KLM Trainer Test Ver 0.4.6

This application is a tool for training RVC models. Currently, it is only available for use with Nvidia GPUs.

*Notice*
The KLM-HF architecture is currently under testing and is not fully functional.

New Features

Batch pitch and embedder extraction
Supports extremely fast extraction with both sequential and overlap multi-GPU processing.
A batch size of 16–32 is recommended during extraction.

Dataset Type (Automatic Embedder Selection)
Speech Only = ContentVec
Vocal Only = KLM-Vocal
Speech + Vocal Mixed = KLM-SVM
Quick Training

Allows fast training with simple input. Convenient for beginners or when creating voice models.

Pretrained Model
OG - Default RVC OG pretrained model
KLM - KLM KPU
Custom - Manually select any RVC-compatible model
Delete old model and start from fresh

If this option is enabled, two warning messages will appear in the CMD window.
This function deletes the existing model and starts training from scratch.

If you want to delete all extracted data (embedding, pitch, etc.)
If you only want to delete TensorBoard events and trained weights
Fast Training

Useful for training standard models (not suitable for large datasets over 2 hours).
Disables all safety mechanisms and focuses resources on training speed.
Automatically keeps dataset cache in VRAM (not suitable for low VRAM GPUs).

Overtraining Warning
When enabled, it calculates the average value at each step and automatically saves the model as "[ModelName]_BestModel.pth" whenever it reaches an optimal value in TensorBoard. Training will not stop and continues until the specified epoch is reached.
If NaN, gradient explosion, or TensorBoard issues occur, a warning will be displayed in the logs.
Midpoint AutoSave (MAS)

Saves checkpoints based on steps instead of epochs.
Weights are not saved, but if the app closes due to Windows updates or errors, training can be resumed from the saved checkpoint.
Setting the interval too low may slow down training.

START / STOP / RESUME

Allows you to start, stop, and resume training.
If MAS is enabled, training can resume from the checkpoint after stopping.

Training ADV

Useful for training pretrained models or large-scale models.

Extract Cuda Devices / Train Cuda Devices

KLM Trainer uses different multi-GPU strategies for extraction and training.
Extraction supports multiple GPUs, but training supports up to two GPUs.

Model Type
Base Model - For standard model training
Pretrained Model - For pretrained model training
Pretrain Model Initialization
From Scratch - Start training from zero with all safety mechanisms enabled
Engine Bootstrap Weight - Fine-tune from an existing pretrained model
Model Architecture
RVC V2 - Standard RVC V2 model
KLM HF - Higher-parameter KLM model (currently in development)
F0 Method

rmvpe

Embedder Model
ContentVEC - Default ContentVEC
KLM Vocal - ContentVEC backbone + 150 hours of vocal data
KLM SVM - ContentVEC backbone + 200 hours of vocal + speech data
Pipeline Mode

Determines dataset extraction method.

Sequential
Similar to the original RVC method, multiple GPUs extract pitch and embeddings sequentially.
Very fast due to batch processing.

Overlap Pitch + Embedder
Each GPU extracts pitch and embeddings separately, and completed GPUs assist others to distribute workload dynamically.

Batch Settings
Pitch Batch / Embed Batch: Recommended 8–32 (64+ may reduce performance)
Audio Loader: Match CPU core count, but do not exceed 32
Embedder Precision

If supported, BF16 provides the fastest performance.

Mel Mask Assistant (MMA)

A helper engine that generates a soft occupancy mask from mel distribution to reduce unwanted energy leakage.
Vocal Safe mode prevents excessive suppression of breath, tail, and high-frequency harmonics.

Mask Profile

Controls the strength of MMA influence.
Vocal Safe minimizes intervention, while Speech applies stronger masking.

Mask Fill Weight

Weight that improves accuracy in regions where energy should exist (recommended: 0.25)

Mask Empty Weight

Weight that suppresses unwanted energy in silent regions.
Too high values may damage breath sounds or sibilance.

Spectral Anti-Mirror Assistant (SAMA)

A helper engine that suppresses high-frequency leakage and mirroring artifacts.
Especially useful for reducing early-stage mirroring when training from scratch.

Upper Leakage Weight: Controls high-frequency leakage suppression
Mirror Pattern Weight: Controls overlapping frequency mirroring suppression

*Note*: Setting values too high may negatively affect training.

Warmup Steps

Used only for pretrained models trained from scratch.
Gradually adjusts the learning rate to prevent gradient explosion in early stages.

TensorBoard

No need to launch separately. Graphs are displayed directly in the UI.