Revert "[CI/Build] Add is_quant_method_supported to control quantiz…

…ation test configurations" (vllm-project#5463)

tests/models/test_aqlm.py

@@ -4,6 +4,7 @@
 """
 
 import pytest
+import torch
 
 from tests.nm_utils.utils_skip import should_skip_test_group
 from tests.quantization.utils import is_quant_method_supported
@@ -63,7 +64,7 @@
 ]
 
 
-@pytest.mark.skipif(not is_quant_method_supported("aqlm"),
+@pytest.mark.skipif(aqlm_not_supported,
                     reason="AQLM is not supported on this GPU type.")
 @pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])
 @pytest.mark.parametrize("dtype", ["half"])

tests/models/test_fp8.py

@@ -72,8 +72,16 @@
     },
 }
 
+fp8_not_supported = True
 
-@pytest.mark.skipif(not is_quant_method_supported("fp8"),
+if torch.cuda.is_available():
+    capability = torch.cuda.get_device_capability()
+    capability = capability[0] * 10 + capability[1]
+    fp8_not_supported = (capability <
+                         QUANTIZATION_METHODS["fp8"].get_min_capability())
+
+
+@pytest.mark.skipif(fp8_not_supported,
                     reason="fp8 is not supported on this GPU type.")
 @pytest.mark.parametrize("model_name", MODELS)
 @pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8"])

tests/models/test_gptq_marlin.py

@@ -11,6 +11,7 @@
 import os
 
 import pytest
+import torch
 
 from tests.models.utils import check_logprobs_close
 from tests.nm_utils.utils_skip import should_skip_test_group
@@ -25,6 +26,14 @@
 
 MAX_MODEL_LEN = 1024
 
+gptq_marlin_not_supported = True
+
+if torch.cuda.is_available():
+    capability = torch.cuda.get_device_capability()
+    capability = capability[0] * 10 + capability[1]
+    gptq_marlin_not_supported = (
+        capability < QUANTIZATION_METHODS["gptq_marlin"].get_min_capability())
+
 MODELS = [
     # act_order==False, group_size=channelwise
     ("robertgshaw2/zephyr-7b-beta-channelwise-gptq", "main"),
@@ -48,7 +57,7 @@
 
 
 @pytest.mark.flaky(reruns=3)
-@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
+@pytest.mark.skipif(gptq_marlin_not_supported,
                     reason="gptq_marlin is not supported on this GPU type.")
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half", "bfloat16"])

tests/models/test_gptq_marlin_24.py

@@ -9,6 +9,7 @@
 from dataclasses import dataclass
 
 import pytest
+import torch
 
 from tests.models.utils import check_logprobs_close
 from tests.nm_utils.utils_skip import should_skip_test_group
@@ -44,7 +45,7 @@ class ModelPair:
 
 
 @pytest.mark.flaky(reruns=2)
-@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin_24"),
+@pytest.mark.skipif(marlin_not_supported,
                     reason="Marlin24 is not supported on this GPU type.")
 @pytest.mark.parametrize("model_pair", model_pairs)
 @pytest.mark.parametrize("dtype", ["half"])

tests/models/test_marlin.py

@@ -19,6 +19,7 @@
 
 import pytest
 import torch
+import torch
 
 from tests.models.utils import check_logprobs_close
 from tests.nm_utils.utils_skip import should_skip_test_group
@@ -46,7 +47,7 @@ class ModelPair:
 
 
 @pytest.mark.flaky(reruns=2)
-@pytest.mark.skipif(not is_quant_method_supported("marlin"),
+@pytest.mark.skipif(marlin_not_supported,
                     reason="Marlin is not supported on this GPU type.")
 @pytest.mark.parametrize("model_pair", model_pairs)
 @pytest.mark.parametrize("dtype", ["half"])

tests/quantization/test_bitsandbytes.py

@@ -5,12 +5,16 @@
 import pytest
 import torch
 
-from tests.quantization.utils import is_quant_method_supported
 from vllm import SamplingParams
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 
+capability = torch.cuda.get_device_capability()
+capability = capability[0] * 10 + capability[1]
 
-@pytest.mark.skipif(not is_quant_method_supported("bitsandbytes"),
-                    reason='bitsandbytes is not supported on this GPU type.')
+
+@pytest.mark.skipif(
+    capability < QUANTIZATION_METHODS['bitsandbytes'].get_min_capability(),
+    reason='bitsandbytes is not supported on this GPU type.')
 def test_load_bnb_model(vllm_runner) -> None:
     with vllm_runner('huggyllama/llama-7b',
                      quantization='bitsandbytes',