#6518 [GUI] Model picker shows "Download size" for a fully downloaded Qwen-family model (IQ3_S); Gyro-S in the same folder shows "Already on this computer"

open · @gdup213 · 1 コメント · GitHub で見る

Setup & installServer & APINVIDIA / CUDAModels & quantsWindows

本文

**Version**

- Strata desktop 0.2.18, control 0.1.64 (commit `6e38f588`), engine 0.1.64 `windows-x86_64-cuda-cuda13`
- Also seen on desktop 0.2.17 / control 0.1.63
- Windows 11 Pro 10.0.26300, RTX 5090 32 GB, 128 GB RAM

**What happens**

In the first-run screen "Select the model you want to run", with the model folder set to a Strata data folder that already holds both models:

- "Gyro · Smallest (Gyro-S)" shows "Already on this computer".
- "High quality (IQ3_S)" shows "83.6 GB Download size".

Both models were downloaded by Strata's own setup into the same data folder. Both IQ3_S shards are complete and have their `.done` marks:

```
<data>\models\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf   54,817,524,224 bytes  (+ .done)
<data>\models\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf   28,800,138,432 bytes  (+ .done)
<data>\models\gyro-Gyro-S\Qwen3.8-Flash-Next-Gyro-S-TQ1_0.gguf             58,487,015,808 bytes  (+ .done)
<data>\models\mmproj-Qwen3.8-Flash-Next-BF16.gguf                             907,543,008 bytes  (+ .done)
```

No `.part` files exist.

**What I expected**

"Already on this computer" on the IQ3_S card, as on the Gyro-S card.

**What actually happens after Continue**

Nothing is downloaded. The setup log reports both shards "already downloaded" and "model files present", and the model loads. `strata launch --preset <iq3_s preset> --plan` also says "Nothing to download". So only the card is wrong, but it tells the user that Continue will download 83.6 GB.

**Cause (from reading the bundled control code)**

`files_here()` in `tools/strata_mcp.py` (about line 1693) sizes each expected file with `S.fam_pin(fam, path)[0]` when the file entry carries no `bytes`. `fam_pin()` in `setup.py` (about line 2274) starts with `fam["sha256"][s.name]`. The `qwen` family has no `sha256` table, so this raises `KeyError('sha256')`. `files_here()` catches it and returns `None`; `install_plan()` then sets `files_complete` to `None`, and the page falls back to the download size.

Families and whether `FAMILIES[family]` has a `sha256` table in control 0.1.64:

| family | has `sha256` | entries |
|---|---|---|
| qwen | no | 0 |
| swift | no | 0 |
| coder | no | 0 |
| nvfp4 | no | 0 |
| unsloth | yes | 50 |
| gyro | yes | 2 |

So I expect the same card fault for every model in `qwen`, `swift`, `coder` and `nvfp4`. I have only observed it for `qwen` IQ3_S.

**Reproduction with Strata's own functions (read-only)**

Run with the bundled Python, `-B`, with `control` and `control/tools` on `sys.path`:

```python
import setup as S, strata_mcp as M
from pathlib import Path
d = Path(r"<data>\models")
print(M.files_here(S, "qwen", "IQ3_S", d, True))    # None
print(M.files_here(S, "gyro", "Gyro-S", d, True))   # {'complete': True, 'left_gb': 0.0, 'total_gb': 59.4}
fam = S.FAMILIES["qwen"]
S.fam_pin(fam, d / "IQ3_S" / "Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf")   # KeyError: 'sha256'
```

**Suggested fix**

Either of:

1. In `files_here()`, when a family has no `sha256` table, take the size from wherever setup pins these files for that family (the pin `download()` reads through `file_pin(url)`), so the size check works for every family.
2. Failing a pinned size, treat a file that exists with its `.done` mark and no `.part` as present, which is the test `download()` itself applies first.

**Related**

- #3996 was a similar "downloaded model shows as not on this computer" report in the TUI, with a different cause (folder-name case).

関連リンク

インストール・モデル・リリースへのサイト内リンク。