Vendored: ROCm/FastFlowLM (open-source FLM runtime)

Vendored from https://github.com/ROCm/FastFlowLM at commit 17a35cb (Merge pull request #659 from ROCm/repo-transfer — repo moved from FastFlowLM org to ROCm, 2026-08-07; MIT license).

The runtime that serves the Q4NX models on the Ryzen AI NPU (the Q4NX pivot). The installed package (/opt/fastflowlm, FLM v0.9.46) matches this tree. This vendor is the reference/archive copy — the build links the installed package (see CMakeLists FLM_BINARY/FLM_CONFIG/FLM_XCLBINS discovery).

What is open in this tree:

• server / CLI / runner / tokenizer / model pull + registry (full sources)
• whisper model (fully open .cpp)
• xclbins (NPU AIE kernels, binaries) + aiebu assembler headers
• companion converter: third_party/FLM_Q4NX_Converter (GGUF -> q4nx writer)

What is NOT yet open (prebuilt .lib/.dll in src/lib/): the NPU model engines (qwen3_npu, llama_npu, ...) and the Q4NX file reader (npu_quantize_block). Declarations in src/include/; re-vendor on upstream sync.