domains / families / science / nemo-mbridge-perf-expert-parallel-overlap

nemo-mbridge-perf-expert-parallel-overlap

separate

Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.

When to use

• Enabling EP overlap to hide dispatch/combine latency
• Tracing throughput regression to EP overlap config change
• Configuring overlap_moe_expert_parallel_comm
• Configuring delay_wgrad_compute
• Using flex dispatcher with DeepEP
• Using flex dispatcher with HybridEP
• Benchmarking plain EP overlap isolation
• Validating HybridEP production-shape overlap

When not to use

• Runs with full activation recompute enabled
• Runs with moe_shared_expert_overlap enabled
• Runs being brought up for correctness
• PyTorch versions below 2.6.0
• Tiny or communication-light runs
• Runs dominated by another wall

Skill page

Source