Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
test(onnx-export): tolerate cross-machine fp32 drift, and skip quantization on Windows
Adeshina Lasisi2d906cbf8f9c
diff --git a/tests/test_onnx_export_batch.py b/tests/test_onnx_export_batch.pyindex 1333b0a0..841edddf 100644--- a/tests/test_onnx_export_batch.py+++ b/tests/test_onnx_export_batch.py@@ -101,6 +101,20 @@ class _TinyEncoder(nn.Module):return types.SimpleNamespace(last_hidden_state=hidden)+class _Skip(Exception):+ """Raised to skip a block this platform cannot run."""+++#: Tolerance for the TINY model in this file, not for a real checkpoint. `verify_batch_dynamic`+#: defaults to 1e-3, which the 322M checkpoint passes at 6.14e-05 -- that default is what bounds+#: quality and it is asserted separately. This 32-dim random model has near-tied logits, and softmax+#: amplifies the difference between one machine's fp32 kernels and another's: measured 4.69e-04 here+#: and 1.03e-02 on a GitHub runner, which straddles 1e-3 and made this suite fail by hardware. The+#: signal it has to separate is enormous by comparison -- a perturbed weight moves probabilities by+#: 6.4e-01 -- so 5e-2 sits 5x above the worst drift observed and 13x below the thing being detected.+TINY_ATOL = 5e-2++def _build_model():torch.manual_seed(0)# head_layers=2 is the shipped default, and the head is where the traced batch got baked in.@@ -155,8 +169,17 @@ finally:# exported a real graph but never quantized it. Deleting that one line leaves both suites green and# makes `python scripts/export_onnx.py --quantize` crash outright with# `InferenceError: Inferred shape and existing shape differ in dimension 0`.-_tmp3 = tempfile.mkdtemp()+# Skipped on Windows: quantizing a graph this size dies there with an illegal instruction inside+# onnxruntime (exit 132), which no `except` can catch. `tests/test_onnx_quantize.py` passes on+# Windows because its graph is a single MatMul node. The `onnx export (quantization, weight-free)`+# job on Linux is where this path is gated, and that is the job the documented `--quantize`+# invocation has to keep green.+_tmp3 = None if sys.platform.startswith("win") else tempfile.mkdtemp()+if _tmp3 is None:+ check_true("quantize/skipped on Windows (covered by the Linux quantization job)", True, "")try:+ if _tmp3 is None:+ raise _Skip()_p3 = os.path.join(_tmp3, "q.onnx")_m3 = _build_model()export_onnx.export_module(_m3, _p3)@@ -169,8 +192,10 @@ try:_s3 = ort.InferenceSession(_int8, providers=["CPUExecutionProvider"])check("quantize/the INT8 graph keeps a symbolic batch axis",_s3.get_inputs()[0].shape[0], "batch_size")+except _Skip:+ passfinally:- shutil.rmtree(_tmp3, ignore_errors=True)+ shutil.rmtree(_tmp3, ignore_errors=True) if _tmp3 else None# Every way of turning the new verification OFF survived: the CLI was pinned only by grepping# __main__ for substrings, and the stubs recorded THAT a function was called, never with what. So@@ -178,8 +203,12 @@ finally:# `atol=1e9` on the INT8 check all shipped a success message while checking nothing. The real# __main__ now runs against a stubbed Agent and the arguments it passes are asserted._seen = {}-_tmp4 = tempfile.mkdtemp()+_tmp4 = None if sys.platform.startswith("win") else tempfile.mkdtemp()+if _tmp4 is None:+ check_true("cli/skipped on Windows (it exports and quantizes; see above)", True, "")try:+ if _tmp4 is None:+ raise _Skip()_p4 = os.path.join(_tmp4, "cli.onnx")def _fake_verify(model, path, atol=None, batches=(1, 2, 3)):@@ -209,8 +238,10 @@ try:_calls[1]["atol"] if len(_calls) > 1 else None, export_onnx.INT8_ATOL)check_true("cli/the INT8 graph is written beside the fp32 one, not over it",len(_calls) > 1 and _calls[1]["path"] != _calls[0]["path"], _calls)+except _Skip:+ passfinally:- shutil.rmtree(_tmp4, ignore_errors=True)+ shutil.rmtree(_tmp4, ignore_errors=True) if _tmp4 else Noneshutil.rmtree(tmp, ignore_errors=True)os.makedirs(tmp, exist_ok=True)@@ -305,7 +336,7 @@ check_true("parity/the comparison is real, not an all-zero graph", _biggest_outp# ---------------------------------------------------------------- verify_batch_dynamic accepts ittry:- export_onnx.verify_batch_dynamic(model, onnx_path)+ export_onnx.verify_batch_dynamic(model, onnx_path, atol=TINY_ATOL)PASS.append("verify/accepts a batch-dynamic export")except SystemExit as error:FAIL.append("verify/accepts a batch-dynamic export: raised %r" % (str(error),))@@ -342,7 +373,7 @@ with torch.no_grad():_w = [q for q in _perturbed.parameters() if q.ndim >= 2][-1]_w.add_(torch.linspace(0.0, 5.0, _w.numel()).reshape(_w.shape))try:- export_onnx.verify_batch_dynamic(_perturbed, onnx_path)+ export_onnx.verify_batch_dynamic(_perturbed, onnx_path, atol=TINY_ATOL)FAIL.append("verify/rejects a graph whose numbers drifted: it was accepted, so the guard is vacuous")except SystemExit as error:check_true("verify/rejects a graph whose numbers drifted",@@ -362,7 +393,7 @@ import contextlib as _contextlib # n_cap = _io.StringIO()with _contextlib.redirect_stdout(_cap):- export_onnx.verify_batch_dynamic(model, onnx_path)+ export_onnx.verify_batch_dynamic(model, onnx_path, atol=TINY_ATOL)_reported = _cap.getvalue()for _name in export_onnx.OUTPUT_NAMES:check_true("verify/compares %s against PyTorch" % _name,@@ -432,7 +463,7 @@ _hidden = {"onnxruntime": None}_saved = {k: sys.modules.get(k) for k in _hidden}sys.modules.update(_hidden)try:- export_onnx.verify_batch_dynamic(model, onnx_path)+ export_onnx.verify_batch_dynamic(model, onnx_path, atol=TINY_ATOL)check_true("cli/missing onnxruntime is explained", False, "no SystemExit was raised")except SystemExit as error:msg = str(error)