MedGemma 1.5\u662f\u57fa\u4e8eGoogle Gemma\u67b6\u6784\u7684\u533b\u5b66\u601d\u7ef4\u94fe\u63a8\u7406\u5f15\u64ce，\u4e13\u95e8\u4e3a\u533b\u5b66\u54a8\u8be2\u3001\u75c5\u7406\u5206\u6790\u548c\u672f\u8bed\u89e3\u91ca\u8bbe\u8ba1\u3002\u8fd9\u4e2a\u672c\u5730\u533b\u7597AI\u95ee\u7b54\u7cfb\u7edf\u91c7\u7528MedGemma-1.5-4B-IT\u6a21\u578b，\u80fd\u591f\u5728\u5b8c\u5168\u79bb\u7ebf\u73af\u5883\u4e0b\u63d0\u4f9b\u63a5\u8fd1\u4e13\u5bb6\u7ea7\u7684\u533b\u7597\u903b\u8f91\u63a8\u7406\u80fd\u529b\u3002<\/p>\n
\u5bf9\u4e8e\u9700\u8981\u5728ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u4e0a\u90e8\u7f72\u7684\u7528\u6237\u6765\u8bf4，\u6700\u5173\u5fc3\u7684\u95ee\u9898\u662f：\u8fd9\u4e2a\u533b\u5b66AI\u6a21\u578b\u80fd\u5426\u5728\u975ex86\u67b6\u6784\u4e0a\u6b63\u5e38\u8fd0\u884c？\u7b54\u6848\u662f\u80af\u5b9a\u7684\u3002\u7ecf\u8fc7\u5b9e\u9645\u6d4b\u8bd5\u9a8c\u8bc1，MedGemma 1.5\u5df2\u7ecf\u6210\u529f\u9002\u914dARM\u67b6\u6784，\u53ef\u4ee5\u5728NVIDIA Grace\u7b49\u670d\u52a1\u5668\u4e0a\u7a33\u5b9a\u8fd0\u884c\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u524d，\u8bf7\u786e\u4fdd\u60a8\u7684ARM\u67b6\u6784\u670d\u52a1\u5668\u6ee1\u8db3\u4ee5\u4e0b\u8981\u6c42：<\/p>\n
\u9996\u5148\u66f4\u65b0\u7cfb\u7edf\u5e76\u5b89\u88c5\u57fa\u7840\u4f9d\u8d56：<\/p>\n
# \u66f4\u65b0\u7cfb\u7edf\u5305\u5217\u8868
\nsudo apt update && sudo apt upgrade -y<\/p>\n
# \u5b89\u88c5\u57fa\u7840\u7f16\u8bd1\u5de5\u5177
\nsudo apt install -y build-essential cmake git wget<\/p>\n
# \u5b89\u88c5Python\u73af\u5883
\nsudo apt install -y python3 python3-pip python3-venv<\/p>\n
\u5bf9\u4e8eNVIDIA Grace\u5e73\u53f0，CUDA\u901a\u5e38\u5df2\u7ecf\u9884\u88c5\u3002\u5982\u679c\u9700\u8981\u624b\u52a8\u5b89\u88c5：<\/p>\n
# \u68c0\u67e5CUDA\u662f\u5426\u5df2\u5b89\u88c5
\nnvidia-smi<\/p>\n
# \u5982\u679c\u672a\u5b89\u88c5，\u4eceNVIDIA\u5b98\u7f51\u4e0b\u8f7dARM\u7248\u672c\u7684CUDA toolkit
\nwget https:\/\/developer.download.nvidia.com\/compute\/cuda\/12.2.0\/local_installers\/cuda_12.2.0_535.54.03_linux_sbsa.run
\nsudo sh cuda_12.2.0_535.54.03_linux_sbsa.run<\/p>\n
\u4e3a\u907f\u514d\u4f9d\u8d56\u51b2\u7a81，\u5efa\u8bae\u4f7f\u7528\u865a\u62df\u73af\u5883：<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883
\npython3 -m venv medgemma-env<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nsource medgemma-env\/bin\/activate<\/p>\n
ARM\u67b6\u6784\u9700\u8981\u5b89\u88c5\u7279\u5b9a\u7248\u672c\u7684PyTorch：<\/p>\n
# \u5b89\u88c5ARM\u517c\u5bb9\u7684PyTorch
\npip3 install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cu118<\/p>\n
# \u9a8c\u8bc1PyTorch\u80fd\u5426\u8bc6\u522bGPU
\npython3 -c "import torch; print(torch.cuda.is_available()); print(torch.version.cuda)"<\/p>\n
\u5b89\u88c5\u8fd0\u884cMedGemma\u6240\u9700\u7684\u5176\u4ed6\u4f9d\u8d56：<\/p>\n
# \u5b89\u88c5Transformers\u548c\u5176\u4ed6AI\u5e93
\npip install transformers>=4.35.0 accelerate>=0.24.0<\/p>\n
# \u5b89\u88c5Web\u754c\u9762\u76f8\u5173\u4f9d\u8d56
\npip install gradio==3.50.0 sentencepiece protobuf<\/p>\n
# \u5b89\u88c5\u533b\u5b66\u4e13\u4e1a\u8bcd\u5e93（\u53ef\u9009）
\npip install medspacy<\/p>\n
\u7531\u4e8e\u6a21\u578b\u6587\u4ef6\u8f83\u5927，\u5efa\u8bae\u4f7f\u7528huggingface_hub\u5e93\u4e0b\u8f7d：<\/p>\n
from huggingface_hub import snapshot_download<\/p>\n
# \u4e0b\u8f7dMedGemma-1.5-4B-IT\u6a21\u578b
\nmodel_path = snapshot_download(
\n "google\/medgemma-1.5-4b-it",
\n local_dir=".\/medgemma-1.5-4b-it",
\n ignore_patterns=["*.msgpack", "*.h5", "*.ot"]
\n)<\/p>\n
\u521b\u5efa\u6d4b\u8bd5\u811a\u672c\u9a8c\u8bc1ARM\u67b6\u6784\u517c\u5bb9\u6027：<\/p>\n
# architecture_test.py
\nimport platform
\nimport torch
\nimport transformers<\/p>\n
print(f"\u7cfb\u7edf\u67b6\u6784: {platform.machine()}")
\nprint(f"PyTorch\u7248\u672c: {torch.__version__}")
\nprint(f"CUDA\u53ef\u7528: {torch.cuda.is_available()}")
\nprint(f"GPU\u6570\u91cf: {torch.cuda.device_count()}")<\/p>\n
if torch.cuda.is_available():
\n print(f"\u5f53\u524dGPU: {torch.cuda.get_device_name(0)}")
\n print(f"GPU\u5185\u5b58: {torch.cuda.get_device_properties(0).total_memory \/ 1024**3:.1f} GB")<\/p>\n
# \u6d4b\u8bd5\u57fa\u672c\u7684tensor\u64cd\u4f5c
\ntensor = torch.randn(3, 3).cuda()
\nprint(f"Tensor\u8fd0\u7b97\u6d4b\u8bd5: {tensor.sum()}")<\/p>\n
\u8fd0\u884c\u6d4b\u8bd5\u811a\u672c\u786e\u8ba4\u73af\u5883\u6b63\u5e38：<\/p>\n
python architecture_test.py<\/p>\n
\u9a8c\u8bc1\u6a21\u578b\u80fd\u5426\u5728ARM\u67b6\u6784\u4e0a\u6b63\u786e\u52a0\u8f7d：<\/p>\n
# model_load_test.py
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM
\nimport torch<\/p>\n
# \u521d\u59cb\u5316tokenizer\u548cmodel
\ntokenizer = AutoTokenizer.from_pretrained(".\/medgemma-1.5-4b-it")
\nmodel = AutoModelForCausalLM.from_pretrained(
\n ".\/medgemma-1.5-4b-it",
\n torch_dtype=torch.float16,
\n device_map="auto"
\n)<\/p>\n
print("\u6a21\u578b\u52a0\u8f7d\u6210\u529f！")
\nprint(f"\u6a21\u578b\u8bbe\u5907: {model.device}")
\nprint(f"\u6a21\u578b\u53c2\u6570\u91cf: {sum(p.numel() for p in model.parameters()):,}")<\/p>\n
\u521b\u5efa\u542f\u52a8\u6587\u4ef6medgemma_launch.py：<\/p>\n
import gradio as gr
\nfrom transformers import AutoTokenizer, AutoModelForCausalLlM
\nimport torch<\/p>\n
# \u52a0\u8f7d\u6a21\u578b
\ntokenizer = AutoTokenizer.from_pretrained(".\/medgemma-1.5-4b-it")
\nmodel = AutoModelForCausalLM.from_pretrained(
\n ".\/medgemma-1.5-4b-it",
\n torch_dtype=torch.float16,
\n device_map="auto"
\n)<\/p>\n
def medical_chatbot(question):
\n # \u6784\u5efa\u533b\u5b66\u95ee\u7b54\u63d0\u793a
\n prompt = f"<start_of_turn>user\\\\n{question}<end_of_turn>\\\\n<start_of_turn>model"<\/p>\n
# \u751f\u6210\u56de\u7b54
\n inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
\n outputs = model.generate(
\n **inputs,
\n max_new_tokens=512,
\n temperature=0.7,
\n do_sample=True
\n )<\/p>\n
# \u89e3\u7801\u5e76\u8fd4\u56de\u7ed3\u679c
\n response = tokenizer.decode(outputs[0], skip_special_tokens=True)
\n return response.split("<start_of_turn>model")[-1].strip()<\/p>\n
# \u521b\u5efaGradio\u754c\u9762
\niface = gr.Interface(
\n fn=medical_chatbot,
\n inputs=gr.Textbox(label="\u533b\u5b66\u95ee\u9898", placeholder="\u8bf7\u8f93\u5165\u60a8\u7684\u533b\u5b66\u95ee\u9898…"),
\n outputs=gr.Textbox(label="MedGemma\u56de\u7b54"),
\n title="🩺 MedGemma 1.5 \u533b\u7597AI\u52a9\u624b – ARM\u67b6\u6784\u7248",
\n description="\u57fa\u4e8eMedGemma-1.5-4B-IT\u7684\u672c\u5730\u533b\u7597\u95ee\u7b54\u7cfb\u7edf，\u8fd0\u884c\u5728ARM\u67b6\u6784\u670d\u52a1\u5668\u4e0a"
\n)<\/p>\n
# \u542f\u52a8\u670d\u52a1
\niface.launch(server_name="0.0.0.0", server_port=6006, share=False)<\/p>\n
\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u542f\u52a8\u533b\u7597\u95ee\u7b54\u7cfb\u7edf：<\/p>\n
python medgemma_launch.py<\/p>\n
\u670d\u52a1\u542f\u52a8\u540e，\u5728\u6d4f\u89c8\u5668\u4e2d\u8bbf\u95ee http:\/\/\u670d\u52a1\u5668IP:6006 \u5373\u53ef\u4f7f\u7528\u533b\u7597\u95ee\u7b54\u529f\u80fd\u3002<\/p>\n
\u9488\u5bf9ARM\u67b6\u6784\u7684\u5185\u5b58\u7279\u6027\u8fdb\u884c\u4f18\u5316：<\/p>\n
# \u5728\u6a21\u578b\u52a0\u8f7d\u65f6\u6dfb\u52a0\u5185\u5b58\u4f18\u5316\u53c2\u6570
\nmodel = AutoModelForCausalLM.from_pretrained(
\n ".\/medgemma-1.5-4b-it",
\n torch_dtype=torch.float16,
\n device_map="auto",
\n low_cpu_mem_usage=True, # \u51cf\u5c11CPU\u5185\u5b58\u4f7f\u7528
\n offload_folder=".\/offload" # \u8bbe\u7f6eoffload\u76ee\u5f55
\n)<\/p>\n
\u4f7f\u7528\u66f4\u597d\u7684\u63a8\u7406\u914d\u7f6e\u63d0\u5347\u901f\u5ea6：<\/p>\n
# \u4f18\u5316\u751f\u6210\u53c2\u6570
\noutputs = model.generate(
\n **inputs,
\n max_new_tokens=512,
\n temperature=0.7,
\n do_sample=True,
\n use_cache=True, # \u542f\u7528\u7f13\u5b58\u52a0\u901f
\n pad_token_id=tokenizer.eos_token_id
\n)<\/p>\n
\u5bf9\u4e8e\u591a\u7528\u6237\u573a\u666f，\u5b9e\u73b0\u6279\u5904\u7406：<\/p>\n
def batch_medical_qa(questions):
\n # \u6279\u91cf\u5904\u7406\u533b\u5b66\u95ee\u9898
\n prompts = [
\n f"<start_of_turn>user\\\\n{q}<end_of_turn>\\\\n<start_of_turn>model"
\n for q in questions
\n ]<\/p>\n
inputs = tokenizer(prompts, return_tensors="pt", padding=True, truncation=True).to(model.device)<\/p>\n
with torch.no_grad():
\n outputs = model.generate(
\n **inputs,
\n max_new_tokens=256,
\n temperature=0.7,
\n do_sample=True
\n )<\/p>\n
responses = []
\n for i in range(len(questions)):
\n response = tokenizer.decode(outputs[i], skip_special_tokens=True)
\n clean_response = response.split("<start_of_turn>model")[-1].strip()
\n responses.append(clean_response)<\/p>\n
return responses<\/p>\n
\u5982\u679c\u9047\u5230\u5185\u5b58\u4e0d\u8db3\u9519\u8bef，\u5c1d\u8bd5\u4ee5\u4e0b\u89e3\u51b3\u65b9\u6848：<\/p>\n
# \u51cf\u5c11\u6a21\u578b\u7cbe\u5ea6
\nmodel = model.half() # \u4f7f\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570<\/p>\n
# \u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9
\nmodel.gradient_checkpointing_enable()<\/p>\n
# \u4f7f\u7528\u5185\u5b58\u6620\u5c04
\nmodel = AutoModelForCausalLM.from_pretrained(
\n ".\/medgemma-1.5-4b-it",
\n torch_dtype=torch.float16,
\n device_map="auto",
\n offload_state_dict=True # \u79bb\u7ebf\u52a0\u8f7d\u72b6\u6001\u5b57\u5178
\n)<\/p>\n
\u5982\u679c\u63a8\u7406\u901f\u5ea6\u8f83\u6162，\u53ef\u4ee5\u5c1d\u8bd5：<\/p>\n
# \u542f\u7528TensorRT\u52a0\u901f（\u5982\u679c\u53ef\u7528）
\nimport torch_tensorrt
\nmodel = torch_tensorrt.compile(model, inputs=[torch_tensorrt.Input((1, 512), dtype=torch.int32)])<\/p>\n
# \u6216\u8005\u4f7f\u7528\u66f4\u597d\u7684GPU\u914d\u7f6e
\ntorch.backends.cudnn.benchmark = True
\ntorch.set_float32_matmul_precision('high')<\/p>\n
\u7ecf\u8fc7\u5728NVIDIA Grace ARM\u67b6\u6784\u670d\u52a1\u5668\u4e0a\u7684\u5168\u9762\u6d4b\u8bd5，MedGemma 1.5\u8868\u73b0\u51fa\u826f\u597d\u7684\u517c\u5bb9\u6027\u548c\u7a33\u5b9a\u6027\u3002\u4ee5\u4e0b\u662f\u9a8c\u8bc1\u7ed3\u679c：<\/p>\n
MedGemma 1.5\u5728ARM\u67b6\u6784\u670d\u52a1\u5668\u4e0a\u7684\u6210\u529f\u90e8\u7f72，\u4e3a\u533b\u7597\u884c\u4e1a\u63d0\u4f9b\u4e86\u66f4\u7075\u6d3b\u7684AI\u89e3\u51b3\u65b9\u6848\u9009\u62e9\u3002\u7279\u522b\u662f\u5728\u5bf9\u6570\u636e\u9690\u79c1\u8981\u6c42\u6781\u9ad8\u7684\u533b\u7597\u573a\u666f，\u672c\u5730\u5316\u90e8\u7f72\u5728ARM\u670d\u52a1\u5668\u4e0a\u65e2\u80fd\u4fdd\u8bc1\u6570\u636e\u5b89\u5168，\u53c8\u80fd\u63d0\u4f9b\u4e13\u4e1a\u7684\u533b\u7597AI\u670d\u52a1\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
MedGemma 1.5\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u9002\u914d\u8fd0\u884c\u53ef\u884c\u6027\u9a8c\u8bc1
\n1. \u9879\u76ee\u6982\u8ff0
\nMedGemma 1.5\u662f\u57fa\u4e8eGoogle Gemma\u67b6\u6784\u7684\u533b\u5b66\u601d\u7ef4\u94fe\u63a8\u7406\u5f15\u64ce，\u4e13\u95e8\u4e3a\u533b\u5b66\u54a8\u8be2\u3001\u75c5\u7406\u5206\u6790\u548c\u672f\u8bed\u89e3\u91ca\u8bbe\u8ba1\u3002\u8fd9\u4e2a\u672c\u5730\u533b\u7597AI\u95ee\u7b54\u7cfb\u7edf\u91c7\u7528MedGemma-1.5-4B-IT\u6a21\u578b，\u80fd\u591f\u5728\u5b8c\u5168\u79bb\u7ebf\u73af\u5883\u4e0b\u63d0\u4f9b\u63a5\u8fd1\u4e13\u5bb6\u7ea7\u7684\u533b\u7597\u903b\u8f91\u63a8\u7406\u80fd\u529b\u3002
\n\u5bf9\u4e8e\u9700\u8981\u5728ARM\u67b6\u6784\u670d\u52a1\u5668（<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[7591,6714,9720,9544],"topic":[],"class_list":["post-86287","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-arm","tag-medgemma"],"yoast_head":"\n