add gfx938

CMakeLists.txt

@@ -42,7 +42,7 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12" "3.13")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx906;gfx926;gfx928;gfx936")
+set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx906;gfx926;gfx928;gfx936;gfx938")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.

vllm/platforms/rocm.py

@@ -132,7 +132,7 @@ def on_mi3xx() -> bool:
 @cache
 def on_gfx9() -> bool:
     GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
-    return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936"])
+    return any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936", "gfx938"])
 
 
 @cache
@@ -154,7 +154,7 @@ def use_rocm_custom_paged_attention(
         sinks: Optional[torch.Tensor] = None) -> bool:
 
     GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
-    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936"])
+    ON_GFX9 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942", "gfx950", "gfx928", "gfx936", "gfx938"])
     ON_GFX11_GFX12 = any(arch in GPU_ARCH for arch in ["gfx11", "gfx12"])
 
     # custom paged attn always supported on V0. On V1, requires sliding window

vllm/utils/__init__.py

@@ -93,7 +93,7 @@ POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
 MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
 
 GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
-SUPPORT_TC = any(arch in GPU_ARCH for arch in ["gfx928", "gfx936"])
+SUPPORT_TC = any(arch in GPU_ARCH for arch in ["gfx928", "gfx936", "gfx938"])
 
 # Constants related to forcing the attention backend selection