Skip to content

v4.7.1 - MiniMax Music

Choose a tag to compare

View all tags
@bghira bghira released this 20 Aug 19:51
· 220 commits to release since this release
v4.7.1
25e841e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Features

  • MiniMax Music 3 - a very difficult campaign
    • Building an open RVQ encoder by distilling the MiniMax Music LM output logits back into an approximate encoder distribution
    • Attempting to inject the RVQ audio codes directly to attempt to exploit any built-in reference conditioning (there isn't any)
    • Putting together a latent refiner to restore poor-quality real audio (it works!)
    • Attempting to build a latent replanner to "remix" and cover existing audio outputs from MiniMax Music 3 or real audio - it can refine AI output audio and restore its quality, but not able to transfer style or cover artists
    • Training a working artist LoRA on a single 48G GPU in a couple hours, testing a number of strength values and caption styles' impact on the base model
  • LTX Video 2.5 finetuning fully supported
  • Tuned configurations for MiniMax H3 are now provided that function at reasonable speeds across the 24/32/48G GPU size range
  • Webshart datasets now can use .txt sidecar captions (textfile caption strategy natively supported)
  • Webshart datasets can be optimised by SimpleTuner via a dataloader config option; txt files are collected into the optimised json indices instead
  • New option --init_lora_step to jump to a previous training runs' correct step count
  • LoRA global_step is recorded in metadata and automatically used for continuing training of a LoRA if detected (set init_lora_step=0 to override this)
  • New dataloader option train_batch_size for per-dataset batch sizing; useful if mixing 512/1024/2048px image sizes or different num_frames video dataset copies
  • MiniMax H3 audio-only training - if MiniMax Music doesn't suit your fancy, you can only define audio datasets for H3 instead, and it will automatically optimise training for the creation of a music model. Example LoRA here which teaches a particular prompt format to make it generate music coherently and reliably.
  • Community-trained Anima 2.9B flavour is now added and available for easy selection from CLI or WebUI configuration management

What's Changed

  • feat: add optional Kubeflow single-GPU worker integration by @duroey in #2874
  • docs: add Kubeflow server options section to all translated OPTIONS files by @bghira with @Copilot in #2881
  • merge release into main by @bghira in #3038
  • minimax h3: negative prompting fixes for validation by @bghira in #3039
  • AnyFlow: improve reproducibility of results by @bghira in #3040
  • AnyFlow: negative prompt handling fixes by @bghira in #3041
  • restore warmup LR correctly across resume world sizes by @bghira in #3045
  • add --init_lora_step to specify the last global_step the initial LoRA weights came from by @bghira in #3043
  • distillation should disable lora_dropout unless force-enabled by user by @bghira in #3046
  • anchor AnyFlow diffusion targets for guided bases by @bghira in #3044
  • add h3 to list of segmented checkpointing supported families by @bghira in #3050
  • anyflow logging for diagnostics by @bghira in #3051
  • minimax h3: batch dimension fix for sparse attention by @bghira in #3049
  • metal-flash-attention fix for quantised attention backend selector by @bghira in #3048
  • add metal-flash-attention install script and docs by @bghira in #3053
  • minimax h3 training config example tuned for H200 by @bghira in #3052
  • initial LTX 2.5 support by @bghira in #3047
  • ideogram4: fix quantisation selector, add base upcast option and vram-specific config examples by @bghira in #3055
  • update huggingface kernels library by @bghira in #3056
  • AnyFlow for Wan 2.1 by @bghira in #3054
  • webshart: accept .txt sidecar captions by @bghira in #3057
  • native single-worker kubeflow integration by @bghira in #2882
  • block lora_init_step being higher than max_train_steps by @bghira in #3063
  • AnyFlow compatibility fixes for ideogram4 by @bghira in #3062
  • improve minimax h3 LoRA swapped-gate detection; write metadata indicating the order by @bghira in #3058
  • AnyFlow serialisation fix for metadata by @bghira in #3061
  • anyflow default guidance settings by @bghira in #3060
  • webshart: retrieve txt based captions and cache them via SaveHooks by @bghira in #3059
  • webshart caption lookup by @bghira in #3064
  • ideogram4: fix hub model uploads since it rejects invalid metadata by @bghira in #3066
  • ideogram4 conditioning keys compatibility with anyflow by @bghira in #3065
  • add gazingstars123 anima 2.9b model by @bghira in #3067
  • Add Gazingstars Anima 2.9B flavour by @bghira in #3069
  • Enable H3 text encoder RamTorch presets by @bghira in #3071
  • Fix pipeline runtime conditioning by @bghira in #3073
  • Add MiniMax Music model family by @bghira in #3074
  • Add validation start after step & epoch gates by @bghira in #3072
  • AnyFlow: trainable+persisted FlowMap delta embedders, minimaxmusic validation wrapper coverage by @bghira in #3075
  • Add real Ideogram 4 unconditional transformer support by @bghira in #3076
  • minimax music: tuned example configs for 24/32/48g vram systems by @bghira in #3077
  • webshart: memoize shard metadata, ignore empty caption cache files by @bghira in #3079
  • Fix RamTorch residency for quantized PEFT layers by @bghira in #3078
  • Default textfile lyrics for MiniMax Music by @bghira in #3080
  • h3 unsloth checkpointing + sparse attn recompile fix by @bghira in #3081
  • Split text embed precompute across ranks by @bghira in #3083
  • Publish validation audio assets by @bghira in #3082
  • Use cached MiniMax Music validation embeds by @bghira in #3084
  • anyflow: swap unconditional aliases instead of popping them by @bghira in #3085
  • Gather batch-parallel validation payloads by @bghira in #3086
  • Support audio-only datasets for LTX-2 and MiniMax-H3 by @bghira in #3087
  • dataset specific train batch size by @bghira in #3088
  • Support variable dataset batch sizes across distributed ranks by @bghira in #3095
  • Validate per-dataset batch sizes on MPS by @bghira in #3097
  • Validate dataset train batch sizes strictly by @bghira in #3089
  • Keep dataset batch size current across cache reloads by @bghira in #3090
  • Guard batch-size changes when resuming sampler state by @bghira in #3093
  • Document per-dataset batch size accounting by @bghira in #3094
  • Wire caption batch sizing and align WebUI eligibility by @bghira in #3096
  • fix audio-only validation upload from h3 model by @bghira in #3098
  • Fix Webshart multi-caption text precache by @bghira in #3102
  • h3: context parallel fix by @bghira in #3109
  • Fix on-demand text embed collation by @bghira in #3103
  • functioning MiniMax Music 3 training (RVQ encoder, LM LoRA and DiT finetuning) by @bghira in #3108
  • merge by @bghira in #3112
  • Bump version from 4.7.0 to 4.7.1 by @bghira in #3113

Full Changelog: v4.7.0...v4.7.1

Contributors

bghira and duroey
Assets 2
Loading