|
|
@@ -173,6 +173,20 @@ DFlash path learns the fp8-KV/block matrix.
|
|
|
|
|
|
## Open actions
|
|
|
|
|
|
+0.1 **#15 RESOLVED (2026-10-17 night): the effort-mapping patch is image-baked.**
|
|
|
+ Per the glm52 precedent (engine-code patches belong in the image, never as
|
|
|
+ per-node hand copies), the 1017 lane now builds
|
|
|
+ `docker-images/dsv41-sglang-1017/` — a thin FROM wrapper over Aiden's
|
|
|
+ production-1.7.1 (arm64 digest-pinned) with `patch_dsv41_effort.py` applied
|
|
|
+ at build. Image tag is versioned (`dsv41-sglang-1017:v1`, checksum label
|
|
|
+ `1b49acd2…`) for rollback; bump to v2 when the patch changes. Verified live:
|
|
|
+ `REASONING_EFFORT_MAPPINGS = {'low': 50, 'high': 75, 'max': 100}` and
|
|
|
+ `high` renders `Reasoning Effort: 75`. Upstream had no fix (Aiden still
|
|
|
+ 1.7.1, eugr repo unchanged since 10-16, sglang upstream silent). Fresh
|
|
|
+ forum find: thread 386121 — TensorFold runs 1017 on dual Sparks (decode
|
|
|
+ excellent on c1; repo x00byte/DeepSeek-V4.1-Flash-1017-Dual-Spark) — a
|
|
|
+ future 2x lane candidate to mine.
|
|
|
+
|
|
|
0. **FULLY RESOLVED (2026-10-17 evening): the launcher now serves the lane.**
|
|
|
The bisect exposed the real root causes, none of which were the b12x stack:
|
|
|
1. the recipe lacked `distributed_backend` → the launcher emitted NO dist
|
