Skip to content

[Core] Whisper enable FULL_DECODE_ONLY CudaGraph #30072

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
vllm-bot merged 8 commits into from
Dec 10, 2025
Merged

[Core] Whisper enable FULL_DECODE_ONLY CudaGraph #30072

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
0353950
init
NickLucche Dec 3, 2025
97eb2c3
mind the padding!
NickLucche Dec 4, 2025
548af89
has_encoder_output check+config changes
NickLucche Dec 4, 2025
21626df
cruft
NickLucche Dec 4, 2025
c06589f
address review
NickLucche Dec 5, 2025
cd13d1f
address piecewise mode
NickLucche Dec 9, 2025
427ad52
skip cg in tests
NickLucche Dec 10, 2025
fc2bedd
todo
NickLucche Dec 10, 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
2 changes: 2 additions & 0 deletions tests/models/multimodal/generation/test_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,8 @@ def run_test(
max_model_len=448,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
# TODO (NickLucche) figure out output differences with non-eager and re-enable
enforce_eager=True,
Copy link
Copy Markdown
Member

@DarkLight1337 DarkLight1337 Dec 10, 2025

Choose a reason for hiding this comment

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

Could you explain a bit why the test needs to set this explicitly?

Copy link
Copy Markdown
Collaborator Author

@NickLucche NickLucche Dec 10, 2025

Choose a reason for hiding this comment

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

@DarkLight1337 There's a subtle difference in output which makes the test fail.


[2025-12-09T18:42:49Z]   -  And the 0-1 pitch on the way to Edgar Martinez. Swung on the line down the left field line for a base hit. Here comes Joy. Here is Junior to third base. They're going to wave him in. The throw to the plate will be late. The Mariners are going to play for the American League Championship. I don't believe it. It just continues. My, oh, my.
--
[2025-12-09T18:42:49Z]   ?                                                                                                                                                                                                                                                                                                                                                    ^
[2025-12-09T18:42:49Z]   +  And the 0-1 pitch on the way to Edgar Martinez. Swung on the line down the left field line for a base hit. Here comes Joy. Here is Junior to third base. They're going to wave him in. The throw to the plate will be late. The Mariners are going to play for the American League Championship. I don't believe it. It just continues. My, oh, my God.

but tbh I think checking token by token a copy-pasted output is too strict.
We can restructure the test to use some new invariant techniques, or just relax it in a separate PR.

Copy link
Copy Markdown
Member

@DarkLight1337 DarkLight1337 Dec 10, 2025

Choose a reason for hiding this comment

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

Yeah we can check logprobs instead. For now can you add a code comment with a TODO to fix this later?

Copy link
Copy Markdown
Collaborator Author

@NickLucche NickLucche Dec 10, 2025

Choose a reason for hiding this comment

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

sure

) as vllm_model:
llm = vllm_model.llm

Expand Down
30 changes: 19 additions & 11 deletions vllm/config/vllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -666,8 +666,9 @@ def has_blocked_weights():

default_config = OPTIMIZATION_LEVEL_TO_CONFIG[self.optimization_level]
self._apply_optimization_level_defaults(default_config)

if (
self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
self.compilation_config.cudagraph_mode.requires_piecewise_compilation()
and self.compilation_config.mode != CompilationMode.VLLM_COMPILE
):
logger.info(
Expand All @@ -692,22 +693,29 @@ def has_blocked_weights():

if current_platform.support_static_graph_mode():
# if cudagraph_mode has full cudagraphs, we need to check support
if (
self.compilation_config.cudagraph_mode.has_full_cudagraphs()
and self.model_config is not None
):
if self.model_config.pooler_config is not None:
if model_config := self.model_config:
if (
self.compilation_config.cudagraph_mode.has_full_cudagraphs()
and 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
elif self.model_config.is_encoder_decoder:
logger.warning_once(
"Encoder-decoder models do not support full cudagraphs. "
"Overriding cudagraph_mode to PIECEWISE."
elif (
model_config.is_encoder_decoder
and self.compilation_config.cudagraph_mode
not in (CUDAGraphMode.NONE, CUDAGraphMode.FULL_DECODE_ONLY)
):
logger.info_once(
"Encoder-decoder models do not support %s. "
"Overriding cudagraph_mode to FULL_DECODE_ONLY.",
self.compilation_config.cudagraph_mode.name,
)
self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_DECODE_ONLY
)
self.compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE

# disable cudagraph when enforce eager execution
if self.model_config is not None and self.model_config.enforce_eager:
Expand Down
11 changes: 10 additions & 1 deletion vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1267,6 +1267,8 @@ def _get_encoder_seq_lens(
if not isinstance(kv_cache_spec, CrossAttentionSpec):
return None, None

# Zero out buffer for padding requests that are not actually scheduled (CGs)
self.encoder_seq_lens.np[:num_reqs] = 0
# Build encoder_seq_lens array mapping request indices to
# encoder lengths for inputs scheduled in this batch
for req_id in num_scheduled_tokens:
Expand Down Expand Up @@ -2764,6 +2766,7 @@ def _determine_batch_execution_and_padding(
# be improved in model runner v2)
force_uniform_decode: bool | None = None,
force_has_lora: bool | None = None,
num_encoder_reqs: int = 0,
) -> tuple[
CUDAGraphMode,
BatchDescriptor,
Expand All @@ -2780,6 +2783,11 @@ def _determine_batch_execution_and_padding(
if force_uniform_decode is None
else force_uniform_decode
)
# Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
# is present). Also, chunked-prefill is disabled, so batch are uniform.
has_encoder_output = (
Comment thread
NickLucche marked this conversation as resolved.
self.model_config.is_encoder_decoder and num_encoder_reqs > 0
)

has_lora = (
len(self.input_batch.lora_id_to_lora_request) > 0
Expand All @@ -2799,7 +2807,7 @@ def _determine_batch_execution_and_padding(
)

cudagraph_mode, batch_descriptor = dispatch_cudagraph(
num_tokens_padded, use_cascade_attn
num_tokens_padded, use_cascade_attn or has_encoder_output
)
num_tokens_padded = batch_descriptor.num_tokens

Expand Down Expand Up @@ -2997,6 +3005,7 @@ def execute_model(
num_scheduled_tokens_np=num_scheduled_tokens_np,
max_num_scheduled_tokens=max_num_scheduled_tokens,
use_cascade_attn=cascade_attn_prefix_lens is not None,
num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
)

logger.debug(
Expand Down