feat: use fa3 mla by default on hopper (#5210)

Co-authored-by: yundai424 <yundai424@gmail.com>
Co-authored-by: hebiao064 <hebiaobuaa@gmail.com>

python/sglang/srt/layers/attention/flashattention_backend.py

@@ -325,7 +325,7 @@ class FlashAttentionBackend(AttentionBackend):
         batch_size = len(seqlens_in_batch)
         device = seqlens_in_batch.device
 
-        if forward_batch.forward_mode.is_decode():
+        if forward_batch.forward_mode.is_decode_or_idle():
             # Draft Decode
             if forward_batch.spec_info is not None:
                 metadata.cache_seqlens_int32 = (
@@ -527,7 +527,9 @@ class FlashAttentionBackend(AttentionBackend):
             else (-1, -1)
         )
         k_descale, v_descale = None, None
-        if self.kv_cache_dtype_str != "auto":
+        # only use kv scaling if: 1) fp8 kv is explicitly enabled, 2) RadixAttention
+        # has corresponding quantization method so that layer.k_scale is not None
+        if self.kv_cache_dtype_str != "auto" and layer.k_scale is not None:
             descale_shape = (forward_batch.batch_size, layer.tp_k_head_num)
             k_descale = layer.k_scale.expand(descale_shape)
             v_descale = layer.v_scale.expand(descale_shape)
@@ -670,10 +672,13 @@ class FlashAttentionBackend(AttentionBackend):
         causal = not layer.is_cross_attention
 
         k_descale, v_descale = None, None
+        # only use kv scaling if: 1) fp8 kv is explicitly enabled, 2) RadixAttention
+        # has corresponding quantization method so that layer.k_scale is not None
         if self.kv_cache_dtype_str != "auto":
-            descale_shape = (forward_batch.batch_size, layer.tp_k_head_num)
-            k_descale = layer.k_scale.expand(descale_shape)
-            v_descale = layer.v_scale.expand(descale_shape)
+            if layer.k_scale is not None:
+                descale_shape = (forward_batch.batch_size, layer.tp_k_head_num)
+                k_descale = layer.k_scale.expand(descale_shape)
+                v_descale = layer.v_scale.expand(descale_shape)
             q = q.to(self.kv_cache_dtype)
 
         if not self.use_mla:
@@ -834,7 +839,7 @@ class FlashAttentionBackend(AttentionBackend):
         """Initialize forward metadata for capturing CUDA graph."""
         metadata = FlashAttentionMetadata()
         device = seq_lens.device
-        if forward_mode.is_decode():
+        if forward_mode.is_decode_or_idle():
             if spec_info is not None:
                 # Draft Decode
                 metadata.cache_seqlens_int32 = self.decode_cuda_graph_metadata[
@@ -937,7 +942,7 @@ class FlashAttentionBackend(AttentionBackend):
         seq_lens = seq_lens[:bs]
         seq_lens_cpu = seq_lens_cpu[:bs]
         req_pool_indices = req_pool_indices[:bs]
-        if forward_mode.is_decode():
+        if forward_mode.is_decode_or_idle():
             metadata = self.decode_cuda_graph_metadata[bs]
 
             if spec_info is not None:

python/sglang/srt/model_executor/model_runner.py

@@ -80,6 +80,7 @@ from sglang.srt.utils import (
     is_cuda,
     is_flashinfer_available,
     is_hip,
+    is_hopper_with_cuda_12_3,
     monkey_patch_p2p_access_check,
     monkey_patch_vllm_gguf_config,
     set_cpu_offload_max_bytes,
@@ -245,7 +246,16 @@ class ModelRunner:
                     "flashinfer" if is_flashinfer_available() else "triton"
                 )
             else:
-                server_args.attention_backend = "triton"
+                if is_hopper_with_cuda_12_3():
+                    if server_args.speculative_eagle_topk is None or (
+                        server_args.speculative_eagle_topk is not None
+                        and server_args.speculative_eagle_topk == 1
+                    ):
+                        server_args.attention_backend = "fa3"
+                    else:
+                        server_args.attention_backend = "triton"
+                else:
+                    server_args.attention_backend = "triton"
             logger.info(
                 f"Attention backend not set. Use {server_args.attention_backend} backend by default."
             )
@@ -263,6 +273,16 @@ class ModelRunner:
             else:
                 raise ValueError(f"MLA optimization not supported on CPU.")
 
+        if (
+            server_args.attention_backend == "fa3"
+            and server_args.kv_cache_dtype == "fp8_e5m2"
+        ):
+            logger.warning(
+                "FlashAttention3 only supports fp8_e4m3 if using FP8; "
+                "Setting attention backend to triton."
+            )
+            server_args.attention_backend = "triton"
+
         if server_args.enable_double_sparsity:
             logger.info(
                 "Double sparsity optimization is turned on. Use triton backend without CUDA graph."
@@ -889,9 +909,6 @@ class ModelRunner:
                 "FlashAttention v3 Backend requires SM>=90. "
                 "Please use `--attention-backend flashinfer`."
             )
-            logger.warning(
-                "FlashAttention v3 Backend is in Beta. FP8 is not supported."
-            )
             from sglang.srt.layers.attention.flashattention_backend import (
                 FlashAttentionBackend,
             )

python/sglang/srt/utils.py

@@ -1828,3 +1828,12 @@ def fast_topk(values, topk, dim):
     else:
         # Use topk for efficiency with larger k values
         return torch.topk(values, topk, dim=dim)
+
+
+def is_hopper_with_cuda_12_3():
+    if not is_cuda():
+        return False
+    is_hopper = torch.cuda.get_device_capability()[0] == 9
+    cuda_version = torch.version.cuda.split(".")
+    is_cuda_compatible = int(cuda_version[0]) == 12 and int(cuda_version[1]) >= 3
+    return is_hopper and is_cuda_compatible