[release/2.6] Cherry-picks from release/2.5#1910

Merged

jithunnair-amd merged 21 commits intorelease/2.6from release2.6_cherry_picks_from_release2.5

Feb 20, 2025

+250,902-371

Commits on Feb 19, 2025

Skipped certain distributed tests (#1383 )

BLOrange-AMD
authored and
jithunnair-amd
committed
[release/2.4] skip pytorch-nightly installation… (#1556 )
jithunnair-amd
committed
[release/2.4] add tlparse into requirements-ci.txt (#1608 ) (#1614 )

dnikolaev-amd
authored and
jithunnair-amd
committed
[ROCm] Use IPT=8 for block radix sort (#1619 ) (#1636 )

jerrymannil
authored and
jithunnair-amd
committed
CONSOLIDATED COMMITS: Enable bf16 with fp32 weights for MIOpen batchnorm and disable BF16 batchnorm with MIOpen for ROCm less then 6.4
jithunnair-amd
committed
[AUTOGENERATED] [release/2.5] [ROCm][layer_norm] Use __builtin_amdgcn_rcpf(x) instead of 1.f/x (#1800 )

rocm-mici
authored and
jithunnair-amd
committed
FEAT:at least one of ROCM_HOME or CUDA_HOME must be None. (#1814 )
jithunnair-amd
committed
[AUTOGENERATED] [release/2.5] [ROCm] Improve performance of reduce sum for 3D shapes (pytorch#143137 ) (#1843 )

rocm-mici
authored and
jithunnair-amd
committed
[AUTOGENERATED] [release/2.5] [ROCM] Enable *_load_dwordx4 ISA for BFloat16 and Half. (#1844 )

rocm-mici
authored and
jithunnair-amd
committed
[release/2.5] fix test_cublas_workspace_explicit_allocation for navi (#1847 )

dnikolaev-amd
authored and
jithunnair-amd
committed
CONSOLIDATED COMMITS: miopen benchmark behavior now better aligns with cudnn and missing miopen version.h

jeffdaily
authored and
jithunnair-amd
committed
[SWDEV-509031] [CP] [64-bit] Int64 casting for UpSampleNearest3D (pytorch#144865 ) (#1869 )

jataylo
authored and
jithunnair-amd
committed
[ROCm] Tune 3d tensor sums when not using fastest dimension (#1871 )

doru1004
authored and
jithunnair-amd
committed
Skip non_standard_bool_values tests (#1879 )

BLOrange-AMD
authored and
jithunnair-amd
committed
CONSOLIDATED COMMITS: Bump to AOTriton 0.8.2b and add head_dim=512 support

xinyazhang
authored and
jithunnair-amd
committed
CONSOLIDATED COMMITS: Turn on TF32 for aten::mm and De-noise tf32 warnings

xw285cornell
authored and
jithunnair-amd
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[release/2.6] Cherry-picks from release/2.5#1910

Commits on Feb 19, 2025

Skipped certain distributed tests (#1383 )

[release/2.4] skip pytorch-nightly installation… (#1556 )

[release/2.4] add tlparse into requirements-ci.txt (#1608 ) (#1614 )

[ROCm] Use IPT=8 for block radix sort (#1619 ) (#1636 )

CONSOLIDATED COMMITS: Enable bf16 with fp32 weights for MIOpen batchnorm and disable BF16 batchnorm with MIOpen for ROCm less then 6.4

[AUTOGENERATED] [release/2.5] [ROCm][layer_norm] Use __builtin_amdgcn_rcpf(x) instead of 1.f/x (#1800 )

FEAT:at least one of ROCM_HOME or CUDA_HOME must be None. (#1814 )

[AUTOGENERATED] [release/2.5] [ROCm] Improve performance of reduce sum for 3D shapes (pytorch#143137 ) (#1843 )

[AUTOGENERATED] [release/2.5] [ROCM] Enable *_load_dwordx4 ISA for BFloat16 and Half. (#1844 )

[release/2.5] fix test_cublas_workspace_explicit_allocation for navi (#1847 )

CONSOLIDATED COMMITS: miopen benchmark behavior now better aligns with cudnn and missing miopen version.h

[SWDEV-509031] [CP] [64-bit] Int64 casting for UpSampleNearest3D (pytorch#144865 ) (#1869 )

[ROCm] Tune 3d tensor sums when not using fastest dimension (#1871 )

Skip non_standard_bool_values tests (#1879 )

CONSOLIDATED COMMITS: Bump to AOTriton 0.8.2b and add head_dim=512 support

CONSOLIDATED COMMITS: Turn on TF32 for aten::mm and De-noise tf32 warnings

Commits on Feb 20, 2025

Use ROCm/triton for triton commit

[ROCm] CK Flash Attention Backend (pytorch#143695 )

Changes to support docker v23

Add QA automation scripts for running PyTorch unit tests

CONSOLIDATED COMMITS: Updates to build on Jammy (UB22.04), Noble (UB24.04) and CentOS