[Feature] Support cute cpp Encoder FA4 (#7016)

* add cute cpp fa4

* 删掉注释

* 修正合并错误

* sm_version放到函数内

* ci错误

Author: mpgemm

fastdeploy/model_executor/layers/attention/flash_mask_attn_backend.py

@@ -30,6 +30,7 @@
 )
 from fastdeploy.model_executor.layers.attention.ops import (
     append_attention,
+    flash_attn_v4,
     flash_mask_attention,
     get_block_shape_and_split_kv_block,
     gqa_rope_write_cache,
@@ -51,6 +52,8 @@
 else:
     merge_prefill_decode_output = None
 
+from fastdeploy.model_executor.utils import get_sm_version
+
 
 @dataclass
 class FlashMaskAttentionMetadata(AttentionMetadata):
@@ -124,6 +127,7 @@ def __init__(
         if fd_config.speculative_config.model_type != "main":
             self.rope_3d = False
         self.max_partition_size: int = int(os.getenv("FLAGS_max_partition_size", "32768"))
+        self.sm_version = get_sm_version()
 
     def get_kv_cache_shape(
         self,
@@ -278,19 +282,30 @@ def forward_mixed(
                 self.rope_3d,
             )
 
-            flash_mask_attention(
-                q,
-                k,
-                v,
-                forward_meta.cu_seqlens_q,
-                forward_meta.attn_cu_seqlens_k,
-                forward_meta.seq_lens_encoder,
-                res_encoder,
-                forward_meta.attn_mask_offsets,
-                self.num_heads,
-                self.kv_num_heads,
-                self.head_dim,
-            )
+            if self.sm_version >= 100:
+                flash_attn_v4(
+                    q,
+                    k,
+                    v,
+                    forward_meta.cu_seqlens_q,
+                    forward_meta.attn_cu_seqlens_k,
+                    res_encoder,
+                    forward_meta.attn_mask_offsets,
+                )
+            else:
+                flash_mask_attention(
+                    q,
+                    k,
+                    v,
+                    forward_meta.cu_seqlens_q,
+                    forward_meta.attn_cu_seqlens_k,
+                    forward_meta.seq_lens_encoder,
+                    res_encoder,
+                    forward_meta.attn_mask_offsets,
+                    self.num_heads,
+                    self.kv_num_heads,
+                    self.head_dim,
+                )
 
         res_decoder = append_attention(
             qkv,

fastdeploy/model_executor/layers/attention/ops/__init__.py

@@ -15,6 +15,7 @@
 """
 
 from .append_attention import append_attention, append_attention_with_output
+from .flash_attn_v4 import flash_attn_v4
 from .flash_mask_attention import flash_mask_attention
 from .get_attn_mask_q import get_attn_mask_q
 from .get_block_shape_and_split_kv_block import get_block_shape_and_split_kv_block
@@ -33,6 +34,7 @@
     "gqa_rope_write_cache",
     "pre_cache_len_concat",
     "init_kv_signal_per_query",
+    "flash_attn_v4",
     "flash_mask_attention",
     "get_attn_mask_q",
 ]

fastdeploy/model_executor/layers/attention/ops/flash_attn_v4.py

@@ -0,0 +1,39 @@
+"""
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+
+from typing import Optional
+
+import paddle
+
+from fastdeploy.model_executor.utils import get_sm_version
+from fastdeploy.platforms import current_platform
+
+
+def flash_attn_v4(
+    q: paddle.Tensor,
+    k: paddle.Tensor,
+    v: paddle.Tensor,
+    cu_seqlens_q: paddle.Tensor,
+    cu_seqlens_k: paddle.Tensor,
+    attn_out: paddle.Tensor,
+    attn_mask_offsets: Optional[paddle.Tensor] = None,
+):
+    if current_platform.is_cuda() and get_sm_version() >= 100:
+        from blackwell_ops import flash_encoder_attn_fwd
+
+        flash_encoder_attn_fwd(q, k, v, cu_seqlens_q, cu_seqlens_k, attn_out, attn_mask_offsets)
+    else:
+        raise NotImplementedError

tests/operators/test_flash_mask_attn.py

@@ -22,7 +22,10 @@
 from fastdeploy.model_executor.layers.attention.flash_attn_backend import (
     flash_attn_func,
 )
-from fastdeploy.model_executor.layers.attention.ops import get_attn_mask_q
+from fastdeploy.model_executor.layers.attention.ops import (
+    flash_attn_v4,
+    get_attn_mask_q,
+)
 from fastdeploy.model_executor.ops.gpu import flash_mask_attention
 
 
@@ -109,6 +112,65 @@ def test_flash_mask_attention(self):
         max_diff = (paddle_attn_out - naive_attn_out).abs().max().item()
         self.assertLessEqual(max_diff, 0.05)
 
+    def causal_attention_naive(self, q_input, k_input, v_input, cu_seq_q, cu_seq_k):
+        """Causal attention reference implementation for flash_attn_v4 testing."""
+        bsz = cu_seq_q.shape[0] - 1
+        q_token_sum, num_head, head_dim = q_input.shape
+        k_token_sum, num_kv_head, _ = k_input.shape
+        gqa_group_size = num_head // num_kv_head
+        qk_scale = 1 / np.sqrt(head_dim)
+        out = paddle.zeros([num_head, q_token_sum, head_dim], q_input.dtype)
+        for bi in range(bsz):
+            q = q_input[cu_seq_q[bi] : cu_seq_q[bi + 1], :, :].transpose([1, 0, 2]).astype("float32").numpy()
+            k = k_input[cu_seq_k[bi] : cu_seq_k[bi + 1], :, :].transpose([1, 2, 0]).astype("float32").numpy()
+            v = v_input[cu_seq_k[bi] : cu_seq_k[bi + 1], :, :].transpose([1, 0, 2]).astype("float32").numpy()
+            qk = np.matmul(q, np.repeat(k, gqa_group_size, 0))
+            qk *= qk_scale
+            condition = np.tril(np.ones(qk.shape), q.shape[1] - k.shape[2])
+            mask = np.ones(condition.shape).astype("float32") * -1000000
+            qk = np.where(condition > 0, qk, mask)
+            qk_max = qk.max(axis=-1, keepdims=True)
+            qk -= qk_max
+            qk = np.exp(qk)
+            exp_sum = qk.sum(axis=-1, keepdims=True)
+            exp_sum_inv = 1.0 / exp_sum
+            temp_out = paddle.to_tensor(np.matmul(qk, np.repeat(v, gqa_group_size, 0)))
+            out[:, cu_seq_q[bi] : cu_seq_q[bi + 1], :] = temp_out * exp_sum_inv
+        return out.transpose([1, 0, 2])
+
+    def test_flash_encoder_attn_fwd(self):
+        if self.sm_version < 100:
+            self.skipTest("Flash Encoder Attention V4 requires SM100+.")
+
+        q_input = paddle.randn([self.q_len, self.num_head, self.head_dim], dtype="bfloat16")
+        k_input = paddle.randn([self.q_len, self.num_kv_head, self.head_dim], dtype="bfloat16")
+        v_input = paddle.randn(k_input.shape, dtype="bfloat16")
+
+        mask = paddle.arange(self.q_len).astype("int32") + 1
+
+        bsz = self.bsz
+        cu_seq_q = paddle.arange(bsz + 1) * self.q_len
+        cu_seq_k = paddle.arange(bsz + 1) * self.q_len
+        cu_seq_q = cu_seq_q.astype("int32")
+        cu_seq_k = cu_seq_k.astype("int32")
+
+        naive_attn_out = self.causal_attention_naive(q_input, k_input, v_input, cu_seq_q, cu_seq_k)
+
+        paddle_attn_out = paddle.empty(q_input.shape, dtype="bfloat16")
+
+        flash_attn_v4(
+            q_input,
+            k_input,
+            v_input,
+            cu_seq_q,
+            cu_seq_k,
+            paddle_attn_out,
+            mask,
+        )
+
+        max_diff = (paddle_attn_out - naive_attn_out).abs().max().item()
+        self.assertLessEqual(max_diff, 0.05)
+
     def test_fa4(
         self,
     ):