PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Enhance llmc CI on XPU (#1483)
chen, suyue19b3c4852fa9
diff --git a/.azure-pipelines/docker/Dockerfile_xpu.devel b/.azure-pipelines/docker/Dockerfile_xpu.develindex efffa549f..e0d41ad13 100644--- a/.azure-pipelines/docker/Dockerfile_xpu.devel+++ b/.azure-pipelines/docker/Dockerfile_xpu.devel@@ -17,7 +17,10 @@ RUN apt-get update && apt-get install -y --no-install-recommends --fix-missing \wget \bc \jq \- vim+ vim \+ python3.12 \+ python3.12-dev \+ python3-pipRUN apt-get install -y software-properties-common \&& add-apt-repository -y ppa:kobuk-team/intel-graphics \diff --git a/.azure-pipelines/scripts/ut/run_ut_xpu.sh b/.azure-pipelines/scripts/ut/run_ut_xpu.shindex 9c2163f95..76f405300 100644--- a/.azure-pipelines/scripts/ut/run_ut_xpu.sh+++ b/.azure-pipelines/scripts/ut/run_ut_xpu.sh@@ -45,7 +45,7 @@ function run_unit_test() {echo "##[endgroup]"done- for test_file in $(find ./test_xpu -name "test*.py" | sort); do+ for test_file in $(find ./test_xpu -name "test*.py" ! -name "test_llmc_integration.py" | sort); dolocal test_basename=$(basename ${test_file} .py)echo "##[group]Running xpu ${test_file}..."@@ -58,6 +58,27 @@ function run_unit_test() {done}+function run_unit_test_vllm() {+ echo "##[group]set up vllm UT env..."+ BUILD_TYPE="nightly" uv pip install -r ./test_xpu/requirements_llmc.txt+ uv pip list+ echo "##[endgroup]"++ auto_round_path=$(python -c 'import auto_round; print(auto_round.__path__[0])')++ for test_file in $(find ./test_xpu -name "test_llmc_integration.py" | sort); do+ local test_basename=$(basename ${test_file} .py)++ echo "##[group]Running xpu vLLM ${test_file}..."+ local ut_log_name="${LOG_DIR}/unittest_xpu_${test_basename}.log"+ numactl --physcpubind="${NUMA_CPUSET:-0-27}" --membind="${NUMA_NODE:-0}" pytest --cov="${auto_round_path}" \+ --cov-report term --html=report.html --self-contained-html \+ --cov-report xml:coverage.xml --cov-append -vs --disable-warnings \+ ${test_file} 2>&1 | tee ${ut_log_name}+ echo "##[endgroup]"+ done+}+function print_summary() {local status=0while IFS= read -r line; do@@ -86,6 +107,7 @@ function collect_log() {function main() {setup_environmentrun_unit_test+ run_unit_test_vllmcollect_logprint_summary}diff --git a/test/test_xpu/requirements_llmc.txt b/test/test_xpu/requirements_llmc.txtnew file mode 100644index 000000000..a76afdd35--- /dev/null+++ b/test/test_xpu/requirements_llmc.txt@@ -0,0 +1,2 @@+llmcompressor @ git+https://github.com/vllm-project/llm-compressor.git@main+compressed-tensors @ git+https://github.com/vllm-project/compressed-tensors.git@maindiff --git a/test/test_xpu/test_llmc_integration.py b/test/test_xpu/test_llmc_integration.pynew file mode 100644index 000000000..2b0310306--- /dev/null+++ b/test/test_xpu/test_llmc_integration.py@@ -0,0 +1,233 @@+import pytest+import torch+from compressed_tensors.quantization import QuantizationArgs, QuantizationScheme+from llmcompressor import oneshot+from llmcompressor.modifiers.autoround import AutoRoundModifier+from transformers import AutoModelForCausalLM, AutoTokenizer++from auto_round.calib_dataset import get_dataset++recipe_str = """+quant_stage:+ quant_modifiers:+ AutoRoundModifier:+ ignore: ["lm_head"]+ iters: 10+ config_groups:+ group_0:+ targets:+ - "Linear"+ input_activations: null+ output_activations: null+ weights:+ num_bits: 4+ type: "int"+ symmetric: true+ strategy: group+ group_size: 128+"""++recipe_modifier_full = AutoRoundModifier(+ ignore=["lm_head"],+ iters=10,+ config_groups={+ "group_0": QuantizationScheme(+ targets=["Linear"],+ weights=QuantizationArgs(num_bits=4, strategy="group", group_size=128),+ )+ },+)+recipe_modifier_nvfp4 = AutoRoundModifier(+ ignore=["lm_head"],+ iters=2,+ scheme="NVFP4",+)++recipe_modifier_mxfp4 = AutoRoundModifier(+ ignore=["lm_head"],+ iters=0,+ scheme="MXFP4",+)++w8a8_dynamic_recipe_modifier = AutoRoundModifier(+ ignore=["lm_head"],+ iters=0,+ config_groups={+ "group_0": QuantizationScheme(+ targets=["Linear"],+ weights=QuantizationArgs(num_bits=8, type="float", strategy="channel"),+ input_activations=QuantizationArgs(num_bits=8, type="float", strategy="token", dynamic=True),+ )+ },+)++w8a8_static_recipe_modifier = AutoRoundModifier(+ ignore=["lm_head"],+ iters=0,+ config_groups={+ "group_0": QuantizationScheme(+ targets=["Linear"],+ weights=QuantizationArgs(num_bits=8, type="float", strategy="tensor"),+ input_activations=QuantizationArgs(num_bits=8, type="float", strategy="tensor"),+ )+ },+)++[email protected](torch.xpu.device_count() < 1, reason="test requires at least 1 XPU")+ "recipe",+ [+ recipe_str,+ recipe_modifier_full,+ recipe_modifier_nvfp4,+ recipe_modifier_mxfp4,+ ],+)+def test_oneshot_application(recipe, tmp_path):+ output = tmp_path / "oneshot_output"+ model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"+ tokenizer = AutoTokenizer.from_pretrained(model)+ dataset = get_dataset(+ tokenizer=tokenizer,+ seqlen=1024,+ nsamples=32,+ )++ oneshot(+ model=model,+ dataset=dataset,+ output_dir=output,+ recipe=recipe,+ )+ model_loaded = AutoModelForCausalLM.from_pretrained(output, device_map="xpu")++ # Check that the model is quantized+ # decompress() will attach a quantization_config to the model+ # as we decompress right away+ quantization_config = model_loaded.config.quantization_config.quantization_config+ assert quantization_config is not None++ # check config is set properly+ assert "lm_head" in quantization_config.ignore+ assert len(quantization_config.config_groups) == 1+ quant_scheme = quantization_config.config_groups["group_0"]+ assert isinstance(quant_scheme, QuantizationScheme)++ weight_args = quantization_config.config_groups["group_0"].weights+ assert isinstance(weight_args, QuantizationArgs)+ assert weight_args.num_bits == 4++ # Check a specific layer is quantized+ targeted_linear_layer = model_loaded.model.layers[2].self_attn.q_proj+ assert hasattr(targeted_linear_layer, "quantization_scheme")++ # Check lm-head is not quantized+ not_targeted = model_loaded.lm_head+ assert not hasattr(not_targeted, "quantization_scheme")++[email protected](torch.xpu.device_count() < 2, reason="test requires at least 2 XPUs")+def test_oneshot_with_device_ids(tmp_path):+ output = tmp_path / "oneshot_output"+ model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"+ tokenizer = AutoTokenizer.from_pretrained(model)+ dataset = get_dataset(+ tokenizer=tokenizer,+ seqlen=512,+ nsamples=4,+ )++ recipe = AutoRoundModifier(+ ignore=["lm_head"],+ iters=10,+ config_groups={+ "group_0": QuantizationScheme(+ targets=["Linear"],+ weights=QuantizationArgs(num_bits=4, strategy="group", group_size=128),+ )+ },+ device_ids="0,1",+ )++ oneshot(+ model=model,+ dataset=dataset,+ output_dir=output,+ recipe=recipe,+ )+ model_loaded = AutoModelForCausalLM.from_pretrained(output, device_map="xpu")++ # Check that the model is quantized+ # decompress() will attach a quantization_config to the model+ # as we decompress right away+ quantization_config = model_loaded.config.quantization_config.quantization_config+ assert quantization_config is not None++ # check config is set properly+ assert "lm_head" in quantization_config.ignore+ assert len(quantization_config.config_groups) == 1+ quant_scheme = quantization_config.config_groups["group_0"]+ assert isinstance(quant_scheme, QuantizationScheme)++ weight_args = quantization_config.config_groups["group_0"].weights+ assert isinstance(weight_args, QuantizationArgs)+ assert weight_args.num_bits == 4++ # Check a specific layer is quantized+ targeted_linear_layer = model_loaded.model.layers[2].self_attn.q_proj+ assert hasattr(targeted_linear_layer, "quantization_scheme")++ # Check lm-head is not quantized+ not_targeted = model_loaded.lm_head+ assert not hasattr(not_targeted, "quantization_scheme")++[email protected](torch.xpu.device_count() < 1, reason="test requires at least 1 XPU")+ "recipe",+ [w8a8_dynamic_recipe_modifier, w8a8_static_recipe_modifier],+)+def test_rtn_oneshot(recipe, tmp_path):+ output = tmp_path / "oneshot_output"+ model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"+ tokenizer = AutoTokenizer.from_pretrained(model)+ dataset = get_dataset(+ tokenizer=tokenizer,+ seqlen=1024,+ nsamples=32,+ )++ oneshot(+ model=model,+ dataset=dataset,+ output_dir=output,+ recipe=recipe,+ )+ model_loaded = AutoModelForCausalLM.from_pretrained(output, device_map="xpu")++ quantization_config = model_loaded.config.quantization_config.quantization_config+ assert quantization_config is not None++ # check config is set properly+ assert "lm_head" in quantization_config.ignore+ assert len(quantization_config.config_groups) == 1+ quant_scheme = quantization_config.config_groups["group_0"]+ assert isinstance(quant_scheme, QuantizationScheme)++ weight_args = quantization_config.config_groups["group_0"].weights+ act_args = quantization_config.config_groups["group_0"].input_activations+ assert isinstance(weight_args, QuantizationArgs)+ assert weight_args.num_bits == recipe.config_groups["group_0"].weights.num_bits+ assert weight_args.strategy == recipe.config_groups["group_0"].weights.strategy+ if act_args is not None:+ assert act_args.num_bits == recipe.config_groups["group_0"].input_activations.num_bits+ assert act_args.strategy == recipe.config_groups["group_0"].input_activations.strategy++ # Check a specific layer is quantized+ targeted_linear_layer = model_loaded.model.layers[2].self_attn.q_proj+ assert hasattr(targeted_linear_layer, "quantization_scheme")++ # Check lm-head is not quantized+ not_targeted = model_loaded.lm_head+ assert not hasattr(not_targeted, "quantization_scheme")