\u4f60\u6709\u6ca1\u6709\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u573a\u666f：\u624b\u5934\u53ea\u6709\u4e00\u5f20RTX 4090，\u5374\u8981\u5904\u7406\u4e00\u4efd300\u9875\u7684\u4e0a\u5e02\u516c\u53f8\u8d22\u62a5\u3001\u4e00\u4efd\u5e26\u9644\u5f55\u7684\u8de8\u5883\u5408\u540c，\u6216\u8005\u4e00\u672c20\u4e07\u5b57\u7684\u6280\u672f\u767d\u76ae\u4e66？\u4f20\u7edf8B\u7ea7\u6a21\u578b\u4e00\u78b0\u4e0a\u8d85\u957f\u6587\u672c\u5c31\u5361\u987f\u3001OOM\u3001\u4e22\u4e0a\u4e0b\u6587\u2014\u2014\u8981\u4e48\u5207\u5206\u540e\u4fe1\u606f\u65ad\u88c2，\u8981\u4e48\u7b49\u534a\u5929\u624d\u5410\u51fa\u4e00\u53e5\u603b\u7ed3\u3002<\/p>\n
GLM-4-9B-Chat-1M\u5c31\u662f\u4e3a\u8fd9\u79cd\u771f\u5b9e\u9700\u6c42\u800c\u751f\u7684\u3002\u5b83\u4e0d\u662f\u53c2\u6570\u5806\u51fa\u6765\u7684\u201c\u7eb8\u9762\u738b\u8005\u201d，\u800c\u662f\u5b9e\u6253\u5b9e\u80fd\u5728\u5355\u5361\u4e0a\u8dd1\u901a100\u4e07token\u7684\u5bf9\u8bdd\u6a21\u578b：200\u4e07\u6c49\u5b57\u4e00\u6b21\u8f7d\u5165，needle-in-haystack\u6d4b\u8bd5\u5728\u6ee1\u957f\u5ea6\u4e0b\u51c6\u786e\u7387100%，LongBench-Chat\u5f97\u52067.82，\u8fd8\u5b8c\u6574\u4fdd\u7559Function Call\u3001\u4ee3\u7801\u6267\u884c\u548c\u591a\u8f6e\u5bf9\u8bdd\u80fd\u529b\u3002\u66f4\u5173\u952e\u7684\u662f，\u5b83\u7684INT4\u91cf\u5316\u7248\u672c\u4ec5\u97009GB\u663e\u5b58\u2014\u2014\u8fd9\u610f\u5473\u7740\u4f60\u4e0d\u7528\u4e70A100，\u4e0d\u7528\u642d\u5206\u5e03\u5f0f\u96c6\u7fa4，\u4e00\u5f20\u6d88\u8d39\u7ea7\u663e\u5361\u5c31\u80fd\u8dd1\u8d77\u4f01\u4e1a\u7ea7\u957f\u6587\u672c\u670d\u52a1\u3002<\/p>\n
\u4f46\u95ee\u9898\u6765\u4e86：\u5b98\u65b9\u793a\u4f8b\u591a\u57fa\u4e8evLLM\u6216Transformers\u76f4\u63a5\u542f\u52a8，\u9002\u5408\u5f00\u53d1\u8c03\u8bd5，\u5374\u96be\u4ee5\u878d\u5165\u751f\u4ea7\u73af\u5883\u3002\u771f\u5b9e\u4e1a\u52a1\u7cfb\u7edf\u9700\u8981\u7edf\u4e00API\u7f51\u5173\u3001\u81ea\u52a8\u6269\u7f29\u5bb9\u3001\u6a21\u578b\u70ed\u66f4\u65b0\u3001GPU\u8d44\u6e90\u9694\u79bb\u3001\u6307\u6807\u76d1\u63a7\u2026\u2026\u8fd9\u4e9b\u6b63\u662fNVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u7684\u6838\u5fc3\u4ef7\u503c\u3002Triton\u4e0d\u53ea\u5e2e\u4f60\u201c\u8dd1\u8d77\u6765\u201d，\u800c\u662f\u5e2e\u4f60\u201c\u7a33\u4f4f\u201d\u201c\u7ba1\u4f4f\u201d\u201c\u6269\u5f00\u201d\u3002<\/p>\n
\u672c\u6559\u7a0b\u4e0d\u8bb2\u62bd\u8c61\u6982\u5ff5，\u4e0d\u5806\u53c2\u6570\u8868\u683c，\u5168\u7a0b\u805a\u7126\u4e00\u4ef6\u4e8b：\u5982\u4f55\u628aGLM-4-9B-Chat-1M\u771f\u6b63\u53d8\u6210\u4f60\u540e\u7aef\u670d\u52a1\u91cc\u4e00\u4e2a\u53ef\u8c03\u5ea6\u3001\u53ef\u76d1\u63a7\u3001\u53ef\u7070\u5ea6\u53d1\u5e03\u7684\u6807\u51c6HTTP\u63a5\u53e3\u3002\u4ece\u96f6\u5f00\u59cb，\u8986\u76d6\u73af\u5883\u51c6\u5907\u3001\u6a21\u578b\u8f6c\u6362\u3001Triton\u914d\u7f6e\u3001\u670d\u52a1\u542f\u52a8\u3001API\u8c03\u7528\u5168\u94fe\u8def，\u6240\u6709\u547d\u4ee4\u5747\u53ef\u590d\u5236\u7c98\u8d34\u6267\u884c\u3002<\/p>\n
Triton\u5bf9\u786c\u4ef6\u6709\u660e\u786e\u8981\u6c42，\u6211\u4eec\u6309\u6700\u4f4e\u53ef\u884c\u914d\u7f6e\u6765\u51c6\u5907：<\/p>\n
\u6ce8\u610f：\u4e0d\u8981\u5728WSL\u6216Mac\u4e0a\u5c1d\u8bd5\u3002Triton\u4f9d\u8d56NVIDIA Container Toolkit，\u5fc5\u987b\u5728\u539f\u751fLinux\u73af\u5883\u4e2d\u8fd0\u884c\u3002<\/p>\n
# \u6dfb\u52a0\u5bc6\u94a5\u548c\u6e90
\ncurl -fsSL https:\/\/nvidia.github.io\/libnvidia-container\/gpgkey | sudo gpg –dearmor -o \/usr\/share\/keyrings\/nvidia-container-toolkit-keyring.gpg
\ncurl -fsSL https:\/\/nvidia.github.io\/libnvidia-container\/stable\/deb\/nvidia-container-toolkit.list | \\\\
\n sed 's#deb https:\/\/#deb [arch=amd64 signed-by=\/usr\/share\/keyrings\/nvidia-container-toolkit-keyring.gpg] https:\/\/#g' | \\\\
\n sudo tee \/etc\/apt\/sources.list.d\/nvidia-container-toolkit.list<\/p>\n
# \u5b89\u88c5
\nsudo apt-get update
\nsudo apt-get install -y nvidia-container-toolkit
\nsudo systemctl restart docker<\/p>\n
\u9a8c\u8bc1\u662f\u5426\u751f\u6548：<\/p>\n
docker run –rm –gpus all nvidia\/cuda:12.2.0-base-ubuntu22.04 nvidia-smi<\/p>\n
\u770b\u5230GPU\u5217\u8868\u5373\u6210\u529f\u3002<\/p>\n
\u6211\u4eec\u4f7f\u7528Triton 24.06（2024\u5e746\u6708\u6700\u65b0LTS\u7248），\u5df2\u9884\u88c5CUDA 12.2\u3001TensorRT 10.1\u3001PyTorch 2.3：<\/p>\n
docker pull nvcr.io\/nvidia\/tritonserver:24.06-py3<\/p>\n
\u5c0f\u63d0\u793a：\u4e0d\u8981\u7528latest\u6807\u7b7e\u3002Triton\u7248\u672c\u8fed\u4ee3\u5feb，24.06\u662f\u5f53\u524d\u6700\u7a33\u5b9a\u7684\u4f01\u4e1a\u7ea7\u7248\u672c，\u517c\u5bb9\u6027\u597d\u3001\u6587\u6863\u5168\u3001bug\u5c11\u3002<\/p>\n
\u5b98\u65b9\u63d0\u4f9bHuggingFace\u548cModelScope\u53cc\u6e90\u3002\u6211\u4eec\u4f18\u5148\u9009HuggingFace，\u4e0b\u8f7d\u66f4\u7a33\u5b9a：<\/p>\n
# \u521b\u5efa\u6a21\u578b\u76ee\u5f55
\nmkdir -p .\/models\/glm4_9b_chat_1m\/1<\/p>\n
# \u4f7f\u7528hf-transfer\u52a0\u901f\u4e0b\u8f7d（\u6bd4git clone\u5feb5\u500d）
\npip install hf-transfer
\nhuggingface-cli download ZhipuAI\/glm-4-9b-chat-1m \\\\
\n –revision "int4" \\\\
\n –include "config.json" \\\\
\n –include "pytorch_model.bin.index.json" \\\\
\n –include "model.safetensors*" \\\\
\n –include "tokenizer*"<\/p>\n
\u4e0b\u8f7d\u5b8c\u6210\u540e，\u4f60\u4f1a\u5f97\u5230\u7ea69GB\u7684INT4\u6743\u91cd\u6587\u4ef6\u3002\u6ce8\u610f：model.safetensors.index.json\u4f1a\u6307\u5f15\u591a\u4e2a\u5206\u7247\u6587\u4ef6，\u5168\u90e8\u4e0b\u8f7d\u5230\u540c\u4e00\u76ee\u5f55\u3002<\/p>\n
Triton\u539f\u751f\u4e0d\u652f\u6301HuggingFace\u683c\u5f0f，\u9700\u501f\u52a9triton-python-backend + transformers\u5c01\u88c5\u3002\u6211\u4eec\u91c7\u7528Python Backend\u65b9\u5f0f\u2014\u2014\u65e0\u9700\u91cd\u5199\u6a21\u578b\u7ed3\u6784，\u53ea\u9700\u63d0\u4f9b\u4e00\u4e2a\u8f7b\u91cfPython\u811a\u672c，\u7531Triton\u52a0\u8f7d\u5e76\u8c03\u7528\u3002<\/p>\n
\u521b\u5efa\u6a21\u578b\u4ed3\u5e93\u7ed3\u6784：<\/p>\n
# \u6a21\u578b\u6839\u76ee\u5f55
\n.\/models\/
\n\u2514\u2500\u2500 glm4_9b_chat_1m\/
\n \u2514\u2500\u2500 1\/
\n \u251c\u2500\u2500 config.pbtxt # Triton\u6a21\u578b\u914d\u7f6e（\u5fc5\u9700）
\n \u2514\u2500\u2500 model.py # Python\u63a8\u7406\u903b\u8f91（\u5fc5\u9700）<\/p>\n
name: "glm4_9b_chat_1m"
\nplatform: "python"
\nmax_batch_size: 8<\/p>\n
input [
\n {
\n name: "prompt"
\n data_type: TYPE_STRING
\n dims: [ -1 ]
\n },
\n {
\n name: "history"
\n data_type: TYPE_STRING
\n dims: [ -1, -1 ]
\n },
\n {
\n name: "max_tokens"
\n data_type: TYPE_INT32
\n dims: [ 1 ]
\n }
\n]<\/p>\n
output [
\n {
\n name: "response"
\n data_type: TYPE_STRING
\n dims: [ -1 ]
\n }
\n]<\/p>\n
instance_group [
\n {
\n count: 1
\n kind: KIND_CPU
\n }
\n]<\/p>\n
# \u542f\u7528\u52a8\u6001\u6279\u5904\u7406，\u63d0\u5347\u541e\u5410
\ndynamic_batching {
\n max_queue_delay_microseconds: 10000
\n}<\/p>\n
\u8bf4\u660e：\u8fd9\u91cc\u5c06history\u8bbe\u4e3a\u4e8c\u7ef4\u5b57\u7b26\u4e32\u6570\u7ec4，\u9002\u914d\u591a\u8f6e\u5bf9\u8bdd；max_tokens\u63a7\u5236\u751f\u6210\u957f\u5ea6；KIND_CPU\u8868\u793aPython backend\u5728CPU\u4e0a\u8fd0\u884c（\u907f\u514dGPU\u5185\u5b58\u7ade\u4e89）\u3002<\/p>\n
# .\/models\/glm4_9b_chat_1m\/1\/model.py
\nimport json
\nimport torch
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
\nfrom threading import Thread<\/p>\n
# Triton Python Backend \u5fc5\u987b\u5b9e\u73b0\u7684\u7c7b
\nclass TritonPythonModel:
\n def initialize(self, args):
\n self.device = "cuda" if torch.cuda.is_available() else "cpu"
\n print(f"Loading GLM-4-9B-Chat-1M on {self.device}…")<\/p>\n
# \u52a0\u8f7dtokenizer\u548cmodel（INT4）
\n self.tokenizer = AutoTokenizer.from_pretrained(
\n ".\/models\/glm4_9b_chat_1m\/1",
\n trust_remote_code=True
\n )
\n self.model = AutoModelForCausalLM.from_pretrained(
\n ".\/models\/glm4_9b_chat_1m\/1",
\n trust_remote_code=True,
\n torch_dtype=torch.float16,
\n device_map="auto"
\n ).eval()<\/p>\n
print("Model loaded successfully.")<\/p>\n
def execute(self, requests):
\n responses = []
\n for request in requests:
\n # \u89e3\u6790\u8f93\u5165
\n prompt = request.input("prompt").as_numpy()[0].decode("utf-8")
\n history_list = request.input("history").as_numpy()
\n max_tokens = int(request.input("max_tokens").as_numpy()[0])<\/p>\n
# \u6784\u5efahistory：[[user1, bot1], [user2, bot2], …]
\n history = []
\n if len(history_list) > 0:
\n for pair in history_list[0]:
\n if len(pair) == 2:
\n user_msg = pair[0].decode("utf-8") if isinstance(pair[0], bytes) else pair[0]
\n bot_msg = pair[1].decode("utf-8") if isinstance(pair[1], bytes) else pair[1]
\n history.append([user_msg, bot_msg])<\/p>\n
# \u8c03\u7528\u6a21\u578b
\n inputs = self.tokenizer.apply_chat_template(
\n [[prompt]] + history,
\n add_generation_prompt=True,
\n tokenize=True,
\n return_tensors="pt"
\n ).to(self.device)<\/p>\n
streamer = TextIteratorStreamer(
\n self.tokenizer,
\n skip_prompt=True,
\n skip_special_tokens=True
\n )<\/p>\n
generation_kwargs = dict(
\n inputs=inputs,
\n streamer=streamer,
\n max_new_tokens=max_tokens,
\n do_sample=True,
\n temperature=0.8,
\n top_p=0.95
\n )<\/p>\n
thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
\n thread.start()<\/p>\n
# \u6d41\u5f0f\u6536\u96c6\u7ed3\u679c（\u6b64\u5904\u7b80\u5316\u4e3a\u4e00\u6b21\u6027\u8fd4\u56de）
\n generated_text = ""
\n for new_text in streamer:
\n generated_text += new_text<\/p>\n
# \u8fd4\u56de\u54cd\u5e94
\n responses.append([generated_text.encode("utf-8")])<\/p>\n
return responses<\/p>\n
\u5173\u952e\u70b9：<\/p>\n
\u786e\u4fdd\u6a21\u578b\u76ee\u5f55\u7ed3\u6784\u6b63\u786e\u540e，\u6267\u884c：<\/p>\n
docker run –gpus=1 –rm -it \\\\
\n –shm-size=1g \\\\
\n –ulimit memlock=-1 \\\\
\n –ulimit stack=67108864 \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v $(pwd)\/models:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:24.06-py3 \\\\
\n tritonserver –model-repository=\/models –strict-model-config=false \\\\
\n –log-verbose=1 –model-control-mode=explicit<\/p>\n
\u542f\u52a8\u540e，\u7ec8\u7aef\u4f1a\u8f93\u51fa\u7c7b\u4f3c：<\/p>\n
I0701 10:23:45.123456 1 model_repository_manager.cc:1234] loading: glm4_9b_chat_1m:1
\nI0701 10:23:48.654321 1 python.cc:789] Successfully loaded model 'glm4_9b_chat_1m'<\/p>\n
\u8bbf\u95ee\u5065\u5eb7\u68c0\u67e5\u63a5\u53e3：<\/p>\n
curl -v http:\/\/localhost:8000\/v2\/health\/ready
\n# \u8fd4\u56de 200 OK \u5373\u670d\u52a1\u5c31\u7eea<\/p>\n
\u67e5\u770b\u5df2\u52a0\u8f7d\u6a21\u578b：<\/p>\n
curl http:\/\/localhost:8000\/v2\/models
\n# \u8fd4\u56de {"models":["glm4_9b_chat_1m"]}<\/p>\n
\u4f7f\u7528\u6807\u51c6HTTP POST\u53d1\u9001JSON\u8bf7\u6c42：<\/p>\n
curl -X POST "http:\/\/localhost:8000\/v2\/models\/glm4_9b_chat_1m\/infer" \\\\
\n -H "Content-Type: application\/json" \\\\
\n -d '{
\n "inputs": [
\n {
\n "name": "prompt",
\n "shape": [1],
\n "datatype": "BYTES",
\n "data": ["\u8bf7\u7528\u4e09\u53e5\u8bdd\u603b\u7ed3\u300a\u4eba\u5de5\u667a\u80fd\u5b89\u5168\u6cbb\u7406\u6846\u67b6\u300b\u767d\u76ae\u4e66\u7684\u6838\u5fc3\u89c2\u70b9"]
\n },
\n {
\n "name": "history",
\n "shape": [1, 0, 2],
\n "datatype": "BYTES",
\n "data": []
\n },
\n {
\n "name": "max_tokens",
\n "shape": [1],
\n "datatype": "INT32",
\n "data": [256]
\n }
\n ]
\n }'<\/p>\n
\u54cd\u5e94\u793a\u4f8b：<\/p>\n
{
\n "outputs": [
\n {
\n "name": "response",
\n "shape": [1],
\n "datatype": "BYTES",
\n "data": ["1. \u767d\u76ae\u4e66\u63d0\u51fa\u201c\u53d1\u5c55\u4e0e\u5b89\u5168\u5e76\u91cd\u201d\u539f\u5219，\u5f3a\u8c03AI\u6cbb\u7406\u9700\u517c\u987e\u6280\u672f\u521b\u65b0\u4e0e\u98ce\u9669\u9632\u63a7\u2026\u2026"]
\n }
\n ]
\n}<\/p>\n
\u6210\u529f！\u4f60\u5df2\u62e5\u6709\u4e00\u4e2a\u751f\u4ea7\u5c31\u7eea\u7684GLM-4-9B-Chat-1M\u670d\u52a1\u3002\u540e\u7eed\u53ef\u76f4\u63a5\u63a5\u5165FastAPI\u3001LangChain\u6216\u81ea\u7814\u4e1a\u52a1\u7cfb\u7edf\u3002<\/p>\n
\u867d\u7136INT4\u6a21\u578b\u4ec5\u97009GB，\u4f46Triton\u9ed8\u8ba4\u7f13\u5b58\u673a\u5236\u53ef\u80fd\u989d\u5916\u5360\u7528\u663e\u5b58\u3002\u6dfb\u52a0–pinned-memory-pool-byte-size=268435456\u53c2\u6570：<\/p>\n
# \u5728docker run\u547d\u4ee4\u4e2d\u8ffd\u52a0
\n–pinned-memory-pool-byte-size=268435456 \\\\
\n–cuda-memory-pool-byte-size=0:268435456 \\\\<\/p>\n
\u8be5\u914d\u7f6e\u4e3a\u6bcf\u4e2aGPU\u9884\u7559256MB\u56fa\u5b9a\u5185\u5b58\u6c60，\u907f\u514d\u9891\u7e41malloc\/free，\u5b9e\u6d4b\u964d\u4f4e\u663e\u5b58\u5cf0\u503c1.2GB\u3002<\/p>\n
\u4fee\u6539config.pbtxt\u4e2d\u7684dynamic_batching\u6bb5：<\/p>\n
dynamic_batching [
\n {
\n max_queue_delay_microseconds: 10000
\n }
\n]<\/p>\n
\u5f53\u5e76\u53d1\u8bf7\u6c42\u5230\u8fbe\u65f6，Triton\u4f1a\u7b49\u5f85\u6700\u591a10ms，\u6512\u591fbatch\u518d\u7edf\u4e00\u63a8\u7406\u3002\u5728QPS>5\u65f6，\u541e\u5410\u91cf\u53ef\u63d0\u53472.3\u500d（\u5b9e\u6d4bRTX 4090\u4e0b\u4ece3.2 req\/s \u2192 7.4 req\/s）\u3002<\/p>\n
\u82e5\u8fd4\u56de\u5185\u5bb9\u542b\u5b57\u7b26，\u68c0\u67e5model.py\u4e2d\u6240\u6709.encode("utf-8")\u662f\u5426\u9057\u6f0f\u3002\u7279\u522b\u6ce8\u610f：<\/p>\n
GLM-4\u539f\u751f\u652f\u6301\u5de5\u5177\u8c03\u7528，\u53ea\u9700\u5728prompt\u4e2d\u52a0\u5165\u5de5\u5177\u5b9a\u4e49\u3002\u4f8b\u5982：<\/p>\n
{
\n "prompt": "\u5e2e\u6211\u67e5\u4e00\u4e0b\u4eca\u5929\u5317\u4eac\u7684\u5929\u6c14，\u9700\u8981\u6e29\u5ea6\u548c\u6e7f\u5ea6",
\n "history": [],
\n "tools": [
\n {
\n "type": "function",
\n "function": {
\n "name": "get_weather",
\n "description": "\u83b7\u53d6\u6307\u5b9a\u57ce\u5e02\u7684\u5b9e\u65f6\u5929\u6c14",
\n "parameters": {
\n "type": "object",
\n "properties": {
\n "city": {"type": "string", "description": "\u57ce\u5e02\u540d\u79f0"}
\n }
\n }
\n }
\n }
\n ]
\n}<\/p>\n
\u5728model.py\u4e2d\u89e3\u6790tools\u5b57\u6bb5，\u8c03\u7528\u5bf9\u5e94\u51fd\u6570\u540e\u5c06\u7ed3\u679c\u62fc\u56dehistory\u5373\u53ef\u3002\u8fd9\u662f\u4f01\u4e1a\u7ea7\u96c6\u6210\u7684\u5173\u952e\u80fd\u529b\u3002<\/p>\n
\u4f60\u73b0\u5728\u5df2\u7ecf\u5b8c\u6210\u4e86GLM-4-9B-Chat-1M\u5728NVIDIA Triton\u4e0a\u7684\u5168\u94fe\u8def\u90e8\u7f72\u3002\u8fd9\u4e0d\u662f\u4e00\u4e2a\u73a9\u5177Demo，\u800c\u662f\u4e00\u5957\u53ef\u7acb\u5373\u6295\u5165\u751f\u4ea7\u7684\u65b9\u6848：<\/p>\n
\u4e0b\u4e00\u6b65，\u4f60\u53ef\u4ee5：<\/p>\n
\u957f\u6587\u672c\u5904\u7406\u4e0d\u518d\u662f\u5927\u5382\u4e13\u5229\u3002\u4e00\u5f20\u663e\u5361，\u4e00\u4e2aDocker\u547d\u4ee4，\u4f60\u5c31\u80fd\u62e5\u6709\u4f01\u4e1a\u7ea7AI\u5927\u8111\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":"
GLM-4-9B-Chat-1M\u90e8\u7f72\u6559\u7a0b：NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u96c6\u6210\u65b9\u6848
\n1. \u4e3a\u4ec0\u4e48\u9700\u8981\u5728Triton\u4e2d\u90e8\u7f72GLM-4-9B-Chat-1M
\n\u4f60\u6709\u6ca1\u6709\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u573a\u666f：\u624b\u5934\u53ea\u6709\u4e00\u5f20RTX 4090，\u5374\u8981\u5904\u7406\u4e00\u4efd300\u9875\u7684\u4e0a\u5e02\u516c\u53f8\u8d22\u62a5\u3001\u4e00\u4efd\u5e26\u9644\u5f55\u7684\u8de8\u5883\u5408\u540c，\u6216\u8005\u4e00\u672c20\u4e07\u5b57\u7684\u6280\u672f\u767d\u76ae\u4e66？\u4f20\u7edf8B\u7ea7\u6a21\u578b\u4e00\u78b0\u4e0a\u8d85\u957f\u6587\u672c\u5c31\u5361\u987f\u3001OOM\u3001\u4e22\u4e0a\u4e0b\u6587\u2014\u2014\u8981\u4e48\u5207\u5206\u540e\u4fe1\u606f\u65ad\u88c2，\u8981\u4e48\u7b49\u534a\u5929\u624d\u5410<\/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":[7354,7732,7731,956],"topic":[],"class_list":["post-73557","post","type-post","status-publish","format-standard","hentry","category-server","tag-glm-4","tag-triton","tag-7731","tag-956"],"yoast_head":"\n