[Bugfix][HiSparse] Reject cudagraph_mode=FULL at startup (#59688)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
64cb683842
commit
54d93af9fb
@@ -502,6 +502,17 @@ def test_hisparse_rejects_pipeline_parallelism(monkeypatch):
|
|||||||
)
|
)
|
||||||
|
|
||||||
|
|
||||||
|
def test_hisparse_rejects_full_cudagraph_mode(monkeypatch):
|
||||||
|
monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
|
||||||
|
monkeypatch.setattr(current_platform, "support_static_graph_mode", lambda: True)
|
||||||
|
monkeypatch.setattr("vllm.config.vllm.HAS_TRITON", True)
|
||||||
|
with pytest.raises(ValueError, match="does not support cudagraph_mode=FULL"):
|
||||||
|
VllmConfig(
|
||||||
|
attention_config=AttentionConfig(hisparse_config=HiSparseConfig()),
|
||||||
|
compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.FULL),
|
||||||
|
)
|
||||||
|
|
||||||
|
|
||||||
def test_hisparse_rejects_disabled_hybrid_kv_cache_manager(monkeypatch):
|
def test_hisparse_rejects_disabled_hybrid_kv_cache_manager(monkeypatch):
|
||||||
monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
|
monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
|
||||||
monkeypatch.setattr("vllm.config.vllm.HAS_TRITON", True)
|
monkeypatch.setattr("vllm.config.vllm.HAS_TRITON", True)
|
||||||
|
|||||||
@@ -1836,6 +1836,12 @@ class VllmConfig:
|
|||||||
raise ValueError(
|
raise ValueError(
|
||||||
"HiSparse does not support decode context parallelism."
|
"HiSparse does not support decode context parallelism."
|
||||||
)
|
)
|
||||||
|
if self.compilation_config.cudagraph_mode == CUDAGraphMode.FULL:
|
||||||
|
raise ValueError(
|
||||||
|
"HiSparse does not support cudagraph_mode=FULL; use "
|
||||||
|
"FULL_AND_PIECEWISE (the default), which captures FULL graphs "
|
||||||
|
"for decode batches."
|
||||||
|
)
|
||||||
if not self.scheduler_config.scheduler_reserve_full_isl:
|
if not self.scheduler_config.scheduler_reserve_full_isl:
|
||||||
# Without it, async loads admitted against free host blocks can
|
# Without it, async loads admitted against free host blocks can
|
||||||
# each wait on host pages the others hold, and waiting requests
|
# each wait on host pages the others hold, and waiting requests
|
||||||
|
|||||||
Reference in New Issue
Block a user