DRAFT: ggml: optimize Flash Next inference paths

[gaetan-puleo]

Overview

Measurements

Device:     Ryzen AI Max+ 395 / <mini-PC or laptop model>
Memory:     128 GB LPDDR5X-8000
Power:      <sustained TDP / platform profile / power governor>
BIOS:       UMA split <n> GB
Kernel:     <uname -r>, amdgpu params: <gttsize / ttm.pages_limit, or none>
Backend:    Vulkan RADV, Mesa <version>   (or: ROCm <version>, HIP_LAUNCH_BLOCKING=<0|1>)
Build:      <cmake flags>
Baseline:   <merge-base commit sha, built and run in this same session>
Change:     <your branch commit sha>
Model:      <HF repo / file>, <quant>

Baseline:

After:

Correctness:

Additional information

Requirements

• I have read and agree with the contributing guidelines
• This change is Strix Halo specific, or justified by measurements on Strix Halo. General llama.cpp improvements
  belong in halo-box/llama.cpp instead
• AI usage disclosure:
• What was NOT verified:

[dzannotti]

Validation: static whitespace checks pass. I did not run the implementation because the PR is draft, based on an old revision, is marked dirty/merge-conflicted, and currently has a failed Windows check. Those are prerequisites to meaningful validation.

Scope: the work is broad generic llama.cpp functionality, not a Strix-only patch. Rebase onto current upstream/staging, resolve the merge state, split independent concerns, and provide CPU/CUDA/Vulkan correctness coverage before requesting a new review.

comment generated by my clanker Codex