·
207 commits
to ascendc_pto
since this release
🎯 Highlights
🤖 AI Agent Skills System (New)
Automated operator development with AI-powered skills:
- tilelang-op-design: Generate design documents
- tilelang-op-generate: Generate runnable code from designs
- tilelang-pass-analyzer: Pass analysis and comparison
- tilelang-error-fixer: Error diagnosis and fixing
🚀 Support DeepSeek V4 kernels
- Sparse Flash Attention (examples/deepseek_v4/sparse_flash_attention.py)
- mHC（examples/deepseek_v4/hc_split_sinkhorn.py）
- Activation Quantization（examples/deepseek_v4/act_quant.py）
- Int8 GEMM（examples/deepseek_v4/int8_gemm.py）
For more details, please refer to CANN community:
https://gitcode.com/cann/cann-recipes-infer/blob/master/docs/models/deepseek-v4/deepseek_v4_tilelang_operator_guide.md
https://gitcode.com/cann/cann-recipes-infer/tree/master/ops/tilelang/ds_v4
🚀 High-Performance Attention Kernels optimization
- FlashAttention: High-performance implementation + optimization guide
- Sparse Flash Attention: Benchmark & optimization guide
🔧 API Enhancements
- Support T.alloc_var allocation (aligned with tilelang mainstream)
- Support clear parameter in T.reduce_xxx to align with tilelang mainstream
- Support T.tile.atomic_add: UB-to-GM atomic accumulation
- Add T.tile.clamp support to bound the upper and lower values
- Add T.tile.sigmoid support for sigmoid operation.
- Refactor sort operations: T.tile.sort/T.tile.merge_sort/T.tile.topk
- Refactor T.tile.gather/T.tile.gather_mask.
- Remove the tmp parameter in the APIs.
- Support slice operation for calculation APIs.
- Support shmem put/get API: Inter-core communication
🛠️ Infrastructure
CI/CD
- Incremental testing (only modified directories)
- /re-test real-time feedback (⏳ → 🚀 → ✅/❌)
- Auto format check (ruff + clang-format)
Installation - Wheel package: pip install tilelang-*.whl
- IDE support: .clangd + compile_flags.txt
📚 Documentation
- TileLang-Ascend Programming Guide
- FlashAttention optimization guide
- 5 video lessons (environment, basics, developer mode, tuning, deployment)
🔧 Bug Fixes
- gemm rw_mask crash
- CrossCorePipeline shared buffer analysis
- softmax/layernorm dtype generalization
- PipelinePlanning empty loops