MiniMax H3 INT8 Fast (ROCm)
Validated MiniMax H3 prequantized INT8 ConvRot loader for AMD ROCm, preserving H3 mixed-precision boundaries.
Nodes (8)
Your text encoder is eating your VRAM — Load CLIP INT8 shrinks it
Load Dual CLIP INT8 — the Flux text stack, without the VRAM bill
Save Int8 CLIP — quantize your encoder once, stop paying for it on every load
Stack ten LoRAs on your INT8 model — without nuking the speedup
Export your quantized model so you stop re-quantizing it
Bake your LoRA into the quantize step, where it belongs
MiniMax H3 on AMD, without the 42GB bill — Load MiniMax H3 INT8 Fast
The INT8-native loader for ComfyUI
MiniMax H3 INT8 Fast for AMD ROCm
DBJ's validated MiniMax H3 loader for the official prequantized INT8 ConvRot FL2VA and Ref2VA diffusion models on AMD ROCm.
comfy node install minimax-h3-int8-fast-rocm
What this release does
MiniMaxH3INT8FastLoader uses ComfyUI's native MiniMax H3 model loading so the model's BF16, FP16, and FP32 precision-sensitive conditioning, AdaLN, patch, audio, and video output layers remain intact. Only the checkpoint's intended INT8 ConvRot transformer projections use the scoped ROCm fast path.
Validated models:
Validated environment: Linux, ComfyUI 0.32.0, comfy-kitchen 0.2.31, and RX 7900 XTX/gfx1100. Other GPUs and ComfyUI versions are unvalidated.
LoRA
For a quantized H3 base, place ComfyUI's LoraLoaderBypassModelOnly immediately after this loader. ComfyUI-format H3 LoRAs add their residual without rewriting the INT8 base weights.
MiniMaxH3INT8FastLoader
→ LoraLoaderBypassModelOnly (optional)
→ ModelAttentionBackend (comfy kitchen attention)
→ MiniMaxH3FastPatch (optional)
→ MiniMaxH3BlockCacheT8 (optional)
→ MiniMaxH3SigmaShift
→ scheduler / guider / sampler
Upstream Turbo LoRAs: lightx2v/Minimax-h3-Turbo. Match FL2V LoRAs with FL2VA and Ref2V LoRAs with Ref2VA, and use the LoRA's stated 4-step or 8-step schedule rather than this release's standard 20-step validation schedule.
Companion packages
Both accelerators are approximate and opt-in. The INT8 loader is the quality baseline.
Provenance
This release is derived from patientx/ComfyUI-INT8-Fast-ROCM, pinned at f8cff25e7fb1c5f873ccf46bacb06698a89d774e. See H3_COMMUNITY_R01.md for the frozen release details.