[Tests] Start oneshot tests on CPU (vllm-project#1555)

## Purpose ##
* Speed up tests by reducing device movement

## Background ##
As of vllm-project#1263, the model is dispatched to different device maps depending
on which pipelines are used. If the model starts on anything but the
CPU, then these dispatches and undispatches create device movement.
Starting on the CPU will ensure no device movement occurs when offloaded
dispatches happen.

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>

Author: 2 people

tests/llmcompressor/recipe/test_recipe_parsing.py

@@ -18,7 +18,6 @@ def setup_model_and_config(tmp_path):
     """
     model = AutoModelForCausalLM.from_pretrained(
         "Xenova/llama2.c-stories110M",
-        device_map="auto",
         torch_dtype="auto",
     )
 

tests/llmcompressor/transformers/finetune/test_oneshot_then_finetune.py

@@ -21,7 +21,7 @@ def setUp(self):
     def test_oneshot_sparsification_then_finetune(self):
         recipe_str = "tests/llmcompressor/transformers/obcq/recipes/test_tiny2.yaml"
         model = AutoModelForCausalLM.from_pretrained(
-            "nm-testing/llama2.c-stories15M", device_map="auto", torch_dtype="auto"
+            "nm-testing/llama2.c-stories15M", torch_dtype="auto"
         )
         dataset = "open_platypus"
         concatenate_data = False
@@ -47,12 +47,11 @@ def test_oneshot_sparsification_then_finetune(self):
         # Explictly decompress the model for training using quantization_config
         model = AutoModelForCausalLM.from_pretrained(
             self.output / "oneshot_out",
-            device_map="auto",
             torch_dtype="auto",
             quantization_config=self.quantization_config,
         )
         distill_teacher = AutoModelForCausalLM.from_pretrained(
-            "nm-testing/llama2.c-stories15M", device_map="auto", torch_dtype="auto"
+            "nm-testing/llama2.c-stories15M", torch_dtype="auto"
         )
         dataset = "open_platypus"
         concatenate_data = False
@@ -88,7 +87,6 @@ def test_oneshot_sparsification_then_finetune(self):
         # Explictly decompress the model for training using quantization_config
         model = AutoModelForCausalLM.from_pretrained(
             output_dir,
-            device_map="auto",
             torch_dtype="auto",
             quantization_config=self.quantization_config,
         )
@@ -112,7 +110,7 @@ def test_oneshot_quantization_then_finetune(self):
         )
 
         model = AutoModelForCausalLM.from_pretrained(
-            "TinyLlama/TinyLlama-1.1B-Chat-v1.0", device_map="auto", torch_dtype="auto"
+            "TinyLlama/TinyLlama-1.1B-Chat-v1.0", torch_dtype="auto"
         )
         dataset = "open_platypus"
         concatenate_data = False
@@ -136,7 +134,6 @@ def test_oneshot_quantization_then_finetune(self):
         quantization_config = CompressedTensorsConfig(run_compressed=False)
         model = AutoModelForCausalLM.from_pretrained(
             output_dir,
-            device_map="auto",
             torch_dtype="auto",
             quantization_config=quantization_config,
         )
@@ -159,7 +156,6 @@ def test_oneshot_quantization_then_finetune(self):
         # test reloading checkpoint and final model
         model = AutoModelForCausalLM.from_pretrained(
             output_dir,
-            device_map="auto",
             torch_dtype="auto",
             quantization_config=quantization_config,
         )

tests/llmcompressor/transformers/obcq/test_consecutive_runs.py

@@ -44,7 +44,6 @@ def _test_consecutive_runs(
 
         first_model = AutoModelForCausalLM.from_pretrained(
             self.output_first,
-            device_map="auto",
             torch_dtype="auto",
             quantization_config=self.quantization_config,
         )
@@ -74,7 +73,6 @@ def _test_consecutive_runs(
         second_model = AutoModelForCausalLM.from_pretrained(
             self.output_second,
             quantization_config=self.quantization_config,
-            device_map="auto",
             torch_dtype="auto",
         )