Update GGML to latest master

[Skyrion9]

• Many fixes & perf updates in general, we're 316 commits behind as of this PR.
• Seems to have some important updates for vulkan including fp16: 1 -> fp16: dot2 which should be more efficient.
  ggml_vulkan: 0 = AMD Radeon RX 6700 (RADV NAVI22) (radv) | uma: 0 | fp16: dot2 | bf16: 0 | warp size: 32 | shared memory: 65536 | int dot: 1 | matrix cores: none
• Makes the auto-patches obsolete (needs testing, I don't have CUDA hardware.)

https://github.com/ggml-org/ggml/tree/57ea0bc119d722d74594196cc5b494a34dd87be4

https://github.com/ggml-org/ggml

Summary by CodeRabbit

Summary by CodeRabbit

• Bug Fixes

  • Improved CUDA convolution handling so large workloads reliably process all output elements.
  • Updated GPU work distribution and launch sizing for more consistent results on large workloads.

• Build Changes

  • Local patches are no longer applied automatically by default during builds; this can be re-enabled when needed.

• Maintenance

  • Updated the bundled ggml component to a newer revision, including related CUDA improvements.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 3546a08f-5119-4579-b738-677f9ea3d70e

📥 Commits

Reviewing files that changed from the base of the PR and between 71955cf and f95fc3e.

📒 Files selected for processing (1)

• ggml

🚧 Files skipped from review as they are similar to previous changes (1)

• ggml

---

📝 Walkthrough

Walkthrough

The PR disables automatic local-patch application by default, updates the ggml revision, and changes CUDA depthwise convolution to use grid-stride processing with capped launch blocks.

Changes

CUDA and build updates

Layer / File(s)	Summary
CUDA convolution patch update ggml, patches/ggml-conv2d-dw-grid-fix.patch	The ggml reference advances. The depthwise CUDA convolution patch adds grid-stride processing, includes <algorithm>, and caps the block count at 65535.
Local patch application default CMakeLists.txt	S2_AUTO_APPLY_LOCAL_PATCHES now defaults to OFF.

Estimated code review effort: 3 (Moderate) | ~20 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the primary change: updating the GGML submodule to a newer master revision.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}