Cross-track lesson · Run & build

Weights, Caches and Volumes

Where a QB2 actually keeps model weights and compiled kernels, and how to make tt-inference-server reuse what already shipped instead of downloading it again.

Your QB2 shipped with Qwen3-32B already on disk. Serving Models on QB2 shows the one command that uses it. This lesson is the layer underneath: what that directory actually is, why the obvious flag is the wrong one, and how to reuse the compiled kernel cache as well as the weights.

What’s actually in that directory

~/data/tt-cache is tt-inference-server’s persistent volume root. Each model it knows about gets a cache_root beneath it, and the one your box shipped with looks like this:

~/data/tt-cache/                                        <- persistent volume root
└── volume_id_tt_transformers-Qwen3-32B-vqb2_launch/    <- one model's cache_root
    ├── weights/Qwen3-32B/                              <- ~62 GB, a normal HF model dir
    ├── tt_metal_cache/cache_Qwen3-32B/P300x2/P150x4/   <- ~30 GB of compiled kernels
    ├── logs/
    └── model_file_symlinks_map/

weights/Qwen3-32B/ holds exactly what you’d expect — config.json, tokenizer.json, model.safetensors.index.json and 17 .safetensors shards. Only the location is unusual, which is why --host-weights-dir works on it directly.

It is not a Hugging Face cache, and this is the part that catches people: a factory QB2 has no ~/.cache/huggingface and no ~/models at all. Reaching for --host-hf-cache therefore resolves HOST_HF_HOME, then HF_HOME, then ~/.cache/huggingface, finds nothing, and downloads all 62 GB again.

Reusing the compiled kernels, and the version trap in the way

--host-weights-dir gets you the weights but not the 30 GB of compiled kernels. For those you point run.py at the whole persistent volume with --host-volume ~/data/tt-cache — but don’t run that yet, because on a stock box it does the opposite of what you want.

run.py doesn’t look inside that directory to see what’s there. It computes the name it expects, as volume_id_<impl>-<model>-v<version>. That version is not the tt-inference-server release number; it comes from the model spec entry matching your model and device. For Qwen3-32B on p300x2 that entry is impl: tt_transformers, version: 0.17.0, so run.py goes looking for volume_id_tt_transformers-Qwen3-32B-v0.17.0.

Your box ships volume_id_tt_transformers-Qwen3-32B-vqb2_launch — a build label, not a version. The names don’t match, so run.py finds nothing to reuse and quietly starts downloading 62 GB into a brand-new volume beside the one it was meant to use. Set up the alias first, then launch.

Step 1 — ask your box which name it wants

Don’t take the version above on faith; it moves between releases, and there are four different Qwen3-32B spec entries.

# HF_TOKEN first: run.py validates it up front whenever --docker-server is passed,
# including on this dry run. Without it you get "⛔ HF_TOKEN not set." and no output.
export HF_TOKEN=hf_...

python3 ~/.local/lib/tt-inference-server/run.py \
  --model Qwen3-32B --tt-device p300x2 \
  --workflow server --docker-server \
  --host-volume ~/data/tt-cache \
  --print-docker-cmd --skip-system-sw-validation

--print-docker-cmd prints the docker run it would issue and starts nothing, so this is safe to run before the alias exists. Read the src= path in the --mount line — that’s the exact directory name it wants. The same output carries TT_CACHE_PATH=.../tt_metal_cache/cache_Qwen3-32B/P300x2, confirming the kernel cache you’re about to alias is the one the container will read.

--skip-system-sw-validation is there only because validation shells out to tt-smi and will stop you before anything prints if the chips are busy. You’re only printing a command, so skipping it is safe — don’t carry that flag into a real run.

Step 2 — alias the shipped directory to that name

Substituting whatever name you actually saw:

Copy the directory name out of the src= path Step 1 printed — don’t retype the v0.17.0 above, which is only what this spec version happened to want:

cd ~/data/tt-cache

# Paste the basename of the src= path from Step 1 here.
WANTED=volume_id_tt_transformers-Qwen3-32B-v0.17.0

ln -s volume_id_tt_transformers-Qwen3-32B-vqb2_launch "$WANTED"
ls -l "$WANTED"        # should resolve to the -vqb2_launch directory

Step 3 — serve, reusing both weights and kernels

export HF_TOKEN=hf_...

python3 ~/.local/lib/tt-inference-server/run.py \
  --model Qwen3-32B --tt-device p300x2 \
  --workflow server --docker-server \
  --host-volume ~/data/tt-cache

Because the alias tracks a version that moves, upgrading tt-inference-server will eventually invalidate it and you’ll be back at step 1. That maintenance cost is exactly why --host-weights-dir is the better default for most people: it’s version-independent, and a one-time kernel compile beats a symlink that silently stops matching.