Rreitsma/minimax-music3-ensemble-blocks

MiniMax Music 3 — Ensemble Blocks

Render K variations of one prompt in a single batched pass — the marginal variation is nearly free.

The MiniMax Music 3 autoregressive stage reads the full 8B language model plus seven 0.6B depth-decoder passes (~23GB of weights) for every audio frame at 25fps: it is memory-bandwidth bound, and that read costs the same whether it serves one variation or four. These blocks decode K variations in lockstep (batch 2K rows), sharing the dominant cost. The flow-matching stage batches across variations too (grouped by exact frame count — no padding).

Measured with THIS block on an RTX 4090 (24GB, bf16, diffusers main, warm, 15s songs, same prompt/seeds):

variations	total	per variation
1	28.9s	28.9s
2	34.9s	17.5s (1.65x)
3	41.5s	13.8s (2.1x)

Each additional variation costs ~6s on a ~29s base — the marginal take is ~21% of a solo render. Discussion: diffusers#14486.

Usage

import torch
from diffusers.modular_pipelines import ModularPipelineBlocks, ComponentsManager

blocks = ModularPipelineBlocks.from_pretrained(
    "Rreitsma/minimax-music3-ensemble-blocks", trust_remote_code=True
)
manager = ComponentsManager()
manager.enable_auto_cpu_offload(device="cuda")
pipe = blocks.init_pipeline("MiniMaxAI/MiniMax-Music3", components_manager=manager)
pipe.load_components(dtype=torch.bfloat16)

out = pipe(
    prompt="Genre: acoustic pop. BPM: 96. Key: C major. Warm female vocals, fingerpicked guitar.",
    lyrics="[verse]\nMorning light filtering through the pine\n[chorus]\nSoftly the world begins to breathe",
    audio_duration=60.0,
    num_variations=3,
    seed=7,            # variation i uses seed + i; omit for random seeds
    output="audios",
)
# out: list of float32 stereo waveforms, one per variation, 44.1kHz, (channels, samples)

Notes & honest caveats

• Quality: every variation's math is row-independent and draws from its own seeded generator using the reference sampling recipe — identical in distribution to solo generation. Batched kernels differ from solo kernels at the floating-point-ulp level, so a given seed may take a different (equally valid) trajectory than it would solo.
• Guidance is standard CFG with the checkpoint's guidance scale. For other guidance techniques, use the default MiniMax Music 3 blocks (with the guider) instead.
• VRAM scales with duration x num_variations (KV cache). On 24GB: 3 variations up to ~1 minute is comfortable; use fewer variations for longer songs. An early EOS in one variation freezes its rows at no cost.
• Deliberately plain eager PyTorch — no torch.compile, no extra dependencies — so it runs wherever diffusers runs. A further-optimized local studio (compiled AR decode, ~2.9x total on a 4090) lives at minimax-music3-studio.

License & credits

Apache-2.0 (portions derived from the diffusers MiniMax Music 3 modular pipeline, Copyright 2026 The MiniMax Team and The HuggingFace Team). Model weights are MiniMax's, CC BY 4.0, fetched separately from MiniMaxAI/MiniMax-Music3.

Built with Claude (Fable 5 Max) following a request from the diffusers maintainers in #14486.