Add Triton kernel benchmark suite#493

Open

jlamypoirier wants to merge 30 commits intomainfrom

worktree-triton_benchmark

Commits on Apr 30, 2026

Add Triton kernel benchmark suite

jlamypoirier
and
claude
committed
Add sparse_linear benchmark and fix sparse matmul kernel correctness

jlamypoirier
and
claude
committed
Add GRPO loss and sparse copy benchmarks

jlamypoirier
and
claude
committed
Fix sparse_copy benchmark: zero phantom rows before correctness comparison

jlamypoirier
and
claude
committed
Fix sparse_copy benchmark correctness without polluting timing

jlamypoirier
and
claude
committed
Fix benchmark timing loops: precompute backward gradients

jlamypoirier
and
claude
committed
Fix sparse_linear backward_grad: zero phantom rows before precomputing

jlamypoirier
and
claude
committed
Restructure rotary kernel: full-head contiguous load + autotune head_block_size

jlamypoirier
and
claude
committed
Fix Triton rotary kernel: remove autotune, simplify to two-half load

jlamypoirier
and
claude
committed
Revert "Fix Triton rotary kernel: remove autotune, simplify to two-half load"
jlamypoirier
committed
Revert "Restructure rotary kernel: full-head contiguous load + autotune head_block_size"
jlamypoirier
committed
Revert "Revert "Restructure rotary kernel: full-head contiguous load + autotune head_block_size""
jlamypoirier
committed
Revert "Revert "Fix Triton rotary kernel: remove autotune, simplify to two-half load""
jlamypoirier
committed
Restore autotuned rotary kernel, add benchmark warmup

jlamypoirier
and
claude
committed
Drop autotune from rotary kernel, add torch.compile inspection script

jlamypoirier
and
claude
committed
Fix benchmark fairness: pre-allocate work buffer, reset between reps

jlamypoirier
and
claude
committed
Fix fwd_bwd benchmark bias: zero logits.grad between reps for PyTorch variants

jlamypoirier
and
claude
committed
Add fp32 references for sparse_copy and sparse_linear benchmarks

jlamypoirier
and
claude
committed
Fix sparse warmup duplication; expose Apex availability flags publicly

jlamypoirier
and
claude
committed
Fix E741 ambiguous variable name; document dead GRPO masks

jlamypoirier
and
claude
committed
Rename abbreviations and use functools.partial in benchmark suite

jlamypoirier
and
claude
committed
Convert inspect_rotary_compile.py to pathlib

jlamypoirier
and
claude
committed
Add sparse linear autograd tests and benchmark smoke test

jlamypoirier
and
claude
committed
Remove benchmark smoke test

jlamypoirier
and
claude
committed
Fix two nits from rename pass

jlamypoirier
and
claude
committed
Fix bench_fn: call reset() during warmup and calibration phases

jlamypoirier
and
claude
committed
Add shapes parameter to bench run() functions and smoke tests

jlamypoirier
and
claude
committed
Refactor benchmark tests: per-kernel parametrization, single timed rep

jlamypoirier
and
claude
committed
Skip compiled variants in benchmark smoke tests

jlamypoirier
and
claude
committed
Skip sparse benchmarks in Triton interpreter, revert broken histogram patch

jlamypoirier
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Triton kernel benchmark suite#493

Add Triton kernel benchmark suite#493
jlamypoirier wants to merge 30 commits intomainfrom
worktree-triton_benchmark

Commits on Apr 30, 2026

Add Triton kernel benchmark suite

Add sparse_linear benchmark and fix sparse matmul kernel correctness

Add GRPO loss and sparse copy benchmarks

Fix sparse_copy benchmark: zero phantom rows before correctness comparison

Fix sparse_copy benchmark correctness without polluting timing

Fix benchmark timing loops: precompute backward gradients

Fix sparse_linear backward_grad: zero phantom rows before precomputing

Restructure rotary kernel: full-head contiguous load + autotune head_block_size

Fix Triton rotary kernel: remove autotune, simplify to two-half load

Revert "Fix Triton rotary kernel: remove autotune, simplify to two-half load"

Revert "Restructure rotary kernel: full-head contiguous load + autotune head_block_size"

Revert "Revert "Restructure rotary kernel: full-head contiguous load + autotune head_block_size""

Revert "Revert "Fix Triton rotary kernel: remove autotune, simplify to two-half load""

Restore autotuned rotary kernel, add benchmark warmup

Drop autotune from rotary kernel, add torch.compile inspection script

Fix benchmark fairness: pre-allocate work buffer, reset between reps

Fix fwd_bwd benchmark bias: zero logits.grad between reps for PyTorch variants

Add fp32 references for sparse_copy and sparse_linear benchmarks

Fix sparse warmup duplication; expose Apex availability flags publicly

Fix E741 ambiguous variable name; document dead GRPO masks

Rename abbreviations and use functools.partial in benchmark suite

Convert inspect_rotary_compile.py to pathlib

Add sparse linear autograd tests and benchmark smoke test

Remove benchmark smoke test

Fix two nits from rename pass

Fix bench_fn: call reset() during warmup and calibration phases

Add shapes parameter to bench run() functions and smoke tests

Refactor benchmark tests: per-kernel parametrization, single timed rep

Skip compiled variants in benchmark smoke tests

Skip sparse benchmarks in Triton interpreter, revert broken histogram patch