From c2040e0f179c5a2405ed967f46c2148a547e864a Mon Sep 17 00:00:00 2001
From: Story Crater Bot <19826264+Riotpiaole@users.noreply.github.com>
Date: Fri, 21 Aug 2026 15:49:49 -0700
Subject: [PATCH] fix: drop reasoning-predictor to TP=1 single-GPU (TP=2 hit a
 vLLM shm_broadcast deadlock, not a compile/timeout issue)

---
 k8s/apps/llm-serving/reasoning.yaml | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)

diff --git a/k8s/apps/llm-serving/reasoning.yaml b/k8s/apps/llm-serving/reasoning.yaml
index 5fde4ae..2ddad69 100644
--- a/k8s/apps/llm-serving/reasoning.yaml
+++ b/k8s/apps/llm-serving/reasoning.yaml
@@ -23,8 +23,20 @@ spec:
       # Requires vLLM >=0.17.0 -- v0.11.0 errors with "Model architectures
       # ['Qwen3_5MoeForConditionalGeneration'] are not supported for now."
       # moe_wna16 is the checkpoint's documented quantization kernel;
-      # compute-capability requirement on sm70 is UNVERIFIED going in --
-      # this rollout is the real test. --kv-cache-dtype stays auto, not
+      # confirmed working on sm70 -- weights loaded fine (10.6GiB/rank at
+      # TP=2, no crash). What actually broke TP=2 was a shm_broadcast
+      # deadlock between the two worker processes post-load (matches a
+      # known vLLM bug class, e.g. vllm-ascend#9405 -- logic bug in the
+      # broadcast ring buffer, unaffected by shm size/timeouts/eager-mode,
+      # all of which were tried there too). Dropped to tensor-parallel-size=1
+      # to sidestep the cross-process sync entirely -- full unsharded
+      # weights (~21GB, roughly 2x the per-rank figure above) fit on one
+      # 32GB V100 with room for KV cache. gpu-memory-utilization raised
+      # accordingly (0.5 was sized for the TP=2 split, too low for
+      # unsharded weights on a single GPU). Second V100 sits idle for now;
+      # pipeline-parallel-size=2 is the next thing to try if it's needed
+      # back, since PP unblocked the previous model instead of TP too.
+      # --kv-cache-dtype stays auto, not
       # fp8_e5m2: V100 has no FP8 tensor cores at all (Hopper/Ada only),
       # hardware-blocked regardless of vLLM version. tool-call-parser
       # changed hermes -> qwen3_coder per the checkpoint's own README
@@ -39,9 +51,9 @@ spec:
       - --quantization=moe_wna16
       - --dtype=float16
       - --kv-cache-dtype=auto
-      - --tensor-parallel-size=2
+      - --tensor-parallel-size=1
       - --max-model-len=16384
-      - --gpu-memory-utilization=0.5
+      - --gpu-memory-utilization=0.85
       - --max-num-seqs=4
       - --enable-chunked-prefill
       - --enable-prefix-caching
@@ -82,11 +94,11 @@ spec:
         limits:
           cpu: '16'
           memory: 36Gi
-          nvidia.com/gpu: '2'
+          nvidia.com/gpu: '1'
         requests:
           cpu: '8'
           memory: 12Gi
-          nvidia.com/gpu: '2'
+          nvidia.com/gpu: '1'
       startupProbe:
         failureThreshold: 240
         httpGet: