b10255

ggml-org/llama.cppb10255Aug 4, 2026by github-actions[bot]same day as b10254

What changed since b10254

low risk

b10255 extends the SYCL oneDNN SDPA fast path to quantized and FP32 KV caches, plus a stream-sync fix.

New features

  • SYCL oneDNN SDPA now accepts Q4_0, Q4_1, Q5_0, Q5_1, Q8_0 and FP32 KV caches by converting K/V to dense FP16 on-device before the fused kernel.
  • Non-F16 KV acceleration is gated to prefill workloads with K >= 1024 and Q >= 32, while F16 KV keeps working at any sequence length.

Improvements

  • Quantized and FP32 KV caches on SYCL can now use the fused oneDNN attention kernel instead of falling back to the generic path.
  • Stream synchronization is now unconditional via wait_and_throw(), removing a class of race conditions in SYCL execution.
  • K and V are always dequantized into separate buffers, eliminating the V_is_K_view aliasing assumption.
  • SYCL.md documentation was trimmed of the unshipped GGML_SYCL_FA_DEBUG reference.

Bug fixes

  • Fixed missing stream synchronization in the SYCL backend that could let work be observed before completion.

Upgrade notes

This is an additive SYCL backend change; no config migration is required and non-SYCL builds are unaffected. SYCL users with quantized or FP32 KV caches should see faster prefill once sequences reach the K >= 1024 / Q >= 32 threshold. BF16 and IQ KV types remain unsupported on this path and continue to use the existing fallback.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Extends SYCL oneDNN Self-Attention (SDPA) support to non-FP16 KV caches (Q4_0–Q8_0 and FP32).

Key Highlights

  • SYCL oneDNN SDPA for Q4_0-Q8_0 and FP32 KV caches
  • Stream synchronization fix
  • Requires K >= 1024 and Q >= 32 for non-F16

Release Notes (excerpt)

<details open> Extended SYCL oneDNN SDPA to non-FP16 KV caches (Q4_0–Q8_0 and FP32) (#25874) * sycl: extend oneDNN SDPA to Q4_0-Q8_0 and F32 KV caches Extends the oneDNN SDPA path (PR #25222) to handle non-F16 KV caches by dequantizing or converting K/V to dense FP16 on-device before feeding them into the SDPA graph. The fused systolic kernel then runs identically to the native FP16 path. Supported KV types: - Q4_0, Q4_1, Q5_0, Q5_1, Q8_0: to_fp16_sycl / to_fp16_nc_sycl…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories