GitLab

[Enhancement] Keep max score attention across blocks in FlashAttention for better numerical stablity (#1269)

* Implement max score retention across blocks in FlashAttention for improved stability

* fix manual pipeline parameters

* Update examples/flash_attention/example_gqa_fwd_varlen.py
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

* fix typo

* more

* fix a previous typo

---------
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

[EXAMPLE] In the flash attention example keep the max of all blocks seen in scores_max numerical stability (#1148)

* Keep the max of all blocks seen in scores_max for stability

* ruff formatting

* [Example] Add page table for gqa decode

* [Example] Page table for varlen decoding

* [Lint]

* [Refactor] Remove redundant code

* [Lint]

* [Lint]

* [Lint]

* fix nsa bwd and atomic

* [Lint]

* [BugFix]
- New implementation for atomicMax and atomicMin using atomicCAS
- PTX version atomicAdd for single 16-byte data
- Modify the test cases

* [Lint]

---------
Co-authored-by: tzj-fxz <tzjfxz@gmail.com>

[BugFix] Refactor attention kernel to handle OOB positions by filling with `-inf` instead of clearing accumulators. (#1222)

* Refactor attention kernel to handle OOB positions by filling with `-inf` instead of clearing accumulators.

* lint

* pre-commit

* Update imports in flash attention test file to use new backward and forward examples for better clarity and consistency.

* [BugFix] Add autotune and exp2 for GDN kernel

* [Lint]

* [Lint]

* fix

* Refactor tensor reshaping in fp8_lighting_indexer.py

- Replaced the allocation of `s_reshaped` with a reshape operation to improve clarity and performance.
- Updated the logic in the computation of `s_reshaped` to utilize the reshaped tensor, enhancing the overall functionality of the attention mechanism.

* Refactor analyzer usage in Layout and Fragment reshaping

- Consolidated analyzer logic in the `Reshape` methods of `LayoutNode` and `FragmentNode` to utilize a fallback analyzer, improving code clarity and preventing potential null dereference issues.
- Updated variable binding and simplification calls to use the selected analyzer consistently, enhancing robustness in shape validation and index computation.

* [Enhancement] Add FP8 support and reproducibility in lighting indexer

* Introduced a manual seed in `test_fp8_lighting_indexer` to ensure reproducible performance.
* Added specializations for `cute::float_e4m3_t` and `cute::float_e5m2_t` in `gemm_mma.h` for enhanced FP8 support across multiple CUDA architectures, ensuring compatibility and improved functionality.ix

* Fix typos in `fp8_lighting_indexer.py` and improve formatting in `gemm_mma.h`

* Corrected a typo in the comment for `test_fp8_lighting_indexer` to enhance clarity.
* Reformatted lines in `gemm_mma.h` for better readability by aligning template specializations across multiple CUDA architectures.

* test fix

* bug fix

* Fix division by zero in RMS normalization

* Fix rsqrt calculation to avoid division by zero

* Add kernel selection option for GEMM v1 in environment settings

- Introduced `TILELANG_USE_GEMM_V1` environment variable to control the selection of GEMM version.
- Added `use_gemm_v1` method in the `Environment` class to determine if GEMM v1 should be used based on the environment variable.
- Updated GEMM function assignment to default to v2, allowing for v1 to be forced via the new environment variable.

* bug fix

* Add kernel selection option for GEMM in environment settings

- Introduced `TILELANG_USE_GEMM_V1` environment variable to allow users to select between GEMM v1 and v2 implementations.
- Updated `gemm` function to default to v2 but switch to v1 if the environment variable is set to a truthy value.
- Added a method `use_gemm_v1` in the `Environment` class to facilitate this selection based on the environment variable.

* Refactor GEMM macro generator to use BufferRegion instead of Buffer

- Updated `wgmma` and `wgmma_rs` methods in `TensorCoreIntrinEmitter` to accept `BufferRegion` parameters instead of `Buffer`.
- Adjusted related calls in `GemmWGMMA` to ensure compatibility with the new parameter types.
- Simplified buffer access logic for better clarity and maintainability.

* Refactor GEMM functions to utilize BufferRegion for improved memory handling

- Updated `run_gemm`, `run_gemm_rs`, `run_gemm_sr`, and `run_gemm_rr` functions to set `num_stages` based on block dimensions, enhancing performance for larger matrices.
- Simplified calls to GEMM functions by removing redundant parameters and ensuring compatibility with BufferRegion.
- Introduced utility functions for converting between Buffer, BufferLoad, and BufferRegion, improving code clarity and maintainability.
- Enhanced error handling for full region checks in GEMM operations to ensure correctness in memory access.

* Refactor GEMM code for improved readability and consistency

- Cleaned up formatting and spacing in GEMM-related files for better readability.
- Standardized comments and code structure across various GEMM functions and macros.
- Enhanced error messages for clarity in buffer region checks.
- Removed redundant lines and improved overall code maintainability.

* Update GEMM correctness evaluation and macro generator for improved functionality

- Modified `N_VALUES` in `correctness_evaluation_sm70.py` to include only relevant sizes for tests.
- Updated test function call in `correctness_evaluation.py` to use `test_gemm_false_true` for better accuracy in testing.
- Refactored buffer handling in `mma_sm70_macro_generator.py` to improve clarity and consistency in shared buffer access.
- Enhanced `gemm_mma_sm70.py` to ensure full region checks for input and output buffers, improving correctness in GEMM operations.

* Refactor GEMM and intrinsic files for improved clarity and functionality

- Removed unused variable `A_stride_last` in `mma_sm70_macro_generator.py` to streamline code.
- Adjusted function signature formatting in `swizzle.py` for better readability.
- Restored the return of `GemmWGMMA` in `__init__.py` for correct GEMM instantiation.
- Removed unused variable `B_buf` in `gemm_mma_sm70.py` to enhance code cleanliness.
- Improved function signature formatting in `language.py` for consistency.

* Enhance GEMM and MMA functionality for FP64 support

- Refactored `GemmNode` to streamline the decision-making process for GEMM instruction selection.
- Added support for FP64 inputs in the MMA dispatcher, enabling new tensor operations.
- Introduced a new layout function for FP64 in `mma_layout.py` to facilitate shared memory storage.
- Updated `TensorCoreIntrinEmitter` to handle FP64 data types, including adjustments for micro tile dimensions and loading mechanisms.
- Enhanced utility functions to accommodate FP64 index mapping for shared memory operations.

* lint fix

* Refactor GEMM correctness evaluation and shared memory alignment handling

- Reverted the GEMM function call in `correctness_evaluation.py` to the original implementation for consistency.
- Added a helper function in `merge_shared_memory_allocations.cc` to streamline the marking of shared variables under alignment scope.
- Enhanced the `VisitExpr_` methods to ensure proper handling of shared memory alignment for `BufferLoadNode` and `VarNode` types.
- Cleaned up commented-out test code in `correctness_evaluation.py` for better readability.

* Enhance GEMM and MMA implementations with region-based memory handling

- Updated GEMM and MMA classes to utilize BufferRegion for input and output buffers, improving memory management and supporting strided GEMM operations.
- Added checks to ensure full region compliance for input buffers, enhancing correctness in matrix multiplication.
- Implemented clear accumulation functionality to reset output buffers before accumulation, ensuring accurate results in GEMM operations.

* Refactor test_tilelang_example_deepseek_v32.py to improve import structure and function calls

- Updated import statements to directly reference modules instead of individual test functions, enhancing clarity.
- Modified function calls to use the new module structure for better organization and maintainability in testing examples.

* Enhance OnArrayDeclaration method to handle repeated buffer declarations

- Updated the OnArrayDeclaration method to merge metadata for buffers that may appear in multiple Allocate statements, improving robustness against upstream transformations.
- Added logic to prefer concrete element data types and record extents when previously unknown, enhancing the handling of buffer declarations.

* Add abbreviation for bfloat16 data type in mfma_macro_generator.py

- Introduced a new abbreviation "bf16" for the bfloat16 data type in the mfma_macro_generator.py file, enhancing clarity and consistency in data type representation.

* Refactor CodeGenTileLangHIP to enhance dtype handling and mfma call generation

- Introduced a mapping function to normalize input data types to their corresponding scalar types, improving compatibility with MfmaTraits.
- Updated the mfma call generation to utilize the new mapping, streamlining the code and enhancing clarity.
- Removed outdated dtype mapping and replaced it with a more flexible approach to support additional data types like FP8.

* lint fix

* Enhance backend configuration in CMakeLists.txt and improve dtype handling in CodeGenTileLangHIP

- Introduced a macro to define backend options for CUDA, ROCM, and Metal, allowing user overrides and caching of settings.
- Updated logic to track user-selected backends and conditionally enable defaults based on environment variables.
- Refactored dtype handling in CodeGenTileLangHIP to streamline mfma call generation and improve clarity.
- Added support for bfloat16 in the mfma_macro_generator.py, enhancing data type representation consistency.

* Update bfloat16 handling in CodeGenTileLangHIP and mfma_macro_generator.py

- Changed the representation of bfloat16 in CodeGenTileLangHIP from "bfloat16x4" to "bfloat16x4_vec" for improved clarity.
- Adjusted the mfma_suffix generation in mfma_macro_generator.py to remove the underscore before "bf16", aligning with HIP intrinsic requirements.

* Change logging level from WARNING to DLOG in LegalizeNegativeIndex for non-negative index checks to reduce log verbosity.

* Refactor attention sink examples to simplify index calculations

- Updated index handling in `example_gqa_sink_bwd_bhsd.py` and `example_mha_sink_bwd_bhsd.py` to eliminate unnecessary local allocations and streamline logic for determining start and end indices.
- Improved readability by using direct calculations instead of local variables for index bounds in pipelined loops.

* Refactor attention sink examples to streamline index calculations

- Simplified index handling in `example_gqa_sink_bwd_bhsd.py`, `example_gqa_sink_fwd_bhsd_wgmma_pipelined.py`, `example_mha_sink_bwd_bhsd.py`, `example_mha_sink_fwd_bhsd_wgmma_pipelined.py`, and `example_mha_sink_fwd_bhsd.py` by removing unnecessary local allocations for start and end indices.
- Enhanced readability by directly calculating index bounds for pipelined loops, improving overall code clarity.

* lint fix

* bugfix

* Refactor reduce operation handling in CUDA and Python

- Removed outdated shared memory reduction logic from `reduce.cc`.
- Introduced fragment allocation and improved buffer handling in `reduce.py` to support shared and fragment scopes.
- Updated CUDA header to define a wider accumulator type for better numerical accuracy.
- Enhanced error handling for buffer scope validation in the reduction process.

* Fix ReduceOpNode to correctly compute AbsMax by using absolute values of inputs

* Enhance unit loop handling by refining annotation checks

- Updated the condition for identifying effectively empty annotations in unit loops to include cases where only the `pragma_unroll_explicit` hint is present.
- Introduced a new method, `IsEffectivelyEmptyAnnotation`, to encapsulate this logic, improving code clarity and maintainability.

* clean clode

* [Example] Add GQA varlen decoding kernel with logits return

* [Example] Support Sink for GQA varlen decoding

* [Example] Add for no-varlen support

* [Tune] Add high performance logits saving

* [Lint]

* [Lint]

* [Rename]

* [Example] Update GQA varlen fwd

* fix

* [Test] Add cp async to avoid register spill

* [BugFix] GQA fwd and bwd
- Fix the undefined behavior of -inf in acc_s
- Fix the causal loop range in varlen scenario

* [TMA] Move on to TMA and locate the register spill issue

* [Debug] Not the reason of zero-assignment. Probably the combination of Parallel op & conditional qkT

* [Debug] The SIMT copy in producer occupies too many registers

* [BugFix] Use 3D lse and delta to avoid illegal instruction

* [Perf] Relaxed order for dQ and SIMT store for dKdV

* [Feat] For atomic add version

* [Lint]

* [Bugfix] Enable code lowering with producer‑copy‑only program (#1168)

* bugfix

* lint fix

* Enhance warp group register allocation to handle missing consumer bodies gracefully. Updated logic to annotate producer side when consumer is absent, ensuring robustness in degenerate warp-specialized patterns.

* Refactor VisitExpr_ method in inject_tma_barrier.cc for improved readability. Adjusted formatting and spacing for clarity in barrier handling logic.

* Update barrier handling in inject_tma_barrier.cc to accommodate newly appended entries. Adjusted the size of the replace vector to ensure it covers the full needed length, and modified the logic for appending barriers based on the updated replace conditions.

* [Bugfix] Support 16bits shfl_sync (#1169)

* Add type-safe warp shuffle helpers for 16-bit float types in common.h

- Introduced generic passthrough functions for warp shuffle operations: `shfl_xor_sync`, `shfl_down_sync`, `shfl_up_sync`, and `shfl_sync`.
- Added specializations for `cutlass::half_t` and `cutlass::bfloat16_t` to ensure type safety during shuffle operations.
- Updated `reduce.h` to utilize the new shuffle functions, enhancing code clarity and maintainability.

* lint fix

* [Testing] Move TMA 1D and test for its functionality (#1167)

* [Testing] Move TMA 1D and test for its functionality

* [Lint]

* [Refactor]: Change the params in pytest to avoid oom error during ci (#1170)

* [Refactor]: Change the params in pytest to avoid oom error during ci

* format

* fix

* Update test_example_cast.py

* Update parameters in test_example_cast

* Update test_example_flash_attention.py

* update

* format

* fix

* fix

* format

* [Bugfix] Fix tvm import path for editable build (#1172)

* [Language] Expose `T.warpgroup_fence_operand` for nvcc code motion (#986)

* remove debug print

* pipeline fix

* use the correct buffer access scope

* rs support

* warp warpgroup_fence_operand

* fix

* fp8 dtype ptx enhance

* mma fix

* TCGEN05 Interface

* tcgen05 support

* rebase

* update

* Enhance TCGEN05 support by adding new intrinsic operations and descriptors. Introduced `ptx_tcgen05_mma_ts` for tensor-memory to shared-memory instructions and `tcgen05_mma_arrive` for signaling barrier completion. Updated existing descriptors and code generation logic to accommodate these changes, ensuring compatibility with new instruction sets. Refactored related allocation functions and improved handling of shared memory descriptors.

* lint fix

* Refactor buffer reference handling in CUDA code generation and update test execution in tilelang. Ensure default annotations for unrolling are set correctly in TIR IR module.

* wgmma fix

---------
Co-authored-by: Zhiwen Mo <zm125@ic.ac.uk>

* [Language] Add Correctness and performance check scripts for V2 (#1174)

* fix

* lint fix

* fix

* lint fix

* fix

* upd

* [Bugfix] Legalize Datatype for mma intrinisc codegen  (#1179)

* fix

* lint fix

* Enhance CUDA code generation by updating register type handling for float data types. Introduced a workaround for TF32 type compatibility and improved the registration of MMA register types for A and B operands.

* [Perf] Add layout and use_tma to boost performance

* [Lint]

* [Note]

---------
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>
Co-authored-by: Yuqi Dong <134183314+yyttt6@users.noreply.github.com>
Co-authored-by: Zhiwen Mo <zm125@ic.ac.uk>

* [Feature] Enhance fill operation to support various buffer types

- Added support for `BufferLoad` in the `fill` function to handle different buffer types.
- Updated `Fill` class to process region descriptors and buffer regions, improving flexibility in buffer handling.
- Introduced checks for static bounds in region definitions to ensure safety during operations.
- Refactored loop induction variable handling in `FillNode` to accommodate sliced regions.

* lint fix

* [Refactor] Improve Python compatibility for ParamSpec and Self

- Added compatibility handling for ParamSpec and Self to support Python versions below 3.10 and 3.11 respectively.
- Updated type annotations across multiple files to ensure consistent usage of typing features.

* [Update] Require Python 3.9 and enhance type annotations

- Updated the minimum required Python version from 3.8 to 3.9 in `pyproject.toml`.
- Removed references to Python 3.8 in classifiers.
- Changed type annotations from `int | None` to `Optional[int]` in multiple example files for better clarity and compatibility.
- Improved import statements to use `collections.abc` for `Iterable` and `contextlib` for `AbstractContextManager` in relevant files.

* [Refactor] Update import statements to enhance type annotations

- Replaced imports from `typing` with `collections.abc` for `Iterable` and `Mapping` in relevant files to improve compatibility and clarity.
- Updated the caching decorator from `functools.lru_cache` to `functools.cache` for better performance in the C++ compiler retrieval function.
- Adjusted import statements in the language proxy file to maintain consistency in type annotations.

* disable rocm rs nt test.

* lint fix

* tilelang frontend v2

* syntax sugar: defining a local var by annotation

* [Refactor] fix type linting warning like `T.float32`

* Add tl.local_var_init for new tl.float32

* allow passing default argument as function annotation

* allow default arguments as annotation

* fix lint error

* minor fix

* [Refactor] refactor tilelang.jit and tilelang.autotune

* minor fix

* minor fix

* minor fix

* fix metal get function name

* add par_compile impl and tests

* Type consistency on tvm datatype
1. isinstance(tl.float32, tvm.DataType) == True
2. Allow `tl.float32` as function annotations
3. Allow `tl.float32` as argument to be passed to `tl.alloc` or other functions

* fix lint error

* add more warning in frontend

* update tvm version

* Minor fix on tvm_ffi annotations

* add document and examples

* fix lint error

* Simplify index calculations in example_chunk_o_bwd.py

Refactor index calculations for dg_last_fragment assignment.

* minor fix

* lint fix

---------
Co-authored-by: Lei Wang <leiwang1999@outlook.com>
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

* [Refactor]: Change the params in pytest to avoid oom error during ci

* format

* fix

* Update test_example_cast.py

* Update parameters in test_example_cast

* Update test_example_flash_attention.py

* update

* format

* fix

* fix

* format

* [Testing] Move TMA 1D and test for its functionality

* [Lint]

* 3rdparty tvm bump

* bump tvm into v0.22.0

* lint fix

* rebase tvm

* Update submodule tvm to latest commit 3085bc4

* Refactor: Update configuration retrieval in CopyNode and adjust test registration in tilelang

* test fix

* add requirement

* atomic_fix

* atomic_fix

* phaseout py39

* optimize

* optimize

* lint fix

* do not clean cache

* do not clean cache

* [Minor] Minor update for Python versions and dependencies

* [Lint] fix lint for py39

* [Lint] fix lint for ROCm

* [Build][CI] Sync CI changes from upstream/sdist

* [Lint] fix lint for ROCm

* [Build][CI] Update `repair-wheel-command`

* [Minor] update abi3audit result format

* [Lint] fix lint for ROCm

* [BugFix] fix build

* [Lint] fix lint for ROCm

* [BugFix] set rpath for libtvm and libtvm_runtime

* [Deps] pin apache-tvm-ffi version

* [Build] set Python 3.9 Limited API for Cython target

* [Build] set Python 3.9 Limited API for Cython target

* [Deps] Restore Python 3.8 support

* [Build] use `apache-tvm-ffi`'s `libtvm_ffi`

* [BugFix] use `;` as delimiter for RPATH on macOS

* [BugFix] use `--ignore-missing-dependencies` for `delocate-wheel`

* [Build] support `sccache` if available

* [Build] add CIBW import test

* [Build][CI] enable ccache for CIBW on Linux

* [BugFix] set rpath for libtvm and libtvm_runtime

* Revert "[Build][CI] enable ccache for CIBW on Linux"

This reverts commit cd9ab57bb5ddd2572c60bcbbebde81480a658fd3.

* [CI] fix perfbench bot

* [BugFix] use Python 3.9 to build wheel

* [Minor] update perfbench bot envs

* [BugFix] fix CIBW environment on Linux

* [CI] skip import test on CentOS 7

* [CI] use Python urllib to download file instead of Wget

---------
Co-authored-by: Xuehai Pan <XuehaiPan@pku.edu.cn>

* [BugFix] Add memory order for split version kernel; Remove torch manual seed

* [Lint] Manual

* [Feature] Add vectorized float16 and float32 conversion support in CUDA codegen

* Implemented handling for conversions between float16 and float32 types, specifically for vectorized operations using __half22float2 and __float22half2_rn.
* Enhanced the existing code to support both directions of conversion based on the lane count.
* Improved overall type handling in the VisitExpr_ method for better compatibility with TileLang.

* [Feature] Add float32 to float8 conversion support in CUDA codegen

* Implemented handling for conversion from float32 to float8 (E4M3/E5M2) in the VisitExpr_ method.
* Added vectorized conversion support using __nv_cvt_float2_to_fp8x2 for float2 to fp8x2 transformations.
* Enhanced type handling for better compatibility with TileLang, particularly for float8 types.

* lint

* fix a bug

* [Enhancement] Support lanes=4 cases and add unit test for vectorized cast

* lint

* [Feature] Refactor bf16 convertion operations and remove legacy compile flags

* lint

* add alloc_reducer gemv example

* test

* [Lint]

* [BugFix] Freeze the memory order of all atomic_add operations

* [Lint]

* [Atomic] Move on to regional atomic add

* [Lint]

* [Cleanup] Remove `tilelang.disable_cache()` calls from example scripts

* lint

* lint

* [Feature] Add GQA backward kernel with varlen input

* [Lint]

* [BugFix] Freeze the memory order of all atomic_add operations

* [Lint]

* [Lint]

* [BugFix] Use release order to boost performance

* recommend using T.dynamic instead of T.symbolic

* lint fix

* lint fix

* [Feature] Support Reduce operators for bitwise and/or/xor

* [Lint]

[Enhancement] Deprecate split&sum in attn bwd examples on Hopper and migrate to vectorized atomic add (#1065)

* [Refactor][Example] Update linear attention examples and add tests

- Refactored the backward and forward linear attention kernels to use shared memory and atomic additions for improved performance.
- Introduced L2 normalization in the main functions of both examples.
- Added a new test suite for the linear attention examples to ensure correctness and performance.
- Updated argument parsing in the main functions for better usability.

* upd docstring for tma atomic add

* lint

* Add flash-linear-attention dependency to requirements.txt

* Rename main function to chunk_linear_attn_bwd

* Rename main function to chunk_linear_attn_fwd

* chore

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

* [CI]:Reduce test shapes to avoid OOM errors during CI.

* rabbit

* Increase number of processes for pytest from 2 to 4

---------
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

[Enhancement] Add support for symbolic dimensions in Cython kernel adapter and improve static shape validation in wrapper (#1024)

* [Enhancement] Add support for symbolic dimensions in Cython kernel adapter and improve static shape validation in wrapper

* [BugFix] Fix shape mismatch and deprecate `T.if()` in fused_moe example

* [Fix] Add `is_symbolic_expr` function to check for symbolic expressions in TIR

- Introduced a new utility function `is_symbolic_expr` to determine if an expression is a symbolic expression, enhancing type checking capabilities.
- Updated shape handling in `CythonKernelAdapter` to utilize the new function, improving handling for symbolic shapes.

* [CI] fix ROCm CI

* feat: add a hook to error out on no test runs

* [BugFix] Phaseout dependency of Triton in sink examples to make CI happy

- Added `benchmark_gqa_sink_fwd.py` and `benchmark_mha_sink_fwd.py` to evaluate performance of GQA and MHA attention mechanisms using Triton.
- Refactored existing attention sink implementations to remove Triton kernel definitions from the reference programs, streamlining the code.
- Updated input generation and benchmarking logic to enhance configurability and performance measurement.
- Improved overall structure and organization of the examples for better clarity and usability.

* [Lint]: [pre-commit.ci] auto fixes [...]

---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* refactor: merge test CI workflow files into one

* chore: set `UV_INDEX_STRATEGY=unsafe-best-match`

* feat: add AST test with Python 3.8

* feat: implement manual caching mechanism for self-hosted runners

* refactor: simplify cache logic for self-hosted runners

* chore: clear uv cache on failure

* chore: print format.sh output to logs

* chore: improve uv caching

* chore: disable parallel test

* chore: use `PYTHONDEVMODE=1` in CI

* feat: enable coredump generation

* fix: fix perfbench condition

* Revert "feat: enable coredump generation"

This reverts commit c52da65cb572932e09905d08c43a39ec3cf47c54.

* chore: move example CI down

* Revert "chore: move example CI down"

This reverts commit 9d8e65055e01d955c5268a9a6705d270c2de0d57.

* chore: skip example `test_example_mha_sink_bwd_bhsd`

* chore: skip example `test_example_gqa_sink_bwd_bhsd`

* fix: fix example argument passing

* fix: loosen test criteria

* chore: rename `CMAKE_CONFIG...

* [Enhancement] Refactor buffer index handling for improved precision and clarity (#668)

- Enhanced buffer index handling to address precision issues by removing redundant operations.
- Streamlined the logic for determining buffer overlaps, ensuring more accurate conflict detection.
- Updated related documentation to reflect changes in buffer management practices.

* Remove obsolete test script for AMD example, streamlining the examples directory.

* Remove unused dtype_size variable in AMD example script to streamline code.

* Add input configuration file and update AMD example script for enhanced flexibility

- Introduced a new input.txt file for configurable parameters.
- Modified the example_amd_flash_attn_fwd.py script to allow for a wider range of configurations, including additional options for num_stages, enable_rasterization, and k_pack.
- Streamlined the main function for better clarity and organization.
- Added a new test script to facilitate running the example with specified parameters.

* Remove input configuration file and obsolete test script; enhance AMD example with swizzle layout annotations

- Deleted input.txt and test.sh files as they are no longer needed.
- Updated example_amd_flash_attn_fwd.py to include swizzle layout annotations for shared memory, improving bank conflict avoidance.
- Reintroduced swizzle usage in the kernel for better performance.

* Refactor AMD example script for FlashAttention-2

- Updated function names for clarity, changing `get_v2_configs` to `get_configs` and `fast_flashattn_v2` to `fast_flashattn`.
- Streamlined the main function by renaming `main_v2` to `main` and adjusting the corresponding calls.
- Removed outdated comments and improved code organization for better readability.

* Refactor formatting in AMD FlashAttention example script

- Improved code readability by adjusting line breaks and indentation in the `fast_flashattn` function.
- Streamlined the `main` function parameter formatting for consistency.
- Removed unnecessary blank lines to enhance overall code organization.

* Update example_amd_flash_attn_fwd.py

* Enhance AMD example script and update CI workflows

- Improved the `example_amd_flash_attn_fwd.py` script for better clarity and organization.
- Added new CI workflows for AMD and documentation publishing.
- Updated various requirements files to include necessary dependencies.
- Introduced new test cases and examples for better coverage and functionality.
- Refactored existing code for improved readability and maintainability.

* Remove redundant tool cache cleanup step in AMD CI workflow

* Remove `torch` dependency from `requirements-rocm.txt` to streamline requirements.

* Add new AMD FlashAttention example and test script

- Introduced `example_amd_flash_attn_bwd.py` for backward attention computation using TileLang.
- Added `test.sh` script to facilitate running the new example with specified parameters.
- Enhanced the overall structure and organization of the example for better clarity and usability.

* Update configurations in `example_amd_flash_attn_fwd.py` for autotuner

- Reduced the number of threads and `num_split_q` options for improved performance.
- Adjusted `panel_size` options to streamline configuration settings.

* Update submodule 'tvm' to commit 6ccc74f622c7ec4ac25d430d0f6546e7b9edb217

* Update submodule 'tvm' to commit 14ff70ab142b9e5a31bbf9c7923c8a697d41e86c

* Add example for AMD Flash Attention backward pass implementation

- Introduced a new example script `example_amd_flash_attn_bwd.py` demonstrating the forward and backward operations of Flash Attention using TileLang.
- Implemented JIT-compiled functions for both forward and backward passes, including preprocessing and postprocessing steps.
- Added a main function to facilitate testing and benchmarking of the attention mechanism with configurable parameters.
- Included reference implementation for validation against PyTorch's attention mechanism.

This addition enhances the examples directory by providing a comprehensive guide for users to understand and utilize Flash Attention in their applications.

* Enhance AMD Flash Attention example with additional testing capabilities

- Updated `example_amd_flash_attn_bwd.py` to include more comprehensive testing features for the Flash Attention implementation.
- Improved the main function to allow for better parameter configuration and benchmarking.
- Added validation checks against PyTorch's attention mechanism to ensure accuracy and reliability of the example.

This update aims to provide users with a more robust tool for understanding and utilizing Flash Attention in their applications.

* Update submodule TVM to commit a64a5926a6e59f5417ef2501f9d88b467337cf6a

* Refactor HIP intrinsic rules to CUDA

- Updated file name from `intrin_rule_hip.cc` to `intrin_rule_cuda.cc` to reflect the change in focus from HIP to CUDA intrinsic rules.
- Adjusted include paths for better organization and clarity in the code structure.

* Update AMD CI workflow to uninstall specific PyTorch packages before installation

- Removed the installation of `flash_attn==2.5.8` to streamline the CI process.
- Added a step to uninstall `torch`, `torchvision`, and `torchaudio` prior to installing pre-release versions, ensuring compatibility and reducing potential conflicts.

* Remove unused shared memory allocations in AMD Flash Attention backward example

- Eliminated the allocation of shared memory for `dv_shared` and `dk_shared` in `example_amd_flash_attn_bwd.py` to streamline memory usage and improve performance.
- This change focuses on optimizing the backward pass implementation by reducing unnecessary memory overhead.

* Remove unnecessary pip uninstall command from AMD CI workflow

- Eliminated the step to uninstall `torch`, `torchvision`, and `torchaudio` in the AMD CI workflow, as it is no longer required for the installation of pre-release versions.
- This change simplifies the CI process and reduces potential overhead during package management.

* Refactor DispatchHIPWarpActiveMask function in HIP intrinsic rules

- Updated the return statement to use std::string for concatenation in the case of 16-bit types, improving code clarity.
- Added a null check for the CallNode pointer in DispatchHIPWarpActiveMask to enhance robustness and prevent potential dereferencing issues.

* Refactor formatting of HIP intrinsic rule registrations

- Adjusted the formatting of TVM_REGISTER_OP calls for better readability by aligning method chaining.
- No functional changes were made; this update focuses on code style improvements to enhance maintainability.

* Update file name and documentation for HIP intrinsic rules

- Renamed the file from `intrin_rule_cuda.cc` to `intrin_rule_hip.cc` to accurately reflect the focus on HIP intrinsic rules.
- Updated the file documentation to clarify its purpose as related to HIP rather than CUDA.

* Enhance DispatchHIPShuffle function with clang-analyzer comments

- Added NOLINTBEGIN and NOLINTEND comments to the DispatchHIPShuffle function to suppress clang-analyzer warnings related to inner pointer usage.
- This change improves code clarity and maintains compliance with static analysis tools.

* lint fix

* fix

* Enhance autotuner configurations in example_amd_flash_attn_fwd.py by adding new block sizes, stages, and panel sizes. Update test script to use relative Python path and adjust parameters for consistency.

* Add backward attention example to test script

- Extended the test.sh script to include a new backward attention example using example_amd_flash_attn_bwd.py.
- Added parameters for batch size, context length, and head dimensions to ensure consistency with the forward example.
- Updated the command for the backward tile example to match the new configuration.

* Refactor FlashAttention implementation in example_amd_flash_attn_bwd.py and example_amd_flash_attn_fwd.py

- Introduced new functions for forward and backward configurations to enhance autotuning capabilities.
- Updated the FlashAttention forward and backward functions to improve performance and maintainability.
- Adjusted test script parameters for consistency and clarity, including the addition of group handling.
- Enhanced the autotuner configurations by refining block sizes and stages for better performance tuning.
- Updated the main function to reflect changes in parameter names and types for better usability.

* Enhance FlashAttention backward implementation in example_amd_flash_attn_bwd.py

- Updated the backward function to return additional outputs, including log-sum-exp (LSE) values for improved gradient calculations.
- Refined autotuner configurations by adding new block sizes and adjusting parameters for better performance tuning.
- Improved shared memory usage in the backward pass to optimize memory access patterns and enhance computational efficiency.
- Updated the main function to reflect changes in parameter handling and ensure consistency with the forward pass.
- Enhanced correctness checks in the main function to include LSE validation alongside gradient checks.

* Enhance FlashAttention backward implementation in example_amd_flash_attn_bwd.py

- Introduced a scaling factor for improved numerical stability in gradient calculations.
- Optimized shared memory usage by adding new shared buffers for intermediate calculations.
- Refined the handling of tensor fragments to improve performance and maintainability.
- Updated the main function to ensure compatibility with the new output parameters for backward operations.
- Removed unnecessary parameters from the test script to streamline execution.

* Refactor FlashAttention implementation in example_amd_flash_attn_bwd.py and example_mha_bwd.py

- Updated the forward and backward functions to improve numerical stability and performance.
- Enhanced shared memory usage by optimizing buffer allocations and reducing unnecessary parameters.
- Adjusted autotuner configurations for better performance tuning and compatibility with new output parameters.
- Added debugging and benchmarking functions for improved correctness verification and performance analysis.
- Updated the main function to reflect changes in parameter handling and ensure consistency across examples.

* Enhance FlashAttention backward implementation in example_amd_flash_attn_bwd.py

- Updated scaling factor application for improved numerical stability in gradient calculations.
- Refined tensor handling to ensure consistency with forward pass operations.
- Optimized atomic operations for writing gradients to dK and dV using fp32 for better precision.
- Adjusted comments for clarity and alignment with standard implementation practices.

* Expand autotuner configurations in example_amd_flash_attn_bwd.py and update test.sh

- Increased the range of block sizes and stages for forward and backward configurations to enhance performance tuning.
- Adjusted the test script to include additional parameters for batch size and head dimensions, ensuring consistency with the forward example.
- Improved comments for clarity and alignment with the updated configurations.

* Enhance performance calculations and benchmarking in example_amd_flash_attn_bwd.py

- Updated FLOPs calculation to account for both forward and backward passes, clarifying the total computational cost.
- Modified benchmarking functions to evaluate the complete forward and backward performance of both reference and Tile-lang implementations.
- Improved comments for better understanding of the performance metrics and implementation details.
- Removed unnecessary parameter from test.sh to streamline execution.

* Remove forward attention test commands from test.sh and retain backward attention execution for streamlined testing.

* Refactor FlashAttention forward and backward implementations in example_amd_flash_attn_bwd.py and example_amd_flash_attn_fwd.py

- Updated the forward function to return both output and log-sum-exp (LSE) values for improved gradient calculations.
- Enhanced autotuner configurations for forward pass, including new parameters for better performance tuning.
- Refined scaling factor calculations for numerical stability in both forward and backward passes.
- Improved comments and documentation for clarity and consistency across implementations.
- Adjusted main function to reflect changes in parameter handling and ensure compatibility with new output requirements.

* Refactor FlashAttention implementation in example_amd_flash_attn_bwd.py

- Removed outdated comments and improved clarity in the code.
- Enhanced the forward function to consistently return output and log-sum-exp (LSE) values.
- Updated autotuner configurations to include new parameters for better performance tuning.
- Refined tensor handling and scaling factor calculations for improved numerical stability.
- Adjusted the main function to ensure compatibility with updated output requirements and parameter handling.

* Enhance FlashAttention backward implementation in example_amd_flash_attn_bwd.py

- Updated configuration parameters for backward calculations, including new options for block sizes, threads, and rasterization.
- Added new parameters (k_pack, qk_coalesced_width, v_coalesced_width) to improve performance tuning and memory access patterns.
- Modified tensor copy operations to utilize coalesced widths for optimized memory loads.
- Enhanced GEMM operations with k_pack for improved computational efficiency.
- Refined the configuration generation logic to accommodate the new parameters, ensuring comprehensive coverage for backward pass scenarios.

* Refactor configuration and tensor operations in example_amd_flash_attn_bwd.py

- Updated backward configuration parameters to include larger block sizes and a wider range of threads for enhanced performance tuning.
- Removed unnecessary parameters (k_pack, qk_coalesced_width, v_coalesced_width) from function signatures and tensor operations to simplify the implementation.
- Optimized tensor copy operations by eliminating coalesced width specifications, streamlining memory access patterns.
- Adjusted GEMM operations to improve computational efficiency without the use of k_pack.

* Enhance HIP code generation and FP8 type support

- Added support for additional FP8 types (e4m3, e4m3b11fnuz, e5m2fnuz, e8m0) in codegen_hip.cc to improve compatibility.
- Updated error logging to include unsupported FP8 type details for better debugging.
- Implemented handling for loop break and no-op register management in HIP within VisitExpr_ method.
- Introduced new FP8 vector types (e5 and e8) in hip_fp8.h for enhanced functionality.
- Added overloads for AtomicAdd in common.h to support both pointer and value arguments.

* Enhance FP8 type support and clarify accumulator handling in HIP

- Expanded FP8 type support in codegen_hip.cc to include additional float8 formats.
- Updated gemm.h to clarify the handling of the accumulator when clear_accum is true.
- Added comments in hip_fp8.h to indicate that E8M0 types are not supported in the current HIP version.

* Remove deprecated files and update print statements for clarity in example_amd_flash_attn_bwd.py

* Update print statement formatting for clarity in example_amd_flash_attn_bwd.py

* Remove redundant verification results summary print statement in example_amd_flash_attn_bwd.py for cleaner output.

* Fix formatting inconsistencies in example_amd_flash_attn_bwd.py and example_amd_flash_attn_fwd.py by adding spaces for improved readability in configuration parameters and print statements.

* Refactor and enhance HIP code generation for improved FP8 support

- Reorganized and cleaned up code in codegen_hip.cc for better readability and maintainability.
- Enhanced handling of FP8 types, including additional formats and improved error logging for unsupported types.
- Updated AtomicAdd function in common.h to streamline its implementation.
- Refined the PrintVecElemLoadExpr method to handle volatile loads more effectively.
- Added function to manage the addition of new functions in the code generation process.

* Fix formatting issue in HIP code generation for MFMA call

- Adjusted the indentation of the MFMA call code block in codegen_hip.cc for improved readability and consistency.

* Refactor HIP code generation and enhance FP8 type handling

- Reintroduced necessary includes and reorganized code in codegen_hip.cc for improved structure and readability.
- Enhanced the GetFP8Type function to support additional FP8 formats and improved error handling for unsupported types.
- Updated PrintType and PrintVecElemLoadExpr methods to better manage type conversions and vector element loading.
- Refined the AddFunction method to streamline function addition in the code generation process.

* Remove unnecessary blank line in example_amd_flash_attn_bwd.py for improved code cleanliness.

* Refactor backward attention implementation in example_amd_flash_attn_bwd.py

- Updated the GEMM operation to use shared memory for improved performance.
- Adjusted parallelization parameters to enhance efficiency in the backward pass.

* Fix formatting by removing an unnecessary blank line in example_amd_flash_attn_bwd.py for improved code cleanliness.

* Add additional test cases for `assert_tl_matmul_correctness` with `float8_e4m3fnuz` and various configurations

* Refactor test case formatting for `assert_tl_matmul_correctness` in `test_tilelang_gemm_mfma_intrinsic.py`

---------
Co-authored-by: xinxyxiao <xinyxiao@amd.com>
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>