Skip to content

[Bugfix] Use PIECEWISE cudagraphs on Blackwell if max_model_len > 131072 #27114

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 5 commits into from
Oct 17, 2025
Merged

[Bugfix] Use PIECEWISE cudagraphs on Blackwell if max_model_len > 131072 #27114

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
322685c
Use PIECEWISE cudagraphs on Blackwell if max_model_len > 131072
mgoin Oct 17, 2025
e83e7af
Cleanup
mgoin Oct 17, 2025
d901ca7
Add comment
mgoin Oct 17, 2025
a8f373f
Merge branch 'main' into disable-trtllm-full-cudagraph-long-context
mgoin Oct 17, 2025
9a57a22
Skip this case if MLA is used
mgoin Oct 17, 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
52 changes: 37 additions & 15 deletions vllm/config/vllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -350,25 +350,47 @@ def __post_init__(self):
self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_AND_PIECEWISE
)
else:
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE

# pooling models and encoder-decoder models
# do not support full cudagraphs
if self.model_config is not None and (
self.model_config.pooler_config is not None
or self.model_config.is_encoder_decoder
):
# if cudagraph_mode has full cudagraphs, we need to check support
if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
# decode context parallel does not support full cudagraphs
if self.parallel_config.decode_context_parallel_size > 1:
logger.warning_once(
"Decode context parallel (DCP) is enabled, which is "
"incompatible with full CUDA graphs. "
"Overriding cudagraph_mode to PIECEWISE."
)
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
elif self.model_config is not None:
if self.model_config.pooler_config is not None:
logger.warning_once(
"Pooling models do not support full cudagraphs. "
"Overriding cudagraph_mode to PIECEWISE."
)
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE

# decode context parallel do not support full cudagraphs now.
if self.parallel_config.decode_context_parallel_size > 1:
logger.warning(
"Decode context parallel (DCP) is enabled, which is "
"incompatible with full CUDA graphs. Set "
"cudagraph_mode to PIECEWISE."
elif self.model_config.is_encoder_decoder:
logger.warning_once(
"Encoder-decoder models do not support full cudagraphs. "
"Overriding cudagraph_mode to PIECEWISE."
)
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
elif (
current_platform.is_cuda()
and current_platform.is_device_capability(100)
and self.model_config.max_model_len > 131072
and not self.model_config.use_mla
):
# Refer to vllm/utils/flashinfer.py::use_trtllm_attention()
logger.warning_once(
"NVIDIA Blackwell TRTLLM attention cannot support "
"max_model_len >= 131072 (found "
f"{self.model_config.max_model_len}), causing dynamic "
"dispatching that breaks full cudagraphs. "
"Overriding cudagraph_mode to PIECEWISE."
)
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
else:
self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE

# disable cudagraph when enforce eager execution
if self.model_config is not None and self.model_config.enforce_eager:
Expand Down