selgis 0.2.6.1

Universal Training Framework for PyTorch and HuggingFace Transformers

Selgis ML

Universal Training Framework for PyTorch and HuggingFace Transformers.

Selgis (Self-Guided Intelligent Stability) is a training framework with automatic failure protection.

---

What is Selgis

03:47 — Training started.
07:00 — Loss: NaN. Training crashed.
07:01 — You realize: 8 hours of work are gone.

Neural network training is fragile. Loss spikes, NaN, OOM and plateaus can destroy hours of computation. Standard trainers log the error and stop — you debug and restart manually.

Selgis automatically:

• Detects anomalies (NaN, spikes)
• Rolls back to stable state
• Lowers learning rate
• Continues training without your intervention

---

Installation

# Base (PyTorch only)
pip install selgis

# Full (Transformers, LoRA, quantization)
pip install "selgis[all]"

# Unsloth support (recommended for LLM training)
pip install unsloth

---

Quick Start

Any PyTorch model

from selgis import Trainer, SelgisConfig
from torch.utils.data import DataLoader

config = SelgisConfig(max_epochs=10)
trainer = Trainer(model, config, train_dataloader)
trainer.train()

LLM with LoRA

from selgis import TransformerTrainer, TransformerConfig

config = TransformerConfig(
    model_name_or_path="Qwen/Qwen2-0.5B",
    use_peft=True,
    peft_config={"r": 16, "target_modules": ["q_proj", "v_proj"]},
    quantization_type="4bit",
)

trainer = TransformerTrainer("Qwen/Qwen2-0.5B", config=config)
trainer.train()

---

Features

1. Self-Healing

Automatic recovery from anomalies:

config = SelgisConfig(
    nan_recovery=True,        # Auto-rollback on NaN/Inf
    spike_threshold=3.0,     # Rollback when loss 3x spike
    min_history_len=10,      # Detection window
)

What happens:

1. Loss becomes NaN — loads last stable state
2. Loss spikes sharply — rollback + LR reduced 50%
3. Optimizer momentum cleared
4. Training continues

2. Memory Optimization

Techniques for large models on small GPUs:

Technique	Savings
4-bit quantization	75%
CPU offload	40%
Gradient checkpointing	40%
LoRA (trainable only)	99.9%
Unsloth	50% less VRAM, 2x faster

config = TransformerConfig(
    quantization_type="4bit",
    cpu_offload=True,
    gradient_checkpointing=True,
    use_peft=True,
    peft_config={"r": 16},
)

2.1 Unsloth (NEW)

~2x faster training with ~50% less VRAM:

config = TransformerConfig(
    model_name_or_path="Qwen/Qwen2-0.5B",
    use_unsloth=True,
    use_peft=True,
    peft_config={"r": 16},
)

Works with: Llama, Qwen, Mistral, Phi, Gemma, Gemma 4.

3. Final Surge

Automatic plateau escape:

config = SelgisConfig(
    patience=5,               # epochs without improvement
    final_surge_factor=5.0,   # LR boost multiplier
)

If 5 epochs no improvement — LR multiplies to escape local minima.

4. LR Finder

Automatic learning rate search:

config = SelgisConfig(
    lr_finder_enabled=True,
    lr_finder_steps=100,
    lr_finder_start=1e-7,
    lr_finder_end=1.0,
)

Leslie Smith style — finds optimal LR in 100 steps.

5. Schedulers

Built-in schedulers:

config = SelgisConfig(
    scheduler_type="cosine_restart",  # cosine, linear, polynomial, constant
    warmup_ratio=0.1,
    min_lr=1e-7,
    t_0=10,
    t_mult=2,
)

6. Mixed Precision

config = SelgisConfig(
    fp16=True,   # FP16 mixed precision
    # bf16=True, # or BF16 for Ampere+
)

7. Gradient Management

config = SelgisConfig(
    grad_clip_norm=1.0,
    # grad_clip_value=0.5,
    gradient_accumulation_steps=4,
)

8. Checkpointing

config = SelgisConfig(
    output_dir="./output",
    save_best_only=True,
    save_total_limit=3,
    state_storage="disk",     # or "memory"
)

9. Callbacks

Extend functionality:

from selgis import (
    LoggingCallback,
    EarlyStoppingCallback,
    CheckpointCallback,
    HistoryCallback,
    WandBCallback,
    SparsityCallback,
)

callbacks = [
    LoggingCallback(log_every=10),
    CheckpointCallback(output_dir="./checkpoints"),
    EarlyStoppingCallback(patience=5, metric="accuracy", mode="max"),
    WandBCallback(project="my-project"),
]

10. Datasets

Unified data API:

from selgis import create_dataloaders, DatasetConfig

# Text (JSONL) - auto-detects format by extension
config = DatasetConfig(
    data_type="text",
    data_path="./data.jsonl",  # .jsonl, .json, .csv, .txt
    max_length=512,
)

# Chat datasets - auto-detects alpaca/sharegpt/messages
config = DatasetConfig(
    data_type="text",
    data_path="./alpaca_data.jsonl",  # auto-detects: alpaca, sharegpt, messages
)
# or manually:
config = DatasetConfig(
    data_type="text",
    data_path="./chat.jsonl",
    chat_format="messages",
    user_role="user",      # custom role (default)
    assistant_role="assistant",
)

# HuggingFace datasets
config = DatasetConfig(
    data_type="text",
    data_path="tatsu-lab/alpaca",  # auto-downloads from HF
)

# Image
config = DatasetConfig(
    data_type="image",
    data_path="./images",
)

# Streaming (large files)
config = DatasetConfig(
    data_type="streaming",
    data_path="./large.jsonl",
    buffer_size=1000,
)

train_loader, eval_loader = create_dataloaders(config)

---

CLI

# Demo mode
selgis train

# From config
selgis train --config config.yaml

# Check device
selgis device

# Run tests
selgis test

---

Configuration

Parameter	Default	Description
max_epochs	100	Max epochs
learning_rate	1e-3	Base LR
batch_size	32	Batch size
nan_recovery	True	Auto-rollback
spike_threshold	3.0	Spike detection
grad_clip_norm	1.0	Gradient clip
save_best_only	True	Save best only
cpu_offload	False	CPU optimizer
final_surge_factor	5.0	LR boost on plateau

---

Examples

Full examples: example_selgis.py

# Basic
from selgis import Trainer, SelgisConfig
config = SelgisConfig(max_epochs=10)
trainer = Trainer(model, config, loader)
trainer.train()

# LoRA
from selgis import TransformerTrainer, TransformerConfig
config = TransformerConfig(model_name_or_path="Qwen/Qwen2-0.5B", use_peft=True)
trainer = TransformerTrainer("Qwen/Qwen2-0.5B", config)
trainer.train()

# Callbacks
from selgis import LoggingCallback, CheckpointCallback
trainer = Trainer(model, config, loader, callbacks=[
    LoggingCallback(log_every=10),
    CheckpointCallback(output_dir="./ckpt"),
])

---

Dependencies

# Base
torch>=2.0, numpy>=1.20, tqdm

# Optional
transformers>=4.30, datasets, accelerate>=0.21.0
peft>=0.5.0
bitsandbytes>=0.41.0
wandb
pytest

---

Limitations

• DeepSpeed — partial support (v0.3.0)
• FSDP — in development

---

Future Plans

• Unsloth integration — DONE (v0.2.6)

  • 2x faster training, 50% less VRAM
  • Llama, Qwen, Mistral, Phi, Gemma, Gemma 4 support
  • Run locally or from HuggingFace

• DeepSpeed full — complete ZeRO, pipeline

• FSDP — Fully Sharded Data Parallel

• Distributed Training — DDP, multi-GPU

• More schedulers — OneCycle, ReduceLROnPlateau

• MLflow integration — W&B alternative

---

Links

• PyPI
• GitHub
• API Docs
• Tests
• Examples

---

License: Apache 2.0