|
ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">在英特尔® 至强® 处理器上实现高性能WOQ INT4推理ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">Qwen2是新一代大型语言模型（LLM）。对于Qwen2，阿里巴巴发布了一列基础语言模型和指令优化语言模型，参数量从0.5亿到720亿不等，包括混合专家模型。在本文中，我们将部署Intel Extension for Transformers来加速英特尔至强处理器上的低比特Qwen2推理。首先，我们使用Intel Extension for Transformers对模型进行4位压缩。我们使用创新的AutoRound算法为Qwen2-1.5B和Qwen2-7B生成INT4权重量化模型。ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">表1显示了两个模型的准确性比较。ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;"> ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">表1.Qwen2 1.5B/7B AutoRound BF16/INT4 准确性比较ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">您可直接从Hugging Face获取INT4模型：ingFang SC", Cambria, Cochin, Georgia, serif;letter-spacing: normal;text-align: start;text-wrap: wrap;" class="list-paddingleft-1">Qwen2–0.5B-instuct-int4-inc Qwen2–1.5B-instuct-int4-inc Qwen2–7B-int4-inc ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">以下示例代码展示了如何使用Intel Extension for Transformers后端Neural Speed进行量化和推理：##pip install auto-round
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen2-1.5B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
from auto_round import AutoRound
bits, group_size, sym = 4, 128, True
device="cpu"
autoround = AutoRound(model, tokenizer, bits=bits, group_size=group_size, sym=sym, device=None)
autoround.quantize()
output_dir = "./qwen2_7b_autoround"
autoround.save_quantized(output_dir)
from transformers import AutoTokenizer
from intel_extension_for_transformers.transformers import AutoModelForCausalLM, GPTQConfig
# Auto Round 量化模型
model_name = "./qwen2_7b_autoround"
prompt = "Once upon a time, a little girl"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
inputs = tokenizer(prompt, return_tensors="pt").input_ids
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
outputs = model.generate(inputs)
ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;">在Qwen2中，所有指令微调模型都在32K长度的上下文中进行训练，并使用如YARN或Dual Chunk Attention等技术推断更长的上下文长度。Intel Extension for Transformers在输入长度达到32K时，下一个令牌延迟仍低于50毫秒：ingFang SC", Cambria, Cochin, Georgia, serif;text-wrap: wrap;"> Intel Extension for Transformers提供了多种方式来在保持可接受准确性的同时最小化令牌延迟。用户友好的类似Transformers的API仅需少量代码更改。还有更高级的功能，如流式LLM、张量并行性和低比特压缩（低至1位），所以我们鼓励您查看Intel Extension for Transformers。
|
