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

[NPU] Fix dead patch_model monkey-patch breaking NPU torch.compile capture (#27774)

Browse Source
This commit is contained in:
Cheng Wan
2026-06-10 01:15:50 -07:00
committed by GitHub
parent 01f10acd06
commit 70c71ba183
1 changed files with 3 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/runner/decode_cuda_graph_runner.py
+3 -5
View File
@@ -35,10 +35,8 @@ import torch
import tqdm import tqdm
from torch.profiler import ProfilerActivity, profile from torch.profiler import ProfilerActivity, profile
from sglang.srt.compilation.torch_compile_decoration import ( from sglang.srt.compilation import torch_compile_decoration
patch_model, from sglang.srt.compilation.torch_compile_decoration import set_torch_compile_config
set_torch_compile_config,
)
from sglang.srt.distributed import get_tensor_model_parallel_rank from sglang.srt.distributed import get_tensor_model_parallel_rank
from sglang.srt.distributed.parallel_state import ( from sglang.srt.distributed.parallel_state import (
graph_capture, graph_capture,
@@ -815,7 +813,7 @@ class DecodeCudaGraphRunner(BaseCudaGraphRunner):
for variant_label, _variant_has_lora in lora_variants: for variant_label, _variant_has_lora in lora_variants:
_set_capture_lora_variant(variant_label) _set_capture_lora_variant(variant_label)
with patch_model( with torch_compile_decoration.patch_model(
self.model_runner.model, self.model_runner.model,
bs in self.compile_bs, bs in self.compile_bs,
num_tokens=bs * self.num_tokens_per_bs, num_tokens=bs * self.num_tokens_per_bs,