{"id":84370,"date":"2026-07-26T12:25:29","date_gmt":"2026-07-26T04:25:29","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/84370.html"},"modified":"2026-07-26T12:25:29","modified_gmt":"2026-07-26T04:25:29","slug":"%e5%a4%9a%e5%8f%b0%e6%9c%8d%e5%8a%a1%e5%99%a8%e5%90%8c%e6%97%b6%e8%ae%ad%e7%bb%83llamfactory-%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b-%e5%9b%bd%e5%ae%b6%e8%b6%85%e7%ae%97%e4%b8%ad%e5%bf%83-slurm","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/84370.html","title":{"rendered":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf"},"content":{"rendered":"

\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97（HPC）\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf<\/h3>\n

\u5b8c\u5168\u53ef\u4ee5\u300210 \u5f20 64GB Blackwell \u8de8 5 \u8282\u70b9\u5206\u5e03\u5f0f\u8bad\u7ec3 14B+LoRA \u4e0d\u4ec5\u53ef\u884c，\u800c\u4e14\u5355\u5361\u663e\u5b58\u975e\u5e38\u5bbd\u88d5\u3002<\/p>\n

\n

1. \u663e\u5b58\u6838\u7b97（\u6253\u6d88\u987e\u8651）<\/h3>\n\n\u9879\u76ee\u5360\u7528\u4f30\u7b97\u8bf4\u660e<\/tr>\n\n\n\n\n\n\n\n
14B \u57fa\u5ea7 (BF16)<\/td>\n~28 GB<\/td>\n\u53c2\u6570\u672c\u8eab<\/td>\n<\/tr>\n
LoRA \u53ef\u8bad\u7ec3\u53c2\u6570 (r=128)<\/td>\n~0.2 GB<\/td>\n\u4f4e\u79e9\u77e9\u9635<\/td>\n<\/tr>\n
\u4f18\u5316\u5668\u72b6\u6001 (AdamW)<\/td>\n~0.6 GB<\/td>\n\u53ea\u5b58 LoRA \u7684 momentum<\/td>\n<\/tr>\n
\u6fc0\u6d3b\u503c (seq=768, bs=4)<\/td>\n~4-8 GB<\/td>\n\u4e0e\u957f\u5ea6\u6b63\u76f8\u5173<\/td>\n<\/tr>\n
KV Cache \/ \u788e\u7247<\/td>\n~2 GB<\/td>\n\u9884\u7559<\/td>\n<\/tr>\n
\u5355\u5361\u603b\u8ba1<\/td>\n< 40 GB<\/td>\n64GB \u5361\u5269\u4f59 24GB+<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u7ed3\u8bba：\u5355\u5361\u5c31\u80fd\u72ec\u7acb\u653e\u4e0b\u5b8c\u6574\u6a21\u578b+LoRA，\u4e0d\u9700\u8981 ZeRO-3 \u5207\u5206\u53c2\u6570\u3002\u591a\u8282\u70b9\u8bad\u7ec3\u672c\u8d28\u662f\u7eaf\u6570\u636e\u5e76\u884c（DDP），\u901a\u4fe1\u91cf\u6781\u5c0f（\u53ea\u540c\u6b65 LoRA \u68af\u5ea6\u7684\u51e0\u767e MB）\u3002<\/p>\n

\n

2. \u7f51\u7edc\u74f6\u9888\u8bc4\u4f30（\u5173\u952e\u524d\u63d0）<\/h3>\n

5 \u53f0\u673a\u5668\u8bad\u7ec3，\u8282\u70b9\u95f4\u7f51\u7edc\u662f\u552f\u4e00\u53ef\u80fd\u5361\u4f60\u7684\u5730\u65b9：<\/p>\n\n\u7f51\u7edc\u7c7b\u578b\u662f\u5426\u53ef\u7528\u5efa\u8bae<\/tr>\n\n\n\n\n\n
InfiniBand (200Gbps+)<\/td>\n\u2705 \u6700\u4f18<\/td>\n\u76f4\u63a5\u8dd1，\u65e0\u74f6\u9888<\/td>\n<\/tr>\n
RDMA RoCE (25Gbps+)<\/td>\n\u2705 \u53ef\u7528<\/td>\n\u8bbe\u7f6e NCCL \u73af\u5883\u53d8\u91cf\u5373\u53ef<\/td>\n<\/tr>\n
\u666e\u901a\u4ee5\u592a\u7f51 (1Gbps)<\/td>\n\u26a0\ufe0f \u80fd\u7528\u4f46\u6162<\/td>\n\u589e\u5927 gradient accumulation，\u51cf\u5c11\u901a\u4fe1\u9891\u6b21<\/td>\n<\/tr>\n
\u8de8\u516c\u7f51\/\u5f02\u5730<\/td>\n\u274c \u4e0d\u5efa\u8bae<\/td>\n\u5ef6\u8fdf\u592a\u9ad8，NCCL \u4f1a\u8d85\u65f6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u68c0\u6d4b\u547d\u4ee4（\u6bcf\u53f0\u6267\u884c）：<\/p>\n

# \u770b\u7f51\u5361\u901f\u5ea6<\/span>
\nethtool<\/span> eth0 |<\/span> grep<\/span> Speed
\n# \u5982\u679c\u662f 25000Mb\/s \u4ee5\u4e0a，\u653e\u5fc3\u8dd1；\u5982\u679c\u662f 1000Mb\/s，\u770b\u4e0b\u65b9\u964d\u7ea7\u65b9\u6848<\/span><\/p>\n

\n

3. \u65b9\u6848 A：\u591a\u8282\u70b9\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e00\u4e2a\u4efb\u52a1（\u63a8\u8350）<\/h3>\n

3.1 \u8bad\u7ec3\u4ee3\u7801\u6700\u5c0f\u6539\u52a8<\/h4>\n

\u4f60\u53ea\u9700\u8981\u5728\u539f\u6709 train.py \u4e0a\u52a0\u51e0\u884c\u5206\u5e03\u5f0f\u521d\u59cb\u5316：<\/p>\n

import<\/span> os
\nimport<\/span> torch
\nimport<\/span> torch.<\/span>distributed as<\/span> dist
\nfrom<\/span> torch.<\/span>nn.<\/span>parallel import<\/span> DistributedDataParallel as<\/span> DDP
\nfrom<\/span> torch.<\/span>utils.<\/span>data import<\/span> DataLoader
\nfrom<\/span> torch.<\/span>utils.<\/span>data.<\/span>distributed import<\/span> DistributedSampler
\nfrom<\/span> transformers import<\/span> AutoModelForCausalLM,<\/span> AutoTokenizer,<\/span> TrainingArguments,<\/span> Trainer
\nfrom<\/span> peft import<\/span> get_peft_model,<\/span> LoraConfig
\nfrom<\/span> datasets import<\/span> load_dataset<\/p>\n

def<\/span> main<\/span>(<\/span>)<\/span>:<\/span>
\n # \u5206\u5e03\u5f0f\u73af\u5883\u53d8\u91cf\u7531 torchrun \u81ea\u52a8\u6ce8\u5165<\/span>
\n local_rank =<\/span> int<\/span>(<\/span>os.<\/span>environ[<\/span>"LOCAL_RANK"<\/span>]<\/span>)<\/span>
\n torch.<\/span>cuda.<\/span>set_device(<\/span>local_rank)<\/span><\/p>\n

# \u521d\u59cb\u5316\u8fdb\u7a0b\u7ec4（NCCL \u662f NVIDIA \u591a\u5361\u901a\u4fe1\u540e\u7aef）<\/span>
\n dist.<\/span>init_process_group(<\/span>backend=<\/span>"nccl"<\/span>)<\/span>
\n global_rank =<\/span> dist.<\/span>get_rank(<\/span>)<\/span>
\n world_size =<\/span> dist.<\/span>get_world_size(<\/span>)<\/span> # \u5e94\u8be5\u662f 10<\/span><\/p>\n

# \u53ea\u5728\u4e3b\u8282\u70b9\u6253\u5370，\u907f\u514d 10 \u4efd\u65e5\u5fd7\u5237\u5c4f<\/span>
\n if<\/span> global_rank ==<\/span> 0<\/span>:<\/span>
\n print<\/span>(<\/span>f"Total GPUs: <\/span>{<\/span>world_size}<\/span><\/span>, Local rank: <\/span>{<\/span>local_rank}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

# \u52a0\u8f7d\u6a21\u578b（\u8ddf\u4e4b\u524d\u4e00\u6837）<\/span>
\n model =<\/span> AutoModelForCausalLM.<\/span>from_pretrained(<\/span>
\n "deepseek-ai\/DeepSeek-R1-Distill-Qwen-14B"<\/span>,<\/span>
\n torch_dtype=<\/span>torch.<\/span>bfloat16,<\/span>
\n device_map=<\/span>{<\/span>""<\/span>:<\/span> local_rank}<\/span>,<\/span> # \u5173\u952e：\u6307\u5b9a\u5f53\u524d\u5361<\/span>
\n trust_remote_code=<\/span>True<\/span>
\n )<\/span><\/p>\n

# LoRA<\/span>
\n lora_config =<\/span> LoraConfig(<\/span>
\n r=<\/span>128<\/span>,<\/span> lora_alpha=<\/span>256<\/span>,<\/span>
\n target_modules=<\/span>[<\/span>"q_proj"<\/span>,<\/span> "k_proj"<\/span>,<\/span> "v_proj"<\/span>,<\/span> "o_proj"<\/span>,<\/span>
\n "gate_proj"<\/span>,<\/span> "up_proj"<\/span>,<\/span> "down_proj"<\/span>]<\/span>,<\/span>
\n lora_dropout=<\/span>0.05<\/span>,<\/span> bias=<\/span>"none"<\/span>,<\/span>
\n task_type=<\/span>"CAUSAL_LM"<\/span>
\n )<\/span>
\n model =<\/span> get_peft_model(<\/span>model,<\/span> lora_config)<\/span><\/p>\n

# \u5173\u952e：\u5305\u88c5\u4e3a DDP\u3002\u6ce8\u610f device_ids \u5fc5\u987b\u8bbe，\u5426\u5219\u9ed8\u8ba4\u7528\u6240\u6709\u53ef\u89c1\u5361<\/span>
\n model =<\/span> DDP(<\/span>model,<\/span> device_ids=<\/span>[<\/span>local_rank]<\/span>,<\/span> output_device=<\/span>local_rank)<\/span><\/p>\n

# \u52a0\u8f7d\u6570\u636e<\/span>
\n dataset =<\/span> load_dataset(<\/span>"json"<\/span>,<\/span> data_files=<\/span>"week1_mixed.jsonl"<\/span>,<\/span> split=<\/span>"train"<\/span>)<\/span><\/p>\n

# \u5173\u952e：DistributedSampler \u4fdd\u8bc1\u6bcf\u53f0\u673a\u5668\u53ea\u5904\u7406\u81ea\u5df1\u90a3\u6279\u6570\u636e<\/span>
\n sampler =<\/span> DistributedSampler(<\/span>dataset,<\/span> num_replicas=<\/span>world_size,<\/span> rank=<\/span>global_rank,<\/span> shuffle=<\/span>True<\/span>)<\/span><\/p>\n

# \u6ce8\u610f：\u8fd9\u91cc\u7684 batch_size \u662f\u5355\u5361 batch_size！<\/span>
\n # \u603b global batch = 4 * 10 * grad_accum<\/span>
\n dataloader =<\/span> DataLoader(<\/span>
\n dataset,<\/span>
\n sampler=<\/span>sampler,<\/span>
\n batch_size=<\/span>4<\/span>,<\/span> # \u5355\u5361 batch<\/span>
\n num_workers=<\/span>4<\/span>,<\/span>
\n pin_memory=<\/span>True<\/span>
\n )<\/span><\/p>\n

# \u4f18\u5316\u5668<\/span>
\n optimizer =<\/span> torch.<\/span>optim.<\/span>AdamW(<\/span>model.<\/span>parameters(<\/span>)<\/span>,<\/span> lr=<\/span>5e-6<\/span>,<\/span> weight_decay=<\/span>0.1<\/span>)<\/span><\/p>\n

# \u8bad\u7ec3\u5faa\u73af<\/span>
\n model.<\/span>train(<\/span>)<\/span>
\n for<\/span> epoch in<\/span> range<\/span>(<\/span>3<\/span>)<\/span>:<\/span>
\n sampler.<\/span>set_epoch(<\/span>epoch)<\/span> # \u5173\u952e：\u4fdd\u8bc1\u591a\u8282\u70b9 shuffle \u968f\u673a\u6027<\/span>
\n for<\/span> step,<\/span> batch in<\/span> enumerate<\/span>(<\/span>dataloader)<\/span>:<\/span>
\n # \u6570\u636e\u8fc1\u79fb\u5230\u5f53\u524d GPU<\/span>
\n input_ids =<\/span> batch[<\/span>"input_ids"<\/span>]<\/span>.<\/span>to(<\/span>local_rank)<\/span>
\n attention_mask =<\/span> batch[<\/span>"attention_mask"<\/span>]<\/span>.<\/span>to(<\/span>local_rank)<\/span>
\n labels =<\/span> batch[<\/span>"labels"<\/span>]<\/span>.<\/span>to(<\/span>local_rank)<\/span><\/p>\n

outputs =<\/span> model(<\/span>input_ids=<\/span>input_ids,<\/span> attention_mask=<\/span>attention_mask,<\/span> labels=<\/span>labels)<\/span>
\n loss =<\/span> outputs.<\/span>loss<\/p>\n

# \u68af\u5ea6\u7d2f\u79ef（\u53ef\u9009，\u51cf\u5c11\u901a\u4fe1\u9891\u6b21）<\/span>
\n loss =<\/span> loss \/<\/span> 4<\/span> # \u5047\u8bbe grad_accum=4<\/span>
\n loss.<\/span>backward(<\/span>)<\/span><\/p>\n

if<\/span> (<\/span>step +<\/span> 1<\/span>)<\/span> %<\/span> 4<\/span> ==<\/span> 0<\/span>:<\/span>
\n optimizer.<\/span>step(<\/span>)<\/span>
\n optimizer.<\/span>zero_grad(<\/span>)<\/span><\/p>\n

if<\/span> global_rank ==<\/span> 0<\/span> and<\/span> step %<\/span> 50<\/span> ==<\/span> 0<\/span>:<\/span>
\n print<\/span>(<\/span>f"Epoch <\/span>{<\/span>epoch}<\/span><\/span>, Step <\/span>{<\/span>step}<\/span><\/span>, Loss: <\/span>{<\/span>loss.<\/span>item(<\/span>)<\/span>:<\/span>.4f<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

# \u53ea\u5728\u4e3b\u8282\u70b9\u4fdd\u5b58<\/span>
\n if<\/span> global_rank ==<\/span> 0<\/span>:<\/span>
\n # \u4fdd\u5b58\u65f6\u8981\u7528 model.module（\u56e0\u4e3a DDP \u5305\u88c5\u4e86\u4e00\u5c42）<\/span>
\n model.<\/span>module.<\/span>save_pretrained(<\/span>f"checkpoint_epoch_<\/span>{<\/span>epoch}<\/span><\/span>"<\/span><\/span>)<\/span>
\n print<\/span>(<\/span>f"Saved checkpoint at epoch <\/span>{<\/span>epoch}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

dist.<\/span>destroy_process_group(<\/span>)<\/span><\/p>\n

if<\/span> __name__ ==<\/span> "__main__"<\/span>:<\/span>
\n main(<\/span>)<\/span><\/p>\n

3.2 \u4e00\u952e\u542f\u52a8\u811a\u672c<\/h4>\n

\u8282\u70b9 1（\u4e3b\u8282\u70b9，\u5047\u8bbe IP \u4e3a 192.168.1.10）：<\/p>\n

# node1.sh<\/span>
\nexport<\/span> NCCL_DEBUG<\/span>=<\/span>INFO # \u8c03\u8bd5\u7528，\u8dd1\u901a\u540e\u53ef\u5220<\/span>
\nexport<\/span> NCCL_IB_DISABLE<\/span>=<\/span>0<\/span> # \u5982\u679c\u6709 IB，\u4fdd\u6301 0；\u5982\u679c\u7eaf\u4ee5\u592a\u7f51，\u8bbe 1<\/span>
\nexport<\/span> NCCL_SOCKET_IFNAME<\/span>=<\/span>eth0 # \u6539\u6210\u4f60\u7684\u5b9e\u9645\u7f51\u5361\u540d（ifconfig \u770b）<\/span><\/p>\n

torchrun \\\\<\/span>
\n –nnodes<\/span>=<\/span>5<\/span> \\\\<\/span>
\n –nproc_per_node<\/span>=<\/span>2<\/span> \\\\<\/span>
\n –rdzv_id<\/span>=<\/span>druggpt_job \\\\<\/span>
\n –rdzv_backend<\/span>=<\/span>c10d \\\\<\/span>
\n –rdzv_endpoint<\/span>=<\/span>192.168<\/span>.1.10:29500 \\\\<\/span>
\n train.py<\/p>\n

\u8282\u70b9 2~5（\u5de5\u4f5c\u8282\u70b9）：<\/p>\n

# node2.sh（\u5176\u4ed6\u8282\u70b9\u552f\u4e00\u533a\u522b\u662f rdzv_endpoint \u6307\u5411\u4e3b\u8282\u70b9）<\/span>
\nexport<\/span> NCCL_DEBUG<\/span>=<\/span>INFO
\nexport<\/span> NCCL_IB_DISABLE<\/span>=<\/span>0<\/span>
\nexport<\/span> NCCL_SOCKET_IFNAME<\/span>=<\/span>eth0<\/p>\n

torchrun \\\\<\/span>
\n –nnodes<\/span>=<\/span>5<\/span> \\\\<\/span>
\n –nproc_per_node<\/span>=<\/span>2<\/span> \\\\<\/span>
\n –rdzv_id<\/span>=<\/span>druggpt_job \\\\<\/span>
\n –rdzv_backend<\/span>=<\/span>c10d \\\\<\/span>
\n –rdzv_endpoint<\/span>=<\/span>192.168<\/span>.1.10:29500 \\\\<\/span>
\n train.py<\/p>\n

\u6267\u884c\u65b9\u5f0f：<\/p>\n

# \u5728 5 \u53f0\u673a\u5668\u4e0a\u540c\u65f6\u6267\u884c\u5404\u81ea\u811a\u672c（\u7528 tmux\/screen\/ssh \u5e76\u884c）<\/span>
\n# \u6216\u8005\u5982\u679c\u4f60\u7528 Kubernetes\/Slurm，\u7531\u8c03\u5ea6\u7cfb\u7edf\u7edf\u4e00\u542f\u52a8<\/span><\/p>\n

\n

4. \u65b9\u6848 B：\u7528 Accelerate \u5e93（\u66f4\u7701\u4e8b）<\/h3>\n

HuggingFace Accelerate \u80fd\u81ea\u52a8\u751f\u6210\u591a\u8282\u70b9\u547d\u4ee4，\u4e0d\u7528\u624b\u5199 DDP：<\/p>\n

# 1. \u5728\u6bcf\u53f0\u673a\u5668\u4e0a\u751f\u6210\u914d\u7f6e<\/span>
\naccelerate config –config_file<\/span> ds_config.yaml
\n# \u4ea4\u4e92\u5f0f\u9009\u62e9：multi-node, 2 GPUs per node, DeepSpeed\/DDP<\/span><\/p>\n

# 2. \u4e3b\u8282\u70b9\u542f\u52a8（Accelerate \u4f1a\u81ea\u52a8\u901a\u8fc7 rdzv \u53d1\u73b0\u5176\u4ed6\u8282\u70b9）<\/span>
\naccelerate launch \\\\<\/span>
\n –config_file<\/span> ds_config.yaml \\\\<\/span>
\n –main_process_ip<\/span> 192.168<\/span>.1.10 \\\\<\/span>
\n –main_process_port<\/span> 29500<\/span> \\\\<\/span>
\n –machine_rank<\/span> 0<\/span> \\\\<\/span>
\n –num_machines<\/span> 5<\/span> \\\\<\/span>
\n –num_processes<\/span> 10<\/span> \\\\<\/span>
\n train.py<\/p>\n

train.py \u51e0\u4e4e\u4e0d\u7528\u6539，\u628a model = DDP(…) \u5220\u6389，Accelerate \u4f1a\u81ea\u52a8\u5904\u7406\u3002<\/p>\n

\n

5. \u65b9\u6848 C：\u5982\u679c\u7f51\u7edc\u5f88\u5dee（\u964d\u7ea7）<\/h3>\n

\u5982\u679c\u8282\u70b9\u95f4\u53ea\u6709\u5343\u5146\u7f51（1Gbps），All-Reduce \u4f1a\u6210\u4e3a\u74f6\u9888\u3002\u6b64\u65f6\u4e0d\u8981\u5206\u5e03\u5f0f\u8bad\u4e00\u4e2a\u4efb\u52a1，\u800c\u662f：<\/p>\n\n\u673a\u5668\u4efb\u52a1\u6570\u636e<\/tr>\n\n\n\n\n
\u8282\u70b9 1-2<\/td>\n\u8bad Week 1（\u5168\u957f70%+Domain30%）<\/td>\n\u6570\u636e\u5206\u7247 A<\/td>\n<\/tr>\n
\u8282\u70b9 3-4<\/td>\n\u8bad Week 2（\u5168\u957f30%+Domain70%）<\/td>\n\u6570\u636e\u5206\u7247 B<\/td>\n<\/tr>\n
\u8282\u70b9 5<\/td>\n\u8bad Week 3（Domain50%+Pocket50%）<\/td>\n\u6570\u636e\u5206\u7247 C<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u8fd9\u4e0d\u662f\u5206\u5e03\u5f0f，\u662f\u5e76\u884c\u8dd1\u5b9e\u9a8c\u3002\u6bcf\u53f0\u673a\u5668\u5185\u90e8 2 \u5f20\u5361\u7528 torchrun –nproc_per_node=2 \u505a\u672c\u5730 DDP\u3002<\/p>\n

\u7ed3\u679c：5 \u4e2a\u5b9e\u9a8c\u540c\u65f6\u8dd1，3 \u5468\u540e\u9009\u6700\u4f18\u6a21\u578b\u3002\u9002\u5408\u7f51\u7edc\u5dee\u3001\u60f3\u5feb\u901f\u8bd5\u9519\u7684\u60c5\u51b5\u3002<\/p>\n

\n

6. \u5173\u952e\u68c0\u67e5\u6e05\u5355（\u8dd1\u524d\u5fc5\u505a）<\/h3>\n

# 1. \u786e\u8ba4 5 \u53f0\u673a\u5668 NCCL \u80fd\u4e92\u76f8\u53d1\u73b0（\u5728\u4e3b\u8282\u70b9\u6267\u884c）<\/span>
\npython -c<\/span> "import torch; print(torch.cuda.device_count())"<\/span> # \u5e94\u8f93\u51fa 2<\/span><\/p>\n

# 2. \u786e\u8ba4\u7aef\u53e3\u4e92\u901a<\/span>
\n# \u5728\u8282\u70b9 2 \u4e0a：telnet 192.168.1.10 29500<\/span>
\n# \u5fc5\u987b\u80fd\u901a，\u5426\u5219 torchrun \u4f1a\u5361\u6b7b<\/span><\/p>\n

# 3. \u786e\u8ba4\u6240\u6709\u8282\u70b9 PyTorch\/CUDA\/NCCL \u7248\u672c\u4e00\u81f4<\/span>
\npython -c<\/span> "import torch; print(torch.__version__, torch.version.cuda)"<\/span><\/p>\n

# 4. \u6570\u636e\u53ef\u89c1\u6027<\/span>
\n# \u6240\u6709\u8282\u70b9\u5fc5\u987b\u80fd\u8bfb\u53d6 week1_mixed.jsonl（NFS \u6302\u8f7d\u6216\u6bcf\u53f0\u672c\u5730\u590d\u5236\u4e00\u4efd）<\/span><\/p>\n

# 5. \u9632\u706b\u5899<\/span>
\nsudo<\/span> ufw allow 29500<\/span>\/tcp # Ubuntu \u793a\u4f8b<\/span><\/p>\n

\n

7. \u663e\u5b58\u4e0e\u901f\u5ea6\u9884\u671f<\/h3>\n\n\u914d\u7f6e\u5355\u5361\u663e\u5b58\u5360\u7528\u5168\u5c40 Batch Size\u9884\u4f30\u901f\u5ea6<\/tr>\n\n\n\n\n
\u5355\u5361\u72ec\u7acb (1×2)<\/td>\n~35 GB<\/td>\n4<\/td>\n1x<\/td>\n<\/tr>\n
\u5355\u673a\u53cc\u5361 (1×2 DDP)<\/td>\n~38 GB<\/td>\n8<\/td>\n~1.9x<\/td>\n<\/tr>\n
5\u673a10\u5361 (5×2 DDP)<\/td>\n~38 GB<\/td>\n40<\/td>\n~9x<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

10 \u5f20\u5361\u7406\u8bba\u4e0a\u80fd\u628a\u4f60\u7684\u8bad\u7ec3\u65f6\u95f4\u538b\u7f29\u5230\u539f\u6765\u7684 1\/9\u300230 \u4e07\u6b65\u5982\u679c\u539f\u6765\u8981 7 \u5929，\u73b0\u5728\u7ea6 18 \u5c0f\u65f6\u4e00\u4e2a epoch\u3002<\/p>\n

\n

8. \u603b\u7ed3<\/h3>\n\n\u4f60\u7684\u60c5\u51b5\u63a8\u8350\u65b9\u6848<\/tr>\n\n\n\n\n
\u8282\u70b9\u95f4\u6709 25Gbps+ \/ IB<\/td>\n\u65b9\u6848 A：5\u673a10\u5361\u5206\u5e03\u5f0f\u8bad\u4e00\u4e2a\u4efb\u52a1<\/td>\n<\/tr>\n
\u8282\u70b9\u95f4\u53ea\u6709\u5343\u5146\u7f51<\/td>\n\u65b9\u6848 C：\u6bcf\u53f0\u72ec\u7acb\u8dd1\u4e0d\u540c Week \u5b9e\u9a8c<\/td>\n<\/tr>\n
\u4e0d\u60f3\u6539\u4ee3\u7801<\/td>\n\u65b9\u6848 B：Accelerate \u81ea\u52a8\u5206\u5e03\u5f0f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u73b0\u5728\u6700\u8be5\u505a\u7684：<\/p>\n

  • \u9009\u4e00\u53f0\u673a\u5668，\u5148\u5355\u673a torchrun –nproc_per_node=2 train.py \u8dd1\u901a<\/li>\n
  • \u786e\u8ba4\u80fd\u8dd1\u540e，5 \u53f0\u540c\u65f6\u542f\u52a8，\u52a0 –nnodes=5<\/li>\n
  • \u89c2\u5bdf NCCL \u65e5\u5fd7，\u786e\u4fdd global rank 0-9 \u5168\u90e8\u4e0a\u7ebf<\/li>\n

    \u9700\u8981\u6211\u63d0\u4f9b DeepSpeed ZeRO-1 \u7684\u591a\u8282\u70b9\u914d\u7f6e\u6587\u4ef6（\u8fdb\u4e00\u6b65\u538b\u69a8\u901a\u4fe1\u6548\u7387），\u6216\u8005 Slurm \u96c6\u7fa4\u7684 sbatch \u63d0\u4ea4\u811a\u672c \u5417？<\/p>\n

    Slurm（\u5168\u79f0 Simple Linux Utility for Resource Management）\u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97（HPC）\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf，\u88ab\u8d85\u7b97 TOP500 \u4e2d\u8d85\u8fc7\u534a\u6570\u7684\u7cfb\u7edf\u91c7\u7528\u3002<\/p>\n

    \u4e00\u3001\u6838\u5fc3\u5b9a\u4f4d\u4e0e\u4f5c\u7528<\/h4>\n

    Slurm \u4f5c\u4e3a\u96c6\u7fa4\u201c\u64cd\u4f5c\u7cfb\u7edf\u201d，\u4e3b\u8981\u89e3\u51b3\u4e09\u5927\u95ee\u9898：<\/p>\n

  • \u8d44\u6e90\u5206\u914d：\u4e3a\u7528\u6237\u5206\u914d\u8ba1\u7b97\u8282\u70b9\u3001CPU\u3001GPU\u3001\u5185\u5b58\u3001\u7f51\u7edc\u7b49\u8d44\u6e90（\u72ec\u5360\/\u5171\u4eab）\u3002<\/li>\n
  • \u4f5c\u4e1a\u6267\u884c：\u63d0\u4f9b\u6846\u67b6\u542f\u52a8\u3001\u8fd0\u884c\u3001\u76d1\u63a7\u5e76\u884c\/\u4e32\u884c\u4f5c\u4e1a（\u5982 MPI\u3001AI \u8bad\u7ec3）\u3002<\/li>\n
  • \u8c03\u5ea6\u4ef2\u88c1：\u7ba1\u7406\u4f5c\u4e1a\u961f\u5217，\u6309\u7b56\u7565\u8c03\u5ea6\u3001\u89e3\u51b3\u8d44\u6e90\u7ade\u4e89\u3002<\/li>\n

    \u4e8c\u3001\u57fa\u672c\u67b6\u6784（\u6838\u5fc3\u7ec4\u4ef6）<\/h4>\n
      \n
    • slurmctld（\u4e2d\u592e\u63a7\u5236\u5668）\n
        \n
      • \u96c6\u7fa4\u5927\u8111：\u76d1\u63a7\u8d44\u6e90\u3001\u7ba1\u7406\u961f\u5217\u3001\u5206\u914d\u8d44\u6e90\u3001\u8c03\u5ea6\u4f5c\u4e1a\u3002<\/li>\n
      • \u652f\u6301\u4e3b\u5907\u5197\u4f59（\u9ad8\u53ef\u7528）\u3002<\/li>\n<\/ul>\n<\/li>\n
      • slurmd（\u8282\u70b9\u5b88\u62a4）\n
          \n
        • \u6bcf\u4e2a\u8ba1\u7b97\u8282\u70b9\u8fd0\u884c\u4e00\u4e2a：\u63a5\u6536\u6307\u4ee4\u3001\u542f\u52a8\u4efb\u52a1\u3001\u76d1\u63a7\u8282\u70b9\u72b6\u6001\u3002<\/li>\n<\/ul>\n<\/li>\n
        • slurmdbd（\u8bb0\u8d26\u6570\u636e\u5e93）\n
            \n
          • \u96c6\u4e2d\u5b58\u50a8\u4f5c\u4e1a\u5386\u53f2\u3001\u8d44\u6e90\u4f7f\u7528\u3001\u7528\u6237\u914d\u989d\u3001\u8d26\u5355\u6570\u636e\u3002<\/li>\n<\/ul>\n<\/li>\n
          • slurmstepd\n
              \n
            • \u4f5c\u4e1a\u6b65\u5b88\u62a4：\u7ba1\u7406\u5355\u4e2a\u4f5c\u4e1a\u5185\u4efb\u52a1\u7684 I\/O\u3001\u4fe1\u53f7\u3001\u8d44\u6e90\u9650\u5236\u3002<\/li>\n<\/ul>\n<\/li>\n
            • \u7528\u6237\u547d\u4ee4\n
                \n
              • sbatch：\u6279\u5904\u7406\u4f5c\u4e1a\u63d0\u4ea4<\/li>\n
              • srun：\u4ea4\u4e92\u5f0f\/\u5b9e\u65f6\u8fd0\u884c<\/li>\n
              • salloc：\u7533\u8bf7\u8d44\u6e90\u5e76\u8fdb\u5165\u4f1a\u8bdd<\/li>\n
              • sinfo：\u67e5\u770b\u96c6\u7fa4\/\u8282\u70b9\u72b6\u6001<\/li>\n
              • squeue：\u67e5\u770b\u6392\u961f\/\u8fd0\u884c\u4f5c\u4e1a<\/li>\n
              • scancel：\u53d6\u6d88\u4f5c\u4e1a<\/li>\n
              • sacct：\u67e5\u8be2\u5386\u53f2\u8bb0\u8d26<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

                \u4e09\u3001\u5173\u952e\u7279\u6027<\/h4>\n
                  \n
                • \u6781\u81f4\u53ef\u6269\u5c55：\u5355\u96c6\u7fa4\u53ef\u7ba1\u7406\u6570\u767e\u4e07 CPU \u6838\u5fc3\u3001\u6570\u4e07\u8282\u70b9\u3002<\/li>\n
                • \u9ad8\u53ef\u7528：\u63a7\u5236\u5668\u4e3b\u5907\u3001\u8282\u70b9\u6545\u969c\u9694\u79bb\u3001\u65e0\u5355\u70b9\u5931\u6548\u3002<\/li>\n
                • \u4e30\u5bcc\u8c03\u5ea6\u7b56\u7565\n
                    \n
                  • FIFO\u3001\u516c\u5e73\u5171\u4eab（Fair-share）\u3001\u56de\u586b\u8c03\u5ea6（Backfill）<\/li>\n
                  • \u4f18\u5148\u7ea7\u3001\u62a2\u5360\u3001\u9884\u7559\u3001GANG \u8c03\u5ea6（\u5206\u65f6）<\/li>\n<\/ul>\n<\/li>\n
                  • \u8d44\u6e90\u7cbe\u7ec6\u5316\u7ba1\u7406\n
                      \n
                    • \u652f\u6301 CPU\/GPU\/\u5185\u5b58\/\u7f51\u5361\/IB \u7b49\u7ed1\u5b9a\u4e0e\u9694\u79bb<\/li>\n
                    • \u5206\u533a（Partition）\u3001QoS\u3001\u7528\u6237\/\u8d26\u6237\u914d\u989d\u3001\u65f6\u9650<\/li>\n<\/ul>\n<\/li>\n
                    • \u4f5c\u4e1a\u9ad8\u7ea7\u7279\u6027\n
                        \n
                      • \u4f5c\u4e1a\u4f9d\u8d56\u3001\u4f5c\u4e1a\u6570\u7ec4\u3001\u6570\u7ec4\u4efb\u52a1\u3001\u4fe1\u53f7\u63a7\u5236<\/li>\n
                      • \u6807\u51c6\u8f93\u51fa\/\u9519\u8bef\u91cd\u5b9a\u5411\u3001\u5b9e\u65f6 attach \u4f5c\u4e1a（sattach）<\/li>\n<\/ul>\n<\/li>\n
                      • \u63d2\u4ef6\u5316：\u8c03\u5ea6\u3001\u8bb0\u8d26\u3001\u8ba4\u8bc1\u3001\u7f51\u7edc\u7b49\u5747\u4e3a\u63d2\u4ef6，\u6613\u6269\u5c55\u3002<\/li>\n
                      • \u5f00\u6e90\u514d\u8d39：GPLv2 \u534f\u8bae，\u793e\u533a\u6d3b\u8dc3\u3002<\/li>\n<\/ul>\n

                        \u56db\u3001\u5178\u578b\u4f7f\u7528\u6d41\u7a0b（\u7528\u6237\u89c6\u89d2）<\/h4>\n
                      • \u7528 sinfo \u770b\u53ef\u7528\u8d44\u6e90\u3002<\/li>\n
                      • \u5199\u811a\u672c（job.sh），\u6307\u5b9a\u8d44\u6e90：#!\/bin\/bash<\/span>
                        \n#SBATCH –job-name=my_train<\/span>
                        \n#SBATCH –nodes=2<\/span>
                        \n#SBATCH –ntasks-per-node=8<\/span>
                        \n#SBATCH –gres=gpu:4<\/span>
                        \n#SBATCH –time=12:00:00<\/span>
                        \npython train.py<\/p>\n<\/li>\n
                      • \u63d0\u4ea4：sbatch job.sh<\/li>\n
                      • \u67e5\u770b：squeue\u3001sacct<\/li>\n
                      • \u53d6\u6d88：scancel <jobid><\/li>\n

                        \u4e94\u3001\u5e94\u7528\u573a\u666f<\/h4>\n
                          \n
                        • \u8d85\u7b97\u4e2d\u5fc3\u3001\u9ad8\u6821\/\u79d1\u7814\u96c6\u7fa4<\/li>\n
                        • AI \u8bad\u7ec3\u96c6\u7fa4（\u5927\u6a21\u578b\u3001\u6df1\u5ea6\u5b66\u4e60）<\/li>\n
                        • \u4e91\u8ba1\u7b97\u3001\u79c1\u6709\u4e91\u8d44\u6e90\u8c03\u5ea6<\/li>\n
                        • \u5927\u89c4\u6a21\u79d1\u5b66\u8ba1\u7b97\u3001\u6570\u503c\u6a21\u62df\u3001\u751f\u7269\u4fe1\u606f\u3001\u6c14\u8c61\u3001CAE \u7b49<\/li>\n<\/ul>\n

                          \u516d\u3001\u4e0e\u540c\u7c7b\u5bf9\u6bd4<\/h4>\n
                            \n
                          • \u6bd4 PBS Pro\u3001LSF \u66f4\u8f7b\u91cf\u3001\u5f00\u6e90\u3001\u751f\u6001\u5f3a\u3002<\/li>\n
                          • \u6bd4 Kubernetes \u66f4\u9002\u5408\u957f\u65f6\u3001\u6279\u5904\u7406\u3001MPI \u5e76\u884c\u3001GPU \u5bc6\u96c6\u7684 HPC \u573a\u666f\u3002<\/li>\n<\/ul>\n

                            \u8981\u4e0d\u8981\u6211\u7ed9\u4f60\u4e00\u4efd\u53ef\u76f4\u63a5\u590d\u5236\u7684 Slurm \u5e38\u7528\u547d\u4ee4\u901f\u67e5\u8868（\u542b\u63d0\u4ea4\u3001\u67e5\u8be2\u3001\u76d1\u63a7\u3001\u53d6\u6d88\u3001\u8d44\u6e90\u67e5\u770b）？<\/p>\n

                            \u8fd9\u4efd\u662f\u8d85\u7cbe\u7b80\u3001\u53ef\u76f4\u63a5\u590d\u5236\u7c98\u8d34\u7684 Slurm \u5e38\u7528\u547d\u4ee4\u901f\u67e5\u8868，\u65e5\u5e38HPC\/AI\u8bad\u7ec3\u5b8c\u5168\u591f\u7528\u3002<\/p>\n

                            \n

                            Slurm \u5e38\u7528\u547d\u4ee4\u901f\u67e5<\/h2>\n

                            \u4e00\u3001\u67e5\u770b\u96c6\u7fa4\u4e0e\u8d44\u6e90<\/h3>\n

                            # \u67e5\u770b\u6240\u6709\u8282\u70b9\/\u5206\u533a\u72b6\u6001<\/span>
                            \nsinfo<\/p>\n

                            # \u53ea\u770b\u7a7a\u95f2\u8282\u70b9<\/span>
                            \nsinfo -N<\/span> -l<\/span> |<\/span> grep<\/span> idle<\/p>\n

                            # \u67e5\u770b GPU \u8d44\u6e90（\u9700\u8981\u96c6\u7fa4\u914d\u7f6e gres）<\/span>
                            \nsinfo -o<\/span> "%N %P %T %G"<\/span><\/p>\n

                            # \u67e5\u770b\u8282\u70b9\u8be6\u7ec6\u4fe1\u606f<\/span>
                            \nscontrol show node<\/span> <<\/span>\u8282\u70b9\u540d><\/span><\/p>\n

                            \n

                            \u4e8c\u3001\u4f5c\u4e1a\u63d0\u4ea4（\u6700\u5e38\u7528）<\/h3>\n

                            1. \u63d0\u4ea4\u6279\u5904\u7406\u811a\u672c<\/h4>\n

                            sbatch job.sh<\/p>\n

                            2. \u5e38\u7528 #SBATCH \u53c2\u6570（\u5199\u5728\u811a\u672c\u5f00\u5934）<\/h4>\n

                            #!\/bin\/bash<\/span>
                            \n#SBATCH –job-name=myjob # \u4f5c\u4e1a\u540d<\/span>
                            \n#SBATCH –partition=your_part # \u5206\u533a<\/span>
                            \n#SBATCH –nodes=1 # \u8282\u70b9\u6570<\/span>
                            \n#SBATCH –ntasks-per-node=8 # \u6bcf\u8282\u70b9\u4efb\u52a1\u6570<\/span>
                            \n#SBATCH –gres=gpu:1 # GPU \u6570\u91cf<\/span>
                            \n#SBATCH –mem=32G # \u603b\u5185\u5b58<\/span>
                            \n#SBATCH –time=12:00:00 # \u6700\u5927\u8fd0\u884c\u65f6\u95f4<\/span>
                            \n#SBATCH –output=%j.out # \u8f93\u51fa\u65e5\u5fd7<\/span>
                            \n#SBATCH –error=%j.err # \u9519\u8bef\u65e5\u5fd7<\/span>
                            \n#SBATCH –mail-type=END # \u7ed3\u675f\u65f6\u53d1\u90ae\u4ef6<\/span><\/p>\n

                            3. \u76f4\u63a5\u4ea4\u4e92\u5f0f\u8fd0\u884c<\/h4>\n

                            srun -p<\/span> \u5206\u533a –gres<\/span>=<\/span>gpu:1 nvidia-smi<\/p>\n

                            4. \u7533\u8bf7\u8d44\u6e90\u8fdb\u5165\u8282\u70b9<\/h4>\n

                            salloc -p<\/span> \u5206\u533a –gres<\/span>=<\/span>gpu:1 –time<\/span>=<\/span>2<\/span>:00:00
                            \n# \u8fdb\u5165\u540e：ssh \u8282\u70b9\u540d<\/span><\/p>\n

                            \n

                            \u4e09\u3001\u67e5\u770b\u4f5c\u4e1a<\/h3>\n

                            # \u67e5\u770b\u81ea\u5df1\u7684\u4f5c\u4e1a<\/span>
                            \nsqueue<\/p>\n

                            # \u67e5\u770b\u6240\u6709\u4eba\u7684\u4f5c\u4e1a<\/span>
                            \nsqueue -a<\/span><\/p>\n

                            # \u53ea\u770b\u81ea\u5df1<\/span>
                            \nsqueue -u<\/span> $USER<\/span><\/p>\n

                            # \u67e5\u770b\u4f5c\u4e1a\u8be6\u7ec6\u4fe1\u606f<\/span>
                            \nscontrol show job <<\/span>jobid><\/span><\/p>\n

                            # \u67e5\u770b\u5386\u53f2\u4f5c\u4e1a（\u7ed3\u675f\u7684\u4f5c\u4e1a）<\/span>
                            \nsacct
                            \nsacct -j<\/span> <<\/span>jobid><\/span>
                            \nsacct -u<\/span> $USER<\/span> –format<\/span>=<\/span>jobid,jobname,partition,state,elapsed<\/p>\n

                            \n

                            \u56db\u3001\u63a7\u5236\u4f5c\u4e1a<\/h3>\n

                            # \u53d6\u6d88\u4e00\u4e2a\u4f5c\u4e1a<\/span>
                            \nscancel <<\/span>jobid><\/span><\/p>\n

                            # \u53d6\u6d88\u81ea\u5df1\u6240\u6709\u4f5c\u4e1a<\/span>
                            \nscancel -u<\/span> $USER<\/span><\/p>\n

                            # \u6682\u505c\/\u6062\u590d\u4f5c\u4e1a<\/span>
                            \nscontrol hold <<\/span>jobid><\/span>
                            \nscontrol release <<\/span>jobid><\/span><\/p>\n

                            # \u4fee\u6539\u4f5c\u4e1a\u6700\u5927\u65f6\u9650<\/span>
                            \nscontrol update jobid<\/span>=<\/span><<\/span>jobid><\/span> timelimit<\/span>=<\/span>24<\/span>:00:00<\/p>\n

                            \n

                            \u4e94\u3001\u67e5\u770b\u8d26\u6237\u4e0e\u914d\u989d<\/h3>\n

                            # \u67e5\u770b\u8d26\u6237\u4fe1\u606f<\/span>
                            \nsacctmgr show assoc user<\/span>=<\/span>$USER<\/span><\/p>\n

                            # \u67e5\u770b QoS \u9650\u5236<\/span>
                            \nsacctmgr show qos<\/p>\n

                            \n

                            \u516d\u3001\u5e38\u7528\u72b6\u6001\u8bf4\u660e<\/h3>\n
                              \n
                            • PD：Pending \u6392\u961f<\/li>\n
                            • R：Running \u8fd0\u884c\u4e2d<\/li>\n
                            • CG：Completing \u6b63\u5728\u7ed3\u675f<\/li>\n
                            • F：Failed \u5931\u8d25<\/li>\n
                            • TO：TimeOut \u8d85\u65f6<\/li>\n
                            • CA：Canceled \u88ab\u53d6\u6d88<\/li>\n<\/ul>\n
                              \n

                              \u5982\u679c\u4f60\u544a\u8bc9\u6211\u4f60\u7528\u7684\u662fCPU \u8fd8\u662f GPU \u8bad\u7ec3，\u6211\u53ef\u4ee5\u76f4\u63a5\u7ed9\u4f60\u4e00\u4e2a\u53ef\u76f4\u63a5\u7528\u7684\u901a\u7528 sbatch \u811a\u672c\u6a21\u677f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                              \u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97（HPC）\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf
                              \n\u5b8c\u5168\u53ef\u4ee5\u300210 \u5f20 64GB Blackwell \u8de8 5 \u8282\u70b9\u5206\u5e03\u5f0f\u8bad\u7ec3 14BLoRA \u4e0d\u4ec5\u53ef\u884c，\u800c\u4e14\u5355\u5361\u663e\u5b58\u975e\u5e38\u5bbd\u88d5\u30021. \u663e\u5b58\u6838\u7b97（\u6253\u6d88\u987e\u8651）\u9879\u76ee\u5360\u7528\u4f30\u7b97\u8bf4\u660e14B \u57fa\u5ea7 (BF16)~28 GB\u53c2\u6570\u672c\u8eabLoRA \u53ef\u8bad\u7ec3\u53c2\u6570 (r128)~0.2 GB<\/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":[227,43,51],"topic":[],"class_list":["post-84370","post","type-post","status-publish","format-standard","hentry","category-server","tag-227","tag-43","tag-51"],"yoast_head":"\n\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/84370.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97（HPC）\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf \u5b8c\u5168\u53ef\u4ee5\u300210 \u5f20 64GB Blackwell \u8de8 5 \u8282\u70b9\u5206\u5e03\u5f0f\u8bad\u7ec3 14BLoRA \u4e0d\u4ec5\u53ef\u884c，\u800c\u4e14\u5355\u5361\u663e\u5b58\u975e\u5e38\u5bbd\u88d5\u30021. \u663e\u5b58\u6838\u7b97（\u6253\u6d88\u987e\u8651）\u9879\u76ee\u5360\u7528\u4f30\u7b97\u8bf4\u660e14B \u57fa\u5ea7 (BF16)~28 GB\u53c2\u6570\u672c\u8eabLoRA \u53ef\u8bad\u7ec3\u53c2\u6570 (r128)~0.2 GB\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/84370.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-26T04:25:29+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/84370.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/84370.html\",\"name\":\"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-26T04:25:29+00:00\",\"dateModified\":\"2026-07-26T04:25:29+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/84370.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/84370.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/84370.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/84370.html","og_locale":"zh_CN","og_type":"article","og_title":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97（HPC）\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf \u5b8c\u5168\u53ef\u4ee5\u300210 \u5f20 64GB Blackwell \u8de8 5 \u8282\u70b9\u5206\u5e03\u5f0f\u8bad\u7ec3 14BLoRA \u4e0d\u4ec5\u53ef\u884c，\u800c\u4e14\u5355\u5361\u663e\u5b58\u975e\u5e38\u5bbd\u88d5\u30021. \u663e\u5b58\u6838\u7b97（\u6253\u6d88\u987e\u8651）\u9879\u76ee\u5360\u7528\u4f30\u7b97\u8bf4\u660e14B \u57fa\u5ea7 (BF16)~28 GB\u53c2\u6570\u672c\u8eabLoRA \u53ef\u8bad\u7ec3\u53c2\u6570 (r128)~0.2 GB","og_url":"https:\/\/www.wsisp.com\/helps\/84370.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-26T04:25:29+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/84370.html","url":"https:\/\/www.wsisp.com\/helps\/84370.html","name":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-26T04:25:29+00:00","dateModified":"2026-07-26T04:25:29+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/84370.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/84370.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/84370.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u591a\u53f0\u670d\u52a1\u5668\u540c\u65f6\u8bad\u7ec3llamfactory \u5927\u8bed\u8a00\u6a21\u578b \u56fd\u5bb6\u8d85\u7b97\u4e2d\u5fc3 Slurm \u662f\u76ee\u524d\u5168\u7403\u6700\u4e3b\u6d41\u7684\u5f00\u6e90\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u96c6\u7fa4\u8d44\u6e90\u7ba1\u7406\u4e0e\u4f5c\u4e1a\u8c03\u5ea6\u7cfb\u7edf"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/84370","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=84370"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/84370\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=84370"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=84370"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=84370"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=84370"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}