owusu/DS-LLM-TEMPLATE-FINETUNING
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

instruct model setup

Browse Source
This commit is contained in:
Your Name
2025-08-28 17:57:59 +00:00
parent 77c563f358
commit d49b4ff2d5
55 changed files with 27760 additions and 326 deletions
Download Patch File Download Diff File Expand all files Collapse all files
unsloth_compiled_cache/UnslothRLOOTrainer.py
+28 -9
View File
@@ -1,7 +1,7 @@
"""
2025.8.4
2025.8.5
4.55.1
2025.8.9
2025.8.10
4.55.4
0.21.0
__UNSLOTH_VERSIONING__
"""
@@ -99,6 +99,7 @@ class UnslothRLOOConfig(RLOOConfig):
default = -1,
metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
)
def __init__(
self,
output_dir = None,
@@ -261,6 +262,7 @@ class UnslothRLOOConfig(RLOOConfig):
ds3_gather_for_generation = True,
vllm_sampling_params = None,
unsloth_num_chunks = -1,
**kwargs,
):
if learning_rate < 1e-7: raise FloatingPointError(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
@@ -270,7 +272,7 @@ class UnslothRLOOConfig(RLOOConfig):
save_strategy = 'no'
if dataset_num_proc is None:
from multiprocessing import cpu_count
dataset_num_proc = min(cpu_count()*2, 2)
dataset_num_proc = max(cpu_count()+4, 2)
if temperature <= 0:
raise MathError('Unsloth: Please set a positive non-zero temperature since your results will be wrong.')
elif temperature >= 10:
@@ -438,6 +440,7 @@ class UnslothRLOOConfig(RLOOConfig):
ds3_gather_for_generation = ds3_gather_for_generation,**kwargs)
self.vllm_sampling_params = vllm_sampling_params
self.unsloth_num_chunks = unsloth_num_chunks
pass
class _UnslothRLOOTrainer(Trainer):
@@ -865,7 +868,7 @@ class _UnslothRLOOTrainer(Trainer):
with torch.no_grad():
pg_clipfrac = (pg_losses2 > pg_losses).float().mean()
prob_dist = torch.nn.functional.softmax(logits, dim=-1)
prob_dist = torch.nn.functional.softmax(logits, dim=-1, dtype = torch.float32).to(logits.dtype)
entropy = torch.logsumexp(logits, dim=-1) - torch.sum(prob_dist * logits, dim=-1)
approxkl = 0.5 * (logprobs_diff**2).mean()
approxkl_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = approxkl
@@ -1125,9 +1128,17 @@ class UnslothRLOOTrainer(_UnslothRLOOTrainer):
from unsloth_zoo.vision_utils import UnslothVisionDataCollator
if not isinstance(data_collator, UnslothVisionDataCollator):
if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
data_collator = TransformersDataCollatorForLanguageModeling(__tokenizer, mlm = False, mlm_probability = 0.0)
data_collator = TransformersDataCollatorForLanguageModeling(
__tokenizer,
mlm = False,
mlm_probability = 0.0,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
data_collator = DataCollatorForSeq2Seq(__tokenizer)
data_collator = DataCollatorForSeq2Seq(
__tokenizer,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
else:
if hasattr(args, 'remove_unused_columns'): args.remove_unused_columns = False
if hasattr(args, 'dataset_text_field'): args.dataset_text_field = ''
@@ -1135,9 +1146,17 @@ class UnslothRLOOTrainer(_UnslothRLOOTrainer):
if not isinstance(data_collator, UnslothVisionDataCollator):
if not hasattr(__tokenizer, 'pad') and hasattr(__tokenizer, 'tokenizer'):
if isinstance(data_collator, DataCollatorForSeq2Seq):
data_collator = DataCollatorForSeq2Seq(__tokenizer.tokenizer)
data_collator = DataCollatorForSeq2Seq(
__tokenizer.tokenizer,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
else:
data_collator = TransformersDataCollatorForLanguageModeling(__tokenizer.tokenizer, mlm = False, mlm_probability = 0.0)
data_collator = TransformersDataCollatorForLanguageModeling(
__tokenizer.tokenizer,
mlm = False,
mlm_probability = 0.0,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
other_metrics = []
from unsloth_zoo.logging_utils import PatchRLStatistics