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>

* [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

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

* lint

* lint

* recommend using T.dynamic instead of T.symbolic

* lint fix

* lint fix

* [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>

* remove debug print

* Remove inline let expressions from the LowerAndLegalize function in phase.py

* add test

* Update sparse MLA examples to support SKV adjustment and correctness checks

- Changed SKV parameter from 32768 to 8192 in sparse MLA backward and forward tests.
- Added check_correctness parameter to test functions for validation of outputs.
- Updated test cases to reflect new SKV values and correctness checks.

* reduce test shape

* Update documentation structure and refactor main function parameters in example_fusedmoe_tilelang.py

- Added a new section for compiler internals in the documentation.
- Refactored the main function in example_fusedmoe_tilelang.py to accept parameters for hidden dimensions, expert configurations, and batch/sequence sizes, improving flexibility and readability.

* Update buffer access checks in merge_shared_memory_allocations.cc

- Changed the condition for buffer access from less than (<) to less than or equal to (<=) to allow access at the same scope level.
- Adjusted the logic for determining the access level when touching buffers to ensure correct handling of scope levels.

* lint fix

* Support pipeline with LetStmt

* lint fix

* • Fix LowerTileOp let handling to avoid LetInline dependency

  - inline let-bound BufferLoad nodes via resolver helpers and structured return
  - remap layouts/buffers using original data vars and only rewrite when needed
  - update pipeline planner to understand let-bound address_of buffers
  - document the new inline behaviour in docs/let_inline_fix.md

* fix for wgmma pipeline with let binding

* lint fix

* test fix

* reduce smem usage.

* let binding enhancement

* fix for dpgm

* fix simplify

* lint fix

* use tilelang.Simplify instead of tir.Simplify

* • Add TL_FORCE_LET_INLINE pass config and gate eager LetInline usage

  - register the new config in builtin headers/registration
  - add helper to pipeline enabling LetInline based on pass context
  - document LetStmt inlining controls and usage

* [Enhancement] Enable fast math optimization in tilelang JIT configurations

- Updated multiple examples and kernel functions to include `pass_configs` for enabling fast math optimization.
- Added support for the `TL_ENABLE_FAST_MATH` configuration option in the built-in operations.
- Enhanced the `LibraryGenerator` to handle the new fast math configuration, ensuring compatibility with existing settings.
- Updated documentation to reflect the changes in fast math handling and deprecation of the `TL_DISABLE_FAST_MATH` option.

* lint fix

* [Refactor] Introduce deprecated_warning utility for improved deprecation handling

- Added a new `deprecated_warning` function to streamline deprecation messages.
- Updated the `LibraryGenerator` to utilize the new function for warning about the deprecated `TL_DISABLE_FAST_MATH` configuration.
- Enhanced the `deprecated` decorator to support phaseout version messaging, improving clarity for users.

* Implement new free stage layout inference.

* Fix bug

* Make replication upcasting and unnormalizable iterators safe.

* Better handling of updating with more replica

* Remove unnecessary check.

* Fix compilation.

* Fix setup.py.

* Simplify development mode.

* Allow ParallelOp layout when there's already a compatible layout specified

* lint fix

* Add ProveFragmentContains function to validate thread access between small and large fragments

This function checks if the threads accessing elements of a smaller fragment are a subset of those accessing a larger fragment, ensuring valid access during updates. The implementation includes deriving thread indices, computing logical indices, and verifying thread mappings.

* Update dependencies in requirements files

* Remove 'thefuzz' from requirements-dev.txt
* Specify exact versions for 'torch' and add 'flash_attn' in requirements-test.txt

* Update CI workflow to use SHA256 hash for requirements file

* Update requirements and CI workflow for flash attention

* Removed specific version for 'torch' in requirements-test.txt
* Added installation of 'flash_attn==2.5.8' in CI workflow to ensure compatibility

* Refactor flash attention import handling in examples

* Removed availability checks for 'flash_attn' in multiple example scripts.
* Simplified import statements for 'flash_attn' to ensure consistent usage across examples.

---------
Co-authored-by: Huanqi Cao <caohuanqi@deepseek.com>

* Update flash-attn version in requirements-test.txt from <=2.2.0 to ==2.5.8

* lint fix

* Remove unused dependencies from requirements-test.txt

* Update import path for padding functions in example MHA forward variable length script

* Refactor code formatting in bert_padding.py for improved readability

* fix CI bugs in hopper

* lint fix

* Update bulk_copy.cc

* Refactor bulk copy logic in LowerBulkCopy function

- Removed unnecessary blank lines for improved code readability.
- Enhanced stride validation by checking for null pointers in global stride calculations, ensuring robustness against symbolic strides.
- Updated pass configuration handling in dynamic tile language tests to streamline dynamic alignment and TMA lower pass settings.

* test fix

* ci fix

* Update flash-attention dependencies and clean up example code

- Downgraded `flash-attn` dependency version in `requirements-test.txt` to `<=2.2.0`.
- Removed unused imports and commented-out code in various example files to enhance readability and maintainability.
- Updated the `flashattn` function signature to include default parameters for `block_M`, `block_N`, `num_stages`, and `threads`.
- Cleaned up the `example_mha_fwd_varlen.py` and `example_mha_bwd_wgmma_pipelined.py` files by removing unnecessary comments and improving code clarity.
- Deleted the `example_mha_inference.py` file as it is no longer needed.

* Update CI workflow to remove `--user` flag from pip install commands

- Removed the `--user` flag from the pip install commands in both the development and testing sections of the CI workflow to ensure proper installation of dependencies in the virtual environment.

* Update CI workflow to include `--no-user` flag in pip install commands

- Added the `--no-user` flag to the pip install commands in both the development and testing sections of the CI workflow to ensure dependencies are installed correctly within the virtual environment.

* Update CI workflow to include `--no-user` flag in pip install command for wheel mode

- Added the `--no-user` flag to the pip install command in the wheel mode section of the CI workflow to ensure dependencies are installed correctly within the virtual environment.

* test fix

* avoid conflict with system environments

* test fix

* add commnets

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

* Add paged block-sparse flash-decoding kernel

* Update example_tilelang_sparse_gqa_decode_paged.py

* lint fix

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

* [Example] Update kernel compilation in examples to use @tilelang.jit

- Refactored multiple examples to eliminate the use of `tilelang.compile` for kernel creation, directly invoking the functions instead.
- Added `@tilelang.jit` decorators with appropriate output indices to enhance performance and maintainability.
- Improved code clarity by simplifying the kernel invocation process across various examples, ensuring consistency in how kernels are defined and executed.

* format

* Update example_tilelang_sparse_gqa_decode_varlen_indice.py

* Update example_dequant_gemm_fine_grained.py

* Update example_gemm_autotune.py

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

* Remove debug print statement from block_sparse_attn_triton.py and implement a timeout handler in autotuner for function execution. This enhances the robustness of the autotuner by allowing it to handle timeouts gracefully.

* Enhance the autotuner module by adding a timeout handler for function execution, improving robustness in handling long-running tasks. This change includes the introduction of a custom TimeoutException and updates to the run_with_timeout function for better signal management.

* Add merge shared memory allocations pass and related configurations

- Introduced a new pass for merging shared memory allocations in GPU kernels, allowing for more efficient memory usage.
- Registered configuration options for debugging and controlling the merging behavior.
- Updated relevant files to integrate the new pass into the TileLang engine and transform modules.
- Adjusted import paths and added documentation for the new functionality.

* Reduce num_stages parameter in GEMM functions from 3 to 1 for improved performance in test_tilelang_kernel_gemm.py

* yes

* [Bugfix] fix the unexpected keyword error of autotune

* format

* test

* [CI] Add Analyzer and blocksparse_attention examples to CI

* format

* try

* try

* try

* try

* t

* format

* d

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

[Example] Handle Scenarios in Which a Threadblock is Assigned Only Invalid Block Indices for Sparse Attention  (#361)

* Fix issue where threadblock with only invalid blocks produces incorrect output.

* fix score scale

* format

* [Example] Add triton block sparse gqa decode

* lint fix

---------
Co-authored-by: LeiWang1999 <leiwang1999@outlook.com>

* add example gqa decode wgmma pipelined

* add sparse gqa

* support num split

* support num split

* add if condition

* add heuristic num split

* clean code

* add ref

* fix bug

* add torch ref

* fix bug

* integrate to torch

* symbolic

* clean mask

* rm actual_num_blocks

* clean code

* get num_sm via torch

* add sparse gqa decode example

* format

* rm example_gqa_decode_wgmma_pipelined.py

* Add license headers to example scripts

* format

* Remove commented-out cache disabling lines

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

* [Refactor] Improve flash attention example and layout comparison logic

- Removed unnecessary annotation for `lse_local_split` in the flash attention example to streamline the code.
- Updated the handling of `lse_local_split` to utilize parallel processing for better performance.
- Refactored kernel compilation and profiling logic to enhance clarity and maintainability in the flash attention example.
- Added a condition in `FragmentNode::IsEqual` to handle broadcast cases, improving the robustness of layout comparisons.

* lint fix

* [Enhancement] Add support for shared memory scope in Fill operation

- Introduced handling for `shared.dyn` and `shared` memory scopes in the Fill operation.
- Implemented parallel operation and layout inference for improved performance in shared memory scenarios.
- Updated thread loop partitioning and vectorization logic to accommodate new memory scope handling.

* [Refactor] Remove deprecated decorator and enhance Cython kernel handling

- Removed the deprecated decorator from the main module and added a new implementation in the utils module for better organization.
- Introduced a pointer map in the Cython kernel adapter to manage pointer arguments, improving runtime shape resolution.
- Updated the Cython kernel wrapper to utilize the new pointer map for handling kernel arguments.
- Enhanced error checking in the tensor utility functions to ensure static shapes are enforced.
- Added a new proxy module for buffer and tensor handling, streamlining the interface for TIR programs.

* [Feature] Add matrix multiplication test and kernel implementation

- Introduced a new test file `test_tilelang_language_ptr.py` that implements a matrix multiplication function using TileLang's primitives.
- The `matmul_test` function defines a kernel for performing tile-level GEMM operations with customizable block sizes and data types.
- Added a `run_matmul` function to compile and execute the kernel, along with a test function to validate the implementation.
- Updated the `proxy.py` file to enhance type handling for buffer and tensor proxies, ensuring compatibility with TIR programs.
- Minor formatting improvements in `deprecated.py` for better readability.

* lint fix

* [Refactor] Update tensor creation in matrix multiplication test

- Replaced `T.Tensor.from_ptr` with `T.make_tensor` in `matmul_test` for improved clarity and consistency.
- Updated imports in `__init__.py` to include `make_tensor`.
- Added `make_tensor` function in `proxy.py` to streamline tensor creation from pointers.

* [Refactor] Update tensor definitions across multiple files

- Replaced instances of `T.Tensor` with updated tensor definitions in various benchmark and example files to enhance consistency and clarity.
- Adjusted tensor shapes and types in functions related to matrix multiplication, attention mechanisms, and other operations.
- Improved documentation in README and example files to reflect changes in tensor usage.

* lint fix

* [Refactor] Update tensor types in attention and matrix multiplication examples

- Replaced instances of `T.Tensor` with `T.SharedTensor` and `T.FragmentTensor` in various attention and matrix multiplication functions to improve consistency and clarity.
- Adjusted tensor definitions in benchmark and example files to align with the new tensor types.
- Enhanced the overall structure and readability of the code by standardizing tensor usage across multiple files.

* lint fix

* [Refactor] Update tensor types in GEMM example and test files

- Replaced instances of `T.Tensor` with `T.LocalTensor` and `T.Buffer` in the GEMM example and related test functions to improve consistency and clarity.
- Enhanced the overall structure of the code by standardizing tensor usage across multiple files, aligning with recent updates in tensor definitions.

* [Refactor] Update tensor usage in customize.py

- Replaced instances of `T.Tensor` with `T.Buffer` in the `reshape` and `view` functions to enhance consistency with recent tensor definitions.
- Improved code clarity by standardizing buffer usage across the file.

* [Refactor] Update tensor types in test_tilelang_transform_annotate_device_regions.py

- Replaced instances of `T.Tensor` with `T.Buffer` in the `before` and `expected` methods of the `TestAnnotateThreadExtent` and `TestAnnotateDeviceScope` classes to enhance consistency with recent tensor definitions.
- Improved code clarity by standardizing buffer usage across the test file.

* [Refactor] Update tensor types to SharedBuffer and FragmentBuffer

- Replaced instances of `T.SharedTensor` and `T.FragmentTensor` with `T.SharedBuffer` and `T.FragmentBuffer` across multiple benchmark, example, and test files to enhance consistency with recent tensor definitions.
- Improved code clarity and structure by standardizing buffer usage in attention and matrix multiplication functions.

* [Refactor] Introduce Tensor alias for Buffer in proxy.py

- Added a new alias `Tensor` for `Buffer` in `proxy.py` to facilitate JIT compilation, ensuring that inputs and outputs are mapped with `torch.Tensor`.
- This change enhances clarity and consistency in tensor usage across the codebase.

* remove llvm build

* [Refactor] Update kernel compilation and profiling in examples

- Replaced `tilelang.lower` with `tilelang.compile` in multiple example scripts to streamline kernel compilation.
- Updated profiling calls to utilize the new `get_profiler` method, enhancing performance measurement consistency.
- Adjusted assertions and benchmarking methods to align with the new profiling structure across various examples, ensuring correctness and clarity in performance evaluations.

* lint fix

* License Update

* [Refactor] Improve code formatting and documentation in CUDA header and HIP runtime files

- Adjusted formatting in `cuda.h` for better readability, including alignment of comments and struct fields.
- Cleaned up whitespace and improved comment clarity in `rt_mod_hip.cc` to enhance code maintainability.

* [Refactor] Enhance formatting and clarity in CUDA header and HIP runtime files

- Improved comment alignment and readability in `cuda.h`.
- Cleaned up whitespace and formatting in `rt_mod_hip.cc` to enhance maintainability.

* lint fix

* lint fix

* lint fix

* lint fix

* fix

* License update

* [Enhancement] Update JITKernel to use artifact for kernel source

- Assigned the generated artifact to `self.artifact` for better management.
- Updated kernel source references to use `artifact.kernel_source` for consistency in execution backend handling.

* lint fix

* Add @tilelang.testing.requires_llvm decorator to vectorization tests

* Enhance setup.py and env.py for library management

- Added functionality to remove original files after copying in CMakeBuild.
- Updated TVM_LIBRARY_PATH in env.py to include the PyPI build library path for better integration.

* Refactor TVM_LIBRARY_PATH assignment for improved readability in env.py

* Refactor CMakeBuild file handling in setup.py

- Added a check to ensure the target library directory exists before copying .so files.
- Improved the logic for creating the target directory and copying files to enhance robustness.

* bugfix

* Rename BuildTLDebug to BuildTileLangCUDAWithoutCompile and update registration. Add @tilelang.testing.requires_llvm decorator to multiple tests for LLVM requirement.

* lint fix

* Enhance TileLang code generation by adding support for device code generation without compilation. Updated `host_codegen` and `device_codegen` functions to include new transformations and registration for `tilelang_hip_without_compile`. Refactored JIT kernel adapters to accommodate host and device modules, improving overall integration and flexibility.

* lint fix

* Add support for C target in device code generation

- Updated `device_codegen_without_compile` to include handling for the C target by registering the `tilelang_cpp` function.

* [Enhancement] Implement auto-clear cache feature based on environment variable

* Added TILELANG_CLEAR_CACHE environment variable to control cache clearing.
* Updated CI workflow to set TILELANG_CLEAR_CACHE during testing.
* Modified cache initialization to clear cache if TILELANG_CLEAR_CACHE is set to true.

* [Refactor] Update kernel invocation and import paths in tests and cache

* Changed kernel invocation in `test_tilelang_kernel_dequantize_gemm.py` to return the result.
* Updated import statements in `test_tilelang_kernel_int4_gemm_mma.py` to use `bitblas` instead of `tilelang`.
* Refactored paths for artifact and parameters in `kernel_cache.py` for better maintainability.

* [Refactor] Clean up whitespace and improve code formatting in kernel_cache.py

* Removed unnecessary blank lines and adjusted spacing for better readability in the KernelCache class.
* Enhanced overall code formatting to align with project standards.

* [Enhancement] Add bfloat16 test case and improve kernel caching logic

* Introduced a new test case for bfloat16 matrix multiplication in `test_tilelang_kernel_gemm_mma_intrinsic.py`.
* Updated `KernelCache` to handle multiple kernel source files and improve error handling during saving and loading.
* Refactored `JITKernel` to support instantiation from a database, enhancing flexibility in kernel management.
* Adjusted `CtypesKernelAdapter` and `CythonKernelAdapter` to utilize the new kernel loading mechanism from the database.
* Improved code formatting and readability across several files.

* lint fix

* Update bfloat16 matrix multiplication test case to use larger dimensions for improved coverage

* Optimize CMake build process with dynamic job count calculation

- Modify build_csrc function to use 90% of available CPU cores
- Ensure at least one job is used during compilation
- Improve build performance by dynamically adjusting parallel job count

* Optimize build_csrc function with multiprocessing module

- Replace os.cpu_count() with multiprocessing.cpu_count()
- Maintain existing 90% CPU utilization logic
- Improve CPU core count calculation for build process

* Add dynamic shape support with out_idx in Cython JIT kernel compilation

- Implement `run_cython_dynamic_shape_with_out_idx` function in test_tilelang_jit_gemm_cython.py
- Update Cython wrapper to handle dynamic symbolic shapes during tensor allocation
- Add support for resolving dynamic shape dimensions using input tensor references
- Enhance flexibility of JIT kernel compilation with symbolic shape handling

* Enhance error reporting for dynamic symbolic shape resolution in Cython JIT kernel

- Add detailed error message when a dynamic symbolic dimension is not found in dynamic_symbolic_map
- Improve debugging by providing context about missing symbolic dimensions
- Maintain existing dynamic shape resolution logic

* Fix Copy operation handling for scalar and multi-dimensional tensors

- Add special handling for scalar tensor copy operations
- Enhance error reporting in MakeIndices method with more detailed diagnostic information
- Improve SIMT loop generation to support zero-dimensional tensors
- Add explicit check and handling for scalar tensor scenarios

* Refactor Copy operation code formatting and improve readability

- Improve code formatting in MakeIndices and MakeSIMTLoop methods
- Add line breaks to enhance readability of complex ICHECK statements
- Simplify code structure in scalar tensor handling
- Remove unnecessary whitespace and improve code alignment

* Simplify GEMM example with direct kernel compilation

- Update copyright header to Tile-AI Corporation
- Remove Profiler import and usage
- Replace tilelang.lower() with tilelang.compile()
- Simplify kernel execution workflow
- Update kernel source retrieval method

* Enhance block sparse attention implementation

- Update `blocksparse_flashattn` to use 2 stages for improved performance.
- Change `block_mask_dtype` from `int8` to `bool` for better memory efficiency.
- Modify condition checks in the kernel to utilize boolean values.
- Introduce a new example for top-k sparse attention and a benchmark for native sparse attention.
- Add support for asynchronous copy in PTX and improve pipeline planning with condition handling.

* Refactor and clean up code formatting across multiple files

- Added whitespace for improved readability in `example_blocksparse_gemm.py`, `example_tilelang_nsa_fwd.py`, and `benchmark_nsa_fwd.py`.
- Enhanced code structure and alignment in `inject_ptx_async_copy.cc` and `pipeline_planning.cc`.
- Updated comments and documentation for clarity in `__init__.py` and `phase.py`.
- Ensured consistent formatting and style across the codebase.

* Remove Torch CPP backend and update execution backend options

- Remove TorchCPPKernelAdapter and related code from JIT modules
- Update execution backend options in jit/__init__.py, kernel.py, and adapter/__init__.py
- Remove "torch_cpp" from supported execution backend literals
- Simplify backend validation and remove unused torch_cpp-related code
。

* lint fix

* Add block sparse attention implementations for TileLang and Triton

- Implement block sparse attention kernels for TileLang and Triton
- Add example scripts for block sparse attention with top-k and threshold-based masking
- Include utility functions for generating sparse attention masks
- Demonstrate causal attention with block-level sparsity
- Add test cases to validate sparse attention implementations against PyTorch reference

* Bump version to 0.1.1

* Refactor block sparse attention examples for improved code quality

- Apply consistent code formatting and style in TileLang and Triton block sparse attention implementations
- Add ruff linter ignore comment for specific line in Triton implementation
- Improve readability by adjusting indentation and line breaks
- Standardize sparse mask generation and test function implementations
- Minor optimizations in test case configurations

* lint

* Remove Torch CPP backend and update execution backend options

- Remove TorchCPPKernelAdapter and related code from JIT modules
- Update execution backend options in jit/__init__.py, kernel.py, and adapter/__init__.py
- Remove "torch_cpp" from supported execution backend literals
- Simplify backend validation and remove unused torch_cpp-related code
。

* lint fix

* Add block sparse attention implementations for TileLang and Triton

- Implement block sparse attention kernels for TileLang and Triton
- Add example scripts for block sparse attention with top-k and threshold-based masking
- Include utility functions for generating sparse attention masks
- Demonstrate causal attention with block-level sparsity
- Add test cases to validate sparse attention implementations against PyTorch reference