echuraev commented on issue #8700:
URL: https://github.com/apache/tvm/issues/8700#issuecomment-904706891
Hello @schell!
I took a look at your problem and I'm not able to reproduce the issue. I
tried on Apple M1 and x86. And on both machines everything works fine. Could
you please provide more details about your problem? Maybe some simple scripts
which reproduces the issue or something like that?
I used the following simple script for running some simple model on metal:
```python
import numpy as np
import math
import os
import tvm
from tvm import relay, auto_scheduler
from tvm import testing
from tvm.contrib import utils, xcode, coreml_runtime, graph_runtime
target = "metal"
target_host = "llvm -mtriple=arm64-apple-darwin20.5.0"
def _get_model(shape, dtype, var_names):
"""Return a model and any parameters it may have."""
a = relay.var(next(var_names), shape=shape, dtype=dtype)
out = relay.op.reduce.mean(a, 0)
params = {}
return out, params
def converter(shape):
print("Shape: {}".format(shape))
dtype = "float32"
# b, data
inputs = {"data": tvm.nd.array(np.random.uniform(-128, 127,
shape).astype(dtype))}
mod, params = _get_model(shape, dtype, iter(inputs))
if isinstance(mod, tvm.relay.expr.Call):
mod = tvm.IRModule.from_expr(mod)
print('mod: ', mod)
with tvm.transform.PassContext(opt_level=3):
graph_module = relay.build(mod['main'], target=target,
target_host=target_host, params=params)
#with auto_scheduler.ApplyHistoryBest("my_mean_model_metal"):
# with tvm.transform.PassContext(opt_level=3,
config={"relay.backend.use_auto_scheduler": True}):
# graph_module = relay.build(mod['main'], target=target,
target_host=target_host, params=params)
return graph_module
def run(graph_module):
ctx = tvm.metal(0)
m =
graph_runtime.graph_executor.GraphModule(graph_module["default"](ctx))
m.run()
if __name__ == "__main__":
shape = (2, 1)
gm = converter(shape)
run(gm)
```
--
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]
