SageMaker HyperPod#
An L2 adapter backed by the ai-toolkit daemon of Amazon SageMaker HyperPod, which provides shared memory access for LMCache.
Required fields:
url: ai-toolkit daemon URL (sagemaker-hyperpod://host:port).
Optional fields:
bucket(string, default"lmcache"): ai-toolkit cache namespace.shared_memory_name(string, default"shared_memory"): Name of shared memory segment.max_concurrent_requests(int, default100): Maximum concurrent HTTP requests.max_connections(int, default256): Total HTTP connection-pool limit.max_connections_per_host(int, default128): Per-host connection limit.timeout_ms(int, default5000): HTTP transport timeout.lease_wait_timeout_ms(int, default1000): How long the daemon may hold a lease request before answering.lease_ttl_ms(int, default30000): Server-side lease lifetime.put_stream_chunk_bytes(int, default65536): HTTP PUT chunk size.max_lease_size_mb(float, optional): Upper bound for accepted lease payloads.use_https(bool, defaultfalse): Enable HTTPS instead of HTTP.
Configuration examples:
# Basic (node IP via the Kubernetes downward API)
--l2-adapter '{"type": "sagemaker-hyperpod", "url": "sagemaker-hyperpod://'"${NODE_IP}"':9200"}'
# Custom namespace and faster miss detection
--l2-adapter '{"type": "sagemaker-hyperpod", "url": "sagemaker-hyperpod://'"${NODE_IP}"':9200", "bucket": "team-a", "lease_wait_timeout_ms": 250}'
Kubernetes Deployment Requirements#
The ai-toolkit daemon is node-local (hostNetwork, port 9200), so the LMCache server must run on the same node as the daemon.
Environment Variable for Node IP#
Add the NODE_IP environment variable to resolve the local node’s IP
address:
env:
- name: NODE_IP
valueFrom:
fieldRef:
fieldPath: status.hostIP
/dev/shm Volume Configuration#
SageMaker Hyperpod requires /dev/shm for high-performance shared memory operations:
volumeMounts:
- name: dshm
mountPath: /dev/shm/shared_memory
subPath: shared_memory
volumes:
- name: dshm
hostPath:
path: /dev/shm