kingjones777/Granite-4.1-8B-ROCmFP4-GGUF

⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL

🚀 39.90 tok/s — ~13% faster than Q4_K_M, ranges disjoint by a wide margin.

Granite-4.1-8B — ROCmFP4 (tier 102 COHERENT) GGUF

A 4-bit ROCmFP4 quantization for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo), quantized from BF16 GGUF — a lossless source, not a requantization of a lower-bit build.

File	granite-4.1-8b-Q4_0_ROCMFP4_COHERENT.gguf
Size	5.162 GB
BPW	4.69
ftype	Q4_0_ROCMFP4_COHERENT (102)

⛔ Requires a llama.cpp with the ROCmFP4 quant types

Q4_0_ROCMFP4_COHERENT (ftype 102) exists only in charlie12345/ROCmFPX, not upstream llama.cpp. Ignore the auto-generated "Use this model" commands above.

---

Measured

Ryzen AI MAX+ 395 (gfx1151, 128 GB unified, ROCm 7.2.4). Median of 3+, warm-up discarded, otherwise-idle box. Correctness at the model's official sampling.

build	size	decode (median)	range
this build	5.162 GB	39.90	[39.87 – 39.93]
Q4_K_M	5.348 GB	35.27	[35.22 – 35.28]

~13% faster, ranges disjoint (39.87 min vs 35.28 max). Also 3.5% smaller.

Correctness: 17×23 ⇒ ✅ 391 · capital of Japan ⇒ ✅ Tokyo · days in 2024 ⇒ ✅ 366

Per-tensor types (audited in the finished file)

token_embd Q6_K · norms F32 · bulk Q4_0_ROCMFP4 (type 100)

granite is dense — no SSM/MoE state to protect.

---

What was NOT measured

• No perplexity run, and no quality A/B against the baseline or the source. The checks above are memorized-fact prompts — necessary but not sufficient; a damaged model can pass them.
• No long-context testing.
• No tool-calling evaluation.

Base model licence inherited; credit for the model goes to its authors.