zijie-tian/nano-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

📝 docs: add XAttention density alignment verification results

Browse Source 
- Add verification doc comparing GPU-only vs Offload mode density
- Test results: 32K (0.37% diff), 64K (0.09% diff) - alignment successful
- Both modes achieve 100% accuracy on RULER niah_single_1

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
This commit is contained in:
Zijie Tian
2026-02-05 01:59:11 +08:00
parent 1eb7521994
commit 54fd302fa8
2 changed files with 143 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
CLAUDE.md
View File

@@ -44,6 +44,7 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
| [`docs/xattn_memory_benchmark.md`](docs/xattn_memory_benchmark.md) | 📊 BENCH: XAttention 内存基准测试，Qwen3-0.6B 32K 在 24GB 显存可行 (gpu-util=0.28) |
| [`docs/xattn_offload_stream_sync_fix.md`](docs/xattn_offload_stream_sync_fix.md) | 🐛 FIX: XAttention Offload stream 同步 bug，Pass1/Pass2 K 数据不一致，compute_stream 包装 |
| [`docs/xattn_density_types.md`](docs/xattn_density_types.md) | 📊 Compute vs Comm density: BSA block (128) vs CPU block (4096) 粒度，聚合效应导致 comm=100% |
| [`docs/xattn_density_alignment_verification.md`](docs/xattn_density_alignment_verification.md) | ✅ VERIFIED: GPU-only vs Offload density 对齐验证 (32K 差异 0.37%, 64K 差异 0.09%) |
## Rules Index