in ("global", "local")
# Per-item merge happens CPU-side in finalize_encoder_cudagraph_output if path == "global":
# using actual num_patches from the batch data. return self._batched_encoder_forward(values["pixel_values"])
pixel_values = values["pixel_values"] else:
patch_pixel_values = values["patch_pixel_values"] return self._batched_encoder_forward(values["patch_pixel_values"])
image_features = self._process_image_features(
self._get_vision_model_output(pixel_values)
)
has_patches = len(patch_pixel_values) > 0
if has_patches:
patch_features = self._process_image_features(
self._get_vision_model_output(patch_pixel_values)
)
# Deterministic single cat: [all_img_flat, all_patch_flat]
img_flat = image_features.reshape(-1, image_features.shape[-1])
if has_patches:
patch_flat = patch_features.reshape(-1, patch_features.shape[-1])
return torch.cat([img_flat, patch_flat], dim=0)
return img_flat
def encoder_eager_forward( def encoder_eager_forward(
self, self,
mm_kwargs: dict[str, Any], mm_kwargs: dict[str, Any],
path: str = "default", path: str = "default",
) -> torch.Tensor: ) -> torch.Tensor:
image_input = Step3VLImagePixelInputs( assert path in ("global", "local")
type="pixel_values", if path == "global":
pixel_values=mm_kwargs["pixel_values"], return self._batched_encoder_forward(mm_kwargs["pixel_values"])
patch_pixel_values=mm_kwargs["patch_pixel_values"], else:
num_patches=mm_kwargs["num_patches"], return self._batched_encoder_forward(mm_kwargs["patch_pixel_values"])
)
vision_embeddings = self._process_image_input(image_input)
return torch.cat(vision_embeddings, dim=0)
def postprocess_encoder_output( def postprocess_encoder_output(
self, self,
@@ -907,38 +884,24 @@ class Step3VLForConditionalGeneration(
batch_mm_kwargs: dict[str, Any] | None = None, batch_mm_kwargs: dict[str, Any] | None = None,
local_output: torch.Tensor | None = None, local_output: torch.Tensor | None = None,
): ):
"""CPU-side per-item merge after graph replay. """CPU-side per-item merge after dual-path graph replay.
The graph output is ``[all_img_flat, all_patch_flat]``. ``output`` contains global-image features and ``local_output``
This method splits the flat output into image and patch features, contains local-patch features (or ``None`` when there are no patches).
then reassembles per-item embeddings using the *actual* batch
``num_patches`` from ``batch_mm_kwargs`` (not the capture-time values).
""" """
num_patches = batch_mm_kwargs["num_patches"] num_patches = batch_mm_kwargs["num_patches"]
hidden = output.shape[-1] hidden = output.shape[-1]
bsz = len(indices) bsz = len(indices)
img_out = self._compute_spatial_tokens(
self.config.vision_config.image_size,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
patch_out = self._compute_spatial_tokens(
504,
self.config.vision_config.patch_size,
self.config.understand_projector_stride,
)
# Valid portion: bsz images, actual_total_patches patches
actual_np = [int(np) for np in num_patches] actual_np = [int(np) for np in num_patches]
total_patches = sum(actual_np) total_patches = sum(actual_np)
img_tokens = bsz * img_out img_tokens = bsz * self.img_output_tokens
patch_tokens = total_patches * patch_out patch_tokens = total_patches * self.patch_output_tokens
img_part = output[:img_tokens].reshape(bsz, img_out, hidden) global_part = output[:img_tokens].reshape(bsz, self.img_output_tokens, hidden)
if total_patches > 0: if total_patches > 0:
patch_part = output[img_tokens : img_tokens + patch_tokens].reshape( patch_part = local_output[:patch_tokens].reshape(
-1, patch_out, hidden -1, self.patch_output_tokens, hidden
) )
else: else:
patch_part = None patch_part = None
@@ -951,7 +914,7 @@ class Step3VLForConditionalGeneration(
if patch_part is not None and np > 0: if patch_part is not None and np > 0:
parts.append(patch_part[cur_patch : cur_patch + np].reshape(-1, hidden)) parts.append(patch_part[cur_patch : cur_patch + np].reshape(-1, hidden))
cur_patch += np cur_patch += np
parts.append(img_part[i].reshape(-1, hidden)) parts.append(global_part[i].reshape(-1, hidden))
merged[idx] = torch.cat(parts, dim=0) if len(parts) > 1 else parts[0] merged[idx] = torch.cat(parts, dim=0) if len(parts) > 1 else parts[0]
out = [merged[i] for i in indices] out = [merged[i] for i in indices]
@@ -969,14 +932,13 @@ class Step3VLForConditionalGeneration(
EncoderCudaGraphReplayBuffers, EncoderCudaGraphReplayBuffers,
) )
# Only patch_pixel_values lives in the values dict; num_patches is assert path in ("global", "local")
# processed CPU-side by finalize_encoder_cudagraph_output. if path == "global":
return EncoderCudaGraphReplayBuffers( values = {"pixel_values": mm_kwargs["pixel_values"]}
values={ else:
"pixel_values": mm_kwargs["pixel_values"], values = {"patch_pixel_values": mm_kwargs["patch_pixel_values"]}
"patch_pixel_values": mm_kwargs["patch_pixel_values"],
}, return EncoderCudaGraphReplayBuffers(values=values)
)
def forward( def forward(
self, self,