Converter folds float→int→float cast round-trips, dropping truncation semantics (CPU too)

[john-rocky]

Summary

The converter cancels float → int → float cast round-trips as if they were no-ops, silently dropping the truncation semantics of the intermediate integer cast. The classic "floor for bounded values" idiom (x + K).long().float() - K therefore compiles to the identity function — on every compute unit, including CPU.

Environment

• coreai-torch 0.4.0, coreai-core 1.0.0b1 (cp312), torch 2.11.0
• macOS 27.0 (build 26A5353q), M4 Max

Minimal repro

import asyncio, shutil
from pathlib import Path
import torch
import coreai.runtime as rt
from coreai_torch import TorchConverter, get_decomp_table

class M(torch.nn.Module):
    def forward(self, x):
        return (x + 64.0).long().float() - 64.0   # floor(x) for x > -64

x = torch.tensor([0.3, 1.7, -0.4, -1.6])
ep = torch.export.export(M().eval(), (x,)).run_decompositions(get_decomp_table())
prog = TorchConverter().add_exported_program(exported_program=ep, input_names=["x"], output_names=["y"]).to_coreai()
prog.optimize()
out = Path("/tmp/cast_pair.aimodel"); shutil.rmtree(out, ignore_errors=True)
prog.save_asset(out, rt.AIModelAssetMetadata())

async def run():
    m = await rt.AIModel.load(out, rt.SpecializationOptions.cpu_only())
    return (await m.load_function("main")({"x": rt.NDArray(x.numpy())}))["y"].numpy()

print(asyncio.run(run()))
# got:      [ 0.3  1.7 -0.4 -1.6]   (identity)
# expected: [ 0.   1.  -1.  -2. ]   (torch eager)

Expected

A float→int cast truncates; the pair is only removable when the value range provably contains integers. Either keep the casts or restrict the cancellation to provably-integer-valued producers.

Notes

One-directional casts consumed by integer-typed ops (e.g. gather indices) behave correctly — only the round-trip is folded. Found while porting RF-DETR's deformable-attention bilinear sampling. Related: with aten.floor unavailable on the GPU delegate (separate issue), this fold also removes the natural workaround.