[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
+30 -39
View File
@@ -10,6 +10,7 @@ Note: Only symmetric GPTQ models are supported after consolidation to Marlin.
import pytest
import torch
from tests.quantization.utils import load_model_without_vllm_runner
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQLinearMethod
from vllm.model_executor.layers.quantization.utils.gptq_utils import (
@@ -28,48 +29,38 @@ MODELS = [
@pytest.mark.parametrize("model_id", MODELS)
def test_gptq_with_dynamic(vllm_runner, model_id: str, monkeypatch):
def test_gptq_with_dynamic(model_id: str, monkeypatch, dist_init, workspace_init):
# `LLM.apply_model` requires pickling a function.
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
linear_method_cls = AutoGPTQLinearMethod
with vllm_runner(
model_id, dtype=torch.float16, max_model_len=2048, enforce_eager=True
) as llm:
model, _ = load_model_without_vllm_runner(
model_id,
dtype=torch.float16,
model_config_kwargs={
"max_model_len": 2048,
"hf_overrides": {"num_hidden_layers": 3},
},
)
def check_model(model):
for name, submodule in model.named_modules():
if name == "lm_head":
assert isinstance(submodule.quant_method, linear_method_cls)
elif name == "model.layers.0.self_attn.qkv_proj":
# The first layer is quantized using bits=4, group_size=128
# desc_act=True
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert config.weight_bits == 4
assert config.group_size == 128
assert config.desc_act
elif name == "model.layers.1.self_attn.qkv_proj":
# The second layer is quantized using bits=8, group_size=32
# desc_act=False
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert (
get_dynamic_override(config, layer_name=name, key="bits") == 8
)
assert (
get_dynamic_override(config, layer_name=name, key="group_size")
== 32
)
assert not get_dynamic_override(
config, layer_name=name, key="desc_act"
)
elif (
name == "model.layers.2.self_attn.qkv_proj"
or name == "model.layers.2.mlp.gate_up_proj"
):
# All other layers (layer index >= 2) are not quantized
assert isinstance(submodule.quant_method, UnquantizedLinearMethod)
llm.apply_model(check_model)
for name, submodule in model.named_modules():
if name == "lm_head":
assert isinstance(submodule.quant_method, linear_method_cls)
elif name == "model.layers.0.self_attn.qkv_proj":
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert config.weight_bits == 4
assert config.group_size == 128
assert config.desc_act
elif name == "model.layers.1.self_attn.qkv_proj":
assert isinstance(submodule.quant_method, linear_method_cls)
config = submodule.quant_method.quant_config
assert get_dynamic_override(config, layer_name=name, key="bits") == 8
assert get_dynamic_override(config, layer_name=name, key="group_size") == 32
assert not get_dynamic_override(config, layer_name=name, key="desc_act")
elif name in (
"model.layers.2.self_attn.qkv_proj",
"model.layers.2.mlp.gate_up_proj",
):
assert isinstance(submodule.quant_method, UnquantizedLinearMethod)