Merge pull request #96 from zzasdf/dev-zzasdf

fix for issue #93 and a memory hack

src/slam_llm/pipeline/finetune_deepspeed.py

@@ -145,6 +145,20 @@ def main(kwargs: DictConfig):
     parameters = filter(lambda p: p.requires_grad, model.parameters())
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
+    # If you are facing problem from limited memory(<=256GB), you can try to replace the above code with the following code
+    # for i in range(rank):
+    #     while not os.path.isfile(f".{i}.done"):
+    #         pass
+    # assert not os.path.isfile(f".{rank}.done"), f".{rank}.done already exists!"
+    # model_factory = get_custom_model_factory(model_config, logger)
+    # model, tokenizer = model_factory(train_config, model_config, **kwargs)
+    # parameters = filter(lambda p: p.requires_grad, model.parameters())
+    # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    # model.half()
+    # with open(f".{rank}.done", "w"):
+    #     pass
+
+
     # Initialize the optimizer and learning rate scheduler
     model_engine, _, _, _ = deepspeed.initialize(
         model=model, model_parameters=parameters, config=deepspeed_config

src/slam_llm/utils/config_utils.py

@@ -69,7 +69,7 @@ def get_dataloader_kwargs(train_config, dataset, tokenizer, mode):
         kwargs = {}
         batch_size = train_config.batch_size_training if mode=="train" else train_config.val_batch_size
         if train_config.batching_strategy == "padding":
-            if train_config.enable_fsdp or train_config.enable_ddp:
+            if train_config.enable_fsdp or train_config.enable_ddp or train_config.enable_deepspeed:
                 kwargs["batch_sampler"] = DistributedLengthBasedBatchSampler(
                     dataset,
                     batch_size=batch_size,
@@ -81,7 +81,7 @@ def get_dataloader_kwargs(train_config, dataset, tokenizer, mode):
                 kwargs["batch_sampler"] = LengthBasedBatchSampler(dataset, batch_size, drop_last=True, shuffle=mode=="train")
             kwargs["collate_fn"] = DataCollatorForSeq2Seq(tokenizer)
         elif train_config.batching_strategy == "packing":
-            if train_config.enable_fsdp or train_config.enable_ddp:
+            if train_config.enable_fsdp or train_config.enable_ddp or train_config.enable_deepspeed:
                 kwargs["sampler"] = DistributedSampler(
                 dataset,
                 rank=dist.get_rank(),
@@ -93,7 +93,7 @@ def get_dataloader_kwargs(train_config, dataset, tokenizer, mode):
             kwargs["collate_fn"] = default_data_collator
         else:
             # raise ValueError(f"Unknown batching strategy: {train_config.batching_strategy}")
-            if train_config.enable_fsdp or train_config.enable_ddp:
+            if train_config.enable_fsdp or train_config.enable_ddp or train_config.enable_deepspeed:
                 kwargs["sampler"] = DistributedSampler(
                 dataset,
                 rank=dist.get_rank(),