Ornith-1.5-35B-A3B-DFlash2
DFlash2 draft for ornith-ai/Ornith-1.5-35B-A3B-MLX-4bit, produced by the
no-op-graft + selector-training recipe:
- Trunk:
ornith-ai/Ornith-1.5-35B-A3B-DFlash(DFlashV1) weights, copied verbatim. - Convolutions:
attention_conv/mlp_convgrafted as identity (unit self-tap, zero dynamic corrections). - Candidate selector:
hidden_projection+ predecessor/successor codebooks trained on features captured from the quantized 4-bit target (741 samples, top-16 hard-negative margin, convs frozen).
Measured (M5 Pro, opencode_sim n=4, engine decode)
| Draft | serial | packed B4 | speedup |
|---|---|---|---|
| Ornith-1.5-35B-A3B-DFlash (V1) | 142.8 tok/s | 147.4 tok/s | 1.05x |
| jzinno/Ornith-1.5-35B-A3B-DFlash2 (bf16-trained) | 78.3 | 114.6 | 1.32x |
| this checkpoint | 190.1 tok/s | 316.8 tok/s | 1.65x |
The bf16-trained DFlash2 underperforms on the 4-bit target — its selector was calibrated to bf16 hidden states. This checkpoint trained on the deployed quant's own features, which is why it scales with packed lanes.
Serving notes
architectures: ["DFlash2DraftModel"], hidden 2048, dynamic 512, conv_group_size 16, block_size 8, selector top_k 16.- Requires the RichEngine
Hidden2048draft context core (draft_conv_h2048).