[bong-water-water-bong]

Docs catch-up after #196-#212.

• README: a Direction note: the engine moves to HRX (Loom kernels) plus the NPU, as geramyL (AMD) proposed, and Vulkan stays the default until the gates in RFC #213 are met. The status lines are updated too: six GGUF architectures answer on the NPU (correct, 0.006-0.17 tok/s), DwarfStar reads 1BP packages, and Laya now classifies conversations with its scorer on HRX at 15-16 ms. The old text said it picked a device directly at 0.38 s.
• docs/PORTING.md: the NPU row gains GGUF on the NPU (npu: GGUF-on-NPU serving route (Qwen2.5-7B, MiniCPM4-8B, MiniCPM5-1B answer ' Paris' on the NPU) #179, NPU GGUF-on-NPU: converter preference, forward re-sync (MLA/MoE), streaming crash fix #181, GGUF-on-NPU: all six architectures answer on --device npu (qwen3moe, qwen35moe, GLM fixes), tokenizer and serve fixes #209); the Laya row gains serve --laya: Laya classifies each conversation, a measured policy picks the device (RFC #186) #187, Laya GGUF scorer: ggmlc's laya on Vulkan, 24 ms a routing decision #206 and Laya scorer on HRX: ggmlc's laya on AMD's backend instead of Vulkan #208.
• Blog: blog/2026-09-29-hrx-and-npu-direction.md, "Where the engine is going: HRX and the NPU". It covers Laya on HRX (538 ms -> 24 ms -> 15-16 ms), W4A4 on MoE experts and --long-model (13.1 against 8.6 effective tok/s at 16K), GGUF on the NPU with its speed stated plainly, 1BP, and where HRX stands against the gates. Every number is from the merged PRs or docs. The site builds locally (39 pages) and the post renders with its three tables.

The wiki Home status was updated separately (523d6c6).

🤖 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 fd2d432

[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⚡ Recommended focus areas for review Outdated Performance Claim The README states that Laya's scorer runs on HRX at 15-16 ms per decision, which was previously 0.38 s. However, the blog post accompanying this PR provides more precise measurements, indicating that the performance has improved to 15-16 ms (with a note that it's 25 ms at 128 tokens). This discrepancy in performance claims should be reconciled to ensure consistency across documentation. > 200 labelled requests) and a measured policy picks the device; the scorer runs on HRX at > 15-16 ms a decision ([docs/laya.md](docs/laya.md)). Step 5, the model registry, has landed: of 332,726 > HF text-generation models with an architecture, 94.88% are mapped to a backend and 64.33% Missing Performance Context The blog post mentions that GGUFs on the NPU are "correct but at 0.006-0.17 tok/s" and that the next step is to implement a fused path. However, it does not provide sufficient context about the performance limitations or the specific reasons for the low throughput, which could be important for users trying to understand the current state of NPU support. GLM's MLA weights. It is not fast yet: 0.006-0.17 tok/s, because the host sequences every operation. The next step is the fused path, which the NPU work already has for Qwen3.6-35B-A3B (16.3-16.5 tok/s decode through the private add-on). Inconsistent Performance Metrics The PORTING.md file lists the performance of the NPU engine as "0.006-0.17 tok/s" for GGUFs, while the blog post provides a more detailed explanation of the performance limitations. The inconsistency in performance metrics should be addressed to ensure clarity and accuracy in the documentation. | 3 | **NPU engine.** Full ELFs only, and the open 16-tile layer kernel built from source | `engine/npu` (`npu_engine_universal.cpp`, `I8Ctx::init_elf`); ELF dispatch table on `backup/iso-build-elf-native-2026-09-22`; kernel on `bench/fastlane-16tile-corrections-2026-09-22` | **3a–3c landed** ([docs/npu.md](npu.md)). 3a (#8): full ELFs generated in C++, reproducing all 5,153 captured contexts. 3b/3c (#9): the lane runtime (logits bit-identical to the reference lane, 24/24 steps, 11.0 ms/token), tokenizer, `1bit unified`, and Qwen3-0.6B served on the NPU (now through `1bit serve`). The XDNA driver and XRT are pinned upstream and built privately (#12). GGUF on the NPU (#179, #181, #209): `1bit serve -m <gguf> --device npu` repacks a GGUF to Q4NX and runs six architectures on the per-op forward, correct but at 0.006-0.17 tok/s. **Open:** the layer kernel and lm-head artifacts are not yet built from source (npu.md, "Open") |