Sample cod Error on Gpu: RuntimeError: cublasSgemm(dev.cublas_handle,

import numpy as np
import dynet_config
dynet_config.set_gpu()

import dynet as dy
from optparse import OptionParser
parser = OptionParser()
parser.add_option("--dynet-gpu", action="store_true")
class OurNetwork(object):
def __init__(self, pc):
self.pW = pc.add_parameters((10,30))
self.pB = pc.add_parameters(10)
self.lookup = pc.add_lookup_parameters((500,10))

def __call__(self, inputs):
lookup = self.lookup
emb_vectors = [lookup[i] for i in inputs]
net_input = dy.concatenate(emb_vectors)
net_output = dy.softmax((self.pW * net_input) + self.pB)
return net_output

def create_network_return_loss(self, inputs, expected_output):
dy.renew_cg()
out = self(inputs)
loss = -dy.log(dy.pick(out, expected_output))
return loss

def create_network_return_best(self, inputs):
dy.renew_cg()
out = self(inputs)
return np.argmax(out.npvalue())

dy.init()
dy.renew_cg()

m = dy.Model()

network = OurNetwork(m)

trainer = dy.SimpleSGDTrainer(m)

for epoch in range(50):
for inp,lbl in (([1,2,3],1), ([3,2,4],2)):
loss = network.create_network_return_loss(inp, lbl)
loss.value()
loss.backward()
trainer.update()
print(loss.value()) # need to run loss.value() for the forward prop

print('Predicted smallest element among {} is {}:'.format([1,2,3], network.create_network_return_best([1,2,3])))

==========================out info=================
[dynet] initializing CUDA
[dynet] CUDA driver/runtime versions are 10.0/10.0
Request for 1 GPU ...
[dynet] Device Number: 0
[dynet] Device name: GeForce RTX 2080 Ti
[dynet] Memory Clock Rate (KHz): 7000000
[dynet] Memory Bus Width (bits): 352
[dynet] Peak Memory Bandwidth (GB/s): 616
[dynet] Memory Free (GB): 1.65767/11.5229
[dynet]
[dynet] Device(s) selected: 0
[dynet] random seed: 3237171193
[dynet] allocating memory: 512MB
[dynet] memory allocation done.
WARNING: Attempting to initialize dynet twice. Ignoring duplicate initialization.
CUBLAS failure in cublasSgemm(dev.cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N, y.d.rows(), y.d.cols() * y.d.batch_elems(), l.d.cols(), dev.kSCALAR_ONE, l.v, l.d.rows(), r.v, r.d.rows(), acc_scalar, y.v, y.d.rows())
13
Traceback (most recent call last):
File "/data/jupyter/hongchao/text2structure/model2note/model/test_dy_gpu.py", line 57, in
loss.value()
File "_dynet.pyx", line 769, in _dynet.Expression.value
File "_dynet.pyx", line 783, in _dynet.Expression.value
RuntimeError: cublasSgemm(dev.cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N, y.d.rows(), y.d.cols() * y.d.batch_elems(), l.d.cols(), dev.kSCALAR_ONE, l.v, l.d.rows(), r.v, r.d.rows(), acc_scalar, y.v, y.d.rows())

Process finished with exit code 1

Start with the reproducer in test_dy_gpu.py and the traceback at loss.value(), then inspect the reported cublasSgemm failure alongside the CUDA initialization output and duplicate-initialization warning. Done means the sample runs through its GPU forward pass without the RuntimeError, or the failure's cause and required usage are documented.

Written by the indexing model from the issue text.