MiniMax-Music-3-Turbo-FP8

⚡ MiniMax Music 3 Turbo FP8 & 8-Step LoRA (ComfyUI Optimized)

⚠️ Notice (v1.0 Community Release - Work in Progress / Experimental Calibration):
This is a v1.0 calibration release created to accelerate local generation on NVIDIA RTX GPUs. Further refinements and community feedback are welcome as we calibrate audio hyperparameters.

---

🚀 Overview & Benchmark Highlights

This repository contains the ultra-optimized FP8 (torch.float8_e4m3fn) weights and the Turbo 8-Step Step-Distillation LoRA for MiniMax Music 3, tailored specifically for lightning-fast inference in ComfyUI.

---

📦 Downloadable Files in this Repo

---

🎛️ Recommended ComfyUI Workflow & Settings

1. Model Connections

[Load Diffusion Model] (minimax_music3_dit_fp8_e4m3fn.safetensors)
       │
       ▼
[Load LoRA] (minimax_music3_turbo_lora_8step.safetensors | strength: 0.85 ⭐ IDEAL)
       │
       ▼
[Patch Sage Attention KJ] (sage_attention: auto)
       │
       ▼
[KSampler]

💡 Pro-Tip on LoRA Strength (0.85 is Ideal)

⭐ LoRA Strength = 0.85 (Recommended Sweet Spot):
Setting the LoRA strength to 0.85 (rather than 1.0) provides the absolute best balance: it accelerates the diffusion sampling to 8-10 steps while letting the base model inject 100% crystal-clear vocal formants, diction, and delicate high-frequency percussion.

---

2. Optimal Node Parameters

A. MiniMax Music3 Text Encode

• clip : Connect to Load Text Encoder with minimax_music3_text_encoder_fp8_e4m3fn.safetensors.
• cfg_scale : 1.5 (Recommended: keeps 100% voice clarity, instrument separation and lyrical articulation).
• max_duration : Adjust to your desired song length (e.g. 120.0 for 2 min, 190.0 for 3 min 10s).

B. KSampler

• steps : 8 (Ultra-fast) or 10 - 12 (Studio Master Quality)
• sampler_name : euler
• scheduler : simple
• cfg : 1.7
• denoise : 1.0

---

🔬 Technical Details & Methodology

1. Text Encoder Quantization: Converted from the pristine unrotated minimax_music3_text_encoder_pruned_bf16.safetensors directly to float8_e4m3fn. All 166 sensitive layers (audio decoder heads, RMSNorms, token embeddings) remain in high-precision BF16 to guarantee zero early-stop anomalies and pitch accuracy.
2. DiT Flow-Matching Quantization: Converted 150 2D matrix weights to FP8 while keeping 224 modulation/norm parameters in FP16.
3. Turbo LoRA Distillation: Trained using Consistency Flow-Matching & Mean Flow Distillation across 144 attention and MLP projection matrices over 20,000 steps with Cosine Warmup scheduling on an NVIDIA RTX 4090.

---

🤝 GitHub Repository

Training scripts and conversion tools are open-sourced at:
👉 GitHub: Guillaume-127/Minimax-music-3-Turbo-8-steps

---

Created by Guillaume-127. Released for the open-source AI music community.