- 25 Feb, 2025 2 commits
-
-
Lei Wang authored
* Add DeepSeek MLA decode example with Flash Attention implementation * Add GEMM SplitK and StreamK example implementations This commit introduces two new example scripts demonstrating advanced GEMM (matrix multiplication) techniques: - `example_tilelang_gemm_splitk.py`: Implements a Split-K GEMM kernel using TileLang - `example_tilelang_gemm_streamk.py`: Implements a Stream-K GEMM kernel using TileLang Both examples showcase different parallel computation strategies for matrix multiplication, with comprehensive testing using PyTorch reference implementations. * Refactor GEMM SplitK and StreamK example implementations Clean up and improve code formatting for the SplitK and StreamK GEMM example scripts: - Remove unused import (Profiler) in splitk example - Simplify line breaks and improve code readability - Standardize indentation and remove unnecessary whitespace - Optimize atomic add and copy operations for better clarity * Add block sparse attention benchmarks for multiple libraries This commit introduces comprehensive block sparse attention benchmarks for different libraries: - TileLang block sparse FMHA implementation - Triton block sparse FMHA implementation - PyTorch reference block sparse FMHA implementation - FlashAttention dense FMHA reference implementation The benchmarks include: - Configurable benchmark parameters (batch size, heads, sequence length, etc.) - Sparse mask generation using top-k and threshold methods - Performance measurement for different sparse attention configurations - Utility functions for mask generation and benchmarking * Refactor block sparse attention benchmarks with code style improvements - Add Ruff linter ignore comments to benchmark files - Improve code formatting and line breaks - Remove unused imports - Standardize print statement formatting - Enhance code readability across multiple library benchmarks * lint fix * Add CUDA atomic operations for BFLOAT16 and update function naming - Implement AtomicAdd functions for BFLOAT16 and BFLOAT16x2 in CUDA common header - Rename existing atomic add functions to use PascalCase (atomicAdd -> AtomicAdd) - Add a new __pack_nv_bfloat162 function for packing BFLOAT16 values - Update kernel and language customization to use new function names - Add return type annotations in profiler module * lint fix * Add example for Group Query Attention (GQA) forward pass using Flash Attention in TileLang This commit introduces a new example script `example_gqa_fwd_bshd.py` that demonstrates: - Group Query Attention (GQA) implementation - Flash Attention forward pass - Performance benchmarking - Configurable parameters for batch, heads, sequence length, and dimension - Autotuning support - Reference implementation comparison * Refactor IR lowering pipeline into modular phases This commit introduces a new module `phase.py` to modularize the IR lowering process by splitting the complex lowering pipeline into two distinct phases: - `LowerAndLegalize`: Handles initial IR legalization and transformation - `OptimizeForTarget`: Applies target-specific optimizations The changes simplify the lowering logic in multiple files by extracting the transformation steps into reusable functions, improving code readability and maintainability. * lintfix
-
Yu Cheng authored
-
- 19 Feb, 2025 2 commits
-
-
Lei Wang authored
* bump version into v0.1.0 * [Enhancement] Add custom develop command for editable installs and update .gitignore * [Documentation] Update README to include system dependencies installation instructions * [Build] Update setup.py to support library file copying for both release and develop modes * [Build] Refactor library file copying logic in setup.py * [Documentation] Remove unnecessary install section header in Installation.md * [Build] Add tox configuration and local distribution script for multi-Python version support * [Build] Improve git submodule update function with better error handling * [Build] Update LLVM configuration path in ROCm installation script * [Build] Add .tox/ to .gitignore for tox testing environment * [Build] Add support for TVM prebuild path configuration in CMakeLists.txt * [Cleanup] Remove unused TVM runtime error codes header * [Cleanup] Fix TVM grid constant type reference in CUDA module * [Cleanup] Remove unused customized_code function from IR module * [Feature] Add TileLang thread synchronization and storage access analysis passes * [Build] Reorder DLL search path directories for more flexible library loading * [Refactor] Improve thread synchronization and library path handling - Rename ThreadSync and TileLangThreadSync functions in C++ code - Update Python docstring for ThreadSync with more detailed description - Reorder library path detection in tilelang environment setup - Minor comment and code cleanup in CUDA and warp specialization modules * [Refactor] Improve thread synchronization code style and formatting - Standardize pointer type spacing in storage_access.h and storage_access.cc - Update whitespace and indentation in thread_storage_sync.cc - Reorder include statements in thread_partial_sync.cc - Minor code formatting improvements across thread synchronization files * [Refactor] Fix global function registration for ThreadSync - Correct global function registration to use ThreadSync instead of TileLangThreadSync - Update TVM global registration to match recent refactoring efforts * [Refactor] Simplify ThreadSync global function registration - Remove unnecessary whitespace in global function registration - Compact the TVM global registration line for ThreadSync * [Feature] Add WebGPU code generation support in TileLang - Implement WebGPU code generator (codegen_webgpu.cc and codegen_webgpu.h) - Add WebGPU target support in lower.py and target.py - Update CMakeLists.txt to include WebGPU codegen source files - Introduce WebGPU-specific code generation for WGSL shader language * [Refactor] Improve WebGPU code generation formatting and readability - Enhance code formatting in codegen_webgpu.cc and codegen_webgpu.h - Standardize pointer type spacing and indentation - Improve line breaks and reduce line length for better readability - Minor code style improvements in WebGPU code generation * [Test] Add WebGPU matrix multiplication code generation test - Implement test_webgpu_codegen.py for WebGPU matrix multiplication - Add assert_gemm_codegen function to validate WebGPU code generation - Include basic matrix multiplication kernel test case * Update README with WebGPU codegen support announcement * Support multi version pypi package build via tox * Add support for CPU device backend with C code generation - Introduce `is_cpu_device_backend` function to detect CPU backend with C code generation - Modify `lower` function to handle special case of CPU device backend - Update host and device call filtering for CPU backend - Add conditional source code generation for C host target - Extend JITKernel to support optional target_host parameter * lint fix * Enhance JIT kernel adapters with CTypes and Torch C++ backends - Add CtypesKernelAdapter with dynamic library generation and kernel wrapping - Implement TorchCPPKernelAdapter for CUDA kernel compilation - Refactor BaseKernelAdapter to support more flexible initialization - Improve error handling and argument processing in kernel adapters - Update adapter initialization to support various execution backends * Refactor and clean up code style in JIT CTypes adapter modules - Apply consistent code formatting and whitespace in CTypes adapter files - Remove unused imports and improve import organization - Enhance readability of code in adapter, libgen, and wrapper modules - Add missing whitespace and improve line breaks - Minor linting and code style improvements across CTypes adapter files * Add test for TileLang JIT GEMM with CTypes backend - Implement comprehensive test for matrix multiplication using CTypes execution backend - Create test functions for GEMM with float16 data type - Add kernel source verification with custom callback - Implement reference implementation using PyTorch for result validation - Support various matrix multiplication configurations (transposition, block sizes) * test fix * Update TileLang JIT callback registration with override parameter - Modify tilelang_callback_cuda_postproc to use @tvm.register_func(override=True) - Ensure proper function registration with ability to replace existing implementations * Reorder TileLang lowering passes for Hopper intrinsics and PTX async copy - Adjust the order of LowerHopperIntrin and InjectPTXAsyncCopy passes - Move these passes to ensure correct synchronization and device preparation * Rebase main * shared.dyn * lint fix * test fix * Add environment variable handling for TileLang template and CUTLASS paths - Introduce fallback logic for TL_TEMPLATE_PATH environment variable - Add support for optional TL_CUTLASS_PATH configuration - Include TODO comment for future environment variable renaming
-
Lei Wang authored
* bump version into v0.1.0 * [Enhancement] Add custom develop command for editable installs and update .gitignore * [Documentation] Update README to include system dependencies installation instructions * [Build] Update setup.py to support library file copying for both release and develop modes * [Build] Refactor library file copying logic in setup.py * [Documentation] Remove unnecessary install section header in Installation.md * [Build] Add tox configuration and local distribution script for multi-Python version support * [Build] Improve git submodule update function with better error handling * [Build] Update LLVM configuration path in ROCm installation script * [Build] Add .tox/ to .gitignore for tox testing environment * [Build] Add support for TVM prebuild path configuration in CMakeLists.txt * [Cleanup] Remove unused TVM runtime error codes header * [Cleanup] Fix TVM grid constant type reference in CUDA module * [Cleanup] Remove unused customized_code function from IR module * [Feature] Add TileLang thread synchronization and storage access analysis passes * [Build] Reorder DLL search path directories for more flexible library loading * [Refactor] Improve thread synchronization and library path handling - Rename ThreadSync and TileLangThreadSync functions in C++ code - Update Python docstring for ThreadSync with more detailed description - Reorder library path detection in tilelang environment setup - Minor comment and code cleanup in CUDA and warp specialization modules * [Refactor] Improve thread synchronization code style and formatting - Standardize pointer type spacing in storage_access.h and storage_access.cc - Update whitespace and indentation in thread_storage_sync.cc - Reorder include statements in thread_partial_sync.cc - Minor code formatting improvements across thread synchronization files * [Refactor] Fix global function registration for ThreadSync - Correct global function registration to use ThreadSync instead of TileLangThreadSync - Update TVM global registration to match recent refactoring efforts * [Refactor] Simplify ThreadSync global function registration - Remove unnecessary whitespace in global function registration - Compact the TVM global registration line for ThreadSync * [Feature] Add WebGPU code generation support in TileLang - Implement WebGPU code generator (codegen_webgpu.cc and codegen_webgpu.h) - Add WebGPU target support in lower.py and target.py - Update CMakeLists.txt to include WebGPU codegen source files - Introduce WebGPU-specific code generation for WGSL shader language * [Refactor] Improve WebGPU code generation formatting and readability - Enhance code formatting in codegen_webgpu.cc and codegen_webgpu.h - Standardize pointer type spacing and indentation - Improve line breaks and reduce line length for better readability - Minor code style improvements in WebGPU code generation * [Test] Add WebGPU matrix multiplication code generation test - Implement test_webgpu_codegen.py for WebGPU matrix multiplication - Add assert_gemm_codegen function to validate WebGPU code generation - Include basic matrix multiplication kernel test case * Update README with WebGPU codegen support announcement * Support multi version pypi package build via tox * Add support for CPU device backend with C code generation - Introduce `is_cpu_device_backend` function to detect CPU backend with C code generation - Modify `lower` function to handle special case of CPU device backend - Update host and device call filtering for CPU backend - Add conditional source code generation for C host target - Extend JITKernel to support optional target_host parameter * lint fix * Enhance JIT kernel adapters with CTypes and Torch C++ backends - Add CtypesKernelAdapter with dynamic library generation and kernel wrapping - Implement TorchCPPKernelAdapter for CUDA kernel compilation - Refactor BaseKernelAdapter to support more flexible initialization - Improve error handling and argument processing in kernel adapters - Update adapter initialization to support various execution backends * Refactor and clean up code style in JIT CTypes adapter modules - Apply consistent code formatting and whitespace in CTypes adapter files - Remove unused imports and improve import organization - Enhance readability of code in adapter, libgen, and wrapper modules - Add missing whitespace and improve line breaks - Minor linting and code style improvements across CTypes adapter files * Add test for TileLang JIT GEMM with CTypes backend - Implement comprehensive test for matrix multiplication using CTypes execution backend - Create test functions for GEMM with float16 data type - Add kernel source verification with custom callback - Implement reference implementation using PyTorch for result validation - Support various matrix multiplication configurations (transposition, block sizes) * test fix * Update TileLang JIT callback registration with override parameter - Modify tilelang_callback_cuda_postproc to use @tvm.register_func(override=True) - Ensure proper function registration with ability to replace existing implementations
-
- 15 Feb, 2025 1 commit
-
-
Lei Wang authored
* bump version into v0.1.0 * [Enhancement] Add custom develop command for editable installs and update .gitignore * [Documentation] Update README to include system dependencies installation instructions * [Build] Update setup.py to support library file copying for both release and develop modes * [Build] Refactor library file copying logic in setup.py * [Documentation] Remove unnecessary install section header in Installation.md * [Build] Add tox configuration and local distribution script for multi-Python version support * [Build] Improve git submodule update function with better error handling * [Build] Update LLVM configuration path in ROCm installation script * [Build] Add .tox/ to .gitignore for tox testing environment * [Build] Add support for TVM prebuild path configuration in CMakeLists.txt * [Cleanup] Remove unused TVM runtime error codes header * [Cleanup] Fix TVM grid constant type reference in CUDA module * [Cleanup] Remove unused customized_code function from IR module * [Feature] Add TileLang thread synchronization and storage access analysis passes * [Build] Reorder DLL search path directories for more flexible library loading * [Refactor] Improve thread synchronization and library path handling - Rename ThreadSync and TileLangThreadSync functions in C++ code - Update Python docstring for ThreadSync with more detailed description - Reorder library path detection in tilelang environment setup - Minor comment and code cleanup in CUDA and warp specialization modules * [Refactor] Improve thread synchronization code style and formatting - Standardize pointer type spacing in storage_access.h and storage_access.cc - Update whitespace and indentation in thread_storage_sync.cc - Reorder include statements in thread_partial_sync.cc - Minor code formatting improvements across thread synchronization files * [Refactor] Fix global function registration for ThreadSync - Correct global function registration to use ThreadSync instead of TileLangThreadSync - Update TVM global registration to match recent refactoring efforts * [Refactor] Simplify ThreadSync global function registration - Remove unnecessary whitespace in global function registration - Compact the TVM global registration line for ThreadSync * [Feature] Add WebGPU code generation support in TileLang - Implement WebGPU code generator (codegen_webgpu.cc and codegen_webgpu.h) - Add WebGPU target support in lower.py and target.py - Update CMakeLists.txt to include WebGPU codegen source files - Introduce WebGPU-specific code generation for WGSL shader language * [Refactor] Improve WebGPU code generation formatting and readability - Enhance code formatting in codegen_webgpu.cc and codegen_webgpu.h - Standardize pointer type spacing and indentation - Improve line breaks and reduce line length for better readability - Minor code style improvements in WebGPU code generation * [Test] Add WebGPU matrix multiplication code generation test - Implement test_webgpu_codegen.py for WebGPU matrix multiplication - Add assert_gemm_codegen function to validate WebGPU code generation - Include basic matrix multiplication kernel test case * Update README with WebGPU codegen support announcement
-
- 14 Feb, 2025 1 commit
-
-
Lei Wang authored
* bump version into v0.1.0 * [Enhancement] Add custom develop command for editable installs and update .gitignore * [Documentation] Update README to include system dependencies installation instructions * [Build] Update setup.py to support library file copying for both release and develop modes * [Build] Refactor library file copying logic in setup.py * [Documentation] Remove unnecessary install section header in Installation.md * [Build] Add tox configuration and local distribution script for multi-Python version support * [Build] Improve git submodule update function with better error handling * [Build] Update LLVM configuration path in ROCm installation script * [Build] Add .tox/ to .gitignore for tox testing environment * [Build] Add support for TVM prebuild path configuration in CMakeLists.txt * [Cleanup] Remove unused TVM runtime error codes header * [Cleanup] Fix TVM grid constant type reference in CUDA module * [Cleanup] Remove unused customized_code function from IR module * [Feature] Add TileLang thread synchronization and storage access analysis passes * [Build] Reorder DLL search path directories for more flexible library loading * [Refactor] Improve thread synchronization and library path handling - Rename ThreadSync and TileLangThreadSync functions in C++ code - Update Python docstring for ThreadSync with more detailed description - Reorder library path detection in tilelang environment setup - Minor comment and code cleanup in CUDA and warp specialization modules * [Refactor] Improve thread synchronization code style and formatting - Standardize pointer type spacing in storage_access.h and storage_access.cc - Update whitespace and indentation in thread_storage_sync.cc - Reorder include statements in thread_partial_sync.cc - Minor code formatting improvements across thread synchronization files * [Refactor] Fix global function registration for ThreadSync - Correct global function registration to use ThreadSync instead of TileLangThreadSync - Update TVM global registration to match recent refactoring efforts * [Refactor] Simplify ThreadSync global function registration - Remove unnecessary whitespace in global function registration - Compact the TVM global registration line for ThreadSync
-
- 03 Feb, 2025 1 commit
-
-
Lei Wang authored
* [Enhancement] Add VectorizeLoop function and update imports for compatibility * [CI][Test] Improve test cases for vectorization and fix typos in parser comments * lint fix * Fix incorrect module reference for VectorizeLoop transformation * Refactor vectorize_loop transformation by removing unused extent mutation logic
-
- 26 Jan, 2025 1 commit
-
-
Lei Wang authored
* implement jit test case * [Dev] implement auto tune test case for matrix multiplication * Implement test for legalize memory access and vectorized loop * lint fix * introduce run_once * Refactor callback function names for consistency and improve code readability * enhance documentations * lint fix * lint fix * lint fix * lint fix * fix formatting issues in rt_mod_hip.cc * add random seed initialization for deterministic testing
-
- 23 Jan, 2025 1 commit
-
-
Lei Wang authored
[Bugfix] Reorder Passes: Place Vectorize Loop Before StorageFlatten and FlattenBuffer to Prevent Redundant Allocations (#37) * installation script fix * readme typo fix * doc fix for dequantize gemm * [Doc] remove CODE_OF_CONDUCT.md and SECURITY.md; update references in CONTRIBUTING.md * [Doc] add unit tests for AnnotateDeviceRegions transform; remove SUPPORT.md * update license * [Enhancement] add tensor supply handling for unsigned integers; improve error message for execution backend assertion * [Refactor] improve code readability by reformatting function signatures and assertions * [Refactor] replace torch.manual_seed with tilelang.testing.set_random_seed for consistency in random seed handling * [Refactor] unify thread binding variable naming across kernel and example files * [Refactor] remove unused thread binding parameter from matrix multiplication functions * [Refactor] remove unused thread binding parameter from matrix multiplication functions * [Refactor] enable main testing function in tilelang kernel gemm test * bug fix * lint fix * [Refactor] reorder vectorize loop
-
- 20 Jan, 2025 1 commit
-
-
Lei Wang authored
* instruction update * replace link with TileLang/tile-lang * [Dev][Adapter] Implement Torch DLPack Kernel Adapter and related utilities * lint fix * Implement JIT Compiler Components * Documents update * lint fix * update logo * install script fix
-
- 17 Jan, 2025 1 commit
-
-
Lei Wang authored
* README.md fixed * test fix * cpu backend update * cpu test case
-
- 11 Jan, 2025 1 commit
-
-
Lei Wang authored
* Add format.sh script for code formatting and linting * docs update * center align the title * lint fix * add ignore * Add .gitignore for 3rdparty directory * Add requirements-dev.txt, requirements-test.txt, and requirements.txt * 3rdparty * Add gemm.h, CMakeLists.txt, _ffi_api.py, __init__.py, runtime.h, reduce.h, loop_partition.h, utils.h, and loop_vectorize.h * Refactor CMakeLists.txt and include statements - Update CMakeLists.txt to use a newer version of CMake and add project name - Remove unnecessary include directories Fix include paths in layout.cc, codegen.cc, codegen.h, rt_mod.cc, frontend_legalize.cc, inject_pipeline.cc, layout_inference.cc, loop_vectorize.cc, and lower_tile_op.cc - Update include paths to use relative paths instead of absolute paths * Update submodule for 3rdparty/tvm * update * load dll first * Refactor CMakeLists.txt and include statements * Refactor CMakeLists.txt and include statements * git keep update * Refactor CMakeLists.txt and include statements * Refactor CMakeLists.txt and include statements * refactor code structure * Update Readme * CMakeLists Customized * update readme * update README * update readme * update usage * with TVM_IMPORT_PYTHON_PATH to handle own tvm build python import * annotate lower transform global func with `transform` prefix * Migrate Simplify Pass from tilelang tvm branch * enhance system environment handling with __init__ and CMake * Initial commit * CODE_OF_CONDUCT.md committed * LICENSE committed * README.md committed * SECURITY.md committed * SUPPORT.md committed * CODE_OF_CONDUCT Commit * LICENSE Commit * SECURITY Commit * SUPPORT Commit * Modify Support * Update README.md * security ci update * remove examples * Update and implement clang-format * add composable kernel components * Migrate from latest update * submodule update * Test update * Update License * Spell check * lint fix * add clang-tidy to apply static analysis for c source * update tilelang examples * Update Install Docs * Refactor filetree * Enhance Install * conflict resloved * annotate_version * Initial Update * test fix * install * Implement setup.py * lint fix * Separate Init * Separate test * docker file commit * add logo * Update Readme and Examples * update readme * update logo * Implement AMD Installation * Add License * Update AMD MI300x Benchmark * update README * update mi300 benchmark scripts * update ignore * enhance build scirpt * update image * enhance setup.py to remove duplicated libraries * remove debug files * update readme * update image * update gemm examples * update flashattention README * readme update * add cmake into requirements * libinfo fix * auto update submodule * lint fix * Fix AMD Build and Test * Update check for transpose attribute for CDNA Arch * typo fix for amd * Implement Matmul Benchmark * Refactor Code * [TypoFix] Fix GEMM Example * [Docs] Init Linear Attention README * [TYPO] Typo fix * [Lint] Lint Fix * enhance example with intrinsics * [Enhancement] Improve Buffer Collection during IR Parser * [Dev] Introduce Current classmethod to get current frame * submodule update * fake test pass update * support thread_extent_api * code optimize * Add GEMM function implementation for matrix multiplication * Update logging format to reflect TileLang in logger messages * Refactor CMakeLists.txt for improved readability and set default build type to Release * Support Gemm SS Primitives Implementation * [README] Upload Tile Language Logo (#5) * update logo * Update README.md to enhance formatting and center the title --------- Co-authored-by:
microsoft-github-operations[bot] <55726097+microsoft-github-operations[bot]@users.noreply.github.com> Co-authored-by:
Microsoft Open Source <microsoftopensource@users.noreply.github.com> Co-authored-by:
Yu Cheng <yu.cheng@pku.edu.cn>
-
