[zhen8838]

Why

Keep the Qwen3-1.7B and Qwen3.5-4B AMX prefill experiments executable with shared HIR definitions and a directly callable runtime module.

What

• Add the four AMX operations used by the example (X/Y loads, FP16 outer product with FP32 accumulation, and Z store) to TileFoundry, together with minimal native instruction wrappers shipped with the package.
• Admit Copy/Fill on Apple core and AMX executor scopes.
• Add whole-model HIR, scheduled HIR, explicit Python runtime twins and C++ projection kernels under examples/qwen_prefill-amx. Build/cache native kernels on first use; prepare GGUF weights from the entry point when needed.
• Keep architecture parameters in Python. No JSON artifacts or standalone build/export/check scripts are included.

Contract

AMX contracts and wrappers belong to TileFoundry; model-specific orchestration and kernels belong to the example. The runtime is hand-written and callable directly after loading resources. CPU host operations use Torch. The example covers batch-one text prefill through logits.

Execution validation on M2 Pro: both full models complete 128-token prefill using existing prepared Q8_0 resources, with finite logits and next-token IDs 1813 / 306. Native kernels build automatically. No tests are added or modified relative to main; no performance measurements were repeated.

Risk

Requires Apple Silicon and Command Line Tools. Execution checks do not establish generation quality. This adds only the AMX instructions needed by the example, without an AMX machine-code generator.