b10322

ggml-org/llama.cppb10322Aug 7, 2026by github-actions[bot]same day as b10321

What changed since b10321

low risk

b10322 adds a SYCL-only optimization that coalesces SSM_CONV window loads, delivering up to ~1.85x speedup on Arc Pro B70.

Improvements

  • SYCL backend now coalesces SSM_CONV window loads, cutting per-run time roughly in half for large interleaved A/B shapes.
  • Memory access patterns were reworked for interleaved A/B tensor layouts in the SSM convolution path.
  • End-to-end prompt processing on Qwen35 27B improved about 2% in llama-bench, while token generation stayed flat.
  • Small or degenerate SSM_CONV shapes (n_t=1) show no regression, indicating the change is scoped to the windowed path.

Upgrade notes

This release is purely a performance change for the SYCL backend; CPU, CUDA, Metal, and other backends are unaffected. No configuration, model, or API changes are required, so a drop-in binary swap is sufficient. The prior Metal NORM/RMS_NORM correctness fix from b10321 remains in place.

Analyzed Oct 29, 2026

Upgrading?

Release notes summary

Optimizes the SYCL backend performance for SSM (State Space Model) convolution operations by coalescing window loads.

Key Highlights

  • Significant performance improvements on Arc Pro B70 hardware, achieving up to 1.85x speedup for SSM_CONV operations.
  • Optimized memory access patterns for interleaved A/B tensor layouts.
  • Benchmark results show consistent throughput gains on Qwen35 27B models.

Release Notes (excerpt)

<details open> sycl: coalesce the ssm_conv window loads (#26612) test-backend-ops perf -o SSM_CONV on an Arc Pro B70, interleaved A/B against master, 6 reps, us/run: ne_a=[515,3328,1,1] ne_b=[4,3328,1,1] n_t=512 97.68 -> 52.95 1.85x ne_a=[937,8192,1,1] ne_b=[4,8192,1,1] n_t=934 516.16 -> 276.13 1.87x ne_a=[4,3328,1,1] ne_b=[4,3328,1,1] n_t=1 2.73 -> 2.71 flat llama-bench on qwen35 27B Q4_K - Medium (48 of its 64 blocks run ssm_conv),…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories