b70-optimization-lab

Reference Lab Storage Layout

For outside readers: every absolute path beginning /mnt/fast-ai/ or /mnt/usb-models/ anywhere in this repository is a path on the maintainer’s reference lab host. Those paths do not exist on your machine and are not expected to. They are recorded so that a result stays traceable to the exact bytes it was measured against, not as instructions you can run unmodified.

When following a recipe from repro/ or community/, substitute your own model directory for any such path. Nothing else about the recipe depends on where the weights happen to live.

For the maintainer: this is the map of what sits where, and which storage has to be mounted before a lane will run.

Last verified: 2026-07-25.

Internal NVMe

/dev/nvme0n1p2, 915 GB ext4, always present.

Path Size Notes
/mnt/fast-ai/llm-models/laguna-s-2.1 70 G Active campaign model: 68 G int4 target, 2.1 G dflash-int4 draft
/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-q8-gguf 57 G Gemma 4 26B Q8 reference/production lane
/mnt/fast-ai/llm-models/gemma4-26b-a4b-it-hf-tokenizer 62 M Tokenizer for the above
/mnt/fast-ai/llm-cache/hf 46 G+ Hugging Face cache
/mnt/fast-ai/bench-results 205 G Run evidence; several CURRENT.md record identities point here
/mnt/fast-ai/llm-optimization-artifacts 23 G Sealed run roots
/mnt/fast-ai/deepseek-v4-corpora 25 G Prompt corpora

Only the active Laguna model and the Gemma 4 26B pair are kept on NVMe as working weights. Everything else was relocated on 2026-07-25 to free the disk, which had reached 98% full.

External USB Drive

/dev/sda2, 3.6 TB NTFS, volume label CorsairExternal, mounted at /mnt/usb-models.

This drive does not auto-mount. If it is not mounted, every path below fails with a “no such file or directory” error that looks like missing data rather than a missing drive. Mount it first:

sudo mount -t ntfs-3g -o rw,uid=1000,gid=1000,umask=022 /dev/sda2 /mnt/usb-models

Note that mount -o remount,rw does not work on this filesystem; ntfs-3g silently keeps the old mode and writes fail with a misleading ENOENT. Unmount and mount fresh instead.

Relocated on 2026-07-25, verified byte-for-byte before the source was removed:

Model Size Files
deepseek-v4-flash-xpu 114 G 110 — K160 target plus DSpark draft, paused record lane
minimax-m2.7-int4-autoround 121 G 78
gemma4-12b-it-int4-autoround-intel 7.8 G 32
gemma4-12b-it-int4-autoround-intel-vllm-compat 32 M 29

Already resident before that move: qwen3.6-27b-fp8-vrfai, qwen36-27b-awq-int4-cyankiwi-8f269fb, qwen3.6-35b-a3b-int4-autoround-abhinand, minimax-m2.7-reap-autoround-w4a16, gemma3-12b-it-int4-autoround-opea, gemma4-12b-it-int4-autoround-vishva007, gemma4-26b-a4b-it-dflash-hf, gemma4-26b-a4b-it-eagle3-gguf, gemma4-26b-a4b-it-qat-gguf, gemma4-e4b-it-gguf. The drive also holds an hf-cache/ tree with further Qwen3.6-27B quantizations.

Old Paths Still Work

Each relocated model left a symlink behind at its original NVMe path:

/mnt/fast-ai/llm-models/deepseek-v4-flash-xpu
    -> /mnt/usb-models/llm-models/deepseek-v4-flash-xpu

So the ~168 documents in this repo that reference /mnt/fast-ai/llm-models/... remain correct while the drive is mounted, including symlinks stored inside the moved trees, which use absolute paths and resolve back through the compatibility link. No recipe, config, or note needed editing.

The single failure mode to remember: with the drive unmounted, those paths look like deleted models rather than an unmounted disk.