February 25, 2026 Efficient Scaling of LLM Training with Flexible Context Parallelism FCP：面向异构长上下文数据的自适应上下文并行策略，支持非2幂并行度、毫秒级调度开销，训练吞吐提升1.46倍 distributed-training context-parallelism long-context llm-training mllm-training load-balancing dynamic-parallelism megatron-lm