tensorflow rnn variable seq length

import tensorflow as tf    
import numpy as np

if __name__ == '__main__':
  np.random.seed(1)  
  # the size of the hidden state for the lstm (notice the lstm uses 2x of this amount so actually lstm will have state of size 2)
  size = 1
  # 2 different sequences total
  batch_size= 2
  # the maximum steps for both sequences is 10
  n_steps = 10
  # each element of the sequence has dimension of 2
  seq_width = 2

  # the first input is to be stopped at 4 steps, the second at 6 steps
  e_stop = np.array([4,6])

  initializer = tf.random_uniform_initializer(-1,1) 

  # the sequences, has n steps of maximum size
  seq_input = tf.placeholder(tf.float32, [n_steps, batch_size, seq_width])
  # what timesteps we want to stop at, notice it's different for each batch hence dimension of [batch]
  early_stop = tf.placeholder(tf.int32, [batch_size])

  # inputs for rnn needs to be a list, each item being a timestep. 
  # we need to split our input into each timestep, and reshape it because split keeps dims by default  
  inputs = [tf.reshape(i, (batch_size, seq_width)) for i in tf.split(0, n_steps, seq_input)]

  cell = tf.nn.rnn_cell.LSTMCell(size, seq_width, initializer=initializer)  
  initial_state = cell.zero_state(batch_size, tf.float32)

  # ========= This is the most important part ==========
  # output will be of length 4 and 6
  # the state is the final state at termination (stopped at step 4 and 6)  
  outputs, state = tf.nn.rnn(cell, inputs, initial_state=initial_state, sequence_length=early_stop)

  # usual crap
  iop = tf.initialize_all_variables()
  session = tf.Session()
  session.run(iop)
  feed = {early_stop:e_stop, seq_input:np.random.rand(n_steps, batch_size, seq_width).astype('float32')}

  print "outputs, should be 2 things one of length 4 and other of 6"
  outs = session.run(outputs, feed_dict=feed)
  for xx in outs:
    print xx

  print "states, 2 things total both of size 2, which is the size of the hidden state"
  st = session.run(state, feed_dict=feed)
  print st

Actually, the len(outs) equals to n_steps instead of "2 things one of length 4 and other of 6". Have any ideas how to make this right?
Note that I replace tf.nn.rnn and tf.nn.rnn_cell.LSTMCell by rnn.rnn and LSTMCell imported from tensorflow.models.rnn and tensorflow.models.rnn.rnn_cell

Thanks much!

How much time does it take for you to run this code?
on tensorflow version 0.8.0 and GTX Titan X (12 GB memory) it takes me about 45 minutes to build the finish the first "run" of the two (which is responsible for building the graph). It takes 0.05 seconds to finish the 2nd run

Sorry if this question is too naive, but where is the training happening?