From 7a0c9452987835aa1a6ab1333819a94cea1c9432 Mon Sep 17 00:00:00 2001
From: Story Crater Bot <19826264+Riotpiaole@users.noreply.github.com>
Date: Fri, 21 Aug 2026 15:19:43 -0700
Subject: [PATCH] fix: enforce-eager and longer startup probe for
 reasoning-predictor (first-time GDN compile was getting killed mid-way)

---
 k8s/apps/llm-serving/reasoning.yaml | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/k8s/apps/llm-serving/reasoning.yaml b/k8s/apps/llm-serving/reasoning.yaml
index d33a568..5fde4ae 100644
--- a/k8s/apps/llm-serving/reasoning.yaml
+++ b/k8s/apps/llm-serving/reasoning.yaml
@@ -52,6 +52,13 @@ spec:
       # for this specific GPTQ-Int4 release -- not hermes.
       - --enable-auto-tool-choice
       - --tool-call-parser=qwen3_coder
+      # skip CUDA graph capture / torch.compile -- this arch's custom ops
+      # (mamba_mixer2, gdn_attention_core) are compiling for the first time
+      # ever on this hardware with no cache, and startupProbe kept killing
+      # the pod mid-compile every ~20min before it could finish. Trade some
+      # runtime throughput for a startup that actually completes; revisit
+      # once this is confirmed working end to end.
+      - --enforce-eager
       - --host=0.0.0.0
       - --port=8080
       env:
@@ -81,7 +88,7 @@ spec:
           memory: 12Gi
           nvidia.com/gpu: '2'
       startupProbe:
-        failureThreshold: 80
+        failureThreshold: 240
         httpGet:
           path: /health
           port: 8080