Intel Releases LLM-Scaler-vLLM PV 1.5, Upgrading vLLM to 0.26 and Expanding Model Support
Intel updates its AI software stack for Arc GPUs, jumping vLLM versions and adding support for Qwen and Gemma models with FP8 quantization.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
Intel software engineers released LLM-Scaler-vLLM PV 1.5 on October 16, an updated containerized Docker stack designed to provide a convenient AI/LLM experience on Intel Arc Graphics hardware. The tool originated from the Project Battlematrix initiative.
The most significant change is the version jump in underlying frameworks: vLLM has been upgraded from 0.14 to 0.26, and PyTorch from 2.10 to 2.12. While still behind the upstream community's latest release (vLLM 0.31), this update brings Intel's stack closer to current mainstream standards.
In terms of model compatibility, the new version adds support for Qwen3.6-27B, Qwen3.8-27B, Qwen3.6-35B-A3B, and the Gemma-4 series. It also now supports online FP8 and symmetric INT4 quantization, as well as FP8 KV cache. Intel states these improvements lead to faster time-to-first-token and optimized LoRA serving performance, though these figures are vendor-reported.