|
|
@@ -72,6 +72,20 @@ args = [
|
|
|
]
|
|
|
|
|
|
[env]
|
|
|
+# Master gate for every TP4 production adapter (sitecustomize's
|
|
|
+# _tp4_launcher check — start-tp4.sh sets this; without it the gated
|
|
|
+# installs (indexer v3, shared_pad_k, l2_prefetch, moe_b12x_next, …)
|
|
|
+# silently stay off and the stock FlashInfer path rejects the 576-wide
|
|
|
+# EP1 experts. Must be set before anything else arms.
|
|
|
+DSV41_LAUNCHER = "tp4"
|
|
|
+# boot.py exports these to the container; the adapters read them from env
|
|
|
+# (not from sglang's flags). TP_SIZE is load-bearing: tp3_pad.py defaults to
|
|
|
+# 3 when unset, pads heads 64->96 under our TP4 and the sparse-MLA kernel
|
|
|
+# aborts on 24 heads/rank. DSV41_TP_PAD=0 is their own TP4 setting (docs/tp4.md:
|
|
|
+# everything divides 4, the pad has nothing to fix).
|
|
|
+TP_SIZE = "4"
|
|
|
+EP_SIZE = "1"
|
|
|
+DSV41_TP_PAD = "0"
|
|
|
# ---- image defaults our path must keep explicit ----
|
|
|
OFFLOAD_MODE = "nvme" # never ram (Engram would evict the model)
|
|
|
STATE_PATH = "/state" # boot.py's state dir (sparks tables, api key)
|
