LMCache MP 0.5.5 with GLM-5.3-Flash fails on non-contiguous tensor or no IPC lookup key

Summary
When attempting to run LMCache MP against VLLM (glm53-flash tag) with lmcache manually upgraded to 0.5.5, there are 2 issues:

• In isolatedIPC: true mode, LMCache fails to initialize, throwing

ValueError: RawCudaIPCWrapper requires a tensor that is contiguous (possibly after permutation); got shape=(11285, 1, 640, 848) stride=(655360, 542720, 848, 1)

• In isolatedIPC: false mode, with a hostPath /dev/shm mounted to both the lmcache and vllm pod, LMCache does initialize normally, but every chat request results in a warning getting logged on the lmcache pod, indicating the request is essentially ignored:

LMCache WARNING: Session <ID> has no lookup ipc key, skipping touch

Details
Using LMCache in-process with VLLM isn't working at all for GLM 5.2/5.3/5.3-Flash models, the 0.5.5 version in MP mode was the first time (in non-isolated IPC mode) the model would come up fully functional, except the LMCache pod ignoring every request.

Steps / Reproduction (if applicable)

1. Create VLLM container image capable of running GLM 5.3 Flash with LMCache 0.5.5
   Build from this Dockerfile (also adds Fastokens):

FROM docker.io/vllm/vllm-openai:glm53-flash
RUN uv pip install --system fastokens
RUN uv pip install --system --upgrade lmcache==0.5.5 awscrt==0.36.2 cuda-bindings==13.3.1 cuda-core==1.0.1 cuda-python==13.3.1 fastapi==0.136.3 google-auth==2.57.1 huggingface-hub==1.30.0 llvmlite==0.47.0 multidict==6.7.1 numba==0.65.0 nvidia-cudnn-cu13==9.20.0.48 nvidia-nvjitlink==13.3.33 nvtx==0.2.15 regex==2026.9.3 torch==2.13.0+cu130 tqdm==4.70.0 transformers==5.16.1 triton==3.7.1

The version of other packages is to ensure those are not touched by the lmcache upgrade.

2. Deploy LMCache Operator nightly (the CRDs in the nightly build support spec.isolatedIPC)

kubectl apply -f https://github.com/LMCache/LMCache/releases/download/operator-nightly-latest/install.yaml

3. Deploy a v0.5.5 LMCacheEngine:

apiVersion: lmcache.lmcache.ai/v1alpha1
kind: LMCacheEngine
metadata:
  name: lmcache-glm-53-flash
  namespace: default
spec:
  isolatedIPC: true
  nodeSelector:
    nvidia.com/gpu.present: "true"
  image:
    repository: lmcache/standalone
    tag: v0.5.5-cu130
  server:
    port: 6555
    chunkSize: 640
    maxWorkers: 32
    hashAlgorithm: blake3
  l1:
    sizeGB: 4
  eviction:
    policy: LRU
    triggerWatermark: 0.8
    evictionRatio: 0.2
  l2Backend:
    raw:
      type: fs
      config:
        base_path: /engine-data/lmc/glm53flash
        num_workers: 32
        max_capacity_gb: 8192
  env:
  - name: PYTHONHASHSEED
    value: "0"
  volumeMounts:
  - mountPath: /engine-data
    name: engine-data
  volumes:
  - hostPath:
      path: /data/engine-data
      type: DirectoryOrCreate
    name: engine-data

4. Deploy model with VLLM

apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-test
  namespace: default
spec:
  replicas: 1
  selector:
    matchLabels:
      engine: vllm
  strategy:
    type: Recreate
  template:
    metadata:
      labels:
        engine: vllm
        lmcache.ai/lmcache-inject: "true"
      annotations:
        lmcache.ai/lmcache-engine: "lmcache-glm-53-flash"
    spec:
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
            - matchExpressions:
              - key: nvidia.com/gpu.present
                operator: In
                values:
                - "true"
      containers:
      - args:
        - /engine-data/cache/hf/hub/models--zai-org--GLM-5.3-Flash/snapshots/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a
        - --port
        - "8000"
        - --served-model-name
        - glm-5.3-flash
        - --max-model-len
        - "270336"
        - --gpu-memory-utilization
        - "0.92"
        - --max-num-seqs
        - "128"
        - --max-num-batched-tokens
        - "16000"
        - --mamba-cache-mode
        - align
        - --mm_encoder_tp_mode
        - data
        - --mm_processor_cache_type
        - shm
        - --mm_processor_cache_gb
        - "32"
        - --chat-template
        - /engine-data/cache/hf/hub/models--zai-org--GLM-5.3-Flash/snapshots/eb9eb208eb0d988989d07a6a12d0fdeb5f52574a/chat_template.jinja
        - --chat-template-content-format
        - string
        - --enable-expert-parallel
        - --tensor-parallel-size
        - "8"
        - --no-enable-flashinfer-autotune
        - --tool-call-parser
        - glm47
        - --enable-auto-tool-choice
        - --reasoning-parser
        - glm45
        env:
        - name: VLLM_DEEP_GEMM_WARMUP
          value: skip
        - name: VLLM_ENGINE_READY_TIMEOUT_S
          value: "3600"
        - name: VLLM_USE_FASTOKENS
          value: "1"
        - name: VLLM_USE_V2_MODEL_RUNNER
          value: "1"
        - name: HF_HOME
          value: /engine-data/cache/hf
        - name: HSA_COREDUMP_FILE
          value: /engine-data/dumps/core.gpu
        - name: PYTHONHASHSEED
          value: "0"
        - name: TILELANG_CACHE_DIR
          value: /engine-data/cache/tilelang
        - name: TMPDIR
          value: /engine-data/cache/tmp
        - name: TRITON_CACHE_DIR
          value: /engine-data/cache/triton
        - name: VLLM_CACHE_ROOT
          value: /engine-data/cache/vllm
        - name: XDG_CACHE_HOME
          value: /engine-data/cache
        image: custom/vllm-openai:glm53-flash-fastokens-lmcache
        imagePullPolicy: IfNotPresent
        name: engine
        ports:
        - containerPort: 8000
          name: http
          protocol: TCP
        resources:
          limits:
            nvidia.com/gpu: 8
        volumeMounts:
        - mountPath: /engine-data
          name: engine-data
      volumes:
      - hostPath:
          path: /data/engine-data
          type: DirectoryOrCreate
        name: engine-data

Expected Outcome / Goal
Isolated IPC mode should come up normally, not error out requiring a tensor that is contiguous. Non-isolated IPC mode should not result in all queries getting ignored with ipc key lookup failures.

Start by reproducing both isolatedIPC modes using the supplied Dockerfile and LMCacheEngine and vLLM deployment manifests. Inspect the RawCudaIPCWrapper failure and the session lookup-key warning, then verify that isolated IPC initializes and non-isolated IPC processes chat requests without either reported error.