v13.0.2 — the install carried 3 GB of CUDA it cannot reach

[13.0.2] — 2026-09-07 — the install carried 3 GB of CUDA it cannot reach

Fixed — the torch stack moved to a rerank extra

A base install resolved sentence-transformers, transformers,
FlagEmbedding and peft. Each of them resolves torch, and torch on Linux
resolves the entire nvidia-cu* set. Measured with uv pip compile for
x86_64-unknown-linux-gnu / py3.12, summing the wheels PyPI serves:

	packages	wheel bytes
before	147	~3,108 MB
after	97	~113 MB

The largest single items were torch (529 MB), nvidia-cudnn-cu13 (528 MB),
nvidia-cublas (404 MB) and triton (237 MB) — none of which a CPU host runs.
FlagEmbedding additionally dragged in datasets, pyarrow and ir-datasets.

This was not merely wasteful. The Glama build sandbox ran
out of disk unpacking nvidia-cudnn-cu13 and the server failed to build there
at all.

And the default configuration cannot reach any of it. MEMORY_MODE=fast — the
default since v11 — sets MEMORY_RERANK_ENABLED=false, and it also sets
MEMORY_ALLOW_OLLAMA_IN_HOT_PATH=false, which is the exact flag gating the
fall-through to SentenceTransformer in Recall._compute. So the base install
shipped 3 GB for two code paths its own defaults forbid, and a third
(ai_layer/verifier.py) that only benchmarks/v11_pipeline.py calls.

Install it when you turn the reranker on with MEMORY_MODE=deep or
MEMORY_RERANK_ENABLED=true:

pip install "total-agent-memory[rerank]"      # pip / uvx / pipx
pip install -r requirements-rerank.txt        # clone / Docker

Both loaders in src/reranker.py now distinguish "not installed" from "failed
to load" and name the extra in the log line instead of reporting a bare
ImportError. Retrieval continues in RRF order without them.

Fixed — every installer warmed a model the server does not use

install.sh, install.ps1, install-codex.ps1 and setup.sh pre-downloaded
all-MiniLM-L6-v2 through sentence_transformers. That is the fallback
model, not the fastembed default the server embeds with, so the warm-up
populated a cache nothing reads and the first real save still downloaded. The
install.sh copy also ran through the system python3 rather than the venv it
had just built. All four now warm FASTEMBED_MODEL through the venv.

EMBEDDING_MODEL=all-MiniLM-L6-v2 was likewise written into every IDE config,
the Dockerfile, docker-compose.yml and docker/run-mcp.sh; it named a model
the install can no longer load and has been dropped. TRANSFORMERS_CACHE and
TORCHINDUCTOR_CACHE_DIR left the image with the stack that used them.

setup.sh — unreferenced by the README since the rebrand — installed a
hand-written dependency list that floored mcp[cli] at >=1.0.0, the exact
floor that broke every install before 13.0.0, and never installed fastembed at
all. It now installs requirements.txt like every other path.

Added

tests/test_dependency_declaration.py gains the mirror of its original guard.
It has always checked that nothing required is missing from the wheel
metadata; it now also checks that the torch stack is not present in it, that
the rerank extra carries everything src/reranker.py imports, that
requirements-rerank.txt and the extra agree, and that no installer warms a
model it cannot load.