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>

* 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

* [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

* [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

* revert split+sum template for MHA backward

* lint

* Update example_mha_bwd.py

* Update example_mha_bwd_wgmma_pipelined.py

* Refactor attention sink examples to support bf16 and user-defined softmax scale

* fix typos

* Adding compile flags for fast math optimizations and enabling BF16 support in both GQA and MHA backward implementations.

* Update backward configuration for GQA and MHA examples to align with flash attention

* Refactor GQA backward implementation to improve atomic add performance

* Allow for slightly larger numerical error for bf16

* upd readme to show bf16 benchmark results

* lint

* fix ci and lint

* fix comments and lint

* refactor atomic add

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

* [Example] Add a new example to support attention sink for MHA

- Introduced a new example script for multi-head attention (MHA) with sliding window attention and sink tokens.
- Added a reference attention function to validate the implementation against PyTorch.
- Included argument parsing for command-line execution of the example.

* [Example] Replace MHA sink forward example with updated implementation

- Removed the old example script for multi-head attention (MHA) with sliding window attention and sink tokens.
- Introduced a new example script that modifies the attention mechanism to enhance performance and maintainability.
- Updated argument parsing and reference functions to align with the new implementation.

* Enhance MHA sink example with sliding window support

- Added a `window_size` parameter to the `flashattn` function to enable sliding window attention.
- Implemented assertions to ensure `window_size` is compatible with `block_N`.
- Updated the main function to include a `tune` option for performance tuning.
- Introduced a new test file to validate both full attention and sliding window scenarios.
- Adjusted FLOPS calculation to account for the sliding window configuration.

* lint

* [Fix] Add checkinf process to fix the bug of swa

* Migrate to BSHD layout to align with triton baselines

* lint

* fix typo

* Refactor MHA sink example to use seq_q and seq_kv parameters to accommodate the new sequence length parameters.

* Add GQA sink example for optimized attention mechanism & lint fix

* fix several typos and bugs

* lint

* fix speed issues of swa

* Add flash attention example with backward pass for BHSD layout and corresponding test cases

* Add backward pass implementation for flash attention with sinks and corresponding test case

* fix lint and typo

* Optimze the calculation of `dsinks`

* Add support for swa backward and update examples

* fix previous typos

* Add example for GQA sink backward pass and update tests for both MHA and GQA sinks

* fix lint

* fix previous typos

* typo