0
代码
代码
Issues
Pull Requests
讨论
Wiki
分析
0
cann-robotcann-robotadd mhc
922a02c5创建于 2月25日历史提交
文件最后提交记录最后更新时间
csrc
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
docs
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
kernel
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
test
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
CMakeLists.txt
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
README.md
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
mhc_res_ops.py
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
setup.py
add mhc Co-authored-by: ElevenLiu<liuzhiwen@autokernel.cn> # message auto-generated for no-merge-commit merge: !1589 merge master into master add mhc Created-by: ElevenLiu Commit-by: LiuEleven;ElevenLiu Merged-by: cann-robot Description: ## 描述 面向昇腾 NPU 的 mHC（Manifold-Constrained Hyper-Connections，流形约束超连接）算子 AscendC 实现，此算子由智子芯元 KernelCAT 智能体生成。 ## 关联的Issue 关联Issue #763 ## 测试 ```bash # C++ cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype # Python LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_pre_ops.py ``` ## 使用方式 ```python import mhc_pre_ext x = torch.randn(B * N, S, D, device='npu') # [batch*streams, seq, dim] h = torch.randn(N, device='npu') # [streams] out = mhc_pre_ext.forward(x, h) # [batch, seq, dim] ``` ## 性能（对比 torch.einsum, Ascend 910B2） | 算子 | 加速比 | |----------|--------| | mhc_pre | 24x ~ 52x | | mhc_post | 2x ~ 5x | | mhc_res | 24x ~ 50x | ## 文档更新 更新了README.md文件 ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [x] 文档更新 - [x] 其他，请描述：新增开源贡献算子 See merge request: cann/ops-transformer!15893 个月前
README.md

mhc_res AscendC Operator

Stream mixing via learned weight matrix (residual path in mHC).

Formula

output[b * N + t, seq, d] = Σ_r (h_res[r, t] × x[b * N + r, seq, d])

Equivalent einsum: torch.einsum('brsd,rt->btsd', x.view(B,N,S,D), h_res).reshape(B*N,S,D)

Notes

  • h_res is a static weight matrix [num_streams, num_streams], shared across all batches and token positions
  • This implements the H^res mixing from mHC paper (Sinkhorn-normalized residual connections)
  • Weight normalization is handled upstream, not in this operator

Build

# 1. Build AscendC kernel
mkdir build && cd build
cmake .. -DSOC_VERSION=ascend910b2
make -j

# 2. Build PyTorch C++ extension
cd ..
source /usr/local/Ascend/ascend-toolkit/set_env.sh
python setup.py build_ext --inplace

Test

source /usr/local/Ascend/ascend-toolkit/set_env.sh

# C++ test
cd build && LD_LIBRARY_PATH=./lib:$LD_LIBRARY_PATH ./test_multi_dtype

# Python test
LD_LIBRARY_PATH=./build/lib:$LD_LIBRARY_PATH python mhc_res_ops.py

API

// C++ kernel
extern "C" void mhc_res_do_fp32(uint32_t blockDim, void* stream,
    uint8_t* input, uint8_t* h_res, uint8_t* output,
    int64_t batch, int64_t seq_len, int64_t dim, int64_t num_streams);

# Python (via C++ extension)
import mhc_res_ext
output = mhc_res_ext.forward(x, h_res)  # x on NPU, returns NPU tensor

# Or use wrapper
from mhc_res_ops import mhc_res, mhc_res_einsum
output = mhc_res(x, h_res)

Performance

On Ascend 910B2, mhc_res achieves 24x~50x speedup over torch.einsum.