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 13, 14:30 UTC
[Attention] Remove max cudagraph size limit of 992 (#27840)
Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
22quinn committed
Nov 7, 2025 at 22:33 UTC
608bb1446285db42fcc3aa33881c8c5dd0646ce0
2 files changed
-14
vllm/v1/attention/backends/flash_attn.py
-7
@@ -244,13 +244,6 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
244
self.max_cudagraph_size = self.compilation_config.max_cudagraph_capture_size
245
246
if self.use_full_cuda_graph and self.aot_schedule:
247
- if self.max_cudagraph_size > 992:
248
- # This condition derives from FA3's internal heuristic.
249
- # TODO(woosuk): Support larger cudagraph sizes.
250
- raise ValueError(
251
- "Capture size larger than 992 is not supported for full cuda graph."
252
- )
253
-
247
self.scheduler_metadata = torch.zeros(
248
vllm_config.scheduler_config.max_num_seqs + 1,
249
dtype=torch.int32,
vllm/v1/attention/backends/mla/flashattn_mla.py
-7
@@ -97,13 +97,6 @@ class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]
97
self.max_cudagraph_size = self.compilation_config.max_cudagraph_capture_size
98
99
if self.use_full_cuda_graph and self.fa_aot_schedule:
100
- if self.max_cudagraph_size > 992:
101
- # This condition derives from FA3's internal heuristic.
102
- # TODO(woosuk): Support larger cudagraph sizes.
103
- raise ValueError(
104
- "Capture size larger than 992 is not supported for full cuda graph."
105
- )
106
-
100
self.scheduler_metadata = torch.zeros(
101
vllm_config.scheduler_config.max_num_seqs + 1,
102
dtype=torch.int32,