Evaluate native HRX and Loom on gfx1100 with matched measurements

Row: BACKEND-ROCM

Evaluate native AMD HRX and Loom for the project on gfx1100. The developer requested hardware measurements before selecting an architecture. Existing vllm.cpp code is a comparator, not a preferred design or correctness oracle.

Build pinned HRX and the author-linked llama.cpp HRX integration. Establish native gfx1100 execution and correctness, compare HRX and conventional HIP on identical supported workloads, separate runtime-submission and kernel/compiler contributions, and assess vllm.cpp integration including Qwen3.5 GDN and quantized arms. Native HRX, its HIP compatibility layer, and selective Loom reuse are all candidates.

Starting anchors: ROCm/hrx-system 6bcd5a4ff111fa5bf160ab9f4592ca8e7cc810b1; AMD-Ecosystem/llama.cpp hrx-graph-develop-v2 at 6319038132ed12f968ea68f37753f705da830ea8; https://github.com/ggml-org/llama.cpp/discussions/27219. Establish exact-pair build compatibility. Spec: .agents/specs/rocm-hrx-evaluation.md, to be committed before implementation.

Record revisions, commands, binary/model hashes, device identity and contention, output correctness, cold-start/JIT and warm prefill/decode/latency/memory. Product correctness uses the applicable pinned oracle and unchanged thresholds. Upstream benchmark or compiler-smoke results are diagnostic evidence, not end-to-end vllm.cpp acceptance.

Finish with an independently reviewed empirical recommendation: adopt, pursue a specific measurable gap, or reject for measured reasons. Missing coverage is engineering scope to investigate. Shipped defaults require the normal acceptance gates.

The BACKEND-ROCM operator owns this evaluation. Task staging and builds stay in ignored build directories, with no /tmp or /dev/shm workspaces. Preserve compact reproducible evidence and remove disposable artifacts when no longer needed.

Start with .agents/specs/rocm-hrx-evaluation.md and the pinned HRX and llama.cpp revisions, then establish the exact-pair builds and native gfx1100 execution. Record reproducible hashes, commands, correctness, startup, warm performance, latency, memory, and contention across the specified workloads. Done means an independently reviewed recommendation to adopt, pursue a measured gap, or reject the approaches.