[ROCm] Enable fp8 query for Triton attention (#60296)

Signed-off-by: xawang <xawang@amd.com>
Co-authored-by: JartX <sagformas@epdcenter.es>

tests/model_executor/model_loader/test_reload.py

@@ -250,16 +250,17 @@ def test_attention_sink_layerwise_reload(monkeypatch, tp_rank, padding):
 def _fp8_reload_unsupported() -> bool:
     """Whether the FP8 reload/online-quantize tests should be skipped.
 
-    ``supports_fp8()`` returns True on MI250 (gfx90a) because the general
-    quantization paths upcast FP8 weights, but gfx90a has no native FP8 and
-    cannot run these reload models, so treat it as unsupported here.
+    ``supports_fp8()`` returns True on MI250 (gfx90a) and RDNA3 (gfx11)
+    because the general quantization paths upcast FP8 weights, but neither has
+    native FP8 and cannot run these reload models, so treat them as
+    unsupported here.
     """
     if not current_platform.supports_fp8():
         return True
     if current_platform.is_rocm():
-        from vllm.platforms.rocm import on_gfx90a
+        from vllm.platforms.rocm import on_gfx11, on_gfx90a
 
-        return on_gfx90a()
+        return on_gfx90a() or on_gfx11()
     return False
 
 

vllm/platforms/rocm.py

@@ -1191,7 +1191,7 @@ class RocmPlatform(Platform):
 
     @classmethod
     def supports_fp8(cls) -> bool:
-        return on_cdna() or on_rdna4()
+        return on_cdna() or on_rdna()
 
     @classmethod
     def is_fp8_fnuz(cls) -> bool:

vllm/v1/attention/backends/triton_attn.py

@@ -549,6 +549,11 @@ class TritonAttentionImpl(AttentionImpl):
         self.use_alibi_sqrt = use_alibi_sqrt
         self.chunk_lookback = chunk_lookback
         self.supports_quant_query_input = current_platform.is_cuda()
+        if current_platform.is_rocm():
+            # Enable GPUs with fp8 MMA support
+            from vllm.platforms.rocm import on_gfx11, on_gfx90a
+
+            self.supports_quant_query_input = not (on_gfx90a() or on_gfx11())
 
         self._kv_quant_mode = get_kv_quant_mode(kv_cache_dtype)
         self._is_per_token_head_quant = self._kv_quant_mode.is_per_token_head