PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
# Copyright (c) 2026 Intel Corporation
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import shutil
import pytest
import torch
from transformers import AutoModelForCausalLM
from auto_round import AutoRound
from auto_round.experimental.qmodules.fake import FakeActQuantLinear
from auto_round.wrapper import WrapperWALayer
pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason="test requires CUDA")
class TestActQuantizationGpu:
@pytest.fixture(autouse=True)
def _save_dir(self, tmp_path):
self.save_dir = str(tmp_path / "saved")
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
def _reload_and_forward(self, quantized_model_path):
model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map="cuda:0", trust_remote_code=True)
assert isinstance(model, torch.nn.Module)
input_ids = torch.randint(0, 1000, (1, 16), device="cuda:0")
with torch.no_grad():
out = model(input_ids)
assert out.logits.shape[0] == 1
def test_w4a8_act_dynamic(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
bits=4,
act_bits=8,
group_size=128,
act_group_size=32,
act_dynamic=True,
sym=True,
iters=1,
nsamples=2,
seqlen=16,
)
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
def test_w8a8_act_dynamic(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
bits=8,
act_bits=8,
group_size=128,
act_group_size=32,
act_dynamic=True,
sym=True,
iters=0,
disable_opt_rtn=True,
nsamples=2,
seqlen=16,
)
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
def test_w4a8_act_static(self, tiny_opt_model_path):
"""Static (non-dynamic) activation quantization path."""
autoround = AutoRound(
tiny_opt_model_path,
bits=4,
act_bits=8,
group_size=128,
act_group_size=128,
act_dynamic=False,
sym=True,
iters=0,
disable_opt_rtn=True,
nsamples=2,
seqlen=16,
)
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
def test_w4a8_act_group128(self, tiny_opt_model_path):
"""Wider act_group_size exercises a different act-hook grouping path."""
autoround = AutoRound(
tiny_opt_model_path,
bits=4,
act_bits=8,
group_size=128,
act_group_size=128,
act_dynamic=True,
sym=True,
iters=1,
nsamples=2,
seqlen=16,
)
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
def test_w8a8_asym(self, tiny_opt_model_path):
"""Asymmetric W8A8 cannot be exported to the auto_round format.
``auto_round`` supports W8A8 export only for the symmetric path; an
asymmetric W8A8 config is rejected when the scheme is resolved (the
default export format has no room for an 8-bit zero point).
"""
with pytest.raises(ValueError, match="8-bit asymmetric weight quantization is not supported"):
AutoRound(
tiny_opt_model_path,
bits=8,
act_bits=8,
group_size=128,
act_group_size=32,
act_dynamic=True,
sym=False,
iters=0,
disable_opt_rtn=True,
nsamples=2,
seqlen=16,
)
def test_w4a8_reload_layer_type_diff_but_output_equal(self, tiny_opt_model_path):
"""Save path keeps wrappers; reload path materializes FakeActQuantLinear, outputs should match."""
old_offload = os.environ.get("AR_DISABLE_OFFLOAD")
os.environ["AR_DISABLE_OFFLOAD"] = "1"
try:
autoround = AutoRound(
tiny_opt_model_path,
bits=4,
act_bits=8,
group_size=128,
act_group_size=32,
act_dynamic=True,
sym=True,
iters=0,
disable_opt_rtn=True,
nsamples=2,
seqlen=16,
)
quantized_model, quantized_model_path = autoround.quantize_and_save(
output_dir=self.save_dir, format="auto_round"
)
quantized_model.eval()
reloaded_model = AutoModelForCausalLM.from_pretrained(
quantized_model_path, device_map="cuda:0", trust_remote_code=True
)
reloaded_model.eval()
src_layer = quantized_model.model.decoder.layers[0].fc2
dst_layer = reloaded_model.model.decoder.layers[0].fc2
assert isinstance(src_layer, WrapperWALayer)
assert isinstance(dst_layer, FakeActQuantLinear)
assert type(src_layer) is not type(dst_layer)
state_path = os.path.join(quantized_model_path, "model.safetensors")
if os.path.exists(state_path):
from safetensors.torch import load_file as load_state_dict
state_dict = load_state_dict(state_path)
else:
state_dict = torch.load(os.path.join(quantized_model_path, "pytorch_model.bin"), weights_only=True)
weight_key = "model.decoder.layers.0.fc2.weight"
bias_key = "model.decoder.layers.0.fc2.bias"
assert weight_key in state_dict
assert bias_key in state_dict
assert "model.decoder.layers.0.fc2.orig_layer.weight" not in state_dict
torch.manual_seed(7)
probe = torch.randn(2, 3, dst_layer.in_features, device="cuda:0", dtype=dst_layer.weight.dtype)
with torch.no_grad():
qdq_probe = dst_layer.qdq_input(probe)
src_out = (
torch.nn.functional.linear(
qdq_probe,
state_dict[weight_key].to(qdq_probe.device, qdq_probe.dtype),
state_dict[bias_key].to(qdq_probe.device, qdq_probe.dtype),
)
.float()
.cpu()
)
dst_out = dst_layer(probe).float().cpu()
torch.testing.assert_close(src_out, dst_out, rtol=1e-3, atol=1e-3)
finally:
if old_offload is None:
os.environ.pop("AR_DISABLE_OFFLOAD", None)
else:
os.environ["AR_DISABLE_OFFLOAD"] = old_offload