Ornith-1.0-9b-ROCmFPX-STRIX_LEAN-GGUF

Ornith-1.0-9B ROCmFPX STRIX_LEAN — GGUF

ROCmFPX Q4_0_ROCMFP4_STRIX_LEAN quant of deepreinforce-ai/Ornith-1.0-9B (qwen35 hybrid SSM+attention, 8.95 B params, 262 144 native ctx).

Built with charlie12345/ROCmFPX on a Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11. Quantized 2026-06-27 with build commit 11d76c2.

This is not a stock llama.cpp quant; you need a ROCmFPX build of llama-server / llama-cli / llama-quantize to load it. The ROCmFP4 weight format is unknown to stock llama.cpp and will fail with unknown quantization.

Scope of these benchmarks — read this first

These numbers are a light baseline, not a thorough ROCmFPX evaluation. The mesh's bench framework is built for production agent workload regression-detection on the local stack, not for the kind of multi-axis sweep that upstream quant maintainers typically publish. Specifically:

• Harness scope is bounded. The numbers below come from llama-bench ctx sweeps + the mesh's mesh_eval (6 tests, 4 deterministic + throughput) + hermes_loop_eval (5 agent scenarios). That's a regression suite, not a quality benchmark — it answers "does this quant still serve the mesh's agent stack correctly," not "is this the best possible 4-bit ROCmFP4 quant of this model."
• Sample sizes are small. Throughput numbers are 3 reps on a single GPU; correctness is 5 prompts × 16 tokens; agent loop is 5 scenarios with one-shot generation. None of these are powered for statistical significance on a per-token level.
• No perplexity / wikitext / MMLU / GSM8K. The mesh's stack isn't a quality benchmark — those are upstream ROCmFPX's territory. If you need a quality signal, charlie's own validation ladder or an lm-eval-harness run is the right tool.
• Single GPU class. All measurements are on a 16 GB RDNA4 (RX 9060 XT, gfx1200). No Strix unified-memory, no CDNA, no multi-GPU, no Vulkan, no CUDA. Cross-hardware generalization is not implied.
• No human eval. Quality is "byte-identical on factual / short deterministic outputs, divergent on high-entropy creative generation" — which is expected for any 4-bit quant, not a quality verdict on this one specifically.

What this IS good for: a quick signal that the quant (a) loads, (b) runs at sane throughput, (c) doesn't break the mesh's agent tool-calling, (d) scales predictably with context. What this is NOT good for: claiming "this is the best quant of this model," reproducing academic benchmark results, or substituting for upstream's validation work.

If you want the rigorous version, charlie's own ROCmFPX brief + the model's stock GGUF variants (e.g. bartowski/deepreinforce-ai_Ornith-1.0-9B-GGUF) are the place to look.

What we measured

Hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11 Software: charlie12345/ROCmFPX main @ 11d76c2 Source GGUF: ornith-1.0-9b-bf16.gguf (BF16, 17.9 GB) from deepreinforce-ai/Ornith-1.0-9B Same-class baseline: stock llama.cpp b9608 Q4_K_M quantized from the same BF16 source

Throughput vs stock Q4_K_M

llama-bench 3 reps, q8_0 KV, fa=on, ngpu-layers=99. Raw JSON: BENCH-strix-lean-ctx-sweep.json and BENCH-stock-Q4_K_M-ctx-sweep.json.

Findings (small-sample, 3 reps — see the scope caveat above):

• STRIX_LEAN beats stock Q4_K_M at every ctx tested on prompt processing.
• Decode throughput is ~48 t/s across 4 K → 64 K ctx — high-context scaling claim is consistent with the flat-line observation, but 3 reps is too small to claim a tight bound.
• Prompt-processing edge narrows as ctx grows (KV cache dominates at 64 K).
• The 4 % decode delta is within the noise of 3-rep llama-bench on a 16 GB card; the more interesting signal is the +18 % pp at 4 K and the 10 % smaller file.

KV cache type sweep (added v0.5.134, head_dim=128)

131 K ctx, fa=on, kv-unified, -np 1:

turbo4 is the production default for any head_dim=128 model in the ROCmFPX build. -1.1 GB VRAM, same speed. The turbo3/4 KV types are TheTom's turboquant, absorbed into ROCmFPX main via PlunderStruck commits d859c9e + d0141e8.

Correctness — STRIX_LEAN vs stock Q4_K_M (5 prompts × 16 tokens, top-10 logprobs, sequential A/B)

Byte-identical on factual / short deterministic outputs (KL < 0.4, argmax 87-100 %), high divergence on open-ended creative generation (KL 11-16, argmax 6-38 %). Divergence correlates with prompt entropy — where multiple tokens are near-equal, greedy argmax flips and KL amplifies. This is expected for any 4-bit quant.

Agent / loop validation (raw JSONs included)

mesh_eval.py 4 deterministic tests + throughput (raw-mesh-eval-ornith-strix-lean.json):

hermes_loop_eval.py 5 scenarios (raw-hermes-loop-ornith-strix-lean.json):

The 4/5 loop is the error_recovery scenario's strict-match failure, not a quant defect. The model behaved correctly.

Quick start

# Build llama.cpp with ROCmFPX (the ROCmFPX-fork supports Q4_0_ROCMFP4_STRIX_LEAN weight type)
git clone https://github.com/charlie12345/ROCmFPX
cd ROCmFPX
cmake -S . -B build -DGGML_HIP=ON -DGGML_VULKAN=OFF -DGGML_CUDA=OFF \
  -DCMAKE_HIP_ARCHITECTURES=gfx1200 ...
cmake --build build --target llama-server llama-cli llama-quantize

# Serve (131 072 ctx, turbo4 KV for head_dim=128, fa=on)
./build/bin/llama-server \
  -m Ornith-1.0-9b-ROCmFPX-STRIX_LEAN.gguf \
  -np 1 -c 131072 \
  -ctk turbo4 -ctv turbo4 \
  -kvo -cram 32768 -fa on

Reproduce the quant

# Source (we used the BF16 gguf; any BF16/F16 gguf of the same parent works)
SRC=/mnt/e/llms-models-data/ornith/ornith-1.0-9b-bf16.gguf

# ROCmFPX llama-quantize (preset is built in; see `llama-quantize --help`)
~/ROCmFPX/build-rdna4/bin/llama-quantize \
  $SRC \
  Ornith-1.0-9b-ROCmFPX-STRIX_LEAN.gguf \
  Q4_0_ROCMFP4_STRIX_LEAN

Quantize time: ~6 min cold, <2 min warm-cache. CPU-only, no GPU required.

Files in this repo

What's NOT in this repo (caveats)

• Stock llama.cpp will not load this file. The ROCmFP4 weight format is unique to charlie12345/ROCmFPX. Use that fork's llama-server/llama-cli/llama-quantize.
• No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200). Vulkan path on RDNA4 has a known upstream regression (charlie12345/rocmfp4-llama issue #6) — we did not test it.
• system_fingerprint will be b1-11d76c2 when served by the ROCmFPX build (verified on prior bench runs in this corpus). If you see a different fingerprint, the wrong binary loaded the file.
• No multi-GPU / tensor-parallel bench. 9 B params at 4.7 GB fits comfortably on a single 16 GB card; no need to split.
• No MTP / speculative-decode bench on this file. Ornith 1.0 9B does not ship with MTP draft heads.
• No vision/multimodal test. Ornith 1.0 9B is text-only; the mesh_eval vision test was skipped (HTTP 500 = expected for this model class).

Provenance

• Source model: deepreinforce-ai/Ornith-1.0-9B — qwen35 hybrid SSM+attention, 8.95 B params, native ctx 262 144
• Source model license: MIT (https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B/blob/main/LICENSE)
• Quantizer: charlie12345/ROCmFPX main @ 11d76c2 (2026-06-27)
• Quantizer license: MIT
• Build hardware: Node B, AMD Ryzen 9 5900XT 16-core, Radeon RX 9060 XT 16 GB (gfx1200), ROCm 7.2.3, NixOS 25.11
• Build tooling: NixOS 25.11, ROCm store paths dynamic-discovered. See the meshina repo's references/nixos-rocm-external-build-recipe.md for the build env setup.
• Bench harnesses: scripts/mesh-bench/mesh_eval.py + scripts/mesh-bench/hermes_loop_eval.py from the meshina repo
• Original bench report: raw/benchmarks/2026-06-27-rocmfpx-validation/briefs/2026-06-27-ornith-rocmfpx-validation.md in the meshina repo

License

• The Ornith 1.0 9B parent model is MIT (per its HF model card).
• The charlie12345/ROCmFPX quantizer is MIT.
• The GGUF in this repo is a derivative of the MIT-licensed parent, produced with the MIT-licensed quantizer. The MIT license is preserved.