wuyii8941 opened a new issue, #19578:
URL: https://github.com/apache/tvm/issues/19578
## Expected behavior
`matmul(A[M,0], B[0,N])` should return an M×N matrix of zeros (sum of zero
terms = 0), consistent with NumPy, PyTorch, and ONNX Runtime.
## Actual behavior
The output buffer is not zeroed, so the result contains uninitialized memory
(garbage values, NaN). The output is non-deterministic across runs.
## Reproducer
```python
import numpy as np
import tvm
from tvm import relax
import tvm.relax.op as R
from tvm.relax.transform import LegalizeOps
bb = relax.BlockBuilder()
a = relax.Var("a", relax.TensorStructInfo((4, 0), "float32"))
b = relax.Var("b", relax.TensorStructInfo((0, 4), "float32"))
with bb.function("main", [a, b]):
with bb.dataflow():
gv = bb.emit_output(bb.emit(R.matmul(a, b)))
bb.emit_func_output(gv)
mod = bb.finalize()
pipeline = tvm.ir.transform.Sequential([LegalizeOps()])
exe = tvm.relax.build(pipeline(mod), target="llvm")
vm = tvm.relax.VirtualMachine(exe, device=tvm.cpu())
out = vm["main"](
tvm.runtime.tensor(np.zeros((4, 0), np.float32), device=tvm.cpu()),
tvm.runtime.tensor(np.zeros((0, 4), np.float32), device=tvm.cpu()),
).numpy()
print(out)
# Non-deterministic garbage, e.g.:
# [[ 6.19e+21 3.07e-41 5.25e+21 3.07e-41]
# [ 6.19e+21 3.07e-41 5.25e+21 3.07e-41]
# [ nan -7.43e-44 nan -1.40e-44]
# [ nan 7.57e-44 nan 4.48e-44]]
print(np.zeros((4, 0)) @ np.zeros((0, 4)))
# [[0. 0. 0. 0.]
# [0. 0. 0. 0.]
# [0. 0. 0. 0.]
# [0. 0. 0. 0.]]
```
## Scope
The same issue affects other reduction operations with a zero-size reduction
axis:
```python
# All return uninitialized memory instead of the identity element:
R.sum(X[4, 0], axis=1) # expected: zeros(4,1)
R.prod(X[4, 0], axis=1) # expected: ones(4,1)
R.matmul(A[2,3,4,0], B[2,3,0,5]) # expected: zeros(2,3,4,5) [batched]
R.nn.conv2d(X[1,0,8,8], W[4,0,3,3]) # expected: zeros(1,4,8,8) [0
in_channels]
R.einsum([A[4,0], B[0,8]], 'ik,kj->ij') # expected: zeros(4,8)
```
Notably, `R.max` and `R.min` over size-0 axes are handled correctly
(returning -inf / +inf).
## Why this matters
The garbage output propagates silently. For example:
```python
mm = R.matmul(A[4,0], B[0,8]) # garbage
sm = R.nn.softmax(mm, axis=-1) # produces [1,0,0,...,0] instead of uniform
```
In dynamic-shape models, K=0 can occur naturally when a filter removes all
elements. The result is silent, non-deterministic incorrect output with no
crash or warning.
## Environment
- TVM commit: 0b0afd8dd (main, 2026-04-24)
- OS: Ubuntu 20.04
- Target: llvm (CPU)
--
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]
