Home

h3-hip.c wiki

HIP port of antirez/h3.c for AMD GPUs (gfx1151 / Strix Halo). The GPU backend is pure HIP so the same MiniMax-H3 CLI and model stack run on ROCm.

Project ident — 864×480, 56 frames, --steps 20 --layers 50 --reuse 1. Click for the MP4.

Project page: alexhegit.github.io/h3-hip.c — speedup ladder, clip gallery, caveats
Code: alexhegit/h3-hip.c
Weights: MiniMaxAI/MiniMax-H3
Upstream: antirez/h3.c · CUDA sibling: alexhegit/h3-spark.c

Current tagged line: v0.9.0.

Clips generated on gfx1151

Every clip below came out of this HIP build on a Strix Halo iGPU. Click a poster for the MP4. H3 cannot render readable captions, so any on-screen title was added with ffmpeg after generate.

T2VA fox — 512² · 22f · --layers 50 --reuse 1	Linux 35 tribute — 864×480 · 56f · untitled
Ident draft — 512×288 · 56f · untitled	AMD developer community — Ref2VA · 864×480 · untitled

Knobs and full prompts: Showcase.

Where it stands (v0.9.0)

Preset	Knobs	E2E	Denoise GPU
fox-s2	512² · 22f · --steps 2 --layers 35 --reuse 1	83–87 s	6.3–6.5 s
fox-fast	512² · 22f · --steps 20 --layers 45 --reuse 2	95 s	28 s (11 evals)

Remaining end-to-end time is mostly NVMe weight I/O, not GPU. Details: Performance.

Try it

git clone https://github.com/alexhegit/h3-hip.c.git
cd h3-hip.c
make -j$(nproc) h3

./h3 --profile -d /path/to/MiniMax-H3 \
  -p "A red fox walks through fresh snow." \
  --width 512 --height 512 --frames 22 \
  --steps 20 --layers 50 --reuse 1 \
  -o outputs/fox.mp4

Full setup: Getting started.

Pages

• Getting started — build, generate, tests
• Showcase — gfx1151 clips and reproduce commands
• Performance — fox-s2 / fox-fast scoreboard
• Known issues — HIP gaps that do not block default T2VA
• CLI — flags that matter on this port

README stays the short front door. This wiki is the longer user guide. Engineering KEEP/REJECT logs stay in-tree under docs/perf/ and are not mirrored here.