rebase on dev and cherry-pick

Author: abhi-mosaic

composer/datasets/__init__.py

@@ -21,6 +21,7 @@
 """
 from composer.datasets.ade20k import ADE20kDatasetHparams as ADE20kDatasetHparams
 from composer.datasets.brats import BratsDatasetHparams as BratsDatasetHparams
+from composer.datasets.c4 import C4DatasetHparams as C4DatasetHparams
 from composer.datasets.cifar10 import CIFAR10DatasetHparams as CIFAR10DatasetHparams
 from composer.datasets.dataloader import DataloaderHparams as DataloaderHparams
 from composer.datasets.dataloader import WrappedDataLoader as WrappedDataLoader

composer/datasets/c4.py

@@ -2,6 +2,7 @@
 
 from composer.datasets.ade20k import ADE20kDatasetHparams
 from composer.datasets.brats import BratsDatasetHparams
+from composer.datasets.c4 import C4DatasetHparams
 from composer.datasets.cifar10 import CIFAR10DatasetHparams
 from composer.datasets.glue import GLUEHparams
 from composer.datasets.imagenet import ImagenetDatasetHparams
@@ -15,7 +16,8 @@
     "cifar10": CIFAR10DatasetHparams,
     "mnist": MNISTDatasetHparams,
     "lm": LMDatasetHparams,
-    "glue": GLUEHparams
+    "glue": GLUEHparams,
+    "c4": C4DatasetHparams,
 }
 
 

composer/datasets/dataset_registry.py

@@ -420,9 +420,10 @@ def validate(self):
         super().validate()
 
         if self.deepspeed is not None:
-            zero_stage = cast(int, self.deepspeed.get("zero_stage", 0))
+            self.deepspeed["zero_stage"] = cast(int, self.deepspeed.get("zero_stage", 0))
+            self.deepspeed["steps_per_print"] = cast(int, self.deepspeed.get("steps_per_print", 1e20))
 
-            if self.deterministic_mode and zero_stage > 0:
+            if self.deterministic_mode and self.deepspeed["zero_stage"] > 0:
                 raise ValueError("Deepspeed with zero stage > 0 is not compatible with deterministic mode")
 
             if isinstance(self.device, CPUDeviceHparams):

composer/trainer/trainer_hparams.py

@@ -0,0 +1,86 @@
+train_dataset:
+  c4:
+    split: train
+    max_samples: 3584000 # Compute-optimal 7.3e9 tok ~= 256[bs] * 14000[ba] * 2048[msl] = 3584000[sa] * 2048[msl]
+    max_seq_len: 2048
+    tokenizer_name: gpt2
+    group_method: concat
+    seed: 17
+    shuffle: true
+    drop_last: true
+val_dataset:
+  c4:
+    split: validation
+    max_samples: 102400 # Approx 100k samples
+    max_seq_len: 2048
+    tokenizer_name: gpt2
+    group_method: concat
+    seed: 17
+    shuffle: false
+    drop_last: false
+model:
+  gpt2:
+    use_pretrained: false
+    tokenizer_name: gpt2
+    model_config:
+      activation_function: gelu_new
+      architectures:
+        - GPT2LMHeadModel
+      attn_pdrop: 0.0
+      bos_token_id: 50256
+      embd_pdrop: 0.0
+      eos_token_id: 50256
+      initializer_range: 0.02
+      layer_norm_epsilon: 1.0e-05
+      model_type: gpt2
+      n_embd: 768
+      n_head: 12
+      n_inner: 3072
+      n_layer: 12
+      n_positions: 2048
+      resid_pdrop: 0.0
+      scale_attn_weights: true
+      summary_activation: null
+      summary_first_dropout: 0.0
+      summary_proj_to_labels: true
+      summary_type: cls_index
+      summary_use_proj: true
+      task_specific_params:
+        text-generation:
+          do_sample: true
+          max_length: 50
+      transformers_version: 4.16.2
+      use_cache: true
+      vocab_size: 50257
+optimizer:
+  decoupled_adamw:
+    lr: 6.0e-4
+    betas:
+      - 0.9
+      - 0.95
+    eps: 1.0e-08
+    weight_decay: 0.0
+schedulers:
+  - cosine_decay_with_warmup:
+      warmup_time: 0.01dur
+loggers:
+  - tqdm: {}
+max_duration: 1ep
+train_batch_size: 256 # 0.5e6 tok ~= 256[bs] * 2048[msl]
+grad_accum: 2 # 256[bs] / 8[devices] / 16[per_gpu_microbatch_size] = 2[ga], assuming 8xA100-80GB
+eval_batch_size: 128 # 128[bs] / 8[devices] = 16[per_gpu_microbatch_size], assuming 8xA100-80GB
+seed: 17
+device:
+  gpu: {}
+dataloader:
+  pin_memory: true
+  persistent_workers: true
+  num_workers: 1
+  timeout: 0
+  prefetch_factor: 2
+deepspeed:
+  zero_stage: 0
+precision: fp16
+grad_clip_norm: 1.0
+validate_every_n_batches: 1000
+validate_every_n_epochs: 1

composer/yamls/models/gpt3_125m.yaml

@@ -4,40 +4,55 @@
 
 import pytest
 
-from composer.datasets import (ADE20kDatasetHparams, BratsDatasetHparams, CIFAR10DatasetHparams, DataloaderHparams,
-                               DatasetHparams, GLUEHparams, ImagenetDatasetHparams, LMDatasetHparams,
+from composer.datasets import (ADE20kDatasetHparams, BratsDatasetHparams, C4DatasetHparams, CIFAR10DatasetHparams,
+                               DataloaderHparams, DatasetHparams, GLUEHparams, ImagenetDatasetHparams, LMDatasetHparams,
                                MNISTDatasetHparams, SyntheticHparamsMixin)
 from composer.trainer.trainer_hparams import dataset_registry
 
 # for testing, we provide values for required hparams fields
 # to initialize test hparams objects
 default_required_fields: Dict[Type[DatasetHparams], Callable[[], DatasetHparams]] = {
     #  hparams with empty dicts have no required fields
-    CIFAR10DatasetHparams: lambda: CIFAR10DatasetHparams(
-        is_train=False,
-        download=False,
-    ),
-    ADE20kDatasetHparams: lambda: ADE20kDatasetHparams(is_train=False),
-    BratsDatasetHparams: lambda: BratsDatasetHparams(is_train=False,),
-    ImagenetDatasetHparams: lambda: ImagenetDatasetHparams(
-        is_train=False,
-        crop_size=224,
-        resize_size=-1,
-    ),
-    MNISTDatasetHparams: lambda: MNISTDatasetHparams(
-        is_train=False,
-        download=False,
-    ),
-    LMDatasetHparams: lambda: LMDatasetHparams(
-        datadir=["hello"],
-        split='train',
-        tokenizer_name='gpt2',
-    ),
-    GLUEHparams: lambda: GLUEHparams(
-        task="rte",
-        tokenizer_name="bert-base-uncased",
-        split="train",
-    ),
+    CIFAR10DatasetHparams:
+        lambda: CIFAR10DatasetHparams(
+            is_train=False,
+            download=False,
+        ),
+    ADE20kDatasetHparams:
+        lambda: ADE20kDatasetHparams(is_train=False),
+    BratsDatasetHparams:
+        lambda: BratsDatasetHparams(is_train=False,),
+    ImagenetDatasetHparams:
+        lambda: ImagenetDatasetHparams(
+            is_train=False,
+            crop_size=224,
+            resize_size=-1,
+        ),
+    MNISTDatasetHparams:
+        lambda: MNISTDatasetHparams(
+            is_train=False,
+            download=False,
+        ),
+    LMDatasetHparams:
+        lambda: LMDatasetHparams(
+            datadir=["hello"],
+            split='train',
+            tokenizer_name='gpt2',
+        ),
+    GLUEHparams:
+        lambda: GLUEHparams(
+            task="rte",
+            tokenizer_name="bert-base-uncased",
+            split="train",
+        ),
+    C4DatasetHparams:
+        lambda: C4DatasetHparams(
+            split="train",
+            max_samples=1000,
+            max_seq_len=100,
+            tokenizer_name="gpt2",
+            group_method="concat",
+        ),
 }