【训练营】BF16核函数优化 - #233

Open

xiaoerlang0359-debug wants to merge 19 commits into

InfiniTensor:masterfrom

xiaoerlang0359-debug:master

Commits on Sep 20, 2026

feat(profiling): add NVTX instrumentation with baseline nsys captures
xiaoerlang0359-debug
committed
34a52a8
chore: ignore local build dirs and drop the unused DeepGEMM gitlink
xiaoerlang0359-debug
committed
1ed5238
chore(scripts): download example assets from hf-mirror
xiaoerlang0359-debug
committed
5c101fb
perf(cuda): serve bf16 weights from Adam-fused shadow copies
xiaoerlang0359-debug
committed
fe5591a
perf(cuda): vectorize Adam memory access
xiaoerlang0359-debug
committed
5754c7e
docs: record kernel optimization analysis and bf16 reference results
xiaoerlang0359-debug
committed
c07805a
perf(cuda): reduce CrossEntropy loss on-device to remove D2H sync
xiaoerlang0359-debug
committed
f09ca48
perf(cuda): drop dead Fill(0) before full-coverage transform kernels
xiaoerlang0359-debug
committed
3bbdd76
perf(cuda): specialize Fill(0) to cudaMemsetAsync off the launch path
xiaoerlang0359-debug
committed
9c1947f
perf(cuda): pass SliceForward metadata by value via kernel parameter space
xiaoerlang0359-debug
committed
c2b2de3
perf(cuda): drop dead grad_a Fill(0) in binary broadcast backward
xiaoerlang0359-debug
committed
384dfc5
perf(cuda): fuse RMSNorm into single forward/backward kernels
xiaoerlang0359-debug
committed
5a50c0c
perf(cuda): update only hit rows in embedding backward and Adam
xiaoerlang0359-debug
committed
6a0b178
style: run clang-format and translate the remaining comments to English
xiaoerlang0359-debug
committed
117be73
chore: ignore nsys/out profiling artifacts
xiaoerlang0359-debug
committed
e152c84
docs: add the nsys timeline figure embedded by the kernel report
xiaoerlang0359-debug
committed
ee39356
docs: split the detailed nsys timeline out of the kernel report
xiaoerlang0359-debug
committed
694c953
docs: restructure the kernel report and add the timeline figures
xiaoerlang0359-debug
committed
178ce0c
fix(cuda): align the mixed binary backward with the logical warp refactor
xiaoerlang0359-debug
committed
fd7d522

Navigation Menu

【训练营】BF16核函数优化 - #233

【训练营】BF16核函数优化#233
xiaoerlang0359-debug wants to merge 19 commits into
InfiniTensor:masterfrom
xiaoerlang0359-debug:master

Commits on Sep 20, 2026

feat(profiling): add NVTX instrumentation with baseline nsys captures

chore: ignore local build dirs and drop the unused DeepGEMM gitlink

chore(scripts): download example assets from hf-mirror

perf(cuda): serve bf16 weights from Adam-fused shadow copies

perf(cuda): vectorize Adam memory access

docs: record kernel optimization analysis and bf16 reference results

perf(cuda): reduce CrossEntropy loss on-device to remove D2H sync

perf(cuda): drop dead Fill(0) before full-coverage transform kernels

perf(cuda): specialize Fill(0) to cudaMemsetAsync off the launch path

perf(cuda): pass SliceForward metadata by value via kernel parameter space

perf(cuda): drop dead grad_a Fill(0) in binary broadcast backward

perf(cuda): fuse RMSNorm into single forward/backward kernels

perf(cuda): update only hit rows in embedding backward and Adam

style: run clang-format and translate the remaining comments to English

chore: ignore nsys/out profiling artifacts

docs: add the nsys timeline figure embedded by the kernel report

docs: split the detailed nsys timeline out of the kernel report

docs: restructure the kernel report and add the timeline figures

fix(cuda): align the mixed binary backward with the logical warp refactor