omg-hybridomga 1.0.1

OMG-hybridOMGa — Ultimate Hybrid LoRA Suite: LoRA, DoRA, QLoRA, LoRA+, rsLoRA, OMGa ve tam eğitim motoru

OMG-hybridOMGa ⚡🧠🔮♻️

Ultimate Hybrid LoRA Suite — LoRA, DoRA, QLoRA, LoRA+, rsLoRA ve OMGa'nın tam eğitim motoruyla birleşimi.

---

Kurulum

# Temel kurulum (sadece torch + transformers + accelerate)
pip install omg-hybridomga

# Tam kurulum (bitsandbytes, peft, trl, datasets dahil)
pip install "omg-hybridomga[full]"

# Her şey dahil
pip install "omg-hybridomga[all]"

---

Özellikler

LoRA Katmanları

Yöntem	Açıklama	Kaynak
LoRA	Standard Low-Rank Adaptation	Hu et al., 2021
DoRA	Weight-Decomposed LoRA	Liu et al., 2024
QLoRA	4/8-bit quantized base weights	Dettmers et al., 2023
LoRA+	Separate LRs for A and B matrices	Hayou et al., 2024
rsLoRA	Rank-Stabilized LoRA	Kalajdzievski, 2023
OMGa ★	OMG Adaptive LoRA — per-token gate, dual-rank	—

Bellek & VRAM Yönetimi

• MemoryMonitor — daemon thread VRAM watchdog
• VRAMGuard — otomatik batch küçültme + akıllı kurtarma
• MorphicMemory™ — Markov tahmini + tensor reincarnation
• 2-bit NF2 quantization (bitsandbytes gerektirmez)

Hız & Derleme

• Accelerator — grad accum, AMP, clip, fused optimizer
• CrystalCore™ — runtime kernel kristalizasyonu
• TritonKernels — RMSNorm, SwiGLU Triton fallback
• torch.compile — fullgraph + cache desteği

Optimizer & Scheduler

• EMA / SWA — Exponential/Stochastic Weight Averaging
• SpectraOptimizer™ — frekans-domain adaptive optimizer
• ResonanceScheduler™ — gradient-spectrum self-tuning LR
• WarmupCosineScheduler / WarmupLinearScheduler

Gradient İyileştirme

• GradientHarmonics™ — wavelet gradient processing
• NeuralProfiler™ — LSTM-based OOM/explode prediction
• LossSpikeDetector — spike tespiti + LR müdahalesi

---

Hızlı Başlangıç

Sadece LoRA Katmanı

from omg_hybridomga import HybridConfig, apply_hybrid_lora

cfg = HybridConfig(method="omga", rank=16, alpha=32)
model = apply_hybrid_lora(model, cfg)

Tam Eğitim Motoru

from omg_hybridomga import HybridOMGa

engine = HybridOMGa(
    "meta-llama/Llama-3.2-3B",
    rank=16,
    method="omga",
    ema=True,
    swa=True,
    crystal_core=True,
    priority_prop=True,
)

model, tokenizer = engine.load()
trainer = engine.get_trainer(dataset)
engine.train(trainer)

LoRA Kaydet / Yükle

from omg_hybridomga import save_hybrid_lora, load_hybrid_lora

save_hybrid_lora(model, "./my_lora_weights")
model = load_hybrid_lora(model, "./my_lora_weights")

Ortam Kontrolü

from omg_hybridomga import check_environment
check_environment()

---

Opsiyonel Bağımlılıklar

Grup	Kurulum	İçerik
full	pip install "omg-hybridomga[full]"	bitsandbytes, peft, trl, datasets
triton	pip install "omg-hybridomga[triton]"	Triton kernel desteği
flash	pip install "omg-hybridomga[flash]"	Flash Attention 2
deepspeed	pip install "omg-hybridomga[deepspeed]"	DeepSpeed ZeRO
all	pip install "omg-hybridomga[all]"	Hepsi

---

Gereksinimler

• Python ≥ 3.9
• PyTorch ≥ 2.1.0
• transformers ≥ 4.40.0
• accelerate ≥ 0.27.0

---

Lisans

Apache 2.0 — Ayrıntılar için LICENSE dosyasına bakın.