> For the complete documentation index, see [llms.txt](https://docs.apolo.us/index/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.apolo.us/index/apolo-console/apps/installable-apps/available-apps/llm-inference/vllm-inference-details.md).

# vLLM Inference details

### Under the Hood

When you specify a multi-GPU preset (like a preset with multiple Nvidia and AMD GPUs), **LLM Inference**:

1. **Determines GPU Provider & Count**
   * AMD MI210 → `gpuProvider=amd`, 2 GPUs.
   * NVIDIA → `gpuProvider=nvidia`, 2 GPUs.
2. **Sets GPU Visibility**
   * On AMD: `HIP_VISIBLE_DEVICES=0,1`, `ROCR_VISIBLE_DEVICES=0,1` (if 2 GPUs).
   * On NVIDIA: `CUDA_VISIBLE_DEVICES=0,1`.
3. **Applies a Default Parallel** arguments (e.g. `--tensor-parallel-size=2`) if the user hasn’t already done so.

#### Environment Variables for AMD

By default, if you select a preset with AMD GPU cards, the chart’s logic sets:

* **`HIP_VISIBLE_DEVICES=0,1...`** and **`ROCR_VISIBLE_DEVICES=0,1...`**` `` ``depending on number of available GPUs: ` Tells ROCm which GPUs are accessible.
* **`TORCH_USE_HIP_DSA=1`**: Enables direct storage access for HIP.
* **`HSA_FORCE_FINE_GRAIN_PCIE=1`** & **`HSA_ENABLE_SDMA=1`**: Improves GPU ↔ Host & GPU ↔ GPU memory transfers.
* **`ROCM_DISABLE_CU_MASK=0`**: All compute units remain active.
* **`VLLM_WORKER_MULTIPROC_METHOD=spawn`**: Avoids “fork” issues on AMD.
* **`NCCL_P2P_DISABLE=0`**: By default, we assume your cluster has correct kernel parameters for GPU–GPU direct memory access. If not, you can pass `--set envAmd.NCCL_P2P_DISABLE=1` to forcibly disable P2P.

### Final Notes

* **NCCL / RCCL** logs will appear in the vLLM container logs. Look for lines referencing peer-to-peer if you see a hang.
* On AMD, if you do have persistent hangs, append `--set "envAmd.NCCL_P2P_DISABLE=1"` to your Apolo command to force fallback GPU communication.
* For the best performance, we keep ROCm version 6.2+ or 6.3+ in sync with your Docker image (`rocm/vllm-ci`)

By combining the right Apolo preset, environment variables, you can reliably run **vLLM** on multiple GPUs—be they AMD or NVIDIA —and get high token throughput for large LLMs.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.apolo.us/index/apolo-console/apps/installable-apps/available-apps/llm-inference/vllm-inference-details.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.