fix zero3 integration (#1897)

* fix zero3 integration

* bump transformers and accelerate too

requirements.txt

@@ -1,10 +1,10 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.12.0
-transformers==4.44.0
+transformers==4.44.2
 tokenizers>=0.19.1
 bitsandbytes==0.43.3
-accelerate==0.33.0
+accelerate==0.34.0
 datasets==2.20.0
 deepspeed==0.14.4
 pydantic==2.6.3

src/axolotl/utils/trainer.py

@@ -425,7 +425,8 @@ def setup_deepspeed_env(cfg, stage=None):
         os.environ["ACCELERATE_DEEPSPEED_ZERO_STAGE"] = str(stage)
         if stage == 3:
             os.environ["ACCELERATE_DEEPSPEED_ZERO3_INIT"] = "true"
-    HfTrainerDeepSpeedConfig(cfg.deepspeed)
+    # If we don't assign this, it doesn't actually get set in the accelerate weakref
+    _ = HfTrainerDeepSpeedConfig(cfg.deepspeed)
 
 
 def setup_fsdp_envs(cfg):