xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-12 06:46:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] [DP/EP] Fix slow execution when BS <= DP (#25407)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Chris Bamford <chrisbam4d@gmail.com>
This commit is contained in:
Matthew Bonanni 2025-09-22 20:26:17 -04:00 committed by GitHub
parent 090197034f
commit ac0048c0ae
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/v1/worker/gpu_model_runner.py
View File

@ -55,7 +55,7 @@ from vllm.sampling_params import SamplingType
from vllm.sequence import IntermediateTensors, PoolerOutput from vllm.sequence import IntermediateTensors, PoolerOutput
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler, from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
GiB_bytes, check_use_alibi, get_dtype_size, GiB_bytes, cdiv, check_use_alibi, get_dtype_size,
is_pin_memory_available, is_pin_memory_available,
length_from_prompt_token_ids_or_embeds, round_up, length_from_prompt_token_ids_or_embeds, round_up,
supports_dynamo) supports_dynamo)
@ -2913,12 +2913,13 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
# Note: Overriding max_query_len to be the prefill tokens # Note: Overriding max_query_len to be the prefill tokens
max_query_len = num_prefill_tokens max_query_len = num_prefill_tokens
elif uniform_decode: elif uniform_decode:
num_reqs = num_tokens // max_query_len assert not create_mixed_batch
num_reqs = cdiv(num_tokens, max_query_len)
assert num_reqs <= max_num_reqs, \ assert num_reqs <= max_num_reqs, \
"Do not capture num_reqs > max_num_reqs for uniform batch" "Do not capture num_reqs > max_num_reqs for uniform batch"
num_scheduled_tokens_list = [max_query_len] * num_reqs num_scheduled_tokens_list = [max_query_len] * num_reqs
if num_tokens % max_query_len != 0: if num_tokens % max_query_len != 0:
num_scheduled_tokens_list[-1] += num_tokens % max_query_len num_scheduled_tokens_list[-1] = num_tokens % max_query_len
else: else:
num_reqs = min(num_tokens, max_num_reqs) num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs min_tokens_per_req = num_tokens // num_reqs

2
vllm/v1/worker/gpu_worker.py
View File

@ -487,7 +487,7 @@ class Worker(WorkerBase):
sort_by="self_cuda_time_total")) sort_by="self_cuda_time_total"))
def execute_dummy_batch(self) -> None: def execute_dummy_batch(self) -> None:
self.model_runner._dummy_run(1) self.model_runner._dummy_run(1, uniform_decode=True)
def add_lora(self, lora_request: LoRARequest) -> bool: def add_lora(self, lora_request: LoRARequest) -> bool:
return self.model_runner.add_lora(lora_request) return self.model_runner.add_lora(lora_request)