feat(llm-serving): run clef-serve so /v1/systemone answers

clef-serve was parked at 0 while candle-vllm took a card. worker-1 holds
four V100s, so the coder takes one and clef takes one, leaving two free
for a TP=2 candle run.

The gateway already routes the clef-flash model to clef-serve:80 on
/v1/systemone, so this single replica count is what turns that route on.

k8s/apps/llm-serving/clef-serve.yaml

@@ -33,6 +33,16 @@ metadata:
     app: clef-serve
     app.kubernetes.io/part-of: llm-serving
 spec:
+  # One card. worker-1 holds four V100s: candle-vllm takes one for the coder, and
+  # this pod takes one, so two stay free for a TP=2 candle run.
+  #
+  # Enabled 2026-11-04 to serve /v1/systemone, which the gateway already routes to
+  # this Service. It was parked at 0 while candle-vllm was brought up on the same
+  # cards, and it served no inference request between 2026-10-30T06:23:40Z and then.
+  #
+  # Cold start is about four minutes: the pod pip-installs on every boot because the
+  # torch base image cannot be pushed to the registry (Cloudflare answers 413 above
+  # 100 MB per layer). startupProbe allows for that.
   replicas: 1
   # Recreate, not RollingUpdate. Every GPU card is allocated, so a rolling update leaves
   # the new pod Pending forever waiting for a card the old pod still holds.

k8s/apps/llm-serving/coder-service.yaml

@@ -0,0 +1,40 @@
+# Stable client-facing name for the coder chat model, deliberately decoupled
+# from the process that serves it.
+#
+# poimen-memory addresses a Service, not a pod and not an InferenceService. The
+# vLLM -> candle-vllm handoff therefore changes this file's selector and nothing
+# else, so the client config never changes again and there is no window in which
+# the host name is wrong.
+#
+# The previous KTML predictor requested two V100s and never reached Ready, so it
+# held half the GPU pool while serving nothing. candle-vllm replaces it.
+#
+# Two ports on purpose:
+#   - 80    matches every other predictor Service in this namespace, and matches
+#           a client base URL that carries no port number.
+#   - 8080  is the container port. Publishing it directly avoids a port
+#           translation that is easy to get wrong in a curl.
+#
+# Do not delete and recreate this Service: a delete drops the ClusterIP, and a
+# client that cached the name during the gap fails until it re-resolves.
+apiVersion: v1
+kind: Service
+metadata:
+  name: coder-predictor
+  namespace: llm-serving
+  labels:
+    app.kubernetes.io/part-of: llm-serving
+    app: coder-predictor
+spec:
+  type: ClusterIP
+  selector:
+    app: candle-vllm
+  ports:
+    - name: http
+      protocol: TCP
+      port: 80
+      targetPort: 8080
+    - name: http-8080
+      protocol: TCP
+      port: 8080
+      targetPort: 8080

k8s/apps/llm-serving/coder.yaml

@@ -0,0 +1,134 @@
+apiVersion: serving.kserve.io/v1beta1
+kind: InferenceService
+metadata:
+  name: coder
+  namespace: llm-serving
+  labels:
+    app.kubernetes.io/name: llm-coder
+    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
+    # Two GPUs are free, so this rolls out beside `reasoning` instead of over it.
+    # Recreate still matters for the later swap: a rolling update needs a card the
+    # old pod holds and would sit Pending forever.
+    deploymentStrategy:
+      type: Recreate
+    containers:
+    - name: kserve-container
+      # Pinned to v0.11.0 on purpose. This is the last image measured to work on this
+      # node: torch 2.8.0+cu128, arch list sm_70..sm_120, driver 580.159.04.
+      # v0.15.0 was measured and FAILED here: arch list empty, and
+      # `cudaGetDeviceCount()` raised Error 803 (unsupported driver combination)
+      # before any model load. Do not raise this tag without re-running that probe.
+      image: vllm/vllm-openai:v0.11.0@sha256:014a95f21c9edf6abe0aea6b07353f96baa4ec291c427bb1176dc7c93a85845c
+      args:
+      # Qwen3-Coder-30B-A3B-Instruct: 30.5B total, 3.3B active. Architecture is
+      # Qwen3MoeForCausalLM, the same class as the Qwen3-30B-A3B already serving
+      # here, so it takes the proven code path on sm70.
+      #
+      # Qwen ships this model only as BF16 (61.1 GB) and FP8 (no FP8 on Volta).
+      # 4-bit GPTQ is the only form that fits one 32 GB card. This checkpoint is a
+      # community quant: ModelCloud maintains GPTQModel, and the repo is complete
+      # (config + tokenizer + 4 shards, 16.9 GB). Verified by loading, not by trust.
+      #
+      # Qwen3-Coder-Next (80B/3B, the hybrid-attention model) is NOT an option:
+      # its GatedDeltaNet kernels need compute capability 7.5+ and the `fla` module,
+      # and sm70 has neither. Measured, not assumed.
+      - --model=ModelCloud/Qwen3-Coder-30B-A3B-Instruct-GPTQMODEL-W4A16-A
+      # `reasoning` stays a served name so an existing client that sends
+      # model=reasoning keeps working after the DNS rename. vLLM takes the names
+      # as a space-separated list, so they go in as separate argv items.
+      - --served-model-name
+      - coder
+      - reasoning
+      - --quantization=gptq
+      - --dtype=float16
+      - --kv-cache-dtype=auto
+      # Native window is 262144 and rope_scaling is null, so 256K needs no YaRN.
+      # One card cannot hold it: measured KV pool on a single V100 was 132,672
+      # tokens. TP=2 halves the weights per card (7.8 GiB each), which frees room
+      # for the larger pool. Measured after boot: see the KV line in the log.
+      - --max-model-len=262144
+      - --tensor-parallel-size=2
+      - --pipeline-parallel-size=1
+      - --gpu-memory-utilization=0.90
+      # A single 262,144-token request consumes the whole pool, so keep few slots.
+      - --max-num-seqs=4
+      - --max-num-batched-tokens=8192
+      - --long-prefill-token-threshold=2048
+      - --enable-chunked-prefill
+      - --enable-prefix-caching
+      # Measured: with the default (custom all-reduce ON) the engine finished CUDA graph
+      # capture in 6s at 17:01:56 and then went silent for 14+ minutes, emitting only
+      # "No available shared memory broadcast block found in 60 seconds". The front-end
+      # never received the ready message. vLLM's custom all-reduce registers its own
+      # CUDA graph addresses and deadlocks on this Volta/PCIe pair. NCCL is the fallback
+      # and is correct here, because TP=2 across two PCIe V100s gains nothing from the
+      # custom kernel anyway.
+      - --disable-custom-all-reduce
+      - --reasoning-parser=qwen3
+      - --enable-auto-tool-choice
+      - --tool-call-parser=hermes
+      - --host=0.0.0.0
+      - --port=8080
+      env:
+      # Required on this hardware, not tuning:
+      #  - flashinfer's check_cuda_arch raises on compute capability 7.0,
+      #    so the sampler must stay off
+      #  - FlashInfer and FlashAttention-2 need sm75+, so Triton 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
+        # TP=2 at 256K needs ~10 min before the API binds: 22s weights, 80s torch.compile,
+        # 85s KV profiling, then graph capture. 120 x 15s = 30 min of headroom.
+        failureThreshold: 120
+        periodSeconds: 15
+      resources:
+        # worker-1 exposes only 63,495 MiB of allocatable memory, so a 24Gi request
+        # never schedules. Two vLLM workers keep the weights on the GPU and need
+        # little host RAM, so the request stays small and the limit allows burst.
+        requests:
+          cpu: "8"
+          memory: 12Gi
+          nvidia.com/gpu: "2"
+        limits:
+          cpu: "16"
+          memory: 40Gi
+          nvidia.com/gpu: "2"
+      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
+        # NCCL all-reduce over TP=2 moves through /dev/shm. The 64Mi default is
+        # too small and shows up as an init hang or a bus error.
+        sizeLimit: 8Gi

k8s/apps/llm-serving/kustomization.yaml

@@ -14,7 +14,10 @@ resources:
   # v0.5.22+, published only on CUDA 13, and CUDA 13 dropped Volta kernels. The model
   # instead runs the implementation it ships (joint_schema_model.py) in fp16 on one V100.
   - clef-serve.yaml
-  - reasoning.yaml
+  # Stable client-facing name for the chat model. poimen-memory points at this Service,
+  # so the vLLM -> candle-vllm handoff is a selector edit here and no client change.
+  - coder-service.yaml
+  - reasoning-long-service.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.

k8s/apps/llm-serving/networkpolicy.yaml

@@ -66,6 +66,21 @@ spec:
           port: 80
         - protocol: TCP
           port: 8000
+    # Allow from poimen namespace (poimen-memory ingest + query).
+    # The client-side egress rule is the other half and must stay in step:
+    # NetworkPolicy poimen/poimen-memory-llm-egress permits port 8080 to pods
+    # labelled part-of=llm-serving, which is the same selector used here. This
+    # policy had no matching ingress rule at all, so the packet was allowed out
+    # and dropped here. The client then saw a bare timeout and the predictor log
+    # stayed empty, which is why this failure looked like a hung inference engine
+    # for 34 days.
+    - from:
+        - namespaceSelector:
+            matchLabels:
+              kubernetes.io/metadata.name: poimen
+      ports:
+        - protocol: TCP
+          port: 8080
     # Allow intra-namespace (pod-to-pod within llm-serving)
     - from:
         - podSelector:

k8s/apps/llm-serving/reasoning-long-service.yaml

@@ -0,0 +1,43 @@
+# Stable client-facing name for the chat model, deliberately decoupled from the
+# process that currently serves it.
+#
+# poimen-memory addresses this Service, not a pod and not an InferenceService. That
+# means the vLLM -> candle-vllm handoff changes only this file's selector. The client
+# config never changes again, so the ingest cutover has no window in which the host
+# name is wrong.
+#
+# Two ports on purpose:
+#   - 80    matches every other predictor Service in this namespace, and matches
+#           poimen-memory's LLM_CHAT_BASE, which carries no port number.
+#   - 8080  is the container port. It is also the only port llm-serving-ingress and
+#           poimen-memory-llm-egress permit, and the port the handoff readiness check
+#           uses. Publishing it directly avoids a port translation that is easy to get
+#           wrong in a curl.
+#
+# The selector names the vLLM reasoning pod today. To hand off, replace the three
+# selector labels with candle-vllm's pod labels and nothing else. Do not delete and
+# recreate this Service: a delete drops the ClusterIP, and a client that cached the
+# name during the gap will fail until it re-resolves.
+apiVersion: v1
+kind: Service
+metadata:
+  name: reasoning-long-predictor
+  namespace: llm-serving
+  labels:
+    app.kubernetes.io/part-of: llm-serving
+    app: reasoning-long-predictor
+spec:
+  type: ClusterIP
+  selector:
+    app.kubernetes.io/name: llm-reasoning
+    component: predictor
+    serving.kserve.io/inferenceservice: reasoning
+  ports:
+    - name: http
+      protocol: TCP
+      port: 80
+      targetPort: 8080
+    - name: http-8080
+      protocol: TCP
+      port: 8080
+      targetPort: 8080

k8s/apps/llm-serving/reasoning.yaml

@@ -30,7 +30,28 @@ spec:
       # 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
+      # Two served names, one pod, on purpose. `reasoning-long` is the name the goal map
+      # uses for this model, and the name qwen-finetune wiring targets. It is NOT a
+      # separate InferenceService: this model already occupies the only free card, and a
+      # second pod on the same weights cannot schedule (4 of 4 GPUs allocated: coder 2,
+      # this 1, clef-serve 1). Serving both names from one process costs nothing and
+      # keeps the two callers working without a second weights copy.
+      #
+      # `reasoning` is deliberately still served here. poimen-memory pins
+      # LLM_CHAT_BASE=reasoning-predictor, so dropping the name here breaks ingest
+      # immediately. coder-predictor also serves `reasoning`; that ambiguity is known
+      # and documented in the PR. Resolve it by moving the client host, not by
+      # removing the name here.
+      #
+      # The two names are separate list entries, NOT one comma-joined value. This
+      # image's parser returns ['reasoning,reasoning-long'] for the comma form, i.e. a
+      # single model whose name contains a comma, and api_server.py then iterates
+      # args.served_model_name to build the model list. Verified against vLLM 0.11.0:
+      #   FlexibleArgumentParser().parse_args([... "--served-model-name", "a", "b"])
+      #     -> ['a', 'b']
+      - --served-model-name
+      - reasoning
+      - reasoning-long
       - --quantization=gptq
       - --dtype=float16
       - --kv-cache-dtype=auto