📝 docs: add XAttention algorithm guide based on COMPASS implementation

- Create docs/xattention_algorithm_guide.md with detailed algorithm explanation
  - Stride reshape (inverse mode) for Q/K interleaved sampling
  - Triton kernels: flat_group_gemm_fuse_reshape, softmax_fuse_block_sum
  - Block selection via find_blocks_chunked with cumulative threshold
  - BSA (block_sparse_attn) dependency for sparse computation
- Update docs/sparse_attention_guide.md XAttention section with accurate description
- Add documentation index entry in CLAUDE.md

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
Zijie Tian
2026-01-20 02:50:03 +08:00
parent 07f5220f40
commit e440c45e73
3 changed files with 395 additions and 25 deletions

View File

@@ -14,6 +14,7 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
| [`docs/sparse_policy_architecture.md`](docs/sparse_policy_architecture.md) | SparsePolicy abstraction: prefill/decode delegation, pipeline modes, policy implementations |
| [`docs/sparse_policy_implementation_guide.md`](docs/sparse_policy_implementation_guide.md) | How to implement custom SparsePolicy: required methods, hooks, ring buffer pipeline pattern |
| [`docs/sparse_attention_guide.md`](docs/sparse_attention_guide.md) | Block sparse attention methods (XAttention, FlexPrefill, MInference, AvgPool, Quest), computation flow, algorithms |
| [`docs/xattention_algorithm_guide.md`](docs/xattention_algorithm_guide.md) | XAttention 算法详解: stride reshape、Triton kernels、BSA 依赖、块选择算法 |
| [`docs/debugging_guide.md`](docs/debugging_guide.md) | PyTorch hooks for debugging, hook positions, tensor comparison, memory profiling |
| [`docs/optimization_guide.md`](docs/optimization_guide.md) | Performance optimizations: sgDMA (15x), Triton merge (4.3x), N-way pipeline (2x) |
| [`docs/known_issues.md`](docs/known_issues.md) | Documented bugs and fixes: partial last block bug, block size 4096 race condition |