- 16 Mar, 2025 1 commit
-
-
Yu Cheng authored
- Replaced instances of `tilelang.lower` and `tilelang.Profiler` with `tilelang.compile` and the new profiler interface in multiple example files. - Enhanced the kernel compilation process to utilize the updated API, improving consistency and maintainability. - Adjusted benchmarking logic to use the new profiler methods for better clarity and functionality in performance testing. - Cleaned up whitespace and improved formatting for better readability across the modified files.
-
- 14 Mar, 2025 2 commits
-
-
Lei Wang authored
* 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. * Add kernel source printing in example_tilelang_nsa_fwd.py and implement IfThenElse node replacement in inject_pipeline.cc - Added a print statement to output the kernel source in `example_tilelang_nsa_fwd.py` for debugging purposes. - Introduced a new function `replace_if_then_else` in `inject_pipeline.cc` to transform IfThenElse nodes while preserving attributes, enhancing the handling of conditional statements in the pipeline. * Refactor condition handling in inject_pipeline.cc - Change the data structure for mapping conditions to statements from a Map to an Array for improved performance and simplicity. - Update condition comparison logic to use StructuralEqual for better accuracy. - Enhance logging to provide detailed insights into condition changes and statement processing. - Adjust final statement construction to utilize the new data structure, ensuring correct handling of conditions and statements. * Improve logging and formatting in inject_pipeline.cc - Enhance logging statements for better clarity on condition changes and statement processing. - Adjust formatting for improved readability, including line breaks and consistent spacing. - Ensure accurate condition comparison and handling in the pipeline logic. * Refactor logging and clean up inject_pipeline.cc - Remove excessive logging statements to streamline the code and improve performance. - Simplify condition handling by eliminating unnecessary log outputs related to condition changes and statement processing. - Maintain the core functionality while enhancing code readability and maintainability. * Update Dockerfiles to specify exact version of libstdcxx-ng - Change installation command in multiple Dockerfiles to use `libstdcxx-ng=12` instead of `libstdcxx-ng-12` for consistency and to avoid potential issues with package resolution. - Ensure all Dockerfiles from cu118 to cu126 reflect this change for uniformity across builds. * Refactor and enhance examples and kernel handling - Adjusted the pipeline stages in `example_blocksparse_gemm.py` from 2 to 1 for improved performance. - Added kernel source printing in `benchmark_nsa_fwd.py` for better debugging and profiling insights. - Updated tensor allocation and parameter handling in `CtypesKernelAdapter` and `CythonKernelWrapper` to cache parameter dtypes and shapes, improving efficiency and clarity. - Enhanced the handling of dynamic shapes in the Cython JIT kernel compilation process. - Modified the benchmark script to accommodate new tensor output parameters and improved batch size defaults for testing. * Update copyright header in Cython wrapper to reflect Tile-AI Corporation * revert change
-
Chenghua authored
* [Example] Modify tuning configurations for FlashAttention example * [Examples] formatting example_gqa_fwd_bshd.py
-
- 13 Mar, 2025 4 commits
-
-
Yu Cheng authored
- Introduce `example_gqa_bwd.py` demonstrating the backward pass of FlashAttention with pipelined execution. - Implement forward and backward functions for FlashAttention, including preprocessing and postprocessing steps. - Enhance argument parsing for batch size, heads, context size, and dimensions. - Include a reference implementation for validation and performance benchmarking.
-
zqh-wz authored
* upgrade cutlass to upstream v3.8.0 * Implement fp8 gemm and add example script * Fix dtype retrieval with map_torch_type for fp8 inputs * Disable vectorization of fp8 values * Make MMA declaration compatible with cutlass 3.4.0+ * Add test for fp8 T.gemm * fix indent * fix indent * Add copyright and license header * Add copyright and license header * lint fix * Refactor matmul_nt and assert_matmul_correctness functions for improved readability by consolidating parameter definitions and adjusting formatting. * clang format lint --------- Co-authored-by:
Lei Wang <34334180+LeiWang1999@users.noreply.github.com> Co-authored-by:
LeiWang1999 <leiwang1999@outlook.com>
-
Lei Wang authored
* 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. * Add kernel source printing in example_tilelang_nsa_fwd.py and implement IfThenElse node replacement in inject_pipeline.cc - Added a print statement to output the kernel source in `example_tilelang_nsa_fwd.py` for debugging purposes. - Introduced a new function `replace_if_then_else` in `inject_pipeline.cc` to transform IfThenElse nodes while preserving attributes, enhancing the handling of conditional statements in the pipeline. * Refactor condition handling in inject_pipeline.cc - Change the data structure for mapping conditions to statements from a Map to an Array for improved performance and simplicity. - Update condition comparison logic to use StructuralEqual for better accuracy. - Enhance logging to provide detailed insights into condition changes and statement processing. - Adjust final statement construction to utilize the new data structure, ensuring correct handling of conditions and statements. * Improve logging and formatting in inject_pipeline.cc - Enhance logging statements for better clarity on condition changes and statement processing. - Adjust formatting for improved readability, including line breaks and consistent spacing. - Ensure accurate condition comparison and handling in the pipeline logic. * Refactor logging and clean up inject_pipeline.cc - Remove excessive logging statements to streamline the code and improve performance. - Simplify condition handling by eliminating unnecessary log outputs related to condition changes and statement processing. - Maintain the core functionality while enhancing code readability and maintainability.
-
Yu Cheng authored
- Introduce `example_gqa_fwd_bshd_wgmma_pipelined.py` demonstrating a pipelined implementation of FlashAttention. - Update sequence length parameter in existing example to 8192 and adjust number of stages for improved performance. - Enhance argument parsing to accommodate new configurations for batch size, heads, and groups.
-
- 12 Mar, 2025 3 commits
-
-
Lei Wang authored
* 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.
-
66RING authored
Expired example code, update readme.
-
Lei Wang authored
* 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
-
- 11 Mar, 2025 1 commit
-
-
Yu Cheng authored
* [Dev][Bugfix] Add RMS Normalization Kernels and Fix Reduce Bug - Implement two RMS normalization implementations in TileLang: * `rms_norm_splitk`: Split-K reduction approach for large matrices * `rms_norm`: Full reduction kernel with simplified implementation - Add reference implementation using PyTorch for validation - Include performance benchmarking for both kernel variants - Demonstrate flexible block size and matrix size configurations * [Examples] Simplify RMS Normalization Kernel Compilation - Remove commented-out code for split-K RMS normalization - Simplify kernel compilation by removing explicit TMA lowering configuration - Update copyright header to Tile-AI Corporation - Streamline main script for RMS normalization example
-
- 10 Mar, 2025 1 commit
-
-
Lei Wang authored
* Update native sparse attention example with scale parameter handling - Add scale parameter processing in native_sparse_attention function - Modify example script to include custom scale value - Update function calls to pass scale parameter - Enhance flexibility of sparse attention implementation * Refactor Triton Native Sparse Attention Example - Improve code formatting and readability in example_triton_nsa_bwd.py - Standardize function and parameter alignment - Remove unnecessary whitespaces and optimize imports - Enhance code style consistency with previous commits
-
- 09 Mar, 2025 1 commit
-
-
Lei Wang authored
* Add kernel caching mechanism to TileLang - Implement a new `cached` function in `tilelang/cache/__init__.py` to cache and reuse compiled kernels - Expose the `cached` function in the main `tilelang/__init__.py` - Add a test case for cached matrix multiplication in `testing/python/cache/test_tilelang_cache_matmul.py` - Provide a `clear_cache()` function to reset the kernel cache when needed * Refactor kernel caching test and implementation - Simplify the `cached` function in `tilelang/cache/__init__.py` - Update test script `test_tilelang_cache_matmul.py` to use `tilelang.testing.main()` - Remove unnecessary whitespace and improve code formatting * Update import for `cached` function in MHA examples - Modify import statement in `example_mha_bwd.py` and `test_tilelang_kernel_mha_bwd.py` - Change import from `tilelang.profiler import cached` to `tilelang import cached` - Align with recent refactoring of kernel caching mechanism * Refactor `cached` function signature in kernel caching - Update function signature to use keyword-only arguments for `target` and `target_host` - Improve parameter order and readability of the `cached` decorator - Maintain existing functionality while enhancing function definition
-
- 07 Mar, 2025 5 commits
-
-
Lei Wang authored
* [Refactor] Update BitBLAS Benchmark with TileLang Carver Imports and Roller Hints Generation - Replace BitBLAS imports with TileLang Carver imports in benchmark_matmul.py - Modify roller hints generation using new TileLang Carver template and utility functions - Update get_roller_hints_from_func to handle None cases and improve return logic - Adjust DefaultPolicy to handle different codegen dictionary formats * [Refactor] Update Thread Binding and Import Statements in TileLang Kernels - Replace T.thread_binding() with T.get_thread_binding() across multiple kernel test files - Update import statements for MMA layout and macro generator in dequantize GEMM and FP8 examples - Move map_torch_type utility function to tilelang.utils.tensor - Remove unnecessary imports and improve code organization * Refactor Native Sparse Attention Example with Enhanced Triton Kernel - Update parallel_nsa_fwd_kernel to support more flexible sparse attention computation - Add support for block counts and offsets in the Triton kernel - Modify kernel grid and computation logic for improved performance - Update example script to use naive_nsa_simple reference implementation - Improve type hints and kernel configuration * Add Native Sparse Attention Examples with Tilelang and Triton Implementations - Introduce new example scripts for native sparse attention: * example_tilelang_nsa_fwd.py: Forward pass implementation using TileLang * example_tilelang_nsa_decode.py: Decoding-specific sparse attention implementation * example_triton_nsa_fwd.py: Triton-based sparse attention forward pass - Update reference.py with naive implementations for sparse attention - Support different sparse attention scenarios including forward pass and inference - Add comprehensive testing and validation against reference implementations * lint fix * Add Variable-Length Native Sparse Attention Examples for TileLang and Triton - Introduce new example scripts for variable-length native sparse attention: * example_tilelang_nsa_fwd_varlen.py: TileLang implementation with variable sequence lengths * example_triton_nsa_fwd_varlen.py: Triton implementation with variable sequence lengths - Update reference.py to support variable-length sparse attention scenarios - Enhance existing sparse attention implementations to handle variable-length inputs - Add comprehensive testing and validation for variable-length sparse attention * Refactor Native Sparse Attention Examples: Code Style and Formatting Improvements - Standardize function and parameter formatting across NSA example files - Improve code readability by adjusting indentation and line breaks - Enhance type hints and parameter alignment - Remove unnecessary whitespaces and optimize imports - Maintain consistent code style across TileLang and Triton implementations
-
You Jiacheng authored
It's slightly faster than T.copy then RS-GEMM, and simpler.
-
Lei Wang authored
* [Refactor] Update BitBLAS Benchmark with TileLang Carver Imports and Roller Hints Generation - Replace BitBLAS imports with TileLang Carver imports in benchmark_matmul.py - Modify roller hints generation using new TileLang Carver template and utility functions - Update get_roller_hints_from_func to handle None cases and improve return logic - Adjust DefaultPolicy to handle different codegen dictionary formats * [Refactor] Update Thread Binding and Import Statements in TileLang Kernels - Replace T.thread_binding() with T.get_thread_binding() across multiple kernel test files - Update import statements for MMA layout and macro generator in dequantize GEMM and FP8 examples - Move map_torch_type utility function to tilelang.utils.tensor - Remove unnecessary imports and improve code organization * Refactor Native Sparse Attention Example with Enhanced Triton Kernel - Update parallel_nsa_fwd_kernel to support more flexible sparse attention computation - Add support for block counts and offsets in the Triton kernel - Modify kernel grid and computation logic for improved performance - Update example script to use naive_nsa_simple reference implementation - Improve type hints and kernel configuration * Add Native Sparse Attention Examples with Tilelang and Triton Implementations - Introduce new example scripts for native sparse attention: * example_tilelang_nsa_fwd.py: Forward pass implementation using TileLang * example_tilelang_nsa_decode.py: Decoding-specific sparse attention implementation * example_triton_nsa_fwd.py: Triton-based sparse attention forward pass - Update reference.py with naive implementations for sparse attention - Support different sparse attention scenarios including forward pass and inference - Add comprehensive testing and validation against reference implementations * lint fix
-
Lei Wang authored
* [Refactor] Update BitBLAS Benchmark with TileLang Carver Imports and Roller Hints Generation - Replace BitBLAS imports with TileLang Carver imports in benchmark_matmul.py - Modify roller hints generation using new TileLang Carver template and utility functions - Update get_roller_hints_from_func to handle None cases and improve return logic - Adjust DefaultPolicy to handle different codegen dictionary formats * [Refactor] Update Thread Binding and Import Statements in TileLang Kernels - Replace T.thread_binding() with T.get_thread_binding() across multiple kernel test files - Update import statements for MMA layout and macro generator in dequantize GEMM and FP8 examples - Move map_torch_type utility function to tilelang.utils.tensor - Remove unnecessary imports and improve code organization * Refactor Native Sparse Attention Example with Enhanced Triton Kernel - Update parallel_nsa_fwd_kernel to support more flexible sparse attention computation - Add support for block counts and offsets in the Triton kernel - Modify kernel grid and computation logic for improved performance - Update example script to use naive_nsa_simple reference implementation - Improve type hints and kernel configuration
-
Lei Wang authored
* [Refactor] Update BitBLAS Benchmark with TileLang Carver Imports and Roller Hints Generation - Replace BitBLAS imports with TileLang Carver imports in benchmark_matmul.py - Modify roller hints generation using new TileLang Carver template and utility functions - Update get_roller_hints_from_func to handle None cases and improve return logic - Adjust DefaultPolicy to handle different codegen dictionary formats * [Refactor] Update Thread Binding and Import Statements in TileLang Kernels - Replace T.thread_binding() with T.get_thread_binding() across multiple kernel test files - Update import statements for MMA layout and macro generator in dequantize GEMM and FP8 examples - Move map_torch_type utility function to tilelang.utils.tensor - Remove unnecessary imports and improve code organization
-
- 06 Mar, 2025 2 commits
-
-
Lei Wang authored
Simplify the control flow in the MLA decode kernel by replacing TileLang's T.If construct with a standard Python if statement. This change improves code readability and maintains the existing logic for handling sequence length constraints during block-wise computation.
-
Yu Cheng authored
* [Dev] Adjust computation logic to avoid precision loss when casting acc_s from float to float16 - Remove redundant `acc_s_0` fragment in flash attention kernel - Simplify memory copy and reduction operations - Reorder memory copy and scaling steps for improved performance - Add Hopper-specific synchronization method in CUDA reduce template - Update reduce operation to use architecture-specific synchronization * [Dev] Add DeepSeek MLA Decoding (Paged+Varlen) kernel and Performance Benchmark Script - Implement comprehensive MLA (Multi-Head Latent Attention) decoding benchmark script - Add support for multiple implementations: Torch, TileLang, FlashMLA, FlashInfer, and Triton - Create flexible configuration for benchmarking different batch sizes, sequence lengths, and head configurations - Implement performance comparison and CSV output for detailed performance analysis - Add command-line argument support for targeted benchmarking and comparison * [Dev] Refactor MLA Paged Decoding Kernel with Improved Block Handling and Precision - Replace `d` parameter with `dv` to clarify value dimension in MLA decoding - Enhance block distribution logic for split KV processing - Improve handling of remaining blocks in split KV computation - Add initialization of `lse_max_local` to prevent potential precision issues - Optimize block start and range calculations for more accurate sequence processing * lint
-
- 05 Mar, 2025 2 commits
-
-
Lei Wang authored
* Change default log level from WARNING to INFO in TileLang initialization * Refactor Flash Attention Variable-Length MHA Example with Cython Backend Support - Update `example_mha_fwd_varlen.py` to use Cython backend for kernel compilation - Remove unused imports and simplify function signature - Modify `flashattn` function to handle max sequence length as a separate argument - Update kernel call to include max sequence length parameter - Improve code readability and remove commented-out code - Add print statement to confirm successful assertion * Refactor code formatting in TileLang lowering and example files - Improve line breaks and code formatting in `lower.py`, `wrapper.py`, and `tensor.py` - Simplify line breaks and reduce unnecessary whitespace - Enhance code readability by adjusting indentation and line breaks - Update example MHA forward pass script with cleaner tensor initialization * Update TileLang kernel test with import path changes for MMA layout and macro generator - Modify import statements in test_tilelang_kernel_dequantize_gemm.py - Replace bitblas imports with tilelang.intrinsics imports for MMA-related utilities - Update main function to use tilelang.testing.main() * Add Block Sparse Attention Examples for TileLang and Triton - Implement block sparse attention kernels for both TileLang and Triton - Add utility functions for generating sparse attention masks using top-k and threshold methods - Support causal and variable-length attention scenarios - Include test cases for different sequence length configurations - Demonstrate block-level sparse attention with configurable parameters * Refactor Block Sparse Attention Examples with Code Style Improvements - Improve code formatting in block_sparse_attn_tilelang.py and block_sparse_attn_triton.py - Enhance readability by adjusting line breaks and indentation - Simplify kernel and function calls with better formatting - Add whitespace and line break improvements for better code clarity * Enhance Layout Plotting with Multi-Replication and Dynamic Visualization - Update plot_layout function to support multiple replications in thread and value mapping - Improve thread and value mapping to handle replicated layouts - Dynamically adjust figure size and legend positioning - Add print statements for saved plot file paths - Modify example fragment_mma_load_a.py to uncomment and enable warp and block layout plotting * Refactor AtomicAdd functions in CUDA common header - Implement a generic template for AtomicAdd function - Specialize templates for half_t, bfloat16_t, and pointer types - Reorganize and clean up existing AtomicAdd implementations - Improve type handling and conversion in atomic operations * Remove unused import in MHA backward test file - Remove unnecessary argparse import from test_tilelang_kenrel_mha_bwd.py - Add blank line for improved code formatting - Minor code cleanup in test file * Add FP8 GEMM Example with TensorCore Intrinsics - Implement a new example for FP8 matrix multiplication using TensorCore intrinsics - Support E4M3 and E5M2 floating-point 8-bit data types - Add README with notes on current FP8 implementation limitations - Include correctness test for FP8 GEMM with different configurations - Demonstrate swizzle layout and pipeline optimizations for FP8 computation
-
Yu Cheng authored
[Dev] Adjust computation logic to avoid precision loss when casting acc_s from float to float16 (#141) - Remove redundant `acc_s_0` fragment in flash attention kernel - Simplify memory copy and reduction operations - Reorder memory copy and scaling steps for improved performance - Add Hopper-specific synchronization method in CUDA reduce template - Update reduce operation to use architecture-specific synchronization
-
- 04 Mar, 2025 2 commits
-
-
Yu Cheng authored
- Add non-split flash attention macro for more flexible kernel generation - Implement `main_no_split` function to handle single-split scenarios - Modify kernel selection logic to dynamically choose between split and non-split implementations
-
Yu Cheng authored
- Update news and MLA performance benchmark in README.md - Move performance benchmark and layout images to a dedicated 'figures' directory - Improve code formatting and image references in documentation
-
- 03 Mar, 2025 3 commits
-
-
Lei Wang authored
* Change default log level from WARNING to INFO in TileLang initialization * Refactor Flash Attention Variable-Length MHA Example with Cython Backend Support - Update `example_mha_fwd_varlen.py` to use Cython backend for kernel compilation - Remove unused imports and simplify function signature - Modify `flashattn` function to handle max sequence length as a separate argument - Update kernel call to include max sequence length parameter - Improve code readability and remove commented-out code - Add print statement to confirm successful assertion * Refactor code formatting in TileLang lowering and example files - Improve line breaks and code formatting in `lower.py`, `wrapper.py`, and `tensor.py` - Simplify line breaks and reduce unnecessary whitespace - Enhance code readability by adjusting indentation and line breaks - Update example MHA forward pass script with cleaner tensor initialization * Update TileLang kernel test with import path changes for MMA layout and macro generator - Modify import statements in test_tilelang_kernel_dequantize_gemm.py - Replace bitblas imports with tilelang.intrinsics imports for MMA-related utilities - Update main function to use tilelang.testing.main() * Add Block Sparse Attention Examples for TileLang and Triton - Implement block sparse attention kernels for both TileLang and Triton - Add utility functions for generating sparse attention masks using top-k and threshold methods - Support causal and variable-length attention scenarios - Include test cases for different sequence length configurations - Demonstrate block-level sparse attention with configurable parameters * Refactor Block Sparse Attention Examples with Code Style Improvements - Improve code formatting in block_sparse_attn_tilelang.py and block_sparse_attn_triton.py - Enhance readability by adjusting line breaks and indentation - Simplify kernel and function calls with better formatting - Add whitespace and line break improvements for better code clarity * Enhance Layout Plotting with Multi-Replication and Dynamic Visualization - Update plot_layout function to support multiple replications in thread and value mapping - Improve thread and value mapping to handle replicated layouts - Dynamically adjust figure size and legend positioning - Add print statements for saved plot file paths - Modify example fragment_mma_load_a.py to uncomment and enable warp and block layout plotting
-
Yu Cheng authored
-
Yu Cheng authored
* [Dev] Add RetNet Linear Attention example * [Dev] Add WgmmaSync rewriter for pipelined WGMMA operations and add MHA WGMMA pipelined example (FA3-like scheduling) This commit introduces a new transformation pass `RewriteWgmmaSync` to optimize warp group matrix multiply accumulate (WGMMA) operations in the TileLang compiler: - Implemented `WgmmaSyncRewriter` in `src/transform/wgmma_sync_rewriter.cc` - Added pass registration for `RewriteWgmmaSync` - Updated `tilelang/engine/phase.py` to include the new transformation pass - Updated `tilelang/transform/__init__.py` to expose the new pass The rewriter intelligently manages synchronization and dependencies between WGMMA operations, improving pipeline efficiency for complex matrix multiplication kernels. * [Bugfix] Fix bug in ThreadTagChecker for warp specialization Improve thread tag validation in warp specialized rewriter to prevent unintended transformations: - Add more precise checks for threadIdx.y and threadIdx.z - Validate thread extent to ensure only single-extent thread bindings are allowed - Prevent warp specialization for multi-extent thread bindings in y and z dimensions * lint * [CI] Add TMA descriptor attribute to transformed module in test case * [Dev] Refactor DeepSeek MLA Decode Example with Non-Split and Split Flash Attention Implementations - Add new `flash_attn` macro for non-split flash attention implementation - Add swizzled layout for tile in shared memory - Use threadblock swizzle to imporve L2 cache hit rate * [Dev] Add DeepSeek MLA Decode Example with Documentation and Performance Benchmarks - Add detailed README.md explaining MLA (Multi-Head Latent Attention) implementation - Include performance benchmark images for batch sizes 64 and 128 - Add layout visualization images for QK and PV operations - Implement torch reference implementations in torch_refs.py - Update example_mla_decode.py with command-line argument support and flexible configuration - Add performance benchmarking and comparison with other implementations
-
- 02 Mar, 2025 2 commits
-
-
Lei Wang authored
* Change default log level from WARNING to INFO in TileLang initialization * Refactor Flash Attention Variable-Length MHA Example with Cython Backend Support - Update `example_mha_fwd_varlen.py` to use Cython backend for kernel compilation - Remove unused imports and simplify function signature - Modify `flashattn` function to handle max sequence length as a separate argument - Update kernel call to include max sequence length parameter - Improve code readability and remove commented-out code - Add print statement to confirm successful assertion * Refactor code formatting in TileLang lowering and example files - Improve line breaks and code formatting in `lower.py`, `wrapper.py`, and `tensor.py` - Simplify line breaks and reduce unnecessary whitespace - Enhance code readability by adjusting indentation and line breaks - Update example MHA forward pass script with cleaner tensor initialization * Update TileLang kernel test with import path changes for MMA layout and macro generator - Modify import statements in test_tilelang_kernel_dequantize_gemm.py - Replace bitblas imports with tilelang.intrinsics imports for MMA-related utilities - Update main function to use tilelang.testing.main() * Add Block Sparse Attention Examples for TileLang and Triton - Implement block sparse attention kernels for both TileLang and Triton - Add utility functions for generating sparse attention masks using top-k and threshold methods - Support causal and variable-length attention scenarios - Include test cases for different sequence length configurations - Demonstrate block-level sparse attention with configurable parameters * Refactor Block Sparse Attention Examples with Code Style Improvements - Improve code formatting in block_sparse_attn_tilelang.py and block_sparse_attn_triton.py - Enhance readability by adjusting line breaks and indentation - Simplify kernel and function calls with better formatting - Add whitespace and line break improvements for better code clarity
-
Lei Wang authored
* Change default log level from WARNING to INFO in TileLang initialization * Refactor Flash Attention Variable-Length MHA Example with Cython Backend Support - Update `example_mha_fwd_varlen.py` to use Cython backend for kernel compilation - Remove unused imports and simplify function signature - Modify `flashattn` function to handle max sequence length as a separate argument - Update kernel call to include max sequence length parameter - Improve code readability and remove commented-out code - Add print statement to confirm successful assertion * Refactor code formatting in TileLang lowering and example files - Improve line breaks and code formatting in `lower.py`, `wrapper.py`, and `tensor.py` - Simplify line breaks and reduce unnecessary whitespace - Enhance code readability by adjusting indentation and line breaks - Update example MHA forward pass script with cleaner tensor initialization
-
- 28 Feb, 2025 3 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 * nas kernel * Enhance Native Sparse Attention Examples with Code Improvements and Parameter Updates - Updated example_tilelang_nsa.py and example_triton_nsa.py with code formatting and style improvements - Increased default number of heads and selected blocks in TileLang NSA example - Added Ruff linter ignore comments to reference.py - Standardized function signatures and improved code readability across NSA implementations * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Refactor DeepSeek MLA Decode Example with Enhanced Flash Attention Implementation - Update flash attention kernel to support positional embeddings (PE) - Modify reference implementation to handle PE and group query attention - Increase default batch size and adjust benchmarking parameters - Improve kernel performance and readability - Add einops and torch operations for more flexible tensor manipulation * Update README.md with corrected Flash MLA Decoding example path - Modify the example link for Flash MLA Decoding to point to the correct directory - Ensure accurate navigation to the DeepSeek MLA decoding example * Refactor Native Sparse Attention Kernel and Improve Utility Functions This commit introduces several improvements: - Simplified native sparse attention kernel by inlining macro functions in example_tilelang_nsa.py - Enhanced error handling in loop_partition.cc with more informative error messages - Updated print.py to support multi-dimensional buffer printing - Improved torch_assert_close in testing/__init__.py with more detailed mismatch reporting - Reduced default absolute tolerance in torch comparison from 1e-3 to 1e-2 - Added shape validation and detailed mismatch information in tensor comparison * Refactor Code Formatting and Improve Utility Functions This commit introduces several code formatting and utility improvements: - Add Ruff linter ignore comment in example_tilelang_nsa.py - Enhance code readability in loop_partition.cc and lower_tile_op.cc with improved line breaks - Simplify print_flat_buffer_with_condition in print.py - Refactor torch_assert_close in testing/__init__.py with improved line formatting * Enhance Buffer Printing Support for Fragment and Shared Memory Buffers This commit improves the print functionality in print.py by: - Adding support for printing fragment memory buffers - Implementing a new print_fragment_buffer_with_condition macro - Extending print_shared_buffer_with_condition for shared memory buffers - Updating the generic print function to handle different buffer scopes * Resolve merge conflict in print.py Remove merge conflict marker and clean up whitespace in the print module * Add Variable-Length Multi-Head Attention (MHA) Example with Flash Attention Support Introduce a new example script `example_mha_fwd_varlen.py` that demonstrates: - Variable-length Multi-Head Attention (MHA) implementation - Flash Attention forward pass with padding mask support - Performance benchmarking for variable-length sequences - Configurable parameters for batch, heads, sequence length, and dimension - Reference implementation comparison with PyTorch and FlashAttention * Refactor Flash Attention Variable-Length MHA Example Improve code formatting and readability in the variable-length multi-head attention example: - Add Ruff linter ignore comment - Enhance code style with consistent formatting - Remove unused imports - Improve line breaks and indentation - Simplify function signatures and lambda expressions
-
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 * nas kernel * Enhance Native Sparse Attention Examples with Code Improvements and Parameter Updates - Updated example_tilelang_nsa.py and example_triton_nsa.py with code formatting and style improvements - Increased default number of heads and selected blocks in TileLang NSA example - Added Ruff linter ignore comments to reference.py - Standardized function signatures and improved code readability across NSA implementations * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Refactor DeepSeek MLA Decode Example with Enhanced Flash Attention Implementation - Update flash attention kernel to support positional embeddings (PE) - Modify reference implementation to handle PE and group query attention - Increase default batch size and adjust benchmarking parameters - Improve kernel performance and readability - Add einops and torch operations for more flexible tensor manipulation * Update README.md with corrected Flash MLA Decoding example path - Modify the example link for Flash MLA Decoding to point to the correct directory - Ensure accurate navigation to the DeepSeek MLA decoding example * Refactor Native Sparse Attention Kernel and Improve Utility Functions This commit introduces several improvements: - Simplified native sparse attention kernel by inlining macro functions in example_tilelang_nsa.py - Enhanced error handling in loop_partition.cc with more informative error messages - Updated print.py to support multi-dimensional buffer printing - Improved torch_assert_close in testing/__init__.py with more detailed mismatch reporting - Reduced default absolute tolerance in torch comparison from 1e-3 to 1e-2 - Added shape validation and detailed mismatch information in tensor comparison * Refactor Code Formatting and Improve Utility Functions This commit introduces several code formatting and utility improvements: - Add Ruff linter ignore comment in example_tilelang_nsa.py - Enhance code readability in loop_partition.cc and lower_tile_op.cc with improved line breaks - Simplify print_flat_buffer_with_condition in print.py - Refactor torch_assert_close in testing/__init__.py with improved line formatting
-
Yu Cheng authored
[Dev][Bugfix] Fix bug in ThreadTagChecker; Add WgmmaSync rewriter and add MHA WGMMA pipelined example (#128) * [Dev] Add RetNet Linear Attention example * [Dev] Add WgmmaSync rewriter for pipelined WGMMA operations and add MHA WGMMA pipelined example (FA3-like scheduling) This commit introduces a new transformation pass `RewriteWgmmaSync` to optimize warp group matrix multiply accumulate (WGMMA) operations in the TileLang compiler: - Implemented `WgmmaSyncRewriter` in `src/transform/wgmma_sync_rewriter.cc` - Added pass registration for `RewriteWgmmaSync` - Updated `tilelang/engine/phase.py` to include the new transformation pass - Updated `tilelang/transform/__init__.py` to expose the new pass The rewriter intelligently manages synchronization and dependencies between WGMMA operations, improving pipeline efficiency for complex matrix multiplication kernels. * [Bugfix] Fix bug in ThreadTagChecker for warp specialization Improve thread tag validation in warp specialized rewriter to prevent unintended transformations: - Add more precise checks for threadIdx.y and threadIdx.z - Validate thread extent to ensure only single-extent thread bindings are allowed - Prevent warp specialization for multi-extent thread bindings in y and z dimensions * lint * [CI] Add TMA descriptor attribute to transformed module in test case
-
- 27 Feb, 2025 1 commit
-
-
Lei Wang authored
* refactor code * enhance tutorial * Enhance error handling and code generation in CUDA and TileLang components This commit introduces several improvements across multiple files: - Added more informative error messages in GEMM layout checks - Updated CUDA codegen to support more flexible function signature generation - Improved TMA descriptor initialization and kernel dispatch logic - Refined library generation and source code parsing utilities - Enhanced error handling in various adapter and wrapper classes * Add thread tag validation for warp specialization Introduce a ThreadTagChecker to validate that a PrimFunc only uses threadIdx.x before applying warp specialization. This prevents unintended transformations on kernels with complex thread binding and provides a clear warning to users about potential issues with warp specialization. * Update TileLang Profiling and Compilation in Flash Decoding Examples Refactor the profiling and compilation workflow in two flash decoding example scripts: - Replace `tilelang.lower()` and `tilelang.Profiler()` with `tilelang.compile()` - Simplify profiler initialization using `get_profiler()` - Update method calls to use the new profiler and compiled kernel objects - Maintain existing performance benchmarking and validation logic * Refactor and clean up code formatting in TileLang testing and adapter modules This commit includes several code style and formatting improvements: - Adjust whitespace and line breaks in test files - Improve code formatting in CUDA source wrapper and adapter utilities - Enhance readability of function calls and argument handling - Remove unnecessary whitespace and standardize indentation - Simplify function signatures and argument parsing * Refactor CUDA codegen and improve code formatting This commit includes several improvements to CUDA code generation and formatting: - Enhance function signature generation in CodeGenTileLangCUDA - Improve code formatting and readability in CUDA-related files - Simplify parameter handling and type annotations - Clean up whitespace and line breaks in codegen and layout files --------- Co-authored-by:Ubuntu <dlisuser@h100testl730RPS.xu5snccwrbtejcqqalluoku5hb.xx.internal.cloudapp.net>
-
- 26 Feb, 2025 2 commits
-
-
Yu Cheng authored
-
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 * nas kernel * Enhance Native Sparse Attention Examples with Code Improvements and Parameter Updates - Updated example_tilelang_nsa.py and example_triton_nsa.py with code formatting and style improvements - Increased default number of heads and selected blocks in TileLang NSA example - Added Ruff linter ignore comments to reference.py - Standardized function signatures and improved code readability across NSA implementations * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers * Refactor DeepSeek MLA Decode Example with Enhanced Flash Attention Implementation - Update flash attention kernel to support positional embeddings (PE) - Modify reference implementation to handle PE and group query attention - Increase default batch size and adjust benchmarking parameters - Improve kernel performance and readability - Add einops and torch operations for more flexible tensor manipulation * Update README.md with corrected Flash MLA Decoding example path - Modify the example link for Flash MLA Decoding to point to the correct directory - Ensure accurate navigation to the DeepSeek MLA decoding example
-
- 25 Feb, 2025 3 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 * nas kernel * Enhance Native Sparse Attention Examples with Code Improvements and Parameter Updates - Updated example_tilelang_nsa.py and example_triton_nsa.py with code formatting and style improvements - Increased default number of heads and selected blocks in TileLang NSA example - Added Ruff linter ignore comments to reference.py - Standardized function signatures and improved code readability across NSA implementations * Add utility math functions for integer operations - Implement `next_power_of_2()` to calculate the next power of 2 for an integer - Add `cdiv()` function for ceiling division of integers
-
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
-
- 24 Feb, 2025 1 commit
-
-
Lei Wang authored
-
- 23 Feb, 2025 1 commit
-
-
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
-
