water1234/MiniMax-H3-Turbo-v4-step600-EMA-MLX

MiniMax-H3 Turbo v4-step600 EMA — MLX Native Streaming Adapter

BF16 MLX-native conversion of larryvrh/MiniMax-H3-Turbo-Lora, pinned to revision 43a74557ac3f6539db8e0f2a959d03feb7a81480 and source file minimax_h3_turbo_v4_step600_ema.safetensors.

Conversion contract

• Source SHA-256: 5f3a626cd72c93a8b9318d6760c510bc5092d2ab13aaba1f932c5bab07a416d3
• MLX monolithic adapter SHA-256: ed679d4ec8f39ab5c71902f3e13af5bc6a4792bd0317d61c5f6171e40411925e
• 518/518 BF16 tensors consumed
• 259/259 LoRA A/B pairs
• Tensor values are exactly equal before/after conversion
• Attention/MLP rank: 64
• Block/final AdaLN rank: 16
• alpha=rank, runtime multiplier 1.0
• Runtime implementation commit: d5c1e7995e54bbc99cd8b5e7dbca7fb41d119a39

The source already uses canonical fused H3 module paths. Conversion validates every tensor and adds MLX runtime metadata plus a streaming index; it does not quantize, merge or numerically alter the adapter.

Files

• adapter.safetensors.index.json: MLX streaming map
• shards/: 50 main-block, two token-refiner, and one final-layer shard
• conversion_manifest.json: source identity, mapping, ranks and per-shard records
• turbo_lora.safetensors: optional monolithic archival copy (duplicates the shard payload)

For the smallest download, clients need the index, manifest and shards/; the monolithic file is optional.

MLX usage

Use this adapter as a directory. Do not pass a fixed --turbo-lora-alpha; the runtime restores alpha=rank from the native contract.

python scripts/generate.py "A cinematic fox running through a misty forest" \
  --checkpoint /path/to/MiniMax-H3/FL2VA \
  --transformer /path/to/MiniMax-H3-MLX-Argus-Calibrated-INT8 \
  --text-encoder /path/to/your/MLX-text-encoder \
  --turbo-lora /path/to/MiniMax-H3-Turbo-v4-step600-EMA-MLX \
  --turbo-lora-scale 1.0 --steps 9 \
  --low-memory --stream-blocks --output out.mp4

--steps 9 means eight denoiser evaluations. --steps 7 is the faster six-evaluation setting.

Validation

The actual converted adapter passed:

• mixed-rank native-provider strict validation
• per-component shard loading
• legacy Turbo streaming regression
• real calibrated INT8 base + real BF16 Turbo finite refiner computation
• all 50 streamed block AdaLN cache computations
• real block-0 attention/MLP application
• final-layer AdaLN application

The adapter remains separate from the INT8 base so its low-rank update is never rounded into quantized weights.

Attribution and license

Turbo training and original release are by larryvrh. This repository changes packaging/runtime layout only. The source model card declares Apache-2.0 for the adapter. Use also requires a compatible MiniMax-H3 base model, whose separate Community License remains applicable.