v0.1.4 - ComfyUI compatibility fix

MiniMax H3 V100 Custom Node v0.1.4

This release fixes compatibility with the updated ComfyUI MiniMax H3 FinalLayer API.

Fixed

• Fixes _patched_final_forward() takes 5 positional arguments but 8 were given.
• Preserves the upstream sigma, sample_sigmas, and shifts arguments and the native PDD output-head selection logic.
• Keeps the FinalLayer hidden state and time embedding on the FP32 safety path.
• Supports the newer optional DiT block attention override.
• Fails closed before FP16 registration if a reimplemented upstream method gains an unknown parameter.

Package scope

The downloadable ZIP is the standalone L3 Custom Node. TE adapters, TE-Speed, and the legacy source patcher are excluded. Install only one MiniMax H3 precision runtime extension.

Validation

• 23 dependency-free tests passed.
• Pinned ComfyUI v0.34.5 and PDD source fixtures passed structure, argument-routing, schedule, and dtype-flow checks.
• Manifest, ZIP contents, and SHA-256 passed verification.
• V100 inference, video/audio quality, speed, and memory remain pending server testing.

SHA-256: 6cdaeeba5e3f4b0cbf62b0a83e5e27ff3853ce6f6ecee656abc8f1b1505e974a

Stop ComfyUI, remove the previous H3 extension folder, extract the ZIP into custom_nodes, and restart. Confirm the startup log reports v0.1.4 runtime profile installed.

v0.1.2 - MiniMax H3 V100 Custom Node

MiniMax H3 V100 Custom Node v0.1.2

Version 0.1.2 makes the ComfyUI Custom Node the recommended installation method for the existing audio-safe V100 mixed-precision profile.

Highlights

• no longer modifies comfy/ldm/minimax/model.py
• requires no --fp16-unet or other launch-argument change
• supports the updated MiniMax H3 audio carry/sampler structure in ComfyUI 0.31.1
• retains FP16 resident QKV/QKV compute and text/video attention with FP32 audio attention and FP32-sensitive paths
• preserves the previous guarded source patchers, restore scripts, and origin/TE source snapshots under legacy_patcher/
• adds structural checks, conflict refusal, fail-closed audio handling, version reporting, tests, and bilingual migration documentation

The Custom Node delivery was reported working on the target V100 with ComfyUI 0.31.1.

Install

1. Restore a clean comfy/ldm/minimax/model.py if an older source patcher was used.
2. Extract the repository folder into ComfyUI/custom_nodes/minimax-h3-v100-patch/.
3. Start ComfyUI normally without --fp16-unet or --force-fp16.

See the English and Chinese READMEs for migration, uninstall, compatibility, and legacy-patcher instructions.

Validation

• Python AST parsing: 8 files passed
• simulated Custom Node loader tests: 6/6 passed
• root manifest: 20/20 entries verified
• legacy manifest: 11/11 entries verified
• legacy origin and origin-to-TE patch/revert round trips passed

Release asset SHA-256:

f468ba78a6ddc6f3e44944f599d334d8866f981a7e80b83030012ae230798657

Special thanks to Amduraznak/minimax-h3-fp16-fix for demonstrating the Custom Node delivery approach that inspired this release. The FP32-base precision strategy remains this project's existing independently tested profile.

Full change: PR #4

v0.1.1 - MiniMax H3 V100 mixed-precision patch

Documentation release with English and Simplified Chinese READMEs. Benchmark context: Windows 10, NVIDIA Tesla V100 32 GB, 150 W power limit; the 317 s baseline includes TE-Speed. The tested patch completed the 0.2 MP, 5 s, 24 fps workload in 206 s. Performance at 300 W is expected to improve but has not been benchmarked. ZIP SHA-256: 3f8fdf7fdfea89f3cadedd5521b80b045fcf5d35f39ef1946c9c8f47ba7be0dc.

v0.1.0 - MiniMax H3 V100 Plan 2

MiniMax H3 V100 mixed-precision patch

Community patch with separate patch and restore launchers for:

• Official/origin MiniMax H3 model.py
• TE-Speed-hooked MiniMax H3 model.py

Tested precision profile

• FP16: QKV projection and optimized attention
• FP32: Q/K RMSNorm, RoPE, attention output projection, MLP, AdaLN and residual accumulation

V100 benchmark

Test setting reported by the tester: 0.2 MP, 5 s, 24 fps.

• Baseline: 317 s
• Attention-only FP16: 260 s
• Plan 2: 206 s
• Measured speedup: 1.54x

Performance depends on the model build, workflow and attention backend.

Safety

• Automatic, variant-specific backup
• TE/origin target detection and wrong-target refusal
• Idempotent patching
• Python AST and SHA-256 verification
• Dedicated TE and origin restore BAT files

Package checksum

minimax_h3_v100_patchers.zip

SHA-256: f626915d8f2c0aaeb626ec72b8c7bdfa9f6f387f141d6d475c03cd6ecb647673