Merge pull request #8 from YerevaNN/fix_ci_tests

Fix ci tests

chemlactica/config/test_configs/fsdp_config.yaml

@@ -1,13 +1,15 @@
 compute_environment: LOCAL_MACHINE
+debug: false
 distributed_type: FSDP
 downcast_bf16: 'no'
 fsdp_config:
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_backward_prefetch: BACKWARD_PRE
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_forward_prefetch: false
   fsdp_offload_params: false
-  fsdp_sharding_strategy: 1
+  fsdp_sharding_strategy: FULL_SHARD
   fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: OPTForCausalLM
 machine_rank: 0
 main_training_function: main
 mixed_precision: bf16
@@ -19,4 +21,3 @@ tpu_env: []
 tpu_use_cluster: false
 tpu_use_sudo: false
 use_cpu: false
-main_process_port: 30001

chemlactica/custom_trainer.py

@@ -1,14 +1,10 @@
-import shutil
 import torch
 import submitit
 from typing import Any, Dict
 import os
 from torch._tensor import Tensor
 from torch.nn.modules import Module
 from transformers import Trainer, TrainingArguments
-from torch.distributed.fsdp.fully_sharded_data_parallel import (
-    FullyShardedDataParallel as FSDP,
-)
 from chemlactica.utils.utils import get_tokenizer
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
 from transformers.utils import is_torch_tpu_available
@@ -41,23 +37,6 @@ def training_step(self, model: Module, inputs: Dict[str, Tensor | Any]) -> Tenso
             self.num_samples_to_print = None
         return super().training_step(model, inputs)
 
-    def _save_checkpoint(self, model, trial, metrics=None):
-        if shutil.disk_usage("/").free > 3 * 1024**3:
-            super()._save_checkpoint(model, trial, metrics=None)
-        else:
-            print("**disk is full didn't save**")
-
-    def _load_from_checkpoint(self, resume_from_checkpoint, model=None):
-        """
-        This code is added because we had a failure when resuming training.
-        Basically, we load the model with fsdp when the model is not fsdp wrapped.
-        In the future versions transformers this issue is handled, by adding an extra check,
-        but not in 4.31.0 version. So this is our manual check addition to solve the problem.
-        """
-        if type(self.model) != FSDP:
-            return
-        return super()._load_from_checkpoint(resume_from_checkpoint, model)
-
     def _build_slurm_eval_command(self, train_command, trial):
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
         run_dir = self._get_output_dir(trial=trial)

chemlactica/utils/model_utils.py

@@ -90,11 +90,10 @@ def load_model(
         model = OPTForCausalLM.from_pretrained(
             from_pretrained, torch_dtype=dtype, attn_implementation=attn_implementation
         )
-        print(type(model.lm_head))
-        model.lm_head = float_casting_decorator(model.lm_head.__class__)(
-            in_features=model.lm_head.in_features,
-            out_features=model.lm_head.out_features,
-        )
+        # model.lm_head = float_casting_decorator(model.lm_head.__class__)(
+        #     in_features=model.lm_head.in_features,
+        #     out_features=model.lm_head.out_features,
+        # )
         # model.lm_head.forward = cast_to_fp32(OPTForCausalLM.lm_head.forward)
 
     if "mistral" in from_pretrained.lower():

test_status.yaml

@@ -1 +1 @@
-86eb89a6651607c56835456eb9d2f0ae6cd222cc: PASS
+d438c06d55ece4b8ccd55e414172079d0e98ef7f: PASS