Modelpublic

pugant/grug-35b-v2-ROCmFP4-STRIX_LEAN

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face

1likes644downloads

Model Card

grug-35b-v2 — ROCmFP4-STRIX_LEAN (Strix Halo / gfx1151)

Version 1.0 — 2026-08-11

TL;DR

grug-35b-v2 (35B params, 3B active per token, Qwen3.5-VL-MoE family) quantized to `Q4_0_ROCMFP4_STRIX_LEAN` (type 106 preset, ~4.29 BPW). Tuned for AMD Strix Halo (gfx1151 / RDNA 3.5) on the ROCmFPX fork family — we serve and benchmark these files on our lab runtime (full source: pugant/strix-nebulosa; upstream: charlie12345/ROCmFPX). Runs the full vision + text multimodal model in ~17.3 GiB.

⚠️ Critical warnings — read before downloading

—Requires a ROCmFPX fork build — see Runtime. Upstream charlie12345/ROCmFPX loads these files too (any build with the custom GGML types, e.g. built via the kyuz0/amd-strix-halo-toolboxes container). The type 106 (Q4_0_ROCMFP4_STRIX_LEAN) tensor format is INVALID in stock llama.cpp — it will refuse to load. See Usage below.
—Profiled for gfx1151 only (Strix Halo / Ryzen AI Max+ 395, RDNA 3.5). Not tested on other GPUs.
—FP4 here is software on RDNA 3.5 (no FP4 silicon units): the win is bandwidth / memory, not raw compute throughput. The Strix Halo ceiling on this MoE is bandwidth-bound, which is exactly where FP4 helps.

Benchmarks

Tested on Strix Halo (AMD Ryzen AI Max+ 395, 128 GB LPDDR5X). Methodology: llama-bench -ngl 999 -fa on -p 512 -n 128 -mmap 0. All benchmarks for this model were run in ROCm containers (HIP backend). Since then we also benchmarked ROCmFPX quants on the Vulkan (RADV) build of the fork — see the comparison table in Qwen3.6-35B-A3B-MTP-Q6_0_ROCMFPX; those numbers are on a different backend and not directly comparable.

Model	Quant	Size	tg128 (tok/s)	pp512 (tok/s)
grug-35b-v2	ROCmFP4-STRIX_LEAN	17.31 GiB	70.92	1418
grug-35b-v2	Q4KM (baseline)	19.70 GiB	61.18	—
Qwen3.6-35B-A3B (production ref)	ROCmFP4-STRIX_LEAN	17.31 GiB	63	—

Speed-up: +16% vs Q4KM (70.92 vs 61.18 tok/s tg128) at −12% size (17.31 vs 19.70 GiB). +12% vs the production Qwen3.6-35B-A3B reference.

System configuration at bench time

Declared for reproducibility:

—Bare metal host: Bosgame BeyondMax Series, Ubuntu 24.04.4 LTS, kernel 7.0.0-28-generic
—CPU power profile: balanced (powerprofilesctl get) — default, NOT forced to `performance`. Representative of an out-of-the-box setup.
—CPU scaling driver: amd-pstate-epp, scaling_governor performance (amd-pstate-epp default), EPP performance
—IOMMU / iGPU power: auto (no manual tuning)

Note: tok/s above were measured on a non-tuned system (power profile balanced). Users who set powerprofilesctl set performance may see slightly higher numbers.

Quantization details

Preset Q4_0_ROCMFP4_STRIX_LEAN (GGUF file_type 106, ~4.29 bits/weight):

—Attention K/V (blk.*.attn_qkv.weight, blk.*.attn_v.weight) → q4_0_rocmfp4 (high-precision path for attention state)
—Token embeddings (token_embd.weight) → Q5_K (preserve vocab fidelity)
—Expert FFN (blk.*.ffn_*_exps.weight) → q4_0_rocmfp4_fast (max speed path; the bulk of MoE weights)
—Other tensors → F32 / Q40ROCMFP4_FAST as appropriate

Reference fork: `charlie12345/ROCmFPX` commit 00d5452.

Serving runtime: see Runtime.

imatrix methodology

Generated with llama-imatrix (256 chunks, 16 threads, CPU-only). Calibration text from `ProCreations/grug-think-v3-10k` — public Apache-2.0 dataset, not gated: anyone can download it to replicate. Many thanks to the grug team for publishing both the model and a clean calibration set.

—510 entries over 733 tensors
—Warning partial data 99.61% during quantization = 1 of 256 experts not activated in calibration (normal for MoE — see tools/imatrix/imatrix.cpp in llama.cpp). Negligible impact.

Files

File	Size	Description
`grug-35b-v2-ROCmFP4-STRIX_LEAN.gguf`	~17.32 GiB	Main model (type 106)
`mmproj-grug-35b-v2-f16.gguf`	~857 MB	Vision projector (F16)
`imatrix-grug-35b-v2.gguf`	~183 MB	Importance matrix (for re-quantization)

Usage

bash

# Requires the kyuz0 Strix Halo toolbox (which builds charlie12345/ROCmFPX)
docker run --rm -p 1234:1234 --device /dev/kfd --device /dev/dri \
  -v /path/to/models:/models rocmfpx-llm-service \
  llama-server \
    -m /models/grug-35b-v2-ROCmFP4-STRIX_LEAN.gguf \
    --mmproj /models/mmproj-grug-35b-v2-f16.gguf \
    -ngl 999 -fa on --jinja -c 32768 --host 0.0.0.0 --port 1234

Notes:

—MTP not enabled for grug. The mtp_num_hidden_layers field is 0 in this model (MTP was removed during fine-tuning), so it cannot be activated.
—The --mmproj flag is required for the vision tower (multimodal). Without it, text-only still works.

How to replicate

Pipeline described in text only (no published scripts):

1.Build the docker-llm-service-convert image from kyuz0/amd-strix-halo-toolboxes + charlie12345/ROCmFPX (commit 00d5452 or later main HEAD — must contain MODEL_ARCH.QWEN35MOE).
2.Download the BF16 safetensors from ProCreations/grug-35b-v2.
3.Convert to GGUF with convert_hf_to_gguf.py (inside the container).
4.Generate the imatrix with llama-imatrix using ProCreations/grug-think-v3-10k (256 chunks).
5.Quantize: llama-quantize <bf16>.gguf <out>.gguf Q4_0_ROCMFP4_STRIX_LEAN 16.

Attribution & model tree

Qwen3.5-VL-MoE (base architecture)
    └── ornith-ai/Ornith-1.0-35B (MIT)
            └── ProCreations/grug-35b-v2 (Apache-2.0)
                    └── this GGUF (ROCmFP4-STRIX_LEAN)

—Base model: `ProCreations/grug-35b-v2` (Apache-2.0)
—Quantization fork: `charlie12345/ROCmFPX` (MIT)
—Container runtime: `kyuz0/amd-strix-halo-toolboxes`

License

Apache-2.0 (inherited from ProCreations/grug-35b-v2). Derivative work: original model and its license are preserved. See `LICENSE` and `NOTICE`.

Acknowledgements

Built on the shoulders of giants:

—kyuz0/amd-strix-halo-toolboxes — Strix Halo container runtime
—charlie12345/ROCmFPX — llama.cpp fork with ROCmFP4 presets (type 106)
—ProCreations — grug-35b-v2 + the grug-think-v3-10k calibration dataset
—ornith-ai / DeepReinforce Team — Ornith-1.0-35B, the base of grug
—llama.cpp community + Kawrakow (imatrix methodology)
—Hardware: Bosgame BeyondMax Series (Strix Halo bare metal host)

Limitations & community feedback

—Speed benchmark only. No perplexity / MMLU / quality eval is included in this release. The MoE structure is preserved bit-for-bit from the BF16 source except for the quantized tensor formats above; quality is expected to track standard Q4KM-class with the ROCmFP4 attention/K-V choices, but this is not measured here.
—Profiled for gfx1151 only. Not tested on other GPUs (no Navi 3 / Navi 4 / data-center MI series numbers — feel free to share yours).
—MTP not activated (plain inference).

We invite the community — especially fellow Strix Halo owners — to test and share quality results. Open a Discussion on this repo.

Citation

bibtex

@misc{grug35b2026,
  title  = {grug-35b-v2},
  author = {ProCreations},
  year   = {2026},
  url    = {https://huggingface.co/ProCreations/grug-35b-v2}
}

Disclaimer

No affiliation with AMD, Qwen, ProCreations, DeepReinforce, unsloth, kyuz0, or charlie12345. Provided as-is, without warranty. Users must comply with the base model license (Apache-2.0).

Runtime

Benchmark environment: one bare-metal AMD Strix Halo (Ryzen AI MAX+ 395, 128 GB) — full dated configuration and measurement policy: BARE-METAL.md

Requires a ROCmFPX fork build (custom tensor types — stock llama.cpp refuses the file). Recommended: our lab build (pugant/strix-nebulosa, main) — reasoning budget and persistent prompt cache on every model; drafter features where the model ships one: see the engine section of its README.

Plain-inference quant (no MTP head: it cannot be activated); the lab build's reasoning budget and persistent prompt cache still apply.

Everything here is experimental and provided as-is, at your own risk.