PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix cuda ut fail (#1237)
Heng Guo7a46aee8cc03
diff --git a/auto_round/formats.py b/auto_round/formats.pyindex 6254f8e39..b2f172f40 100644--- a/auto_round/formats.py+++ b/auto_round/formats.py@@ -493,7 +493,7 @@ class AutoGPTQFormat(OutputFormat):@OutputFormat.register("auto_awq")class AutoAWQFormat(OutputFormat):- support_schemes = ["W4A16", "W2A16", "W3A16", "W8A16", "BF16", "W2A16G64", "W2A16G32"]+ support_schemes = ["W4A16"]format_name = "auto_awq"@staticmethoddiff --git a/docs/step_by_step.md b/docs/step_by_step.mdindex 260980895..06e65c56a 100644--- a/docs/step_by_step.md+++ b/docs/step_by_step.md@@ -152,11 +152,10 @@ adopted within the community, **only 4-bits quantization is supported**. Please|export format | supported scheme ||--------------|------------------||**auto_round** | W4A16, W2A16, W3A16, W8A16, MXFP4, MXFP8, NVFP4, FPW8A16, W2A16G64, W2A16G32, FP8_STATIC, BF16|-|**auto_awq / auto_round:auto_awq** | W4A16, W2A16, W3A16, W8A16, BF16, W2A16G64, W2A16G32 |+|**auto_awq / auto_round:auto_awq** | W4A16||**auto_gptq / auto_round:auto_gptq / auto_round:gptqmodel**|W4A16, W2A16, W3A16, W8A16, BF16, W2A16G64, W2A16G32||**llm_compressor / auto_round:llm_compressor** | MXFP4, MXFP8, NVFP4, FPW8A16, FP8_STATIC ||**gguf** | GGUF:Q4_0, GGUF:Q4_1, GGUF:Q5_0, GGUF:Q5_1, GGUF:Q2_K_S, GGUF:Q3_K_S, GGUF:Q3_K_M, GGUF:Q3_K_L, GGUF:Q4_K_S, GGUF:Q4_K_M, GGUF:Q5_K_S, GGUF:Q5_K_M, GGUF:Q6_K, GGUF:Q8_0 |-|**itrex / itrex_xpu** | W4A16, W2A16, W3A16, W8A16, BF16, W2A16G64, W2A16G32 ||**fake** | all scheme|### Hardware Compatibilitydiff --git a/test/helpers.py b/test/helpers.pyindex 89b832c6d..6bd5ad30c 100644--- a/test/helpers.py+++ b/test/helpers.py@@ -10,6 +10,7 @@ from auto_round.utils import get_attr, llm_load_model, mllm_load_model, set_attr# Automatic choose local path or model name.def get_model_path(model_name: str) -> str:+ model_name = model_name.rstrip("/")ut_path = f"/tf_dataset/auto_round/models/{model_name}"local_path = f"/models/{model_name.split('/')[-1]}"diff --git a/test/test_cpu/export/test_gguf_format.py b/test/test_cpu/export/test_gguf_format.pyindex 5b96716eb..7304f22c7 100644--- a/test/test_cpu/export/test_gguf_format.py+++ b/test/test_cpu/export/test_gguf_format.py@@ -10,6 +10,9 @@ from auto_round import AutoRoundfrom ...helpers import get_model_path, get_tiny_model+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+class TestGGUF:@@ -26,7 +29,7 @@ class TestGGUF:def test_basic_usage(self, tiny_gemma_model_path, tiny_qwen_model_path):python_path = sys.executableres = os.system(- f"cd .. && {python_path} -m auto_round --model {tiny_gemma_model_path} "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {tiny_gemma_model_path} "f" --bs 16 --iters 0 --nsamples 1 --format gguf:q4_k_m")if res > 0 or res == -1:@@ -34,7 +37,7 @@ class TestGGUF:shutil.rmtree("./saved", ignore_errors=True)res = os.system(- f"cd .. && {python_path} -m auto_round --model {tiny_qwen_model_path}"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {tiny_qwen_model_path}"f" --bs 16 --iters 1 --nsamples 1 --format fake,gguf:q4_0")if res > 0 or res == -1:@@ -162,7 +165,7 @@ class TestGGUF:# for gguf_format in ["gguf:q4_0", "gguf:q4_1", "gguf:q4_k_m", "gguf:q6_k"]:for gguf_format in ["gguf:q4_k_m"]:res = os.system(- f"cd .. && {python_path} -m auto_round --model {model_name} "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {model_name} "f" --bs 16 --iters 1 --nsamples 1 --seqlen 16 --format {gguf_format}")if res > 0 or res == -1:@@ -170,7 +173,7 @@ class TestGGUF:shutil.rmtree("../../tmp_autoround", ignore_errors=True)res = os.system(- f"cd .. && {python_path} -m auto_round --model {model_name}"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {model_name}"f" --bs 16 --iters 0 --nsamples 1 --seqlen 16 --format fake,{gguf_format}")if res > 0 or res == -1:@@ -179,7 +182,7 @@ class TestGGUF:# test mixed q2_k_sres = os.system(- f"cd .. && {python_path} -m auto_round --model {model_name}"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {model_name}"f" --bs 16 --iters 0 --nsamples 1 --seqlen 16 --scheme GGUF:Q2_K_MIXED")if res > 0 or res == -1:diff --git a/test/test_cpu/utils/test_cli_usage.py b/test/test_cpu/utils/test_cli_usage.pyindex 62607c83a..6ba676936 100644--- a/test/test_cpu/utils/test_cli_usage.py+++ b/test/test_cpu/utils/test_cli_usage.py@@ -4,8 +4,12 @@ import sysfrom ...helpers import get_model_path+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+class TestAutoRoundCmd:+@classmethoddef setup_class(self):pass@@ -21,24 +25,24 @@ class TestAutoRoundCmd:python_path = sys.executable# Test llm script- res = os.system(f"cd .. && {python_path} -m auto_round -h")+ res = os.system(f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round -h")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && {python_path} -m auto_round --model {tiny_opt_model_path} --seqlen 32 --iter 2 --nsamples 1 --format auto_gptq,auto_round --output_dir ./saved --tasks piqa"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {tiny_opt_model_path} --seqlen 32 --iter 2 --nsamples 1 --format auto_gptq,auto_round --output_dir ./saved --tasks piqa")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && {python_path} -m auto_round --model {tiny_opt_model_path} --seqlen 8 --iter 1 --nsamples 1 --eval_task_by_task --tasks openbookqa --bs 32"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model {tiny_opt_model_path} --seqlen 8 --iter 1 --nsamples 1 --eval_task_by_task --tasks openbookqa --bs 32")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && {python_path} -c 'from auto_round.__main__ import run_light; run_light()' --seqlen 8 --iter 2 --nsamples 8 --output_dir ./saved --tasks lambada_openai"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -c 'from auto_round.__main__ import run_light; run_light()' --seqlen 8 --iter 2 --nsamples 8 --output_dir ./saved --tasks lambada_openai")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"@@ -46,23 +50,23 @@ class TestAutoRoundCmd:# test mllm script# test auto_round_mllm --eval help- res = os.system(f"cd .. && {python_path} -m auto_round --eval -h")+ res = os.system(f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --eval -h")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"# test auto_round_mllm --lmms help- res = os.system(f"cd .. && {python_path} -m auto_round --eval --lmms -h")+ res = os.system(f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --eval --lmms -h")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && {python_path} -m auto_round --mllm --model {tiny_qwen_vl_model_path} --iter 2 --nsamples 2 --seqlen 32 --format auto_round --output_dir ./saved"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --mllm --model {tiny_qwen_vl_model_path} --iter 2 --nsamples 2 --seqlen 32 --format auto_round --output_dir ./saved")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && {python_path} -m auto_round --mllm --iter 2 --nsamples 2 --model {tiny_qwen_vl_model_path} --seqlen 32 --format auto_round"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --mllm --iter 2 --nsamples 2 --model {tiny_qwen_vl_model_path} --seqlen 32 --format auto_round"" --quant_nontext_module --output_dir ./saved ")if res > 0 or res == -1:diff --git a/test/test_cuda/advanced/test_fp8_input.py b/test/test_cuda/advanced/test_fp8_input.pyindex 9590692a4..ec3dc6bf3 100644--- a/test/test_cuda/advanced/test_fp8_input.py+++ b/test/test_cuda/advanced/test_fp8_input.py@@ -64,8 +64,8 @@ class TestAutoRound:# print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0]))def test_small_model_rtn(self):- model, tokenizer = self.tiny_fp8_model()- ar = AutoRound(model=model, tokenizer=tokenizer, iters=0)+ model_name = get_model_path("qwen/Qwen3-0.6B-FP8")+ ar = AutoRound(model=model_name, iters=0)_, folder = ar.quantize_and_save(output_dir=self.save_dir)model_args = f"pretrained={self.save_dir}"result = simple_evaluate(model="hf", model_args=model_args, tasks="lambada_openai", batch_size="auto")@@ -75,8 +75,8 @@ class TestAutoRound:shutil.rmtree(self.save_dir, ignore_errors=True)def test_small_model_iters1(self):- model, tokenizer = self.tiny_fp8_model()- ar = AutoRound(model=model, tokenizer=tokenizer, iters=1)+ model_name = get_model_path("qwen/Qwen3-0.6B-FP8")+ ar = AutoRound(model=model_name, iters=1)_, folder = ar.quantize_and_save(output_dir=self.save_dir)model_args = f"pretrained={self.save_dir}"result = simple_evaluate(model="hf", model_args=model_args, tasks="lambada_openai", batch_size="auto")@@ -86,25 +86,25 @@ class TestAutoRound:shutil.rmtree(self.save_dir, ignore_errors=True)def test_medium_model_rtn(self):- model, tokenizer = self.tiny_fp8_model()- ar = AutoRound(model=model, tokenizer=tokenizer, iters=0)+ model_name = get_model_path("qwen/Qwen3-0.6B-FP8")+ ar = AutoRound(model=model_name, iters=0)_, folder = ar.quantize_and_save(output_dir=self.save_dir)model_args = f"pretrained={self.save_dir}"result = simple_evaluate(model="hf", model_args=model_args, tasks="lambada_openai", batch_size="auto")print(result["results"]["lambada_openai"]["acc,none"])- assert result["results"]["lambada_openai"]["acc,none"] > 0.55+ assert result["results"]["lambada_openai"]["acc,none"] > 0.33shutil.rmtree(self.save_dir, ignore_errors=True)def test_medium_model_rtn_with_lm_head(self):- model, tokenizer = self.tiny_fp8_model()+ model_name = get_model_path("qwen/Qwen3-0.6B-FP8")layer_config = {"lm_head": {"bits": 4}}- ar = AutoRound(model=model, tokenizer=tokenizer, iters=0, layer_config=layer_config)+ ar = AutoRound(model=model_name, iters=0, layer_config=layer_config)_, folder = ar.quantize_and_save(output_dir=self.save_dir)model_args = f"pretrained={self.save_dir}"result = simple_evaluate(model="hf", model_args=model_args, tasks="lambada_openai", batch_size="auto")print(result["results"]["lambada_openai"]["acc,none"])- assert result["results"]["lambada_openai"]["acc,none"] > 0.55+ assert result["results"]["lambada_openai"]["acc,none"] > 0.33shutil.rmtree(self.save_dir, ignore_errors=True)@@ -135,9 +135,9 @@ class TestAutoRound:def test_diff_datatype(self):for scheme in ["NVFP4", "MXFP4"]:- model, tokenizer = self.tiny_fp8_model()+ model_name = get_model_path("qwen/Qwen3-0.6B-FP8")for iters in [0, 1]:print(f"Testing scheme: {scheme}, iters: {iters}")- ar = AutoRound(model=model, tokenizer=tokenizer, iters=iters, scheme=scheme)+ ar = AutoRound(model_name, iters=iters, scheme=scheme)ar.quantize_and_save(output_dir=self.save_dir)shutil.rmtree(self.save_dir, ignore_errors=True)diff --git a/test/test_cuda/advanced/test_multiple_card.py b/test/test_cuda/advanced/test_multiple_card.pyindex ca4dd7cd5..00863c2eb 100644--- a/test/test_cuda/advanced/test_multiple_card.py+++ b/test/test_cuda/advanced/test_multiple_card.py@@ -354,7 +354,7 @@ class TestAutoRound:@multi_carddef test_mllm_device_map(self):- model_name = get_model_path("qwen/Qwen2-VL-2B-Instruct/")+ model_name = get_model_path("qwen/Qwen2-VL-2B-Instruct")from auto_round import AutoRoundMLLMdevice_map = "0,1"diff --git a/test/test_cuda/advanced/test_multiple_card_calib.py b/test/test_cuda/advanced/test_multiple_card_calib.pyindex fedb3f328..06c869a86 100644--- a/test/test_cuda/advanced/test_multiple_card_calib.py+++ b/test/test_cuda/advanced/test_multiple_card_calib.py@@ -7,6 +7,9 @@ import pytestfrom auto_round.testing_utils import multi_card+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+def get_accuracy(data):match = re.search(r"\|acc\s+\|[↑↓]\s+\|\s+([\d.]+)\|", data)@@ -41,7 +44,7 @@ class TestAutoRound:##test llm scriptres = os.system(- f"cd .. && {python_path} -m auto_round --model /models/Meta-Llama-3.1-8B-Instruct --devices '0,1' --quant_lm_head --iters 1 --nsamples 1 --output_dir None"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model /models/Meta-Llama-3.1-8B-Instruct --devices '0,1' --quant_lm_head --iters 1 --nsamples 1 --output_dir None")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"@@ -52,7 +55,7 @@ class TestAutoRound:##test llm scriptres = os.system(- f"cd .. && {python_path} -m auto_round --model facebook/opt-125m --scheme NVFP4 --devices '0,1' --iters 1 --nsamples 1 --enable_torch_compile --low_gpu_mem_usage"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {python_path} -m auto_round --model facebook/opt-125m --scheme NVFP4 --devices '0,1' --iters 1 --nsamples 1 --enable_torch_compile --low_gpu_mem_usage")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"diff --git a/test/test_cuda/backends/test_marlin_backend.py b/test/test_cuda/backends/test_marlin_backend.pyindex 3f480c90f..793fe3bca 100644--- a/test/test_cuda/backends/test_marlin_backend.py+++ b/test/test_cuda/backends/test_marlin_backend.py@@ -7,11 +7,12 @@ from transformers import AutoModelForCausalLM, AutoTokenizerfrom auto_round import AutoRound, AutoRoundConfigfrom auto_round.eval.evaluation import simple_evaluate_user_model-from ...helpers import model_infer+from ...helpers import get_model_path, model_inferclass TestAutoRoundMarlinBackend:save_dir = "./saved"+ model_name = get_model_path("facebook/opt-125m")@pytest.fixture(autouse=True, scope="class")def setup_and_teardown_class(self):diff --git a/test/test_cuda/export/test_gguf.py b/test/test_cuda/export/test_gguf.pyindex 94d3c7461..4146cb938 100644--- a/test/test_cuda/export/test_gguf.py+++ b/test/test_cuda/export/test_gguf.py@@ -12,6 +12,9 @@ from auto_round.testing_utils import require_gguffrom ...helpers import get_model_path, get_tiny_model, save_tiny_model+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+class TestAutoRound:save_dir = "./saved"@@ -56,7 +59,7 @@ class TestAutoRound:save_dir = os.path.join(os.path.dirname(__file__), "saved")res = os.system(- f"cd .. && {sys.executable} -m auto_round --model {tiny_qwen_model_path} --iter 2 "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {sys.executable} -m auto_round --model {tiny_qwen_model_path} --iter 2 "f"--output_dir {save_dir} --nsample 2 --format gguf:q4_0 --device 0")print(save_dir)@@ -64,8 +67,8 @@ class TestAutoRound:from llama_cpp import Llama- gguf_file = os.listdir(f"{save_dir}/tmp_tiny_qwen_model_path-gguf")[0]- llm = Llama(f"{save_dir}/tmp_tiny_qwen_model_path-gguf/{gguf_file}", n_gpu_layers=-1)+ gguf_file = os.listdir(f"{save_dir}/tiny_qwen_model_path-gguf")[0]+ llm = Llama(f"{save_dir}/tiny_qwen_model_path-gguf/{gguf_file}", n_gpu_layers=-1)output = llm("There is a girl who likes adventure,", max_tokens=32)print(output)shutil.rmtree(save_dir, ignore_errors=True)@@ -155,7 +158,7 @@ class TestAutoRound:autoround.quantize_and_save(output_dir=quantized_model_path, format="gguf:q4_0")assert "mmproj-model.gguf" in os.listdir("./saved")file_size = os.path.getsize("./saved/Qwen2-VL-2B-Instruct-Q4_0.gguf") / 1024**2- assert abs(file_size - 4242) < 5.0+ assert abs(file_size - 894) < 5.0file_size = os.path.getsize("./saved/mmproj-model.gguf") / 1024**2assert abs(file_size - 2580) < 5.0shutil.rmtree("./saved", ignore_errors=True)diff --git a/test/test_cuda/integrations/test_transformers.py b/test/test_cuda/integrations/test_transformers.pyindex feb2516f7..638cb7396 100644--- a/test/test_cuda/integrations/test_transformers.py+++ b/test/test_cuda/integrations/test_transformers.py@@ -36,7 +36,7 @@ if is_torch_available():# @slow@require_torch_gpu@require_accelerate-class AutoRoundTest:+class TestAutoRound:model_name = "OPEA/Qwen2.5-1.5B-Instruct-int4-sym-inc"input_text = "There is a girl who likes adventure,"EXPECTED_OUTPUTS = set()@@ -205,3 +205,8 @@ class AutoRoundTest:text = "There is a girl who likes adventure,"inputs = tokenizer(text, return_tensors="pt").to(model.device)tokenizer.decode(model.generate(**inputs, max_new_tokens=5)[0])+++# FAILED export/test_gguf.py::TestAutoRound::test_gguf_format - AssertionError: qwen2 tuning fail+# FAILED export/test_gguf.py::TestAutoRound::test_all_format - SystemExit: 1+# FAILED export/test_gguf.py::TestAutoRound::test_vlm_gguf - AttributeError: 'Qwen2VLForConditionalGeneration' object has no attribute 'last_layer_name_to_block_name'diff --git a/test/test_cuda/models/test_support_vlms.py b/test/test_cuda/models/test_support_vlms.pyindex 3358c8226..4f9373ca5 100644--- a/test/test_cuda/models/test_support_vlms.py+++ b/test/test_cuda/models/test_support_vlms.py@@ -9,8 +9,12 @@ from PIL import Imagefrom auto_round import AutoRoundConfig # # must import for auto-round formatfrom auto_round.testing_utils import require_gptqmodel, require_package_version_ut, require_vlm_env+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+class TestSupportVLMS:+@classmethoddef setup_class(self):self.save_dir = os.path.join(os.path.dirname(__file__), "ut_saved")@@ -26,7 +30,7 @@ class TestSupportVLMS:model_path = "/models/Qwen2-VL-2B-Instruct/"# test tuneres = os.system(- f"cd .. && {self.python_path} -m auto_round --mllm "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {self.python_path} -m auto_round --mllm "f"--model {model_path} --iter 2 --output_dir {self.save_dir} --device {self.device}")assert not (res > 0 or res == -1), "qwen2 tuning fail"@@ -81,7 +85,7 @@ class TestSupportVLMS:model_path = "/models/Phi-3.5-vision-instruct/"## test tuneres = os.system(- f"cd .. && {self.python_path} -m auto_round --mllm "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {self.python_path} -m auto_round --mllm "f"--model {model_path} --iter 2 --output_dir {self.save_dir} --device {self.device}")assert not (res > 0 or res == -1), "Phi-3.5 tuning fail"@@ -129,7 +133,7 @@ class TestSupportVLMS:model_path = "/models/Phi-3.5-vision-instruct/"## test tuneres = os.system(- f"cd .. && {self.python_path} -m auto_round --mllm "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {self.python_path} -m auto_round --mllm "f"--model {model_path} --iter 2 --quant_nontext_module "f"--nsample 64 --seqlen 32 "f"--format auto_awq --output_dir {self.save_dir} --device {self.device}"@@ -177,7 +181,7 @@ class TestSupportVLMS:model_path = "/models/glm-4v-9b/"## test tuneres = os.system(- f"cd .. && {self.python_path} -m auto_round "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {self.python_path} -m auto_round "f"--model {model_path} --iter 1 --output_dir {self.save_dir} --device {self.device}")assert not (res > 0 or res == -1), "glm-4v-9b tuning fail"@@ -186,7 +190,7 @@ class TestSupportVLMS:model_path = "/models/granite-vision-3.2-2b"## test tuneres = os.system(- f"cd .. && {self.python_path} -m auto_round "+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' {self.python_path} -m auto_round "f"--model {model_path} --iter 1 --output_dir {self.save_dir} --device {self.device}")assert not (res > 0 or res == -1), "granite-vision-3.2-2b tuning fail"diff --git a/test/test_cuda/quantization/test_mix_bits.py b/test/test_cuda/quantization/test_mix_bits.pyindex 82da57008..614a5d2b1 100644--- a/test/test_cuda/quantization/test_mix_bits.py+++ b/test/test_cuda/quantization/test_mix_bits.py@@ -242,7 +242,6 @@ class TestAutoRound:}autoround = AutoRound(tiny_opt_model_path,- self.tokenizer,scheme="W4A16",iters=2,seqlen=2,diff --git a/test/test_cuda/schemes/test_scheme.py b/test/test_cuda/schemes/test_scheme.pyindex 368737946..739a742a6 100644--- a/test/test_cuda/schemes/test_scheme.py+++ b/test/test_cuda/schemes/test_scheme.py@@ -42,7 +42,7 @@ class TestAutoRound:ar.quantize()def test_mxfp4(self, tiny_opt_model_path):- ar = AutoRound(tiny_opt_model_path, scheme="MXFP4", nsamples=1, iters=1)+ ar = AutoRound(tiny_opt_model_path, scheme="MXFP4_RCEIL", nsamples=1, iters=1)assert ar.bits == 4assert ar.act_bits == 4assert ar.data_type == "mx_fp"diff --git a/test/test_cuda/utils/test_alg_ext.py b/test/test_cuda/utils/test_alg_ext.pyindex 5ddc02ed4..d2e8e4bfb 100644--- a/test/test_cuda/utils/test_alg_ext.py+++ b/test/test_cuda/utils/test_alg_ext.py@@ -10,6 +10,9 @@ from auto_round.eval.evaluation import simple_evaluate_user_modelfrom ...helpers import get_model_path+AUTO_ROUND_PATH = __file__.split("/")+AUTO_ROUND_PATH = "/".join(AUTO_ROUND_PATH[: AUTO_ROUND_PATH.index("test")])+class TestAlgExt:save_folder = "./saved"@@ -49,13 +52,13 @@ class TestAlgExt:python_path = sys.executableres = os.system(- f"cd .. && CUDA_VISIBLE_DEVICES=0 {python_path} -m auto_round --model {tiny_opt_model_path} --iters 1 --device auto --enable_alg_ext --avg_bits 2 --options=W2A16,W4A16 --ignore_scale_zp_bits --nsamples 1 --seqlen 32"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' CUDA_VISIBLE_DEVICES=0 {python_path} -m auto_round --model {tiny_opt_model_path} --iters 1 --device auto --enable_alg_ext --avg_bits 2 --options=W2A16,W4A16 --ignore_scale_zp_bits --nsamples 1 --seqlen 32")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"res = os.system(- f"cd .. && CUDA_VISIBLE_DEVICES=0 {python_path} -m auto_round --model {tiny_opt_model_path} --iters 1 --device auto --enable_alg_ext --avg_bits 5.5 --options=mxfp4,mxfp8 --ignore_scale_zp_bits --enable_torch_compile --nsamples 1 --seqlen 32"+ f"PYTHONPATH='AUTO_ROUND_PATH:$PYTHONPATH' CUDA_VISIBLE_DEVICES=0 {python_path} -m auto_round --model {tiny_opt_model_path} --iters 1 --device auto --enable_alg_ext --avg_bits 5.5 --options=mxfp4,mxfp8 --ignore_scale_zp_bits --enable_torch_compile --nsamples 1 --seqlen 32")if res > 0 or res == -1:assert False, "cmd line test fail, please have a check"