sglang-torch-profiler-analysis

To identify dominant GPU kernels in an SGLang torch profiler trace, you can analyze the trace to triage prefill, decode, or extend stages, mapping GPU kernels to CPU operations and Python source locations for bottleneck diagnosis.

Finding kernel fusion opportunities in SGLang profiles involves matching trace evidence against source-backed SGLang, FlashInfer, diffusion, MoE, and attention optimization catalogs before labeling an opportunity as novel.

Yes, you can trigger profiling against a running SGLang server for stage-specific performance triage, or apply the workflow offline to existing trace files and profile directories to capture prefill or decode workloads.

Overlap analysis compares graph-off mapping traces with formal graph-on traces to assess hidden work, dependency risk, and practical overlap headroom, requiring conservative single-trace or mapping-and-formal two-trace analysis.

If an SGLang decode benchmark shows unexpectedly low throughput, trace analysis can determine whether attention, communication, metadata preparation, or MoE kernels dominate and whether an existing fusion or overlap path is disabled.

Analysis of SGLang torch profiler traces requires backend validation before recommending optimizations, and demands conservative single-trace or formal two-trace analysis to properly assess dependency risks across disaggregated workloads.