Skip to content

README.md

Latest commit

 

History

History
32 lines (21 loc) · 1.04 KB

README.md

File metadata and controls

32 lines (21 loc) · 1.04 KB

GraphGPT

Training uses HuggingFace-style arguments flattened to YAML (see pygfm/public/cli/yaml_flat_to_argv.py).

Install

cd /path/to/repo
pip install -e .

Experiments

Stage Typical entry YAML template
Stage-1 alignment (train_mem) python scripts/graphgpt/run_with_config.py -c configs/graphgpt/train_mem_template.yaml configs/graphgpt/train_mem_template.yaml
Smoke configs/graphgpt/train_mem_smoke.yaml 1-step sanity check

Shell equivalents (multi-GPU): scripts/graphgpt/tune_script/graphgpt_stage1_1gpu.sh, graphgpt_stage2_1gpu.sh.

Data prep helpers: prepare_data.py, ensure_stage1_files.py, prepare_models.py.

Export YAML

python scripts/graphgpt/run_with_config.py --export-default-yaml configs/graphgpt/_defaults.yaml
python scripts/graphgpt/run_with_config.py --export-run-yaml configs/graphgpt/_run.yaml -c configs/graphgpt/train_mem_smoke.yaml

Environment

Set GRAPHGPT_NO_FLASH_ATTN=1 if FlashAttention is unavailable.