Module glm5_next_vision 

2026-09-25: Bind the GLM-5.3 vision tower (*.visual.*) as BF16 device pointers.

Owner: model-arch weight loader (GLM-5.3). Invariants:

• Every bound tensor is BF16; any other dtype fails the load (bf16).
• No tensor is copied: each field is the store’s own pointer.
• The patch-embed conv weight is read as [hidden_size, 3 * temporal_patch_size * patch_size^2] and the downsample conv weight as [out_hidden_size, hidden_size * spatial_merge_size^2], over the same bytes; check_shape compares the leading dim and the product of the trailing dims.

Functions

load_glm5_next_vision

2026-09-25: The body of Glm5NextWeightLoader::load_vision_encoder.