num_threads attribute isn't respected during evaluation

[bauks]

Environment info

Operating System: Ubuntu 14.04.4
CPU: 4x Intel Xeon 24 core processors
Python version: 2.7.6
lightgbm version 2.1.2 installed via pip

When calling predict or predict_proba on an LGBMClassifier model, it seems like the number of threads used corresponds to the number of threads most recently used when training a lightgbm model. Is there a way to change the number of threads used during evaluation? I've tried setting the model.num_threads attribute directly as well as calling model.set_params(num_threads=4) but neither seem to affect the number of threads used, while I seem to be able to influence the number of threads used when evaluating model A by training a new model B with a different number of threads.
For example:

import lightgbm as lgb
import numpy as np

x = np.random.random((100000, 1000))
y = np.random.randint(0, 2, size=100000)

m1 = lgb.LGBMClassifier(num_threads=12)
m1.fit(x, y) # 12 threads used

for _ in range(10):
    preds = m1.predict(x) # 12 threads used

m2 = lgb.LGBMClassifier(num_threads=24)
m2.fit(x, y) # 24 threads used

for _ in range(10):
    preds = m1.predict(x) # 24 threads used!

It seems like the issue could be resolved by calling omp_set_num_threads with the appropriate argument in the c api somewhere before the model is actually evaluated.

[StrikerRUS]

Please pass num_threads argument via **kwargs in predict method.

Refer to #1534.