[CI/Build][REDO] Add is_quant_method_supported to control quantization test configurations (#5466)

This commit is contained in:
Michael Goin
2024-06-13 15:18:08 +00:00
committed by GitHub
parent c2637a613b
commit 23ec72fa03
8 changed files with 32 additions and 71 deletions
+2 -10
View File
@@ -8,8 +8,8 @@ import pytest
import torch
from transformers import AutoTokenizer
from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
os.environ["TOKENIZERS_PARALLELISM"] = "true"
@@ -67,16 +67,8 @@ EXPECTED_STRS_MAP = {
},
}
fp8_not_supported = True
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,
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
reason="fp8 is not supported on this GPU type.")
@pytest.mark.parametrize("model_name", MODELS)
@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8"])