bump transformers for fsdp-grad-accum fix, remove patch (axolotl-ai-c…

…loud#2079)

requirements.txt

@@ -1,12 +1,12 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.13.2
-transformers==4.46.2
+transformers==4.46.3
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
 accelerate==1.1.0
 datasets==3.1.0
-deepspeed==0.15.3
+deepspeed==0.15.4
 pydantic==2.6.3
 addict
 fire

src/axolotl/utils/trainer.py

@@ -16,9 +16,6 @@
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
-from axolotl.monkeypatch.trainer_fsdp_grad_accum import (
-    patch_training_loop_for_fsdp_grad_accum,
-)
 from axolotl.utils.distributed import reduce_and_broadcast
 from axolotl.utils.environment import check_cuda_p2p_ib_support
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
@@ -496,11 +493,6 @@ def prepare_opinionated_env(cfg):
 def setup_trainer(
     cfg, train_dataset, eval_dataset, model, tokenizer, processor, total_num_steps
 ):
-    if cfg.fsdp:
-        try:
-            patch_training_loop_for_fsdp_grad_accum()
-        except AssertionError:
-            pass
     if cfg.rl in ["dpo", "ipo", "orpo", "kto", "simpo"]:
         trainer_builder = HFRLTrainerBuilder(cfg, model[0], tokenizer, processor)
         trainer_builder.model_ref = model[1]