Temporary fix for CircleCI errors #378
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
added 2 commits
September 5, 2022 16:46
Codecov Report
@@ Coverage Diff @@
## main #378 +/- ##
==========================================
- Coverage 93.24% 91.19% -2.06%
==========================================
Files 75 75
Lines 4336 4337 +1
==========================================
- Hits 4043 3955 -88
- Misses 293 382 +89
Flags with carried forward coverage won't be shown. Click here to find out more.
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. |
blefaudeux
pushed a commit
to blefaudeux/xformers
that referenced
this pull request
Sep 6, 2022
* Enable masking in memory-efficient attention (facebookresearch#333) * Add attention bias in memory-efficient attention * Add gradient for attn_mask support * Add CPU implementation * clang-format * Add benchmark scripts * Add extra loop in benchmarks * Move zeros array out of helper function * clang-format * Enable dropout in memory-efficient attention (facebookresearch#334) * Merge compute_scaling_coeffs and update_scaling_coeffs into a single function It wasn't needed to break it in two functions to begin with * Add CUDA implementation for dropout * clang-format * Make p be drop probability * Only CUDA supports dropout * Add benchmarks * Remove unused variables * Fix test * Cleanups and comments * Fix masking corner case when full block is masked (facebookresearch#339) * Add cutlass 2.9 - 858c735856a7f17bd33fe438ec76d3c9f0234e7f * Option to load from shared memory for PredicatedTileIterator * Add cutlass include dir * Ignore files in third-party for flake8/coverage * third-party -> third_party * Address comments * Revert some un-needed mods * Add attention_forward_generic.cu * Add tests * Fix duplicate calculations on baseline for mem efficient transformers * Always run all linters in CI * clang-format attention_forward_generic.cu * Benchmark: Add possibility to compare benchmarks * [isort] Ignore third_party * black autoformat * Black again + ignore third_party properly * black * Fix memory leak between the 2 benchmarks in backward * Exclude third_party/ without using pyproject.toml as it imposes isolated build which is a pain * Remove progress bar when finished * mypy * flake8 * Save results to shared folder in home location * run black * clang-format with 'run-clang-format.py' * Fix cutlass build for arch>=75 * Set tests precision for gradient more accurately * Fix precision margin * Revert changes to black * [feat] Fix importing xformers when not built (facebookresearch#351) authored-by: danthe3rd <danthe3rd@users.noreply.github.com> * Update black to 22.3.0 * Tweak precision for mem_eff_attention test * mem-efficient impl for f16 (facebookresearch#352) Co-authored-by: danthe3rd <danthe3rd> * Add support for f16 with tensorcores [sm70/sm75/sm80] (facebookresearch#354) * Add support for f16 with tensorcores * sm75 minimum for tensorcores * Run tests with CUDA_LAUNCH_BLOCKING=1 * Support sm70 properly * Disable tensorcore when not correctly aligned - and use 32bit accessors Co-authored-by: danthe3rd <danthe3rd> Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> * Optimize backward of memory-efficient attention by ~20% (facebookresearch#355) * Optimize backward by 15% by using equivalent formulation * Unify everything into single kernel * Remove unused implementation * clang-format * Remove unused tensor * Display results as we progress during benchmark (facebookresearch#357) Co-authored-by: danthe3rd <danthe3rd> * RFC: Ops dispatch (facebookresearch#356) * Ops dispatch * CI: Fix doc build * memory_efficient_attention raises when no implementation is available * type: ignore * Fix torch.device/str comparison * Make mypy happy Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> Co-authored-by: danthe3rd <danthe3rd> * [A100/f32] Use TensorCores for Q.K_t matmul with FastF32 (facebookresearch#358) * Use TensorCores for MM0 on Float as well * Use MultiStage MMA when available - change to FastF32 rather than FastF16 * Better alignment calculation * Just use regular f32, no fastf32 * Hackfix to handle alignment * HeuristicsMM0 -> GemmTypeQK * No longer use f16 for matmul * Add some doc * Typo * Fix build <sm80 * Alignment check based on current device compute capability * Use TORCH_INTERNAL_ASSERT Co-authored-by: danthe3rd <danthe3rd> * FlashAttention implem and dispatch (facebookresearch#360) * FlashAttention implem WIP * Fix flashattention forward+backward * Fix forward/backward for FlashAttention * Enable tests (more permissive) for f16 backward * Fix CI * flashattn only supports Sm75 and above * Fix CI2 * Disable K=128 when below sm80 for flashattn Co-authored-by: danthe3rd <danthe3rd> * Misc performance improvements for generic mem-efficient attention (facebookresearch#361) * 3% speedup by calculating mi from registers * Also compute m_prime/s_prime and exponentiate from registers * Support for Simt tiles * Fix TensorOp for V100 * Fix for A100 * Fix Simt alignment calculation * clang-format * WarpReduction before atomic call for Simt Co-authored-by: danthe3rd <danthe3rd> Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> * Update flashattention to support bf16 (facebookresearch#363) * Update flashattention to support bf16 * bfloat16 only on sm80 and above Co-authored-by: danthe3rd <danthe3rd> * Flashattn causal (facebookresearch#364) * Implement causal memory-efficient attention with FlashAttention * Update benchmarks * Fix mypy Co-authored-by: danthe3rd <danthe3rd> * Option to disable flashattention (long to build) (facebookresearch#362) * Option to disable flashattention (long to build) * Update setup.py Co-authored-by: danthe3rd <danthe3rd> * Remove code duplicate in attention_scaling_coefs_updater.h (facebookresearch#367) Co-authored-by: danthe3rd <danthe3rd> * Update .gitmodules (facebookresearch#366) * MemoryEff attention forward: Properly fuse matmul and enable TensorCores on the second matmul (facebookresearch#368) * Generic backwards * Guard backward to sm75 only * bounds checking for gradV * clang-format * Fused gemm working for Sm80/Sm75 f16/f32 * WIP * Volta TensorOp for f16 * Working on A100 again * SIMT working * Code cleanup 1 * Code cleanup2 * BUGFIX for shared memory limit * Remove code * clang-format * Remove code again * Remove draft of backward * Enforce alignment for fp16 * Fix tests * Fix constraint on seq length when not using tensorcores * Fix alignment requirements for V100/tensorcores * Clang-format * Update xformers/components/attention/csrc/cuda/attention_forward_generic.cu Co-authored-by: Francisco Massa <fvsmassa@gmail.com> * Address comments from fmassa Co-authored-by: danthe3rd <danthe3rd> Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> Co-authored-by: Francisco Massa <fvsmassa@gmail.com> * Update install instructions with submodule (facebookresearch#365) * Generic backward implem with cutlass (facebookresearch#371) * Old bw code * P100: gradV working * gk/gq working (at least for small values of M, and on P100/f16) * Further restrict supported values for bw * Fix storage into smem for Simt * More tooling for pruint/debug * Remove tests we dont need for now * Tests pass on P100 :D * 4 warps per block * Restraint on q length * Use tensorcores on V100 for f16 * Support dynamic smem for bw * Handle alignment and different dtype/arch * Fix NaNS by initializing shared memory * bw.py * Fix launch bounds * Faster 'computeDi' * minus_lse can operate on arrays * Output number of regs used etc... * Code cleanup * Hackfix for alignment check during forward * zFill to avoid nans in Sm80 + fix launch bounds * COde cleanup1 * clang-format * Fix tests * Add benchmark for K=64 Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> Co-authored-by: danthe3rd <danthe3rd> * Cutlass as submodule (facebookresearch#375) * Make cutlass be back at 858c735856a7f17bd33fe438ec76d3c9f0234e7f * Remove cutlass * Update submodules * Add submodule (properly) * spaces / tab * Make submodule init be recursive * Fix bad rebase * Bump tolerance for backward (facebookresearch#377) * Add verbose flag to CI builds (facebookresearch#376) * Add verbose flag to CI builds * Spurious change to rebuild cache * Add ninja * Ninja wasn't visible before, install through conda * Debugging * Source env * One more try * Forgot to uncomment a line * Another try * Cleanup * Fix for FlashAttention dispatch It requires device capability >= 7.5 * Remove generated file * Address some reviewer feedback Remove unused function and typo fix * Perf improvement on backward (facebookresearch#378) * Fast again on V100 * Fix correctness - missing syncthreads * Get rid of AttentionInfo Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> Co-authored-by: danthe3rd <danthe3rd@users.noreply.github.com> Co-authored-by: dan_the_3rd <43445237+danthe3rd@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Removes functorch install that was causing failures. Will follow up with PR to fix those failures, but created this PR for now to get CircleCI green.
Also fixes some mypy errors.
Before submitting
PR review
Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.