# 各模型支持的最小卡数和序列长度 #### 模型训练时间、集群规模预测 针对不同模型、集群规格（Snt9b B3/B2/B1、Snt9b23）、数据集大小等不同导致训练时间、集群卡数规模不同，如训练过程中对卡数或训练时间有要求可根据以下公式计算预测： - 训练时间(秒)：Time=Tok~总~/(TPS\*N~卡数~)，计算出训练时间为范围值，仅参考。 - 训练卡数：N~卡数~=Tok~总~/(Time\*TPS)，N~卡数~\>8时则需取整为8的倍数且不小于[模型最小卡数配置]。 参数解释： 1. 总Tokens (Tok~总~)：这通常需要对数据集进行预处理，包括分词（tokenization），与数据集大小、数据集遍历次数（Epochs）、序列长度、模型有关系，如填充（Padding）导致无效Token增加等多因素。 - **总tokens数（通过训练步数）** =Seql×数据集样本总数，Seql分为动态、固定两种，一般动态Seql总tokens总数小于以上计算值。设置参数详见[表2 Llama-Factory参数](https://support.huaweicloud.com/bestpractice-modelarts/modelarts_llm_train_590620.html#ZH-CN_TOPIC_0000002430726609__table19503184993916)-packing参数或，一般MindSpeed-LLM与Llama-Factory默认为固定Seq。 2. TPS：各个模型吞吐值(token/s/p)、训练卡数可在benchmark表格中查找，benchmark中的吞吐值均为固定Seq测出基线值，benchmark表格可向华为工程师咨询。 #### 模型最小卡数配置 不同模型推荐的训练参数和计算规格要求如下表所示，目前仅提供微调（SFT）及训练（PT）阶段卡数配置。一般Snt9b规格为单节点8卡，Snt9b23规格为单机8卡=16\*DIE，其中1\*DIE等效于Snt9b中的1卡，Snt9b23规格实际训练过程中设置并行策略时2\*DIE为最小单位；以下配置仅参考，一般小于8卡使用8卡训练，用户可基于卡数配置浮动调动。 **表格中"-"代表不支持** **，规格与卡数中的 4\*Ascend在Snt9b中表示4卡，Snt9b23表示4\*DIE，以此类推。** 表1模型最小卡数配置 | 支持模型参数量 | 训练策略类型 | 序列长度SEQ_LEN | MindSpeed-LLM规格卡数/DIE || Llama-Factory规格卡数/DIE || VeRL规格卡数/DIE || MindSpeed-RL规格卡数/DIE || MindSpeed-MM规格卡数/DIE || | 支持模型参数量 | 训练策略类型 | 序列长度SEQ_LEN | Snt9b | Snt9b23 | Snt9b | Snt9b23 | Snt9b | Snt9b23 | Snt9b | Snt9b23 | Snt9b | Snt9b23 | |:-----------------------------------------------------------:|:---------------------------------------------------------:|:--------------:|:----------:|:---------:|:---------:|:---------:|:----------:|:----------:|:---|:---|:---|:---| | llama3.1-8b | full | 4096/8192 | 4\*Ascend || 8\*Ascend || - | - | - | - | - | - | | llama3.1-8b | lora | 4096/8192 | 4\*Ascend || 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | llama3.1-70b | full | 4096 | 32\*Ascend || 64\*Ascend || - | - | - | - | - | - | | llama3.1-70b | lora | 4096 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | llama3.1-70b | full | 8192 | 64\*Ascend || 64\*Ascend || - | - | - | - | - | - | | llama3.1-70b | lora | 8192 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | llama3.2-1b | full/lora | 4096/8192 | 1\*Ascend | 2\*Ascend | 1\*Ascend | 1\*Ascend | - | - | - | - | - | - | | llama3.2-3b | full | 4096/8192 | 2\*Ascend || 4\*Ascend || - | - | - | - | - | - | | llama3.2-3b | lora | 4096/8192 | 1\*Ascend | 2\*Ascend | 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2-0.5b | full/lora | 4096/8192 | 1\*Ascend | 2\*Ascend | 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2-1.5b | full/lora | 4096/8192 | 1\*Ascend | 2\*Ascend | - || - | - | - | - | - | - | | qwen2-7b | full | 4096 | 4\*Ascend || 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2-7b | lora | 4096 | 4\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen2-7b | full | 8192 | 8\*Ascend || 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2-7b | lora | 8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen2-72b | full | 4096 | 32\*Ascend || 64\*Ascend || - | - | - | - | - | - | | qwen2-72b | lora | 4096 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen2-72b | full | 8192 | 64\*Ascend || 64\*Ascend || - | - | - | - | - | - | | qwen2-72b | lora | 8192 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen2.5-0.5b | full/lora | 4096/8192 | 1\*Ascend | 2\*Ascend | 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2.5- 1.5b | full/lora | 4096/8192 | 1\*Asce nd | 2\*Ascend | - || - | - | 8\*Ascend || - | - | | qwen2.5-7b | full | 4096 | 4\*Ascend || 8\*Ascend || 8\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | - | - | | qwen2.5-7b | lora | 4096 | 2\*Ascend || 1\*Ascend | 2\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | - | - | | qwen2.5-7b | full | 8192 | 8\*Ascend || 8\*Ascend || 8\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | - | - | | qwen2.5-7b | lora | 8192 | 2\*Ascend || 1\*Ascend | 2\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | 8\*Ascend | - | - | | qwen2.5-14b | full | 4096 | 8\*Ascend || 8\*Ascend || 8\*Ascend | 8\*Ascend | - | - | - | - | | qwen2.5-14b | lora | 4096 | 4\*Ascend || 4\*Ascend || 8\*Ascend | 8\*Ascend | - | - | - | - | | qwen2.5-14b | full | 8192 | 8\*Ascend || 16\*Ascend || 8\*Ascend | 8\*Ascend | - | - | - | - | | qwen2.5-14b | lora | 8192 | 8\*Ascend || 4\*Ascend || 8\*Ascend | 8\*Ascend | - | - | - | - | | qwen2.5-32b | full | 4096 | 16\*Ascend || 32\*Ascend || 16\*Ascend | 16\*Ascend | 16\*Ascend | 16\*Ascend | - | - | | qwen2.5-32b | lora | 4096 | 16\*Ascend || 8\*Ascend || 16\*Ascend | 16\*Ascend | 16\*Ascend | 16\*Ascend | - | - | | qwen2.5-32b | full | 8192 | 16\*Ascend || 32\*Ascend || 16\*Ascend | 16\*Ascend | 16\*Ascend | 16\*Ascend | - | - | | qwen2.5-32b | lora | 8192 | 16\*Ascend || 16\*Ascend || 16\*Ascend | 16\*Ascend | 16\*Ascend | 16\*Ascend | - | - | | qwen2.5-72b | full | 4096 | 32\*Ascend || 64\*Ascend || - | - | - | - | - | - | | qwen2.5-72b | lora | 4096 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen2.5-72b | full | 8192 | 64\*Ascend || 64\*Ascend || - | - | - | - | - | - | | qwen2.5-72b | lora | 8192 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen2vl-2b | full | 4096/8192 | - || 2\*Ascend || - | - | - | - | - | - | | qwen2vl-2b | lora | 4096/8192 | - || 1\*Ascend || - | - | - | - | - | - | | qwen2vl-7b | full | 4096/8192 | - || 8\*Ascend || - | - | - | - | - | - | | qwen2vl-7b | lora | 4096/8192 | - || 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | qwen2vl-72b | full | 1024 | - || 32\*Ascend || - | - | - | - | - | - | | qwen2vl-72b | lora | 1024 | - || 16\*Ascend || - | - | - | - | - | - | | qwen2.5_vl-3b | full | 1024 | - || - || - | - | - | - | 8\*Ascend || | qwen2.5_vl-7b | full | 1024/4096/8192 | - || 8\*Ascend || 8\*Ascend | 8\*Ascend | - | - | 8\*Ascend || | qwen2.5_vl-7b | lora | 4096 | - || 1\*Ascend | 2\*Ascend | 8\*Ascend | 8\*Ascend | - | - | - | - | | qwen2.5_vl-32b | full | 4096 | - || 32\*Ascend || 16\*Ascend || - | - | - | - | | qwen2.5_vl-32b | full | 8192 | - || 64\*Ascend || - | - | - | - | - | - | | qwen2.5_vl-32b | lora | 4096/8192 | - || 16\*Ascend || - | - | - | - | - | - | | qwen2.5_vl-72b | full | 4096/8192 | - || 64\*Ascend || - | - | - | - | - | - | | qwen2.5_vl-72b | lora | 4096/8192 | - || 32\*Ascend || - | - | - | - | - | - | | qwen3-0.6b | full/lora | 4096/8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen3-1.7b | full/lora | 4096/8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen3-4b | full/lora | 4096/8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen3-8b | full/lora | 4096/8192 | 8\*Ascend || 8\*Ascend || 8\*Ascend || - | - | - | - | | qwen3-14b | full/lora | 4096/8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen3-32b | full | 4096 | 16\*Ascend || 32\*Ascend || 16\*Ascend || - | - | - | - | | qwen3-32b | full | 8192 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen3-32b | lora | 4096 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | qwen3-32b | lora | 8192 | 8\*Ascend || 16\*Ascend || - | - | - | - | - | - | | qwen3_moe-30B_A3B | full | 4096 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen3_moe-30B_A3B | full | 8192 | 32\*Ascend || 64\*Ascend || - | - | - | - | - | - | | qwen3_moe-30B_A3B | lora | 4096/8192 | 16\*Ascend || 32\*Ascend || - | - | - | - | - | - | | qwen3_moe-235B_A22B | full | 4096 | 256\*Ascend || 512\*Ascend || - | - | - | - | - | - | | qwen3_moe-235B_A22B | lora | 4096 | 128\*Ascend || 256\*Ascend || - | - | - | - | - | - | | glm4-9b | full | 4096/8192 | 8\*Ascend || 8\*Ascend || - | - | - | - | - | - | | glm4-9b | lora | 4096/8192 | 2\*Ascend || 1\*Ascend | 2\*Ascend | - | - | - | - | - | - | | mixtral-8x7b | full | 4096/8192 | 16\*Ascend || - || - | - | - | - | - | - | | DeepSeek-V3/R1 | full | 4096 | 512\*Ascend || - || - | - | - | - | - | - | | DeepSeek-V3/R1 | lora | 4096 | 64\*Ascend || - || - | - | - | - | - | - | | internvl2.5-8b | full/lora | 4096/8192 | - || 8\*Ascend || - | - | - | - | - | - | | internvl2.5-38b | full | 4096/8192 | - || 32\*Ascend || - | - | - | - | - | - | | internvl2.5-38b | lora | 4096/8192 | - || 16\*Ascend || - | - | - | - | - | - | | internvl2.5-78b | full | 4096 | - || 32\*Ascend || - | - | - | - | - | - | | internvl2.5-78b | full | 8192 | - || 64\*Ascend || - | - | - | - | - | - | | internvl2.5-78b | lora | 4096 | - || 16\*Ascend || - | - | - | - | - | - | | internvl2.5-78b | lora | 8192 | - || 32\*Ascend || - | - | - | - | - | - | | gemma3-27b | full | 4096 | - || 16\*Ascend || - | - | - | - | - | - | | gemma3-27b | full | 8192 | - || 48\*Ascend || - | - | - | - | - | - | | gemma3-27b | lora | 4096/8192 | - || 16\*Ascend || - | - | - | - | - | - | ![](https://support.huaweicloud.com/bestpractice-modelarts/public_sys-resources/note_3.0-zh-cn.png) - Llama-Factory使用的zero并行会将优化器、梯度、权重在多卡上切分，因此集群规模的大小会影响最佳配置与性能。 - 当mindspeed-llm上开启分布式优化器并行时，优化器参数会在集群所有机器上切分共享，因此最优配置会和卡数相关。 - 当前benchmark是综合考虑了最小可运行卡数和最优性能平衡情况下测试出的配置，实际情况中可以根据集群规模大小和性能取舍进行参数调整。