Skip to content

[Model] Fix Qwen3VL and Qwen3Omni after torch.compile changes #27705

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Oct 29, 2025
Merged

[Model] Fix Qwen3VL and Qwen3Omni after torch.compile changes #27705

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
70d5ece
[Model] Fix Qwen3VL and Qwen3Omni after torch.compile changes
lgeiger Oct 29, 2025
54f7b08
add
ywang96 Oct 29, 2025
f462fb6
Merge branch 'main' into fix-qwen3
ywang96 Oct 29, 2025
fa6b59e
precommit
ywang96 Oct 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions vllm/model_executor/models/qwen2_5_vl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -836,10 +836,8 @@ def compute_attn_mask_seqlen(
self,
cu_seqlens: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
max_seqlen, seqlens = (
torch.zeros(1, device=cu_seqlens.device),
torch.zeros(1, device=cu_seqlens.device),
)
max_seqlen = torch.zeros([], device=cu_seqlens.device)
seqlens = torch.zeros(1, device=cu_seqlens.device)
Comment on lines -839 to +840
Copy link
Contributor Author

@lgeiger lgeiger Oct 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

max_seqlen is actually a scalar value so this updates the dummy values to match this.

if (
self.attn_backend == _Backend.FLASH_ATTN
or self.attn_backend == _Backend.ROCM_AITER_FA
Expand Down
14 changes: 8 additions & 6 deletions vllm/model_executor/models/qwen3_omni_moe_thinker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -223,8 +223,8 @@ def forward(
x: torch.Tensor,
cu_seqlens: torch.Tensor,
rotary_pos_emb: torch.Tensor,
max_seqlen: int | None = None, # Only used for Flash Attention
seqlens: list[int] | None = None, # Only used for xFormers
max_seqlen: torch.Tensor, # Only used for Flash Attention
seqlens: torch.Tensor, # Only used for xFormers
) -> torch.Tensor:
x = x + self.attn(
self.norm1(x),
Expand Down Expand Up @@ -488,12 +488,13 @@ def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor:
def compute_attn_mask_seqlen(
self,
cu_seqlens: torch.Tensor,
) -> tuple[int | None, list[int] | None]:
max_seqlen, seqlens = None, None
) -> tuple[torch.Tensor, torch.Tensor]:
max_seqlen = torch.zeros([], device=cu_seqlens.device)
seqlens = torch.zeros(1, device=cu_seqlens.device)
if self.attn_backend == _Backend.FLASH_ATTN:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
elif self.attn_backend == _Backend.XFORMERS:
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
return max_seqlen, seqlens

def forward(
Expand Down Expand Up @@ -1114,6 +1115,7 @@ def get_placeholder_str(cls, modality: str, i: int) -> str | None:

def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
self.vllm_config = vllm_config # needed for torch compile forward context
thinker_config: Qwen3OmniMoeThinkerConfig = (
vllm_config.model_config.hf_config.thinker_config
)
Expand Down
13 changes: 7 additions & 6 deletions vllm/model_executor/models/qwen3_vl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -231,8 +231,8 @@ def forward(
x: torch.Tensor,
cu_seqlens: torch.Tensor,
rotary_pos_emb: torch.Tensor,
max_seqlen: int | None = None, # Only used for Flash Attention
seqlens: list[int] | None = None, # Only used for xFormers
max_seqlen: torch.Tensor, # Only used for Flash Attention
seqlens: torch.Tensor, # Only used for xFormers
) -> torch.Tensor:
x = x + self.attn(
self.norm1(x),
Expand Down Expand Up @@ -512,15 +512,16 @@ def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor:
def compute_attn_mask_seqlen(
self,
cu_seqlens: torch.Tensor,
) -> tuple[int | None, list[int] | None]:
max_seqlen, seqlens = None, None
) -> tuple[torch.Tensor, torch.Tensor]:
max_seqlen = torch.zeros([], device=cu_seqlens.device)
seqlens = torch.zeros(1, device=cu_seqlens.device)
if (
self.attn_backend == _Backend.FLASH_ATTN
or self.attn_backend == _Backend.ROCM_AITER_FA
):
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
elif self.attn_backend == _Backend.XFORMERS:
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
return max_seqlen, seqlens

def forward(
Expand Down