README, NOTICE: acknowledge FastFlowLM

[bong-water-water-bong]

The NPU route runs FastFlowLM's Q4NX models, and Lemonade's onebit recipe now downloads FastFlowLM/*-NPU2 (#72, fork #7). FastFlowLM's README asks projects to acknowledge it with Powered by [FastFlowLM](https://github.com/ROCm/FastFlowLM). This adds:

• that line and a row in the README's thanks table;
• a NOTICE entry: the code is MIT; the engine reads the Q4NX format and runs the models on its own full-ELF kernels, without FastFlowLM's runtime or kernel binaries.

🤖 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 f12b5df

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 72 - Partially compliant Compliant requirements: The build's own lax kernels are the default for NPU inference 1bit serve finds the lax kernels without a flag The lookup order is implemented -DONEBIT_NPU_LAX=ON builds the lax kernels into the build tree -DONEBIT_NPU_LAX_KERNELS=<dir> points at an existing build The change is measured on strixhalo with specific md5 hashes and performance metrics Documentation in docs/npu-lax.md and docs/lemonade.md is updated Non-compliant requirements: The small NPU models are not yet supported reasoning_content on the NPU route is not yet supported Our own HF repos are not yet supported Requires further human verification: The small NPU models support The reasoning_content on the NPU route Our own HF repos support 7 - Partially compliant Compliant requirements: Pinned sources for ROCm/hrx, ROCm/hrx-system, and llama.cpp fork Patches 0001–0003 for build plumbing and runtime fix Build with -DONEBIT_HRX=ON builds all three from the build tree Lemonade wiring for 1bit lemonade to run llamacpp-hrx on HRX20 and llamacpp/Vulkan on Vulkan0 New hrx_device option in Lemonade Verified on Strix Halo with reproducible kernels and speed parity Default build with ONEBIT_HRX=OFF is unchanged Non-compliant requirements: zaya Q4NX through Lemonade is not yet supported A relocatable package for HRX libraries is not yet supported HRX in CI is not yet supported Requires further human verification: zaya Q4NX through Lemonade support A relocatable package for HRX libraries HRX in CI support

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Missing Documentation Update The PR adds acknowledgment for FastFlowLM in the README but does not update the documentation files (docs/npu-lax.md and docs/lemonade.md) to reflect the new default behavior of using the build's own lax kernels. This is a requirement from the ticket description. | 11 | [ROCm/FastFlowLM](https://github.com/ROCm/FastFlowLM) | The Q4NX NPU model format and its models on Hugging Face (`FastFlowLM/*-NPU2`), which the engine's NPU route runs on its own kernels | MIT | Powered by [FastFlowLM](https://github.com/ROCm/FastFlowLM) Incomplete License Information The NOTICE file includes a note about FastFlowLM's MIT license for code, but it does not fully reflect the license information for the models themselves, which are under different licenses (e.g., Qwen's Apache-2.0). The description should be more precise about the licensing of the models and their weights. License: MIT (code). The engine's NPU route reads FastFlowLM's Q4NX model format, and Lemonade's onebit recipe downloads FastFlowLM's models from Hugging Face (FastFlowLM/*-NPU2; the weights keep their own licenses, e.g. Qwen's Apache-2.0). The engine runs them on its own kernels as full ELFs; it does not use FastFlowLM's runtime or ship its kernel binaries. Powered by FastFlowLM, as FastFlowLM asks.