TeleFuser

Integrate an external model or pipeline into TeleFuser while preserving upstream behavior and reusing TeleFuser's existing pipeline, stage, model-loading, configuration, example, CLI, and service interfaces. Use for new model support, new pipeline integration, or porting an upstream inference implementation.

Optimize an existing TeleFuser pipeline using measured bottlenecks and current repository-supported ops, parallelism, caching, quantization, compilation, or offload mechanisms. Use for latency, throughput, GPU memory, OOM, multi-GPU, or inference-performance work after a correct baseline exists.

Profile a TeleFuser pipeline progressively from stage timing to isolated kernel analysis and optional Nsight Compute diagnosis. Use when investigating latency, throughput, memory, output cadence, slow stages, or GPU kernel bottlenecks before optimization.