Cosmos3 Edge ViT Runner#

class Cosmos3EdgeViTRunner : public trt_edgellm::rt::Qwen3VLViTRunner#

Cosmos3-Edge reasoner vision encoder: SigLIP2 (packed patches) + Qwen3-VL-style PatchMerger. Follows the qwen3_vl visual engine I/O convention at temporalPatchSize == 1 (per-frame Linear patch embed, per-frame vision spans/timestamps, learned pos-emb via the fast-pos-emb inputs, no rotary input, no deepstack), so the Qwen3-VL runner behavior is inherited unchanged.

Public Functions

virtual bool validateAndFillConfig(
std::string const &engineDir
) override#

Base load + SigLIP2 sanity: temporalPatchSize must be 1 and the engine’s packed-patch input dim must equal the Linear patch-embed width 3 * patchSize * patchSize.