diff --git a/compiler_opt/rl/compilation_runner.py b/compiler_opt/rl/compilation_runner.py
index 91060f80..6c6498dd 100644
--- a/compiler_opt/rl/compilation_runner.py
+++ b/compiler_opt/rl/compilation_runner.py
@@ -30,7 +30,7 @@
 import tensorflow as tf
 
 _COMPILATION_TIMEOUT = flags.DEFINE_integer(
-    'compilation_timeout', 60,
+    'compilation_timeout', 120,
     'Max duration (in seconds) after which we cancel any compilation job.')
 _QUIET = flags.DEFINE_bool(
     'quiet', True, 'Whether or not to compile quietly (hiding info logging)')
diff --git a/compiler_opt/rl/registry.py b/compiler_opt/rl/registry.py
index 85f8097a..a04d9275 100644
--- a/compiler_opt/rl/registry.py
+++ b/compiler_opt/rl/registry.py
@@ -33,6 +33,7 @@
 # to trigger gin registration.
 import compiler_opt.rl.inlining  # pylint: disable=unused-import
 import compiler_opt.rl.regalloc  # pylint: disable=unused-import
+import compiler_opt.rl.unroll  # pylint: disable=unused-import
 
 types = tfa.typing.types
 
diff --git a/compiler_opt/rl/unroll/__init__.py b/compiler_opt/rl/unroll/__init__.py
new file mode 100644
index 00000000..b292fb9f
--- /dev/null
+++ b/compiler_opt/rl/unroll/__init__.py
@@ -0,0 +1,38 @@
+# coding=utf-8
+# Copyright 2020 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Implementation of the 'loop unroll' problem."""
+
+import gin
+
+from compiler_opt.rl import problem_configuration
+from compiler_opt.rl.unroll import config
+from compiler_opt.rl.unroll import unroll_runner
+
+
+@gin.register(module='configs')
+class LoopUnrollConfig(problem_configuration.ProblemConfiguration):
+  """Expose the regalloc eviction components."""
+
+  def get_runner_type(self):
+    return unroll_runner.LoopUnrollRunner
+
+  def get_signature_spec(self):
+    return config.get_unroll_signature_spec()
+
+  def get_preprocessing_layer_creator(self):
+    return config.get_observation_processing_layer_creator()
+
+  def get_nonnormalized_features(self):
+    return config.get_nonnormalized_features()
diff --git a/compiler_opt/rl/unroll/config.py b/compiler_opt/rl/unroll/config.py
new file mode 100644
index 00000000..203f4eaf
--- /dev/null
+++ b/compiler_opt/rl/unroll/config.py
@@ -0,0 +1,76 @@
+# coding=utf-8
+# Copyright 2020 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Loop unroll training config."""
+
+import gin
+import tensorflow as tf
+from tf_agents.specs import tensor_spec
+from tf_agents.trajectories import time_step
+from compiler_opt.rl import feature_ops
+
+
+# pylint: disable=g-complex-comprehension
+@gin.configurable()
+def get_unroll_signature_spec():
+  """Returns (time_step_spec, action_spec) for LLVM loop unroll."""
+  # LINT.IfChange
+  observation_spec = dict(
+      (key, tf.TensorSpec(dtype=tf.int64, shape=(), name=key))
+      for key in ('loop_size', 'trip_count', 'is_innermost_loop',
+                  'preheader_blocksize', 'bb_count', 'num_of_loop_latch',
+                  'load_inst_count', 'store_inst_count', 'logical_inst_count',
+                  'cast_inst_count'))
+  reward_spec = tf.TensorSpec(dtype=tf.float32, shape=(), name='reward')
+  time_step_spec = time_step.time_step_spec(observation_spec, reward_spec)
+  action_spec = tensor_spec.BoundedTensorSpec(
+      dtype=tf.int64, shape=(), name='unroll_count')
+
+  return time_step_spec, action_spec
+
+
+@gin.configurable
+def get_observation_processing_layer_creator(quantile_file_dir=None,
+                                             with_sqrt=True,
+                                             with_z_score_normalization=True,
+                                             eps=1e-8):
+  """Wrapper for observation_processing_layer."""
+  quantile_map = feature_ops.build_quantile_map(quantile_file_dir)
+
+  def observation_processing_layer(obs_spec):
+    """Creates the layer to process observation given obs_spec."""
+
+    # I guess we discard rewards when observation?
+    if obs_spec.name in ('icache_pressure', 'latency'):
+      return tf.keras.layers.Lambda(feature_ops.discard_fn)
+
+    # for boolean features, use feature_ops.identity_fn
+    if obs_spec.name in ('is_innermost_loop'):
+      return tf.keras.layers.Lambda(feature_ops.identity_fn)
+
+    # Do we need to define some layer here to normalize 'loop_size'
+    # and instruction count features (e.g. 'load_inst_count').
+    # Bigger loops expect more instruction counts, and we need to
+    # normalize this?
+
+    quantile = quantile_map[obs_spec.name]
+    return tf.keras.layers.Lambda(
+        feature_ops.get_normalize_fn(quantile, with_sqrt,
+                                     with_z_score_normalization, eps))
+
+  return observation_processing_layer
+
+
+def get_nonnormalized_features():
+  return ['reward', 'is_innermost_loop']
diff --git a/compiler_opt/rl/unroll/gin_configs/behavioral_cloning_nn_agent.gin b/compiler_opt/rl/unroll/gin_configs/behavioral_cloning_nn_agent.gin
new file mode 100644
index 00000000..d38f63aa
--- /dev/null
+++ b/compiler_opt/rl/unroll/gin_configs/behavioral_cloning_nn_agent.gin
@@ -0,0 +1,32 @@
+import gin.tf.external_configurables
+import compiler_opt.rl.constant
+import compiler_opt.rl.gin_external_configurables
+import compiler_opt.rl.unroll.config
+import tf_agents.agents.behavioral_cloning.behavioral_cloning_agent
+import tf_agents.networks.q_network
+
+include 'compiler_opt/rl/unroll/gin_configs/common.gin'
+
+train_eval.agent_name=%constant.AgentName.BEHAVIORAL_CLONE
+train_eval.num_iterations=100000
+train_eval.batch_size=64
+train_eval.train_sequence_length=1
+
+unroll.config.get_observation_processing_layer_creator.with_sqrt = False
+unroll.config.get_observation_processing_layer_creator.with_z_score_normalization = False
+
+create_agent.policy_network = @q_network.QNetwork
+
+QNetwork.preprocessing_combiner=@tf.keras.layers.Concatenate()
+QNetwork.fc_layer_params=(40, 40, 20)
+QNetwork.dropout_layer_params=(0.2, 0.2, 0.2)
+QNetwork.activation_fn=@tf.keras.activations.relu
+
+tf.train.AdamOptimizer.learning_rate = 0.001
+tf.train.AdamOptimizer.epsilon = 0.0003125
+
+BehavioralCloningAgent.optimizer = @tf.train.AdamOptimizer()
+BehavioralCloningAgent.epsilon_greedy = 0.1
+BehavioralCloningAgent.gradient_clipping = None
+BehavioralCloningAgent.debug_summaries = True
+BehavioralCloningAgent.summarize_grads_and_vars = True
diff --git a/compiler_opt/rl/unroll/gin_configs/common.gin b/compiler_opt/rl/unroll/gin_configs/common.gin
new file mode 100644
index 00000000..7d213ba8
--- /dev/null
+++ b/compiler_opt/rl/unroll/gin_configs/common.gin
@@ -0,0 +1,11 @@
+config_registry.get_configuration.implementation=@configs.LoopUnrollConfig
+
+clang_path=None
+llvm_objcopy_path=None
+parse_reward_script_path=None
+latency_coefficient=None
+
+runners.LoopUnrollRunner.clang_path=%clang_path
+runners.LoopUnrollRunner.llvm_objcopy_path=%llvm_objcopy_path
+runners.LoopUnrollRunner.parse_reward_script_path=%parse_reward_script_path
+runners.LoopUnrollRunner.latency_coefficient=%latency_coefficient
diff --git a/compiler_opt/rl/unroll/unroll_runner.py b/compiler_opt/rl/unroll/unroll_runner.py
new file mode 100644
index 00000000..70753614
--- /dev/null
+++ b/compiler_opt/rl/unroll/unroll_runner.py
@@ -0,0 +1,194 @@
+# coding=utf-8
+# Copyright 2020 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Module for collect data of loop unroll."""
+
+import base64
+import io
+import os
+import tempfile
+from typing import Dict, Optional, Tuple
+
+import gin
+import tensorflow as tf
+
+from google.protobuf import struct_pb2  # pytype: disable=pyi-error
+from compiler_opt.rl import compilation_runner
+from compiler_opt.rl import corpus
+
+
+@gin.configurable(module='runners')
+class LoopUnrollRunner(compilation_runner.CompilationRunner):
+  """Class for collecting data for loop partial unroll.
+
+  Usage:
+  runner = LoopUnrollRunner(
+               clang_path, llvm_objcopy_path, parse_reward_script_path,
+              moving_average_decay_rate)
+  policy_reward = unroll.collect_data(
+      ir_path, tf_policy_path, default_reward, moving_average_reward)
+  """
+
+  def __init__(self, llvm_objcopy_path: str, parse_reward_script_path: str,
+               latency_coefficient: str, *args, **kwargs):
+    super().__init__(*args, **kwargs)
+    self._llvm_objcopy_path = llvm_objcopy_path
+    self._parse_reward_script_path = parse_reward_script_path
+    self._latency_coefficient = float(latency_coefficient)
+
+  def compile_fn(
+      self, module_spec: corpus.ModuleSpec, tf_policy_path: str,
+      reward_only: bool, cancellation_manager: Optional[
+          compilation_runner.WorkerCancellationManager]
+  ) -> Dict[str, Tuple[tf.train.SequenceExample, float]]:
+    """Run loop unroll for the given IR file under the given policy.
+
+    Args:
+      module_spec: a ModuleSpec.
+      tf_policy_path: path to TF policy directory on local disk.
+      reward_only: whether to only return reward (icache pressure and latency)
+      cancellation_manager: handler for early termination by killing any running
+      processes
+
+    Returns:
+      For loop unroll, the result is in module level. IWS and Latency is
+      already weighted by the probability to be executed, checkout
+      parse_reward.py and code embedded under AsmPrinter.cpp for more detail).
+
+      Since the reward is calculated at late stage in a compiler that is after
+      inlining some functions may be inlined and not be found for some loops,
+      so we sum all functions into a single float, reward_total.
+
+      The function returns in the format:
+      {
+        "loop1_key": (loop1_features, reward_total),
+        "loop2_key": (loop2_features, reward_total),
+        ...,
+        "loopN_key": (loopN_features, reward_total)
+      }
+      - reward_total: sum of IWS and Latency of all functions in this module
+
+    Early return:
+      The function early returns when the compiled module doesn't record any
+      logs or the log file doesn't record any loop. This happens when
+      `LoopUnrollPass` is not triggered or no loop triggered "partial unroll"
+      in the pass.
+    """
+    working_dir = tempfile.mkdtemp()
+
+    # The compiler will log input feature (loop properties) and decision
+    # (unroll count) into the specified log path
+    log_path = os.path.join(working_dir, 'log')
+
+    # The compilation will generate object files, and our augmentation under
+    # AsmPrinter.cpp will create section data `llvm_block_data`.
+    object_path = os.path.join(working_dir, 'object')
+    # llvm-objcopy extracts the section data from object to data
+    data_path = os.path.join(working_dir, 'data')
+    # Reward parsing script parses data into parsed_reward
+    parsed_reward_path = os.path.join(working_dir, 'parsed_reward')
+
+    try:
+      # Construct command to execute clang
+      command_line = []
+
+      # parameters for MLGO unroll
+      command_line.extend([self._clang_path] + list(module_spec.exec_cmd) + [
+          '-mllvm', '-mlgo-unroll-mode=training', '-mllvm',
+          '-mlgo-unroll-training-log=' +
+          log_path, '-mllvm', '-calc-reward', '-o', object_path
+      ])
+
+      # Under `training mode`...
+      # If model path is provided, compiler will use ModelUnderTrainingRunner
+      # Otherwise, compiler will use NoInferenceModelRunner
+      if tf_policy_path:
+        command_line.extend(
+            ['-mllvm', 'mlgo-unroll-train-model=' + tf_policy_path])
+
+      print('Command to execute clang: ', command_line)
+
+      # run clang
+      compilation_runner.start_cancellable_process(command_line,
+                                                   self._compilation_timeout,
+                                                   cancellation_manager)
+
+      # A module may not generate a log if none of the loops go into the
+      # LoopUnroll decision. Early return here if log_path cannot be found.
+      if not os.path.exists(log_path):
+        print('Early return, log file not found.')
+        return {}
+
+      # A log file may not have anything inside when none of the loops goes
+      # into PartialUnroll decision. Early return a log file is created but
+      # nothing inside.
+      if os.path.getsize(log_path) == 0:
+        print('Early return, log file contains nothing.')
+        return {}
+
+      # Run llvm-objcopy to get section data
+      command_line = [
+          self._llvm_objcopy_path,
+          '--dump-section=.llvm_block_data.=' + data_path, object_path
+      ]
+      print('Command to get section data: ', command_line)
+      compilation_runner.start_cancellable_process(command_line,
+                                                   self._compilation_timeout,
+                                                   cancellation_manager)
+
+      # Run parse_reward.py to get reward
+      command_line = [
+          self._parse_reward_script_path, data_path, parsed_reward_path
+      ]
+      print('Command to parse reward: ', command_line)
+      compilation_runner.start_cancellable_process(command_line,
+                                                   self._compilation_timeout,
+                                                   cancellation_manager)
+
+      # Sum rewards of all functions into a single float
+      reward_total = 0
+      with io.open(parsed_reward_path, 'r', encoding='utf-8') as reward_f:
+        for line in reward_f.readlines():
+          line = line[:-1]  # strip end-line
+          items = line.split(',')
+          assert len(items) == 3
+          # function_name = items[0] (commented out because currently unused)
+          iws = float(items[1])
+          latency = float(items[2])
+          reward_total = reward_total + (
+              iws + latency * self._latency_coefficient)
+
+      if reward_only:
+        return {'default': (None, reward_total)}
+
+      result = {}
+
+      # Read training log, fill them in to result.
+      sequence_examples = struct_pb2.Struct()
+      with io.open(log_path, 'rb') as log_f:
+        sequence_examples.ParseFromString(log_f.read())
+
+      for key, value in sequence_examples.fields.items():
+        entry = tf.train.SequenceExample()
+        entry.ParseFromString(base64.b64decode(value.string_value))
+
+        if not entry.HasField('feature_lists'):
+          continue
+
+        result[key] = (entry, reward_total)
+
+    finally:
+      tf.io.gfile.rmtree(working_dir)
+
+    return result
diff --git a/compiler_opt/tools/sparse_bucket_generator.py b/compiler_opt/tools/sparse_bucket_generator.py
index d8e66217..2bb95a04 100644
--- a/compiler_opt/tools/sparse_bucket_generator.py
+++ b/compiler_opt/tools/sparse_bucket_generator.py
@@ -170,7 +170,7 @@ def main(_) -> None:
   parser_fn = create_tfrecord_parser_fn(sequence_features)
   dataset = dataset.map(parser_fn, num_parallel_calls=tf.data.AUTOTUNE)
   data_list = np.array(list(dataset.as_numpy_iterator()), dtype=object)
-  data_list = np.transpose(data_list, [1, 0])
+  data_list = np.transpose(data_list, [1, 0, 2])
 
   with mp.Pool(FLAGS.parallelism) as pool:
     feature_names = list(sorted(sequence_features))