Red Hat Summit Red Hat SummitSupportConsoleDevelopersStart a trialContact

Chapter 2. Version 3.3.1 release notes

Red Hat AI Inference Server 3.3.1 is a maintenance release containing security fixes, bug fixes, and minor enhancements.

The following container images are available from registry.redhat.io:

  • registry.redhat.io/rhaiis/vllm-cuda-rhel9:3.3.1
  • registry.redhat.io/rhaiis/model-opt-cuda-rhel9:3.3.1
  • registry.redhat.io/rhaiis/vllm-rocm-rhel9:3.3.1

2.1. Security fixes
Copy link

Red Hat AI Inference Server 3.3.1 addresses the following CVEs:

2.2. Bug fixes
Copy link

Streaming tool calls with Mistral models returned invalid JSON
Streaming tool calls using Mistral models through the Anthropic-compatible /v1/messages endpoint returned invalid JSON, preventing clients from parsing responses. The JSON serialization for streaming tool calls is corrected. Mistral models using OpenAI-compatible endpoints and non-Mistral models using the /v1/messages endpoint were not affected.
Quantized Llama-4 models produced incorrect results
Quantization scales were not permuted correctly for attention layers in Llama-4 models, causing accuracy collapse in models such as Llama-Guard-4-12B when attention layers were quantized. Quantization scales are permuted correctly for Llama-4 attention layers.
Encoder models failed on AMD ROCm AI accelerators
The Triton attention backend did not support encoder self-attention, causing encoder-only models, encoder-decoder models, and Whisper speech-to-text models to fail on AMD ROCm AI accelerators. The Triton attention backend supports encoder self-attention.
GPT-OSS models returned empty content in multi-turn conversations
The reasoning parser incorrectly matched markers from previous messages in multi-turn conversations, causing GPT-OSS models to return content: null when using json_object response format. The reasoning parser correctly handles multi-turn conversations.
Malformed json_schema requests returned HTTP 500 instead of HTTP 400
When response_format type was json_schema but the json_schema field was missing, an assertion error caused the server to return HTTP 500 Internal Server Error instead of HTTP 400 Bad Request. The server validates the json_schema field before processing and returns the correct HTTP 400 Bad Request error.
Large Mixture of Experts models crashed due to integer overflow
An int32 overflow in fused MoE stride computation caused large Mixture of Experts models to crash or produce silent data corruption. The stride computation uses overflow-safe arithmetic.
Cascade attention caused numerical instability
Cascade attention was enabled by default and caused numerical instability in some workloads, resulting in unreliable model outputs. Cascade attention is disabled by default to match the upstream vLLM configuration.

2.3. Enhancements
Copy link

BART encoder-decoder model support for CUDA AI accelerators
The BART plugin enables inference serving for BART-based summarization and translation models on CUDA AI accelerators.
Llama-Nemotron embedding model support
The llama-nemotron-embed-1b-v2 embedding model is supported for inference serving.
Custom encoder support for classification models
Classification models can use custom encoders for improved inference performance.
Red Hat logoGithubredditYoutubeTwitter

Learn

Try, buy, & sell

Communities

About Red Hat

We deliver hardened solutions that make it easier for enterprises to work across platforms and environments, from the core datacenter to the network edge.

Making open source more inclusive

Red Hat is committed to replacing problematic language in our code, documentation, and web properties. For more details, see the Red Hat Blog.

About Red Hat Documentation

We help Red Hat users innovate and achieve their goals with our products and services with content they can trust. Explore our recent updates.

Legal Notice

Theme

© 2026 Red HatBack to top