Skip to content

[Feature request]: Add C++ multimodal benchmark (VLM perf testing without Python) #2373

Description

@amd-mingw

model_benchmark (C++) only benchmarks text-only decoder models via OgaTokenizer + AppendTokenSequences. There is no C++ equivalent of benchmark/python/benchmark_multimodal.py for vision-language and other multimodal models.

This gap matters for environments where Python is unavailable or undesirable (Android, embedded, CI perf gates, Foundry Local–style deployments) but multimodal models still need end-to-end perf measurement (vision prefill + text decode).

Any plan to support this feature? Thanks!

Metadata

Metadata

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions