bump transformers 4.57.3

examples/llama-3/diffusion/sft-1b.yaml

@@ -30,7 +30,7 @@ eval_sample_packing: true
 gradient_accumulation_steps: 4
 micro_batch_size: 4
 num_epochs: 1
-warmup_ratio: 0.1
+warmup_steps: 0.1
 
 optimizer: adamw_8bit
 lr_scheduler: cosine
@@ -44,7 +44,7 @@ resume_from_checkpoint:
 sdp_attention: true
 
 logging_steps: 1
-save_strategy: epoch
+save_strategy: best
 eval_strategy: epoch
 
 special_tokens:

requirements.txt

@@ -13,7 +13,7 @@ packaging==23.2
 huggingface_hub>=0.36.0
 peft>=0.18.0
 tokenizers>=0.22.1
-transformers==4.57.1
+transformers==4.57.3
 accelerate==1.11.0
 datasets==4.4.1
 deepspeed>=0.17.0

src/axolotl/core/trainers/base.py

@@ -631,9 +631,7 @@ class AxolotlTrainer(
             logs["tokens_per_second_per_gpu"] = round(
                 self.state.last_tokens_per_second.item() / self.args.logging_steps, 2
             )
-
+            logs["total_tokens"] = int(self.state.total_tokens.item())
-            if hasattr(self.state, "total_tokens"):
-                logs["total_tokens"] = int(self.state.total_tokens.item())
 
         del self._stored_metrics[train_eval]