search

bigcode-project / Megatron-LM

Ongoing research training transformer models at scale
Other
371 stars 48 forks source link

HumanEval greedy (sequential) generation using a server #83

Open loubnabnl opened 10 months ago

loubnabnl commented 10 months ago

Run generations for humaneval problems using Megatron (sequential generation for greedy mode)

# deploy the server on a GPU node (adapt paths/ports..)
bash examples/run_text_generation_starcoder.sh
# run and save generations (on the same node) to be executed with bigcode-evaluation-harness
python /fsx/loubna/bigcode_2/code/pr/Megatron-LM/tools/run_requests_humaneval.py