mimo26-spark: a15 — relax sglang RoCEnante gate from world==8 to b12x's 2..16

The sglang port carried SG17's TP8-only guard while the b12x core accepts
worlds 2..16 (vLLM lanes already run RoCEnante on 4 nodes). a15 widens it,
keeps the StatelessProcessGroup exclusion and the two-HCA check.
mimo26-flash-mopd-4x: GMU 0.88 + SGLANG8_ROCE_ALLREDUCE=1 for the trial;
2x/rl/8x recipes pinned to a15 (identical behavior until the env is set).

docker-images/mimo26-spark/build.sh

@@ -7,7 +7,7 @@
 # [build.versions] overrides (recipe wins over the image defaults).
 set -euo pipefail
 cd "$(dirname "$0")"
-TAG="${TAG:-mimo26-spark:a14}"
+TAG="${TAG:-mimo26-spark:a15}"
 
 ARGS=()
 while IFS= read -r line; do

docker-images/mimo26-spark/patches/sglang/rocenante.diff

@@ -62,8 +62,8 @@ index 46aef5b945..fddacd7257 100644
              del data
  
 +        if roce_enabled and os.environ.get("SGLANG8_ROCE_ALLREDUCE", "0") == "1":
-+            if self.world_size != 8 or isinstance(group, StatelessProcessGroup):
-+                raise RuntimeError("SG17 RoCEnante requires the TP8 Gloo exchange group")
++            if not (2 <= self.world_size <= 16) or isinstance(group, StatelessProcessGroup):
++                raise RuntimeError("SG17 RoCEnante needs a Gloo exchange group of 2..16 ranks")
 +            from b12x.comm import roce
 +
 +            self.roce = roce.AllReduce.from_exchange_group(

recipes/mimo26-flash-mopd-2x/recipe.toml

@@ -31,7 +31,7 @@
 
 engine = "sglang"
 name = "MiMo-V2.6-Flash-MOPD (SGLang DFlash, 2x Spark TP2)"
-docker_image = "mimo26-spark:a14"
+docker_image = "mimo26-spark:a15"
 docker_image_dir = "mimo26-spark"
 model = "XiaomiMiMo/MiMo-V2.6-Flash-MOPD"
 parent_model = "XiaomiMiMo/MiMo-V2.6-Flash-MOPD"

recipes/mimo26-flash-mopd-4x/recipe.toml

@@ -11,13 +11,13 @@
 # GMU trials (SGLang accounts mem-fraction-static more precisely than vLLM's
 # ~0.84 ceiling on this stack — 2x booted clean at 0.90 with 8 GB GPU + 7 GB
 # host headroom, so pushing further on TP4 is cheap): 0.90 -> 0.92 -> 0.94.
-# Current: 0.94 (prod). Measured 2026-09-28: 0.92 → pool 3,971,456, GPU 6.3
+# Current: 0.88 (prod) + RoCEnante trial. Historical trial data: Measured 2026-09-28: 0.92 → pool 3,971,456, GPU 6.3
 # GB free; 0.94 → pool 4,138,432, GPU 5.35 GB free, host ~4 GB during the
 # 515K-token probe (reclaimable page cache, not an OOM signal — the probe
 # served clean in 340 s). SGLang's static fraction is accounted tightly:
 # both values boot and serve without a hitch, so the knob is free to push.
 #
-# Image: docker-images/mimo26-spark (a14) — same as the 2x lane; at TP4 the
+# Image: docker-images/mimo26-spark (a15) — same as the 2x lane; at TP4 the
 # FP8 tile fallbacks of the 2x regime apply (RoCEnante one-shot all-reduce
 # is TP8-only, stock NCCL over the 200G fabric is used).
 #
@@ -27,7 +27,7 @@
 
 engine = "sglang"
 name = "MiMo-V2.6-Flash-MOPD (SGLang DFlash, 4x Spark TP4)"
-docker_image = "mimo26-spark:a14"
+docker_image = "mimo26-spark:a15"
 docker_image_dir = "mimo26-spark"
 model = "XiaomiMiMo/MiMo-V2.6-Flash-MOPD"
 parent_model = "XiaomiMiMo/MiMo-V2.6-Flash-MOPD"
@@ -49,7 +49,7 @@ args = [
     "--enable-multimodal",
     "--moe-runner-backend", "marlin",
     "--moe-a2a-backend", "none",
-    "--mem-fraction-static", "0.94",
+    "--mem-fraction-static", "0.88",
     "--context-length", "524288",
     "--max-running-requests", "8",
     "--chunked-prefill-size", "8192",
@@ -87,8 +87,11 @@ host = "~/.cache/mimo26-jit/sglang"
 container = "/root/.cache/sglang"
 
 [env]
-# RoCEnante overlay is TP8-only (needs the TP8 Gloo exchange group) — stock
-# NCCL on TP4.
+# RoCEnante (b12x one-shot RoCE all-reduce): the sglang port hard-gated it to
+# world=8 ("TP8 Gloo exchange group") while the b12x core supports 2..16 — the
+# vLLM lanes already ran it on 4 nodes. a15 relaxes the sglang guard to the
+# core range, so TP4 can try it; the two-active-HCA check is untouched.
+SGLANG8_ROCE_ALLREDUCE = "1"
 # a13/a14 tuning switches
 MIMO26_FP8_OPROJ = "1"
 MIMO26_FP8_DRAFT = "1"

recipes/mimo26-flash-rl-2x/recipe.toml

@@ -26,7 +26,7 @@
 
 engine = "sglang"
 name = "MiMo-V2.6-Flash-RL (SGLang DFlash, 2x Spark TP2)"
-docker_image = "mimo26-spark:a14"
+docker_image = "mimo26-spark:a15"
 docker_image_dir = "mimo26-spark"
 model = "XiaomiMiMo/MiMo-V2.6-Flash-RL"
 parent_model = "XiaomiMiMo/MiMo-V2.6-Flash-RL"

recipes/mimo26-pro-rl-8x/recipe.toml

@@ -24,7 +24,7 @@
 
 engine = "sglang"
 name = "MiMo-V2.6-Pro-RL (SGLang DFlash, 8x Spark TP8)"
-docker_image = "mimo26-spark:a14"
+docker_image = "mimo26-spark:a15"
 docker_image_dir = "mimo26-spark"
 model = "XiaomiMiMo/MiMo-V2.6-Pro-RL"
 parent_model = "XiaomiMiMo/MiMo-V2.6-Pro-RL"