Some inputs hang the whole embedding service on Qwen3-Embedding-0.6B

[lambdaupb]

System Info

On some inputs (example attached) for Qwen/Qwen3-Embedding-0.6B, the tei server embedding service stops responding.
The HTTP request with the borked input does not get a response and other requests will not be served.

In wireshark I can see that the request is sent, but there is nothing coming back from the server.

There is 0 CPU or GPU activity.

Information

• Docker
• The CLI directly

Tasks

• An officially supported command
• My own modifications

Reproduction

On both cuda and ipex cpu.

# docker run --rm --gpus all -p 8000:80 -v /opt/tei_data:/data --pull always ghcr.io/huggingface/text-embeddings-inference:cpu-ipex-latest --model-id Qwen/Qwen3-Embedding-0.6B 
cpu-ipex-latest: Pulling from huggingface/text-embeddings-inference
Digest: sha256:f1c8b121b9b0582b7ae3e4180f2eebcf968aa8c8de3a1965969b61870c6a48a5
Status: Image is up to date for ghcr.io/huggingface/text-embeddings-inference:cpu-ipex-latest

and

# docker run --rm --gpus all -p 8000:80 -v /opt/tei_data:/data --pull always ghcr.io/huggingface/text-embeddings-inference:1.8 --model-id Qwen/Qwen3-Embedding-0.6B
1.8: Pulling from huggingface/text-embeddings-inference
Digest: sha256:8aeb97215f29e0ed48647384af89661c36cee04120c2d4e86b5a3aead47611fa
Status: Image is up to date for ghcr.io/huggingface/text-embeddings-inference:1.8

curl -vvvv -H "Content-Type: application/json" -d @tei_qwen_embed_0.6b_broken_input.txt http://inference:8000/embed

tei_qwen_embed_0.6b_broken_input.txt

Expected behavior

The HTTP request should not hang and it also should not block other embed requests.

[alvarobartt]

Hey @lambdaupb thanks for reporting! I'll try to reproduce and come back to you, but it's certainly odd so I'll investigate 🤗