feat(llm-serving): serve Qwen3-30B-A3B int4 as reasoning, on one card

Replaces the Qwen3-32B dense deployment, which needed two V100s for 131k context.
The MoE carries 96 KiB of KV per token against the dense model's 256 KiB, so the same
64k window fits on one card.

Measured on this cluster before the switch: 15.55 GiB of weights, 137k tokens of KV
cache, 2.09x concurrency at 65536 tokens per request, correct answer over a 5921-token
prompt, and clean reasoning_content separation through the gateway.

Three settings are load-bearing on sm_70 rather than tuning:
  - VLLM_USE_FLASHINFER_SAMPLER=0  flashinfer's check_cuda_arch raises on CC 7.0
  - VLLM_ATTENTION_BACKEND=TRITON_ATTN  FlashInfer and FA2 both need sm75+
  - deploymentStrategy Recreate  a rolling update deadlocks once every card is allocated

The manifest is committed as well as applied: reasoning.yaml had never been in git.

k8s/apps/llm-serving/kustomization.yaml

@@ -26,12 +26,10 @@ resources:
   #
   # Re-add it when the model gets an Ampere-or-newer GPU. The manifest itself is
   # kept in git (k8s/apps/llm-serving/clef-flash.yaml) with the same reasoning.
-  # ornith and reasoning are live but unmanaged: their manifests were never committed
-  # to this directory, so listing them made `kustomize build` fail and left the whole
-  # Application in ComparisonError. Remove the entries rather than the models. The
-  # Application runs with prune=false, so an object that is no longer in the desired
-  # state is left alone, not deleted. Re-export both from live state when they are
-  # next edited.
+  - reasoning.yaml
+  # ornith is retired: its InferenceService still reports Ready=True while running zero
+  # replicas, because desired equals available at 0. Its manifest was never committed,
+  # so it is left unmanaged and is not listed here. Delete the CR when convenient.
   - embeddings.yaml
   - reranker.yaml
   - qwen-cpu.yaml

k8s/apps/llm-serving/reasoning.yaml

@@ -0,0 +1,102 @@
+apiVersion: serving.kserve.io/v1beta1
+kind: InferenceService
+metadata:
+  name: reasoning
+  namespace: llm-serving
+  labels:
+    app.kubernetes.io/name: llm-reasoning
+    app.kubernetes.io/part-of: llm-serving
+  annotations:
+    serving.kserve.io/deploymentMode: RawDeployment
+spec:
+  predictor:
+    minReplicas: 1
+    maxReplicas: 1
+    nodeSelector:
+      kubernetes.io/hostname: worker-1
+    # Every GPU card is allocated, so a rolling update deadlocks: the new pod needs a
+    # card the old pod still holds and sits Pending forever. Recreate frees the card
+    # first. The cost is a short gap in service for this model only.
+    deploymentStrategy:
+      type: Recreate
+    containers:
+    - name: kserve-container
+      image: vllm/vllm-openai:v0.11.0@sha256:014a95f21c9edf6abe0aea6b07353f96baa4ec291c427bb1176dc7c93a85845c
+      args:
+      # Qwen3-30B-A3B is a 30.5B MoE with 3.3B active per token. On a V100 that matters:
+      # the 32B dense model it replaces needed two cards for the same context, this needs
+      # one, because its KV cache is 96 KiB/token against the dense model's 256 KiB.
+      #
+      # Measured on this cluster, one V100-32GB, 64k context: 15.55 GiB of weights,
+      # 137k tokens of KV cache, 2.09x concurrency at 65536 tokens per request.
+      - --model=Qwen/Qwen3-30B-A3B-GPTQ-Int4
+      - --served-model-name=reasoning
+      - --quantization=gptq
+      - --dtype=float16
+      - --kv-cache-dtype=auto
+      # Native window is 40960. YaRN factor 2.0 reaches 81920, so 64k is inside it.
+      - --rope-scaling={"rope_type":"yarn","factor":2.0,"original_max_position_embeddings":40960}
+      - --tensor-parallel-size=1
+      - --pipeline-parallel-size=1
+      - --max-model-len=65536
+      - --gpu-memory-utilization=0.90
+      - --max-num-seqs=8
+      - --max-num-batched-tokens=8192
+      - --long-prefill-token-threshold=2048
+      - --enable-chunked-prefill
+      - --enable-prefix-caching
+      - --reasoning-parser=qwen3
+      - --enable-auto-tool-choice
+      - --tool-call-parser=hermes
+      - --host=0.0.0.0
+      - --port=8080
+      env:
+      # Both are required on this hardware, not tuning:
+      #  - flashinfer's check_cuda_arch raises AttributeError on compute capability 7.0
+      #    (flashinfer/jit/core.py:55, minor.isdigit() on an int), so the sampler must stay off
+      #  - FlashInfer and FlashAttention-2 both need sm75+, so the Triton backend serves attention
+      - name: VLLM_USE_FLASHINFER_SAMPLER
+        value: "0"
+      - name: VLLM_ATTENTION_BACKEND
+        value: TRITON_ATTN
+      - name: HF_HOME
+        value: /mnt/models
+      ports:
+      - containerPort: 8080
+        protocol: TCP
+      readinessProbe:
+        httpGet:
+          path: /health
+          port: 8080
+        periodSeconds: 10
+      startupProbe:
+        httpGet:
+          path: /health
+          port: 8080
+        # Weight load took 18s from cache and 148s on a cold download, and the first
+        # torch.compile adds about 70s. 80 x 15s leaves ample room for a cold start.
+        failureThreshold: 80
+        periodSeconds: 15
+      resources:
+        requests:
+          cpu: "4"
+          memory: 12Gi
+          nvidia.com/gpu: "1"
+        limits:
+          cpu: "8"
+          memory: 36Gi
+          nvidia.com/gpu: "1"
+      volumeMounts:
+      - mountPath: /mnt/models
+        name: models
+      - mountPath: /dev/shm
+        name: shm
+    runtimeClassName: nvidia
+    volumes:
+    - name: models
+      persistentVolumeClaim:
+        claimName: llm-models
+    - name: shm
+      emptyDir:
+        medium: Memory
+        sizeLimit: 2Gi