Measured on Yggdrasil: loading by repo id leaves the model's _name_or_path
equal to the Llama repo id, so prepare_for_tokenization wraps the text in
the chat header and the vector matches assembly A (cosine 1.0, 0.909 to the
reference B). Loading from the snapshot paths reproduces the reference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The encoder child is CPU only (GPU placement, GpuBusy and plan_placement
are deleted) and assembles LLM2Vec B: Llama base, MNTP merged, then the
supervised adapter. The cache identity names assembly, device, dtype,
torch/transformers/peft/accelerate versions, Kimodo commit and the child's
own sha256 beside the three revisions. VILI_THREADS (1..16) caps torch in
the worker and the child. The worker binds port 0 and reports it in its
ready line; errors name a class and place, never the request.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
text_encoding.py holds the CultCache embedding store (gamecult.vili.text_embedding.v0
through cultcache-py), the encoder identity and key, plan_placement with its typed
gpu-busy refusal, and CachedTextEncoder, which sends a request's misses to one child
call. kimodo-text-encode.py is that child: it assembles Kimodo's LLM2Vec encoder in
bf16 with its own loader (operator ruling encoder-own-loader; peft's offload path
raises TypeError in Kimodo's loader), encodes each text alone, writes a MessagePack
answer and exits. GPU reserve 3 GiB: four cold runs on Raven's 16 GB card produced
identical vectors with no CUDA error. The worker and daemon are not wired yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
