langtrain-ai 1.0.4

The unified Python SDK for training, aligning, and deploying LLMs — text and vision.

The unified Python SDK for training, aligning, and deploying LLMs

---

pip install langtrain-ai              # cloud API + dataset intelligence
pip install langtrain-ai[train]       # + local GPU training (text LLMs)
pip install langtrain-ai[vision]      # + vision LLMs (LLaVA, Qwen-VL, …)
pip install langtrain-ai[all]         # everything

What's inside

Module	What it does
FastLanguageModel	Unsloth-style API for text LLMs — local or cloud
FastVisionModel	Same API for vision LLMs (LLaVA, Qwen-VL, InternVL, …)
AdaptiveRankTrainer	Novel algorithm: SpectralLoRA + dynamic rank + TurboQuant
DatasetIntelligence	Drop any file → auto model + training config
LangtrainClient	Cloud API: fine-tune, deploy, chat, GPU, models
lt CLI	lt login, lt fine-tune, lt analyze, lt gpu

Quick start

Drop any dataset — get a training config

from langtrain import DatasetIntelligence

report = DatasetIntelligence.analyze("my_data.jsonl")
report.print_summary()
# ┌─────────────────────────────────────────────┐
# │  Task type    instruction  (87% confidence) │
# │  Domain       medical                       │
# │  Samples      2,400                         │
# │  Model        meta-llama/Llama-3.1-8B       │
# │  Method       adaptive_rank  rank=16        │
# │  TurboQuant   ✓ polar_quant+qjl             │
# └─────────────────────────────────────────────┘

Local training with AdaptiveRank

from langtrain import FastLanguageModel, AdaptiveRankConfig

config = AdaptiveRankConfig(
    initial_rank=16,
    max_rank=64,           # grows/shrinks based on gradient variance
    use_turboquant_kv=True # PolarQuant 3-bit KV cache
)

model, tokenizer = FastLanguageModel.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(model, method="adaptive_rank", config=config)
FastLanguageModel.train(model, tokenizer, dataset, output_dir="./output")

Cloud API

from langtrain import LangtrainClient

client = LangtrainClient(api_key="lt_...")

# Check account + GPU options
print(client.me())
print(client.gpu.available())

# Fine-tune
job = client.fine_tune(
    model="meta-llama/Llama-3.1-8B",
    dataset_id="ds_xyz",
    method="adaptive_rank",
)
for step in job.stream():
    print(step)

CLI

lt login                              # authenticate
lt whoami                             # account + GPU availability
lt gpu                                # list GPU options
lt analyze my_data.jsonl              # dataset intelligence
lt fine-tune llama-3.1-8b data.jsonl  # launch training
lt jobs                               # list jobs
lt models                             # list models

Why Langtrain vs Unsloth?

	Unsloth	Langtrain
Text LLMs	✓	✓ FastLanguageModel
Vision LLMs	✗	✓ FastVisionModel
Training algorithm	vanilla QLoRA	AdaptiveRank (SpectralLoRA + dynamic rank)
Dataset analysis	✗	✓ DatasetIntelligence (7-pass, auto model pick)
KV cache compression	✗	✓ TurboQuant (6× memory, 8× speed)
Cloud training	✗	✓ langtrain-server (A100/H100)
RL alignment	✗	✓ DPO / GRPO / PPO / Constitutional AI
CLI	✗	✓ lt fine-tune, lt analyze, lt gpu

---

Made by Langtrain AI · Docs · Discord