support using fa4 on deepseek on blackwell (#9928)

python/sglang/srt/entrypoints/engine.py

@@ -666,6 +666,13 @@ def _set_envs_and_config(server_args: ServerArgs):
     if os.environ.get("TRTLLM_ENABLE_PDL", "1") != "0":
         os.environ["TRTLLM_ENABLE_PDL"] = "1"
 
+    if os.environ.get("CUTE_DSL_LOG_LEVEL") is None:
+        # Default to warning level, to avoid too many logs
+        os.environ["CUTE_DSL_LOG_LEVEL"] = "30"
+    if os.environ.get("CUTE_DSL_LOG_TO_CONSOLE") is None:
+        # Need to set log to console, otherwise the log level won't take effect
+        os.environ["CUTE_DSL_LOG_TO_CONSOLE"] = "1"
+
     # Can also be passed as argument
     os.environ["SGLANG_RUN_ID"] = (
         f"sglang-run-{time.time()}-{random.randint(0, 100000000)}"

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

@@ -305,6 +305,7 @@ class FlashAttentionBackend(AttentionBackend):
         speculative_step_id=0,
         topk=0,
         speculative_num_steps=0,
+        fa_impl_ver=3,
     ):
         super().__init__()
 
@@ -338,6 +339,8 @@ class FlashAttentionBackend(AttentionBackend):
         )
         self.speculative_step_id = speculative_step_id
 
+        self.fa_impl_ver = fa_impl_ver
+
         # Local attention settings
         self.attention_chunk_size = (
             model_runner.attention_chunk_size
@@ -712,6 +715,8 @@ class FlashAttentionBackend(AttentionBackend):
 
         # For fa3 interface version compatibility, we put new fields into conditional keyword args
         kwargs = {}
+        if self.fa_impl_ver != 3:
+            kwargs["ver"] = self.fa_impl_ver
         if sinks is not None:
             kwargs["sinks"] = sinks
 
@@ -738,6 +743,7 @@ class FlashAttentionBackend(AttentionBackend):
 
         # Use Flash Attention for prefill
         if not self.use_mla:
+            assert self.fa_impl_ver in [3], "Only FA3 support here"
             # Do multi-head attention
             key_cache, value_cache = forward_batch.token_to_kv_pool.get_kv_buffer(
                 layer.layer_id
@@ -830,6 +836,7 @@ class FlashAttentionBackend(AttentionBackend):
                         softmax_scale=layer.scaling,
                         causal=False,
                         return_softmax_lse=True,
+                        **kwargs,
                     )
                 else:
                     # MHA for extend part of sequence without attending prefix kv cache
@@ -844,6 +851,7 @@ class FlashAttentionBackend(AttentionBackend):
                         softmax_scale=layer.scaling,
                         causal=True,
                         return_softmax_lse=forward_batch.mha_return_lse,
+                        **kwargs,
                     )
                 if forward_batch.mha_return_lse:
                     output, lse, *rest = output
@@ -851,6 +859,7 @@ class FlashAttentionBackend(AttentionBackend):
                     return output, lse
                 return output
             else:
+                assert self.fa_impl_ver in [3], "Only FA3 support here"
                 # Do absorbed multi-latent attention
                 kv_cache = forward_batch.token_to_kv_pool.get_key_buffer(
                     layer.layer_id
@@ -939,6 +948,7 @@ class FlashAttentionBackend(AttentionBackend):
         k_rope: Optional[torch.Tensor] = None,
         sinks: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
+        assert self.fa_impl_ver in [3], "Only FA3 support decoding"
         if k is not None:
             assert v is not None
             if save_kv_cache:
@@ -985,6 +995,8 @@ class FlashAttentionBackend(AttentionBackend):
 
         # For fa3 interface version compatibility, we put new fields into conditional keyword args
         kwargs = {}
+        if self.fa_impl_ver != 3:
+            kwargs["ver"] = self.fa_impl_ver
         if sinks is not None:
             kwargs["sinks"] = sinks
 

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

@@ -21,6 +21,7 @@ class HybridAttnBackend(AttentionBackend):
         self.model_runner = model_runner
         self.prefill_backend = prefill_backend
         self.decode_backend = decode_backend
+        self.data_type = model_runner.kv_cache_dtype
 
     def _select_backend(self, forward_mode: ForwardMode) -> AttentionBackend:
         """

python/sglang/srt/model_executor/model_runner.py

@@ -516,6 +516,7 @@ class ModelRunner:
                     "aiter",
                     "flashinfer",
                     "fa3",
+                    "fa4",
                     "triton",
                     "flashmla",
                     "cutlass_mla",
@@ -1800,6 +1801,15 @@ class ModelRunner:
             )
 
             return FlashAttentionBackend(self)
+        elif backend_str == "fa4":
+            assert (
+                self.use_mla_backend
+            ), "FlashAttention v4 Support is at an early stage, only MLA model supported now"
+            from sglang.srt.layers.attention.flashattention_backend import (
+                FlashAttentionBackend,
+            )
+
+            return FlashAttentionBackend(self, fa_impl_ver=4)
         elif backend_str == "cutlass_mla":
             from sglang.srt.layers.attention.cutlass_mla_backend import (
                 CutlassMLABackend,

python/sglang/srt/models/deepseek_v2.py

@@ -1124,6 +1124,9 @@ class DeepseekV2AttentionMLA(nn.Module):
                 return AttnForwardMethod.MHA_CHUNKED_KV
             else:
                 return _dispatch_mla_subtype()
+        elif attention_backend == "fa4":
+            # TODO(cicirori): use FA4 MHA for DeepSeekV3 for now
+            return AttnForwardMethod.MHA_CHUNKED_KV
         elif attention_backend == "trtllm_mla":
             original_mode = getattr(forward_batch, "_original_forward_mode", None)
             if (

python/sglang/srt/server_args.py

@@ -96,6 +96,7 @@ ATTENTION_BACKEND_CHOICES = [
     # NVIDIA specific
     "cutlass_mla",
     "fa3",
+    "fa4",
     "flashinfer",
     "flashmla",
     "trtllm_mla",

sgl-kernel/python/sgl_kernel/_fa4_interface.py

@@ -4,9 +4,15 @@
 # [2025-07-04] Version in Cute-DSL, for Hopper and Blackwell. You'd need to install nvidia-cutlass-dsl==4.1.0.
 
 
+import copy
+import gc
+import logging
 import math
 from typing import Optional, Tuple
 
+logger = logging.getLogger(__name__)
+
+
 import cuda.bindings.driver as cuda
 import cutlass
 import cutlass.cute as cute
@@ -20,6 +26,22 @@ def maybe_contiguous(x):
     return x.contiguous() if x is not None and x.stride(-1) != 1 else x
 
 
+def _reason_recompile(compile_key, jit_func):
+    compile_cache = jit_func.compile_cache
+    compile_key_map = jit_func.compile_key_map
+    if not compile_cache:
+        return "not compiled yet"
+    for k, v in compile_cache.items():
+        if k == compile_key:
+            continue
+        if len(k) != len(compile_key):
+            continue
+        for i in range(len(k)):
+            if k[i] != compile_key[i]:
+                return f"diff at '{compile_key_map[i]}': {k[i]} vs {compile_key[i]} "
+    return "unknown reason"
+
+
 torch2cute_dtype_map = {
     torch.float16: cutlass.Float16,
     torch.bfloat16: cutlass.BFloat16,
@@ -254,6 +276,9 @@ def _flash_attn_fwd(
         compute_capability,
     )
     if compile_key not in _flash_attn_fwd.compile_cache:
+        logger.info(
+            f"Compiling FA4 kernel with reason: {_reason_recompile(compile_key, _flash_attn_fwd)}"
+        )
         if compute_capability == 9:
             assert page_table is None, "paged KV not supported on SM 9.0"
             # fa_fwd = FlashAttentionForwardSm80(
@@ -335,8 +360,85 @@ def _flash_attn_fwd(
 
 
 _flash_attn_fwd.compile_cache = {}
+_flash_attn_fwd.compile_key_map = [
+    "dtype",
+    "head_dim",
+    "head_dim_v",
+    "qhead_per_kvhead",
+    "causal",
+    "softcap is not None",
+    "lse is None",
+    "cu_seqlens_q is None",
+    "cu_seqlens_k is None",
+    "seqused_q is None",
+    "seqused_k is None",
+    "page_table is not None",
+    "window_size_left is not None",
+    "window_size_right is not None",
+    "learnable_sink is not None",
+    "m_block_size",
+    "n_block_size",
+    "num_threads",
+    "pack_gqa",
+    "compute_capability",
+]
+
+
+def warmup_flash_attn(f):
+    """
+    Decorator for flash_attn_varlen_func:
+    - On the first call, run several warmup passes with different flag combinations
+    - Warmups are executed sequentially to minimize peak GPU memory usage
+    - Does not modify user-provided tensors (clones data)
+    - Easy to extend with more compile-key dimensions
+    """
+    done = False
+
+    def _clone_args(args, kwargs):
+        """Clone tensor arguments to avoid sharing storage; deepcopy for others."""
+
+        def maybe_clone(x):
+            if isinstance(x, torch.Tensor):
+                return x.clone()
+            return copy.deepcopy(x)
+
+        return tuple(maybe_clone(a) for a in args), {
+            k: maybe_clone(v) for k, v in kwargs.items()
+        }
+
+    def _run_warmups(args, kwargs):
+        """Run warmup calls sequentially and release memory after each."""
+        base_args, base_kwargs = _clone_args(args, kwargs)
+
+        # Warmup combinations for return_softmax_lse and causal
+        combos = [
+            dict(return_softmax_lse=False, causal=False),
+            dict(return_softmax_lse=False, causal=True),
+            dict(return_softmax_lse=True, causal=False),
+            dict(return_softmax_lse=True, causal=True),
+        ]
+
+        for combo in combos:
+            wa, wk = _clone_args(base_args, base_kwargs)
+            wk.update(combo)
+            with torch.cuda.stream(torch.cuda.current_stream()):
+                f(*wa, **wk)
+            del wa, wk
+            torch.cuda.empty_cache()
+            gc.collect()
+
+    def wrapper(*args, **kwargs):
+        nonlocal done
+        if not done:
+            logger.info("Running flash_attn_varlen_func warmup passes...")
+            _run_warmups(args, kwargs)
+            done = True
+        return f(*args, **kwargs)
+
+    return wrapper
 
 
+@warmup_flash_attn
 def flash_attn_varlen_func(
     q: torch.Tensor,
     k: torch.Tensor,