pytorch/benchmark

Pull Requests Commits

Move FBGEMM to e5d0c9448774e6bc577e7f210ecbec56b7a69f10

htyu committed 2 years ago

b27dd7ab

Add asymmetric shapes to test sum Triton kernels (#2310)

jananisriram committed 2 years ago

bbeba0a1

Fix ci by skipping the accuracy test (#2306)

xuzhao9 committed 2 years ago

abed1e7d

Remove torchvision and torchaudio dependency when testing userbenchmark. (#2309)

xuzhao9 committed 2 years ago

f5d3de13

Make CUDA graph benchmarking overridable on a per-op basis

int3 committed 2 years ago

51eca7b7

Add colfax_cutlass backend to flash_attention operator (#2296)

xuzhao9 committed 2 years ago

d5f0a12b

Silently disable default saved tensor hooks during tracing (#123196)

xmfan committed 2 years ago

48223b8f

Add FP8 blockwise triton kernel (#2304)

htyu committed 2 years ago

fc298af7

Add coloring to fx graph print out (#128476)

Chillee committed 2 years ago

0e9c5594

Add --baseline flag to override baseline

int3 committed 2 years ago

19ac8767

Deploy the flash_attention operator CI on H100

xuzhao9 committed 2 years ago

5831be0b

Refactor code for sum Triton kernels (#2303)

jananisriram committed 2 years ago

339ccfde

Extend support to varying block sizes on both dimensions for 2D matrices (#2302)

jananisriram committed 2 years ago

f4cbf782

Fix a metrics reporting issue for FP8 row-wise

htyu committed 2 years ago

10e4bc4a

Always run baseline benchmark first, even with --only

int3 committed 2 years ago

95b4202e

Use cuda graphs for benchmarking

int3 committed 2 years ago

b91caadb

Extend support for x_only = True extra metrics through canonical name retrieval (#2300)

jananisriram committed 2 years ago

86904ca0

Add jagged_sum operator for unpadded nested tensors to TritonBench (#2299)

jananisriram committed 2 years ago

576b2b29

Change the llama max_batch_size larger than default eval batch size (#2283)

leslie-fang-intel committed 2 years ago

a7ae1fb3

Change Dynamo's custom ops warning message to be less spammy (#128456)

zou3519 committed 2 years ago

ab687a03

enable fx graph cache on torchbench (#128239)

masnesral committed 2 years ago

04a1339a

Revert "Make torch_geometric models compatible with export (#123403)" (#128377)

chunyuan-w committed 2 years ago

f29ff892

Add support for reducing across the middle dimension for 3D matrices using the sum Triton kernel (#2297)

jananisriram committed 2 years ago

55c975ea

Compile FBGEMM on H100 (#2298)

xuzhao9 committed 2 years ago

c8d6c2a9

Add support for x_only = True extra metrics during table generation (#2294)

jananisriram committed 2 years ago

217d8f8f

Add FBGEMM submodule (#2293)

xuzhao9 committed 2 years ago

abb45ae0

Add support for reducing across individual dimensions for 2D matrices using the sum Triton kernel (#2295)

jananisriram committed 2 years ago

3ecaae9d

Trace through nn.Module dunder methods for UnspecializedNNModule (#126578)

anijain2305 committed 2 years ago

c13df576

Fix 'get_real_value' on placeholder nodes (#127698)

BowenBao committed 2 years ago

c4b0ffcb

Fix 'get_attr' call in dynamo 'run_node' (#127696)

BowenBao committed 2 years ago

2553812e

Older