docs/lemonade.md: the embedding checklist (measured)

[bong-water-water-bong]

The gap audit for embedding the engine fully in Lemonade. Lemonade's LLM suite ran through onebit on the pinned fork (7650b4f), with an audit copy that claims every feature so no test is skipped:

• Vulkan and HRX: 19 of 31 pass.
• NPU: 3 of 31. Lemonade hands the engine a GGUF, which the NPU route can't run.

Six gaps are listed, each marked with where its fix goes (engine or fork). Fixes follow as separate PRs.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit f1ade64

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Unmeasured Claim The PR claims that the gaps were measured on Strix Halo on 2026-09-25, but does not provide specific performance metrics or test results to back up the claim. The document states that the run used a pinned fork and the engine from main, but does not include concrete data such as throughput, latency, or accuracy numbers that would validate the measurements. ## Embedding checklist What "fully embedded" still needs, measured on Strix Halo on 2026-09-25. The run used the pinned fork (`third_party/lemonade`, `7650b4f`), the engine from `main`, and Lemonade's LLM suite (`test/server_llm.py --wrapped-server onebit --backend <device>`). An audit copy of the suite claimed every feature for `onebit`, so each test ran instead of being skipped. | device | pass | fail | |---|---|---| | Vulkan | 19 / 31 | Responses API (2), embeddings (3), reranking (3), slots, tokenize, echo, generation parameters | | HRX | 19 / 31 | the same 12 | | NPU | 3 / 31 | everything that loads a model: Lemonade hands the engine a GGUF, and the NPU runs only 1bit NPU model directories | 1. **Declare what already works.** Tool calls (plain and streaming), `stop`, and async chat/completions pass on Vulkan and HRX, but the recipe declares them unsupported. (fork) 2. **Test on the device asked for.** Lemonade's test harness passes `--backend` to llamacpp, sd-cpp and the others, but not to `onebit`, so every `onebit` run used the default device. (fork) 3. **Forward the rest of llama-server's API.** `1bit serve` forwards only chat and completions, so `/v1/responses`, `/slots` and `/tokenize` return 404. (engine) 4. **Embedding and reranking models.** The recipe declares only `chat`, so Lemonade refuses embedding and reranking models. `1bit serve` serves them only as companions of a chat model (`--embed`, `--rerank`). (engine + fork) 5. **The NPU from Lemonade.** Lemonade needs to download and hand over an NPU model directory, and `1bit serve --device npu` needs to take it. (engine + fork) 6. **Beyond the LLM suite:** - image generation: `1bit comfy` behind Lemonade's image endpoint; - the Laya router; - MLX; - CUDA through ZINC, which needs an NVIDIA box. `echo` and generation parameters also fail, but Lemonade's own llamacpp recipe declares neither, so they aren't gaps.