Dynamo KVBM

The Dynamo KVBM is a distributed KV-cache block management system designed for scalable LLM inference. It cleanly separates memory management from inference runtimes (vLLM, TensorRT-LLM, and SGLang), enabling GPU↔CPU↔Disk/Remote tiering, asynchronous block offload/onboard, and efficient block reuse.

Feature Highlights

• Distributed KV-Cache Management: Unified GPU↔CPU↔Disk↔Remote tiering for scalable LLM inference.
• Async Offload & Reuse: Seamlessly move KV blocks between memory tiers using GDS-accelerated transfers powered by NIXL, without recomputation.
• Runtime-Agnostic: Works out-of-the-box with vLLM, TensorRT-LLM, and SGLang via lightweight connectors.
• Memory-Safe & Modular: RAII lifecycle and pluggable design for reliability, portability, and backend extensibility.

Installation

pip install kvbm

See the support matrix for version compatibility questions.

Build from Source

The pip wheel is built through a Docker build process:

# Render and build the Docker image with KVBM enabled (from the dynamo repo root)
python container/render.py --framework dynamo --target runtime --output-short-filename
docker build --build-arg ENABLE_KVBM="true" -f container/rendered.Dockerfile .

Once built, you can either:

Option 1: Run and use the container directly

./container/run.sh --framework none -it

Option 2: Extract the wheel file to your local filesystem

# Create a temporary container from the built image
docker create --name temp-kvbm-container local-kvbm:latest

# Copy the KVBM wheel to your current directory
docker cp temp-kvbm-container:/opt/dynamo/wheelhouse/ ./dynamo_wheelhouse

# Clean up the temporary container
docker rm temp-kvbm-container

# Install the wheel locally
pip install ./dynamo_wheelhouse/kvbm*.whl

Note that the default pip wheel built is not compatible with CUDA 13 at the moment.

Integrations

Environment Variables

Disk Storage Configuration

Why special configuration may be needed:

Some filesystems (e.g., Lustre, certain network filesystems) don't support fallocate(), which KVBM uses for fast disk space allocation. Additionally, KVBM uses O_DIRECT I/O for GPU DirectStorage (GDS) performance, which requires strict 4096-byte alignment.

Setup for filesystems without fallocate() support:

export DYN_KVBM_DISK_CACHE_DIR=/mnt/storage/kvbm_cache
export DYN_KVBM_DISK_ZEROFILL_FALLBACK=true  # Enables zero-fill fallback when fallocate() unsupported

What happens:

• Without ZEROFILL_FALLBACK=true: Disk cache allocation may fail with "Operation not supported"
• With ZEROFILL_FALLBACK=true: KVBM writes zeros using page-aligned buffers compatible with O_DIRECT requirements

Troubleshooting: If you encounter "write all error" or EINVAL (errno 22), try disabling O_DIRECT: export DYN_KVBM_DISK_DISABLE_O_DIRECT=true

vLLM

DYN_KVBM_CPU_CACHE_GB=100 vllm serve \
  --kv-transfer-config '{"kv_connector":"DynamoConnector","kv_role":"kv_both","kv_connector_module_path":"kvbm.vllm_integration.connector"}' \
  Qwen/Qwen3-8B

See the KVBM configuration reference for all KVBM settings.

TensorRT-LLM

cat >/tmp/kvbm_llm_api_config.yaml <<EOF
cuda_graph_config: null
kv_cache_config:
  enable_partial_reuse: false
  free_gpu_memory_fraction: 0.80
kv_connector_config:
  connector_module: kvbm.trtllm_integration.connector
  connector_scheduler_class: DynamoKVBMConnectorLeader
  connector_worker_class: DynamoKVBMConnectorWorker
EOF

DYN_KVBM_CPU_CACHE_GB=100 trtllm-serve Qwen/Qwen3-8B \
  --host localhost --port 8000 \
  --backend pytorch \
  --extra_llm_api_options /tmp/kvbm_llm_api_config.yaml

See the KVBM configuration reference for all KVBM settings.

📚 Docs

• KVBM Configuration
• NIXL Overview