Optimizing MiniMax H3 on AMD GPUs: Benchmarks and Pitfalls
frq2000 · reddit · 2026-08-07
The author shared detailed optimization experiences for running the MiniMax H3 video generation model locally on an AMD Radeon RX 7900 XT (20GB).
- Baseline: Using ComfyUI with INT8/INT4 quantized models, generating a 10-second 1056×608 video takes 58 minutes.
- Working Optimizations: Enabled AOTriton Flash Attention, MIOpen cache, disabled mmap (fixes an extreme slowdown loading safetensors on AMD), and configured layer-wise VRAM loading.
- Rejected Solutions: SageAttention v1 actually slows down performance on RDNA3; Sol-Attn node causes hangs due to incompatibility; DynamicVRAM fails due to a known ROCm bug.
- To be tested: Listed several ComfyUI nodes and Triton kernels for potential further speedups, asking the community for more experiences.
Related event: Running MiniMax H3 on AMD GPUs: Tests and Optimization(2 posts)→
More from Infra
- Amazon Proposes Per-Request Validated Evidence Caching for Production RAG — _reachsumit · 2026-10-16
- LinkedIn's TransX report: sequence transduction for recsys, +6.02% CTR in A/B tests — _reachsumit · 2026-10-16
- llama.cpp PR boosts CPU inference speed with merged AVX2 tiled accumulators — jacek2023 · 2026-10-16
- Zero's next-gen IVM engine runs in a single DO, live-querying a 100GiB R2 database — threepointone · 2026-10-16
- Meme: driving to the chip factory in 2028 captures AI's compute anxiety — beffjezos · 2026-10-16
- One volunteer maintains the time zone file behind 4 billion phones — danbri · 2026-10-16