docs: add re-download warning for late-added jangtq_runtime.safetensors codebook sidecar
Browse files
README.md
CHANGED
|
@@ -19,6 +19,26 @@ tags:
|
|
| 19 |
- jangtq2
|
| 20 |
- reap
|
| 21 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
<p align="center">
|
| 24 |
<a href="https://osaurus.ai"><img src="./osaurus-x-banner.png" alt="Osaurus AI"></a>
|
|
|
|
| 19 |
- jangtq2
|
| 20 |
- reap
|
| 21 |
---
|
| 22 |
+
> ## ⚠️ IMPORTANT — Re-download required if you pulled before 2026-04-25
|
| 23 |
+
>
|
| 24 |
+
> This bundle was published in two commits: the initial commit had only the
|
| 25 |
+
> quantized weight shards, and the codebook + Hadamard-rotation sidecar
|
| 26 |
+
> (`jangtq_runtime.safetensors`) was uploaded ~2.1h later.
|
| 27 |
+
>
|
| 28 |
+
> **If your local copy is missing `jangtq_runtime.safetensors`,** the Swift
|
| 29 |
+
> JANGTQ loader (and any decoder that reads codebooks from a sidecar
|
| 30 |
+
> rather than re-deriving them at load time) will produce incoherent output.
|
| 31 |
+
>
|
| 32 |
+
> **Action:** re-pull the bundle (or just download the single file):
|
| 33 |
+
> ```bash
|
| 34 |
+
> hf download OsaurusAI/MiniMax-M2.7-Small-JANGTQ jangtq_runtime.safetensors --local-dir <your-dir>
|
| 35 |
+
> ```
|
| 36 |
+
> Bundles without this file but with `model.safetensors.index.json` listing
|
| 37 |
+
> only `*.tq_packed`/`.tq_norms`/`.tq_bits` triplets are also fine for the
|
| 38 |
+
> **Python** runtime — codebooks are recomputed from `(in_features, bits)`
|
| 39 |
+
> via `jang_tools.turboquant.codebook.compute_codebook` at load time. The
|
| 40 |
+
> sidecar exists for the Swift native loader.
|
| 41 |
+
|
| 42 |
|
| 43 |
<p align="center">
|
| 44 |
<a href="https://osaurus.ai"><img src="./osaurus-x-banner.png" alt="Osaurus AI"></a>
|