apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  annotations:
    serving.kserve.io/deploymentMode: RawDeployment
  labels:
    app.kubernetes.io/name: llm-reasoning
    app.kubernetes.io/part-of: llm-serving
  name: reasoning
  namespace: llm-serving
spec:
  predictor:
    containers:
    - args:
      # Swapped off DeepSeek-R1-Distill-Qwen-32B: tool_choice="auto" (what pi
      # sends) hit a documented vLLM architecture conflict for R1-family
      # models -- the model narrated fake tool-call completions in its
      # <think> block instead of emitting real tool_calls, regardless of
      # parser combo tried (deepseek_v3 400s outright, hermes parsed but the
      # model itself never called out to the real tool-call path). Root
      # cause is upstream in the R1 distillation, not this config -- moving
      # to a Qwen3-family model with a Kimi-K2.6 reasoning distillation
      # instead, since Qwen3's own tool-call format is natively supported.
      #
      # No pre-quantized AWQ/GPTQ/bnb checkpoint exists for this specific
      # distilled model (only a GGUF, which is llama.cpp/Ollama-only and not
      # usable here) -- pointing --quantization=bitsandbytes at the full
      # bf16 checkpoint directly, which makes vLLM quantize on load instead
      # of requiring a pre-quantized repo. This on-the-fly bnb path is
      # well-trodden for dense models but less battle-tested for MoE
      # (this model is 35B total / ~3B active) -- watch first boot closely;
      # if it OOMs or errors on the MoE expert weights, that's the likely
      # cause.
      - --model=lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled
      - --served-model-name=reasoning
      - --quantization=bitsandbytes
      - --trust-remote-code
      - --dtype=bfloat16
      - --kv-cache-dtype=auto
      - --tensor-parallel-size=1
      - --max-model-len=16384
      - --gpu-memory-utilization=0.90
      - --max-num-seqs=4
      - --enable-chunked-prefill
      - --enable-prefix-caching
      # qwen3 parser handles this family's <think> reasoning blocks (best
      # match for this architecture; unverified against this exact
      # checkpoint -- if it 400s or fails to strip <think> tags, that's the
      # first thing to check).
      - --reasoning-parser=qwen3
      # hermes previously verified (on the old model) to work against a
      # Qwen tokenizer's plain-text tool-call patterns without needing
      # special tokens; Qwen3's native tool-call format is also
      # hermes-style, so kept as-is.
      - --enable-auto-tool-choice
      - --tool-call-parser=hermes
      # vLLM 0.11.0's native OffloadingConnector -- spills KV cache blocks
      # to CPU DRAM instead of discarding them on preemption (max-num-seqs=4
      # + max-model-len=16384 means concurrent long sequences compete for
      # the same GPU KV space). No extra dependency, built into vLLM core.
      # num_cpu_blocks=2000 hung the pod at startup ("Allocating 64 CPU
      # tensors..." then nothing -- 64 is this model's layer count, one
      # pinned host tensor per layer, each sized for every CPU block; 2000
      # was oversized enough to stall pinning that much host memory, likely
      # blowing well past the pod's memory limit). Dropped to a small,
      # known-safe starting point -- confirm it actually comes up healthy,
      # then watch real host memory usage and raise it deliberately rather
      # than guessing a round number again. block_size=128 tokens matches
      # vLLM's own example.
      # Note: 0.11.0 ships the original (fragmented, small-transfer-block)
      # version of this connector -- 0.12.0 consolidates KV data into one
      # contiguous block per request and is reported an order of magnitude
      # faster for this specific feature, so this is a real but not yet
      # optimal implementation until the image gets bumped.
      - --kv-transfer-config={"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"num_cpu_blocks":32,"block_size":128}}
      - --host=0.0.0.0
      - --port=8080
      env:
      - name: VLLM_USE_FLASHINFER_SAMPLER
        value: '0'
      - name: VLLM_ATTENTION_BACKEND
        value: TRITON_ATTN
      - name: HF_HOME
        value: /mnt/models
      image: vllm/vllm-openai:v0.11.0@sha256:014a95f21c9edf6abe0aea6b07353f96baa4ec291c427bb1176dc7c93a85845c
      name: kserve-container
      ports:
      - containerPort: 8080
        protocol: TCP
      readinessProbe:
        httpGet:
          path: /health
          port: 8080
        periodSeconds: 10
      resources:
        limits:
          cpu: '16'
          memory: 36Gi
          nvidia.com/gpu: '1'
        requests:
          cpu: '8'
          memory: 12Gi
          nvidia.com/gpu: '1'
      startupProbe:
        failureThreshold: 80
        httpGet:
          path: /health
          port: 8080
        periodSeconds: 15
      volumeMounts:
      - mountPath: /mnt/models
        name: models
      - mountPath: /dev/shm
        name: shm
    deploymentStrategy:
      type: Recreate
    maxReplicas: 2
    minReplicas: 2
    nodeSelector:
      kubernetes.io/hostname: worker-1
    runtimeClassName: nvidia
    volumes:
    - name: models
      persistentVolumeClaim:
        claimName: llm-models
    - emptyDir:
        medium: Memory
        sizeLimit: 2Gi
      name: shm