card: Hub revision and verification note
Browse files
README.md
CHANGED
|
@@ -321,6 +321,8 @@ Four folders under `gpu-pipelined/`. A decision needs one decoder and one tower.
|
|
| 321 |
| `decider_2b_vision_g256_vision_fp16w32/` | tower, 256×256 → 64 rows (`.aimodel`) | 660,263,631 | `4a0de698…cede1` |
|
| 322 |
| `decider_2b_vision_g448_vision_fp16w32/` | tower, 448×448 → 196 rows (`.aimodel`) | 662,696,647 | `9bf6023d…d9bdd` |
|
| 323 |
|
|
|
|
|
|
|
| 324 |
No AOT asset ships: the Swift runtime specializes the `.aimodel` correctly on the Mac (the JIT row
|
| 325 |
above). A Mac AOT compile of the ship decoder is 10.2 GB, and the `prefill` function adds 3.8 GB of
|
| 326 |
it (int8lin compiles to 6.2 GB with `main` alone, 10.0 GB with both). The source `config.json`
|
|
|
|
| 321 |
| `decider_2b_vision_g256_vision_fp16w32/` | tower, 256×256 → 64 rows (`.aimodel`) | 660,263,631 | `4a0de698…cede1` |
|
| 322 |
| `decider_2b_vision_g448_vision_fp16w32/` | tower, 448×448 → 196 rows (`.aimodel`) | 662,696,647 | `9bf6023d…d9bdd` |
|
| 323 |
|
| 324 |
+
Uploaded 2026-09-29 as Hub revision `4948e3231035df85c7b03c90b766eb17a08d9451`; after the upload every file's size and sha256 matched the staged copy (the 6 LFS files by the Hub's own hash, the 18 small files by download), and `conversion/zoo_verify.py` reports the two decoder bundles PASS (the towers carry no tokenizer and are skipped).
|
| 325 |
+
|
| 326 |
No AOT asset ships: the Swift runtime specializes the `.aimodel` correctly on the Mac (the JIT row
|
| 327 |
above). A Mac AOT compile of the ship decoder is 10.2 GB, and the `prefill` function adds 3.8 GB of
|
| 328 |
it (int8lin compiles to 6.2 GB with `main` alone, 10.0 GB with both). The source `config.json`
|