Field report: MiniMax H3 full 15s generation on a single R9700 32GB with only 48GB system RAM (Ubuntu / ROCm / Docker) — OOM debugging notes
Sharing a field report from a memory-constrained AMD workstation, in case it saves someone a few days:
Setup
- AMD Radeon AI PRO R9700 32GB (gfx1201), Ryzen 9 9950X, 48GB system RAM
- Ubuntu 26.04.1, Docker, base image
rocm/pytorch:rocm10.1.0_..._pytorch_release_2.14.0, ComfyUI v0.33.2 - MiniMax H3 FL2VA FP8 + Qwen3-VL 32B FP8 encoder (Comfy-Org weights + rzgar encoder), no Turbo, no step reduction
Result
- 608×352, 24 FPS, 362 frames = 15.083s, with 1 native audio track, 20 steps
- 442.26s end-to-end, Docker
OOMKilled=false, verified by decoding the MP4 frame by frame - Container peaks: 34.00GiB RAM (at limit) + 5.71GiB swap; GPU hotspot 98°C
What actually broke, and what fixed it
The GPU was never the problem. The failures were all in host/container memory during the Qwen-encoder → H3-DiT switch:
- 30GiB container OOM on the full workflow, while H3-only loading succeeded — the peak is in the transition, not in either model alone.
- Raising container RAM 34→38GiB made it worse: with
--memory-swaptotal still capped at 42GiB, container swap shrank 8→4GiB. - Two 294-frame attempts were stopped by my own safety monitor when host swap ran out — not model OOMs.
- Fix: +16GiB host SSD swap (≈23.5GiB total), back to the 34GiB container baseline,
--cache-none, and dropping--fast-disk.
Staged runs with full data (39→124→209→294→362 frames), the working Docker/ComfyUI config, and cgroup diagnostic commands are in the repo: https://github.com/wwx0817/minimax-h3-r9700
To be clear about scope: this is not a first R9700+H3 run (Windows and dual-GPU reports exist, and boxwrench/R9700's single-card work predates this), and it's n=1 at 608×352 — an engineering/debugging record, not a benchmark. Model revisions/SHA-256 and the full Dockerfile are still missing; listed as open items in the repo.