instruct model setup

unsloth_compiled_cache/UnslothPPOTrainer.py

@@ -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__
 """
@@ -106,6 +106,7 @@ class UnslothPPOConfig(PPOConfig):
         default = -1,
         metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
     )
+    
     def __init__(
         self,
         output_dir = None,
@@ -270,6 +271,7 @@ class UnslothPPOConfig(PPOConfig):
         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!')
@@ -279,7 +281,7 @@ class UnslothPPOConfig(PPOConfig):
             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:
@@ -449,6 +451,7 @@ class UnslothPPOConfig(PPOConfig):
             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 _UnslothPPOTrainer(Trainer):
@@ -944,7 +947,7 @@ class _UnslothPPOTrainer(Trainer):
                                 pg_clipfrac = masked_mean(
                                     (pg_losses2 > pg_losses).float(), ~padding_mask[micro_batch_inds]
                                 )
-                                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
@@ -1202,7 +1205,7 @@ class UnslothPPOTrainer(_UnslothPPOTrainer):
             print('Unsloth: Switching to float32 training since model cannot work with float16')
             force_float32 = True
         mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
-        dtype = getattr(model.config, 'torch_dtype', None)
+        dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
         if dtype is None: dtype = model.get_input_embeddings().dtype
         from unsloth_zoo.utils import _get_dtype
         dtype = _get_dtype(dtype)
@@ -1268,9 +1271,17 @@ class UnslothPPOTrainer(_UnslothPPOTrainer):
         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 = ''
@@ -1278,9 +1289,17 @@ class UnslothPPOTrainer(_UnslothPPOTrainer):
         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