Getting started

Getting started (HIP)

One source tree. Three supported HIP offload ISAs. You pass HIP_ARCH yourself; the Makefile does not probe the GPU.

HIP_ARCH	GPU	Runtime DiT	Runtime SDPA
gfx1151	Strix Halo (RDNA, wave32)	INT8 weights (hipBLAS), BF16 activations	rocWMMA
gfx90a	MI210 / MI250X (CDNA2, wave64)	BF16 weights (hipBLAS BF16 GEMM)	MFMA flash (BF16 QK, FP16 PV, FP32 accum)
gfx942	MI300X (CDNA3, wave64)	same as gfx90a	same as gfx90a

Tagged v0.10.1 on main covers the dual-ISA line; main after that adds gfx942 plus VAE/INT8/sampler knobs. gfx1151 fox-s2 v0.9.0 md5 1731f95c4aa582597cf83d57f46b8f9e is the historical gate. On current main the default 512 px VAE tile changes fox-s2 to 34507f072c5cabbde6592b3f70b8fa35.

Build

git clone https://github.com/alexhegit/h3-hip.c.git
cd h3-hip.c
git checkout v0.10.1

# Strix Halo
make HIP_ARCH=gfx1151 -j$(nproc) h3

# MI210
make HIP_ARCH=gfx90a -j$(nproc) h3

# MI300X
make HIP_ARCH=gfx942 -j$(nproc) h3

./h3 --info -d /path/to/MiniMax-H3

make clean does not need HIP_ARCH. After switching ISA, make clean then rebuild. A missing HIP_ARCH on any non-clean target is a hard error.

Halo fox-s2 md5 gate (v0.9.0 hash; override H3_FOX_S2_MD5 on current main):

make HIP_ARCH=gfx1151 halo-regression

MI210 unit + smokes:

make HIP_ARCH=gfx90a hip-test

Set H3_MODEL if weights are not at the Makefile default (/home/amd/HF-MODELS/MiniMax-H3).

Weights

Official MiniMax-H3. T2VA uses text_encoder/, transformer/, vae/, and the audio decoder. See T2VA pipeline for dtypes (checkpoint BF16/F32; INT8 is runtime-only on gfx1151).

First T2VA

MODEL=/path/to/MiniMax-H3

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

Three fox presets (all complete MP4s, not stubs):

Name	Knobs	What you get
fox-s2	--steps 2 --layers 35 --reuse 1	HIP A/B; gfx1151 md5 gate
fox-fast	--steps 20 --layers 45 --reuse 2	~0.9 s complete clip; 11 DiT evals
fox showcase	--steps 20 --layers 50 --reuse 1	README / wiki gallery fox

For a ~15 s clip, keep fox-fast knobs and raise --seconds 15 at 864×480 (Long video). Timings: docs/PERFORMANCE.md.

Optional --token-reduction (off by default, same flag as h3-spark.c): pairs video tokens in middle DiT blocks. Faster long T2VA; visible quality loss; fox-s2 md5 gate does not apply. Generate prints a warning when on.

On a multi-GPU box bind one card: H3_HIP_DEVICE=N (default 0). Do not run two weight-streaming T2VA jobs against the same NVMe at once.

Optional knobs

H3_INT8_MLP=1          # gfx90a/gfx942: INT8 DiT (gfx1151 default)
H3_INT8_MLP=0          # gfx1151: keep BF16 DiT weights
H3_GPU_SAMPLER=1       # GPU Euler sampler (opt-in on HIP)
H3_TOKEN_REDUCTION=1   # same as --token-reduction
H3_INT8_VAE=1          # INT8 Video VAE weights (VRAM; fox VAE wall may not fall)
H3_VAE_TILE_PIXELS=272 # restore v0.9.0-sized VAE tiles
H3_SDPA_CDNA_FLASH=0   # gfx90a: hipBLAS score-matrix SDPA fallback
H3_SDPA_CDNA_FP16_PV=0 # gfx90a flash: BF16 PV instead of FP16