GitLab

* [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

[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

* [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

* 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

* 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

* [CI][Test] Add test cases for tilelang transform MultiVersionBuffer and WarpSpecialized

* Relax the mismatch ratio restrictions in the flash_linear_attention and mha tests

* [Dev] Add mha backward example

* [Dev] Add mla decode example

* bug fix

* Add triton impl

* Add gqa decode example

* [Dev] Add GQA decode example

* lint

* delete unused triton example

* set default profiler to 'auto'