Phi-4-Reasoning-Vision-15B-ROCmFP4-GGUF

⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL

⚠️ -fa off is required — flash attention breaks the vision path on gfx1151.

🚀 24.62 tok/s — +9.4% faster and 0.5 GiB smaller than Q4_K_M, ranges disjoint.

Phi-4-Reasoning-Vision-15B — ROCmFP4 (tier 102 COHERENT) GGUF

A 4-bit ROCmFP4 quantization for AMD gfx1151 (Ryzen AI MAX+ 395 / Strix Halo), quantized from F16 GGUF (27.96 GiB) — a lossless source, not a requantization of a lower-bit build.

⛔ 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.

---

All quant variants

Three builds, all measured in one session on one box with one binary (Ryzen AI MAX+ 395, gfx1151, ROCm 7.2.4, ROCmFPX-2809dc5, -fa off) — so these rows are directly comparable. Median of 3, warm-up discarded, otherwise-idle box.

⚠️ The 4-bit build is ~1.7× faster and 43% smaller. The 8-bit builds exist for accuracy headroom, not throughput. The two 8-bit builds are within noise of each other (14.08 vs 14.15, ranges touching) — this model has no MTP draft head, and AGENT's benefit shows up in draft acceptance, so there is nothing here for it to win.

⚠️ -fa off is mandatory — flash attention breaks the vision path on gfx1151. The BF16 mmproj (862 MiB) ships in every one of these repos and is required for vision.

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.

+9.4%, ranges disjoint. Vision verified: an 8×8 red PNG is correctly described as Red.

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

Per-tensor types (audited in the finished file)

output.weight Q6_K · token_embd Q6_K · norms F32 · bulk TYPE_100 (160) · vision tower kept at BF16 (mmproj, 823 MiB — included)

Serve with -fa off and the bundled BF16 mmproj.

---

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.