Merge branch 'main' into t5_pipeline_parallelism

megatron/training.py

@@ -193,7 +193,7 @@ def update_train_iters(args):
     print_rank_0('setting training iterations to {}'.format(args.train_iters))
 
 
-def get_model(model_provider_func, model_type):
+def get_model(model_provider_func, model_type, wrap_with_ddp=True):
     """Build the model."""
     args = get_args()
     args.model_type = model_type
@@ -272,22 +272,24 @@ def get_model(model_provider_func, model_type):
     if args.fp16 or args.bf16:
         model = [Float16Module(model_module, args) for model_module in model]
 
-    if args.DDP_impl == 'torch':
-        i = torch.cuda.current_device()
-        model = [torchDDP(model_module, device_ids=[i], output_device=i,
-                          process_group=mpu.get_data_parallel_group())
-                 for model_module in model]
-        return model
+    if wrap_with_ddp:
+        if args.DDP_impl == 'torch':
+            i = torch.cuda.current_device()
+            model = [torchDDP(model_module, device_ids=[i], output_device=i,
+                              process_group=mpu.get_data_parallel_group())
+                     for model_module in model]
 
-    if args.DDP_impl == 'local':
-        model = [LocalDDP(model_module,
-                          args.accumulate_allreduce_grads_in_fp32,
-                          args.use_contiguous_buffers_in_ddp)
-                 for model_module in model]
-        return model
+        elif args.DDP_impl == 'local':
+            model = [LocalDDP(model_module,
+                              args.accumulate_allreduce_grads_in_fp32,
+                              args.use_contiguous_buffers_in_local_ddp)
+                     for model_module in model]
+
+        else:
+            raise NotImplementedError('Unknown DDP implementation specified: '
+                                      '{}. Exiting.'.format(args.DDP_impl))
 
-    raise NotImplementedError('Unknown DDP implementation specified: {}. '
-                              'Exiting.'.format(args.DDP_impl))
+    return model
 
 
 def get_learning_rate_scheduler(optimizer):
@@ -380,11 +382,10 @@ def train_step(forward_step_func, data_iterator,
     timers = get_timers()
 
     # Set grad to zero.
-    if args.DDP_impl == 'local' and args.use_contiguous_buffers_in_ddp:
+    if args.DDP_impl == 'local' and args.use_contiguous_buffers_in_local_ddp:
         for partition in model:
             partition.zero_grad_buffer()
-    else:
-        optimizer.zero_grad()
+    optimizer.zero_grad()
 
     forward_backward_func = get_forward_backward_func()
     losses_reduced = forward_backward_func(
@@ -392,7 +393,7 @@ def train_step(forward_step_func, data_iterator,
         optimizer, timers, forward_only=False)
 
     # Empty unused memory
-    if args.empty_unused_memory_each_iter >= 1:
+    if args.empty_unused_memory_level >= 1:
         torch.cuda.empty_cache()
 
     # All-reduce if needed.
@@ -443,7 +444,7 @@ def train_step(forward_step_func, data_iterator,
         skipped_iter = 1
 
     # Empty unused memory
-    if args.empty_unused_memory_each_iter >= 2:
+    if args.empty_unused_memory_level >= 2:
         torch.cuda.empty_cache()
 
     if mpu.is_pipeline_last_stage(ignore_virtual=True):
@@ -755,7 +756,7 @@ def evaluate(forward_step_func, data_iterator, model, verbose=False):
                 timers=None, forward_only=True)
 
             # Empty unused memory
-            if args.empty_unused_memory_each_iter >= 1:
+            if args.empty_unused_memory_level >= 1:
                 torch.cuda.empty_cache()
 
             if mpu.is_pipeline_last_stage(ignore_virtual=True):

tasks/zeroshot_gpt/evaluate.py

@@ -205,7 +205,7 @@ def main():
             args.task))
 
     # Set up model and load checkpoint.
-    model = get_model(get_model_provider(eval_metric))
+    model = get_model(get_model_provider(eval_metric), wrap_with_ddp=False)
     if args.load is not None:
         _ = load_checkpoint(model, None, None)
 

tools/generate_samples_gpt.py → tools/run_text_generation_server.py

@@ -14,37 +14,31 @@
 # limitations under the License.
 
 """Sample Generate GPT"""
-
 import os
 import sys
 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
                                              os.path.pardir)))
-
+import socket
 from megatron import get_args
 from megatron import print_rank_0
-from megatron import get_tokenizer
 from megatron import mpu
 from megatron.checkpointing import load_checkpoint
 from megatron.initialize import initialize_megatron
 from megatron.model import GPTModel
 from megatron.training import get_model
-from megatron.text_generation_utils import generate_and_write_samples_unconditional
-from megatron.text_generation_utils import generate_samples_input_from_file
-from megatron.text_generation_utils import generate_samples_interactive
-
+from megatron.text_generation_server import MegatronServer
+from megatron.text_generation_utils import generate
+import torch
 
 def model_provider(pre_process=True, post_process=True):
     """Build the model."""
 
     print_rank_0('building GPT model ...')
-    model = GPTModel(num_tokentypes=0, parallel_output=False,
-                     pre_process=pre_process, post_process=post_process)
+    model = GPTModel(num_tokentypes=0, parallel_output=False, pre_process=pre_process, post_process=post_process)
 
     return model
 
-
 def add_text_generate_args(parser):
-    """Text generation arguments."""
     group = parser.add_argument_group(title='text generation')
 
     group.add_argument("--temperature", type=float, default=1.0,
@@ -57,26 +51,10 @@ def add_text_generate_args(parser):
                        help='Top k sampling.')
     group.add_argument("--out-seq-length", type=int, default=1024,
                        help='Size of the output generated text.')
-    group.add_argument("--sample-input-file", type=str, default=None,
-                       help='Get input from file instead of interactive mode, '
-                       'each line is an input.')
-    group.add_argument("--sample-output-file", type=str, default=None,
-                       help='Output file got from --sample-input-file')
-    group.add_argument("--num-samples", type=int, default=0,
-                       help='Number of samples to generate unconditionally, '
-                       'defaults to 0 and interactive conditional sampling')
-    group.add_argument("--genfile", type=str,
-                       help='Output file when generating unconditionally')
-    group.add_argument("--recompute", action='store_true',
-                       help='During generation recompute all attention '
-                       'instead of using previously computed keys/values.')
-
     return parser
 
 
-def main():
-    """Main program."""
-
+if __name__ == "__main__":
     initialize_megatron(extra_args_provider=add_text_generate_args,
                         args_defaults={'tokenizer_type': 'GPT2BPETokenizer',
                                        'no_load_rng': True,
@@ -86,27 +64,20 @@ def main():
     if args.num_layers_per_virtual_pipeline_stage is not None:
         print("Interleaved pipeline schedule is not yet supported for text generation.")
         exit()
-
-    # Set up model and load checkpoint.
-    model = get_model(model_provider)
+    # Set up model and load checkpoint
+    model = get_model(model_provider, wrap_with_ddp=False)
 
     if args.load is not None:
         _ = load_checkpoint(model, None, None)
 
     assert len(model) == 1, "Above condition should have caught this"
     model = model[0]
-
-    # Generate samples.
-    if args.num_samples == 0:
-        args.micro_batch_size = 1
-        if args.sample_input_file != None:
-            generate_samples_input_from_file(model)
-        else:
-            generate_samples_interactive(model)
-    else:
-        generate_and_write_samples_unconditional(model)
-
-
-if __name__ == "__main__":
-
-    main()
+    if mpu.is_pipeline_first_stage() and mpu.get_tensor_model_parallel_rank() == 0:
+        server = MegatronServer(model)
+        server.run("0.0.0.0")
+
+    while True:
+        choice = torch.cuda.LongTensor(1)
+        torch.distributed.broadcast(choice, 0)
+        if choice[0].item() == 0:
+            generate(model)

tools/text_generation_cli.py

+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+import json
+import sys
+import urllib2
+class PutRequest(urllib2.Request):
+    '''class to handling putting with urllib2'''
+
+    def get_method(self, *args, **kwargs):
+        return 'PUT'
+
+if __name__ == "__main__":
+    url = sys.argv[1]
+    while True:
+        sentence = raw_input("Enter prompt: ")
+        max_len = int(input("Enter number tokens output: "))
+        data = json.dumps({"sentences": [sentence], "max_len":max_len})
+        req = PutRequest(url, data, {'Content-Type': 'application/json'})
+        response = urllib2.urlopen(req)
+        resp_sentences = json.load(response)
+        print("Megatron Response: ")
+        print(resp_sentences["sentences"][0])