init

tests/quantization/test_configs.py

@@ -0,0 +1,73 @@
+"""Tests whether Marlin models can be loaded from the autogptq config.
+
+Run `pytest tests/quantization/test_configs.py --forked`.
+"""
+
+from dataclasses import dataclass
+
+import pytest
+
+from vllm.config import ModelConfig
+
+
+@dataclass
+class ModelPair:
+    model_marlin: str
+    model_gptq: str
+
+
+# Model Id // Quantization Arg // Expected Type
+MODEL_ARG_EXPTYPES = [
+    # AUTOGPTQ
+    # compat: autogptq <=0.7.1 is_marlin_format: bool
+    # Model Serialized in Marlin Format should always use Marlin kernel.
+    ("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", None, "marlin"),
+    ("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "marlin", "marlin"),
+    ("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "gptq", "marlin"),
+    ("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "awq", "ERROR"),
+    # Model Serialized in Exllama Format.
+    ("TheBloke/Llama-2-7B-Chat-GPTQ", None, "gptq_marlin"),
+    ("TheBloke/Llama-2-7B-Chat-GPTQ", "marlin", "gptq_marlin"),
+    ("TheBloke/Llama-2-7B-Chat-GPTQ", "gptq", "gptq"),
+    ("TheBloke/Llama-2-7B-Chat-GPTQ", "awq", "ERROR"),
+    # compat: autogptq >=0.8.0 use checkpoint_format: str
+    # Model Serialized in Marlin Format should always use Marlin kernel.
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", None, "marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "marlin", "marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "gptq", "marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "awq", "ERROR"),
+    # Model Serialized in Exllama Format.
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", None, "gptq_marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "marlin", "gptq_marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "gptq", "gptq"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "awq", "ERROR"),
+
+    # AUTOAWQ
+    ("TheBloke/OpenHermes-2.5-Mistral-7B-AWQ", None, "awq"),
+    ("TheBloke/OpenHermes-2.5-Mistral-7B-AWQ", "awq", "awq"),
+    ("TheBloke/OpenHermes-2.5-Mistral-7B-AWQ", "marlin", "ERROR"),
+    ("TheBloke/OpenHermes-2.5-Mistral-7B-AWQ", "gptq", "ERROR"),
+]
+
+
+@pytest.mark.parametrize("model_arg_exptype", MODEL_ARG_EXPTYPES)
+def test_auto_gptq(model_arg_exptype: str) -> None:
+    model_path, quantization_arg, expected_type = model_arg_exptype
+
+    try:
+        model_config = ModelConfig(model_path,
+                                   model_path,
+                                   tokenizer_mode="auto",
+                                   trust_remote_code=False,
+                                   seed=0,
+                                   dtype="float16",
+                                   revision=None,
+                                   quantization=quantization_arg)
+        found_quantization_type = model_config.quantization
+    except ValueError:
+        found_quantization_type = "ERROR"
+
+    assert found_quantization_type == expected_type, (
+        f"Expected quant_type == {expected_type} for {model_path}, "
+        f"but found {found_quantization_type} "
+        f"for no --quantization {quantization_arg} case")

tests/quantization/test_fp8.py

@@ -0,0 +1,24 @@
+"""Tests whether FP8 computation is enabled correctly.
+
+Run `pytest tests/quantization/test_fp8.py --forked`.
+"""
+import pytest
+import torch
+
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
+from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
+
+capability = torch.cuda.get_device_capability()
+capability = capability[0] * 10 + capability[1]
+
+
+@pytest.mark.skipif(
+    capability < QUANTIZATION_METHODS["fp8"].get_min_capability(),
+    reason="FP8 is not supported on this GPU type.")
+def test_load_fp16_model(vllm_runner) -> None:
+    llm = vllm_runner("facebook/opt-125m", quantization="fp8")
+
+    model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model
+    fc1 = model.model.decoder.layers[0].fc1
+    assert isinstance(fc1.quant_method, Fp8LinearMethod)
+    assert fc1.weight.dtype == torch.float8_e4m3fn