Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

Fix compatibility with latest compressed_tensors compressor refactor (#1576)

Yi Liu49d2dbb0de14
diff --git a/auto_round/export/export_to_llmcompressor/export.py b/auto_round/export/export_to_llmcompressor/export.py
index e5fe46abf..61685377c 100644
--- a/auto_round/export/export_to_llmcompressor/export.py
+++ b/auto_round/export/export_to_llmcompressor/export.py
@@ -97,11 +97,42 @@ def _get_quant_format(model):
return None
+def _compress_and_set_format(layer, scheme, device=None):
+ """Compress a layer and set its quantization format, with backward compatibility.
+
+ Supports both old (<0.9.0) and new (>=0.9.0) compressed_tensors APIs:
+ - New API: compress_module() handles compression, state dict replacement, status, and format.
+ - Old API: NaiveQuantizationCompressor instance + manual param management + set_per_module_format.
+ """
+ try:
+ from compressed_tensors.compressors import compress_module # pylint: disable=E0401
+
+ compress_module(layer)
+ except (ImportError, TypeError):
+ from compressed_tensors.compressors import NaiveQuantizationCompressor # pylint: disable=E0401
+ from compressed_tensors.config.format import set_per_module_format # pylint: disable=E0401
+ from compressed_tensors.quantization import QuantizationStatus # pylint: disable=E0401
+ from compressed_tensors.utils import ( # pylint: disable=E0401
+ delete_offload_parameter,
+ register_offload_parameter,
+ )
+
+ compressor = NaiveQuantizationCompressor()
+ q_state_dict = compressor.compress(layer.state_dict(), names_to_scheme={"": scheme}, show_progress=False)
+
+ for param_name, _ in list(layer.named_parameters(recurse=False)):
+ delete_offload_parameter(layer, param_name)
+
+ for param_name, value in q_state_dict.items():
+ param = torch.nn.Parameter(value, requires_grad=False)
+ register_offload_parameter(layer, param_name, param, device)
+
+ layer.quantization_status = QuantizationStatus.COMPRESSED
+ set_per_module_format(layer)
+
+
def pack_layer(name, model, device=None):
- from compressed_tensors.compressors import NaiveQuantizationCompressor # pylint: disable=E0401
- from compressed_tensors.config.format import set_per_module_format # pylint: disable=E0401
from compressed_tensors.quantization import QuantizationStatus # pylint: disable=E0401
- from compressed_tensors.utils import delete_offload_parameter, register_offload_parameter # pylint: disable=E0401
layer = get_module(model, name)
if type(layer) not in SUPPORTED_LAYER_TYPES and not isinstance(layer, WrapperWALayer): ##already packed
@@ -135,20 +166,7 @@ def pack_layer(name, model, device=None):
setattr(layer, "weight_zero_point", torch.nn.Parameter(zp.to(weight_device), requires_grad=False))
delattr(layer, "scale")
- compressor = NaiveQuantizationCompressor()
- q_state_dict = compressor.compress(layer.state_dict(), names_to_scheme={"": scheme}, show_progress=False)
-
- # remove any existing parameters
- for name, _ in list(layer.named_parameters(recurse=False)):
- delete_offload_parameter(layer, name)
-
- # replace with compressed parameters
- for name, value in q_state_dict.items():
- param = torch.nn.Parameter(value, requires_grad=False)
- register_offload_parameter(layer, name, param, device)
-
- layer.quantization_status = QuantizationStatus.COMPRESSED
- set_per_module_format(layer)
+ _compress_and_set_format(layer, scheme, device)
@torch.no_grad()
diff --git a/test/test_cpu/advanced/test_low_precision_input_model.py b/test/test_cpu/advanced/test_low_precision_input_model.py
index 48aa59b24..ccdf0634f 100644
--- a/test/test_cpu/advanced/test_low_precision_input_model.py
+++ b/test/test_cpu/advanced/test_low_precision_input_model.py
@@ -1,3 +1,4 @@
+import pytest
import torch
from auto_round.utils.weight_handler import (
@@ -14,6 +15,7 @@ class TestCompressedTensor:
mxfp4_model_path = "QuixiAI/Llama-3.2-1B-MXFP4"
fp8_block_model_path = "RedHatAI/Qwen3-0.6B-FP8-BLOCK"
+ @pytest.mark.skip(reason="CompressedLinear removed in compressed_tensors PR #610, see #1578")
def test_fp8_block(self):
model = get_tiny_model(get_model_path(self.fp8_block_model_path))
assert (
@@ -26,6 +28,7 @@ class TestCompressedTensor:
type(model.model.layers[0].mlp.up_proj) is torch.nn.Linear
), "CompressedLinear layer was not converted to Linear"
+ @pytest.mark.skip(reason="CompressedLinear removed in compressed_tensors PR #610, see #1578")
def test_nvfp4(self):
model = get_tiny_model(get_model_path(self.nvfp4_model_path))
assert (
@@ -38,6 +41,7 @@ class TestCompressedTensor:
type(model.model.layers[0].mlp.up_proj) is torch.nn.Linear
), "CompressedLinear layer was not converted to Linear"
+ @pytest.mark.skip(reason="CompressedLinear removed in compressed_tensors PR #610, see #1578")
def test_mxfp4(self):
model = get_tiny_model(get_model_path(self.mxfp4_model_path))
assert (
diff --git a/test/test_cpu/requirements.txt b/test/test_cpu/requirements.txt
index 80f981849..a6ce699a7 100644
--- a/test/test_cpu/requirements.txt
+++ b/test/test_cpu/requirements.txt
@@ -5,8 +5,7 @@ sentencepiece
torchvision
pillow
numba
-llmcompressor @ git+https://github.com/vllm-project/llm-compressor.git@370c04c34191c0e5fb88cf889e16578bf8c7fe20
-compressed-tensors==0.14.1a20260313 # temporary pin for llmcompressor
+llmcompressor @ git+https://github.com/vllm-project/llm-compressor.git@main
lm_eval >= 0.4.10 # for transformers >= 5.0.0
diffusers
protobuf
diff --git a/test/test_cuda/requirements_llmc.txt b/test/test_cuda/requirements_llmc.txt
index d03556a6b..b11e53b7a 100644
--- a/test/test_cuda/requirements_llmc.txt
+++ b/test/test_cuda/requirements_llmc.txt
@@ -1,3 +1,2 @@
-llmcompressor @ git+https://github.com/vllm-project/llm-compressor.git@370c04c34191c0e5fb88cf889e16578bf8c7fe20
-compressed-tensors==0.14.1a20260313 # temporary pin for llmcompressor
+llmcompressor @ git+https://github.com/vllm-project/llm-compressor.git@main
parameterized