K2-Horizon-MoVA-36B-A4B ROCmFP4: dual-expert MoE, 4 tiers measured on gfx1151

README.md

@@ -0,0 +1,120 @@
+---
+license: apache-2.0
+base_model:
+  - IFM/K2-Horizon-MoVA-36B-A4B
+base_model_relation: quantized
+tags:
+  - gguf
+  - rocm
+  - vulkan
+  - rocmfp4
+  - strix-halo
+  - gfx1151
+  - amd
+  - moe
+  - reasoning
+  - k2-horizon
+pipeline_tag: text-generation
+---
+
+# K2-Horizon-MoVA-36B-A4B — ROCmFP4 for AMD Strix Halo (gfx1151)
+
+Four ROCmFP4/ROCmFPX tiers of [`IFM/K2-Horizon-MoVA-36B-A4B`](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B),
+quantised from the official BF16 GGUF and measured on a Ryzen AI Max+ 395 (Radeon 8060S, gfx1151,
+128 GB unified, ROCm 7.2.4). This is the **dual-expert** `k2_horizon` MoVA architecture — 36B total
+parameters, **~4B active** — with two independent expert systems: a 64-way MoVA value-expert bank
+(top-4) alongside a 100-way MoE feed-forward (top-8), 48 layers, hidden 2560, and it is a
+**reasoning** model.
+
+**Every number on this card was measured on these exact files.** Nothing is estimated. As an
+A4B MoE, it decodes far faster than a dense 32B on the same hardware.
+
+## Which file should I use?
+
+| file | ftype | size | decode (HIP) | prefill (HIP) | use it for |
+|---|---|---|---|---|---|
+| `K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_FAST.gguf` | 103 | 18.57 GiB | **41.0 t/s** | **1238 t/s** | **default — fastest on both axes** |
+| `K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_STRIX_LEAN.gguf` | 106 | 18.67 GiB | 41.0 t/s | 1227 t/s | same speed, Q5_K embeddings |
+| `K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_COHERENT.gguf` | 102 | 19.81 GiB | 39.5 t/s | 1114 t/s | most bits in the 4-bit family |
+| `K2-Horizon-MoVA-36B-A4B-Q8_0_ROCMFPX_AGENT.gguf` | 115 | 36.45 GiB | 28.9 t/s | 1183 t/s | 8-bit, agent/tool-call routing |
+
+**Start with `FAST`** — fastest on both axes here, all four answer correctly. Take `AGENT` only if
+you specifically want 8-bit weights.
+
+## Quick start — the backend is one flag
+
+```bash
+# HIP / ROCm — best prefill (long prompts, RAG, agentic)
+llama-server -m K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_FAST.gguf \
+  -dev ROCm0 -fa on -ngl 999 --no-mmap -fit off -np 1 \
+  -b 2048 -ub 1024 -t 16 --poll 100 -c 8192 --jinja
+
+# Vulkan — SAME FILE. Only -dev changes.
+llama-server -m K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 ...
+```
+
+### ⛔ You need a build with BOTH `k2_horizon` and the ROCmFP4 types
+
+These declare `general.architecture = k2-horizon` **and** use the ROCmFP4/ROCmFPX tensor types
+(ggml 100-119). No single public build has both yet:
+
+- **`k2_horizon` arch** (incl. the MoVA dual-expert graph) — [MBZUAI-IFM llama.cpp fork](https://github.com/MBZUAI-IFM/llama.cpp/tree/model/K2Horizon); upstream PR in progress.
+- **ROCmFP4/ROCmFPX types** — [`ROCmFPX`](https://github.com/ROCmFPX/ROCmFPX).
+
+This repo ships **`k2-horizon-on-rocmfpx.patch`** — the arch applied onto a ROCmFPX base, the exact
+tree these files were built and verified against:
+
+```bash
+git clone https://github.com/ROCmFPX/ROCmFPX.git && cd ROCmFPX
+curl -fLO https://huggingface.co/kingjones777/K2-Horizon-MoVA-36B-A4B-ROCmFP4-GGUF/resolve/main/k2-horizon-on-rocmfpx.patch
+git apply k2-horizon-on-rocmfpx.patch
+cmake -B build -DGGML_HIP=ON -DGGML_VULKAN=ON -DAMDGPU_TARGETS=gfx1151 -DGGML_NATIVE=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j$(nproc)
+```
+
+## ⛔ Serving note — this is a reasoner
+
+MoVA-36B emits chain-of-thought into `reasoning_content` before answering. A short budget returns
+**HTTP 200 with empty `content` and the tokens in `reasoning`** — the model thinking, not a broken
+file. Give it room: ~500 tokens for short answers and tool calls, more for long-form. Verified: at
+40 tokens `content` was empty; at 500 it answered correctly.
+
+## Head handling
+
+Quantised at each tier's native routing with no forced `--output-tensor-type` override; the tiers
+are posted as-built.
+
+## Reproduction block
+
+```
+model     : IFM/K2-Horizon-MoVA-36B-A4B  (official BF16 GGUF, K2-Horizon-36B-BF16.gguf, 74.9 GB)
+            arch k2-horizon MoVA (dual-expert MoE): 36B total / ~4B active,
+            mova_num_experts 64 top-4 + num_experts 100 top-8, 48 layers, hidden 2560
+port      : MBZUAI-IFM/llama.cpp @ model/K2Horizon (35999d1) k2_horizon arch
+            applied onto ROCmFPX fork @ 85d8f7e  (patch shipped in this repo)
+quantize  : llama-quantize  <in> <out> <ftype>   (no head-type override)
+            Q4_0_ROCMFP4_COHERENT (102) · Q4_0_ROCMFP4_FAST (103)
+            Q4_0_ROCMFP4_STRIX_LEAN (106) · Q8_0_ROCMFPX_AGENT (115)
+serve     : llama-server -dev {ROCm0|Vulkan0} -fa on -ngl 999 --no-mmap -fit off
+            -np 1 -b 2048 -ub 1024 -t 16 --poll 100 -c 8192 --jinja
+box       : Ryzen AI Max+ 395 / Radeon 8060S (gfx1151) / 128 GB unified, ROCm 7.2.4
+verified  : dual-expert graph loads on GPU, generates correct output
+            ("capital of France" -> "Paris"; "capital of Japan" -> "Tokyo")
+measured  : 2026-09-04
+```
+
+## SHA256
+
+```
+0f11adaada7b8b9f7aa785561f412d2de85a555fc1e2ffeddf71f972b8d576e4  K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_COHERENT.gguf
+5618a8e20c3300dcb38ec7217c612edcd8fb091b23aa9e024ca7063b235e80bf  K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_FAST.gguf
+7fea90c3aa739652d764416741e85f563ca355ab29c8a8eb89e753a37df8e79f  K2-Horizon-MoVA-36B-A4B-Q4_0_ROCMFP4_STRIX_LEAN.gguf
+e653f3855ed2c0c347e14e22593d49ced9118bf2115a7717ed2e34e5e6ff722c  K2-Horizon-MoVA-36B-A4B-Q8_0_ROCMFPX_AGENT.gguf
+```
+
+## Credits
+
+**[IFM / MBZUAI](https://huggingface.co/IFM)** — the K2-Horizon-MoVA-36B-A4B base model
+(Apache-2.0) and the `k2_horizon` llama.cpp architecture. ROCmFP4/ROCmFPX quantisation types are the
+work of **[ROCmFPX](https://github.com/ROCmFPX/ROCmFPX)**. This repository adds only the ROCmFP4
+quant ladder, the arch-on-ROCmFPX patch, and the measurements.