Merge branch 'hotfix/model-device-auto-error' into 'export/v1-2-2'

Fix `device="auto"` crash in evaluation

See merge request onecomp/onecomp-lab!96

Author: FKKimura

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Change log
 
+## [v1.2.2] 2026-07-10
+
+### Bug Fix
+
+- Fixed `device="auto"` evaluation crashes in `Runner.calculate_perplexity()` and related paths by adding `ModelConfig.get_device()` and using a resolved `torch.device` for PyTorch operations such as `model.to()` and `empty_cache()`. Hugging Face `device_map="auto"` remains unchanged for model loading, but PyTorch no longer receives the raw `"auto"` string.
+
 ## [v1.2.1] 2026-07-03
 
 ### Security

onecomp/model_config.py

@@ -11,6 +11,7 @@
 import torch
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
 
+from .utils.device import get_default_device
 from .utils.dtype import needs_bfloat16
 
 try:
@@ -114,6 +115,12 @@ def load_model(self, device_map=None):
         self.logger.info("Model loaded with dtype=%s", next(model.parameters()).dtype)
         return model
 
+    def get_device(self) -> torch.device:
+        """Return a concrete torch.device for PyTorch operations."""
+        if self.device == "auto":
+            return get_default_device()
+        return torch.device(self.device)
+
     def load_tokenizer(self):
         """Load the tokenizer"""
 

onecomp/runner.py

@@ -327,7 +327,7 @@ def check(self):
 
         # MPS device validation: only GPTQ (or AutoBitQuantizer whose
         # candidates are all GPTQ, without DBF fallback) is supported on MPS
-        device = self.model_config.device
+        device = self.model_config.get_device()
         if is_mps_device(device):
             if self.multi_gpu:
                 raise ValueError("multi_gpu is not supported on MPS device.")
@@ -1139,24 +1139,24 @@ def _calculate_evaluation(
             tokenizer = self.model_config.load_tokenizer()
             original_result = eval_function(model=model, tokenizer=tokenizer, **eval_args)
             del model, tokenizer
-            empty_cache(self.model_config.device)
+            empty_cache(self.model_config.get_device())
 
         if quantized_model:
             try:
                 logger.info("Evaluating quantized model (%s)...", eval_name)
                 if self.quantized_model is not None:
                     model = self.quantized_model
-                    model.to(self.model_config.device)
+                    model.to(self.model_config.get_device())
                     tokenizer = self.model_config.load_tokenizer()
                     quantized_result = eval_function(model=model, tokenizer=tokenizer, **eval_args)
                     model.to("cpu")
                     del tokenizer
                 else:
                     model, tokenizer = self.create_quantized_model(quantizer=quantizer)
-                    model.to(self.model_config.device)
+                    model.to(self.model_config.get_device())
                     quantized_result = eval_function(model=model, tokenizer=tokenizer, **eval_args)
                     del model, tokenizer
-                empty_cache(self.model_config.device)
+                empty_cache(self.model_config.get_device())
             except NotImplementedError:
                 logger.warning(
                     "This quantization method does not support creating a quantized model; "
@@ -1171,7 +1171,7 @@ def _calculate_evaluation(
             self.update_model_weights(model, quantizer=quantizer)
             dequantized_result = eval_function(model=model, tokenizer=tokenizer, **eval_args)
             del model, tokenizer
-            empty_cache(self.model_config.device)
+            empty_cache(self.model_config.get_device())
 
         return original_result, dequantized_result, quantized_result
 
@@ -2117,7 +2117,7 @@ def analyze_cumulative_error(
             )
             # Release fragmented GPU memory from previous operations (e.g., run())
             gc.collect()
-            empty_cache(self.model_config.device)
+            empty_cache(self.model_config.get_device())
 
             model = self.model_config.load_model()
             input_device = next(model.parameters()).device
@@ -2141,7 +2141,7 @@ def analyze_cumulative_error(
                 )
                 # Release fragmented GPU memory from previous operations (e.g., run())
                 gc.collect()
-                empty_cache(self.model_config.device)
+                empty_cache(self.model_config.get_device())
 
                 model = self.model_config.load_model()
                 input_device = next(model.parameters()).device