CIRU DualView · Ornith 1.0 35B v1

CIRU DualView + Ornith 1.0 35B

This is the immutable ROCmFPX runtime used by the first public CIRU DualView
model:

Ornith1.0-35b-CIRU-DUALVIEW-FPX7+Q8-MTP on Hugging Face

DualView keeps packed Q7_0_ROCMFPX source weights for the decode lane and
losslessly expands the stored signed codes into a canonical Q8 compute view for
the gfx1151 prefill lane. The Ornith release adds topology-selected canonical
Q8 quality islands and a byte-exact official Q8 MTP head.

Validated headline, using the retained target-only research GGUF:

• PP4096: 1,236.156 tok/s
• Native TG256: 48.049 tok/s
• PPL: 8.461721
• Mean KLD versus official Ornith Q8_0: 0.009796

The public integrated GGUF is exactly 33,536,832,416 bytes with SHA-256
7834fb92d451235c123973ba40eccea156db28af8d42ae13178c7efcc4d0177a.
It is not memory-light: the validated runtime paths measured roughly
68.20–73.08 GB of GTT.

Build

sudo apt update
sudo apt install -y build-essential cmake git ninja-build pkg-config \
  libcurl4-openssl-dev

git clone --branch dualview-ornith-35b-v1 --single-branch \
  https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
JOBS="$(nproc)" ./scripts/build-strix-dualview.sh

Before loading the model:

./build-strix-dualview/bin/test-q7-q8-view
./build-strix-dualview/bin/test-quantize-fns
./build-strix-dualview/bin/llama-bench --list-devices

Linux ROCm/HIP on Strix Halo gfx1151 is the validated target. Windows, other
AMD targets, Vulkan, and stock upstream llama.cpp are not qualified for this
DualView compute-shadow release.

• Visual introduction
• Complete runtime guide
• Full Ornith research record