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

default num_chunks == -1

Browse source
This commit is contained in:
Daniel Han 2025-02-19 23:51:06 -08:00
commit 2ff2cdad63
2 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

6
unsloth/models/rl.py
View file

@ -127,12 +127,12 @@ class Unsloth{RLConfig_name}({RLConfig_name}):
metadata = {{'help': 'vLLM SamplingParams'}},
)
unsloth_num_chunks : Optional[int] = field(
default = 1,
metadata = {{'help': 'Chunk size to reduce memory usage'}},
default = -1,
metadata = {{'help': 'Chunk size to reduce memory usage. -1 is most efficient.'}},
)
def __init__({RLConfig_arguments},
vllm_sampling_params = None,
unsloth_num_chunks = 1,
unsloth_num_chunks = -1,
**kwargs,
):
{RLConfig_extra_args}

2
unsloth/models/rl_replacements.py
View file

@ -177,7 +177,7 @@ def grpo_trainer__get_per_token_logps(function_name, function):
if function_name != "_get_per_token_logps": return function
def _get_per_token_logps(self, model, input_ids, attention_mask, logits_to_keep):
if self.args.unsloth_num_chunks != 1: return None
return None # Unsloth efficient GRPO
if not hasattr(self, '_autocast_dtype'):
self._autocast_dtype = torch.float16 if os.environ.get('ACCELERATE_MIXED_PRECISION', 'fp16') == 'fp16' else torch.bfloat16
with torch.amp.autocast(device_type = 'cuda', dtype = self._autocast_dtype):