Ornith-1.5-122B — ROCmFPX Q4 GGUF

ornith-ai/Ornith-1.5-122B quantized to two 4-bit ROCmFPX types. ROCmFPX is a llama.cpp fork with AMD-oriented low-bit weight types for Vulkan and HIP/ROCm.

• Q4_0_ROCMFP4, balanced: the closest 4-bit file to the original. Start here.
• Q4_0_ROCMFP4_STRIX: the recipe ROCmFPX tuned for speed on AMD Strix Halo (Ryzen AI Max, gfx1151); smaller, a bit less exact.

Download one variant, not the whole repo. The repo holds both versions (146 GB together); you need only one: balanced is 80 GB, Strix is 65 GB, plus 0.9 GB for image input. Each is split into parts below 50 GB; point llama.cpp at the first part and it loads the rest. Image input works with the included F16 vision projector (--mmproj).

# balanced (80 GB)
hf download Harsha001/Ornith-1.5-122B-ROCmFPX --include "*Q4_0_BALANCED-*" "mmproj*" --local-dir Ornith-1.5-122B
# or Strix (65 GB)
hf download Harsha001/Ornith-1.5-122B-ROCmFPX --include "*Q4_0_STRIX-*" "mmproj*" --local-dir Ornith-1.5-122B

Memory on a 128 GB Strix Halo: the GPU can only use the share of memory given to it. Give it 96 GB (in the BIOS on Windows, or through the GTT/TTM limit on Linux): then either variant fits with room for context. With a 64/64 split neither fits with usable context (Strix would leave about 3 GB).

These files only load in the ROCmFPX fork. Stock llama.cpp, LM Studio and Ollama do not know these types.

Files

The files are named ROCmFP4-Q4_0_BALANCED and ROCmFP4-Q4_0_STRIX so that Hugging Face's size panel counts the two versions separately; it reads both full type names as the same label. The weights inside are Q4_0_ROCMFP4 and the Q4_0_ROCMFP4_STRIX recipe, unchanged.

Measured quality

Measured against the BF16 model (228 GB) on WikiText-2 test, 12 chunks of 512 tokens, CPU backend. KLD is the mean KL divergence of the token distribution from BF16 (lower is better). Same top is how often the quant picks the same next token as BF16.

This model is more sensitive to 4-bit quantization than most: even the standard Q4_K_M, made from the same BF16 weights and importance matrix, moves clearly away from BF16. Twelve chunks is a small sample, so take the numbers as a guide (± about 0.01 on KLD).

How this was made

1. convert_hf_to_gguf.py --no-mtp --outtype bf16 from the ROCmFPX fork (fb08d7c). The multi-token-prediction head is left out, because the fork does not use it for decoding.
2. Importance matrix: bartowski's Ornith-1.5-122B-imatrix.gguf (573 chunks × 512 tokens of his calibration set v6). Computing our own from the 228 GB BF16 model on a CPU with 121 GB of RAM would take about 12 hours. Thanks to bartowski.
3. llama-quantize --imatrix from ROCmFPX (fb08d7c), from the BF16 weights, one run per type.
4. Perplexity and KL divergence on the CPU backend. The fork's CUDA kernels for these types are not correct on NVIDIA, so no GPU numbers are used.
5. Each file was checked with a short chat prompt and an image prompt before upload.

Run

git clone https://github.com/charlie12345/ROCmFPX && cd ROCmFPX
cmake -B build -DGGML_VULKAN=ON          # or -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151
cmake --build build --target llama-server -j
build/bin/llama-server -m Ornith-1.5-122B-ROCmFP4-Q4_0_BALANCED-00001-of-00002.gguf --mmproj mmproj-Ornith-1.5-122B-f16.gguf -ngl 999 -fa on --jinja

License

MIT, same as the base model. All credit for the model goes to Ornith AI.