Zijie Tian
|
232fcf043e
|
📝 docs: add GPU-only density alignment test results
Document test results verifying XAttention density calculation in
GPU-only mode matches independent xattn_estimate calls.
Test results (Llama-3.1-8B-Instruct, threshold=0.9):
- 4k: Layer 0 density 63.8%, verified ✅
- 8k: Layer 0 density 65.0%, verified ✅
- 16k: Layer 0 density 61.6%, verified ✅
- 32k: Layer 0 density 50.2%, verified ✅
- 64k: Layer 0 density 37.0%, verified ✅
All tests show exact match (attn_sums diff=0, mask exact match).
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>
|
2026-02-02 11:22:34 +08:00 |
|