PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Update autogptq exporting (#310)
wenhuach21bb07dc455048
diff --git a/README.md b/README.mdindex 46329b35f..e19a310a0 100644--- a/README.md+++ b/README.md@@ -297,6 +297,7 @@ release most of the models ourselves.| meta-llama/Meta-Llama-3.1-8B | [model-kaitchup-autogptq-sym-int4*](https://huggingface.co/kaitchup/Meta-Llama-3.1-8B-autoround-gptq-4bit-sym) || Qwen/Qwen-VL | [accuracy](./examples/multimodal-modeling/Qwen-VL/README.md), [recipe](./examples/multimodal-modeling/Qwen-VL/run_autoround.sh)| Qwen/Qwen2-7B | [model-autoround-sym-int4](https://huggingface.co/Intel/Qwen2-7B-int4-inc), [model-autogptq-sym-int4](https://huggingface.co/Intel/Qwen2-7B-int4-inc) |+| THUDM/glm-4-9b-chat | [recipe](./docs/glm-4-9b-chat-recipe) || Qwen/Qwen2-57B-A14B-Instruct | [model-autoround-sym-int4](https://huggingface.co/Intel/Qwen2-57B-A14B-Instruct-int4-inc),[model-autogptq-sym-int4](https://huggingface.co/Intel/Qwen2-57B-A14B-Instruct-int4-inc) || 01-ai/Yi-1.5-9B | [model-LnL-AI-autogptq-int4*](https://huggingface.co/LnL-AI/Yi-1.5-9B-4bit-gptq-autoround) || 01-ai/Yi-1.5-9B-Chat | [model-LnL-AI-autogptq-int4*](https://huggingface.co/LnL-AI/Yi-1.5-9B-Chat-4bit-gptq-autoround) |diff --git a/auto_round/export/export_to_autogptq/export.py b/auto_round/export/export_to_autogptq/export.pyindex e39ffc695..2ca99b123 100644--- a/auto_round/export/export_to_autogptq/export.py+++ b/auto_round/export/export_to_autogptq/export.py@@ -52,6 +52,13 @@ from tqdm import tqdmfrom concurrent.futures import ThreadPoolExecutorfrom auto_round.utils import get_autogptq_packing_qlinear+BLOCK_PATTERNS = [ ## copy from transformers optimum+ "transformer.h",+ "model.decoder.layers",+ "gpt_neox.layers",+ "model.layers",+]+def pack_layer(name, model, layer_config, backend, pbar):with tctl.threadpool_limits(limits=1):@@ -124,10 +131,21 @@ def save_quantized_as_autogptq(output_dir, inplace=True, backend="auto_gptq:exllif processor is not None:processor.save_pretrained(output_dir)##check module quantized in block, this may have bug for mixed precision quantization+ quantization_config = kwargs["serialization_dict"]if bool(quant_block_list):all_blocks = quant_block_list+ flattened_list = [item for sublist in all_blocks for item in sublist]+ common_prefix = os.path.commonprefix(flattened_list).rstrip('.')+ if common_prefix not in BLOCK_PATTERNS:+ logger.error(f"auto-gptq format may not support loading this quantized model")+ quantization_config['block_name_to_quantize'] = common_prefixelse:all_blocks = get_block_names(model)+ flattened_list = [item for sublist in all_blocks for item in sublist]+ common_prefix = os.path.commonprefix(flattened_list).rstrip('.')+ if common_prefix not in BLOCK_PATTERNS:+ quantization_config['block_name_to_quantize'] = common_prefix+all_to_quantized = Truemodules_in_block_to_quantize = []for block_names in all_blocks:@@ -162,7 +180,7 @@ def save_quantized_as_autogptq(output_dir, inplace=True, backend="auto_gptq:exllpassif output_dir is None:return model- quantization_config = kwargs["serialization_dict"]+quantization_config["quant_method"] = "gptq"quantization_config.pop("dataset", None) ## pile-10k is not supported in gptqquantization_config["desc_act"] = False ## for autogptq API@@ -203,4 +221,3 @@ def save(model: torch.nn.Module, save_dir: str, max_shard_size: str = "5GB", safif hasattr(model, "config") and hasattr(model.config, "quantization_config"):with open(os.path.join(save_dir, config_file), "w", encoding="utf-8") as f:json.dump(model.config.quantization_config, f, indent=2)-diff --git a/auto_round/script/llm.py b/auto_round/script/llm.pyindex 7bf5163a8..292b1162d 100644--- a/auto_round/script/llm.py+++ b/auto_round/script/llm.py@@ -233,7 +233,7 @@ def tune(args):args.format = "auto_round"supported_formats = ["auto_round", "auto_gptq", "auto_awq", "auto_round:gptq", "auto_round:auto_gptq","auto_round:auto_gptq:marlin", "auto_round:gptq:marlin", "auto_round:auto_awq",- "auto_round:awq", "auto_awq", "itrex", "iterx_xpu", "fake"]+ "auto_round:awq", "auto_gptq:marlin", "itrex", "iterx_xpu", "fake"]formats = args.format.replace(' ', '').split(",")for format in formats:if format not in supported_formats:diff --git a/auto_round/script/mllm.py b/auto_round/script/mllm.pyindex e62d72e28..6264f9361 100644--- a/auto_round/script/mllm.py+++ b/auto_round/script/mllm.py@@ -205,6 +205,9 @@ def tune(args):supported_formats = ["auto_round", "auto_round:gptq", "auto_round:auto_gptq","auto_round:auto_gptq:marlin", "auto_round:gptq:marlin", "auto_round:auto_awq","auto_round:awq"]+ if not args.quant_nontext_module:+ supported_formats.append("auto_gptq","auto_gptq:marlin")+formats = args.format.replace(' ', '').split(",")for format in formats:if format not in supported_formats:@@ -278,7 +281,6 @@ def tune(args):raise ValueError(f"{format} is not supported for lm-head quantization, please change to {auto_round_formats}")-if args.quant_lm_head and args.low_gpu_mem_usage:print(f"warning, low_gpu_mem_usage=False is strongly recommended if the whole model could be loaded to "f"gpu")diff --git a/docs/glm-4-9b-chat-recipe.md b/docs/glm-4-9b-chat-recipe.mdnew file mode 100644index 000000000..abadb9868--- /dev/null+++ b/docs/glm-4-9b-chat-recipe.md@@ -0,0 +1,142 @@+## Model Details++This is an int4 recipe with group_size 128 of [THUDM/glm-4-9b-chat](https://huggingface.co/THUDM/glm-4-9b-chat) generated by [intel/auto-round](https://github.com/intel/auto-round). For GPTQ format, please load the model with revision `d45e33e`++## How To Use++### INT4 Inference on CPU/CUDA/HPU++```python+from auto_round import AutoRoundConfig ## must import for auto-round format+from transformers import AutoModelForCausalLM, AutoTokenizer+import torch++quantized_model_dir = "Intel/glm-4-9b-chat-int4-inc"++backend = "auto" ##cuda, hpu, cpu(supported in auto_round>0.3.1),cuda:marlin(supported in auto_round>0.3.1 'pip install -v gptqmodel --no-build-isolation')+quantization_config = AutoRoundConfig(+ backend=backend+)+model = AutoModelForCausalLM.from_pretrained(quantized_model_dir,+ device_map=backend.split(':')[0],+ torch_dtype=torch.float16,+ quantization_config=quantization_config,+ trust_remote_code=True+ ).eval()++tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir, trust_remote_code=True)+query = "请介绍一下智谱华章科技有限公司"+inputs = tokenizer.apply_chat_template([{"role": "user", "content": query}],+ add_generation_prompt=True,+ tokenize=True,+ return_tensors="pt",+ return_dict=True+ )+inputs = inputs.to(model.device)++gen_kwargs = {"max_length": 50, "do_sample": False, "top_k": 1}##change this to follow official usage+with torch.no_grad():+ outputs = model.generate(**inputs, **gen_kwargs)+ outputs = outputs[:, inputs['input_ids'].shape[1]:]+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))++##请介绍一下智谱华章科技有限公司+#智谱华章科技有限公司是一家专注于人工智能、大数据、云计算等前沿技术领域的创新型企业。公司成立于2016年,总部位于中国北京,是一家集技术研发、产品开发、+++##9.8大还是9.11大+##9.8比9.11小。在数值上,9.8小于9.11。+++##Once upon a time,++#In a land where the sun kissed the horizon with a golden glow and the stars whispered secrets to the night, there was a village nestled among rolling hills and whispering forests. This was a place+++##There is a girl who likes adventure,+##That's quite the intriguing starting point! If you're looking to create a story or a character, here's a brief introduction to a girl who likes adventure:++##---++##**Name:**+```++++### Evaluate the model++pip3 install lm-eval==0.4.5++pip3 install langdetect,immutabledict,antlr4-python3-runtime==4.11++```bash+auto-round --eval --eval_bs 16 --tasks leaderboard_ifeval,leaderboard_mmlu_pro,gsm8k,lambada_openai,hellaswag,piqa,winogrande,truthfulqa_mc1,openbookqa,boolq,arc_easy,arc_challenge,cmmlu,ceval-valid++```++| Metric | BF16 | INT4(6.4G) | INT4-quanted-lm-head(5.5G) |+| ------------------------------------------ | ------ | ---------- | -------------------------- |+| Avg | 0.6260 | 0.6230 | 0.6204 |+| leaderboard_mmlu_pro 5shot | 0.3678 | 0.3616 | 0.3610 |+| leaderboard_ifeval inst_level_strict_acc | 0.5504 | 0.5600 | 0.5588 |+| leaderboard_ifeval prompt_level_strict_acc | 0.4067 | 0.4233 | 0.4067 |+| cmmlu | 0.7213 | 0.7137 | 0.7086 |+| ceval-valid | 0.7065 | 0.7058 | 0.6909 |+| lambada_openai | 0.6608 | 0.6493 | 0.6470 |+| hellaswag | 0.6195 | 0.6137 | 0.6134 |+| winogrande | 0.7561 | 0.7545 | 0.7522 |+| piqa | 0.8030 | 0.7976 | 0.8003 |+| truthfulqa_mc1 | 0.4223 | 0.4223 | 0.4284 |+| openbookqa | 0.3560 | 0.3640 | 0.3580 |+| boolq | 0.8691 | 0.8606 | 0.8578 |+| arc_easy | 0.8241 | 0.8249 | 0.8203 |+| arc_challenge | 0.5469 | 0.5341 | 0.5444 |+| gsm8k 5shot strict match | 0.7794 | 0.7597 | 0.7589 |++++### Generate the model++Here is the sample command to generate the model. AutoRound should include this pr https://github.com/intel/auto-round/pull/304++```bash+auto-round \+--model THUDM/glm-4-9b-chat \+--iter 1000 \+--nsamples 512 \+--disable_eval \+--format "auto_round,auto_gptq" \+--model_dtype "fp16" \+--output_dir "./tmp_autoround"+```++copy all the *.py file to the quantized_model folder++For gptq format, need to add "block_name_to_quantize":"transformer.encoder.layers" to config.json, we only tested it on transformers==4.46.1++## Ethical Considerations and Limitations++The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.++Therefore, before deploying any applications of the model, developers should perform safety testing.++## Caveats and Recommendations++Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.++Here are a couple of useful links to learn more about Intel's AI software:++* Intel Neural Compressor [link](https://github.com/intel/neural-compressor)+* Intel Extension for Transformers [link](https://github.com/intel/intel-extension-for-transformers)++## Disclaimer++The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.++++## Cite++@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }++[arxiv](https://arxiv.org/abs/2309.05516) [github](https://github.com/intel/auto-round)\ No newline at end of filediff --git a/examples/language-modeling/main.py b/examples/language-modeling/main.pyindex 25b50e101..bc0309791 100644--- a/examples/language-modeling/main.py+++ b/examples/language-modeling/main.py@@ -170,7 +170,7 @@ if __name__ == '__main__':args.format = "auto_round"supported_formats = ["auto_round", "auto_gptq", "auto_awq", "auto_round:gptq", "auto_round:auto_gptq","auto_round:auto_gptq:marlin", "auto_round:gptq:marlin", "auto_round:auto_awq",- "auto_round:awq", "auto_awq", "itrex", "iterx_xpu", "fake"]+ "auto_round:awq", "auto_gptq:marlin", "itrex", "iterx_xpu", "fake"]formats = args.format.replace(' ', '').split(",")for format in formats:if format not in supported_formats: