Zest v0.2.0

Zest accelerates long prompts for many GGUF models through Lemonade's
llama.cpp backend, using an AMD Ryzen AI NPU. The GPU handles attention and
writes the reply. No ONNX or Q4NX export or per-model NPU kernel compilation
is required. Coverage follows the bundled build and supported operations;
speedups vary by model.

This is an independent community add-on, tested with Lemonade 2026.43.3
and the bundled b10944 backend.

Windows and Linux packages include both zest and npu. Either command runs
the same launcher with the same options.

Use it with Lemonade on Windows

1. Download the Windows ZIP below and extract it. Keep the files together.
2. Quit your existing Lemonade tray server and double-click Start-Lemonade.cmd.
   Choose Lemonade's server executable if asked, then your local GGUF model.
3. In Lemonade App, set Settings > Base URL to http://127.0.0.1:13315
   and save. Select the imported model and send a long prompt.

The launcher keeps its configuration separate and uses the GPU if the NPU
can't start. It doesn't include Lemonade, model weights, or AMD drivers.
The ZIP has an internal checksum manifest; its .sha256 file is attached.

Website
· Lemonade setup and troubleshooting
· Test report

Before you use it

We tested Windows 11 on a Ryzen AI 9 HX PRO 370 with Radeon 890M graphics.
Other Ryzen AI hardware, including Strix Halo, hasn't been verified by this
project. Linux builds and host tests pass; Linux NPU hardware and Lemonade
integration remain unverified.

The NPU keeps extra converted weight copies in RAM and changes numerical
results through block rounding. Small prompt chunks stay on the GPU.
Speed depends on the model. If the NPU fails during a request, failed work
is retried on the GPU and later requests use the GPU. The CPU reference
remains available if the GPU cannot run a failed piece. This build includes
the GPU recovery fix already merged into main.

Direct llama.cpp use

On Windows, run npu llama-server -m model.gguf from the extracted ZIP.
On Linux, unpack the Ubuntu archive, enter its directory, and run
./npu llama-server -m model.gguf. You can also use zest in either command.
See the technical guide
for commands, build instructions, accuracy checks, and standalone benchmarks.

Zest v0.2.0-rc2

For current Windows and Linux downloads, use Zest v0.2.0.

This earlier prerelease is retained with its original packages and verification records.

Zest v0.2.0-rc1

For current Windows and Linux downloads, use Zest v0.2.0.

This earlier prerelease is retained with its original packages and verification records.

Zest for Lemonade (Windows preview r2)

The project is now Zest. Get the current Zest preview. This historical release retains its original archive filenames and checksums.

ggml-xdna accelerates long prompts for many GGUF models through Lemonade's
llama.cpp backend, using an AMD Ryzen AI NPU. The GPU handles attention and
writes the reply. No ONNX or Q4NX export or per-model NPU kernel compilation
is required. Coverage follows the bundled build and supported operations;
speedups vary by model.

This is an independent community add-on, tested with Lemonade 2026.43.3
and the bundled b10944 backend.

Use it with Lemonade on Windows

1. Download the Windows ZIP below and extract it. Keep the files together.
2. Quit your existing Lemonade tray server and double-click Start-Lemonade.cmd.
   Choose Lemonade's server executable if asked, then your local GGUF model.
3. In Lemonade App, set Settings > Base URL to http://127.0.0.1:13315
   and save. Select the imported model and send a long prompt.

The launcher keeps its configuration separate and uses the GPU if the NPU
can't start. It doesn't include Lemonade, model weights, or AMD drivers.
The ZIP has an internal checksum manifest; its .sha256 file is attached.

Website and walkthrough
? Lemonade setup and troubleshooting
? Test report

Before you use it

We tested Windows 11 on a Ryzen AI 9 HX PRO 370 with Radeon 890M graphics.
Other Ryzen AI hardware, including Strix Halo, hasn't been verified by this
project. Linux builds and host tests pass; Linux NPU hardware and Lemonade
integration remain unverified.

The NPU keeps extra converted weight copies in RAM and changes numerical
results through block rounding. Small prompt chunks stay on the GPU.
Speed depends on the model. If the NPU fails during a request, failed work
is retried on the GPU and later requests use the GPU. The CPU reference
remains available if the GPU cannot run a failed piece. This build includes
the GPU recovery fix already merged into main.

The desktop walkthrough uses screenshots from the actual app. It shows
prompt entry and the answer, but isn't a real-time latency recording.

Direct llama.cpp use

Still supported on Windows and Linux. The npu launcher is included in
both packages. See the technical guide
for commands, build instructions, accuracy checks, and standalone benchmarks.

Checks for this preview

• Fresh build with main's GPU recovery fix: all 12 applicable host/NPU tests passed.
• Final ZIP extracted to a new path with spaces: all 96 manifest files verified.
• Packaged launcher and saved settings exercised through the official Lemonade server.
• Normal NPU chat: three correct answers. Injected failure: three correct answers,
  with first content in 4.28 seconds on the affected request. The arithmetic
  check returned an equation rather than the requested integer in both checks.
• Windows and Linux CI build/test and packaging jobs passed.
• Website tested from 320 to 1536 pixels, including keyboard controls,
  clipboard, video, and reduced motion. Accessibility scan: no violations.

The interactive file-picker selection itself wasn't automated.

ZIP SHA-256: 36c0187a883cd48cd26f82b47b95f443e528f20638121ab7809724af1342689e

The backend was rebuilt from engine source at dd628e9. Its SHA-256 is
9ab00b2a12592b981410218e26e7211d0095073ce2c8de263f270fca0e62de0c.
The attached verification receipt records the final package and source commit.

Zest for Lemonade (2026-10-21 preview)

The project is now Zest. Get the current Zest preview. This historical release retains its original archive filenames and checksums.

Windows preview of ggml-xdna served through Lemonade's GGUF/llama.cpp path.

Unpack the ZIP and follow docs/lemonade.md. tools/start-lemonade.ps1 accepts the official Lemonade server executable and an existing local GGUF. It uses its own configuration directory, probes the NPU, and selects the GPU when the NPU is unavailable. This preview requires Lemonade 2026.43.3 and the bundled llama.cpp b10944; it does not replace Lemonade's default backend or require source patches.

Verified on Windows with Ryzen AI 9 HX PRO 370 / Radeon 890M:

• Clean source build: 11 applicable host/NPU tests passed.
• Every file in the unpacked package verified against its checksum manifest.
• Actual Lemonade desktop chat: full long-prompt prefill with NPU calls visible in Lemonade's log panel, successful answers, and cached follow-up behavior.
• Model comparison: mean KL divergence 0.004834, same top token 97.745% on the documented fixture.
• Startup-unavailable and runtime-failure checks returned correct answers through Lemonade. Both GPU and NPU returned the same arithmetic answer with an integer-only formatting violation; the strict failures are retained.

Known limits: a runtime NPU failure can leave the affected prompt finishing slowly on the CPU (about 63 seconds in this test); later prompts use the GPU. Model coverage, extra weight-copy memory, block rounding and model-dependent speed remain as documented. Linux NPU hardware and Strix Halo are unverified. The earlier three-pair 1.267x result is one warm workload, not a general speed claim or a comparison against Lemonade's latest GPU defaults.

The attached walkthrough is assembled from actual desktop screenshots showing prompt entry and the returned answer; it is not a real-time latency recording. The ZIP includes the complete validation records and source instructions. No model weights, AMD drivers or Lemonade installers are included.

Source and review: #23

ZIP SHA-256: 84e724311a5b3e81ed4cd5bd9942819609a0cf7f5332b65cce6b7b67e5aebd04

CI: Windows and Linux build/test and packaging checks passed on source commit 4fd9233. See https://github.com/Cyronius/zest/pull/23/checks . Linux hardware NPU testing remains unverified.