Add generic InfiniOps operator performance adapter

[baominghelly]

Summary

• add a generic InfiniOps operator performance adapter for ten common operators
• describe operators declaratively and route them through binary, cast, concat, and matrix builder families
• move repeated JSON fields and vendor runtime string mappings into common constants
• route Cambricon and Ascend operator cases through lazy-loaded device plugins
• propagate accuracy and runtime-availability failures through the executor
• report latency, tensor accuracy, estimated TFLOPS, and estimated bandwidth

Supported Operators

• Binary: add, sub, mul, div
• Conversion and concat: cast, cat
• Matrix: gemm, matmul, mm, linear

Model-specific and stateful operators such as flash attention, rotary embedding, and KV cache updates are intentionally deferred. Their optional inputs, mixed dtypes, multiple outputs, and aliasing rules need separate contracts and correctness references.

Why

The source branch implemented each operator with a separate setup method and duplicated platform mappings and JSON field strings throughout the adapter. It also assumed implementation slot 8 existed whenever no native implementation was found, which can terminate the InfiniOps C++ dispatcher.

The refactored adapter uses four reusable builder families and only invokes implementation indices returned by active_implementation_indices. Missing runtime implementations become actionable per-test failures instead of process-level aborts.

Input generation and result aggregation utilities were split into draft PR #46 so this PR remains focused on adapter integration.

Validation

Software checks:

• python3 -m pytest -q tests/test_infiniops_adapter.py tests/test_executor_adapter_response.py tests/test_operator_flops.py: 16 passed
• python3 -m compileall -q infinimetrics tests main.py: passed
• Black 23.9.1 check: passed
• Flake8 7.3.0 check: passed

Cambricon hardware smoke test:

• Hardware: one Cambricon MLU590-M9C
• Runtime: Python 3.10.8, PyTorch 2.1.0, torch-mlu 1.25.3+torch2.1.0, InfiniOps 0.1.0
• Method: float16 small shapes, 2 warmup iterations, 5 measured iterations, full main.py -> Dispatcher -> Executor -> InfiniOpsAdapter path
• Runtime-supported add/sub/div/mm: 12/12 passed
• All result files reported tensor accuracy PASS and populated latency, TFLOPS, and bandwidth metrics
• Operators without an active MLU implementation return clean per-test failures rather than invoking an unregistered slot

The latency observations are smoke-test evidence only, not a performance baseline.

[baominghelly]

GitHub automatically closed this PR when its head branch was renamed to feat/cambricon-op-performance-adapter. Review continues in draft PR #47 with the same commit SHA and diff.