Modelpublic

kingjones777/Granite-4.2-30B-ROCmFPX-Q8_0-AGENT-GGUF

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes80downloads

Model Card

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

I built this agent-tier 8-bit ROCmFPX quantization of ibm-granite/granite-4.2-30b on my Strix Halo box. Granite 4.2 ships native function calling with thinking/no-thinking modes; the agent tier is the one I would serve for tool work.

The file


ftype	`115` — `Q8_0_ROCMFPX_AGENT`
size	30,709,052,096 bytes (28.60 GiB)
bpw	8.39
architecture	`granite`
tensors	579
context	131,072
token embedding	Q8_0
`output.weight`	Q8_0 (protected)
sha256	`dd46cdeccfe80af5af2bdbc7bfceca8504fd67bc654b8ba2f60389bd2c555e55`

Type histogram, read from the finished file:

Q8_0 x295, Q8_0_ROCMFPX x155, F32 x129

What this build type is — and what it protects

Q8_0_ROCMFPX_AGENT (ftype 115) is my agent/tool-call coherent 8-bit routing. It keeps the LM head, token embeddings and the attention-output projections at full plain Q8_0 (295 tensors here) and routes the rest through the ROCmFPX 8-bit UE4M3-scale layout (155 tensors). My plain tier keeps zero tensors at plain Q8_0 — the routing difference is exactly what the extra ~500 MiB buys.

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 29,283.02 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:


offload	24 / 64 layers on ROCm0, rest CPU-mmap (`-ngl 24 -c 2048`)
generation (server-reported)	0.069 t/s over 48 tokens (692.7 s)
prompt processing	19 tokens in 16.3 s
MemAvailable	16.5 GiB before load → 0.45 GiB after (lowest headroom of the run; 8 seats stayed healthy; 16.5 GiB recovered after the test)

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 minerals. Now, explain in two clear sentences what the process of photosynthesis does. Answer: Photosynthesis converts light… (self-directed Q/A scaffold — real structured generation)

⚠️ Stock llama.cpp will not load this file

Q8_0_ROCMFPX_AGENT is a custom tensor format that exists only in the ROCmFPX fork of llama.cpp.

bash

llama-server -m granite-4.2-30b-Q8_0_ROCMFPX_AGENT.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 (0.0693 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:

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

All my quants of Granite-4.2-30B

build	what it is	size	tok/s (full GPU offload)
`STRIX_LEAN`	my leaner 4-bit tier, Q6_K head — smallest of my 4-bit builds, the one most people want	14.68 GiB	13.99
`COHERENT`	my 4-bit ROCmFP4 tier with the Q6_K-protected head — the balance I run day to day	15.54 GiB	13.03
`Q8_0`	straight 8-bit ROCmFPX — highest fidelity I publish	28.12 GiB	7.10
`Q8_0-AGENT`	8-bit ROCmFPX with the agent-tuned tensor set — for tool-calling work where precision matters	28.60 GiB	6.92

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