Move: recipe parsing test from e2e/ to main test suite

tests/e2e/recipe.yaml → tests/llmcompressor/recipe/recipe.yaml

@@ -8,10 +8,6 @@ quant_stage:
       - - ['re:.*gate_proj', 're:.*up_proj']
         - re:.*post_attention_layernorm
     GPTQModifier:
-      sequential_update: false
+      targets: ["Linear"]
       ignore: [lm_head]
-      config_groups:
-        group_0:
-          weights: {num_bits: 8, type: int, symmetric: true, strategy: channel}
-          input_activations: {num_bits: 8, symmetric: false}
-          targets: [Linear]
+      scheme: W8A8

tests/llmcompressor/recipe/test_recipe_parsing.py

@@ -0,0 +1,98 @@
+from pathlib import Path
+
+import pytest
+from transformers import AutoModelForCausalLM
+
+from llmcompressor import oneshot
+from llmcompressor.core.session_functions import reset_session
+from llmcompressor.modifiers.quantization.gptq import GPTQModifier
+from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
+from llmcompressor.modifiers.smoothquant.utils import DEFAULT_SMOOTHQUANT_MAPPINGS
+from tests.testing_utils import requires_gpu
+
+
+@pytest.fixture
+def setup_model_and_config(tmp_path):
+    """
+    Loads a test model and returns common arguments used in oneshot runs.
+    """
+    model = AutoModelForCausalLM.from_pretrained(
+        "Xenova/llama2.c-stories110M",
+        device_map="auto",
+        torch_dtype="auto",
+    )
+
+    return {
+        "model": model,
+        "dataset": "ultrachat-200k",
+        "output_dir": str(tmp_path / "compressed_output"),
+        "splits": {"calibration": "train_gen[:10]"},
+        "max_seq_length": 2048,
+        "pad_to_max_length": False,
+        "num_calibration_samples": 8,
+    }
+
+
+def recipe_variants():
+    """
+    Returns three ways of defining the compression recipe:
+    - Python modifier objects
+    - YAML string
+    - Path to a YAML file
+    """
+    python_modifiers = [
+        SmoothQuantModifier(
+            smoothing_strength=0.8,
+            mappings=DEFAULT_SMOOTHQUANT_MAPPINGS,
+        ),
+        GPTQModifier(
+            targets="Linear",
+            scheme="W8A8",
+            ignore=["lm_head"],
+        ),
+    ]
+
+    yaml_string = """
+DEFAULT_stage:
+  DEFAULT_modifiers:
+    SmoothQuantModifier:
+      smoothing_strength: 0.8
+      mappings:
+      - - ['re:.*q_proj', 're:.*k_proj', 're:.*v_proj']
+        - re:.*input_layernorm
+      - - ['re:.*gate_proj', 're:.*up_proj']
+        - re:.*post_attention_layernorm
+    GPTQModifier:
+      targets: Linear
+      ignore: [lm_head]
+      scheme: W8A8
+"""
+
+    yaml_path = str(Path(__file__).parent / "recipe.yaml")
+
+    return [python_modifiers, yaml_string, yaml_path]
+
+
+@requires_gpu
+@pytest.mark.regression
+@pytest.mark.parametrize(
+    "recipe",
+    recipe_variants(),
+    ids=["modifier-objects", "yaml-string", "yaml-file"],
+)
+def test_oneshot_accepts_multiple_recipe_formats(setup_model_and_config, recipe):
+    """
+    Integration test: verifies oneshot runs successfully with different recipe formats.
+    """
+    oneshot(
+        recipe=recipe,
+        **setup_model_and_config,
+        save_compressed=True,
+    )
+
+    output_path = Path(setup_model_and_config["output_dir"])
+    assert output_path.exists() and any(
+        output_path.iterdir()
+    ), f"No output artifacts found in: {output_path}"
+
+    reset_session()