Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Tiny upstream pr #1094

Merged
merged 704 commits into from
Sep 10, 2024
Merged

Tiny upstream pr #1094

merged 704 commits into from
Sep 10, 2024
This pull request is big! We’re only showing the most recent 250 commits.

Commits on Feb 5, 2024

  1. Merge branch 'ck-tiled-fa' into develop

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    dc0e67a View commit details
    Browse the repository at this point in the history

  2. Building xformers using ck-tiled as default

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    58e6101 View commit details
    Browse the repository at this point in the history

  3. Merge branch 'ck-tiled-fa' into develop

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    1276abc View commit details
    Browse the repository at this point in the history

  4. ensure ck_decoder does not dispatch

    @tenpercent
    tenpercent committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    389dfb4 View commit details
    Browse the repository at this point in the history

  5. Add disable_on_rocm on some test scripts

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    f8d9043 View commit details
    Browse the repository at this point in the history

  6. Merge branch 'ck-tiled-fa' into develop

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    78df6a9 View commit details
    Browse the repository at this point in the history

  7. Update to test_mem_eff_attention.py

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    6dae63c View commit details
    Browse the repository at this point in the history

  8. Merge branch 'ck-tiled-fa' into develop

    @qianfengz
    qianfengz committed Feb 5, 2024
    Configuration menu
    Copy the full SHA
    a7ed88c View commit details
    Browse the repository at this point in the history

Commits on Feb 6, 2024

  1. Merge pull request #16 from ROCm/fix_test_attn_bias_padded 

    ensure ck_decoder does not dispatch in test_attn_bias_padded
    @qianfengz
    qianfengz authored Feb 6, 2024
    Configuration menu
    Copy the full SHA
    20e178a View commit details
    Browse the repository at this point in the history

  2. apply isort

    @tenpercent
    tenpercent committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    0624c92 View commit details
    Browse the repository at this point in the history

  3. apply black

    @tenpercent
    tenpercent committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    b8ebf08 View commit details
    Browse the repository at this point in the history

  4. fix flake8 suggestions

    @tenpercent
    tenpercent committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    3b33c5d View commit details
    Browse the repository at this point in the history

  5. add license headers and reapply black

    @tenpercent
    tenpercent committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    0a9c933 View commit details
    Browse the repository at this point in the history

  6. Merge pull request #17 from ROCm/linters 

    Apply the existing linters (1/n)
    @qianfengz
    qianfengz authored Feb 6, 2024
    Configuration menu
    Copy the full SHA
    47367a4 View commit details
    Browse the repository at this point in the history

  7. Merge pull request #10 from ROCm/enable-ci 

    add rocm_ci workflow
    @qianfengz
    qianfengz authored Feb 6, 2024
    Configuration menu
    Copy the full SHA
    fb46611 View commit details
    Browse the repository at this point in the history

  8. Tiny update to rocm_ci.yml

    @qianfengz
    qianfengz committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    28d3672 View commit details
    Browse the repository at this point in the history

  9. Add conditional compiling for cuda-depending codes in ROCM

    @qianfengz
    qianfengz committed Feb 6, 2024
    Configuration menu
    Copy the full SHA
    12fb41c View commit details
    Browse the repository at this point in the history

Commits on Feb 7, 2024

  1. Update to benchmark scripts

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    a9d83c6 View commit details
    Browse the repository at this point in the history

  2. Rename the one script file

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    9ab3831 View commit details
    Browse the repository at this point in the history

  3. Revert "Add conditional compiling for cuda-depending codes in ROCM" 

    This reverts commit 12fb41c.
    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    243dc6a View commit details
    Browse the repository at this point in the history

  4. Update to scripts

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    3240ba1 View commit details
    Browse the repository at this point in the history

  5. Change and add readme for tests and benchmarks

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    0c51af1 View commit details
    Browse the repository at this point in the history

  6. Remove the stuffs for supporting old ck

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    f36c93b View commit details
    Browse the repository at this point in the history

  7. Remove old composable_kernel from submodule list

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    9e4582d View commit details
    Browse the repository at this point in the history

  8. Remove folder third_party/composable_kernel

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    356cafd View commit details
    Browse the repository at this point in the history

  9. Merge branch 'develop' into dev_to_upstream

    @qianfengz
    qianfengz committed Feb 7, 2024
    Configuration menu
    Copy the full SHA
    8415b00 View commit details
    Browse the repository at this point in the history

Commits on Feb 8, 2024

  1. Rename the folder

    @qianfengz
    qianfengz committed Feb 8, 2024
    Configuration menu
    Copy the full SHA
    79c554c View commit details
    Browse the repository at this point in the history

  2. Remove unused script file

    @qianfengz
    qianfengz committed Feb 8, 2024
    Configuration menu
    Copy the full SHA
    2be6c04 View commit details
    Browse the repository at this point in the history

Commits on Feb 9, 2024

  1. apply black

    @tenpercent
    tenpercent committed Feb 9, 2024
    Configuration menu
    Copy the full SHA
    61d875a View commit details
    Browse the repository at this point in the history

  2. pacify mypy

    @tenpercent
    tenpercent committed Feb 9, 2024
    Configuration menu
    Copy the full SHA
    4616121 View commit details
    Browse the repository at this point in the history

  3. fix clang-format

    @tenpercent
    tenpercent committed Feb 9, 2024
    Configuration menu
    Copy the full SHA
    832e223 View commit details
    Browse the repository at this point in the history

  4. reapply black

    @tenpercent
    tenpercent committed Feb 9, 2024
    Configuration menu
    Copy the full SHA
    2b2967e View commit details
    Browse the repository at this point in the history

Commits on Feb 12, 2024

  1. Merge pull request #3 from tenpercent/lints 

    Force merging; please review later @qianfengz
    @tenpercent
    tenpercent authored Feb 12, 2024
    Configuration menu
    Copy the full SHA
    89fb7d6 View commit details
    Browse the repository at this point in the history

Commits on Feb 13, 2024

  1. fix lints

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    3c9d4e5 View commit details
    Browse the repository at this point in the history

  2. make test_splitk_reference run on cpu

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    1d474c5 View commit details
    Browse the repository at this point in the history

  3. add ck modules to docs

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    d38a684 View commit details
    Browse the repository at this point in the history

  4. try fixing nvidia build by re-including sparse24 cpp folder into exte… 

    …nsion sources
    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    eccbf54 View commit details
    Browse the repository at this point in the history

  5. update cutlass to upstream commit

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    1ef6c20 View commit details
    Browse the repository at this point in the history

  6. update flash-attention to upstream commit

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    9dfec0d View commit details
    Browse the repository at this point in the history

  7. simplify setup.py

    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    9fcda18 View commit details
    Browse the repository at this point in the history

  8. Merge branch 'main' of https://github.com/facebookresearch/xformers i… 

    …nto dev_upstream
    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    01c2bfd View commit details
    Browse the repository at this point in the history

  9. remove duplicate run_batched_infer_causalmask_attnbias_dispatched<f16… 

    …, true, true, 128>
    @tenpercent
    tenpercent committed Feb 13, 2024
    Configuration menu
    Copy the full SHA
    58d38d4 View commit details
    Browse the repository at this point in the history

Commits on Feb 14, 2024

  1. add hip version and pytorch hip arch list to xformers build info

    @tenpercent
    tenpercent committed Feb 14, 2024
    Configuration menu
    Copy the full SHA
    07183f0 View commit details
    Browse the repository at this point in the history

  2. fix build

    @tenpercent
    tenpercent committed Feb 14, 2024
    Configuration menu
    Copy the full SHA
    993a90c View commit details
    Browse the repository at this point in the history

  3. patch around the unhappy path in get_hip_version

    @tenpercent
    tenpercent committed Feb 14, 2024
    Configuration menu
    Copy the full SHA
    d4a374b View commit details
    Browse the repository at this point in the history

Commits on Feb 15, 2024

  1. skip test_grad_checkpointing for triton_splitk since it doesn't have … 

    …bwop
    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    ff59f19 View commit details
    Browse the repository at this point in the history

  2. re-enable test_mqa_forward since ck tiled is the current implementation

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    81bcfd5 View commit details
    Browse the repository at this point in the history

  3. make skip test_wrong_alignment more generic

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    a0f7f27 View commit details
    Browse the repository at this point in the history

  4. reapply black

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    a0d8dcc View commit details
    Browse the repository at this point in the history

  5. simplify test_decoder

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    bc7035c View commit details
    Browse the repository at this point in the history

  6. put python version check inside triton_splitk op

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    f02d0d4 View commit details
    Browse the repository at this point in the history

  7. fix logic

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    77a6c13 View commit details
    Browse the repository at this point in the history

  8. cleanup python3.9 checks in tests

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    a7cd678 View commit details
    Browse the repository at this point in the history

  9. cleanup test_attentions

    @tenpercent
    tenpercent committed Feb 15, 2024
    Configuration menu
    Copy the full SHA
    dea783d View commit details
    Browse the repository at this point in the history

Commits on Feb 16, 2024

  1. cleanup test_checkpoint as test running on cpu does not depend on gpu… 

    … platform
    @tenpercent
    tenpercent committed Feb 16, 2024
    Configuration menu
    Copy the full SHA
    acd6b7a View commit details
    Browse the repository at this point in the history

  2. fix lints

    @tenpercent
    tenpercent committed Feb 16, 2024
    Configuration menu
    Copy the full SHA
    f467a1d View commit details
    Browse the repository at this point in the history

  3. try fixing win build by conditional import of triton in triton op

    @tenpercent
    tenpercent committed Feb 16, 2024
    Configuration menu
    Copy the full SHA
    d758eac View commit details
    Browse the repository at this point in the history

Commits on Feb 17, 2024

  1. re-enable test_triton_layernorm as it passes

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    21f1904 View commit details
    Browse the repository at this point in the history

  2. re-enable test_triton_blocksparse as it passes

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    d880c36 View commit details
    Browse the repository at this point in the history

  3. cleanup test_sparse_tensors

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    059c84f View commit details
    Browse the repository at this point in the history

  4. cleanup test_custom_ops

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    8aa0bdc View commit details
    Browse the repository at this point in the history

  5. reapply black

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    5bc7bbe View commit details
    Browse the repository at this point in the history

  6. cleanup test_core_attention

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    5b4ebe4 View commit details
    Browse the repository at this point in the history

  7. benchmark ck ops on rocm only

    @tenpercent
    tenpercent committed Feb 17, 2024
    Configuration menu
    Copy the full SHA
    473ebc7 View commit details
    Browse the repository at this point in the history

Commits on Feb 19, 2024

  1. Merge branch 'main' of https://github.com/facebookresearch/xformers i… 

    …nto dev_upstream
    @tenpercent
    tenpercent committed Feb 19, 2024
    Configuration menu
    Copy the full SHA
    2a7272e View commit details
    Browse the repository at this point in the history

  2. fix mypy

    @tenpercent
    tenpercent committed Feb 19, 2024
    Configuration menu
    Copy the full SHA
    5d3247f View commit details
    Browse the repository at this point in the history

Commits on Feb 20, 2024

  1. Merge branch 'dev_upstream' of https://github.com/ROCm/xformers into … 

    …develop
    @tenpercent
    tenpercent committed Feb 20, 2024
    Configuration menu
    Copy the full SHA
    9be7f8d View commit details
    Browse the repository at this point in the history

Commits on Feb 21, 2024

  1. fix lint: black

    @tenpercent
    tenpercent committed Feb 21, 2024
    Configuration menu
    Copy the full SHA
    58b0f75 View commit details
    Browse the repository at this point in the history

  2. fix lints: mypy

    @tenpercent
    tenpercent committed Feb 21, 2024
    Configuration menu
    Copy the full SHA
    03b7294 View commit details
    Browse the repository at this point in the history

  3. split-k decoder: move all tunable parameters to the top of cpp file

    @tenpercent
    tenpercent committed Feb 21, 2024
    Configuration menu
    Copy the full SHA
    0666088 View commit details
    Browse the repository at this point in the history

  4. apply clang-format

    @tenpercent
    tenpercent committed Feb 21, 2024
    Configuration menu
    Copy the full SHA
    04eec8d View commit details
    Browse the repository at this point in the history

Commits on Feb 22, 2024

  1. Rename HDim/headdim to MaxK/maxk

    @qianfengz
    qianfengz committed Feb 22, 2024
    Configuration menu
    Copy the full SHA
    a02ab9b View commit details
    Browse the repository at this point in the history

  2. Move some headers files to ck examples for later reusing

    @qianfengz
    qianfengz committed Feb 22, 2024
    Configuration menu
    Copy the full SHA
    fd36725 View commit details
    Browse the repository at this point in the history

  3. Merge branch 'develop' of https://github.com/ROCm/xformers into develop

    @qianfengz
    qianfengz committed Feb 22, 2024
    Configuration menu
    Copy the full SHA
    41f5ada View commit details
    Browse the repository at this point in the history

  4. Replace using qs_ks_vs pipeline by qr_ks_vs pipeline while HeadDim is… 

    … 256 for better performance
    @qianfengz
    qianfengz committed Feb 22, 2024
    Configuration menu
    Copy the full SHA
    d8384c1 View commit details
    Browse the repository at this point in the history

Commits on Feb 23, 2024

  1. rm test_ck_7

    @tenpercent @qianfengz
    tenpercent authored and qianfengz committed Feb 23, 2024
    Configuration menu
    Copy the full SHA
    10346df View commit details
    Browse the repository at this point in the history

Commits on Feb 26, 2024

  1. Merge branch 'main' of https://github.com/facebookresearch/xformers i… 

    …nto develop
    @tenpercent
    tenpercent committed Feb 26, 2024
    Configuration menu
    Copy the full SHA
    bbfe112 View commit details
    Browse the repository at this point in the history

Commits on Mar 5, 2024

  1. Merge branch 'main' into develop

    @qianfengz
    qianfengz committed Mar 5, 2024
    Configuration menu
    Copy the full SHA
    dd3f4a9 View commit details
    Browse the repository at this point in the history

Commits on Mar 12, 2024

  1. dump kernel resource usage to compilation logs similar to nv

    @tenpercent
    tenpercent committed Mar 12, 2024
    Configuration menu
    Copy the full SHA
    08b4159 View commit details
    Browse the repository at this point in the history

Commits on Mar 13, 2024

  1. Merge branch 'facebookresearch:main' into develop

    @tenpercent
    tenpercent authored Mar 13, 2024
    Configuration menu
    Copy the full SHA
    ce99d22 View commit details
    Browse the repository at this point in the history

Commits on Mar 19, 2024

  1. Merge pull request #4 from ROCm/move-splitk-tune-params 

    split-k decoder: move all tunable parameters to the top of cpp file
    @qianfengz
    qianfengz authored Mar 19, 2024
    Configuration menu
    Copy the full SHA
    7637c61 View commit details
    Browse the repository at this point in the history

Commits on Mar 20, 2024

  1. Add the c++ extension to the latest change of ck_tile/dev fwd kernel … 

    …(added droppout)
    @qianfengz
    qianfengz committed Mar 20, 2024
    Configuration menu
    Copy the full SHA
    2da2927 View commit details
    Browse the repository at this point in the history

Commits on Mar 27, 2024

  1. Add the c++ extension to use ck_tile/dev/ fmha bwd kernel

    @qianfengz
    qianfengz committed Mar 27, 2024
    Configuration menu
    Copy the full SHA
    9189e45 View commit details
    Browse the repository at this point in the history

  2. Update to add dropout for fmah backward

    @qianfengz
    qianfengz committed Mar 27, 2024
    Configuration menu
    Copy the full SHA
    28e713d View commit details
    Browse the repository at this point in the history

  3. Update in attention.cpp to align efficient_attention_backward_ck inte… 

    …rface parameters
    @qianfengz
    qianfengz committed Mar 27, 2024
    Configuration menu
    Copy the full SHA
    4ef7eba View commit details
    Browse the repository at this point in the history

  4. Enable BwdOp in ck.py

    @qianfengz
    qianfengz committed Mar 27, 2024
    Configuration menu
    Copy the full SHA
    48a5f3e View commit details
    Browse the repository at this point in the history

Commits on Mar 28, 2024

  1. Support grad_out to have different strides as out

    @qianfengz
    qianfengz committed Mar 28, 2024
    Configuration menu
    Copy the full SHA
    2e45012 View commit details
    Browse the repository at this point in the history

  2. Merge branch 'facebookresearch:main' into develop

    @tenpercent
    tenpercent authored Mar 28, 2024
    Configuration menu
    Copy the full SHA
    b382f23 View commit details
    Browse the repository at this point in the history

Commits on Mar 29, 2024

  1. Force seqstart_q/seqstart_k to be in device memory in ck.py

    @qianfengz
    qianfengz committed Mar 29, 2024
    Configuration menu
    Copy the full SHA
    566d26f View commit details
    Browse the repository at this point in the history

  2. Remove duplicated codes in ck_tiled_fmha_grouped_forward.h/infer.h

    @qianfengz
    qianfengz committed Mar 29, 2024
    Configuration menu
    Copy the full SHA
    fc6c4a6 View commit details
    Browse the repository at this point in the history

  3. Use optimized async pipeline where 8x headdim length is assumed

    @qianfengz
    qianfengz committed Mar 29, 2024
    Configuration menu
    Copy the full SHA
    ff0db07 View commit details
    Browse the repository at this point in the history

Commits on Mar 30, 2024

  1. Fix in batched_infer

    @qianfengz
    qianfengz committed Mar 30, 2024
    Configuration menu
    Copy the full SHA
    0f4a171 View commit details
    Browse the repository at this point in the history

Commits on Apr 1, 2024

  1. Update to track ck_tile/opt_padding_fa_train_xformers branch

    @qianfengz
    qianfengz committed Apr 1, 2024
    Configuration menu
    Copy the full SHA
    0d6b915 View commit details
    Browse the repository at this point in the history

  2. Update rocm_ci.yml 

    configuring the self-hosted runner
    @tenpercent
    tenpercent authored Apr 1, 2024
    Configuration menu
    Copy the full SHA
    df43559 View commit details
    Browse the repository at this point in the history

  3. Update to use the newer FmhaFwdEpilogue

    @qianfengz
    qianfengz committed Apr 1, 2024
    Configuration menu
    Copy the full SHA
    4713576 View commit details
    Browse the repository at this point in the history

  4. Merge branch 'facebookresearch:main' into develop

    @tenpercent
    tenpercent authored Apr 1, 2024
    Configuration menu
    Copy the full SHA
    9c2f5ce View commit details
    Browse the repository at this point in the history

  5. Update rocm_ci.yml 

    add option to manually trigger workflow
    @tenpercent
    tenpercent authored Apr 1, 2024
    Configuration menu
    Copy the full SHA
    a745c45 View commit details
    Browse the repository at this point in the history

  6. Update rocm_ci.yml 

    remove condition which skips ci unless github event contains string 'rocm'
    @tenpercent
    tenpercent authored Apr 1, 2024
    Configuration menu
    Copy the full SHA
    95d0260 View commit details
    Browse the repository at this point in the history

  7. copy rocm_ci workflow from main branch

    @tenpercent
    tenpercent committed Apr 1, 2024
    Configuration menu
    Copy the full SHA
    4069efe View commit details
    Browse the repository at this point in the history

  8. Update rocm_ci.yml 

    Bump upload-artifact version
    @tenpercent
    tenpercent authored Apr 1, 2024
    Configuration menu
    Copy the full SHA
    724354c View commit details
    Browse the repository at this point in the history

Commits on Apr 2, 2024

  1. Update to use the newer FmhaFwdEpilogue for grouped infer/forward

    @qianfengz
    qianfengz committed Apr 2, 2024
    Configuration menu
    Copy the full SHA
    b1a1009 View commit details
    Browse the repository at this point in the history

Commits on Apr 3, 2024

  1. Temporarily disable the using of QRKSVSAsync() pipeline

    @qianfengz
    qianfengz committed Apr 3, 2024
    Configuration menu
    Copy the full SHA
    97e4e20 View commit details
    Browse the repository at this point in the history

  2. Update rocm_ci.yml 

    add a daily run
    @tenpercent
    tenpercent authored Apr 3, 2024
    Configuration menu
    Copy the full SHA
    e98877a View commit details
    Browse the repository at this point in the history

  3. Implement the ck_rand_uniform interface for generating random number … 

    …tensor
    @qianfengz
    qianfengz committed Apr 3, 2024
    Configuration menu
    Copy the full SHA
    6fbd05d View commit details
    Browse the repository at this point in the history

Commits on Apr 7, 2024

  1. Add dropout to the infer path (needed by xformers test_dropout)

    @qianfengz
    qianfengz committed Apr 7, 2024
    Configuration menu
    Copy the full SHA
    2ef3b3f View commit details
    Browse the repository at this point in the history

Commits on Apr 8, 2024

  1. Update to support test_dropout and test_dropout_backward tests

    @qianfengz
    qianfengz committed Apr 8, 2024
    Configuration menu
    Copy the full SHA
    930bb25 View commit details
    Browse the repository at this point in the history

Commits on Apr 9, 2024

  1. Update the padding method in batched_backward.h

    @qianfengz
    qianfengz committed Apr 9, 2024
    Configuration menu
    Copy the full SHA
    bdbc956 View commit details
    Browse the repository at this point in the history

  2. Update the OGradDotO kernel padding method

    @qianfengz
    qianfengz committed Apr 9, 2024
    Configuration menu
    Copy the full SHA
    44fff29 View commit details
    Browse the repository at this point in the history

  3. Change the backward padding checking condition

    @qianfengz
    qianfengz committed Apr 9, 2024
    Configuration menu
    Copy the full SHA
    d5c2d88 View commit details
    Browse the repository at this point in the history

Commits on Apr 10, 2024

  1. Add batch_stride_lse/d parameters to adapt grouped mode forward/backw… 

    …ard to [num_batches, H, MaxSeqlenQ] layout
    @qianfengz
    qianfengz committed Apr 10, 2024
    Configuration menu
    Copy the full SHA
    ce9c23c View commit details
    Browse the repository at this point in the history

  2. Fill the grad_bias in advance

    @qianfengz
    qianfengz committed Apr 10, 2024
    Configuration menu
    Copy the full SHA
    dafea78 View commit details
    Browse the repository at this point in the history

Commits on Apr 11, 2024

  1. Add support for kHasBiasGrad as instance template

    @qianfengz
    qianfengz committed Apr 11, 2024
    Configuration menu
    Copy the full SHA
    06ad689 View commit details
    Browse the repository at this point in the history

  2. Remove using hdim_stride_do in fmha backward

    @qianfengz
    qianfengz committed Apr 11, 2024
    Configuration menu
    Copy the full SHA
    bdd6291 View commit details
    Browse the repository at this point in the history

  3. Force kPadSeqLenQ/kPadSeqLenK to be true in batched-backward to save … 

    …compiling time
    @qianfengz
    qianfengz committed Apr 11, 2024
    Configuration menu
    Copy the full SHA
    410f814 View commit details
    Browse the repository at this point in the history

Commits on Apr 12, 2024

  1. Fix missing passing of {philox_seed, philox_offset} in inference path

    @qianfengz
    qianfengz committed Apr 12, 2024
    Configuration menu
    Copy the full SHA
    2712dff View commit details
    Browse the repository at this point in the history

Commits on Apr 14, 2024

  1. Use SimplifiedGenericAttentionMask to replace GenericAttentionMask

    @qianfengz
    qianfengz committed Apr 14, 2024
    Configuration menu
    Copy the full SHA
    7c27a82 View commit details
    Browse the repository at this point in the history

  2. Shorten the instance file names

    @qianfengz
    qianfengz committed Apr 14, 2024
    Configuration menu
    Copy the full SHA
    46c491e View commit details
    Browse the repository at this point in the history

  3. Rename the template parameters

    @qianfengz
    qianfengz committed Apr 14, 2024
    Configuration menu
    Copy the full SHA
    4c6c08d View commit details
    Browse the repository at this point in the history

Commits on Apr 15, 2024

  1. Simplify the names of the dispatch class and interfaces

    @qianfengz
    qianfengz committed Apr 15, 2024
    Configuration menu
    Copy the full SHA
    411ccd6 View commit details
    Browse the repository at this point in the history

Commits on Apr 16, 2024

  1. Changes to reuse the kernel files under ck_tile examples/91_tile_prog… 

    …ram/fmha folder
    @qianfengz
    qianfengz committed Apr 16, 2024
    Configuration menu
    Copy the full SHA
    812a529 View commit details
    Browse the repository at this point in the history

  2. Update test_mem_eff_attention.py for test_dropout/test_dropout_backwa… 

    …rd/test_backward on rocm
    @qianfengz
    qianfengz committed Apr 16, 2024
    Configuration menu
    Copy the full SHA
    51b4223 View commit details
    Browse the repository at this point in the history

  3. Tiny change to the philox_cuda_state input setting

    @qianfengz
    qianfengz committed Apr 16, 2024
    Configuration menu
    Copy the full SHA
    d10ef79 View commit details
    Browse the repository at this point in the history

  4. Allocate logsumexp to ensure aligned access by each thread-group

    @qianfengz
    qianfengz committed Apr 16, 2024
    Configuration menu
    Copy the full SHA
    25bd720 View commit details
    Browse the repository at this point in the history

  5. Add checking for query/key headdim size attention_backward_generic

    @qianfengz
    qianfengz committed Apr 16, 2024
    Configuration menu
    Copy the full SHA
    abfdc27 View commit details
    Browse the repository at this point in the history

Commits on Apr 22, 2024

  1. Using ck_tile/opt_padding_fa_train_pr2 and synchronize the backward c… 

    …odes with the changes
    @qianfengz
    qianfengz committed Apr 22, 2024
    Configuration menu
    Copy the full SHA
    ff95367 View commit details
    Browse the repository at this point in the history

  2. Enable using async pipeline in the batched inference path for perform… 

    …ance
    @qianfengz
    qianfengz committed Apr 22, 2024
    Configuration menu
    Copy the full SHA
    93469ab View commit details
    Browse the repository at this point in the history

Commits on Apr 23, 2024

  1. Re-organize cpp instances for calling fmha infer kernel

    @qianfengz
    qianfengz committed Apr 23, 2024
    Configuration menu
    Copy the full SHA
    2c8626b View commit details
    Browse the repository at this point in the history

  2. Re-organize cpp instances for calling fmha forward kernel

    @qianfengz
    qianfengz committed Apr 23, 2024
    Configuration menu
    Copy the full SHA
    bdd716c View commit details
    Browse the repository at this point in the history

  3. Re-organize cpp instances for calling fmha backward kernel

    @qianfengz
    qianfengz committed Apr 23, 2024
    Configuration menu
    Copy the full SHA
    44d4592 View commit details
    Browse the repository at this point in the history

  4. Position the composable_kernel_tiled to ck_tile/opt_padding_fa_train … 

    …branch
    @qianfengz
    qianfengz committed Apr 23, 2024
    Configuration menu
    Copy the full SHA
    51ca91b View commit details
    Browse the repository at this point in the history

  5. Update to synchronize with the latest commits in ck_tile/opt_padding_… 

    …fa_train
    @qianfengz
    qianfengz committed Apr 23, 2024
    Configuration menu
    Copy the full SHA
    1693683 View commit details
    Browse the repository at this point in the history

Commits on Apr 26, 2024

  1. update submodule to public

    @carlushuang
    carlushuang committed Apr 26, 2024
    Configuration menu
    Copy the full SHA
    b7aa908 View commit details
    Browse the repository at this point in the history

  2. Merge pull request #7 from ROCm/origin/test_opt_padding_train_public 

    update submodule to public
    @qianfengz
    qianfengz authored Apr 26, 2024
    Configuration menu
    Copy the full SHA
    9a878d9 View commit details
    Browse the repository at this point in the history

Commits on May 6, 2024

  1. Update to the criteria for padding seqlen_k in batched infer/forward

    @qianfengz
    qianfengz committed May 6, 2024
    Configuration menu
    Copy the full SHA
    b4fa26d View commit details
    Browse the repository at this point in the history

  2. Keep latest track of ck-tile commits

    @qianfengz
    qianfengz committed May 6, 2024
    Configuration menu
    Copy the full SHA
    ee7950f View commit details
    Browse the repository at this point in the history

Commits on May 8, 2024

  1. Tiny fixing to the decoder including

    @qianfengz
    qianfengz committed May 8, 2024
    Configuration menu
    Copy the full SHA
    74dfdfe View commit details
    Browse the repository at this point in the history

Commits on May 9, 2024

  1. Position the ck-tiled to ck_tile/opt_padding branch

    @qianfengz
    qianfengz committed May 9, 2024
    Configuration menu
    Copy the full SHA
    410757e View commit details
    Browse the repository at this point in the history

  2. Merge branch 'test_opt_padding_train' of https://github.com/ROCm/xfor… 

    …mers into test_opt_padding_train
    @qianfengz
    qianfengz committed May 9, 2024
    Configuration menu
    Copy the full SHA
    fa155eb View commit details
    Browse the repository at this point in the history

  3. Merge branch 'develop' into test_opt_padding_train

    @qianfengz
    qianfengz committed May 9, 2024
    Configuration menu
    Copy the full SHA
    77514d5 View commit details
    Browse the repository at this point in the history

Commits on May 11, 2024

  1. Enable some attn_bias types which were previously disabled by old-ck … 

    …in ck.py
    @qianfengz
    qianfengz committed May 11, 2024
    Configuration menu
    Copy the full SHA
    92924d4 View commit details
    Browse the repository at this point in the history

Commits on May 14, 2024

  1. Add script generate_instances.py which helps to generate instances

    @qianfengz
    qianfengz committed May 14, 2024
    Configuration menu
    Copy the full SHA
    23f64bd View commit details
    Browse the repository at this point in the history

Commits on May 15, 2024

  1. Simplify logic for seqstart_q/k 

    566d26f has put the seqstart_k/q on device. So simplify the logic here.

The upstream xformers don't have this optmization and is copying the seqstart_q/k every iterations. We'd like this change to get in and then merge to upstream.
    @xw285cornell
    xw285cornell committed May 15, 2024
    Configuration menu
    Copy the full SHA
    d94b2c1 View commit details
    Browse the repository at this point in the history

  2. Add Async pipeline to grouped mode inference path

    @qianfengz
    qianfengz committed May 15, 2024
    Configuration menu
    Copy the full SHA
    2486b56 View commit details
    Browse the repository at this point in the history

  3. Use explict true for kPadSeqLenQ/kPadHeadDimQ/kPadHeadDimV templates … 

    …for the Async pipeline
    @qianfengz
    qianfengz committed May 15, 2024
    Configuration menu
    Copy the full SHA
    18b43c9 View commit details
    Browse the repository at this point in the history

Commits on May 16, 2024

  1. Merge pull request #11 from xw285cornell/develop 

    Simplify logic for seqstart_q/k
    @qianfengz
    qianfengz authored May 16, 2024
    Configuration menu
    Copy the full SHA
    cf6cca0 View commit details
    Browse the repository at this point in the history

Commits on May 21, 2024

  1. Synchronize to the update of composable_kernel_tiled for better perfo… 

    …rmance
    @qianfengz
    qianfengz committed May 21, 2024
    Configuration menu
    Copy the full SHA
    14f7abe View commit details
    Browse the repository at this point in the history

Commits on May 23, 2024

  1. Update rocm_ci.yml - clean up dangling images after ci run

    @tenpercent
    tenpercent committed May 23, 2024
    Configuration menu
    Copy the full SHA
    ee4aa87 View commit details
    Browse the repository at this point in the history

Commits on May 25, 2024

  1. Avoid unused-const-variable warning 

    Our compiler will error on unused-const-variable warning. So just fix this
    @xw285cornell
    xw285cornell committed May 25, 2024
    Configuration menu
    Copy the full SHA
    b0b5547 View commit details
    Browse the repository at this point in the history

Commits on May 29, 2024

  1. Tiny change in the BlockTile/Shape setting overriddings

    @qianfengz
    qianfengz committed May 29, 2024
    Configuration menu
    Copy the full SHA
    dfc196d View commit details
    Browse the repository at this point in the history

  2. Merge branch 'develop' of https://github.com/ROCm/xformers into develop

    @qianfengz
    qianfengz committed May 29, 2024
    Configuration menu
    Copy the full SHA
    2490166 View commit details
    Browse the repository at this point in the history

Commits on Jun 12, 2024

  1. try to align fmha C++ extension to the ck_tile in ck develop branch

    @qianfengz
    qianfengz committed Jun 12, 2024
    Configuration menu
    Copy the full SHA
    f50861a View commit details
    Browse the repository at this point in the history

Commits on Jun 13, 2024

  1. Synchronize composable_kernel_tiled to latest ck develop

    @qianfengz
    qianfengz committed Jun 13, 2024
    Configuration menu
    Copy the full SHA
    76fb485 View commit details
    Browse the repository at this point in the history

Commits on Jun 14, 2024

  1. Use FmhaFwdTilePartitioner_HBS only with seqlen_k padded cases

    @qianfengz
    qianfengz committed Jun 14, 2024
    Configuration menu
    Copy the full SHA
    1f3add7 View commit details
    Browse the repository at this point in the history

Commits on Jun 16, 2024

  1. Merge branch 'main' into develop

    @qianfengz
    qianfengz committed Jun 16, 2024
    Configuration menu
    Copy the full SHA
    ed226f4 View commit details
    Browse the repository at this point in the history

Commits on Jun 17, 2024

  1. Tiny fix/change to make test_forward/test_backward/test_dropout/test_… 

    …dropout_backward_ck pass
    @qianfengz
    qianfengz committed Jun 17, 2024
    Configuration menu
    Copy the full SHA
    9df93e5 View commit details
    Browse the repository at this point in the history

  2. Fix compiling issue with regard to Invoker definitions in forward_dec… 

    …oder/forward_decoder_split operators
    @qianfengz
    qianfengz committed Jun 17, 2024
    Configuration menu
    Copy the full SHA
    d6ccfa1 View commit details
    Browse the repository at this point in the history

Commits on Jun 18, 2024

  1. Keep using -Woverloaded-virtual

    @qianfengz
    qianfengz committed Jun 18, 2024
    Configuration menu
    Copy the full SHA
    a7c7475 View commit details
    Browse the repository at this point in the history

  2. Fix clang-format for headers and cpp files

    @qianfengz
    qianfengz committed Jun 18, 2024
    Configuration menu
    Copy the full SHA
    b157b49 View commit details
    Browse the repository at this point in the history

  3. Fix format in python scripts

    @qianfengz
    qianfengz committed Jun 18, 2024
    Configuration menu
    Copy the full SHA
    b2fb213 View commit details
    Browse the repository at this point in the history

  4. Add noqa: C801 for generate_instances.py

    @qianfengz
    qianfengz committed Jun 18, 2024
    Configuration menu
    Copy the full SHA
    fdf8b8e View commit details
    Browse the repository at this point in the history

Commits on Jun 19, 2024

  1. Align dispatch_bw with main branch

    @qianfengz
    qianfengz committed Jun 19, 2024
    Configuration menu
    Copy the full SHA
    633a161 View commit details
    Browse the repository at this point in the history

  2. Align ops/fmha/common.py with main branch

    @qianfengz
    qianfengz committed Jun 19, 2024
    Configuration menu
    Copy the full SHA
    00cf683 View commit details
    Browse the repository at this point in the history

Commits on Jun 20, 2024

  1. Synchronize the thirty-party/composable_kernel_tiled to latest ck_til… 

    …e commits for better performance
    @qianfengz
    qianfengz committed Jun 20, 2024
    Configuration menu
    Copy the full SHA
    252844d View commit details
    Browse the repository at this point in the history

  2. Relax the atol for test_forward and test_dropout due to the using of … 

    …packed fp16_2_fp32 conversion in ck_tile
    @qianfengz
    qianfengz committed Jun 20, 2024
    Configuration menu
    Copy the full SHA
    610909e View commit details
    Browse the repository at this point in the history

Commits on Jul 1, 2024

  1. Generate html report for tests run with rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 1, 2024
    Configuration menu
    Copy the full SHA
    10bf99c View commit details
    Browse the repository at this point in the history

  2. archive test results when tests have failed

    @tenpercent
    tenpercent authored Jul 1, 2024
    Configuration menu
    Copy the full SHA
    16bb10b View commit details
    Browse the repository at this point in the history

  3. Always clean up dangling docker images in rocm_ci

    @tenpercent
    tenpercent authored Jul 1, 2024
    Configuration menu
    Copy the full SHA
    29c782b View commit details
    Browse the repository at this point in the history

Commits on Jul 2, 2024

  1. Bump python to 3.11 in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    782d5a3 View commit details
    Browse the repository at this point in the history

  2. Disable flash attention tests rocm_ci.yml 

    Since the op is broken; tbd either make the op work, or disable it on ROCm
    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    bd8ca1b View commit details
    Browse the repository at this point in the history

  3. Try to fix rocm_ci.yml 

    Init must be called before activation
    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    77beb19 View commit details
    Browse the repository at this point in the history

  4. try to fix rocm_ci.yml flow by overriding PATH

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    b0ae707 View commit details
    Browse the repository at this point in the history

  5. Fix setup.py path in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    d2eeaf0 View commit details
    Browse the repository at this point in the history

  6. cd to xformers dir before running install in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    a62c93e View commit details
    Browse the repository at this point in the history

  7. Use pip to install xformers in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    d3ae25f View commit details
    Browse the repository at this point in the history

  8. Possibly fix python version resolution in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    d4e6abc View commit details
    Browse the repository at this point in the history

  9. Set the correct path for pytest in rocm_ci.yml

    @tenpercent
    tenpercent authored Jul 2, 2024
    Configuration menu
    Copy the full SHA
    490b63d View commit details
    Browse the repository at this point in the history

  10. remove test_reference_splitk as it was moved to a different file duri… 

    …ng the first upstream

remove test_mqa_forward from develop, as the test fails in develop and doesn't run upstream

remove reference attention splitk from the test file; it exists in test_splitk_reference

sync test_mem_eff_attention with upstream
    @tenpercent
    tenpercent committed Jul 2, 2024
    Configuration menu
    Copy the full SHA
    addd2f2 View commit details
    Browse the repository at this point in the history

Commits on Jul 3, 2024

  1. make sure ck operators have a name to be visible in the dispatcher

    @tenpercent
    tenpercent committed Jul 3, 2024
    Configuration menu
    Copy the full SHA
    33810ff View commit details
    Browse the repository at this point in the history

Commits on Jul 8, 2024

  1. fix sm version checks to happen only on CUDA, not ROCm

    @tenpercent
    tenpercent committed Jul 8, 2024
    Configuration menu
    Copy the full SHA
    f3faa1a View commit details
    Browse the repository at this point in the history

  2. (2/n) fix sm version checks to happen only on CUDA, not ROCm

    @tenpercent
    tenpercent committed Jul 8, 2024
    Configuration menu
    Copy the full SHA
    04e9481 View commit details
    Browse the repository at this point in the history

Commits on Jul 9, 2024

  1. Merge pull request #13 from xw285cornell/xdwang-develop 

    Avoid unused-const-variable warning
    @qianfengz
    qianfengz authored Jul 9, 2024
    Configuration menu
    Copy the full SHA
    9440282 View commit details
    Browse the repository at this point in the history

Commits on Jul 18, 2024

  1. Remove _check_large_shapes checking in fmha/ck.py (facebookresearch#1067 

    )
    @qianfengz
    qianfengz committed Jul 18, 2024
    Configuration menu
    Copy the full SHA
    bd49f48 View commit details
    Browse the repository at this point in the history

  2. make xformers install editable to fix cpp extensions detection

    @tenpercent
    tenpercent committed Jul 18, 2024
    Configuration menu
    Copy the full SHA
    0d1d1be View commit details
    Browse the repository at this point in the history

Commits on Jul 23, 2024

  1. Update to using the improved fmha-bwd (compiling passed)

    @qianfengz
    qianfengz committed Jul 23, 2024
    Configuration menu
    Copy the full SHA
    9390d6a View commit details
    Browse the repository at this point in the history

  2. Update to get 80% of the test_backward and test_dropout_backward_ck c… 

    …ases passed
    @qianfengz
    qianfengz committed Jul 23, 2024
    Configuration menu
    Copy the full SHA
    22fce7e View commit details
    Browse the repository at this point in the history

Commits on Jul 25, 2024

  1. Replace the using of ConvertGradQ by using torch tensor type converting

    @qianfengz
    qianfengz committed Jul 25, 2024
    Configuration menu
    Copy the full SHA
    463a475 View commit details
    Browse the repository at this point in the history

  2. Change the tile settings for MaxK=32

    @qianfengz
    qianfengz committed Jul 25, 2024
    Configuration menu
    Copy the full SHA
    3427a6f View commit details
    Browse the repository at this point in the history

Commits on Jul 26, 2024

  1. Fix padding setting bug in grouped_backward

    @qianfengz
    qianfengz committed Jul 26, 2024
    Configuration menu
    Copy the full SHA
    fbc7c50 View commit details
    Browse the repository at this point in the history

  2. Change -DCK_FMHA_FWD_FAST_EXP2=1 to -DCK_TILE_FMHA_FWD_FAST_EXP2=1

    @qianfengz
    qianfengz committed Jul 26, 2024
    Configuration menu
    Copy the full SHA
    6e08666 View commit details
    Browse the repository at this point in the history

  3. Point the composable_kernel_tiled submodule to ck_tile/fa_bwd_opt branch

    @qianfengz
    qianfengz committed Jul 26, 2024
    Configuration menu
    Copy the full SHA
    94ab599 View commit details
    Browse the repository at this point in the history

Commits on Jul 27, 2024

  1. Disable flshattF and flshattB on ROCM

    @qianfengz
    qianfengz committed Jul 27, 2024
    Configuration menu
    Copy the full SHA
    830697c View commit details
    Browse the repository at this point in the history

  2. Add -mllvm and -enable-post-misched=0 compiling options for ROCM on s… 

    …etup.py
    @qianfengz
    qianfengz committed Jul 27, 2024
    Configuration menu
    Copy the full SHA
    afd7e02 View commit details
    Browse the repository at this point in the history

  3. Disable flshattF and flshattB on ROCM

    @qianfengz
    qianfengz committed Jul 27, 2024
    Configuration menu
    Copy the full SHA
    e67de41 View commit details
    Browse the repository at this point in the history

Commits on Jul 28, 2024

  1. Update to support separate grad_q_f32_strides do to the API change in… 

    … the fmd_bwd_kernel
    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    d72c2b3 View commit details
    Browse the repository at this point in the history

  2. Use old method for setting BlockDropout due to the revert in fmha_fwd… 

    …_kernel
    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    5ddff31 View commit details
    Browse the repository at this point in the history

  3. Tiny fix in grouped_backward

    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    cf2b622 View commit details
    Browse the repository at this point in the history

  4. Use packed tensor allocation for grad_q_f32

    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    112aaed View commit details
    Browse the repository at this point in the history

  5. Update to the ConvertGradQ kernel calling

    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    dd83c62 View commit details
    Browse the repository at this point in the history

  6. Tiny update

    @qianfengz
    qianfengz committed Jul 28, 2024
    Configuration menu
    Copy the full SHA
    3e9b99d View commit details
    Browse the repository at this point in the history

Commits on Jul 29, 2024

  1. Fix the parameter location in grouped_backward

    @qianfengz
    qianfengz committed Jul 29, 2024
    Configuration menu
    Copy the full SHA
    019448e View commit details
    Browse the repository at this point in the history

Commits on Aug 5, 2024

  1. Adjust headdim128 tile shapes for better performance

    @qianfengz
    qianfengz committed Aug 5, 2024
    Configuration menu
    Copy the full SHA
    c55966a View commit details
    Browse the repository at this point in the history

  2. Update backward kernel calling due to adding of nhead_stride_dk/nhead… 

    …_stride_dv parameters
    @qianfengz
    qianfengz committed Aug 5, 2024
    Configuration menu
    Copy the full SHA
    e22829a View commit details
    Browse the repository at this point in the history

  3. Synchronize with CK to use separate pipeline for kPadHeadDim true of … 

    …false situtation
    @qianfengz
    qianfengz committed Aug 5, 2024
    Configuration menu
    Copy the full SHA
    cae1b77 View commit details
    Browse the repository at this point in the history

Commits on Aug 6, 2024

  1. Use convertDQ kernel

    @qianfengz
    qianfengz committed Aug 6, 2024
    Configuration menu
    Copy the full SHA
    e564f5e View commit details
    Browse the repository at this point in the history

Commits on Aug 7, 2024

  1. Update to use unpadded lse layout

    @qianfengz
    qianfengz committed Aug 7, 2024
    Configuration menu
    Copy the full SHA
    b043765 View commit details
    Browse the repository at this point in the history

  2. Add explicit headdim256 instances for fmha backward

    @qianfengz
    qianfengz committed Aug 7, 2024
    Configuration menu
    Copy the full SHA
    c9e7595 View commit details
    Browse the repository at this point in the history

  3. Add leaked headdim256 instance references

    @qianfengz
    qianfengz committed Aug 7, 2024
    Configuration menu
    Copy the full SHA
    4a7b7dc View commit details
    Browse the repository at this point in the history

  4. Change to generate.py and the re-generate the instance files using it

    @qianfengz
    qianfengz committed Aug 7, 2024
    Configuration menu
    Copy the full SHA
    1ad9cbe View commit details
    Browse the repository at this point in the history

  5. Change to generate.py to generate instances refences and uses the gen… 

    …erated reference headers
    @qianfengz
    qianfengz committed Aug 7, 2024
    Configuration menu
    Copy the full SHA
    7db2aa4 View commit details
    Browse the repository at this point in the history

Commits on Aug 8, 2024

  1. Relax the RTOL of ckFwOp from 4e-4 to 3e-3 due to one big result case

    @qianfengz
    qianfengz committed Aug 8, 2024
    Configuration menu
    Copy the full SHA
    73dbf32 View commit details
    Browse the repository at this point in the history

Commits on Aug 12, 2024

  1. Change to use .h rather than .hpp as suffix for generated header files

    @qianfengz
    qianfengz committed Aug 12, 2024
    Configuration menu
    Copy the full SHA
    0e6d0c3 View commit details
    Browse the repository at this point in the history

  2. Fix in .gitignore

    @qianfengz
    qianfengz committed Aug 12, 2024
    Configuration menu
    Copy the full SHA
    914ccc5 View commit details
    Browse the repository at this point in the history

  3. Update to bwd setting to use only IGLP pipeline

    @qianfengz
    qianfengz committed Aug 12, 2024
    Configuration menu
    Copy the full SHA
    8503f87 View commit details
    Browse the repository at this point in the history

  4. Synchronize to latest ck_tile fix and align the headdim64 tile shape … 

    …setting
    @qianfengz
    qianfengz committed Aug 12, 2024
    Configuration menu
    Copy the full SHA
    bfe164d View commit details
    Browse the repository at this point in the history

  5. Reformat the generated instances cpp files

    @qianfengz
    qianfengz committed Aug 12, 2024
    Configuration menu
    Copy the full SHA
    f75c3b2 View commit details
    Browse the repository at this point in the history

  6. Merge pull request #18 from ROCm/fa_bwd_opt_test 

    Add integration of improved Fmha-bwd
    @qianfengz
    qianfengz authored Aug 12, 2024
    Configuration menu
    Copy the full SHA
    520e6ed View commit details
    Browse the repository at this point in the history

Commits on Aug 13, 2024

  1. Fix to the backward Trait

    @qianfengz
    qianfengz committed Aug 13, 2024
    Configuration menu
    Copy the full SHA
    bc3db99 View commit details
    Browse the repository at this point in the history

  2. Set occupancy to -1 to avoid the compiling warning

    @qianfengz
    qianfengz committed Aug 13, 2024
    Configuration menu
    Copy the full SHA
    fa6d8b3 View commit details
    Browse the repository at this point in the history

  3. Revert "Set occupancy to -1 to avoid the compiling warning" 

    This reverts commit fa6d8b3.
    @qianfengz
    qianfengz committed Aug 13, 2024
    Configuration menu
    Copy the full SHA
    c5c7cce View commit details
    Browse the repository at this point in the history

Commits on Aug 14, 2024

  1. Add environment variable and compiler definition to control the gener… 

    …ating of headdim256 instances
    @qianfengz
    qianfengz committed Aug 14, 2024
    Configuration menu
    Copy the full SHA
    d230433 View commit details
    Browse the repository at this point in the history

  2. Add --ignore-hd256 argument to generate_instance.py and some update i… 

    …n this script
    @qianfengz
    qianfengz committed Aug 14, 2024
    Configuration menu
    Copy the full SHA
    82a07ae View commit details
    Browse the repository at this point in the history

Commits on Aug 15, 2024

  1. Add environment variable ENABLE_HIP_FMHA_RTN_BF16_CONVERT to enable u… 

    …sing rtn bf16 conversion
    @qianfengz
    qianfengz committed Aug 15, 2024
    Configuration menu
    Copy the full SHA
    38593d6 View commit details
    Browse the repository at this point in the history

  2. Remove commented lines in test_mem_eff_attention.py

    @qianfengz
    qianfengz committed Aug 15, 2024
    Configuration menu
    Copy the full SHA
    15dc911 View commit details
    Browse the repository at this point in the history

  3. Synchronize to latest ck_tile commit

    @qianfengz
    qianfengz committed Aug 15, 2024
    Configuration menu
    Copy the full SHA
    367274c View commit details
    Browse the repository at this point in the history

Commits on Aug 16, 2024

  1. apply black

    @tenpercent
    tenpercent committed Aug 16, 2024
    Configuration menu
    Copy the full SHA
    f7b28c5 View commit details
    Browse the repository at this point in the history

  2. apply flake8

    @tenpercent
    tenpercent committed Aug 16, 2024
    Configuration menu
    Copy the full SHA
    fd82f20 View commit details
    Browse the repository at this point in the history

  3. fix mypy

    @tenpercent
    tenpercent committed Aug 16, 2024
    Configuration menu
    Copy the full SHA
    7d21800 View commit details
    Browse the repository at this point in the history

  4. revert disable flash operator on rocm

    @tenpercent
    tenpercent committed Aug 16, 2024
    Configuration menu
    Copy the full SHA
    d6b6456 View commit details
    Browse the repository at this point in the history

  5. Synchronize to ck_tile latest commit again

    @qianfengz
    qianfengz committed Aug 16, 2024
    Configuration menu
    Copy the full SHA
    87188ea View commit details
    Browse the repository at this point in the history

Commits on Aug 17, 2024

  1. Re-position the composable_kernel submodule to the develop branch

    @qianfengz
    qianfengz committed Aug 17, 2024
    Configuration menu
    Copy the full SHA
    5be80a3 View commit details
    Browse the repository at this point in the history

  2. Merge pull request #20 from tenpercent/develop 

    Fix lints
    @qianfengz
    qianfengz authored Aug 17, 2024
    Configuration menu
    Copy the full SHA
    cee0980 View commit details
    Browse the repository at this point in the history

  3. Avoid the Async pipeline when khasBias is true

    @qianfengz
    qianfengz committed Aug 17, 2024
    Configuration menu
    Copy the full SHA
    2a5c141 View commit details
    Browse the repository at this point in the history

  4. clang-format for two files

    @qianfengz
    qianfengz committed Aug 17, 2024
    Configuration menu
    Copy the full SHA
    2874842 View commit details
    Browse the repository at this point in the history

Commits on Aug 20, 2024

  1. Change allocation of grouped mode lse from [H, M] to [1, H, M] to mat… 

    …ch the xformers scripts
    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    7a91589 View commit details
    Browse the repository at this point in the history

  2. Change in generate_instances.py so that this scripts can be called fr… 

    …om flexible location
    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    66efb2c View commit details
    Browse the repository at this point in the history

  3. Add manual for generate_instances.py (.md)

    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    c19b1f5 View commit details
    Browse the repository at this point in the history

  4. Modification in GENERATE_INSTANCES.md

    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    b450d01 View commit details
    Browse the repository at this point in the history

  5. Fix in GENERATE_INSTANCES.md

    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    07dc8e7 View commit details
    Browse the repository at this point in the history

  6. Update GENERATE_INSTANCES.md

    @qianfengz
    qianfengz authored Aug 20, 2024
    Configuration menu
    Copy the full SHA
    72bf603 View commit details
    Browse the repository at this point in the history

  7. clean-up commented codes

    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    e397974 View commit details
    Browse the repository at this point in the history

  8. Revert "Change allocation of grouped mode lse from [H, M] to [1, H, M… 

    …] to match the xformers scripts"

This reverts commit 7a91589.
    @qianfengz
    qianfengz committed Aug 20, 2024
    Configuration menu
    Copy the full SHA
    7a04357 View commit details
    Browse the repository at this point in the history

Commits on Aug 22, 2024

  1. Merge branch 'main' into develop-asorb-upstream

    @qianfengz
    qianfengz committed Aug 22, 2024
    Configuration menu
    Copy the full SHA
    2923301 View commit details
    Browse the repository at this point in the history

  2. Merge pull request #22 from ROCm/develop-asorb-upstream 

    Develop asorb upstream
    @qianfengz
    qianfengz authored Aug 22, 2024
    Configuration menu
    Copy the full SHA
    84b50ac View commit details
    Browse the repository at this point in the history

Commits on Aug 26, 2024

  1. Merge remote-tracking branch 'fair/main' into merge-xformers-0826

    @tenpercent
    tenpercent committed Aug 26, 2024
    Configuration menu
    Copy the full SHA
    e0e6863 View commit details
    Browse the repository at this point in the history

Commits on Aug 27, 2024

  1. Merge pull request #23 from tenpercent/merge-xformers-0826 

    Merge facebookresearch xformers into rocm (08/26/24)
    @tenpercent
    tenpercent authored Aug 27, 2024
    Configuration menu
    Copy the full SHA
    e1387a4 View commit details
    Browse the repository at this point in the history

Commits on Sep 3, 2024

  1. Synchronize to latest ck develop for using the latest RTN bf16 convert

    @qianfengz
    qianfengz committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    77a2c24 View commit details
    Browse the repository at this point in the history

  2. Add c++ extension compiling options for better performance on ROCM 6.2

    @qianfengz
    qianfengz committed Sep 3, 2024
    Configuration menu
    Copy the full SHA
    4e51efa View commit details
    Browse the repository at this point in the history

Commits on Sep 5, 2024

  1. Use the same rocm_ci.yml as upstream

    @qianfengz
    qianfengz committed Sep 5, 2024
    Configuration menu
    Copy the full SHA
    887996a View commit details
    Browse the repository at this point in the history

  2. Use the same ck.py as upstream

    @qianfengz
    qianfengz committed Sep 5, 2024
    Configuration menu
    Copy the full SHA
    7c06b55 View commit details
    Browse the repository at this point in the history

Commits on Sep 6, 2024

  1. Reformat setup.py

    @qianfengz
    qianfengz committed Sep 6, 2024
    Configuration menu
    Copy the full SHA
    2efa6cd View commit details
    Browse the repository at this point in the history