ymwangg opened a new issue #8441:
URL: https://github.com/apache/tvm/issues/8441
It looks like the native tvm implementation of nn.dense does not handle
dynamic shapes correctly though using libs such as mkl, cublas has no issues.
The following is the code to reproduce this issue.
```python
import tvm
from tvm import relay
from tvm.relay import create_executor, Any
import numpy as np
A = relay.var("A",shape=[Any(), Any()],dtype="float32")
B = relay.var("B",shape=[Any(), Any()],dtype="float32")
C = relay.nn.dense(A, relay.transpose(B))
f = relay.Function([A, B], C)
mod = tvm.IRModule.from_expr(f)
for target in ["llvm -libs=mkl", "llvm"]:
dev = tvm.device(target,0)
executor = create_executor(kind="vm", mod=mod, device=dev, target=target)
a = np.random.uniform(size=[10,10]).astype("float32")
b = np.random.uniform(size=[10,10]).astype("float32")
res = executor.evaluate()(a,b).asnumpy()
print(np.sum(res))
ref = np.matmul(a,b)
print(np.sum(ref))
np.testing.assert_allclose(res, ref, rtol=1e-5)
```
Please note nn.batch_matmul works correctly for such cases not using libs:
```python
import tvm
from tvm import relay
from tvm.relay import create_executor, Any
import numpy as np
A = relay.var("A",shape=[1, Any(), Any()],dtype="float32")
B = relay.var("B",shape=[1, Any(), Any()],dtype="float32")
C = relay.nn.batch_matmul(A, relay.transpose(B, axes=[0,2,1]))
f = relay.Function([A, B], C)
mod = tvm.IRModule.from_expr(f)
for target in ["llvm"]:
dev = tvm.device(target,0)
executor = create_executor(kind="vm", mod=mod, device=dev, target=target)
a = np.random.uniform(size=[1,10,10]).astype("float32")
b = np.random.uniform(size=[1,10,10]).astype("float32")
res = executor.evaluate()(a,b).asnumpy()
print(np.sum(res))
ref = np.matmul(a,b)
print(np.sum(ref))
np.testing.assert_allclose(res, ref, rtol=1e-5)
```
--
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]
