[WIP] Support new accelerate api

README.md

@@ -33,7 +33,7 @@ Lastly, if you want to study the effect of multitask prompted training (a.k.a. i
 - T-Zero ++: https://huggingface.co/bigscience/T0pp
 - T-Zero Single Prompt: https://huggingface.co/bigscience/T0_single_prompt
 - T-Zero Original Task Only: https://huggingface.co/bigscience/T0_original_task_only
-- T-Zero 3B: https://huggingface.co/bigscience/T0_3B
+- T-Zero 3B: https://huggingface.co/bigscience/T0_3Bx§
 
 ## Citation
 

debug/results.json

@@ -0,0 +1,8 @@
+{
+    "dataset_name": "super_glue",
+    "dataset_config_name": "cb",
+    "template_name": "GPT-3 style",
+    "evaluation": {
+        "accuracy": 0.39285714285714285
+    }
+}

evaluation/run_eval.py

@@ -51,7 +51,6 @@ def parse_args():
     parser.add_argument(
         "--dataset_name",
         type=str,
-        default=None,
         help="The name of the dataset to use (via the datasets library).",
         required=True,
     )
@@ -61,12 +60,17 @@ def parse_args():
         default=None,
         help="The configuration name of the dataset to use (via the datasets library).",
     )
+    parser.add_argument(
+        "--template_config_name",
+        type=str,
+        default=None,
+        help="The name of the dataset_config_name of the template we want to use, example: use XNLI En prompts for XNLI Fr",
+    )
     parser.add_argument(
         "--template_name",
         type=str,
         default=None,
-        help="The template/prompt name",
-        required=True,
+        help="The template/prompt name. If None, we run all templates.",
     )
     parser.add_argument(
         "--max_length",
@@ -128,115 +132,38 @@ def parse_args():
         action="store_true",
         help="Activate debug mode and run training only with a subset of data.",
     )
-    parser.add_argument(
-        "--parallelize",
-        action="store_true",
-        help=(
-            "If passed, will call `model.parallelize` which splits the model on all GPUs available when applicable (model parallelism). "
-            "Note that this feature is still experimental in HF Transformers."
-        ),
-    )
-    args = parser.parse_args()
-
-    return args
 
+    args = parser.parse_args()
 
-def main():
-    args = parse_args()
+    # TODO @thomasw21 hack!
+    if args.dataset_config_name == "None":
+        args.dataset_config_name = None
 
-    # Initialize the accelerator. We will let the accelerator handle device placement for us.
-    accelerator = Accelerator()
-    # Make one log on every process with the configuration for debugging.
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO,
-    )
-    logger.info(accelerator.state)
-
-    # Setup logging, we only want one process per machine to log things on the screen.
-    # accelerator.is_local_main_process is only True for one process per machine.
-    logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
-    if accelerator.is_local_main_process:
-        datasets.utils.logging.set_verbosity_warning()
-        transformers.utils.logging.set_verbosity_info()
-    else:
-        datasets.utils.logging.set_verbosity_error()
-        transformers.utils.logging.set_verbosity_error()
+    return args
 
+def run_template(template_name, prompts, model, tokenizer, raw_datasets, accelerator: Accelerator, args):
 
     # Handle the output directory creation
-    if accelerator.is_main_process:
-        os.makedirs(args.output_dir, exist_ok=True)
-    accelerator.wait_for_everyone()
-
-    # In distributed evaluation, the load_dataset function guarantee that only one local process can concurrently
-    # download the dataset.
-    if args.dataset_name is not None:
-        # Downloading and loading a dataset from the hub.
-        if args.dataset_name == "anli":
-            raw_datasets = load_dataset(args.dataset_name, split=args.dataset_config_name)
-        else:
-            raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name, split="validation")
-    #TODO(Victor): enable loading pre-processed dataset from https://huggingface.co/datasets/bigscience/P3
-
-    # Trim a number of evaluation examples
-    if args.debug:
-        raw_datasets = raw_datasets.select(range(min(len(raw_datasets),100)))
-
-    column_names = raw_datasets.column_names
-
-
-    # Load pretrained model and tokenizer
-    #
-    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
-    # download model & vocab.
-    if args.config_name:
-        config = AutoConfig.from_pretrained(args.config_name)
-    elif args.model_name_or_path:
-        config = AutoConfig.from_pretrained(args.model_name_or_path)
-    else:
-        raise ValueError(
-            "Either `args.config_name` or `args.model_name_or_path` should be provided."
-        )
-
-    if args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
-    elif args.model_name_or_path:
-        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
-    else:
-        raise ValueError(
-            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
-            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+    result_dir = None
+    if args.output_dir is not None and accelerator.is_main_process:
+        paths = [
+            args.dataset_name,
+            args.dataset_config_name,
+            template_name,
+        ]
+        result_dir = os.path.join(
+            args.output_dir,
+            *[path.replace(" ", "_") for path in paths if path is not None]
         )
+        os.makedirs(result_dir, exist_ok=True)
 
-    if tokenizer.pad_token is None:
-        for token in [tokenizer.eos_token, tokenizer.bos_token, tokenizer.sep_token]:
-            if token is not None:
-                tokenizer.pad_token = token
-        if tokenizer.pad_token is None:
-            raise ValueError("Please define a pad token id.")
-
+    template = prompts[template_name]
 
-    model = ModelBase.from_config(
-        config=config,
-        model_name_or_path=args.model_name_or_path,
-        parallelize=args.parallelize
-    )
 
     # Preprocessing the datasets.
     # First we tokenize all the texts.
     padding = "max_length" if args.pad_to_max_length else False
-
-    # Get the prompt to apply and the possible targets.
-    # TODO(Victor): If pulling from pre-processed data, remove this logic.
-    prompts = DatasetTemplates(
-        f"{args.dataset_name}"
-        if args.dataset_config_name is None
-        else f"{args.dataset_name}/{args.dataset_config_name}"
-    )
-    template = prompts[args.template_name]
-
+    column_names = raw_datasets.column_names
     def preprocess_function(examples):
         bs = len(examples[column_names[0]])
 
@@ -265,8 +192,9 @@ def preprocess_function(examples):
         tokenized_targets = [
             tokenizer(
                 ans_choi,
-                padding=True,
-                max_length=args.target_max_length,
+                # padding is on the right here.
+                padding=False,
+                max_length=args.max_length,
                 truncation=True,
             )
             for ans_choi in answer_choices_texts
@@ -319,17 +247,16 @@ def preprocess_function(examples):
 
     eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)
 
-
-    # Use the device given by the `accelerator` object.
-    if not args.parallelize:
-        model.to(accelerator.device)
-
     # Prepare everything with our `accelerator`.
     eval_dataloader = accelerator.prepare(eval_dataloader)
 
-
     # Metrics
-    metric = load_metric("accuracy")
+    metric = load_metric(
+        "accuracy",
+        process_id=accelerator.process_index,
+        num_process=accelerator.num_processes,
+        experiment_id=f"{args.dataset_name}_{args.dataset_config_name}_{args.template_name}"
+    )
 
     # Eval!
     total_batch_size = args.per_device_eval_batch_size * accelerator.num_processes
@@ -359,14 +286,119 @@ def preprocess_function(examples):
     results = {
         "dataset_name": args.dataset_name,
         "dataset_config_name": args.dataset_config_name,
-        "template_name": args.template_name,
-        "evaluation": eval_metric
+        "template_name": template_name,
+        "evaluation": eval_metric,
+        "arguments": str(args)
     }
     if accelerator.is_main_process:
-        if args.output_dir is not None:
-            with open(os.path.join(args.output_dir, "results.json"), "w") as f:
-                json.dump(results, f, indent=4)
+        if result_dir is not None:
+            with open(os.path.join(result_dir, "results.json"), "w") as f:
+                json.dump(results, f, indent=2)
+
+def main():
+    args = parse_args()
+
+    # Initialize the accelerator. We will let the accelerator handle device placement for us.
+    accelerator = Accelerator()
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state)
+
+    # Setup logging, we only want one process per machine to log things on the screen.
+    # accelerator.is_local_main_process is only True for one process per machine.
+    logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
+    if accelerator.is_local_main_process:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        transformers.utils.logging.set_verbosity_error()
+
+    accelerator.wait_for_everyone()
+
+    # In distributed evaluation, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    # Downloading and loading a dataset from the hub.
+    if args.dataset_name == "anli":
+        raw_datasets = load_dataset(args.dataset_name, split=args.dataset_config_name)
+    else:
+        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name, split="validation")
+    #TODO(Victor): enable loading pre-processed dataset from https://huggingface.co/datasets/bigscience/P3
+
+    # Trim a number of evaluation examples
+    if args.debug:
+        raw_datasets = raw_datasets.select(range(min(len(raw_datasets),100)))
+
+    # Load pretrained model and tokenizer
+    #
+    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if args.config_name:
+        config = AutoConfig.from_pretrained(args.config_name)
+    elif args.model_name_or_path:
+        config = AutoConfig.from_pretrained(args.model_name_or_path)
+    else:
+        raise ValueError(
+            "Either `args.config_name` or `args.model_name_or_path` should be provided."
+        )
+
+    if args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer, padding_side="left")
+    elif args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer, padding_side="left")
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if tokenizer.pad_token is None:
+        for token in [tokenizer.eos_token, tokenizer.bos_token, tokenizer.sep_token]:
+            if token is not None:
+                tokenizer.pad_token = token
+        if tokenizer.pad_token is None:
+            raise ValueError("Please define a pad token id.")
+
+
+    model = ModelBase.from_config(
+        config=config,
+        model_name_or_path=args.model_name_or_path
+    )
+    model = accelerator.prepare_model(model)
 
+    # Get the prompt to apply and the possible targets.
+    # TODO(Victor): If pulling from pre-processed data, remove this logic.
+
+    if args.dataset_config_name is None or args.template_config_name is None or args.dataset_name == "anli":
+        prompt_dataset_name = f"{args.dataset_name}"
+    elif args.template_config_name is not None:
+        prompt_dataset_name = f"{args.dataset_name}/{args.template_config_name}"
+    else:
+        prompt_dataset_name = f"{args.dataset_name}/{args.dataset_config_name}"
+
+    prompts = DatasetTemplates(
+        prompt_dataset_name
+    )
+
+    if args.template_name is not None:
+        template_names = [args.template_name]
+    else:
+        template_names = prompts.all_template_names
+
+    for template_name in template_names:
+        run_template(
+            template_name=template_name,
+            prompts=prompts,
+            model=model,
+            tokenizer=tokenizer,
+            raw_datasets=raw_datasets,
+            accelerator=accelerator,
+            args=args
+        )
 
 if __name__ == "__main__":
     main()

t0/model.py

@@ -27,7 +27,7 @@ def from_config(config, **kwargs) -> "ModelBase":
         raise NotImplementedError
 
 class EncoderDecoderModel(ModelBase):
-    def __init__(self, config, model_name_or_path: Optional[str], parallelize: bool, **kwargs):
+    def __init__(self, config, model_name_or_path: Optional[str], **kwargs):
         """
 
         Args:
@@ -46,11 +46,9 @@ def __init__(self, config, model_name_or_path: Optional[str], parallelize: bool,
             )
         else:
             logger.info("Training new model from scratch")
-            self._model = AutoModelForSeq2SeqLM.from_config(config)
-
-        if parallelize:
-            assert torch.cuda.is_available(), "You need at least 1 GPU to call `parallelize` (even though if there is only 1 GPU, there won't be any model parallelism)."
-            self._model.parallelize()
+            self._model = AutoModelForSeq2SeqLM.from_config(
+                config,
+            )
 
 
     def forward(self, batch) -> torch.Tensor:
@@ -78,19 +76,23 @@ def __init__(self, config, model_name_or_path: Optional[str], **kwargs):
             )
         else:
             logger.info("Training new model from scratch")
-            self._model = AutoModelForCausalLM.from_config(config)
+            self._model = AutoModelForCausalLM.from_config(
+                config,
+            )
 
     def forward(self, batch):
+        device = batch["input_ids"].device
         _, prefix_length = batch["input_ids"].shape
+
         model_inputs = {
             "input_ids": torch.cat([batch["input_ids"], batch["labels"]], dim=-1),
             "attention_mask": torch.cat([batch["attention_mask"], batch["labels_attention_mask"]], dim=-1),
         }
         # Set position ids correctly to take care of padding tokens between inputs_ids and labels
-        # Empty attention_mask is a forbidden value, ie full of zeros. In fact the first element should be 1 as the input
-        #   cannot be empty
-        assert torch.all(model_inputs["attention_mask"][:,0] == 1), "First element in the attention mask should be 1."
-        position_ids = torch.cumsum(model_inputs["attention_mask"].to(torch.long), dim=-1) - 1
+        position_ids = torch.maximum(
+            torch.cumsum(model_inputs["attention_mask"].to(torch.long), dim=-1) - 1,
+            torch.zeros(1, dtype=torch.long, device=device)[None, None]
+        )
         model_inputs["position_ids"] = position_ids
 
         logits = self._model(**model_inputs).logits[:, prefix_length-1:-1]