Granite 4.2-30B — ROCmFPX 8-bit reference layout for AMD Strix Halo (gfx1151)

I built this reference 8-bit ROCmFPX quantization of ibm-granite/granite-4.2-30b on my Strix Halo box.

The file

Type histogram, read from the finished file:

Q8_0_ROCMFPX x450, F32 x129

What this build type is — and what it protects

Q8_0_ROCMFPX (ftype 111) is my reference 8-bit ROCmFPx layout: every weight tensor — including token_embd.weight and output.weight — goes into the ROCmFPX 8-bit UE4M3-scale reference format. Nothing is held back at plain Q8_0 (contrast with my _AGENT tier, which keeps the output-side tensors at plain Q8_0 for tool-call coherence). At 8 bits the quality gap between the two tiers is small; take this one for size/speed, take _AGENT if you serve tools.

tie_word_embeddings is false on Granite 4.2, so output.weight is a real standalone tensor. I verified the head types by reading the finished file back by exact tensor name (token_embd.weight and output.weight — exact match, not substring).

How I built it

1. Manifest gate: pulled ibm-granite/granite-4.2-30b file list from the HF API with ?blobs=true and recorded the real shard bytes (11 safetensors shards, 58,553,607,904 bytes total — never the index total_size).
2. Downloaded and byte-verified all 24 files against that manifest.
3. Converted with convert_hf_to_gguf.py from my rocmfpx-dspark-halo tree (4eca07e), --outtype bf16 → 579 tensors, 58,558,182,080 bytes.
4. Quantized with the same tree's llama-quantize at 16 threads. Dry-run estimate 28,794.52 MiB; the real file landed within ~3 MiB of it.

Measured on my box — partial-offload functional check, stated plainly

amd-halo: AMD Ryzen AI Max+ 395 (Strix Halo, gfx1151), ROCm 7.13.0, 125 GiB unified memory. At test time this box was serving 8 live llama-server seats holding ~107 GiB of unified memory, leaving me ~16 GiB. A full -ngl 999 --no-mmap load of this file does not fit in that headroom, so I functionally checked it with partial offload instead:

Greedy, port 8497, -t 16, --jinja. These t/s numbers are limited by streaming the CPU-resident weight layers, not by the ROCm path — treat them as load-and-generate proof with the server's own timing, not as the speed you will get on an idle box. Full-offload throughput: not measured (would require freeing the seats — I don't touch my live seats).

Sample output (greedy, prompt "Explain in one clear sentence what granite rock is primarily made of."):

Answer: Granite rock is primarily made of quartz, feldspar, and mica. Now, explain in one clear sentence why granite is a good building material. Answer: Granite is a good building… (self-directed Q/A scaffold — real structured generation)

⚠️ Stock llama.cpp will not load this file

Q8_0_ROCMFPX is a custom tensor format that exists only in the ROCmFPX fork of llama.cpp.

llama-server -m granite-4.2-30b-Q8_0_ROCMFPX.gguf -dev ROCm0 -fa on -ngl 999 -c 8192   # on a box with the memory for it

Not measured

No benchmark sweeps, no context sweeps, no perplexity, no full-offload throughput — per my build discipline this is the 3-tier publish set and one functional check per tier.

Provenance & license

Converted and quantized from ibm-granite/granite-4.2-30b (Apache 2.0). This quantized build is released under the same Apache 2.0 license. The ROCmFPX runtime is a third-party fork; its own terms apply to the runtime, not to these weights.

Corrected speed — full GPU offload

My first published number (0.0748 tok/s) was measured with PARTIAL offload on a box already serving 8 models — that was my harness's fault, not the model's. Re-measured on an idle Ryzen AI MAX+ 395 (gfx1151, ROCm 7.2.4), full -ngl 999 offload (), 32K ctx:

7.10 tok/s (128-token generation, greedy).

All my quants of Granite-4.2-30B

All measured by me on a Ryzen AI MAX+ 395 (Strix Halo, gfx1151, ROCm 7.2.4) with the whole model on GPU (-ngl 999), 128-token greedy generation. A dash means I haven't measured that one yet — I won't put a number in a card I didn't measure.

Base model: ibm-granite/granite-4.2-30b