docs(flash-next-seat): orcarouter 2026-09-18 update is a V100 repack script only; weights identical, nothing to adopt

2026-09-26 00:02:11 -07:00
parent ddd67df2e9
commit e4cab7f430
1 changed files with 9 additions and 0 deletions
@@ -17,6 +17,15 @@ stream with async prefetch.
 | Speculative decoding | **MTP k=3** — 60.4% acceptance, mean acceptance length 2.81 (measured here, n=5) |
 | Gateway wiring | **8 aliases** — gen, gen-reasoning, summarizer(-large), classifier, chat-judge, image-judge, qwen-image-bench |

+> **Upstream check, 2026-09-26:** orcarouter's only commit since our 2026-09-14
+> download (`c1209bda`) is `38efbff` (2026-09-18). It **adds `repack_sm70.py`**, a
+> helper that repacks the model for V100 (experts-only NVFP4, bf16 elsewhere), and
+> nothing else. Every weight, config and card file has the same content hash, so
+> there is nothing to adopt. Upstream still ships a bf16 PLE and
+> `qwen_sparse_attention` layer names, so both in-house fixes below remain
+> required. Re-check with the HF tree API (`/api/models/<repo>/tree/<rev>`,
+> gated, needs `$HF_TOKEN`) rather than re-downloading.
+
 ## Deploy

 ```bash