domains / families / science / nemo-mbridge-perf-cuda-graphs

nemo-mbridge-perf-cuda-graphs

separate

Validate and use CUDA graph capture in Megatron Bridge, including local full-iteration graphs and Transformer Engine scoped graphs for attention, MLP, and MoE modules.

When to use

• Reducing host-driver overhead via CUDA graphs
• tracing a crash or regression to a CUDA graph config change
• cuda_graph_impl
• full iteration graph
• TE scoped graph
• graphed callables
• CUDA graph capture

Skill page

Source