Laguna-S-2.1-Chadrock-ROCmFP4-StrixKVSpine-V4-GGUF
High-performance tuned Laguna S 2.1 Chadrock ROCmFP4 quant with custom ROCmFP4 tensor types + Laguna architecture support, optimized for AMD Strix Halo / Ryzen AI Max+ 395 systems.
⚠️ Important Engine Requirement
[!CAUTION] This quant uses custom ROCmFP4 tensor formats + Laguna architecture support. It does NOT load in stock upstream
llama.cpp. You MUST use the specialized Ciru ROCmFPX Runtime V3.
- Repository: github.com/ciru-ai/ROCmFPX
- Branch:
agent/laguna-s21-runtime-v3 - Pinned Commit:
54f5fe06c74350fb8b6aec21d8749071bc195bdb
🖥️ System Requirements & Specs
- Target Hardware: AMD Ryzen AI Max+ 395 / Radeon 8060S (Strix Halo
gfx1151) with 128 GB Unified Memory. - Backend: Vulkan (
Vulkan0via Mesa RADV driver). - Max Stable Context: 131,072 tokens (131K).
- Model File:
laguna-s-2.1-ROCmFP4-StrixKVSpine-v4.gguf(65.4 GB). - KV Cache Precision: F16 / F16.
⚡ How to Run (Validated Strix Halo Profile)
1. Clone & Pin Runtime
git clone --branch agent/laguna-s21-runtime-v3 --depth 1 https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout --detach 54f5fe06c74350fb8b6aec21d8749071bc195bdb
2. Install Vulkan Deps & Build
scripts/install-laguna-vulkan-deps.sh --install
JOBS=8 BUILD_TYPE=Release scripts/build-laguna-strix-vulkan.sh
3. Launch Supervised Endpoint (Recommended)
HOST=0.0.0.0 PORT=8089 scripts/run-laguna-vulkan-supervised.sh /path/to/laguna-s-2.1-ROCmFP4-StrixKVSpine-v4.gguf
This applies driver preflight, DeviceLost safeguards, safe 131K context profile, Flash Attention, and F16 KV cache.
🌐 Open WebUI / PrivChat Integration
Connect Open WebUI (Settings → Connections → OpenAI API):
- OpenAI API Base URL:
- Docker:
http://host.docker.internal:8089/v1orhttp://172.18.0.1:8089/v1 - Host / Browser:
http://127.0.0.1:8089/v1
- Docker:
- API Key:
sk-no-key-required(or any string) - Model Name:
laguna-s21-rocmfp4-strixkvspine-v4
📊 Benchmark Summary
Tested on AMD Ryzen AI Max+ 395 / Radeon 8060S with 128 GB LPDDR5X-8533:
| Metric | Result |
|---|---|
| Generation Speed (Decode) | 33.1 – 38.3 t/s |
| Prompt Eval Speed (Prefill) | 79.3 – 92.7 t/s |
| Time to First Token (TTFT) | ~560 – 618 ms |
| Validated Max Context | 131,072 tokens (131K) |
| Default Thinking | Off |
🧾 Provenance & Integrity
- Source: poolside Laguna-S-2.1, converted and quantized with the ROCmFPX toolchain (
Q4_0_ROCMFP4_StrixKVSpine-v4preset) - SHA256:
ea1d854a72c47ec8e72c16ea91b8ff3cd5e1620b834df175f683c86f27dc26d6 - Tuned serving profile ships in halofpx:
halofpx pull laguna-s21 && halofpx serve -m laguna-s21(Vulkan0 backend, 131K context)