biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core/DBO][1/N] Add Dual-Batch Overlap mechanism to VLLM (#23693)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Co-authored-by: yewentao256 <zhyanwentao@126.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Sage Moore
2025-09-16 09:21:48 -07:00
committed by GitHub
parent 08369289af
commit 567939953b
22 changed files with 1257 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/config/__init__.py
View File

@@ -2848,6 +2848,14 @@ class VllmConfig:
"when cudagraph_mode piecewise cudagraphs is used, "\
f"cudagraph_mode={self.compilation_config.cudagraph_mode}"
if self.parallel_config.enable_dbo:
a2a_backend = envs.VLLM_ALL2ALL_BACKEND
assert a2a_backend == "deepep_low_latency", \
"Microbatching currently only supports the deepep_low_latency "\
f"all2all backend. {a2a_backend} is not supported. To fix set "\
"the VLLM_ALL2ALL_BACKEND environment variable to "\
"deepep_low_latency and install the DeepEP kerenls."
if not self.instance_id:
self.instance_id = random_uuid()[:5]

8
vllm/config/parallel.py
View File

@@ -137,6 +137,14 @@ class ParallelConfig:
disable_custom_all_reduce: bool = False
"""Disable the custom all-reduce kernel and fall back to NCCL."""
enable_dbo: bool = False
"""Enable microbatching for the model executor."""
dbo_decode_token_threshold: int = 32
"""The threshold for microbatching. If the number of tokens in the
request is greater than this threshold, microbatching will be used.
Otherwise, the request will be processed in a single batch."""
ray_workers_use_nsight: bool = False
"""Whether to profile Ray workers with nsight, see https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler."""