|
|
@@ -57,6 +57,9 @@ TEB pin now **v2.8.0** (97 scenarios).
|
|
|
(`GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD`, `GLM-5.3-NVFP4-CSF`,
|
|
|
`Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD`) under the LIL license.
|
|
|
- **Action:** proceed with the MiMo 4x rebase on this image (plan unchanged).
|
|
|
+ **DONE 2026-10-15:** re-served on dgx58-switch, recipe re-pinned to
|
|
|
+ `e21ba936…` (recipes `a09e85f`), bench **90/100** v2.8.0, tg 66.2/80.7
|
|
|
+ c1/c2 (+19% decode vs the dev-20261004 anchor). Reports `040e9f2`.
|
|
|
DCP2 is a candidate datapoint for the 2x lane (context ceiling) — but it
|
|
|
needs the CSF checkpoint + their toolchain, so treat as a separate
|
|
|
experiment lane, not a kindling change.
|
|
|
@@ -113,8 +116,8 @@ TEB pin now **v2.8.0** (97 scenarios).
|
|
|
|
|
|
## Recommended order (updated)
|
|
|
|
|
|
-1. **MiMo 4x rebase** — 2026.10.4 validates the pre-pulled image; do it when
|
|
|
- dgx3/4/6/8 free (they're off now).
|
|
|
+1. **MiMo 4x rebase** — **done 2026-10-15** (dgx58-switch, 90/100, +19%
|
|
|
+ decode; shift log 2026-10-15).
|
|
|
2. **v0.32.0 stable → kindling base re-pin** — flashinfer 0.7.1 + SM12x
|
|
|
NVFP4 KV + DSpark scheduler/DFlash graph fixes in one bump; bench-gate
|
|
|
vs 95/100.
|
