Refactor(linear): split LinearBackward kernel into 3 independent kernels#142
Open
chen2021673 wants to merge 3 commits intomasterfrom
Open
Refactor(linear): split LinearBackward kernel into 3 independent kernels#142chen2021673 wants to merge 3 commits intomasterfrom
chen2021673 wants to merge 3 commits intomasterfrom
Conversation
Move grad_flags logic from kernel to autograd layer. The monolithic LinearBackward kernel is replaced by LinearBackwardInput, LinearBackwardWeight, and LinearBackwardBias — each a pure compute operation with no autograd-related parameters.
chen2021673
force-pushed
the
split_linear_backward
branch
from
8f64209 to
66e45dc
Compare
Move needs_input_grad logic from kernel to autograd layer. The monolithic MatmulBackward kernel is replaced by MatmulBackwardInput1 and MatmulBackwardInput2.
chen2021673
force-pushed
the
split_linear_backward
branch
from
66e45dc to
be6eed3
Compare
…ls; rename MatmulBackwardInput1/2 - Add gemm.cuh / gemm.cu: GemmParams struct + GemmCuda() dispatch (cublasGemmEx or cublasGemmStridedBatchedEx based on batch_count), GetCublasHandle(), GetCudaStream() shared across all GEMM-using kernels - Split matmul kernels (CPU + CUDA) out of linear.cc / linear.cu into dedicated matmul.cc / matmul.cu; linear.* now only contains the four Linear kernels - Rename MatmulBackwardInput1 → MatmulBackwardInput, MatmulBackwardInput2 → MatmulBackwardOther for semantic clarity matching MatmulForward(input, other) parameter names - Rewrite outer.cu to use GemmCuda() (OuterForward + bf16 backward paths); keep cublasSgemv for the fp32 backward path (more efficient, bf16 unsupported)
chen2021673
force-pushed
the
split_linear_backward
branch
from
283d083 to
23d301b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概述
完成了 Linear/Matmul/Outer kernel 的架构重构：
核心思路是关注点分离——把是否应当计算的决策从 kernel 层上移到 autograd 层，让 kernel 成为纯计算函数；同时在底层抽象出统一的 GEMM 原语，消除重复的 cuBLAS 调用样板。
具体改动