Anemone220 opened a new issue, #20178:
URL: https://github.com/apache/tvm/issues/20178
### Expected behavior
The Relax LLVM compilation pipeline should handle this valid ONNX model
without an internal assertion.
### Actual behavior
The script constructs a valid ONNX model. ONNX checker, ONNX Runtime session
construction, and `from_onnx` succeed. Compiling it with the shown Relax
pipeline fails in `FuseTIR`:
```text
Traceback (most recent call last):
...
File ".../src/s_tir/transform/renew_defs.cc", line 167, in AddDefRemap
TVM_FFI_ICHECK(remap_.count(source) == 0)
tvm.error.InternalError: Check failed: (remap_.count(source) == 0) is false:
```
### Environment
* OS: macOS 15.6 (Darwin 24.6.0, arm64)
* Python: 3.12.2
* TVM: c7b458e946bc4266915da582457476bdcd9705ae (tag v0.26.0; package
reports 0.26.dev0)
* ONNX: 1.17.0
* ONNX Runtime: 1.21.1
* Target: `llvm`
* Frontend: `tvm.relax.frontend.onnx.from_onnx`
### Steps to reproduce
The following self-contained script constructs and compiles the model:
```python
#!/usr/bin/env python3
"""Reproduce the FuseTIR/RenewDefs Flatten-Slice-Cast failure."""
import onnx
import onnxruntime as ort
import tvm
from onnx import TensorProto, helper
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
def main() -> None:
x = helper.make_tensor_value_info("x", TensorProto.FLOAT, ["n", 8])
out = helper.make_tensor_value_info("out", TensorProto.INT64, [1, 1])
initializers = [
helper.make_tensor("starts", TensorProto.INT64, [1], [0]),
helper.make_tensor("ends", TensorProto.INT64, [1], [1]),
helper.make_tensor("axes", TensorProto.INT64, [1], [1]),
helper.make_tensor("steps", TensorProto.INT64, [1], [1]),
]
nodes = [
helper.make_node("Flatten", ["x"], ["flat"], axis=0),
helper.make_node(
"Slice",
["flat", "starts", "ends", "axes", "steps"],
["sliced"],
),
helper.make_node("Cast", ["sliced"], ["out"], to=TensorProto.INT64),
]
graph = helper.make_graph(
nodes,
"renew_defs_flatten_slice_cast",
[x],
[out],
initializers,
)
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("",
18)])
onnx.checker.check_model(model)
ort.InferenceSession(model.SerializeToString(),
providers=["CPUExecutionProvider"])
mod = from_onnx(model, opset=18, keep_params_in_input=True)
pipeline = tvm.transform.Sequential(
[
relax.backend.DispatchSampling(),
relax.backend.DispatchSortScan(),
relax.transform.LegalizeOps(),
relax.transform.AnnotateTIROpPattern(),
relax.transform.FoldConstant(),
relax.transform.FuseOps(fuse_opt_level=2),
relax.transform.FuseTIR(),
relax.transform.RewriteDataflowReshape(),
relax.transform.ToNonDataflow(),
relax.transform.RemovePurityChecking(),
relax.transform.CallTIRRewrite(),
relax.transform.StaticPlanBlockMemory(),
relax.transform.LowerAllocTensor(),
relax.transform.KillAfterLastUse(),
relax.transform.LowerRuntimeBuiltin(),
relax.transform.ComputePrimValue(),
relax.transform.VMShapeLower(emit_err_ctx=True),
relax.transform.AttachGlobalSymbol(),
]
)
tvm.compile(mod, target="llvm", relax_pipeline=pipeline)
if __name__ == "__main__":
main()
```
### Analysis
The frontend expands the float-to-int `Cast` into an
`isfinite`/`where`/`astype` sequence. With the symbolic input dimension `n`,
the normal fusion pipeline reaches `RenewDefs` and fails on a duplicate remap
entry. Replacing `n` with a concrete dimension avoids this failure on this
checkout.
### Triage
* needs-triage
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
