[XPU] Fix decode graph runner is_current_stream_capturing on non-CUDA devices (#35050)

docker/xpu.Dockerfile

@@ -74,6 +74,7 @@ RUN pip install --no-cache-dir torch==2.13.0+xpu torchvision==0.28.0+xpu torchau
 
 RUN echo "Cloning ${SG_LANG_BRANCH} from ${SG_LANG_REPO}" && \
     git clone --branch ${SG_LANG_BRANCH} --single-branch ${SG_LANG_REPO} sglang && \
+    git -C sglang fetch --tags --force origin && \
     cd sglang && cd python && \
     cp pyproject_xpu.toml pyproject.toml && \
     pip install --no-cache-dir ".[dev,diffusion]" --extra-index-url https://download.pytorch.org/whl/xpu && \

python/sglang/srt/model_executor/runner/decode_cuda_graph_runner.py

@@ -467,9 +467,11 @@ class DecodeCudaGraphRunner(BaseCudaGraphRunner):
     def _record_in_graph_metadata_prep_done(self):
         # Purely a marker at this point in the graph; where the shared reads
         # actually end is the attn backend's call.
-        if not torch.cuda.is_current_stream_capturing():
+        if not self.device_module.is_current_stream_capturing():
             # Warmup shares this body. Breakable capture still plants: it opens
             # segment 1 on context entry and every segment re-arms the node.
+            # Routed through device_module so XPU (torch.xpu) is picked up
+            # instead of hitting torch.cuda dummy stubs on non-CUDA builds.
             return
         if self.in_graph_metadata_prep_done is None:
             self.in_graph_metadata_prep_done = make_external_event(self.device_module)