README
Ornith-1.0-35B — ROCmFP4 STRIX_LEAN, DFlash baked in (cross-model draft)
A single-file, self-accelerating GGUF. Notably: the embedded DFlash draft was trained for a different model — Qwen3.6-35B-A3B — and still runs at 98.9% acceptance here, because Ornith shares its architectural lineage (Qwen3.5-MoE) with the draft's intended target. See z-lab/dflash#149 for the write-up on this transfer finding.
No --model-draft, no --spec-type flag — point -m at this file and speculative decoding just happens.
llama-server -m Ornith-1.0-35B-STRIX_LEAN-DFLASH.gguf -ngl 999 -fa on --jinja -c 65536
Requirements
Needs a ROCmFP4-aware build and DFlash-graft support for embedded drafts:
- gsrunion/rocmfp4-llama branch
dflash-graft, validated on AMD Strix Halo / gfx1151
On first load the server extracts the draft to a small cached sidecar next to the model (~1 second, one-time).
Measured performance
AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB unified LPDDR5X), server-timing, self-accelerating load:
| tok/s | acceptance | |
|---|---|---|
| Baked single-file, cross-model draft | 111.6 | 98.9% (430/435) |
| Two-file cross-model config | 112.5 | 97–98% |
| Plain LEAN, no draft | 63.1 | — |
+77% over plain decode, entirely from a draft model that was never trained on Ornith.
Credits
- Base model: deepreinforce-ai — Ornith-1.0-35B (MIT)
- DFlash draft (cross-model): z-lab — trained for Qwen3.6-35B-A3B
- ROCmFP4 quant formats: Hal0ai; fork base: charlie12345
- Vision (mmproj) not included — DFlash and mmproj cannot currently coexist in this serving stack; use the base ROCmFPX release for multimodal.
Contributors
2 commits
Similar repos
Linked in READMEs
README
Ornith-1.0-35B — ROCmFP4 STRIX_LEAN, DFlash baked in (cross-model draft)
A single-file, self-accelerating GGUF. Notably: the embedded DFlash draft was trained for a different model — Qwen3.6-35B-A3B — and still runs at 98.9% acceptance here, because Ornith shares its architectural lineage (Qwen3.5-MoE) with the draft's intended target. See z-lab/dflash#149 for the write-up on this transfer finding.
No --model-draft, no --spec-type flag — point -m at this file and speculative decoding just happens.
llama-server -m Ornith-1.0-35B-STRIX_LEAN-DFLASH.gguf -ngl 999 -fa on --jinja -c 65536
Requirements
Needs a ROCmFP4-aware build and DFlash-graft support for embedded drafts:
- gsrunion/rocmfp4-llama branch
dflash-graft, validated on AMD Strix Halo / gfx1151
On first load the server extracts the draft to a small cached sidecar next to the model (~1 second, one-time).
Measured performance
AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB unified LPDDR5X), server-timing, self-accelerating load:
| tok/s | acceptance | |
|---|---|---|
| Baked single-file, cross-model draft | 111.6 | 98.9% (430/435) |
| Two-file cross-model config | 112.5 | 97–98% |
| Plain LEAN, no draft | 63.1 | — |
+77% over plain decode, entirely from a draft model that was never trained on Ornith.
Credits
- Base model: deepreinforce-ai — Ornith-1.0-35B (MIT)
- DFlash draft (cross-model): z-lab — trained for Qwen3.6-35B-A3B
- ROCmFP4 quant formats: Hal0ai; fork base: charlie12345
- Vision (mmproj) not included — DFlash and mmproj cannot currently coexist in this serving stack; use the base ROCmFPX release for multimodal.
Similar repos
Linked in READMEs
Contributors
2 commits