Chapter 2. New features and enhancements

New versions of vLLM and LLM Compressor are included in this release:
- vLLM v0.9.0.1
  - 900+ upstream commits since vLLM v0.8.4. New features include: FP8 fused Mixture of Experts (MoE) kernels, 14 new models supported, /server_info endpoint, dynamic LoRA hot reload.
- LLM Compressor v0.5.1
The Red Hat AI Inference Server container base is now built on PyTorch 2.7 and Triton 3.2.
Red Hat AI Inference Server is now fully supported on FIPS-compliant Red Hat Enterprise Linux (RHEL) hosts.
The Red Hat AI Inference Server supported product and hardware configurations have been expanded. For more information, see Supported product and hardware configurations.

Expand

Table 2.1. AI accelerator performance highlights
Feature	Benefit	Supported GPUs
Blackwell support	Runs on NVIDIA B200 compute capability 10.0 GPUs with FP8 kernels and full CUDA Graph acceleration	NVIDIA Blackwell
FP8 KV-cache on ROCm	Roughly twice as large context windows with no accuracy loss	All AMD GPUs
Skinny GEMMs	Roughly 10% lower inference latency	AMD MI300X
Full CUDA Graph mode	6–8% improved average Time Per Output Token (TPOT) for small models.	NVIDIA A100 and H100
Auto FP16 fallback	Stable runs on pre-Ampere cards without manual flags, for example, NVIDIA T4 GPUs	Older NVIDIA GPUs

2.1. New models enabled
링크 복사

Red Hat AI Inference Server 3.1 expands capabilities by enabling the following models:

/server_info REST endpoint: Query model, KV cache, and device settings for observability and automation.
Dynamic LoRA hot reload: Swap fine-tuned adapters from a URL with zero downtime.
vllm-bench CLI: "Ship-in-container" tool for instant latency and throughput sizing.
Faster incremental detokenization: Streaming responses start twice as fast on CUDA and ROCm GPUs.
torch.compile caching: Cached first prompt compilation shortens warm-up times across host restarts.

Lower total cost of ownership (TCO): FP8/INT8 kernels and skinny GEMMs allow the same GPUs serve more tokens per second.
Larger models on AMD GPUs: ROCm now matches CUDA for FP8 and fused MoE model performance, making AMD MI300X a first-class deployment target.
Operational agility: LoRA hot swap and /server_info endpoints allow for continuous integration and deployment for model fine-tuning without pod restarts.