| from __future__ import annotations |
|
|
| import hashlib |
| import importlib.util |
| import json |
| import os |
| import re |
| import sys |
| import types |
| from pathlib import Path |
|
|
| import numpy as np |
|
|
| APP = Path(os.environ.get("TBENCH_APP", "/app")) |
| TESTS = Path(os.environ.get("TBENCH_TESTS", "/tests")) |
| FIX = TESTS / "fixtures" |
|
|
| BASE_HASHES = { |
| "vendor/onnx/onnx/reference/ops/op_tensor_scatter.py": "2f12049a0582cfb4e5c9017c189b1ea27d872d550b97cd5156008d79b9fbed5c", |
| "vendor/onnx/onnx/defs/tensor/defs.cc": "3f418e462e60b2fedaebe7a4ce2311ece098c961422ffc9e397bbd8be9697321", |
| "vendor/onnx/onnx/reference/op_run.py": "dbe6f43327f69396d27d48b121f54361f55d9ea67b793cc121314ee3ba389fbd", |
| "vendor/onnx/onnx/reference/ops/_helpers.py": "5b588246f91181be2ea9c7b93055493a6d08aa4fc5b6f147b7b7c2f58786b59d", |
| "vendor/onnx/onnx/reference/ops/_op.py": "5f6682ef95f32d46e8f8e01a47a18867f938e7419a61538426f0c3ca3e888965", |
| } |
|
|
|
|
| def sha256(path: Path) -> str: |
| return hashlib.sha256(path.read_bytes()).hexdigest() |
|
|
|
|
| def load_operator(): |
| onnx_mod = types.ModuleType("onnx") |
| reference_mod = types.ModuleType("onnx.reference") |
| op_run_mod = types.ModuleType("onnx.reference.op_run") |
| op_run_mod.OpRun = type("OpRun", (), {}) |
| sys.modules.setdefault("onnx", onnx_mod) |
| sys.modules.setdefault("onnx.reference", reference_mod) |
| sys.modules["onnx.reference.op_run"] = op_run_mod |
| source = APP / "vendor/onnx/onnx/reference/ops/op_tensor_scatter.py" |
| spec = importlib.util.spec_from_file_location("test_tensor_scatter", source) |
| assert spec and spec.loader |
| module = importlib.util.module_from_spec(spec) |
| spec.loader.exec_module(module) |
| return module.TensorScatter() |
|
|
|
|
| def recompute(case): |
| past = np.asarray(case["past_cache"], dtype=np.int64) |
| update = np.asarray(case["update"], dtype=np.int64) |
| indices = np.asarray(case["write_indices"], dtype=np.int64) |
| axis = int(case["axis"]) % past.ndim |
| assert past.ndim == update.ndim == 2 |
| assert all(past.shape[i] == update.shape[i] for i in range(2) if i != axis) |
| assert update.shape[axis] <= past.shape[axis] |
| out = past.copy() |
| for prefix in np.ndindex(past.shape[:axis]): |
| b = prefix[0] |
| for j in range(update.shape[axis]): |
| pos = int(indices[b]) + j |
| if case["mode"] == "circular": |
| pos %= past.shape[axis] |
| elif case["mode"] != "linear": |
| raise ValueError("unsupported mode") |
| out[prefix + (pos,)] = update[prefix + (j,)] |
| return out |
|
|
|
|
| def test_gate_artifact_schema_and_integrity(): |
| artifact = APP / "output.json" |
| assert artifact.is_file() |
| payload = json.loads(artifact.read_text()) |
| assert payload["output_schema_version"] == "tensor_scatter_inference.v1" |
| assert [r["name"] for r in payload["results"]] == ["batch_wrap", "linear_append"] |
| public = json.loads((APP / "fixtures/kv_cases.json").read_text()) |
| expected = [recompute(case).tolist() for case in public["cases"]] |
| assert [r["values"] for r in payload["results"]] == expected |
| assert sha256(APP / "vendor/onnx/onnx/reference/ops/op_tensor_scatter.py") != BASE_HASHES["vendor/onnx/onnx/reference/ops/op_tensor_scatter.py"] |
| assert sha256(APP / "vendor/onnx/onnx/defs/tensor/defs.cc") != BASE_HASHES["vendor/onnx/onnx/defs/tensor/defs.cc"] |
| assert sha256(APP / "fixtures/kv_cases.json") == "3f5542fc823784faae35d727d3bf35bdff5760320ffc27b47ff27f57424648b0" |
| for rel, digest in BASE_HASHES.items(): |
| if rel.endswith("op_tensor_scatter.py") or rel.endswith("defs.cc"): |
| continue |
| assert sha256(APP / rel) == digest |
|
|
|
|
| def test_core_circular_batch_identity(): |
| operator = load_operator() |
| cases = json.loads((FIX / "hidden_cases.json").read_text())["cases"] |
| actual = operator._run( |
| np.asarray(cases[0]["past_cache"], dtype=np.int64), |
| np.asarray(cases[0]["update"], dtype=np.int64), |
| np.asarray(cases[0]["write_indices"], dtype=np.int64), |
| cases[0]["mode"], |
| cases[0]["axis"], |
| )[0] |
| expected = recompute(cases[0]) |
| assert actual.shape == (6, 4) |
| assert np.array_equal(actual, expected) |
| assert actual[:, 0].tolist() == [12, 23, -30, 41, 52, 63] |
|
|
|
|
| def test_core_linear_and_negative_axis(): |
| operator = load_operator() |
| cases = json.loads((FIX / "hidden_cases.json").read_text())["cases"] |
| for case in cases[1:]: |
| actual = operator._run( |
| np.asarray(case["past_cache"], dtype=np.int64), |
| np.asarray(case["update"], dtype=np.int64), |
| np.asarray(case["write_indices"], dtype=np.int64), |
| case["mode"], |
| case["axis"], |
| )[0] |
| assert np.array_equal(actual, recompute(case)) |
| assert operator._run(np.zeros((1, 4), dtype=np.int64), np.array([[9, 8]]), np.array([3]), "circular", -1)[0].tolist() == [[8, 0, 0, 9]] |
|
|
|
|
| def test_edge_shape_and_mode_errors(): |
| operator = load_operator() |
| with np.testing.assert_raises(ValueError): |
| operator._run(np.zeros((2, 4), dtype=np.int64), np.zeros((3, 3), dtype=np.int64), np.array([0, 0]), "circular", 1) |
| with np.testing.assert_raises(ValueError): |
| operator._run(np.zeros((2, 4), dtype=np.int64), np.zeros((2, 5), dtype=np.int64), np.array([0, 0]), "linear", 1) |
| with np.testing.assert_raises(ValueError): |
| operator._run(np.zeros((2, 4), dtype=np.int64), np.zeros((2, 2), dtype=np.int64), np.array([0, 0]), "diagonal", 1) |
|
|
|
|
| def test_edge_cpp_contract_and_python_build(): |
| source = (APP / "vendor/onnx/onnx/defs/tensor/defs.cc").read_text() |
| block = source[source.index("static constexpr const char* TensorScatter_ver24_doc"):source.index("ONNX_OPERATOR_SET_SCHEMA(\n TensorScatter", source.index("static constexpr const char* TensorScatter_ver24_doc"))] |
| assert "np.asarray(cache_idx)" not in block |
| assert re.search(r"write_indices\[batch_idx\]\s*\+\s*sequence_idx\)\s*%\s*max_sequence_length", block) |
| compile_targets = [APP / "run_kv_inference.py", APP / "vendor/onnx/onnx/reference/ops/op_tensor_scatter.py"] |
| import py_compile |
| for target in compile_targets: |
| py_compile.compile(str(target), doraise=True) |
| assert sha256(APP / "run_kv_inference.py") == "3d42dd5e67e02f193342176ddcdcddd6819499bc871eb16b76a792cb840d1d58" |
|
|
