Commit 23b2f19
feat: fix trinity_format.zig and add .tri loader integration
Fixes:
- Add TritWeight alias to prometheus_seed.zig
- Add toInt(), fromEncoding(), toEncoding() methods to Trit enum
- Fix enum case names (Pos/Neg/Zero vs pos/neg/zero)
New features:
- BitNetWeights struct for loaded model weights
- loadTriFile() function to load .tri format models
- getTriInfo() function to read model metadata
- loadFromTriFile() in bitnet_pipeline.zig for full model loading
New tests (21 total, all passing):
- create and load tri file
- generation produces valid tokens
- sampling with different temperatures
- KV cache grows correctly
- RoPE rotates vectors
Performance:
- Single layer: 6.4-8.3 ms
- 28 layers: 178-234 ms/token
- Throughput: 4.3-5.6 tok/s
- SIMD matmul: 1.04 GFLOPS
Co-authored-by: Ona <no-reply@ona.com>1 parent 3390f01 commit 23b2f19
4 files changed
Lines changed: 591 additions & 6 deletions
Binary file not shown.
0 commit comments