# AGILLM2-fast-training · `5L.py`

Autoregressive (AR-only) single-file trainer/decoder using the Qwen3 tokenizer

**Repo:** [https://huggingface.co/OpenTransformer/AGILLM2-fast-training](https://huggingface.co/OpenTransformer/AGILLM2-fast-training)
**Org:** [https://huggingface.co/OpenTransformer](https://huggingface.co/OpenTransformer)
**Contact:** [OpenTransformers@proton.me](mailto:OpenTransformers@proton.me)

## Overview

`5L.py` is a ~single-file PyTorch training and inference script for language models with:

* **AR-only** training/decoding
* **Qwen3** tokenizer by default (override via `TOKENIZER_ID`)
* **Progressive block growth**, **AMP/FP8 autocast**, **OOM backoff**
* **Time-based checkpointing** only (monotonic, resume-safe)
* **Sampling controls:** top-k/top-p/min-p, greedy, repetition/presence/frequency penalties, no-repeat-ngrams
* **Chinchilla-style target token estimator** using all enabled params (core + AR head)

The goal is **minimal surface area** with production-lean features so you can train quickly, resume safely, and decode reliably on commodity GPUs or cloud nodes.

## Features

* **Presets:** `small`, `smallx2`, `base`
* **Attention:** Low-rank MHA with ALiBi relative bias
* **Determinism helpers:** seed management, checkpoint metadata (RNG states)
* **Tokenizer safety:** adds `[PAD]` if missing; handles EOS fallbacks
* **Streaming data:** uses `datasets` streaming for large corpora

## Requirements

* Python 3.10+
* PyTorch 2.2+ (CUDA build if using NVIDIA GPUs)
* `transformers`, `datasets`, `tqdm`
* CUDA-capable GPU recommended; script also runs CPU-only for smoke tests

Install:

```bash
pip install torch --index-url https://download.pytorch.org/whl/cu121  # pick your CUDA/CPU wheel
pip install transformers datasets tqdm
```

## Quick start

### 1) Set tokenizer (optional)

Default is Qwen3:

```bash
export TOKENIZER_ID="Qwen/Qwen3-235B-A22B-Thinking-2507"
```

Use any compatible tokenizer:

```bash
export TOKENIZER_ID="qwen/qwen2.5-7b"
```

### 2) Train

Minimal example on SlimPajama (streaming):

```bash
python 5L.py train \
  --preset small \
  --source cerebras/SlimPajama-627B \
  --amp \
  --save_dir ckpts_joint \
  --save_every_sec 7200
```

Targets and steps:

```bash
# Let script compute Chinchilla-style target tokens automatically
python 5L.py train --preset small --amp

# Or cap by steps
python 5L.py train --preset small --steps 20000 --amp
```

Warm start / resume:

```bash
# Warm-start from a prior final.pt (shape-safe copy of matching tensors)
python 5L.py train --preset small --warmstart_from ckpts_joint/final.pt

# Full resume (optimizer, scaler, seen tokens, timers)
python 5L.py train --resume ckpts_joint/step00050000.pt
```

Progressive block growth:

```bash
python 5L.py train \
  --preset small \
  --auto_grow \
  --grow_plan "576,640,768,896,1024" \
  --grow_every_steps 50000
```

FP8 fast path:

```bash
# Try FP8; if not supported, fall back to bf16
python 5L.py train --preset small --fp8-only --fp8-fallback
```

### 3) Inference

```bash
python 5L.py infer \
  --mode ar \
  --ckpt ckpts_joint/final.pt \
  --preset small \
  --prompt "Explain ALiBi in simple terms." \
  --max_new 120 \
  --top_p 0.9 --top_k 50 \
  --repetition_penalty 1.1 \
  --no_repeat_ngram_size 3
```

Greedy decode:

```bash
python 5L.py infer --mode ar --ckpt ckpts_joint/final.pt --preset small \
  --prompt "What is progressive block growth in training?" --greedy --max_new 80
```

FP8 during decode (if supported):

```bash
python 5L.py infer --mode ar --ckpt ckpts_joint/final.pt --preset small \
  --prompt "Summarize transformer attention variants." --fp8-only --fp8-fallback
```

## Presets

```text
small    : d=512, layers=8,  heads=16, rank=64
smallx2  : d=512, layers=16, heads=16, rank=64
base     : d=768, layers=12, heads=24, rank=96
```

Use `--x2` during training to double layers of an inferred previous config.

## Checkpointing & Resume

* **Saves** only by **time interval** (`--save_every_sec`, default 24h) to avoid step-based drift.
* `final.pt` includes: core, AR head, optimizer, AMP scaler, cfg, RNG states, and metadata.
* **Resume** with `--resume <path>` to restore optimizer/scaler/wall-clock cadence.
* **Warm start** only copies shape-matched tensors (safe if your topology changed).

Artifacts:

* `ckpts_joint/stepXXXXXXXX.pt`
* `ckpts_joint/latest.json` with canonical latest path and step

## Data

Default streaming dataset:

* `cerebras/SlimPajama-627B` (train split, streaming enabled).
  Replace `--source` with any `datasets`-compatible corpus that yields `{"text": ...}`.

EOS handling: if tokenizer’s `eos_token_id` is missing, uses `sep_token_id`; if a sample doesn’t end with EOS, one is appended.

## Sampling controls

* `--temperature`, `--top_k`, `--top_p`, `--min_p`
* `--repetition_penalty`, `--presence_penalty`, `--frequency_penalty`, `--penalty_last_n`
* `--no_repeat_ngram_size`

Greedy mode (`--greedy`) overrides sampling.

## FP8 / AMP

* `--fp8-only` attempts `float8_e4m3fn` autocast
* `--fp8-fallback` continues with bf16 if FP8 unsupported
* Otherwise use `--amp` for bf16/fp16 autocast
* `torch.backends.cuda.matmul.allow_tf32=True` is enabled when available

## OOM backoff & block growth

* On CUDA OOM, the script **halves** `BLOCK` (down to 128), empties cache, and retries the step.
* With `--auto_grow`, the script periodically attempts to **increase** `BLOCK` along your `--grow_plan`.

## Token targets (Chinchilla-style)

If `--target_tokens` is unspecified, the script computes `25 × (enabled parameters)` using **all** trainable params (core + AR head). This provides a rough target for total tokens to consume.

## Repro tips

* Pin a specific tokenizer via `TOKENIZER_ID`
* Log your `--preset`, `--block`, and `--grow_plan`
* Keep `save_every_sec` stable between resumes for monotonic cadence
* Record CUDA/cuDNN versions in your run logs for reproducibility

## Limitations

* AR-only trainer (no encoder-decoder, no multimodal)
* Low-rank MHA path; FlashAttention not included
* Single-GPU by default; multi-GPU DDP not wired in this file
* Safety/guardrails are out of scope here (this is a trainer, not a hosted chat product)

## Roadmap (planned)

* Optional DDP with NCCL/RCCL/HCCL backends
* FlashAttention path when available across vendors
* Export helpers (Safetensors, GGUF) for downstream serving

## Responsible Use

* Ensure your dataset usage complies with its license and applicable laws.
* Models trained with this script can generate incorrect or biased outputs.
  Evaluate and align according to your deployment requirements.

## Citation

If this script or training pipeline helps your work, consider citing the repo:

```bibtex
@software{OpenTransformer_AGILLM2_fast_training_2025,
  title   = {AGILLM2-fast-training: Single-file AR-only trainer/decoder (5L.py)},
  author  = {OpenTransformers},
  year    = {2025},
  url     = {https://huggingface.co/OpenTransformer/AGILLM2-fast-training}
}
```

---

**Support / Contracts**
We provide **custom development** and **end-to-end training** services (data prep → training → evaluation → deployment).
Email: **[OpenTransformers@proton.me](mailto:OpenTransformers@proton.me)**
Org page: [https://huggingface.co/OpenTransformer](https://huggingface.co/OpenTransformer)