[ Misc ] Refactor Marlin Python Utilities (#6082)

Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>

tests/quantization/test_compressed_tensors.py

@@ -6,7 +6,6 @@ Run `pytest tests/quantization/test_compressed_tensors.py`.
 import pytest
 import torch
 
-from vllm import SamplingParams
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
     CompressedTensorsLinearMethod, CompressedTensorsW4A16Sparse24,
     CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
@@ -57,12 +56,14 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
         assert qkv_proj.weight_scale.dtype is torch.float32
         assert qkv_proj.input_scale.dtype is torch.float32
 
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
+        assert output
+
 
 def test_compressed_tensors_no_enforce_eager(vllm_runner):
     model_path = "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change"
     with vllm_runner(model_path) as llm:
-        sampling_params = SamplingParams()
-        output = llm.generate("Hello world!", sampling_params=sampling_params)
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
         assert output
 
 
@@ -84,13 +85,16 @@ def test_compressed_tensors_w8a8_dynanmic_per_token(vllm_runner, model_args):
         assert qkv_proj.scheme.strategy == strategy
         assert qkv_proj.weight.dtype is torch.int8
 
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
+        assert output
+
 
 @pytest.mark.parametrize(
     "wNa16_args",
     [("nm-testing/tinyllama-oneshot-w4a16-channel-v2", "channel", None, 8),
      ("nm-testing/tinyllama-oneshot-w4a16-group128-v2", "group", 128, 8),
      ("nm-testing/tinyllama-oneshot-w8a16-per-channel", "channel", None, 4)])
-def test_compressed_tensors_w4a16(vllm_runner, wNa16_args):
+def test_compressed_tensors_wNa16(vllm_runner, wNa16_args):
     model, strategy, group, pack_factor = wNa16_args
     with vllm_runner(model) as llm:
         model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model  # noqa: E501
@@ -101,12 +105,15 @@ def test_compressed_tensors_w4a16(vllm_runner, wNa16_args):
         assert isinstance(qkv_proj.scheme, CompressedTensorsWNA16)
 
         assert qkv_proj.scheme.strategy == strategy
-        assert qkv_proj.scheme.group_size == group
+        assert qkv_proj.scheme.group_size == (-1 if group is None else group)
 
         assert qkv_proj.weight_packed.dtype is torch.int32
         assert qkv_proj.weight_scale.dtype is torch.float16
         assert qkv_proj.weight_packed.pack_factor == pack_factor
 
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
+        assert output
+
 
 def test_compressed_tensors_w4a16_marlin24(vllm_runner):
     model_path = "nm-testing/llama7b-one-shot-2_4-w4a16-marlin24-t"
@@ -120,8 +127,7 @@ def test_compressed_tensors_w4a16_marlin24(vllm_runner):
         assert isinstance(qkv_proj.scheme, CompressedTensorsW4A16Sparse24)
         assert qkv_proj.weight_packed.dtype is torch.int32
 
-        sampling_params = SamplingParams()
-        output = llm.generate("Hello world!", sampling_params=sampling_params)
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
         assert output
 
 
@@ -142,6 +148,5 @@ def test_compressed_tensors_fp8(vllm_runner):
         assert len(qkv_proj.input_scale.shape) == 0
         assert len(qkv_proj.weight_scale.shape) == 0
 
-        sampling_params = SamplingParams()
-        output = llm.generate("Hello world!", sampling_params=sampling_params)
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
         assert output