let hf trainer handle torch compile (axolotl-ai-cloud#516)

* let hf trainer handle torch compile

* remove torch compile checks, include option for backend

* suppress torch errors to get further

* require min torch version of 2.1.0 for torch compile to work

---------

Co-authored-by: Aman Karmani <aman@tmm1.net>

README.md

@@ -519,6 +519,10 @@ wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_step
 # where to save the finished model to
 output_dir: ./completed-model
 
+# whether to use torch.compile and which backend to use
+torch_compile:  # bool
+torch_compile_backend:  # Optional[str]
+
 # training hyperparameters
 gradient_accumulation_steps: 1
 micro_batch_size: 2

src/axolotl/train.py

@@ -80,10 +80,6 @@ def train(
 
     model.config.use_cache = False
 
-    if torch.__version__ >= "2" and sys.platform != "win32":
-        LOG.info("Compiling torch model")
-        model = torch.compile(model)
-
     # go ahead and presave, so we have the adapter config available to inspect
     if peft_config:
         LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")

src/axolotl/utils/trainer.py

@@ -11,6 +11,7 @@
 from typing import Optional, Union
 
 import numpy as np
+import torch
 import torch.cuda
 import transformers
 from datasets import Dataset, set_caching_enabled
@@ -604,6 +605,21 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
     if cfg.greater_is_better:
         training_arguments_kwargs["greater_is_better"] = cfg.greater_is_better
 
+    if cfg.torch_compile:
+        if torch.__version__ < "2.1.0":  # pylint: disable=protected-access
+            LOG.warning("torch>=2.1.0 required for torch_compile to work properly")
+        else:
+            import torch._dynamo  # pylint: disable=redefined-outer-name
+
+            torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
+                True
+            )
+            training_arguments_kwargs["torch_compile"] = cfg.torch_compile
+            if cfg.torch_compile_backend:
+                training_arguments_kwargs[
+                    "torch_compile_backend"
+                ] = cfg.torch_compile_backend
+
     # DDP Config
     if cfg.ddp_timeout:
         training_arguments_kwargs["ddp_timeout"] = cfg.ddp_timeout