P1-B: BufferHandle memory-model ABI + user-facing Tensor cutover - #1453

Draft

YunjiQin wants to merge 10 commits into

hw-native-sys:mainfrom

YunjiQin:p1-b-buffer-handle-abi

Commits on Jul 30, 2026

Add the BufferHandle/BufferRef memory-model ABI + owner create_buffer

YunjiQin
and
claude
committed
Flip the live dispatch wire to BufferRef; three-way consumer split; device/remote/strided

YunjiQin
and
claude
committed
Add resource APIs + ABI doc + cpput migration; remove TensorTaskArgs; L2 in-process materialization

YunjiQin
and
claude
committed
Move device-memory ownership from Orchestrator to Worker (thin delegate)

YunjiQin
and
claude
committed
Add make_ref_arg + comm-domain VMM_WINDOW buffers; migrate examples & scene tests to BufferRef/handle

YunjiQin
and
claude
committed
Make ChipStorageTaskArgs chip-only; delete create_host_buffer; sub scalars; orch.alloc→handle

YunjiQin
and
claude
committed
Add the §4.1 capability matrix + access⊆granted dispatch guards (P1)

YunjiQin
and
claude
committed
Update: slim the buffer ABI, validate it, and stage host copies
YunjiQin
committed
Add: submit-time argument guard and a copy-on-write backend tag
YunjiQin
committed
Refactor: rename Worker.make_ref_arg to make_tensor_arg
YunjiQin
committed