Ornith-1.5-9B-DFlash2
DFlash2 draft for ornith-ai/Ornith-1.5-9B-MLX-4bit, produced by the
no-op-graft + selector-training recipe:
- Trunk:
ornith-ai/Ornith-1.5-9B-DFlash(DFlashV1) weights, copied verbatim. - Convolutions:
attention_conv/mlp_convgrafted as identity (unit self-tap, zero dynamic corrections) — the runtime gets identity convs for free. - Candidate selector:
hidden_projection+ predecessor/successor codebooks trained on features captured from the quantized 4-bit target (832 samples, top-16 hard-negative margin, convs frozen).
The selector is the trained component — trained on the deployed quant's hidden states, not the bf16 target's, which is why it transfers.
Measured (M5 Pro, opencode_sim n=4)
| Draft | serial | packed B4 | speedup |
|---|---|---|---|
| Ornith-1.5-9B-DFlash (V1) | ~132 tok/s | ~161 tok/s | ~1.8x |
| this checkpoint | 161.7 tok/s | 768.7 tok/s | 4.43x |
Diverse-traffic pipeline (file-read prompts, 4 concurrent lanes): ~212 tok/s vs V1's ~161 — smaller but real on non-repetitive content.
Serving notes
architectures: ["DFlash2DraftModel"], hidden 4096, dynamic 1024, conv_group_size 16, block_size 8, selector top_k 16.- Requires the RichEngine
Hidden4096draft context core (draft_conv_h4096).