b9973

ggml-org/llama.cppb9973Jul 12, 2026by github-actions[bot]same day as b9972

What changed since b9972

low risk

b9973 narrows DeepSeek V4 cache clearing to per-sequence scope instead of wiping the full context.

Improvements

  • DeepSeek V4 cache invalidation now targets only the affected sequence, avoiding unnecessary full-context resets and preserving other sequences' cached state.

Bug fixes

  • Corrected DeepSeek V4 cache clearing so it no longer discards the entire context when only one sequence needs invalidation.

Upgrade notes

This is a single targeted fix on top of b9972 with no API, config, or model-format changes; a drop-in binary swap is sufficient. Users running DeepSeek V4 with multiple concurrent sequences benefit most, since other sequences' caches are no longer flushed. No migration steps required.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Fixes the cache clearing logic for DeepSeek V4 to clear cache only for specific sequences instead of the entire context, improving efficiency.

Key Highlights

  • Optimized cache clearing for DeepSeek V4
  • Now clears cache per sequence rather than globally

Release Notes (excerpt)

<details open> DeepseekV4: clear cache only for seq rather than full (#25521) </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:** - Android arm64 (CPU) **Windows:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories