The run.py Template

Whether you compile on a CPU node using the csrun_cpu script, or run on the Cerebras system using the csrun_wse script, you must pass a full Python command along with its command line arguments, as an argument to the script.

This section presents an example Python template run.py with a detailed description of the supported options and flags. Note that this applies to the pipelinedworkflow, not for weight streaming.

Note

The run.py described below is an example template only. If you are developing in TensorFlow, you can organize your code whichever way that best suits you, as long as you use CerebrasEstimator. If you are developing in PyTorch you can use the run function, common to all implementations in the Cerebras Model Zoo git repository. You can find more about the run function in the adapting-pytorch-to-cs document.

For TensorFlow code, see the following diagram showing a simplified run.py example using the CerebrasEstimator and the CGC flow.

../../_images/run-py-cerebrasestimator-cgc.jpg

Fig. 1 Cerebras Graph Compiler for the CS system

Syntax

python run.py -p --params <PATH-TO-YAML-FILE> \
              -o --model_dir <PATH-TO-MODEL-DIR> \
              --cs_ip <CS system-IP-ADDRESS> \
              --steps <INTEGER> \
              --max_steps <INTEGER> \
              -m --mode <MODE-TO-RUN> \
              --compile_only \
              --device <SPECIFY-DEVICE-TO-RUN-ON> \
              --checkpoint_path <INITIALIZE WEIGHTS FROM> \

where:

Arguments

params

  • -p --params: Required. String. Path to the YAML file that contains the model parameters.

    For example: --params configs/params_bert_base_msl512.yaml.

model_dir

  • -o --model_dir: Optional. String. The location where your model and all the outputs such as checkpoints and event files are stored. If the directory exists, then the weights are loaded from the checkpoint file. Same as the model_dir passed to the tf.estimator. Default value is current directory of execution. See also tf.estimator.Estimator.

cs_ip

  • --cs_ip: Optional. The IP address of the CS system. Format should be IP-ADDRESS:PORT. Default value is None. This option is ignored on GPU.

steps

  • --steps: Optional. Integer. The number of steps to run the train mode. Runs for the specified number of steps either while starting from the beginning or while continuing from a checkpoint. Default value is None.

max_steps

  • --max_steps: Optional. Integer. The total number of steps to run the train mode, or for training in the train_and_eval mode. If the run is continuing from a checkpoint that was made at or after max_steps, then the run will stop. If the run is continuing from a checkpoint that was made at less than max_steps, then it will run for the number of steps remaining between the checkpoint and the max_steps. Default value is None.

eval_steps

  • --eval_steps: Integer. Optional on the GPU but required when running on the CS system. The total number of steps to run eval or eval_all modes, or for evaluation in train_and_eval mode. Runs once for the specified number.

mode

  • -m --mode: Required. String. Set the mode for your neural network. Allowed choices are:

    • train: In this mode the compiler will compile, and run the training on the CS system. If cs_ip is not specified, then runs the training on the CPU or GPU.

    • eval: In this mode the evaluation will run on CPU or GPU. For some neural network models, this is experimentally supported on the CS system.

    • eval_all: In this mode the evaluation will run on CPU or GPU for all the available checkpoints. This mode is not yet supported on the CS system.

    • train_and_eval: In this mode the training and evaluation will be run on CPU or GPU.

    • predict: In this mode the prediction (inference) will run on CPU or GPU. For some neural network models, this is experimentally supported on the CS system.

compilation flags

  • --validate_only: In this mode the compiler will stop after the kernel matching phase. Available in Pipelined execution mode for TensorFlow.

  • --compile_only: In this mode the compiler will continue after the kernel matching until it finishes. If the compile is successful it will generate the CS system bitstream. Available in Pipelined execution mode.

device

  • --device: Optional. String. This option should be used only to specify a GPU device. The compiler will compile on CPU, and will run on the GPU device specified in this setting.

    For example, --device /gpu:0 will run on the GPU 0.

checkpoint_path

  • --checkpoint_path: Optional. String. The weights are initialized from the checkpoint specified with this option. Default value is None. If this option is used with an initial checkpoint, and if the model_dir already contains a checkpoint, then the compiler will alert you to:

    • Either provide an empty model_dir so the weights are initialized using the value provided to this checkpoint_path option, or

    • Remove this option checkpoint_path in order to initialize weights from the model_dir.

Example: BERT run.py

Shown below is the run.py example code of the BERT model.

  1 # Example run.py script for BERT
  2
  3 # Copyright 2021 Cerebras Systems.
  4 #
  5 # Licensed under the Apache License, Version 2.0 (the "License");
  6 # you may not use this file except in compliance with the License.
  7 # You may obtain a copy of the License at
  8 #
  9 #     http://www.apache.org/licenses/LICENSE-2.0
 10 #
 11 # Unless required by applicable law or agreed to in writing, software
 12 # distributed under the License is distributed on an "AS IS" BASIS,
 13 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 14 # See the License for the specific language governing permissions and
 15 # limitations under the License.
 16
 17 import argparse
 18 import os
 19 import sys
 20
 21 import tensorflow as tf
 22
 23 # Relative path imports
 24 sys.path.append(os.path.join(os.path.dirname(__file__), '../../..'))
 25 from common_zoo.estimator.tf.cs_estimator import CerebrasEstimator
 26 from common_zoo.estimator.tf.run_config import CSRunConfig
 27 from common_zoo.run_utils import (
 28     check_env,
 29     create_warm_start_settings,
 30     get_csconfig,
 31     get_csrunconfig_dict,
 32     is_cs,
 33     save_params,
 34     save_predictions,
 35     update_params_from_args,
 36 )
 37 from transformers.bert.tf.data import eval_input_fn, train_input_fn
 38 from transformers.bert.tf.model import model_fn
 39 from transformers.bert.tf.utils import get_params
 40
 41
 42 def create_arg_parser(default_model_dir):
 43     """
 44     Create parser for command line args.
 45
 46     :param str default_model_dir: default value for the model_dir
 47     :returns: ArgumentParser
 48     """
 49     parser = argparse.ArgumentParser()
 50     parser.add_argument(
 51         "-p",
 52         "--params",
 53         required=True,
 54         help="Path to .yaml file with model parameters",
 55     )
 56     parser.add_argument(
 57         "-o",
 58         "--model_dir",
 59         default=default_model_dir,
 60         help="Model directory where checkpoints will be written. "
 61         + "If directory exists, weights are loaded from the checkpoint file.",
 62     )
 63     parser.add_argument(
 64         "--cs_ip",
 65         default=None,
 66         help="CS system IP address, defaults to None. Ignored on GPU.",
 67     )
 68     parser.add_argument(
 69         "--steps",
 70         type=int,
 71         default=None,
 72         help=(
 73             "Number of steps to run mode train."
 74             + " Runs repeatedly for the specified number."
 75         ),
 76     )
 77     parser.add_argument(
 78         "--max_steps",
 79         type=int,
 80         default=None,
 81         help=(
 82             "Number of total steps to run mode train or for defining training"
 83             + " configuration for train_and_eval. Runs incrementally till"
 84             + " the specified number."
 85         ),
 86     )
 87     parser.add_argument(
 88         "--eval_steps",
 89         type=int,
 90         default=None,
 91         help=(
 92             "Number of total steps to run mode eval, eval_all or for defining"
 93             + " eval configuration for train_and_eval. Runs once for"
 94             + " the specified number."
 95         ),
 96     )
 97     parser.add_argument(
 98         "-m",
 99         "--mode",
100         required=True,
101         choices=["train", "eval", "eval_all", "train_and_eval", "predict",],
102         help=(
103             "Can train, eval, eval_all, train_and_eval, or predict."
104             + "  Train, eval, and predict will compile and train if on CS system,"
105             + "  and just run locally (CPU/GPU) if not on CS system."
106             + "  train_and_eval will run locally."
107             + "  Eval_all will run eval locally for all available checkpoints."
108         ),
109     )
110     parser.add_argument(
111         "--validate_only",
112         action="store_true",
113         help="Compile model up to kernel matching.",
114     )
115     parser.add_argument(
116         "--compile_only",
117         action="store_true",
118         help="Compile model completely, generating compiled executables.",
119     )
120     parser.add_argument(
121         "--device",
122         default=None,
123         help="Force model to run on a specific device (e.g., --device /gpu:0)",
124     )
125     parser.add_argument(
126         "--checkpoint_path",
127         default=None,
128         help="Checkpoint to initialize weights from.",
129     )
130
131     return parser
132
133
134 def validate_runtime_params(params):
135     # check validate_only/compile_only
136     assert not (
137         params["validate_only"] and params["compile_only"]
138     ), "Please only use one of validate_only and compile_only."
139     if params["validate_only"] or params["compile_only"]:
140         assert params["mode"] in [
141             "train",
142             "eval",
143             "predict",
144         ], "Can only validate/compile model in train, eval, or predict mode."
145
146     # check for gpu optimization flags
147     if (
148         params["mode"] not in ["compile_only", "validate_only"]
149         and not is_cs(params)
150         and not params["enable_gpu_optimizations"]
151     ):
152         tf.compat.v1.logging.warn(
153             "Set enable_gpu_optimizations to True in training params "
154             "to improve GPU performance."
155         )
156
157
158 def run(
159     args,
160     params,
161     model_fn,
162     train_input_fn=None,
163     eval_input_fn=None,
164     predict_input_fn=None,
165     output_layer_name=None,
166 ):
167     """
168     Set up estimator and run based on mode
169
170     :params dict params: dict to handle all parameters
171     :params tf.estimator.EstimatorSpec model_fn: Model function to run with
172     :params tf.data.Dataset train_input_fn: Dataset to train with
173     :params tf.data.Dataset eval_input_fn: Dataset to validate against
174     :params tf.data.Dataset predict_input_fn: Dataset to run inference on
175     :params str output_layer_name: name of the output layer to be excluded
176         from weight initialization when performing fine-tuning.
177     """
178     # update and validate runtime params
179     runconfig_params = params["runconfig"]
180     update_params_from_args(args, runconfig_params)
181     validate_runtime_params(runconfig_params)
182     # save params for reproducibility
183     save_params(params, model_dir=runconfig_params["model_dir"])
184
185     # get cs-specific configs
186     cs_config = get_csconfig(params.get("csconfig", dict()))
187     # get runtime configurations
188     use_cs = is_cs(runconfig_params)
189     csrunconfig_dict = get_csrunconfig_dict(runconfig_params)
190
191     stack_params = dict()
192     if (
193         use_cs
194         or runconfig_params["validate_only"]
195         or runconfig_params["compile_only"]
196     ):
197         from cerebras.pb.stack.full_pb2 import FullConfig
198
199         config = FullConfig()
200         if params['train_input']['max_sequence_length'] <= 128:
201             config.matching.kernel.no_dcache_spill_splits = True
202         stack_params['config'] = config
203
204     # prep cs1 run environment, run config and estimator
205     check_env(runconfig_params)
206     est_config = CSRunConfig(
207         cs_ip=runconfig_params["cs_ip"],
208         cs_config=cs_config,
209         stack_params=stack_params,
210         **csrunconfig_dict,
211     )
212     warm_start_settings = create_warm_start_settings(
213         runconfig_params, exclude_string=output_layer_name
214     )
215     est = CerebrasEstimator(
216         model_fn=model_fn,
217         model_dir=runconfig_params["model_dir"],
218         config=est_config,
219         params=params,
220         warm_start_from=warm_start_settings,
221     )
222
223     # execute based on mode
224     if runconfig_params["validate_only"] or runconfig_params["compile_only"]:
225         if runconfig_params["mode"] == "train":
226             input_fn = train_input_fn
227             mode = tf.estimator.ModeKeys.TRAIN
228         elif runconfig_params["mode"] == "eval":
229             input_fn = eval_input_fn
230             mode = tf.estimator.ModeKeys.EVAL
231         else:
232             input_fn = predict_input_fn
233             mode = tf.estimator.ModeKeys.PREDICT
234         est.compile(
235             input_fn, validate_only=runconfig_params["validate_only"], mode=mode
236         )
237     elif runconfig_params["mode"] == "train":
238         est.train(
239             input_fn=train_input_fn,
240             steps=runconfig_params["steps"],
241             max_steps=runconfig_params["max_steps"],
242             use_cs=use_cs,
243         )
244     elif runconfig_params["mode"] == "eval":
245         est.evaluate(
246             input_fn=eval_input_fn,
247             checkpoint_path=runconfig_params["checkpoint_path"],
248             steps=runconfig_params["eval_steps"],
249             use_cs=use_cs,
250         )
251     elif runconfig_params["mode"] == "eval_all":
252         ckpt_list = tf.train.get_checkpoint_state(
253             runconfig_params["model_dir"]
254         ).all_model_checkpoint_paths
255         for ckpt in ckpt_list:
256             est.evaluate(
257                 eval_input_fn,
258                 checkpoint_path=ckpt,
259                 steps=runconfig_params["eval_steps"],
260                 use_cs=use_cs,
261             )
262     elif runconfig_params["mode"] == "train_and_eval":
263         train_spec = tf.estimator.TrainSpec(
264             input_fn=train_input_fn, max_steps=runconfig_params["max_steps"]
265         )
266         eval_spec = tf.estimator.EvalSpec(
267             input_fn=eval_input_fn,
268             steps=runconfig_params["eval_steps"],
269             throttle_secs=runconfig_params["throttle_secs"],
270         )
271         tf.estimator.train_and_evaluate(est, train_spec, eval_spec)
272     elif runconfig_params["mode"] == "predict":
273         sys_name = "cs" if use_cs else "tf"
274         file_to_save = f"predictions_{sys_name}_{est_config.task_id}.npz"
275         predictions = est.predict(
276             input_fn=predict_input_fn,
277             checkpoint_path=runconfig_params["checkpoint_path"],
278             num_samples=runconfig_params["predict_steps"],
279             use_cs=use_cs,
280         )
281         save_predictions(
282             model_dir=runconfig_params["model_dir"],
283             outputs=predictions,
284             name=file_to_save,
285         )
286
287
288 def main():
289     """
290     Main function
291     """
292     default_model_dir = os.path.join(
293         os.path.dirname(os.path.abspath(__file__)), "model_dir"
294     )
295     parser = create_arg_parser(default_model_dir)
296     args = parser.parse_args(sys.argv[1:])
297     params = get_params(args.params)
298     run(
299         args=args,
300         params=params,
301         model_fn=model_fn,
302         train_input_fn=train_input_fn,
303         eval_input_fn=eval_input_fn,
304     )
305
306
307 if __name__ == "__main__":
308     tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.INFO)
309     main()