minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Fix] Avoid applying cuda graph input-buffer registry on non-cuda devices (#27549)

Browse Source
This commit is contained in:
Zaili Wang
2026-06-09 00:24:55 -07:00
committed by GitHub
parent a287ab83c0
commit d981b7b9c4
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/model_runner.py
+3 -3
View File
@@ -3204,7 +3204,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
forward_batch: ForwardBatch, forward_batch: ForwardBatch,
pp_proxy_tensors=None, pp_proxy_tensors=None,
) -> Union[LogitsProcessorOutput, PPProxyTensors]: ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
if not self.server_args.enable_pdmux: if not self.server_args.enable_pdmux and self.device == "cuda":
forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors) forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors)
# Set extra arguments # Set extra arguments
pdmux_override = False pdmux_override = False
@@ -3294,7 +3294,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
ret = self.piecewise_cuda_graph_runner.replay(forward_batch, **kwargs) ret = self.piecewise_cuda_graph_runner.replay(forward_batch, **kwargs)
return (ret, can_run_graph) return (ret, can_run_graph)
if not self.server_args.enable_pdmux: if not self.server_args.enable_pdmux and self.device == "cuda":
forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors) forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors)
# Launch model forward # Launch model forward
@@ -3355,7 +3355,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
# called from the idle path can re-read a prior batch's req_pool # called from the idle path can re-read a prior batch's req_pool
# indices and trigger SWA mapping use-after-free. # indices and trigger SWA mapping use-after-free.
if forward_batch.batch_size > 0: if forward_batch.batch_size > 0:
if not self.server_args.enable_pdmux: if not self.server_args.enable_pdmux and self.device == "cuda":
forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors) forward_batch = self._eager_fb_view(forward_batch, pp_proxy_tensors)
self.attn_backend.init_forward_metadata(forward_batch) self.attn_backend.init_forward_metadata(forward_batch)
else: else: