[Perf][DSv4.1] Run decoder replay layers in CUDA graphs and trim in PIECEWISE graphs (#59532)

Signed-off-by: Juntian Liu <Juntianl777@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Juntian Liu
2026-10-05 22:11:45 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 21d93d0d8c
commit eb2901b735
10 changed files with 445 additions and 71 deletions
+6
View File
@@ -1077,6 +1077,12 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.lora_config, self
),
)
self.model_state.capture_inner_cudagraphs(
input_buffers,
self.block_tables,
self.attn_groups,
self.kv_cache_config,
)
if self.speculator is not None:
with use_workspace_lane(self._draft_workspace_lane):
self.speculator.capture()