[CI] Speed up quantization test group (#53291)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
fxmarty-amd
2026-09-01 13:15:24 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent e7cf4730d6
commit d1c15e589d
20 changed files with 824 additions and 725 deletions
+16 -20
View File
@@ -8,6 +8,7 @@ Run `pytest tests/quantization/test_quant_lm_head_true.py --forked`.
import pytest
import torch
from tests.quantization.utils import load_model_without_vllm_runner
from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQLinearMethod
from vllm.model_executor.layers.vocab_parallel_embedding import (
UnquantizedEmbeddingMethod,
@@ -23,29 +24,24 @@ MODELS_QUANT = [
@pytest.mark.parametrize("model_id, lm_head_quantized", MODELS_QUANT)
def test_lm_head(
vllm_runner,
model_id: str,
lm_head_quantized: bool,
monkeypatch,
dist_init,
workspace_init,
) -> None:
# `LLM.apply_model` requires pickling a function.
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
with vllm_runner(
model_id, dtype=torch.float16, max_model_len=2048, enforce_eager=True
) as vllm_model:
def check_model(model):
lm_head_layer = model.lm_head
if lm_head_quantized:
assert isinstance(
lm_head_layer.quant_method,
AutoGPTQLinearMethod,
)
else:
assert isinstance(
lm_head_layer.quant_method, UnquantizedEmbeddingMethod
)
vllm_model.apply_model(check_model)
print(vllm_model.generate_greedy(["Hello my name is"], max_tokens=4)[0][1])
model, _ = load_model_without_vllm_runner(
model_id,
dtype=torch.float16,
model_config_kwargs={
"max_model_len": 2048,
"hf_overrides": {"num_hidden_layers": 3},
},
)
lm_head_layer = model.lm_head
if lm_head_quantized:
assert isinstance(lm_head_layer.quant_method, AutoGPTQLinearMethod)
else:
assert isinstance(lm_head_layer.quant_method, UnquantizedEmbeddingMethod)