A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 06:00 UTC
[Ascend]: Fixed the issue where OOT Platform vllm-ascend could not enable SP in Eager mode (#28935)
Signed-off-by: leo-pony <nengjunma@outlook.com>
Nengjun Ma committed
Dec 2, 2025 at 04:02 UTC
eaf81485eda13efc10e1c38f4094978eae04e77f
2 files changed
+15
-2
vllm/config/compilation.py
+7
@@ -855,6 +855,13 @@ class CompilationConfig:
855
self.compute_bs_to_padded_graph_size()
856
857
def set_splitting_ops_for_v1(self):
858
+ # To compatible with OOT hardware plugin platform (for example vllm-ascend)
859
+ # which currently only supports sequence parallelism in eager mode.
860
+ if self.mode != CompilationMode.VLLM_COMPILE:
861
+ if self.splitting_ops is None:
862
+ self.splitting_ops = []
863
+ return
864
+
865
# NOTE: this function needs to be called only when mode is
866
# CompilationMode.VLLM_COMPILE
867
assert self.mode == CompilationMode.VLLM_COMPILE, (
vllm/config/vllm.py
+8
-2
@@ -797,8 +797,7 @@ class VllmConfig:
797
), "MTP with cp_kv_cache_interleave_size > 1 is not supported now."
798
799
# Do this after all the updates to compilation_config.mode
800
- if self.compilation_config.mode == CompilationMode.VLLM_COMPILE:
801
- self.compilation_config.set_splitting_ops_for_v1()
800
+ self.compilation_config.set_splitting_ops_for_v1()
801
802
if self.compilation_config.pass_config.enable_sequence_parallelism:
803
# With pipeline parallelism or dynamo partitioning,
@@ -806,6 +805,13 @@ class VllmConfig:
805
# Use custom rms norm to unblock. In the future,
806
# the pass will operate on higher-level IR to avoid the issue.
807
# TODO: https://github.com/vllm-project/vllm/issues/27894
808
+ if self.compilation_config.mode != CompilationMode.VLLM_COMPILE:
809
+ logger.warning(
810
+ "Sequence parallelism is enabled, but running in wrong "
811
+ "vllm compile mode: %s.",
812
+ self.compilation_config.mode,
813
+ )
814
+
815
is_fullgraph = (
816
self.compilation_config.use_inductor_graph_partition
817
or len(self.compilation_config.splitting_ops) == 0