[AMD] Fix Llama 4 FP8 accuracy issues on MI300X (#7699)

2025-08-22 13:13:45 -07:00
parent e5638573c1
commit f445a1d9a3
5 changed files with 212 additions and 17 deletions
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -2228,7 +2228,10 @@ class ServerArgs:
                # use bf16 for mxfp4 triton kernels
                self.dtype = "bfloat16"
        elif "Llama4" in model_arch:
-            assert self.attention_backend == "fa3", "fa3 is required for Llama4 model"
+            assert self.attention_backend in {
+                "fa3",
+                "aiter",
+            }, "fa3 or aiter is required for Llama4 model"
        elif model_arch in [
            "Gemma2ForCausalLM",
            "Gemma3ForCausalLM",