Jens/ModelForge
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial public ModelForge release

Browse Source
This commit is contained in:
Jens
2026-09-01 21:30:16 +02:00
commit 7082ab955a
490 changed files with 104252 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/architecture/VISION_EMBEDDINGS.md
+9
View File
@@ -0,0 +1,9 @@
# vision.embedding@1
`vision.embedding@1` accepts a bounded batch of image or text items. Each item has exactly one modality. Images are PNG/JPEG, base64 encoded at the gateway boundary and constrained to 4096×4096 pixels in the runtime.
The response exposes normalized vectors, their dimension, and an immutable `EmbeddingSpace` identifier. Equal dimensions never imply compatible spaces. Identity binds the exact artifact revision and digest, RuntimeProfile fingerprint, processor, input image size, normalization and pooling/projection semantics.
The preferred SigLIP2 SO400M candidate is retained as discovery evidence but is not forced into a shared 16 GiB GPU envelope. The M9 safe baseline is SigLIP2 Base patch16/224, loaded offline by `transformers_siglip2`.
Evaluation uses a deterministic, locally generated image/text retrieval set and typed Recall@K, MRR, latency and peak-VRAM evidence. Fixtures never leave the local environment.