From a2c6696bfee0ec2275dc08b15eee154ed0e9c0c7 Mon Sep 17 00:00:00 2001 From: Ricardo Decal Date: Thu, 7 Aug 2025 00:29:13 -0700 Subject: [PATCH] [Docs] Factor out troubleshooting to its own guide; add section for Ray Observability (#21578) Signed-off-by: Ricardo Decal --- docs/serving/distributed_serving.md | 29 +++++++++++---------- docs/serving/distributed_troubleshooting.md | 16 ++++++++++++ 2 files changed, 31 insertions(+), 14 deletions(-) create mode 100644 docs/serving/distributed_troubleshooting.md diff --git a/docs/serving/distributed_serving.md b/docs/serving/distributed_serving.md index 08d889a00d2cf..fc9d9f8a34347 100644 --- a/docs/serving/distributed_serving.md +++ b/docs/serving/distributed_serving.md @@ -128,12 +128,17 @@ vllm serve /path/to/the/model/in/the/container \ --tensor-parallel-size 16 ``` -## Troubleshooting distributed deployments +## Optimizing network communication for tensor parallelism -To make tensor parallelism performant, ensure that communication between nodes is efficient, for example, by using high-speed network cards such as InfiniBand. To set up the cluster to use InfiniBand, append additional arguments like `--privileged -e NCCL_IB_HCA=mlx5` to the `run_cluster.sh` script. Contact your system administrator for more information about the required flags. One way to confirm if InfiniBand is working is to run `vllm` with the `NCCL_DEBUG=TRACE` environment variable set, for example `NCCL_DEBUG=TRACE vllm serve ...`, and check the logs for the NCCL version and the network used. If you find `[send] via NET/Socket` in the logs, NCCL uses a raw TCP socket, which is not efficient for cross-node tensor parallelism. If you find `[send] via NET/IB/GDRDMA` in the logs, NCCL uses InfiniBand with GPUDirect RDMA, which is efficient. +Efficient tensor parallelism requires fast inter-node communication, preferably through high-speed network adapters such as InfiniBand. +To set up the cluster to use InfiniBand, append additional arguments like `--privileged -e NCCL_IB_HCA=mlx5` to the + helper script. +Contact your system administrator for more information about the required flags. ## Enabling GPUDirect RDMA +GPUDirect RDMA (Remote Direct Memory Access) is an NVIDIA technology that allows network adapters to directly access GPU memory, bypassing the CPU and system memory. This direct access reduces latency and CPU overhead, which is beneficial for large data transfers between GPUs across nodes. + To enable GPUDirect RDMA with vLLM, configure the following settings: - `IPC_LOCK` security context: add the `IPC_LOCK` capability to the container's security context to lock memory pages and prevent swapping to disk. @@ -175,21 +180,17 @@ spec: ... ``` -Efficient tensor parallelism requires fast inter-node communication, preferably through high-speed network adapters such as InfiniBand. To enable InfiniBand, append flags such as `--privileged -e NCCL_IB_HCA=mlx5` to `run_cluster.sh`. For cluster-specific settings, consult your system administrator. +!!! tip "Confirm GPUDirect RDMA operation" + To confirm your InfiniBand card is using GPUDirect RDMA, run vLLM with detailed NCCL logs: `NCCL_DEBUG=TRACE vllm serve ...`. -To confirm InfiniBand operation, enable detailed NCCL logs: + Then look for the NCCL version and the network used. -```bash -NCCL_DEBUG=TRACE vllm serve ... -``` - -Search the logs for the transport method. Entries containing `[send] via NET/Socket` indicate raw TCP sockets, which perform poorly for cross-node tensor parallelism. Entries containing `[send] via NET/IB/GDRDMA` indicate InfiniBand with GPUDirect RDMA, which provides high performance. - -!!! tip "Verify inter-node GPU communication" - After you start the Ray cluster, verify GPU-to-GPU communication across nodes. Proper configuration can be non-trivial. For more information, see [troubleshooting script][troubleshooting-incorrect-hardware-driver]. If you need additional environment variables for communication configuration, append them to `run_cluster.sh`, for example `-e NCCL_SOCKET_IFNAME=eth0`. Setting environment variables during cluster creation is recommended because the variables propagate to all nodes. In contrast, setting environment variables in the shell affects only the local node. For more information, see . + - If you find `[send] via NET/IB/GDRDMA` in the logs, then NCCL is using InfiniBand with GPUDirect RDMA, which *is* efficient. + - If you find `[send] via NET/Socket` in the logs, NCCL used a raw TCP socket, which *is not* efficient for cross-node tensor parallelism. !!! tip "Pre-download Hugging Face models" If you use Hugging Face models, downloading the model before starting vLLM is recommended. Download the model on every node to the same path, or store the model on a distributed file system accessible by all nodes. Then pass the path to the model in place of the repository ID. Otherwise, supply a Hugging Face token by appending `-e HF_TOKEN=` to `run_cluster.sh`. -!!! tip - The error message `Error: No available node types can fulfill resource request` can appear even when the cluster has enough GPUs. The issue often occurs when nodes have multiple IP addresses and vLLM can't select the correct one. Ensure that vLLM and Ray use the same IP address by setting `VLLM_HOST_IP` in `run_cluster.sh` (with a different value on each node). Use `ray status` and `ray list nodes` to verify the chosen IP address. For more information, see . +## Troubleshooting distributed deployments + +For information about distributed debugging, see [Troubleshooting distributed deployments](distributed_troubleshooting.md). diff --git a/docs/serving/distributed_troubleshooting.md b/docs/serving/distributed_troubleshooting.md new file mode 100644 index 0000000000000..bd45f010ed2ae --- /dev/null +++ b/docs/serving/distributed_troubleshooting.md @@ -0,0 +1,16 @@ +# Troubleshooting distributed deployments + +For general troubleshooting, see [Troubleshooting](../usage/troubleshooting.md). + +## Verify inter-node GPU communication + +After you start the Ray cluster, verify GPU-to-GPU communication across nodes. Proper configuration can be non-trivial. For more information, see [troubleshooting script][troubleshooting-incorrect-hardware-driver]. If you need additional environment variables for communication configuration, append them to , for example `-e NCCL_SOCKET_IFNAME=eth0`. Setting environment variables during cluster creation is recommended because the variables propagate to all nodes. In contrast, setting environment variables in the shell affects only the local node. For more information, see . + +## No available node types can fulfill resource request + +The error message `Error: No available node types can fulfill resource request` can appear even when the cluster has enough GPUs. The issue often occurs when nodes have multiple IP addresses and vLLM can't select the correct one. Ensure that vLLM and Ray use the same IP address by setting `VLLM_HOST_IP` in (with a different value on each node). Use `ray status` and `ray list nodes` to verify the chosen IP address. For more information, see . + +## Ray observability + +Debugging a distributed system can be challenging due to the large scale and complexity. Ray provides a suite of tools to help monitor, debug, and optimize Ray applications and clusters. For more information about Ray observability, visit the [official Ray observability docs](https://docs.ray.io/en/latest/ray-observability/index.html). For more information about debugging Ray applications, visit the [Ray Debugging Guide](https://docs.ray.io/en/latest/ray-observability/user-guides/debug-apps/index.html). For information about troubleshooting Kubernetes clusters, see the +[official KubeRay troubleshooting guide](https://docs.ray.io/en/latest/serve/advanced-guides/multi-node-gpu-troubleshooting.html).