Skip to content

Update gpt-oss mi300,mi325,mi355 for vllm 0.17.0 #268

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Rohan138 wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −7
Open

Update gpt-oss mi300,mi325,mi355 for vllm 0.17.0 #268

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 8 additions & 7 deletions OpenAI/GPT-OSS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,16 +69,15 @@ You can launch GPT-OSS model serving with vLLM using:
```bash
vllm serve openai/gpt-oss-120b
```
However, for optimal performance, applying the configuration below can deliver additional speedups and efficiency gains. These configurations were validated on the [vLLM 0.14.1 release](https://github.com/vllm-project/vllm/releases/tag/v0.14.1).
However, for optimal performance, applying the configuration below can deliver additional speedups and efficiency gains. These configurations were validated on the [vLLM 0.17.0 release](https://github.com/vllm-project/vllm/releases/tag/v0.17.0).

```bash
export HSA_NO_SCRATCH_RECLAIM=1
export AMDGCN_USE_BUFFER_OPS=0
export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
export VLLM_ROCM_USE_AITER_MHA=0
export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4

vllm serve openai/gpt-oss-120b --tensor-parallel-size=8 --gpu-memory-utilization 0.95 --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' --block-size=64 --disable-log-request
vllm serve openai/gpt-oss-120b --tensor-parallel-size=8 --attention-backend ROCM_AITER_UNIFIED_ATTN -cc.pass_config.fuse_rope_kvcache=True -cc.use_inductor_graph_partition=True --gpu-memory-utilization 0.95 --block-size=64
```
* `export HSA_NO_SCRATCH_RECLAIM=1` is only needed on the server with old GPU firmware. If the GPU firmware version is less than 177 by the following command, you need to set `export HSA_NO_SCRATCH_RECLAIM=1` for better performance.
```bash
Expand All @@ -92,13 +91,15 @@ rocm-smi --showfw | grep MEC | head -n 1 | awk '{print $NF}'

```bash
export HSA_NO_SCRATCH_RECLAIM=1
export AMDGCN_USE_BUFFER_OPS=0
export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
export VLLM_ROCM_USE_AITER_MHA=0
export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4

vllm serve openai/gpt-oss-120b --tensor-parallel-size=8 --gpu-memory-utilization 0.95 --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' --block-size=64 --disable-log-request --async-scheduling
vllm serve amd/gpt-oss-120b-w-mxfp4-a-fp8 --tensor-parallel-size=8 --attention-backend ROCM_AITER_UNIFIED_ATTN -cc.pass_config.fuse_rope_kvcache=True -cc.use_inductor_graph_partition=True --gpu-memory-utilization 0.95 --block-size=64
```

* [`amd/gpt-oss-120b-w-mxfp4-a-fp8`](https://huggingface.co/amd/gpt-oss-120b-w-mxfp4-a-fp8) is a Quark-quantized version of openai/gpt-oss-120b with support for fp8-quantized activations with static scales.

#### Known Issues
- When you encounter this error `The link interface of target "torch::nvtoolsext" contains: CUDA::nvToolsExt but the target was not found.` Please double check your pytorch version has suffix `+cu128`.
- If the output you see is garbage, that might be because you haven't properly set `CUDA_HOME`. The CUDA version needs to be greater than or equal to 12.8 and must be the same for installation and serving.
Expand Down