MiniMax H3 Troubleshooting: OOM, Slow Steps & Errors

Fast diagnosis

Start with the symptom, not random flags

Symptom	Likely area	First check	Guide section
OOM before sampling	Text encoder or model load	Encoder precision, free VRAM, model handoff	Model handoff
Each step becomes slower	DynamicVRAM, RAM leak, offload thrashing	RAM/swap growth and per-step log times	Runaway step time
Whole machine freezes	Pinned memory or exhausted RAM/swap	System memory and pinned-memory flags	System freeze
Model not found	Wrong folder or stale loader list	File path and exact loader filename	Models and nodes
`hipMemAddressReserve` error	AMD DynamicVRAM/ROCm interaction	ROCm, Torch, Triton and launch flags	AMD and ROCm

Symptom 1

Per-step time keeps increasing

An RTX 4070 Ti 12GB, 64GB RAM, Windows and ComfyUI 0.30.1 community report described a GGUF run starting around 18 seconds per step and eventually exceeding 400 seconds per step. The author associated the behavior with DynamicVRAM and a GGUF RAM leak.

Community-reported test flags

--disable-dynamic-vram --disable-pinned-memory

Before testing, save the original launch command. Compare per-step time, system RAM and swap with only this change, then revert if performance gets worse.

Symptom 2

MiniMax H3 freezes the whole PC

One Ubuntu report linked a full-system lock to pinned-memory pressure rather than a simple CUDA OOM. The author's stable launch included the following flags:

--disable-pinned-memory
--disable-smart-memory
--cache-none
--fast-disk

Do not paste the full set blindly. Test pinned memory first, monitor RAM and swap, and keep remote access or a local recovery path available when reproducing a hard-lock report.

Symptom 3

OOM when the diffusion model starts

A 16GB RTX 5070 Ti report found that the text encoder remained allocated when the diffusion model loaded. The author inserted a KJNodes `VRAM_Debug` barrier and forced cache and garbage collection before the handoff.

Confirm whether the OOM appears after text encoding but before the first diffusion step.
Log free VRAM immediately before both model loads.
Test an explicit unload or cleanup barrier at the handoff.
Keep the workflow copy so the cleanup node can be removed if it causes a new error.

Symptom 4

Model not found or missing nodes

Diffusion model

ComfyUI/models/diffusion_models/

Place the MiniMax H3 diffusion file here, then select its exact installed filename in the workflow loader.

Text encoder

ComfyUI/models/text_encoders/

Qwen3-VL encoder variants belong here. A workflow may reference a filename you did not download.

Video and audio VAE

ComfyUI/models/vae/

Keep video and audio VAE files distinct and match the nodes used by the chosen workflow.

Missing custom node

Open ComfyUI Manager, install the repository named by the missing node, restart ComfyUI, and reload the JSON. Do not install unrelated node packs at random.

See the complete ComfyUI setup and workflow guide.

Symptom 5

AMD `hipMemAddressReserve` or slower attention

An RX 7900 XT author reported DynamicVRAM failing with `hipMemAddressReserve` and SageAttention v1 running slower, while another participant reported DynamicVRAM working with different flags and software. This is a real community conflict, not a settled fix.

Record the exact ROCm, Torch, Triton and ComfyUI versions.
Start with the author's known-running AOTriton path before adding attention replacements.
A/B test DynamicVRAM with the same input, seed, frames and model files.
Watch system RAM; some reports mention more than 40GB usage under offload.

Read the AMD discussion and its conflicting results.

Clean debugging loop

Make every change reversible

1

Save baseline

Copy the workflow and launch command before changing anything.

2

Locate stage

Record the last completed node or log line and current RAM/VRAM.

3

Change one thing

One flag, model or workflow node per test keeps the result interpretable.

4

Keep evidence

Save logs and undo the change if it does not improve the exact symptom.

FAQ

Common MiniMax H3 errors

Why does MiniMax H3 get slower every step?

One RTX 4070 Ti 12GB community report described per-step time rising from about 18 seconds to more than 400 seconds with a GGUF workflow. The author associated it with DynamicVRAM and a RAM leak and reported improvement after disabling DynamicVRAM and pinned memory. Treat this as an environment-specific test, not a universal fix.

Why does MiniMax H3 freeze the whole PC?

A community Ubuntu 24.04 report linked a full-system lock to pinned-memory pressure. The author stabilized that setup with flags disabling pinned and smart memory plus cache changes. Save your current launch command and test one flag at a time because the same flags may reduce performance elsewhere.

How do I fix model not found in MiniMax H3 ComfyUI?

Confirm that diffusion models are in models/diffusion_models, text encoders are in models/text_encoders, and video or audio VAEs are in models/vae. Then select the installed filenames in the loader nodes and restart ComfyUI if the file list is stale.