Commits

Branch selector

 main

User selector

derdrdirk

Datepicker

All time

Commit History

Commits on May 22, 2025

• [XLA:GPU] Implement ApplyConfig for CustomKernel autotuner backend.

  Show description for 5b66909
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  5b66909

  Copy full SHA for 5b66909

Commits on May 21, 2025

• [XLA:GPU] Implement ApplyConf for CublasLt autotuner backend.

  Show description for 1bc9654
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  1bc9654

  Copy full SHA for 1bc9654

• [XLA:GPU] Add ApplyConfig method to the Cublas autotuner backend.

  Show description for 28ff4c3
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  28ff4c3

  Copy full SHA for 28ff4c3

• [XLA:GPU] Call CublasBackend, CublasLtBackend and CustomKernelBackend from FissionBackend. Return BackendConfigs instead of HloModules.

  Show description for 47a4714
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  47a4714

  Copy full SHA for 47a4714

Commits on May 19, 2025

• [XLA:GPU] CuDNN autotuner backend.

  Show description for 2de87d0
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  2de87d0

  Copy full SHA for 2de87d0

Commits on May 16, 2025

• [XLA:GPU] Implement autotuner fission backend. The backend can unfuse fusion instructions into HloModules with subsituted cublas, cublasLt and custom kernels calls.

  Show description for 02a29ed
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  02a29ed

  Copy full SHA for 02a29ed

Commits on May 15, 2025

• [XLA:GPU] Restrict CustomKernel autotuning backend to fusion instructions with a fusion_backend_config. Return kernel_index=0 as default kernel. Clean up unit test hlo.

  Show description for 6d1c195
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  6d1c195

  Copy full SHA for 6d1c195

Commits on May 14, 2025

• [XLA:GPU] Use Redzone buffers in cublas autotuning backend.

  Show description for 3fa8a10
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  3fa8a10

  Copy full SHA for 3fa8a10

Commits on May 12, 2025

• [XLA:GPU] Only support cuBLAS custom calls in cuBLAS autotuning backend.

  Show description for 1dee1b6
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  1dee1b6

  Copy full SHA for 1dee1b6

• [XLA:GPU] Add custom kernel autotuning backend.

  Show description for 470cd5b
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  470cd5b

  Copy full SHA for 470cd5b

Commits on May 9, 2025

• [XLA:GPU] Add cuBLASLt autotuner backend. The backend returns configs for cuBLASLt custom calls.

  Show description for db0f2cd
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  db0f2cd

  Copy full SHA for db0f2cd

• [XLA:GPU] Add absl::StatusOr to GetSupportedConfigs for Autotuning backends. StatusOr will help in writing cleaner code for early returns.

  Show description for dfded7b
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  dfded7b

  Copy full SHA for dfded7b

Commits on May 7, 2025

• [XLA:GPU] Refactor RedzoneBuffers. Make pararameters of RedzoneBuffers explicit and decouple RedzoneBuffers from AutotunerCompileUtil and AutotunerCompileConfig.

  Show description for bfeacf2
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  bfeacf2

  Copy full SHA for bfeacf2

Commits on May 6, 2025

• [XLA:GPU] Add CuBLAS autotuner backend.

  Show description for 86c64c6
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  86c64c6

  Copy full SHA for 86c64c6

Commits on Oct 15, 2024

• Fix internal test and roll-forward PR #16975: Add a few related optimization passes for fp8 gemm custom-calls.

  Show description for e099ee1
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  e099ee1

  Copy full SHA for e099ee1

Commits on Oct 9, 2024

• Remove C++ template requirement from PTX CustomKernel.

  Show description for 5246600
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  5246600

  Copy full SHA for 5246600

Commits on Oct 8, 2024

• Add PTX CustomKernel.

  Show description for f33223c
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  f33223c

  Copy full SHA for f33223c

Commits on Sep 30, 2024

• Refactor gemm_fusion_autotuner fusion rewriter nested if-else to use early return pattern.

  Show description for aaf7036
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  aaf7036

  Copy full SHA for aaf7036

Commits on Sep 24, 2024

• Add custom kernel fusion to gemm fusion autotuner.

  Show description for d893dd6
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  d893dd6

  Copy full SHA for d893dd6

Commits on Sep 19, 2024

• Add custom kernel fusion to gemm fusion autotuner.

  Show description for 7f1e216
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  7f1e216

  Copy full SHA for 7f1e216

• Allow CustomKernelFusionRewriter to manualy specify the kernel index.

  Show description for 45dea71
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  45dea71

  Copy full SHA for 45dea71

Commits on Sep 16, 2024

• Reverts 67325a0f76c0e989c2648dc57b13fbf2884f995b

  Show description for 1765b13
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  1765b13

  Copy full SHA for 1765b13

Commits on Sep 13, 2024

• Add custom kernel fusion to gemm fusion autotuner.

  Show description for 67325a0
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  67325a0

  Copy full SHA for 67325a0

Commits on Sep 12, 2024

• Add comments to GemmFusionAutotuner and rename TilingConfig to BackendConfig.

  Show description for 4dc1333
  

  derdrdirk

  authored and

  Google-ML-Automation

  committed
  4dc1333

  Copy full SHA for 4dc1333

Commits on Sep 11, 2024

• Remove unused BF16xBF16ToBF16 Ampere and Hopper cutlass kernels.

  Show description for a9e61a1
  

  derdrdirk

  authored and

  copybara-github

  committed
  a9e61a1

  Copy full SHA for a9e61a1

• Add lookup table for Cutlass upcast gemm kernel selection.

  Show description for 4d7cbd1
  

  derdrdirk

  authored and

  copybara-github

  committed
  4d7cbd1

  Copy full SHA for 4d7cbd1

Commits on Sep 4, 2024

• Default enable custom kernel fusions for V100 GPUs and test that we run triton gemms for using A100 GPUS and custom kernel fusions using V100.

  Show description for 1060b6b
  

  derdrdirk

  authored and

  copybara-github

  committed
  1060b6b

  Copy full SHA for 1060b6b

Commits on Sep 3, 2024

• Remove BF16xF32ToF32 Cutlass kernel. The kernel is not required for the model performance we optimize for.

  Show description for 13ab160
  

  derdrdirk

  authored and

  copybara-github

  committed
  13ab160

  Copy full SHA for 13ab160

• Restrict Custom Kernel Fusion pattern matching for upcast cutlass kernels to row major layouts.

  Show description for df276fb
  

  derdrdirk

  authored and

  copybara-github

  committed
  df276fb

  Copy full SHA for df276fb

Commits on Aug 21, 2024

• Restrict Custom Kernel Fusion Cutlass upcast Pattern matcher to a list of precompiled kernels.

  Show description for 0ff3b27
  

  derdrdirk

  authored and

  copybara-github

  committed
  0ff3b27

  Copy full SHA for 0ff3b27

Commits on Aug 6, 2024

• [XLA:GPU] Expose Cutlass GemmUniversalMode and batch size arguments. This enables the use of split-k gemms in custom fusions.

  Show description for 5978d3f
  

  derdrdirk

  authored and

  copybara-github

  committed
  5978d3f

  Copy full SHA for 5978d3f

Commits on Aug 2, 2024

• [XLA:GPU] Canonicalize type conversions for custom kernel fusions and add bf16xs8_to_f32 upcast gemm kernel.

  Show description for 13a6b00
  

  derdrdirk

  authored and

  copybara-github

  committed
  13a6b00

  Copy full SHA for 13a6b00

• Enable GemmWithUpcast Cutlass Fusion Kernels. Add a Bf16 x Bf16 to F32 kernel.

  Show description for 7bfa523
  

  derdrdirk

  authored and

  copybara-github

  committed
  7bfa523

  Copy full SHA for 7bfa523

Commits on Jul 26, 2024

• [XLA:GPU] Add Custom Kernel Fusion Autotuner HLO pass.

  Show description for 7cad716
  

  derdrdirk

  authored and

  copybara-github

  committed
  7cad716

  Copy full SHA for 7cad716

Commits on Jul 25, 2024

• Use kernel_index from backend_config to select custom fusion kernel.

  Show description for 540e996
  

  derdrdirk

  authored and

  copybara-github

  committed
  540e996

  Copy full SHA for 540e996

Pagination

Previous

Next