pytorch/benchmark

Pull Requests Commits

Run single iteration when collecting ncu traces

xuzhao9 committed 1 year ago

f2f0b30f

make --dump-ir also dump sass (#2460)

davidberard98 committed 1 year ago

46ab2e2b

Log structured logging overhead to dynamo compile (kinda) (#2454)

jamesjwu committed 1 year ago

9c7aacc4

Increase multiplier to 3 for Inductor AMP FP16 benchmark correctness check (#135932)

jiayisunx committed 1 year ago

c1755f53

use npt.NDArray instead of np.ndarray in type annotations

igorsugak committed 1 year ago

c97859eb

Update mypy to 1.11.2 (#133816)

Skylion007 committed 1 year ago

7e1ba8d5

Remove ignored modes workaround (#135502)

mlazos committed 1 year ago

6b1c40bc

Trace enter/exit of TorchFunctionModes (#135422)

mlazos committed 1 year ago

99ea924e

Trace torch function modes entered outside of torch.compile (#133137)

mlazos committed 1 year ago

2536a65b

SKIP llama for dynamic size testing (#135960)

leslie-fang-intel committed 1 year ago

c709128f

Add remote cache time saved to compilation metrics (#2449)

jamesjwu committed 1 year ago

47413ed5

Support fwd_no_grad mode

bertmaher committed 1 year ago

195d1195

Support nn.Module in cast_tensors

bertmaher committed 1 year ago

f52e79e9

ignore mark_dynamic() in export (#2451)

pianpwk committed 1 year ago

1808fbca

Add HSTU kernel repo (#2452)

xuzhao9 committed 1 year ago

7acad50f

Implement bf16xint16 kernel (#2349)

davidberard98 committed 1 year ago

1e6003dd

tritonbench bf16xint16 matmul template (#2348)

davidberard98 committed 1 year ago

ebd00aac

Fix dynamo benchmark skip logic for cpu device (#135193)

zxd1997066 committed 1 year ago

8351474b

Turn on TMA by default for row-wise GEMM (#2450)

htyu committed 1 year ago

e95390a1

reland map/zip iterator related changes (#135074)

williamwen42 committed 1 year ago

5ee35999

Retire CompileProfiler (#135133)

anijain2305 committed 1 year ago

cdceb782

adding buck file for triton-lang/kernels

adamomainz committed 1 year ago

ac0cf987

adding arc lint opt in for benchmarking (#2448)

adamomainz committed 1 year ago

7bb00692

Support for proxying frozen dataclasses (#134846)

mlazos committed 1 year ago

8c5abe8e

Add tma option for FB8 row-wise GEMM.

htyu committed 1 year ago

71b6e37d

Stop adding useless prefix to error message here, you're pushing the important info off the screen. (#133108)

ezyang committed 1 year ago

a95718f5

Add to torchbench relative path search (#134871)

yiming0416 committed 1 year ago

2927a2b8

adding llama inference specific shapes for flash attention

adamomainz committed 1 year ago

fdb33106

move loop ordering after fusion (#126254)

shunting314 committed 1 year ago

49bcc381

Switch benchmarking to use export non-strict mode (#130977)

desertfire committed 1 year ago

1f4461fb

Newer Older