|
|
@@ -4,10 +4,13 @@
|
|
|
# sparkring tp4 profile; wiring/cluster facts stay ours.
|
|
|
#
|
|
|
# Image (anonymous digests pull):
|
|
|
-# ghcr.io/fujitsupolycom/sparkring@sha256:71d41057…5620ad
|
|
|
-# line dev-20261004-kraken-cuda1342-nccl2323-status034 (karmic-kraken
|
|
|
-# base, the active lab GLM lineage) + the SparkRing fixes we do
|
|
|
-# NOT have in our nightly:
|
|
|
+# ghcr.io/fujitsupolycom/sparkring@sha256:e21ba936…acd76f
|
|
|
+# = the SparkRing **2026.10.4** release image (line
|
|
|
+# dev-20261013-dcp2prefill-status038-sircl-tp6; release record pins this
|
|
|
+# digest). Rebase from 71d41057… (dev-20261004-kraken, status034):
|
|
|
+# runtime-status 0.3.8 (serving stats on the status page), sircl 0.6.0,
|
|
|
+# DCP2 prefill work, TP6 profile added; every 2026.10.2 profile admitted
|
|
|
+# unchanged (their release note). The fixes below still apply:
|
|
|
# * B12X selection-cache correction (restarted ranks reuse tuning;
|
|
|
# their record: DS4.1 restart 200-225s vs 652-741s),
|
|
|
# * MiMo vision attention sinks (in-softmax sinks as trained — without
|
|
|
@@ -43,7 +46,7 @@
|
|
|
|
|
|
engine = "vllm"
|
|
|
name = "MiMo-V2.6-Flash-MOPD (SparkRing installer image, 4x switched)"
|
|
|
-docker_image = "ghcr.io/fujitsupolycom/sparkring@sha256:71d410571407fef3ce2959c6d392f5a2c3f44b757b856e853a71f6e3295620ad"
|
|
|
+docker_image = "ghcr.io/fujitsupolycom/sparkring@sha256:e21ba936d4fd6bfeb6e7a651e54a03ac461bf4e1e6f5721c1a34845c80acd76f"
|
|
|
entrypoint = ["python3", "/opt/sparkring/toolchain/toolchain.py", "serve"]
|
|
|
# Floating model: the engine resolves the org/model id via the NAS HF cache
|
|
|
# (refs/main -> latest snapshot; the flat snapshot dir carries dflash/ for the
|
