Apply deepseek cuda rope

.github/workflows/pr-test-amd.yml

@@ -95,6 +95,7 @@ jobs:
           docker exec ci_sglang pip uninstall sgl-kernel -y || true
           docker exec -w /sglang-checkout/sgl-kernel ci_sglang bash -c "rm -f pyproject.toml && mv pyproject_rocm.toml pyproject.toml && python3 setup_rocm.py install"
           docker exec ci_sglang pip install -e "python[dev_hip]"
+          docker exec ci_sglang pip install py-spy || true
 
           docker exec -w / ci_sglang git clone https://github.com/merrymercy/human-eval.git
           docker exec -w /human-eval ci_sglang pip install -e .

python/sglang/srt/layers/rotary_embedding.py

@@ -645,7 +645,10 @@ def _compute_cos_sin_cache(self) -> torch.Tensor:
         cache = torch.cat((cos, sin), dim=-1)
         return cache
 
-    def forward(
+    def forward_hip(self, *args, **kwargs):
+        return self.forward_native(*args, **kwargs)
+
+    def forward_native(
         self,
         positions: torch.Tensor,
         query: torch.Tensor,