Crate taconite_clip 

CLIP ViT-H/14 (laion/CLIP-ViT-H-14-laion2B-s32B-b79K) image and text embeddings on an AMD XDNA NPU.

The bundle iron/applications/clip_vit_h14/export_clip.py writes holds every compiled IRON kernel and the model’s weights (the NPU ones pre-packed); this crate replays the forward the Python app runs:

	NPU	host (here)
image	all 32 layers (tower.rs)	resize / crop / normalise (preprocess.rs), the patch embedding (f32), CLS + position embedding, pre- and post-LayerNorm, the projection
text	all 24 layers	BPE tokenizer, token + position embedding, final LayerNorm at the end token, the projection

Clip::encode_images / Clip::encode_texts give the projected embeddings, logits CLIP’s scaled cosine similarities.

Modules

npu

The NPU side: every kernel of the bundle loaded once, kernels naming the same xclbin sharing its hardware context (8 in all, of NPU2’s 16), and the device buffers they run over.

preprocess

Image preprocessing, as HF’s CLIPImageProcessor (torchvision backend):

tower

One transformer tower on the device, as iron/applications/clip_vit_h14/clip_npu.py’s NpuTower: per layer

Structs

Clip

Config

Model constants (the manifest’s param records).

Timing

Wall time per stage, in first-seen order; NPU dispatch time is kept under npu:<kernel>.

Enums

Error

Constants

VERSION

Functions

logits

CLIP’s logits [n_images, n_texts]: logit_scale * cos(image, text).

normalize

Rows of x scaled to unit length.

softmax

Row-wise softmax.