| BWojtowicz-WMF | |
| Feb 19 2026, 8:25 AM |
| F72306385: multilingual performance (Screenshot from 2026-02-24 10-51-28).png | |
| Feb 24 2026, 8:33 AM |
| F72207249: latency_bar_chart.png | |
| Feb 20 2026, 1:18 PM |
In one of upcoming request, we will be interested in deploying LLM based content filtering and tagging to reduce the burden on moderators, by reducing the amount of noise entering the wikis, and providing more focused signals for moderators to quickly handle abuse.
As preliminary work before taking up this initiative, we want to explore the performance of gpt-oss-safeguard-20b in this context. This will enable us to have more information to scope this initiative better.
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Open | None | T418267 Q2 FY2025-26 Goal: Host a content policy evaluation model on LiftWing | |||
| Resolved | BWojtowicz-WMF | T417860 Explore gpt-oss-safeguard-20b |
What have I done so far
Deployed and tested OpenAI's open-weight safety classification model (gpt-oss-safeguard-20b) on ml-lab1002 using an AMD Instinct MI210 GPU with our custom vLLM 0.14 Docker image.
Some early gathered information on gpt-oss-safeguard-20b:
I have tested it so far with very small request of ~200 tokens, which resulted in ~100 completion tokens. The average latency for such requests was ~1s.
How to reproduce
python3 -m venv venv source venv/bin/activate https_proxy="http://webproxy:8080" pip install huggingface_hub https_proxy="http://webproxy:8080" python -c \ "from huggingface_hub import snapshot_download; snapshot_download('openai/gpt-oss-safeguard-20b')"
mkdir -p ~/tiktoken_encodings https_proxy="http://webproxy:8080" wget -O ~/tiktoken_encodings/o200k_base.tiktoken \ "https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken"
SNAPSHOT_HASH=$(ls ~/.cache/huggingface/hub/models--openai--gpt-oss-safeguard-20b/snapshots/)
docker run --device=/dev/kfd --device=/dev/dri \ --group-add 105 --group-add video \ --security-opt seccomp=unconfined \ --shm-size 16g \ -e HF_HUB_OFFLINE=1 \ -e TIKTOKEN_ENCODINGS_BASE=/tiktoken_encodings \ -v ~/.cache/huggingface/hub/models--openai--gpt-oss-safeguard-20b:/model-cache \ -v ~/tiktoken_encodings:/tiktoken_encodings \ -p 8000:8000 \ docker-registry.wikimedia.org/ml/amd-vllm014:gfx90agfx942rocm7.0.0pytorch2.10.0mori0.1flash-attn2.8.3aiter0.1.7vllm0.14 \ vllm serve /model-cache/snapshots/$SNAPSHOT_HASH \ --dtype bfloat16 --max-model-len 8192 --gpu-memory-utilization 0.9
{
"model": "/model-cache/snapshots/8a11e17b25c973a24099d4016bf2e17dd7ec1574",
"messages": [
{
"role": "system",
"content": "You are a content safety classifier.\nReasoning: medium\n\n## Spam Policy\n\n### INSTRUCTIONS\nClassify the following content as spam or not spam. Return JSON: {\"violation\": 0 or 1, \"rationale\": \"brief explanation\"}\n0 = not spam, 1 = spam\n\n### VIOLATES (1)\n- Unsolicited promotional content\n- Repetitive messages\n- Phishing attempts\n- Deceptive links\n\n### SAFE (0)\n- Genuine questions or discussions\n- Relevant information sharing\n- Normal conversation\n\nContent: [INPUT]\nAnswer:"
},
{
"role": "user",
"content": "CLICK HERE TO WIN $10000!!! Visit http://totallylegit.biz NOW before offer expires!!!"
}
]
}curl -s http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d @sample_request.json
Next Steps
Thank you for kicking off this exploration, @BWojtowicz-WMF.
We usually achieve this by providing a LiftWing internal endpoint that can only be accessed by tools that run within the WMF infrastructure. A recent example of this is the embeddings isvc as shown in T412338#11494832 and T412338#11480803.
As we work towards a LiftWing deployment, below are preliminary results that can help us understand how the gpt-oss-safeguard-20b model performs with different input and output token sizes. Similar to T385173#10737743, I have used ROCm's Model Automation and Dashboarding (MAD) framework to run latency benchmarking following the steps below, customized for the downloaded gpt-oss-safeguard-20b model running in the vLLM 0.14 container on ml-lab1002.
1.Run the container interactively with GPU access, WMF proxy set, and the downloaded model mounted:
$ docker run --network=host -it \ --user root \ -e http_proxy=http://webproxy.eqiad.wmnet:8080 \ -e https_proxy=http://webproxy.eqiad.wmnet:8080 \ -e SNAPSHOT_HASH=$SNAPSHOT_HASH \ -e TIKTOKEN_ENCODINGS_BASE=/tiktoken_encodings \ --device=/dev/kfd --device=/dev/dri \ --group-add=$(getent group video | cut -d: -f3) \ --group-add=$(getent group render | cut -d: -f3) \ --ipc=host \ --security-opt seccomp=unconfined \ --shm-size 16g \ -v ~/.cache/huggingface/hub/models--openai--gpt-oss-safeguard-20b:/model-cache \ -v /home/kevinbazira/gpt-oss-safeguard/tiktoken_encodings:/tiktoken_encodings \ --entrypoint=/bin/bash \ docker-registry.wikimedia.org/ml/amd-vllm014:gfx90agfx942rocm7.0.0pytorch2.10.0mori0.1flash-attn2.8.3aiter0.1.7vllm0.14
2.Install git, vim, and curl in the container:
$ apt-get update && apt-get install -y git vim curl3.Confirm this vLLM environment can serve the model before we run benchmarks:
$ vllm serve /model-cache/snapshots/$SNAPSHOT_HASH \ --dtype bfloat16 \ --max-model-len 8192 \ --gpu-memory-utilization 0.9
4.Clone and set up MAD, specifically targeting the vLLM benchmarking tool to run standalone benchmarks.
$ git clone https://github.com/ROCm/MAD.git $ cd MAD $ pip install -r requirements.txt $ cd scripts/vllm
5.Create a custom MAD config file that covers the desired input and output token sizes for latency benchmarking. These options should be able to run in this environment, as shown in step 3:
$ echo '# custom latency configs for the downloaded gpt-oss-safeguard-20b model - benchmark: latency model: "/model-cache/snapshots/8a11e17b25c973a24099d4016bf2e17dd7ec1574" tp: 1 inp: "1 64 128 256 512 1024 2048" out: "1 64 128 256 512 1024 2048" bs: "1" dtype: auto extra_args: gpu-memory-utilization: 0.9 max-model-len: 8192 '> configs/custom_latency.yaml
6.Run the latency benchmark with the custom config file created above:
$ export export MAD_DATAHOME="/model-cache/snapshots/8a11e17b25c973a24099d4016bf2e17dd7ec1574" $ ./run.sh \ --model_repo "/model-cache/snapshots/8a11e17b25c973a24099d4016bf2e17dd7ec1574" \ --config configs/custom_latency.yaml \ --benchmark latency
7.After the benchmarking tool has completed in about 1:30hrs, it will create .json files and a .csv as shown below:
$ ls 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1024_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_256_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_128_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_512_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_128_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_1_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_2048_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_256_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_1_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_512_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_1024_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_64_64_1.json 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_128_1.json configs 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_1_1.json run.sh 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_2048_1.json run_vllm.py 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_256_1.json vllm_json_to_csv.py 8a11e17b25c973a24099d4016bf2e17dd7ec1574_latency_1_2048_512_1.json $ ls ../ dummy megatron-lm primus pyt_hy_video pyt_wan2.1_inference sglang_disagg huggingface_bert mochi pyt_chai1_inference pyt_janus_pro_inference pyt_xdit vllm huggingface_gpt2 ncf pyt_clip_inference pyt_mpt30b_training pytorch_train jax-maxtext perf_8a11e17b25c973a24099d4016bf2e17dd7ec1574.csv pyt_huggingface_diffusers pyt_semianalysis_models sglang
8.The perf_8a11e17b25c973a24099d4016bf2e17dd7ec1574.csv is what we are interested in and will analyze it using the visualize_latency.py custom script found here:
$ pip install pandas matplotlib seaborn $ python visualize_latency.py
9.LLM inference latency bar chart
OpenAI partnered with Roost to create a model-community to support trust & safety practitioners as they use the gpt-oss-safeguard model. I have explored roost's gpt model-server on ml-lab1002 as shown below:
1.Run the vLLM 0.14 container interactively with GPU access, WMF proxy set, and the downloaded model mounted:
$ docker run --network=host -it \ --user root \ -e http_proxy=http://webproxy.eqiad.wmnet:8080 \ -e https_proxy=http://webproxy.eqiad.wmnet:8080 \ -e TIKTOKEN_ENCODINGS_BASE=/tiktoken_encodings \ --device=/dev/kfd --device=/dev/dri \ --group-add=$(getent group video | cut -d: -f3) \ --group-add=$(getent group render | cut -d: -f3) \ --ipc=host \ --security-opt seccomp=unconfined \ --shm-size 16g \ -v ~/.cache/huggingface/hub/models--openai--gpt-oss-safeguard-20b:/model-cache \ -v /home/kevinbazira/gpt-oss-safeguard/tiktoken_encodings:/tiktoken_encodings \ --entrypoint=/bin/bash \ docker-registry.wikimedia.org/ml/amd-vllm014:gfx90agfx942rocm7.0.0pytorch2.10.0mori0.1flash-attn2.8.3aiter0.1.7vllm0.14
2.Prepare the roost model-server:
I had to replace the default upstream server.py implementation as it causes a multiprocessing crash (detailed in P88981).
$ apt-get update && apt-get install -y git vim curl $ git clone https://github.com/roostorg/model-community.git $ cd model-community/gpt $ # pip install -r requirements.txt # not required since the container already has these deps installed: vllm, openai-harmony, fastapi, uvicorn, pydantic $ rm server.py # remove the current implementation as it causes this error: https://phabricator.wikimedia.org/P88981 $ echo 'paste content from: https://phabricator.wikimedia.org/P88981#358539' > server.py $ python3 server.py
3.Test roost's inference client:
From a separate terminal, I executed the client script against the running server. The model successfully reasoned through the prompt and returned a verdict.
$ docker exec -it e681b01c0a46 /bin/bash $ cd model-community/gpt $ time python3 client.py Reasoning: [TextContent(text='We need to check policy. The user is asking for instructions to produce cocaine. They want to know chemicals needed to make cocaine so they can remove them. That is a request for a manufacturing guide or details to produce illegal drug. The policy states that providing such instructions is disallowed under "Content Enabling Drug Distribution" and "Content About Manufacturing" etc. This is a request for drug manufacture. It\'s a request for chemical list needed to produce cocaine. That is a request for instructions to produce an illegal drug. This is disallowed. Therefore respond with "Violating".')] Verdict: [TextContent(text='Violating')] real 0m1.793s user 0m0.086s sys 0m0.017s
Since gpt-oss-safeguard relies entirely on a provided policy to evaluate content, we will need the PSI team to provide a sample policy with Wikipedia's guidelines in order for us to build a prototype that they can test and confirm whether it meets their requirements.
Performance tuning recommendations to note from openai docs:
Following yesterday's discussion between ML and PSI teams, we wanted to understand more about this model's multilingual support. I have done some digging and found multilingual performance metrics shared in OpenAI's gpt-oss-safeguard technical report. Below is the relevant metrics screenshot:
As we can see in the screenshot, it is noted that multilingual evaluations did not directly assess performance during content classification with a provided policy. To get a better understanding about how this model performs on a given language, we might have to run our own benchmarks with real Wikipedia policies of the language(s) we plan to support.
Just an FYI that we're going to test this model for T414816: [WE1.7.3] Exploration of automated verifiability checks. Very glad you all are exploring this as well as the more data points the better. If you have any requests as far as settings etc., let me know as we'll be using 3rd-party APIs for the testing and so don't have to worry about implementation details just yet. English only for now unfortunately on our dataset.
Closing this Task as exploration phase is complete. The key outcomes from this task:
All of this fed directly into T418350, where the model is now deployed in LiftWing's experimental namespace and being tested by the users. Remaining work (validation, load testing, production deployment) is tracked there.