b10307

ggml-org/llama.cppb10307Aug 7, 2026by github-actions[bot]same day as b10306

What changed since b10306

low risk

b10307 fixes SYCL NVFP4 UE4M3 scale-factor parsing, treating unsigned scales correctly instead of as signed.

Improvements

  • Correct scale decoding improves numerical accuracy for NVFP4-quantized models running on the SYCL backend.

Bug fixes

  • SYCL GPU code no longer decodes per-group NVFP4 scaling factors through the signed E4M3 path; the UE4M3 bytes are now handled as unsigned.
  • Prevents wrong dequantization of NVFP4 weights on Intel GPUs caused by misread scale bytes.

Upgrade notes

This is a drop-in patch with no config or API changes; SYCL users running NVFP4 models should upgrade to get correct results. The SYCL GLU performance work from the previous build is retained unchanged. Non-SYCL backends are unaffected.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

This release fixes a critical parsing bug in the SYCL backend where NVFP4 quantization format scaling factors were incorrectly being treated as signed values instead of unsigned values.

Key Highlights

  • Fixed SYCL parsing of NVFP4 UE4M3 quantization format (#25608)
  • Corrected handling of unsigned scaling factors

Release Notes (excerpt)

<details open> sycl: fix UE4M3 parsing (#25608) The NVFP4 quantization format stores a scaling factor for every group of 16 weights, packed into a single UE4M3 byte. The SYCL GPU code was converting these scale values using the E4M3 path, but that's *signed*, and these are unsigned values. </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories