HOARDED FORTUNE GROUP PTE. LTD. / 新加坡鸿德集团 Distributed training 分布式训练方案 REQUIREMENTS CHECKLIST / 双语需求清单 Project requirements record. Signed quotations and service agreements define pricing, capacity and delivery terms. 项目需求记录。签署的报价与服务协议规定价格、容量和交付条件。 Model footprint / 模型需求 Provide parameter scale, precision, sequence length, global/micro batch, optimiser and data/tensor/pipeline parallelism assumptions. 提供参数规模、精度、序列长度、全局及微批大小、优化器,以及数据、张量和流水线并行假设。 Customer input / 客户填写:________________ Dataset and schedule / 数据与训练计划 Define dataset format, preprocessing, tokens or samples, checkpoint cadence and target training window. Identify reproducibility and access constraints. 明确数据格式、预处理、Token 或样本量、检查点周期及训练窗口,并说明复现与访问限制。 Customer input / 客户填写:________________ Memory budget / 内存预算 Peak-memory measurement; batch/sequence profile; parallel strategy. 峰值内存测量、批量及序列特征、并行策略。 Required scope / 所需范围:________________ Communication path / 通信路径 Topology; collectives; rank count; transport; tuning baseline. 拓扑、集合操作、进程数、传输协议、调优基线。 Required scope / 所需范围:________________ Training data lifecycle / 训练数据生命周期 Dataset version; input rate; checkpoint size; restart test. 数据版本、输入速率、检查点大小、重启测试。 Required scope / 所需范围:________________ Training productivity / 训练产出 Report tokens or samples per second and step time with model, precision, batch, sequence and scale stated. 报告每秒 Token 或样本数、训练步耗时,并说明模型、精度、批量、序列及规模。 Agreed target and conditions / 约定目标及条件:________________ Communication evidence / 通信证据 NCCL test results distinguish operation time, algorithm bandwidth and bus bandwidth; retain the test context. NCCL 测试区分操作时间、算法带宽及总线带宽,需保留测试上下文。 Agreed target and conditions / 约定目标及条件:________________ Checkpoint impact / 检查点影响 Validate checkpoint write time and restart recovery, recording the recovery point and lost work. 验证检查点写入耗时与重启恢复,并记录恢复点和工作量损失。 Agreed target and conditions / 约定目标及条件:________________ Additional review / 补充审阅 Region and access locations / 区域及访问地点:________________ Contracting entity and approved contact / 签约主体及认可联系:________________ Delivery window and project gates / 交付窗口及阶段条件:________________ Currency, billing unit, term and budget basis / 币种、计费单位、期限及预算口径:________________ Support and responsibility split / 支持及责任分工:________________ Data retention, export and deletion / 数据保留、导出及删除:________________ Acceptance owner and evidence format / 验收负责人及证据格式:________________ Credentials, identity documents and sensitive datasets use designated secure channels. 凭据、身份证明文件及敏感数据集通过指定安全渠道交付。