ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Update rl.py

Browse source
This commit is contained in:
Daniel Han 2025-12-15 22:43:16 -08:00
commit 4e04e56b82
1 changed files with 1 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

3
unsloth/models/rl.py
View file

@ -741,6 +741,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
"generation_kwargs": {},
"bf16": False,
"fp16": False,
"report_to" : "none",
"include_tokens_per_second": False,
"include_num_input_tokens_seen": False,
"auto_find_batch_size": False, # Auto /2 batch size - too many people complained so removing
@ -907,8 +908,6 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
for process_extra_arg in process_extra_args:
extra_args += process_extra_arg(old_RLTrainer_source, old_RLConfig_source)
# Edit report_to and default it to nothing if max_steps is like 60
# Create RLConfig args
extra_args = extra_args.split("\n")
extra_args = "\n".join(" " * 8 + x for x in extra_args)