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

Add llmc ut back (#1974)

Yi Liud763ae04c28b
diff --git a/test/test_xpu/test_llmc_integration.py b/test/test_xpu/test_llmc_integration.py
index 3dd908bd3..3e68bd144 100644
--- a/test/test_xpu/test_llmc_integration.py
+++ b/test/test_xpu/test_llmc_integration.py
@@ -127,61 +127,60 @@ def test_oneshot_application(recipe, tmp_path, tiny_tiny_llama_model_path):
assert not hasattr(not_targeted, "quantization_scheme")
-#
-# @pytest.mark.skipif(torch.xpu.device_count() < 2, reason="test requires at least 2 XPUs")
-# def test_oneshot_with_device_ids(tmp_path, tiny_tiny_llama_model_path):
-# output = tmp_path / "oneshot_output"
-# model = tiny_tiny_llama_model_path
-# 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() < 2, reason="test requires at least 2 XPUs")
+def test_oneshot_with_device_ids(tmp_path, tiny_tiny_llama_model_path):
+ output = tmp_path / "oneshot_output"
+ model = tiny_tiny_llama_model_path
+ 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")
@pytest.mark.skipif(torch.xpu.device_count() < 1, reason="test requires at least 1 XPU")