[Ops][Feature] Add fused FFN linear op (ffn_linear) for Ascend950

Signed-off-by: chenchris2 <1349418798@qq.com>

Author: chenchris2

csrc/build_aclnn.sh

@@ -225,6 +225,7 @@ elif [[ "$SOC_VERSION" =~ ^ascend950 ]]; then
         "k2q_csr"
         "sparse_attention_score"
         "mla_prolog_v3"
+        "ffn"
     )
 
     CUSTOM_OPS=$(IFS=';'; echo "${CUSTOM_OPS_ARRAY[*]}")

csrc/ffn/CMakeLists.txt

@@ -0,0 +1,18 @@
+# -----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# -----------------------------------------------------------------------------------------------------------
+
+file(GLOB SUBDIRECTORIES LIST_DIRECTORIES true RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
+# 遍历子目录
+foreach(SUBDIR ${SUBDIRECTORIES})
+  # 检查子目录中是否存在 CMakeLists.txt
+  if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${SUBDIR}/CMakeLists.txt)
+    add_subdirectory(${SUBDIR})
+  endif()
+endforeach()

csrc/ffn/README.md

@@ -0,0 +1,55 @@
+# FFN(arch35 fused variant)
+
+`y = act(x @ W1^T + b1) @ W2^T + b2`,act ∈ {gelu, silu, swiglu},bf16/fp16。
+
+本目录为 **Ascend950PR(A5/arch35)融合实现**。上游通用 FFN 算子见
+[ops-transformer 原仓](https://gitcode.com/cann/ops-transformer/tree/master/ffn/ffn)。
+
+| 产品 | 是否支持 |
+| :--- | :---: |
+| Ascend 950PR / 950DT(A5, arch35) | √ |
+| Atlas A2/A3、310P 等 | 需另行适配(本树仅含 arch35 kernel) |
+
+## 构建
+
+算子已加入 `csrc/build_aclnn.sh` 的 `ascend950` 列表,随仓库标准链路构建:
+
+```bash
+# 路径 A(全链路,需与仓 requirements 匹配的 torch/torch_npu 环境):
+export SOC_VERSION=ascend950
+pip install -e .
+
+# 路径 B(单算子,调试用):
+cd csrc && bash build.sh --pkg --ops=ffn --soc=ascend950 -j8
+```
+
+## 调用
+
+```python
+import torch
+import torch_npu
+from vllm_ascend.utils import enable_custom_op
+
+enable_custom_op()
+
+y = torch.ops._C_ascend.ffn_linear(
+    x,           # [M,K] bf16
+    weight1,     # Linear 布局 [H,K](swiglu 为 [2H,K]);canonical [K,H] gelu/silu 亦支持
+    weight2,     # [N,H]
+    bias1, bias2,# 可选(swiglu 无 bias)
+    "gelu",      # "gelu" | "silu" | "swiglu"
+    0,           # inner_precise
+)
+```
+
+Meta kernel 已注册(`torch_binding_meta.cpp`),支持 symbolic tracing / ACLGraph capture。
+精度测试:`tests/e2e/nightly/single_node/ops/singlecard_ops/test_ffn_linear.py`
+(含 meta 契约、确定性、性能 smoke)。950PR 上对标 bf16 小算子链有性能收益。
+
+## 已知限制
+
+- swiglu 仅支持 Linear 布局 weight1 `[2H,K]`;gelu/silu 支持 canonical `[K,H]` 与
+  linear `[N,K]` 双布局;**全方阵(H==K==N)歧义默认 linear**
+  (PyTorch Linear 惯例,与独立 ops-transformer 仓有意分叉)。
+- 符号解析:3rd matmul tiling 头引用的 `Ops::Base::ToString` 按 csrc 框架惯例链接
+  CANN `libops_base.so`(FindOPBASE)解析。

csrc/ffn/ffn/CMakeLists.txt

@@ -0,0 +1,19 @@
+# -----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# -----------------------------------------------------------------------------------------------------------
+
+file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
+if(NOT ENABLE_TEST AND NOT BENCHMARK)
+    list(REMOVE_ITEM CURRENT_DIRS tests)
+endif()
+foreach(SUB_DIR ${CURRENT_DIRS})
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
+        add_subdirectory(${SUB_DIR})
+    endif()
+endforeach()

csrc/ffn/ffn/README.md

@@ -0,0 +1,14 @@
+# FFN(arch35 fused variant)
+
+`y = act(x @ W1^T + b1) @ W2^T + b2`,act ∈ {gelu, silu, swiglu},bf16/fp16。
+
+本目录为 **Ascend950PR(A5/arch35)融合实现**的单算子源码树(op_host / op_kernel,
+含自持 3rd mat_mul_v3 拷贝)。上游通用 FFN 算子(支持矩阵、MoE/量化等完整功能)见
+[ops-transformer 原仓](https://gitcode.com/cann/ops-transformer/tree/master/ffn/ffn)。
+
+| 产品 | 是否支持 |
+| :--- | :---: |
+| Ascend 950PR / 950DT(A5, arch35) | √ |
+| Atlas A2/A3、310P 等 | 需另行适配(本树仅含 arch35 kernel) |
+
+构建/调用/基线/限制见[上级 README](../README.md) 与 `docs/`(aclnn API 文档、设计介绍)。