Skip to content

[Bugfix][ROCm] Fix ViT rotary embeddings for torch.compile compatibility on ROCm #27748

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 6 commits into from
Nov 4, 2025
+8 −5
Merged

[Bugfix][ROCm] Fix ViT rotary embeddings for torch.compile compatibility on ROCm #27748

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
002d181
fix vit rotary embedding on ROCm platform to be torch compile compatible
vllmellm Oct 29, 2025
e723dc2
pre-commit fixes
vllmellm Oct 29, 2025
bc6e808
remove rotary embedding func from flash_attn lib and retrieve the ori…
vllmellm Oct 30, 2025
725f07d
keep triton rotary embedding for models that are not torch compile co…
vllmellm Oct 31, 2025
d4d130e
fix dropout_p dtype for glm vision models when aiter package enabled
vllmellm Oct 31, 2025
f0efc00
Merge remote-tracking branch 'origin/main' into vit-rotary-embed-fix
vllmellm Nov 3, 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
11 changes: 7 additions & 4 deletions vllm/model_executor/layers/rotary_embedding/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,11 @@ def dispatch_rotary_emb_function(
if current_platform.is_cuda():
return apply_rotary_emb

if current_platform.is_rocm():
Copy link
Contributor

@tjtanaa tjtanaa Oct 30, 2025

Choose a reason for hiding this comment

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

@vllmellm @DarkLight1337 What do you think? Should we keep the triton rotary embedding for backwards compatibility at this moment, by adding a condition torch.compiler.is_compiling() instead as there are other model definition files that are not using torch.compile yet.

E.g. model definition files that are not using torch.compile

  • vllm/model_executor/models/dots_ocr.py
  • vllm/model_executor/models/ernie45_vl.py
  • vllm/model_executor/models/glm4_1v.py
  • vllm/model_executor/models/qwen2_vl.py
  • vllm/model_executor/models/siglip2navit.py

Copy link
Contributor

@tjtanaa tjtanaa Oct 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Another thing to note at this point is that the torch.compile has been disabled as there are issues with dynamic slicing.

So if we add this condition torch.compiler.is_compiling(), then we don't need to postpone this PR while ensuring no performance regression for both cases torch.compile is enabled/disabled.

Right now, they are exploring new fixes e.g. #27764

Copy link
Contributor

@Lucaskabela Lucaskabela Oct 30, 2025

Choose a reason for hiding this comment

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

My two cents: I see the benefit of gating with torch.compiler.is_compiling(), especially for models we don't have covered with torch.compile yet, so I am in favor of that approach as opposed to completely removing

Copy link
Member

@ywang96 ywang96 Oct 30, 2025

Choose a reason for hiding this comment

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

+1 on adding a condition on checking if it's torch compiled.

On a side note - is this PR really considered a bugfix? I thought after #27760 things should already work on AMD. Did I miss something here?

Copy link
Contributor

@tjtanaa tjtanaa Oct 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

@ywang96 PR #27760 is a bugfix by reverting the torch.compile PR. If the torch.compile is reenable, AMD needs this bugfix PR to address the incompatibility of the flash_attn.rotary_embedding triton kernel with torch.compile. The discussion of current comment thread is to address the case where other models still haven't had torch.compile support.

Copy link
Contributor

@tjtanaa tjtanaa Oct 30, 2025

Choose a reason for hiding this comment

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

And in this PR we found that torch.compile-ed rotary embedding is faster than triton implementation on ROCm. I think it will be the same case on CUDA.

Copy link
Member

@ywang96 ywang96 Oct 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Yea I was just confirming whether the current main branch is broken on AMD (because it'll affect our release), and it seems that it's not because we currently disabled it, correct?

Copy link
Contributor

@tjtanaa tjtanaa Oct 30, 2025

Choose a reason for hiding this comment

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

@ywang96 On current main branch, the unit tests are healthy and the Qwen3 VL accuracy are normal. 👍

Copy link
Contributor Author

@vllmellm vllmellm Oct 31, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

@tjtanaa @Lucaskabela those model path are not using this dispatching function. it is only used in qwen2_vl.py. While qwen2.5_vl.py and glm4_1v.py are importing the apply_rotary_pos_emb_vision function from qwen2_vl.py and inside this function the dispatch is used.

vllm/vllm/model_executor/models/qwen2_vl.py

Lines 314 to 320 in 3696050

def apply_rotary_pos_emb_vision(t: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
rotary_emb_function = dispatch_rotary_emb_function(default=apply_rotary_emb_torch)
t_ = t.float()
cos = freqs.cos()
sin = freqs.sin()
output = rotary_emb_function(t_, cos, sin).type_as(t)
return output

@DarkLight1337 @Lucaskabela @tjtanaa we may decide to keep the triton embedding function here only for one case that is glm4_1v.py or not ?

# if torch compile is not enabled
# use rotary embedding function from flash_attn package
# otherwise use the naive pytorch embedding implementation
# is faster when torch compile is enabled.
if current_platform.is_rocm() and not torch.compiler.is_compiling():
if find_spec("flash_attn") is not None:
from flash_attn.ops.triton.rotary import apply_rotary

Expand All @@ -87,11 +91,10 @@ def dispatch_rotary_emb_function(
"flash_attn is not installed. Falling back to PyTorch "
"implementation for rotary embeddings."
)

if default is not None:
return default
else:
return apply_rotary_emb_torch

return apply_rotary_emb_torch


# yarn functions
Expand Down
2 changes: 1 addition & 1 deletion vllm/model_executor/models/glm4_1v.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -370,7 +370,7 @@ def forward(
cu_seqlens_k=cu_seqlens,
max_seqlen_q=max_seqlen,
max_seqlen_k=max_seqlen,
dropout_p=0,
dropout_p=0.0,
causal=False,
)

Expand Down