Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
350 lines
11 KiB
JSON
350 lines
11 KiB
JSON
{
|
|
"schema_version": 1,
|
|
"family": "minimax_music3",
|
|
"display_name": "MiniMax Music 3",
|
|
"description": "MiniMax Music 3 text-to-music generation with lyrics conditioning.",
|
|
"category": "audio_generation",
|
|
"status": "experimental",
|
|
"tasks": [
|
|
"music"
|
|
],
|
|
"modes": [
|
|
"offline"
|
|
],
|
|
"languages": [
|
|
"auto"
|
|
],
|
|
"capabilities": {
|
|
"music": [
|
|
"lyrics",
|
|
"style_control"
|
|
]
|
|
},
|
|
"runtime": {
|
|
"tags": [
|
|
"gguf"
|
|
]
|
|
},
|
|
"options": {
|
|
"request": [
|
|
{
|
|
"name": "lyrics",
|
|
"type": "string",
|
|
"description": "Lyrics to sing. Structure tags such as [verse] and [chorus] are supported.",
|
|
"required": true
|
|
},
|
|
{
|
|
"name": "duration_sec",
|
|
"type": "float",
|
|
"description": "Positive AR frame budget expressed in seconds; not a hard final waveform duration limit.",
|
|
"required": false,
|
|
"min": 0.1,
|
|
"default": 20.0
|
|
},
|
|
{
|
|
"name": "num_inference_steps",
|
|
"type": "int",
|
|
"description": "Flow matching Euler steps per chunk.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 30
|
|
},
|
|
{
|
|
"name": "guidance_scale",
|
|
"type": "float",
|
|
"description": "Flow transformer classifier-free guidance scale. Must be > 0.",
|
|
"required": false,
|
|
"min": 0.0,
|
|
"default": 1.7
|
|
},
|
|
{
|
|
"name": "ar_guidance_scale",
|
|
"type": "float",
|
|
"description": "Autoregressive semantic and depth CFG scale. Must be > 0.",
|
|
"required": false,
|
|
"min": 0.0,
|
|
"default": 1.5
|
|
},
|
|
{
|
|
"name": "top_k",
|
|
"type": "int",
|
|
"description": "Top-k sampling for semantic and residual code sampling.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 50
|
|
},
|
|
{
|
|
"name": "seed",
|
|
"type": "int",
|
|
"description": "Generation seed.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 0
|
|
},
|
|
{
|
|
"name": "flow_uncond_interval",
|
|
"type": "int",
|
|
"description": "Evaluate unconditional flow CFG every N-th step, reusing the guidance delta between evaluations. Values <= 1 disable reuse and evaluate both branches every step.",
|
|
"required": false,
|
|
"default": 1
|
|
},
|
|
{
|
|
"name": "flow_uncond_warmup",
|
|
"type": "int",
|
|
"description": "Initial flow steps that evaluate both CFG branches before guidance-delta reuse; nonpositive values add no warmup.",
|
|
"required": false,
|
|
"default": 2
|
|
},
|
|
{
|
|
"name": "ensemble_takes",
|
|
"type": "int",
|
|
"description": "Independent takes in a batched AR pass using consecutive seeds; flow/vocoder decode each take. Named outputs take_01 through take_NN are emitted only when ensemble_takes > 1; otherwise only primary audio is returned.",
|
|
"required": false,
|
|
"min": 1,
|
|
"max": 16,
|
|
"default": 1
|
|
},
|
|
{
|
|
"name": "ensemble_prefix_frames",
|
|
"type": "int",
|
|
"description": "Shared intro frames before ensemble takes fork, approximately 25 frames per second. Zero disables intro locking.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 0
|
|
},
|
|
{
|
|
"name": "flow_chunk_hop_frames",
|
|
"type": "int",
|
|
"description": "Flow chunk hop in AR frames, approximately 25 per second. Zero uses checkpoint configuration; positive values are capped at the configured chunk size.",
|
|
"required": false,
|
|
"min": 0,
|
|
"default": 0
|
|
}
|
|
],
|
|
"session": [
|
|
{
|
|
"name": "weight_type",
|
|
"type": "enum",
|
|
"values": [
|
|
"native",
|
|
"bf16",
|
|
"f16",
|
|
"q8_0",
|
|
"q4_0",
|
|
"q4_k"
|
|
],
|
|
"description": "Shared weight storage type.",
|
|
"required": false,
|
|
"default": "native"
|
|
},
|
|
{
|
|
"name": "language_model_gguf",
|
|
"type": "string",
|
|
"description": "Language model component GGUF file relative to the model root.",
|
|
"required": false,
|
|
"default": "language_model_q4_0.gguf"
|
|
},
|
|
{
|
|
"name": "rvq_depth_decoder_gguf",
|
|
"type": "string",
|
|
"description": "RVQ depth decoder component GGUF file relative to the model root.",
|
|
"required": false,
|
|
"default": "rvq_depth_decoder_q8_0.gguf"
|
|
},
|
|
{
|
|
"name": "flow_transformer_gguf",
|
|
"type": "string",
|
|
"description": "Flow transformer component GGUF file relative to the model root.",
|
|
"required": false,
|
|
"default": "transformer_q4_0.gguf"
|
|
},
|
|
{
|
|
"name": "graph_arena_mb",
|
|
"type": "int",
|
|
"description": "Host graph-metadata arena size in MiB, not device workspace; legacy graph_context_mb remains accepted.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 32
|
|
},
|
|
{
|
|
"name": "weight_context_mb",
|
|
"type": "int",
|
|
"description": "Weight context size in MiB.",
|
|
"required": false,
|
|
"min": 1,
|
|
"default": 32
|
|
},
|
|
{
|
|
"name": "mem_saver",
|
|
"type": "bool",
|
|
"description": "Load large generation stages only while they are needed to reduce peak VRAM.",
|
|
"required": false,
|
|
"default": true
|
|
},
|
|
{
|
|
"name": "perf_mode",
|
|
"type": "enum",
|
|
"values": ["off", "cuda_perf"],
|
|
"description": "Performance preset. cuda_perf selects attention-prefix buckets of 256 rows. Explicit controls override preset values.",
|
|
"required": false,
|
|
"default": "off"
|
|
},
|
|
{
|
|
"name": "kv_cache_alignment",
|
|
"type": "int",
|
|
"description": "Attention-prefix bucket size. 1 preserves full-capacity attention; cuda_perf selects 256. Retains all valid history, but graph shapes may change sampled output. Explicit values override the preset.",
|
|
"required": false,
|
|
"min": 1
|
|
},
|
|
{
|
|
"name": "pipeline_overlap",
|
|
"type": "bool",
|
|
"description": "Overlap AR decoding with condition/flow/vocoder work on a second backend stream. Requires mem_saver=false; mem_saver=true is an error. Long requests and nondefault chunk hops use sequential execution.",
|
|
"required": false,
|
|
"default": false
|
|
}
|
|
],
|
|
"load": []
|
|
},
|
|
"package_defaults": {
|
|
"download": {
|
|
"kind": "huggingface_snapshot",
|
|
"repo": "audio-cpp/MiniMax-Music3-GGUF",
|
|
"revision": "main",
|
|
"gated": false
|
|
}
|
|
},
|
|
"packages": [
|
|
{
|
|
"id": "minimax_music3_q4_0",
|
|
"display_name": "MiniMax Music 3 Q4_0 GGUF",
|
|
"default": true,
|
|
"format": "gguf",
|
|
"precision": "q4_0",
|
|
"target_directory": "MiniMax-Music3-GGUF",
|
|
"files": [
|
|
"config.json",
|
|
"config/language_model.json",
|
|
"config/rvq_depth_decoder.json",
|
|
"config/condition_encoder.json",
|
|
"config/transformer.json",
|
|
"config/vocoder.json",
|
|
"tokenizer/tokenizer.json",
|
|
"tokenizer/tokenizer_config.json",
|
|
"language_model_q4_0.gguf",
|
|
"rvq_depth_decoder_q8_0.gguf",
|
|
"condition_encoder.gguf",
|
|
"transformer_q4_0.gguf",
|
|
"vocoder.gguf"
|
|
]
|
|
},
|
|
{
|
|
"id": "minimax_music3_q8_0",
|
|
"display_name": "MiniMax Music 3 Q8_0 GGUF",
|
|
"format": "gguf",
|
|
"precision": "q8_0",
|
|
"target_directory": "MiniMax-Music3-GGUF",
|
|
"files": [
|
|
"config.json",
|
|
"config/language_model.json",
|
|
"config/rvq_depth_decoder.json",
|
|
"config/condition_encoder.json",
|
|
"config/transformer.json",
|
|
"config/vocoder.json",
|
|
"tokenizer/tokenizer.json",
|
|
"tokenizer/tokenizer_config.json",
|
|
"language_model_q8_0.gguf",
|
|
"rvq_depth_decoder_q8_0.gguf",
|
|
"condition_encoder.gguf",
|
|
"transformer_q8_0.gguf",
|
|
"vocoder.gguf"
|
|
]
|
|
},
|
|
{
|
|
"id": "minimax_music3_bf16",
|
|
"display_name": "MiniMax Music 3 BF16 GGUF",
|
|
"format": "gguf",
|
|
"precision": "bf16",
|
|
"target_directory": "MiniMax-Music3-GGUF",
|
|
"files": [
|
|
"config.json",
|
|
"config/language_model.json",
|
|
"config/rvq_depth_decoder.json",
|
|
"config/condition_encoder.json",
|
|
"config/transformer.json",
|
|
"config/vocoder.json",
|
|
"tokenizer/tokenizer.json",
|
|
"tokenizer/tokenizer_config.json",
|
|
"language_model_bf16.gguf",
|
|
"rvq_depth_decoder_bf16.gguf",
|
|
"condition_encoder.gguf",
|
|
"transformer_bf16.gguf",
|
|
"vocoder.gguf"
|
|
]
|
|
}
|
|
],
|
|
"dependencies": [],
|
|
"ui": {
|
|
"recommended_package": "minimax_music3_q4_0",
|
|
"tags": [
|
|
"Music",
|
|
"GGUF"
|
|
],
|
|
"docs": [
|
|
"docs/models/minimax_music3.md",
|
|
"docs/music_generation.md",
|
|
"docs/gguf.md"
|
|
]
|
|
},
|
|
"sources": [
|
|
{
|
|
"format": "gguf",
|
|
"roots": {
|
|
"model": "."
|
|
},
|
|
"files": {
|
|
"root_config": "model:config.json",
|
|
"language_model_config": "model:config/language_model.json",
|
|
"depth_decoder_config": "model:config/rvq_depth_decoder.json",
|
|
"condition_encoder_config": "model:config/condition_encoder.json",
|
|
"transformer_config": "model:config/transformer.json",
|
|
"vocoder_config": "model:config/vocoder.json",
|
|
"tokenizer_json": "model:tokenizer/tokenizer.json",
|
|
"tokenizer_config": "model:tokenizer/tokenizer_config.json"
|
|
}
|
|
},
|
|
{
|
|
"format": "safetensors",
|
|
"roots": {
|
|
"model": "."
|
|
},
|
|
"files": {
|
|
"root_config": "model:config.json",
|
|
"language_model_config": "model:config/language_model.json",
|
|
"depth_decoder_config": "model:config/rvq_depth_decoder.json",
|
|
"condition_encoder_config": "model:config/condition_encoder.json",
|
|
"transformer_config": "model:config/transformer.json",
|
|
"vocoder_config": "model:config/vocoder.json",
|
|
"tokenizer_json": "model:tokenizer/tokenizer.json",
|
|
"tokenizer_config": "model:tokenizer/tokenizer_config.json"
|
|
},
|
|
"tensors": {
|
|
"language_model_weights": {
|
|
"source": "model:language_model/model.safetensors.index.json"
|
|
},
|
|
"depth_decoder_weights": {
|
|
"source": "model:rvq_depth_decoder/diffusion_pytorch_model.safetensors"
|
|
},
|
|
"condition_encoder_weights": {
|
|
"source": "model:condition_encoder/diffusion_pytorch_model.safetensors"
|
|
},
|
|
"transformer_weights": {
|
|
"source": "model:transformer/diffusion_pytorch_model.safetensors.index.json"
|
|
},
|
|
"vocoder_weights": {
|
|
"source": "model:vocoder/diffusion_pytorch_model.safetensors"
|
|
}
|
|
}
|
|
}
|
|
]
|
|
}
|