feat(gpu): serve 6 models on worker-1 via KServe — vLLM v0.11.0 (bitsandbytes) + Ollama + TEI, plus RuntimeClass/privileged-PSA prereqs and a local-NVMe StorageClass, working around Volta sm_70 limits

Layout on 4x Tesla V100 32GB (PCIe, no NVLink), all TP=1: GPU0+1 vLLM DeepSeek-R1-Distill-Qwen-32B bnb-nf4 (2 replicas) GPU2 Ollama ornith:35b + qwen2.5:3b-instruct (co-resident) GPU3 vLLM Qwen2.5-Math-PRM-7B (reward model) CPU TEI nomic-embed-text-v2-moe, bge-reranker-base Volta constraints, each verified against live output rather than config: - vLLM pinned v0.11.0: sm_70 dropped from CUDA_SUPPORTED_ARCHS at v0.11.1. - AWQ hard-rejected (needs sm_75). GPTQ passes vLLM's min_capability=60 gate but is NUMERICALLY WRONG on sm_70 — emits garbage logits. Proven by an fp16 control run producing correct text on an identical backend. bitsandbytes nf4 verified correct by output. - flashinfer's check_cuda_arch() crashes on any sm_7x (calls .isdigit() on an int) -> VLLM_USE_FLASHINFER_SAMPLER=0. - xformers has no sm_70 kernel for V1's paged-attention bias, and V0 was removed in v0.11.0 -> TRITON_ATTN. - Ornith is Qwen3.5-MoE hybrid-attention; vLLM added that arch after dropping Volta, so no build has both -> Ollama, which also multiplexes a second model on the same card for free. Cluster prereqs that were absent: - RuntimeClass nvidia: the Talos toolkit extension registers the containerd handler but not the k8s object; without it every pod is rejected at admission. - gpu-system pinned to privileged PSA: a device plugin cannot satisfy the cluster-default baseline, it must mount hostPath. - device-plugin affinity=null: the chart requires NFD labels that do not exist here, so it matched zero nodes and reported desiredNumberScheduled=0 silently. - Recreate strategy on GPU services: with GPUs allocated exactly 4/4, a RollingUpdate surge pod has no card and deadlocks the rollout. - longhorn-llm-local SC (1 replica, strict-local, disk tag llm): the default 3-replica class could not place the volume at all (every control-plane disk was at its over-provisioning ceiling), and this keeps ~60GB of weights on worker-1's own NVMe instead of reading them over the network. deploy-gpu-serving.sh sequences ArgoCD syncs (or helm/kubectl in --manual mode) and never applies a manifest absent from git; doctor/unstick/teardown stages exist so this is diagnosable without ad-hoc kubectl archaeology.
2026-08-13 07:02:53 -07:00
parent 2d7127b37e
commit 4fb1c6feeb
15 changed files with 1200 additions and 0 deletions
@@ -0,0 +1,133 @@
+# Reasoning engine — DeepSeek-R1-Distill-Qwen-32B, GPTQ INT4, vLLM.
+#
+# TP=1 with 2 data-parallel replicas (GPU0 + GPU1) rather than one TP=2 engine:
+# worker-1 has NO NVLink, so tensor-parallel's per-token all-reduce would cross
+# PCIe on every decode step. Two independent replicas need zero inter-GPU
+# communication and KServe load-balances them behind one Service.
+#
+# vLLM is pinned to v0.11.0 — the LAST release that compiles sm_70 (Volta)
+# kernels. v0.11.1 dropped 7.0 from CUDA_SUPPORTED_ARCHS. Do not bump this
+# without re-checking CMakeLists.txt, or every pod dies with "no kernel image".
+apiVersion: serving.kserve.io/v1beta1
+kind: InferenceService
+metadata:
+  name: reasoning
+  labels:
+    app.kubernetes.io/name: llm-reasoning
+    app.kubernetes.io/part-of: llm-serving
+spec:
+  predictor:
+    minReplicas: 2
+    # Recreate, not the default RollingUpdate: GPUs are allocated exactly 4/4,
+    # so a surge pod has no card to claim and sits Pending while the old pod is
+    # never torn down — a deadlock. Recreate tears down first, accepting a brief
+    # gap during updates.
+    deploymentStrategy:
+      type: Recreate
+    maxReplicas: 2
+    nodeSelector:
+      kubernetes.io/hostname: worker-1
+    runtimeClassName: nvidia
+    containers:
+      - name: kserve-container
+        image: vllm/vllm-openai:v0.11.0@sha256:014a95f21c9edf6abe0aea6b07353f96baa4ec291c427bb1176dc7c93a85845c
+        args:
+          # GPTQ, NOT AWQ. vLLM hard-refuses AWQ below compute capability 75:
+          #   "The quantization method awq is not supported for the current GPU.
+          #    Minimum capability: 75. Current capability: 70."
+          # V100 is sm_70. GPTQ's min capability is 60, so it runs. (gptq_marlin
+          # needs 80 and fp8 needs 80 — both also out.) Same 19.3GB footprint.
+          # desc_act=False in this build: no activation reordering, faster.
+          - --model=unsloth/DeepSeek-R1-Distill-Qwen-32B-bnb-4bit
+          - --served-model-name=reasoning
+          # bitsandbytes nf4. GPTQ passed vLLM's min_capability=60 check but was
+          # numerically WRONG on sm_70 (garbage logits) — proven by the fp16
+          # control run producing correct text with the identical backend. bnb
+          # declares min_capability=70, but treat that as unverified until the
+          # output itself is checked.
+          # NOTE the repo sets bnb_4bit_compute_dtype=bfloat16, which Volta does
+          # not have; --dtype=float16 must override it.
+          - --quantization=bitsandbytes
+          # Volta has no bf16 — must be explicit, the repo's weights are bf16.
+          - --dtype=float16
+          # No FP8 KV on Volta; stays fp16.
+          - --kv-cache-dtype=auto
+          - --tensor-parallel-size=1
+          - --max-model-len=16384
+          # VRAM budget on a 32GiB V100: 0.92 => ~29.4GiB, minus ~18GiB of GPTQ
+          # weights leaves ~11GiB for KV + activations. One full 32K sequence
+          # costs 32768 x 256KB = 8GiB of KV, so 8 concurrent full-length
+          # sequences is not physically possible here — 4 is honest, and a
+          # serial single-user harness never needs more.
+          - --gpu-memory-utilization=0.90
+          - --max-num-seqs=4
+          # Smooths Volta's slow prefill (no FlashAttention2 on sm_70).
+          - --enable-chunked-prefill
+          - --enable-prefix-caching
+          # Splits <think>…</think> into its own channel.
+          - --reasoning-parser=deepseek_r1
+          - --host=0.0.0.0
+          - --port=8080
+        env:
+          # FlashAttention2 requires sm_80; Volta must fall back to xformers.
+          # flashinfer's check_cuda_arch() has an upstream bug that crashes on
+          # ANY sm_7x GPU: `elif major == 7 and minor.isdigit()` calls .isdigit()
+          # on an int, so instead of reporting "unsupported" it raises
+          #   AttributeError: 'int' object has no attribute 'isdigit'
+          # and engine init dies. Default is None (auto-detect), which walks
+          # straight into that path. 0 disables the flashinfer sampler outright.
+          # Only affects the generate runner — the verifier (pooling) never hits
+          # the sampler, which is why it started fine and this did not.
+          - name: VLLM_USE_FLASHINFER_SAMPLER
+            value: "0"
+          # TRITON_ATTN, not XFORMERS. On sm_70 every xformers kernel is
+          # rejected for V1's paged-attention bias type:
+          #   fa2F / triton_splitKF -> require sm_80
+          #   cutlassF              -> supports sm_70 but not
+          #                            PagedBlockDiagonalCausalWithOffsetPaddedKeysMask
+          # -> NotImplementedError kills EngineCore on the FIRST request, which
+          # takes the whole pod down (vLLM treats engine death as fatal).
+          # V0, whose hand-written paged kernels did support sm_70, was REMOVED
+          # in v0.11.0, so VLLM_USE_V1=0 has nothing to fall back to.
+          # Triton JIT-compiles for the local arch, so it is the last option.
+          - name: VLLM_ATTENTION_BACKEND
+            value: TRITON_ATTN
+          - name: HF_HOME
+            value: /mnt/models
+        ports:
+          - containerPort: 8080
+            protocol: TCP
+        resources:
+          requests:
+            cpu: "8"
+            memory: 8Gi
+            nvidia.com/gpu: "1"
+          limits:
+            cpu: "16"
+            memory: 16Gi
+            nvidia.com/gpu: "1"
+        volumeMounts:
+          - name: models
+            mountPath: /mnt/models
+          - name: shm
+            mountPath: /dev/shm
+        startupProbe:
+          httpGet:
+            path: /health
+            port: 8080
+          # Cold start pulls ~18Gi of weights over Longhorn, then loads to VRAM.
+          periodSeconds: 15
+          failureThreshold: 80
+        readinessProbe:
+          httpGet:
+            path: /health
+            port: 8080
+          periodSeconds: 10
+    volumes:
+      - name: models
+        persistentVolumeClaim:
+          claimName: llm-models
+      - name: shm
+        emptyDir:
+          medium: Memory
+          sizeLimit: 2Gi