[ROCm][Bugfix] Reduce CUDA graph divergences (#57229)

Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
Matt
2026-09-17 05:12:13 +00:00
committed by GitHub
parent 0bfc7a15d0
commit 2d8a5c5741
5 changed files with 8 additions and 90 deletions
+5 -1
View File
@@ -98,6 +98,7 @@ from vllm.v1.worker.worker_base import CompilationTimes, WorkerBase
from vllm.v1.worker.workspace import init_workspace_manager
from ...model_executor.model_loader import TensorizerLoader
from .gpu.cudagraph_utils import has_compiled_submodule
from .gpu.warmup import warmup_kernels
from .utils import request_memory
@@ -772,7 +773,10 @@ class Worker(WorkerBase):
def compile_or_warm_up_model(self) -> CompilationTimes:
warmup_sizes: list[int] = []
if self.vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE:
if (
self.vllm_config.compilation_config.mode == CompilationMode.VLLM_COMPILE
and has_compiled_submodule(self.model_runner.get_model())
):
# warm up sizes that are not in cudagraph capture sizes,
# but users still want to compile for better performance,
# e.g. for the max-num-batched token size in chunked prefill.