Zijie Tian ac1ccbceaa feat: add XAttention sparse policy integration ...

Integrate COMPASS XAttention algorithm into nano-vllm's CPU offload
execution path. Uses FlashAttention with native GQA support for
offload mode.

New files:
- nanovllm/kvcache/sparse/utils.py: find_blocks_chunked() utility
- nanovllm/kvcache/sparse/kernels.py: Triton kernels for XAttention
- nanovllm/kvcache/sparse/xattn.py: XAttentionPolicy implementation

Modified:
- nanovllm/config.py: Add XATTN configuration parameters
- nanovllm/engine/model_runner.py: Support XATTN policy
- nanovllm/kvcache/sparse/__init__.py: Register XAttentionPolicy
- tests/test_ruler.py: Add --sparse-policy parameter

Test results (32k ruler):
- NIAH tasks: 12/12 (100%)
- QA/Recall tasks: 11/15 (73%)
- Overall: 23/27 (85%)

Co-Authored-By: Claude <noreply@anthropic.com>

2026-01-14 10:04:46 +08:00

4.3 KiB

Raw Blame History

View Raw