Projects项目

Proposed runtime changes to keep imbalanced data-parallel decode compatible with captured microbatches and logical output ordering.提出运行时修改，使不均衡的数据并行解码兼容已捕获的微批布局，并保持逻辑输出顺序。

• Runtime
• CUDA Graphs
• Distributed Systems
• MoE

A readable PyTorch runtime serves the real DeepSeek-V4.1-Flash on 4 × B200 and matches the official reference within 0.05 logits across every scheduling mode.用可读的 PyTorch 运行时在 4 × B200 上服务真实 DeepSeek-V4.1-Flash，在所有调度模式下与官方参考实现的 logits 差距小于 0.05。

• Inference
• Runtime
• Attention
• MoE
• Triton
• Blackwell

An opt-in NVLink Engram-fetch proposal enables a single-node execution path when NCCL GIN is unavailable.提出可选的 NVLink Engram 读取方案，为无法使用 NCCL GIN 的节点提供单节点执行路径。

• Co-design
• Inference
• Memory Systems
• GPU Interconnects

A merged session-aware vLLM runner replays workload traces while preserving request order, arrival times and tool waits.已合并的会话级 vLLM 执行器，在重放负载轨迹时保留请求顺序、到达时间与工具等待。

• Workload Replay
• Inference
• Benchmarking
• Runtime

A Qwen2.5 integration proposal adds model-graph, tokenizer and execution-path support to the Ascend inference framework.通过 Qwen2.5 集成提案，为昇腾推理框架补充模型计算图、分词器与执行路径支持。

• Inference
• Model Integration
• Ascend
• Runtime