[feat] support fa3 backend for pd disaggregated (#2695)

* support fa3 backend run in pd disaggregated

* support fa3 backend run in pd disaggregated

* support fa3 backend run in pd disaggregated

* support fa3 backend run in pd disaggregated

* delete use_fast_ffn

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

@@ -17,10 +17,16 @@
 from .append_attention import append_attention
 from .get_block_shape_and_split_kv_block import \
     get_block_shape_and_split_kv_block
+from .gqa_rope_write_cache import gqa_rope_write_cache
 from .init_signal_layerwise import init_signal_layerwise
 from .open_shm_and_get_meta_signal import open_shm_and_get_meta_signal
+from .pre_cache_len_concat import pre_cache_len_concat
 
 __all__ = [
-    "get_block_shape_and_split_kv_block", "append_attention",
-    "open_shm_and_get_meta_signal", "init_signal_layerwise"
+    "get_block_shape_and_split_kv_block",
+    "append_attention",
+    "open_shm_and_get_meta_signal",
+    "init_signal_layerwise",
+    "gqa_rope_write_cache",
+    "pre_cache_len_concat",
 ]

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

@@ -0,0 +1,66 @@
+"""
+# 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.platforms import current_platform
+
+
+def gqa_rope_write_cache(
+        qkv: paddle.Tensor,
+        key_cache: paddle.Tensor,
+        value_cache: paddle.Tensor,
+        cu_seqlens_q: paddle.Tensor,
+        cu_seqlens_k: paddle.Tensor,
+        rotary_embs: paddle.Tensor,
+        seq_lens_this_time: paddle.Tensor,
+        seq_lens_encoder: paddle.Tensor,
+        seq_lens_decoder: paddle.Tensor,
+        padding_offsets: paddle.Tensor,
+        cum_offsets: paddle.Tensor,
+        block_tables: paddle.Tensor,
+        kv_batch_ids: paddle.Tensor,
+        kv_tile_ids_per_batch: paddle.Tensor,
+        kv_num_blocks: paddle.Tensor,
+        cache_batch_ids: paddle.Tensor,
+        cache_tile_ids_per_batch: paddle.Tensor,
+        cache_num_blocks: paddle.Tensor,
+        cache_k_quant_scales: Optional[paddle.Tensor] = None,
+        cache_v_quant_scales: Optional[paddle.Tensor] = None,
+        cache_k_dequant_scales: Optional[paddle.Tensor] = None,
+        cache_v_dequant_scales: Optional[paddle.Tensor] = None,
+        cache_k_zp: Optional[paddle.Tensor] = None,
+        cache_v_zp: Optional[paddle.Tensor] = None,
+        kv_signal_data: Optional[paddle.Tensor] = None,
+        kv_token_num: int = 1,
+        max_seq_len: int = 0,
+        cache_quant_type: str = "none"):
+    if current_platform.is_cuda():
+        from fastdeploy.model_executor.ops.gpu import gqa_rope_write_cache
+        q, k, v, qkv_ = gqa_rope_write_cache(
+            qkv, key_cache, value_cache, cu_seqlens_q, cu_seqlens_k,
+            rotary_embs, seq_lens_this_time, seq_lens_encoder,
+            seq_lens_decoder, padding_offsets, cum_offsets, block_tables,
+            kv_batch_ids, kv_tile_ids_per_batch, kv_num_blocks,
+            cache_batch_ids, cache_tile_ids_per_batch, cache_num_blocks,
+            cache_k_quant_scales, cache_v_quant_scales, cache_k_dequant_scales,
+            cache_v_dequant_scales, cache_k_zp, cache_v_zp, kv_signal_data,
+            kv_token_num, max_seq_len, cache_quant_type)
+        return q, k, v, qkv_
+    else:
+        raise NotImplementedError()

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

@@ -0,0 +1,36 @@
+"""
+# 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.
+"""
+
+import paddle
+
+from fastdeploy.platforms import current_platform
+
+
+def pre_cache_len_concat(seq_lens_decoder: paddle.Tensor,
+                         seq_lens_this_time: paddle.Tensor,
+                         max_dec_len: int = 0,
+                         block_size: int = 64):
+    if current_platform.is_cuda():
+        from fastdeploy.model_executor.ops.gpu import pre_cache_len_concat
+        out = pre_cache_len_concat(seq_lens_decoder, seq_lens_this_time,
+                                   max_dec_len, block_size)
+        return out
+    else:
+        raise NotImplementedError()