Skip to main content

NexusLoRA — Unified LLM Fine-Tuning Engine (FastLoRA × NexusTrain)

Project description

NexusLoRA ⚡🛡️🧠🔮

Unified LLM Fine-Tuning Engine — FastLoRA v4 + NexusTrain v1 tek dosyada birleştirilmiş hali.


Kurulum

# Temel kurulum
pip install nexuslora

# Eğitim araçlarıyla birlikte (önerilen)
pip install nexuslora[train]

# Tam kurulum (flash-attn dahil)
pip install nexuslora[full]

# Geliştirici kurulumu
git clone https://github.com/fastloraoffical/nexuslora
cd nexuslora
pip install -e ".[dev]"

Hızlı Başlangıç

from nexuslora import NexusLoRA

# Tam özellikli — tüm NexusLoRA modülleri aktif
nl = NexusLoRA("meta-llama/Llama-3.2-3B",
               nexus_enabled=True, nexus_power=1.0)
model, tokenizer = nl.load()
trainer = nl.get_trainer(train_dataset)
nl.train(trainer)

Sadece temel LoRA (NexusTrain modülleri kapalı)

nl = NexusLoRA("meta-llama/Llama-3.2-3B", nexus_enabled=False)
model, tokenizer = nl.load()
trainer = nl.get_trainer(train_dataset)
nl.train(trainer)

Seçici modül aktivasyonu

nl = NexusLoRA(
    "meta-llama/Llama-3.2-3B",
    nexus_enabled=True,
    nexus_power=1.0,
    # Çakışma önleme:
    torch_compile=False,             # CrystalCore™ aktif olacak
    nexus_crystal_core=True,
    mixed_precision_optimize=False,  # ChromaticPrecision™ aktif olacak
    nexus_chromatic_precision=True,
    lr_scheduler="constant",         # ResonanceScheduler™ aktif olacak
)

Özellikler

FastLoRA Çekirdeği

Özellik Açıklama
Custom Triton Kernels RMSNorm, SwiGLU, RoPE fused implementasyonlar
2-bit / 4-bit / 8-bit Quant bitsandbytes bağımsız 2-bit dahil
Mixture of Experts Sparse MoE router + otomatik patch
CPU/NVMe Offloading ZeRO-Infinity tarzı bellek yönetimi
OOM Recovery Eğitimi kesmeden OOM'dan devam
Optuna AutoTune lr / rank / batch otomatik optimizasyonu
UnstoppableTrainer Her hatadan otomatik kurtarma

NexusTrain Modülleri

Modül Açıklama
CrystalCore™ Runtime kernel kristalizasyonu
MorphicMemory™ Markov tahminli tensor yeniden kullanımı
SpectraOptimizer™ FFT tabanlı AdamW üstü optimizer
ResonanceScheduler™ Gradient spektrumundan öz-ayarlı LR
ChromaticPrecision™ Per-layer dinamik dtype ataması
GradientHarmonics™ Wavelet tabanlı gradient işleme
NeuralProfiler™ LSTM ile OOM/explode tahmini
CrystalPipeline™ Dinamik grad-accum + async checkpoint
ZeroWaste™ Ölü parametre eliminasyonu
UniversalAdapter™ HF/TRL/PEFT/DeepSpeed otomatik patch

Çakışma Uyarıları

FastLoRA NexusTrain Öneri
torch_compile=True nexus_crystal_core=True Birini kapatın
paged_optimizer=True nexus_spectra_optimizer=True Küçük VRAM'de spectra'yı kapatın
mixed_precision_optimize=True nexus_chromatic_precision=True Birini kapatın
dynamic_batch_scaling=True nexus_crystal_pipeline=True Birini kapatın
smart_checkpoint=True nexus_crystal_pipeline=True Birini kapatın
loss_spike_detection=True nexus_neural_profiler=True Birini kapatın
gradient_noise_monitor=True nexus_gradient_harmonics=True Monitor yanıltıcı olur
mem_defrag=True nexus_morphic_memory=True Zararsız; defrag interval 2x artırın

API Referansı

nl = NexusLoRA(model_name, **kwargs)   # Konfigürasyon
model, tokenizer = nl.load()           # Model yükle + tüm optimizasyonlar
trainer = nl.get_trainer(dataset)      # Trainer oluştur
nl.train(trainer)                      # Eğit
nl.save("./output")                    # Kaydet
nl.push_to_hub("kullanici/model")      # HuggingFace Hub
nl.merge_and_unload()                  # LoRA merge
response = nl.generate("prompt")       # Inference
nl.nexus_async_checkpoint()            # Async checkpoint
nl.profile()                           # Benchmark
nl.stop()                              # Temiz kapatma

Gereksinimler

  • Python ≥ 3.9
  • PyTorch ≥ 2.1.0
  • Transformers ≥ 4.40.0
  • Accelerate ≥ 0.27.0

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

nexuslora-1.0.0.tar.gz (44.9 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

nexuslora-1.0.0-py3-none-any.whl (44.3 kB view details)

Uploaded Python 3

File details

Details for the file nexuslora-1.0.0.tar.gz.

File metadata

  • Download URL: nexuslora-1.0.0.tar.gz
  • Upload date:
  • Size: 44.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.7

File hashes

Hashes for nexuslora-1.0.0.tar.gz
Algorithm Hash digest
SHA256 367c7bd576965991d980f413e72dbbe337ee720c4b6a1f3759f07e47639f4198
MD5 9fa2bfb31568c807850607edb9ede7b0
BLAKE2b-256 2a60e02721d7b57963f558d9341ed935a0df4629fadd48b5ea7ee0eccbc54337

See more details on using hashes here.

File details

Details for the file nexuslora-1.0.0-py3-none-any.whl.

File metadata

  • Download URL: nexuslora-1.0.0-py3-none-any.whl
  • Upload date:
  • Size: 44.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.7

File hashes

Hashes for nexuslora-1.0.0-py3-none-any.whl
Algorithm Hash digest
SHA256 551d27fa432b8fb8e83d43fd40d7fba4adc537e5f5fd964ce069e89f92c55862
MD5 b313a6ca371e0d1fee799d34eb11dde1
BLAKE2b-256 01c3871d20855ab55303539c9b8efe48c308afe73fb7b7ebbb4734f21ec887ff

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page