Deprecate oneshot_device arg in favor of device_map

src/llmcompressor/args/model_arguments.py

@@ -1,5 +1,7 @@
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import Dict, Optional, Union
+
+import torch
 
 
 @dataclass
@@ -80,8 +82,11 @@ class ModelArguments:
         metadata={"help": "Whether to compress sparse models during save"},
     )
     oneshot_device: Optional[str] = field(
-        default="cuda:0",
-        metadata={"help": "Device to run oneshot calibration on"},
+        default=None,
+        metadata={
+            "help": "This field is deprecated, please use `device_map` instead. "
+            "Device to run oneshot calibration on"
+        },
     )
     model_revision: str = field(
         default="main",
@@ -90,3 +95,6 @@ class ModelArguments:
             "(can be a branch name, tag name or commit id)"
         },
     )
+    device_map: Dict[str, Union[int, str, torch.device]] = field(
+        default="auto", metadata={}
+    )

src/llmcompressor/pytorch/model_load/helpers.py

@@ -16,7 +16,6 @@
     "initialize_recipe",
     "save_model_and_recipe",
     "copy_python_files_from_model_cache",
-    "fallback_to_cpu",
     "parse_dtype",
     "get_session_model",
     "get_completed_stages",
@@ -87,22 +86,6 @@ def save_model_and_recipe(
     copy_python_files_from_model_cache(model, save_path)
 
 
-def fallback_to_cpu(device: str) -> str:
-    """
-    Takes in a device string and forces it to cpu if cuda is not available
-
-    :param device: device id to check
-    :return: device modified for CUDA status
-    """
-    if "cuda" in device and not torch.cuda.is_available():
-        logger.warning(
-            f"Requested {device} but CUDA is not available, falling back to CPU"
-        )
-        return "cpu"
-
-    return device
-
-
 def parse_dtype(dtype_arg: Union[str, torch.dtype]) -> torch.dtype:
     """
     :param dtype_arg: dtype or string to parse

src/llmcompressor/transformers/finetune/text_generation.py

@@ -41,7 +41,6 @@
 )
 from llmcompressor.core import pre_initialize_structure, reset_session
 from llmcompressor.pytorch.model_load.helpers import (
-    fallback_to_cpu,
     get_session_model,
     initialize_recipe,
     parse_dtype,
@@ -150,10 +149,19 @@ def parse_args(**kwargs):
     # raise depreciation warnings
     if data_args.remove_columns is not None:
         warnings.warn(
-            "`remove_columns` argument is depreciated. When tokenizing datasets, all "
+            "`remove_columns` argument is deprecated. When tokenizing datasets, all "
             "columns which are invalid inputs the tokenizer will be removed",
             DeprecationWarning,
         )
+    if model_args.oneshot_device is not None:
+        warnings.warn(
+            "`oneshot_device` argument is deprecated and will no longer fall back to "
+            "CPU. Please use `device_map` instead",
+            DeprecationWarning,
+        )
+        device_map_default = ModelArguments.__dataclass_fields__["device_map"].default
+        if model_args.device_map == device_map_default:
+            model_args.device_map = model_args.oneshot_device
 
     # silently assign tokenizer to processor
     if model_args.tokenizer:
@@ -233,25 +241,20 @@ def initialize_model_from_path(
         else model_args.model_name_or_path
     )
 
-    # Fallback to CPU if GPU requested and not available
-    model_args.oneshot_device = fallback_to_cpu(model_args.oneshot_device)
-
     # Trainer handles device assignment for FSDP and training, don't do mapping here
     # if running oneshot outside of FSDP, apply user device settings
-
     fsdp_enabled = os.environ.get("ACCELERATE_USE_FSDP", "false") == "true"
-
-    device_map = model_args.oneshot_device
     if not fsdp_enabled and training_args is not None and training_args.do_train:
-        device_map = "auto"
+        logger.log("Detected FSDP training, setting device map to `auto`")
+        model_args.device_map = "auto"
 
     model_kwargs = {
         "config": config,
         "cache_dir": model_args.cache_dir,
         "revision": model_args.model_revision,
         "use_auth_token": True if model_args.use_auth_token else None,
         "torch_dtype": parse_dtype(model_args.precision),
-        "device_map": device_map,
+        "device_map": model_args.device_map,
         "trust_remote_code": model_args.trust_remote_code_model,
     }