xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 17:26:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Fix DP/EP hang (#25906)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson 2025-09-30 00:18:59 -04:00 committed by GitHub
parent 61aedb5ffe
commit 23194d83e8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 15 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/v1/worker/gpu_model_runner.py
View File

@ -3075,13 +3075,19 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
# We currently only microbatch if the number of tokens is # We currently only microbatch if the number of tokens is
# over a certain threshold. # over a certain threshold.
if self.parallel_config.enable_dbo and allow_microbatching: if self.parallel_config.enable_dbo and allow_microbatching:
ubatch_slices, num_tokens_after_padding = ubatch_split( ubatch_slices, ubatch_num_tokens_after_padding = ubatch_split(
num_scheduled_tokens, num_scheduled_tokens,
total_num_scheduled_tokens, total_num_scheduled_tokens,
total_num_scheduled_tokens, total_num_scheduled_tokens,
uniform_decode=uniform_decode, uniform_decode=uniform_decode,
vllm_config=self.vllm_config, vllm_config=self.vllm_config,
) )
# Currently when DBO is enabled `ubatch_split` returns
# the num_tokens_after_padding for a single ubatch, but we have 2
# TODO(sage,lucas): this is cruft that should be addressed in the
# padding refactor.
if ubatch_num_tokens_after_padding is not None:
num_tokens_after_padding = ubatch_num_tokens_after_padding * 2
# If we failed to microbatch, currently need to resynchronize # If we failed to microbatch, currently need to resynchronize
# TODO(lucas,sage): we should be able to avoid this second sync by # TODO(lucas,sage): we should be able to avoid this second sync by
@ -3198,7 +3204,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
# filter out the valid batch descriptor # filter out the valid batch descriptor
_cg_mode, batch_descriptor = self.cudagraph_dispatcher.dispatch( _cg_mode, batch_descriptor = self.cudagraph_dispatcher.dispatch(
BatchDescriptor(num_tokens=num_tokens, BatchDescriptor(num_tokens=num_tokens_after_padding,
uniform_decode=uniform_decode)) \ uniform_decode=uniform_decode)) \
if not is_profile else (CUDAGraphMode.NONE, None) if not is_profile else (CUDAGraphMode.NONE, None)
if cudagraph_runtime_mode is not None: if cudagraph_runtime_mode is not None:
@ -3212,7 +3218,13 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
cudagraph_runtime_mode = _cg_mode cudagraph_runtime_mode = _cg_mode
if ubatch_slices is not None: if ubatch_slices is not None:
num_tokens = num_tokens // 2 # Adjust values to reflect a single ubatch.
# TODO(sage,lucas): this is cruft that should be addressed in
# the padding refactor.
num_tokens_after_padding = ubatch_slices[0].num_tokens
if num_tokens_across_dp is not None:
num_tokens_across_dp[:] = num_tokens_after_padding
with self.maybe_randomize_inputs(input_ids), set_forward_context( with self.maybe_randomize_inputs(input_ids), set_forward_context(
attn_metadata, attn_metadata,
self.vllm_config, self.vllm_config,