Repository navigation
[Bug] [Relax][ONNX] Float8 Cast drops saturate and merges distinct programs #20936
Description
Expected behavior
Cast-19 distinguishes saturate=0 and 1 for FLOAT8E4M3FN overflow. Casting 1000 and casting back should give NaN and 448, respectively.
Actual behavior
Full ONNX checking and ReferenceEvaluator distinguish both models, while from_onnx produces structurally identical Relax IR. The Cast converter never reads saturate. This demonstrates semantic information loss during conversion; it does not attribute every Float8 backend difference to this attribute or claim a native Float8 kernel was executed.
Environment
TVM 0.28.dev0, LLVM 18.1.3, Linux x86_64; Python 3.12.3, NumPy 2.5.4, ONNX 1.22.0. Installed binary build revision is unknown. Current main was inspected statically, not rebuilt.
Steps to reproduce
import numpy as np
import onnx
import tvm
from onnx import TensorProto as T, helper as h
from onnx.reference import ReferenceEvaluator
from tvm.relax.frontend.onnx import from_onnx
x = np.array([1., 1000.], dtype="float32")
modules, expected = [], []
for saturate in [0, 1]:
model = h.make_model(h.make_graph([
h.make_node("Cast", ["x"], ["f8"], to=T.FLOAT8E4M3FN, saturate=saturate),
h.make_node("Cast", ["f8"], ["y"], to=T.FLOAT)], "float8_cast",
[h.make_tensor_value_info("x", T.FLOAT, [2])],
[h.make_tensor_value_info("y", T.FLOAT, [2])]),
opset_imports=[h.make_opsetid("", 19)], ir_version=10)
onnx.checker.check_model(model, full_check=True)
ref = ReferenceEvaluator(model).run(None, {"x": x})[0]
assert ref[0] == 1
expected.append(ref)
modules.append(from_onnx(model, keep_params_in_input=False))
print("saturate", saturate, "reference", ref.tolist())
assert np.isnan(expected[0][1]) and expected[1][1] == 448
tvm.ir.assert_structural_equal(modules[0], modules[1])
print("distinct ONNX Cast semantics import to structurally identical Relax IR")
print(modules[0].script())Triage
- needs-triage
Activity