Expose PDI output format for alternative runtimes on npu1

[erwei-xilinx]

Summary

Exposes output_format='pdi' on npu1/Phoenix (motivated by #53), producing a raw aie.pdi plus an insts.bin instruction sidecar for an alternative runtime that loads PDIs directly (not XRT). The PDI path is the same internal flow as xclbin — the xclbin just wraps the PDI in an extra layer.

• Bump pinned mlir-air to 2026071405+c7ee42f — the first release wheel containing the aircc --output-format=pdi / --pdi-name flags from Xilinx/mlir-air PR #1729. Older wheels lack the flag, so the bump must land in the same PR.
• Config: accept 'pdi' in npu_config.output_format (env AMD_TRITON_NPU_OUTPUT_FORMAT), alongside elf/xclbin.
• aircc: new PDI branch in _aircc_compile passing --output-format pdi --pdi-name aie.pdi -i insts.bin; artifact cache get/put generalized to a format-neutral bin_path so pdi and xclbin share the insts.bin plumbing.
• Launcher: PDI targets a non-XRT runtime, so the in-repo XRT launcher can't dispatch a .pdi. It now throws a clear error when the device binary ends in .pdi (mirroring mlir-air's XRTBackend.load()). Export artifacts with AMD_TRITON_NPU_COMPILE_ONLY=1 and hand the .pdi + .insts.bin to the target runtime.

Test plan

• Reinstalled triton-xdna with the bumped AIR; confirmed aircc --help now lists --output-format=pdi / --pdi-name.
• i8 matmul compile-time regression check on npu2 (Strix), cold compile of 256×256×256 with correctness verified: old AIR median 4.51s vs new AIR median 4.61s — within noise, no regression.
• End-to-end on npu2 (all PASS): matmul_i8_m64_n64_k64, matmul_i8_m128_n64_k64, matmul_bf16_m64_n64_k64, vec-add, relu.
• PDI export path (AMD_TRITON_NPU_OUTPUT_FORMAT=pdi + AMD_TRITON_NPU_COMPILE_ONLY=1) on npu1 hardware — not exercised here (npu2 defaults to elf; PDI is npu1-oriented and export-only).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR exposes a new NPU backend output format (output_format='pdi') intended for npu1/Phoenix workflows where an alternative (non-XRT) runtime consumes a raw aie.pdi plus an insts.bin sidecar, and updates the pinned mlir-air revision to a build that supports the necessary aircc flags.

Changes:

• Bump the pinned mlir-air commit/timestamp to a build that includes aircc --output-format=pdi / --pdi-name.
• Extend backend configuration parsing/validation to accept output_format="pdi" via npu_config and AMD_TRITON_NPU_OUTPUT_FORMAT.
• Add a PDI compilation branch in driver._aircc_compile, generalize artifact caching to bin_path, and make the generated XRT launcher fail early with a clear error when given a .pdi.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
utils/mlir-air-hash.txt	Updates pinned mlir-air commit/timestamp to a build that provides the new aircc PDI flags.
amd_triton_npu/backend/driver.py	Adds PDI compile support and format-neutral caching changes; adds an explicit runtime error when attempting to XRT-launch a .pdi.
amd_triton_npu/backend/config.py	Accepts and validates output_format="pdi" from config/env and documents the new option.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[erwei-xilinx]

Addressed the Copilot review in 9e0577b:

• Real bug (multilaunch.py): _get_cached_aircc_artifacts / _put_aircc_artifacts now return bin_path (not xclbin_path), but MultiLaunch.compile still read cached["xclbin_path"] on both the cache-hit (multilaunch.py:216) and cache-miss (:229) paths — a KeyError for xclbin/pdi multi-launch compiles. Updated both call sites to cached["bin_path"].
• Docstrings: updated _get_cached_aircc_artifacts (driver.py) and MultiLaunch.compile (multilaunch.py) to reflect the elf / xclbin / pdi shapes.

Verified no remaining dict-key references to xclbin_path in the backend (the only remaining xclbin_path refs are the local path variable inside _aircc_compile).