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

unsloth_num_chunks

Browse source
This commit is contained in:
Daniel Han 2025-02-17 22:17:11 -08:00
commit e554eb6f28
2 changed files with 7 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
unsloth/models/rl.py
View file

@ -117,6 +117,10 @@ class Unsloth{RLConfig_name}({RLConfig_name}):
default = None,
metadata = {{'help': 'vLLM SamplingParams'}},
)
unsloth_num_chunks : Optional[int] = field(
default = 1,
metadata = {{'help': 'Chunk size to reduce memory usage'}},
)
def __init__({RLConfig_arguments},
vllm_sampling_params = None,
**kwargs,

5
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):
return None
if self.args.unsloth_num_chunks != 1: return None
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):
@ -240,7 +240,8 @@ def grpo_trainer_compute_loss(function_name, function):
)
else:
loss, completion_length, mean_kl = grpo_accumulated_loss(
self, _input_ids, logits_to_keep, completion_mask, advantages, n_chunks = 2,
self, _input_ids, logits_to_keep, completion_mask, advantages,
n_chunks = self.args.unsloth_num_chunks,
)
# Log the metrics