update

training/consts.py

@@ -0,0 +1,74 @@
+DEFAULT_INPUT_MODEL = "EleutherAI/pythia-6.9b"
+SUGGESTED_INPUT_MODELS = [
+    "EleutherAI/pythia-2.8b",
+    "EleutherAI/pythia-6.9b",
+    "EleutherAI/pythia-12b",
+    "EleutherAI/gpt-j-6B",
+]
+INTRO_BLURB = (
+    "Below is an instruction that describes a task. Write a response that appropriately completes the request."
+)
+INSTRUCTION_KEY = "### Instruction:"
+INPUT_KEY = "Input:"
+RESPONSE_KEY = "### Response:"
+END_KEY = "### End"
+RESPONSE_KEY_NL = f"{RESPONSE_KEY}\n"
+DEFAULT_SEED = 42
+
+# This is a training prompt that does not contain an input string.  The instruction by itself has enough information
+# to respond.  For example, the instruction might ask for the year a historic figure was born.
+PROMPT_NO_INPUT_FORMAT = """{intro}
+
+{instruction_key}
+{instruction}
+
+{response_key}
+{response}
+
+{end_key}""".format(
+    intro=INTRO_BLURB,
+    instruction_key=INSTRUCTION_KEY,
+    instruction="{instruction}",
+    response_key=RESPONSE_KEY,
+    response="{response}",
+    end_key=END_KEY,
+)
+
+# This is a training prompt that contains an input string that serves as context for the instruction.  For example,
+# the input might be a passage from Wikipedia and the intruction is to extract some information from it.
+PROMPT_WITH_INPUT_FORMAT = """{intro}
+
+{instruction_key}
+{instruction}
+
+{input_key}
+{input}
+
+{response_key}
+{response}
+
+{end_key}""".format(
+    intro=INTRO_BLURB,
+    instruction_key=INSTRUCTION_KEY,
+    instruction="{instruction}",
+    input_key=INPUT_KEY,
+    input="{input}",
+    response_key=RESPONSE_KEY,
+    response="{response}",
+    end_key=END_KEY,
+)
+
+# This is the prompt that is used for generating responses using an already trained model.  It ends with the response
+# key, where the job of the model is to provide the completion that follows it (i.e. the response itself).
+PROMPT_FOR_GENERATION_FORMAT = """{intro}
+
+{instruction_key}
+{instruction}
+
+{response_key}
+""".format(
+    intro=INTRO_BLURB,
+    instruction_key=INSTRUCTION_KEY,
+    instruction="{instruction}",
+    response_key=RESPONSE_KEY,
+)

training/generate.py

@@ -0,0 +1,239 @@
+import logging
+import re
+from typing import List, Tuple
+
+import numpy as np
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    Pipeline,
+    PreTrainedModel,
+    PreTrainedTokenizer,
+)
+
+from transformers.utils import is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+
+from .consts import END_KEY, PROMPT_FOR_GENERATION_FORMAT, RESPONSE_KEY
+
+logger = logging.getLogger(__name__)
+
+
+def load_model_tokenizer_for_generate(
+    pretrained_model_name_or_path: str,
+) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
+    """Loads the model and tokenizer so that it can be used for generating responses.
+
+    Args:
+        pretrained_model_name_or_path (str): name or path for model
+
+    Returns:
+        Tuple[PreTrainedModel, PreTrainedTokenizer]: model and tokenizer
+    """
+    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, padding_side="left", cache_dir="/media/siiva/DataStore/LLMs/cache/dollyV2")
+    model = AutoModelForCausalLM.from_pretrained(
+        pretrained_model_name_or_path, device_map="auto", trust_remote_code=True, cache_dir="/media/siiva/DataStore/LLMs/cache/dollyV2"
+    )
+    return model, tokenizer
+
+
+def get_special_token_id(tokenizer: PreTrainedTokenizer, key: str) -> int:
+    """Gets the token ID for a given string that has been added to the tokenizer as a special token.
+
+    When training, we configure the tokenizer so that the sequences like "### Instruction:" and "### End" are
+    treated specially and converted to a single, new token.  This retrieves the token ID each of these keys map to.
+
+    Args:
+        tokenizer (PreTrainedTokenizer): the tokenizer
+        key (str): the key to convert to a single token
+
+    Raises:
+        RuntimeError: if more than one ID was generated
+
+    Returns:
+        int: the token ID for the given key
+    """
+    token_ids = tokenizer.encode(key)
+    if len(token_ids) > 1:
+        raise RuntimeError(f"Expected only a single token for '{key}' but found {token_ids}")
+    return token_ids[0]
+
+
+class InstructionTextGenerationPipeline(Pipeline):
+    def __init__(
+        self, *args, do_sample: bool = True, max_new_tokens: int = 256, top_p: float = 0.92, top_k: int = 0, **kwargs
+    ):
+        """Initialize the pipeline
+
+        Args:
+            do_sample (bool, optional): Whether or not to use sampling. Defaults to True.
+            max_new_tokens (int, optional): Max new tokens after the prompt to generate. Defaults to 128.
+            top_p (float, optional): If set to float < 1, only the smallest set of most probable tokens with
+                probabilities that add up to top_p or higher are kept for generation. Defaults to 0.92.
+            top_k (int, optional): The number of highest probability vocabulary tokens to keep for top-k-filtering.
+                Defaults to 0.
+        """
+        super().__init__(*args, do_sample=do_sample, max_new_tokens=max_new_tokens, top_p=top_p, top_k=top_k,
+                         **kwargs)
+
+    def _sanitize_parameters(self,
+                             return_full_text: bool = None,
+                             **generate_kwargs):
+        preprocess_params = {}
+
+        # newer versions of the tokenizer configure the response key as a special token.  newer versions still may
+        # append a newline to yield a single token.  find whatever token is configured for the response key.
+        tokenizer_response_key = next(
+            (token for token in self.tokenizer.additional_special_tokens if token.startswith(RESPONSE_KEY)), None
+        )
+
+        response_key_token_id = None
+        end_key_token_id = None
+        if tokenizer_response_key:
+            try:
+                response_key_token_id = get_special_token_id(self.tokenizer, tokenizer_response_key)
+                end_key_token_id = get_special_token_id(self.tokenizer, END_KEY)
+
+                # Ensure generation stops once it generates "### End"
+                generate_kwargs["eos_token_id"] = end_key_token_id
+            except ValueError:
+                pass
+
+        forward_params = generate_kwargs
+        postprocess_params = {
+            "response_key_token_id": response_key_token_id,
+            "end_key_token_id": end_key_token_id
+        }
+
+        if return_full_text is not None:
+            postprocess_params["return_full_text"] = return_full_text
+
+        return preprocess_params, forward_params, postprocess_params
+
+    def preprocess(self, instruction_text, **generate_kwargs):
+        prompt_text = PROMPT_FOR_GENERATION_FORMAT.format(instruction=instruction_text)
+        inputs = self.tokenizer(
+            prompt_text,
+            return_tensors="pt",
+        )
+        inputs["prompt_text"] = prompt_text
+        inputs["instruction_text"] = instruction_text
+        return inputs
+
+    def _forward(self, model_inputs, **generate_kwargs):
+        input_ids = model_inputs["input_ids"]
+        attention_mask = model_inputs.get("attention_mask", None)
+
+        if input_ids.shape[1] == 0:
+            input_ids = None
+            attention_mask = None
+            in_b = 1
+        else:
+            in_b = input_ids.shape[0]
+
+        generated_sequence = self.model.generate(
+            input_ids=input_ids.to(self.model.device),
+            attention_mask=attention_mask,
+            pad_token_id=self.tokenizer.pad_token_id,
+            **generate_kwargs,
+        )
+
+        out_b = generated_sequence.shape[0]
+        if self.framework == "pt":
+            generated_sequence = generated_sequence.reshape(in_b, out_b // in_b, *generated_sequence.shape[1:])
+        elif self.framework == "tf":
+            generated_sequence = tf.reshape(generated_sequence, (in_b, out_b // in_b, *generated_sequence.shape[1:]))
+
+        instruction_text = model_inputs.pop("instruction_text")
+        return {"generated_sequence": generated_sequence, "input_ids": input_ids, "instruction_text": instruction_text}
+
+    def postprocess(self, model_outputs, response_key_token_id, end_key_token_id, return_full_text: bool = False):
+
+        generated_sequence = model_outputs["generated_sequence"][0]
+        instruction_text = model_outputs["instruction_text"]
+
+        generated_sequence: List[List[int]] = generated_sequence.numpy().tolist()
+        records = []
+        for sequence in generated_sequence:
+
+            # The response will be set to this variable if we can identify it.
+            decoded = None
+
+            # If we have token IDs for the response and end, then we can find the tokens and only decode between them.
+            if response_key_token_id and end_key_token_id:
+                # Find where "### Response:" is first found in the generated tokens.  Considering this is part of the
+                # prompt, we should definitely find it.  We will return the tokens found after this token.
+                try:
+                    response_pos = sequence.index(response_key_token_id)
+                except ValueError:
+                    logger.warn(f"Could not find response key {response_key_token_id} in: {sequence}")
+                    response_pos = None
+
+                if response_pos:
+                    # Next find where "### End" is located.  The model has been trained to end its responses with this
+                    # sequence (or actually, the token ID it maps to, since it is a special token).  We may not find
+                    # this token, as the response could be truncated.  If we don't find it then just return everything
+                    # to the end.  Note that even though we set eos_token_id, we still see the this token at the end.
+                    try:
+                        end_pos = sequence.index(end_key_token_id)
+                    except ValueError:
+                        end_pos = None
+
+                    decoded = self.tokenizer.decode(sequence[response_pos + 1 : end_pos]).strip()
+
+            if not decoded:
+                # Otherwise we'll decode everything and use a regex to find the response and end.
+
+                fully_decoded = self.tokenizer.decode(sequence)
+
+                # The response appears after "### Response:".  The model has been trained to append "### End" at the
+                # end.
+                m = re.search(r"#+\s*Response:\s*(.+?)#+\s*End", fully_decoded, flags=re.DOTALL)
+
+                if m:
+                    decoded = m.group(1).strip()
+                else:
+                    # The model might not generate the "### End" sequence before reaching the max tokens.  In this case,
+                    # return everything after "### Response:".
+                    m = re.search(r"#+\s*Response:\s*(.+)", fully_decoded, flags=re.DOTALL)
+                    if m:
+                        decoded = m.group(1).strip()
+                    else:
+                        logger.warn(f"Failed to find response in:\n{fully_decoded}")
+
+            # If the full text is requested, then append the decoded text to the original instruction.
+            # This technically isn't the full text, as we format the instruction in the prompt the model has been
+            # trained on, but to the client it will appear to be the full text.
+            if return_full_text:
+                decoded = f"{instruction_text}\n{decoded}"
+
+            rec = {"generated_text": decoded}
+
+            records.append(rec)
+
+        return records
+
+
+def generate_response(
+    instruction: str,
+    *,
+    model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizer,
+    **kwargs,
+) -> str:
+    """Given an instruction, uses the model and tokenizer to generate a response.  This formats the instruction in
+    the instruction format that the model was fine-tuned on.
+
+    Args:
+        instruction (str): _description_
+        model (PreTrainedModel): the model to use
+        tokenizer (PreTrainedTokenizer): the tokenizer to use
+
+    Returns:
+        str: response
+    """
+
+    generation_pipeline = InstructionTextGenerationPipeline(model=model, tokenizer=tokenizer, **kwargs)
+    return generation_pipeline(instruction)[0]["generated_text"]

training/trainer.py

@@ -0,0 +1,330 @@
+# Copyright 2023 Databricks, Inc.
+
+# 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 logging
+from functools import partial
+from pathlib import Path
+from typing import Any, Dict, List, Tuple, Union
+
+import click
+import numpy as np
+from datasets import Dataset, load_dataset
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    DataCollatorForLanguageModeling,
+    PreTrainedTokenizer,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+
+from .consts import (
+    DEFAULT_INPUT_MODEL,
+    DEFAULT_SEED,
+    PROMPT_WITH_INPUT_FORMAT,
+    PROMPT_NO_INPUT_FORMAT,
+    END_KEY,
+    INSTRUCTION_KEY,
+    RESPONSE_KEY_NL,
+)
+
+logger = logging.getLogger(__name__)
+ROOT_PATH = Path(__file__).parent.parent
+DATABRICKS_DOLLY_15K_PATH = ROOT_PATH / "data" / "databricks-dolly-15k.jsonl"
+
+
+class DataCollatorForCompletionOnlyLM(DataCollatorForLanguageModeling):
+    def torch_call(self, examples: List[Union[List[int], Any, Dict[str, Any]]]) -> Dict[str, Any]:
+        batch = super().torch_call(examples)
+
+        # The prompt ends with the response key plus a newline.  We encode this and then try to find it in the
+        # sequence of tokens.  This should just be a single token.
+        response_token_ids = self.tokenizer.encode(RESPONSE_KEY_NL)
+
+        labels = batch["labels"].clone()
+
+        for i in range(len(examples)):
+
+            response_token_ids_start_idx = None
+            for idx in np.where(batch["labels"][i] == response_token_ids[0])[0]:
+                response_token_ids_start_idx = idx
+                break
+
+            if response_token_ids_start_idx is None:
+                raise RuntimeError(
+                    f'Could not find response key {response_token_ids} in token IDs {batch["labels"][i]}'
+                )
+
+            response_token_ids_end_idx = response_token_ids_start_idx + 1
+
+            # Make pytorch loss function ignore all tokens up through the end of the response key
+            labels[i, :response_token_ids_end_idx] = -100
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def preprocess_batch(batch: Dict[str, List], tokenizer: AutoTokenizer, max_length: int) -> dict:
+    return tokenizer(
+        batch["text"],
+        max_length=max_length,
+        truncation=True,
+    )
+
+
+def load_training_dataset() -> Dataset:
+    logger.info(f"Loading dataset from {DATABRICKS_DOLLY_15K_PATH}")
+    dataset = load_dataset("json", data_files=str(DATABRICKS_DOLLY_15K_PATH))["train"]
+    logger.info("Found %d rows", dataset.num_rows)
+
+    def _add_text(rec):
+        instruction = rec["instruction"]
+        response = rec["response"]
+        context = rec.get("context")
+
+        if not instruction:
+            raise ValueError(f"Expected an instruction in: {rec}")
+
+        if not response:
+            raise ValueError(f"Expected a response in: {rec}")
+
+        # For some instructions there is an input that goes along with the instruction, providing context for the
+        # instruction.  For example, the input might be a passage from Wikipedia and the instruction says to extract
+        # some piece of information from it.  The response is that information to extract.  In other cases there is
+        # no input.  For example, the instruction might be open QA such as asking what year some historic figure was
+        # born.
+        if context:
+            rec["text"] = PROMPT_WITH_INPUT_FORMAT.format(instruction=instruction, response=response, input=context)
+        else:
+            rec["text"] = PROMPT_NO_INPUT_FORMAT.format(instruction=instruction, response=response)
+        return rec
+
+    dataset = dataset.map(_add_text)
+
+    return dataset
+
+
+def load_tokenizer(pretrained_model_name_or_path: str = DEFAULT_INPUT_MODEL) -> PreTrainedTokenizer:
+    logger.info(f"Loading tokenizer for {pretrained_model_name_or_path}")
+    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path)
+    tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.add_special_tokens({"additional_special_tokens": [END_KEY, INSTRUCTION_KEY, RESPONSE_KEY_NL]})
+    return tokenizer
+
+
+def load_model(
+    pretrained_model_name_or_path: str = DEFAULT_INPUT_MODEL, *, gradient_checkpointing: bool = False
+) -> AutoModelForCausalLM:
+    logger.info(f"Loading model for {pretrained_model_name_or_path}")
+    model = AutoModelForCausalLM.from_pretrained(
+        pretrained_model_name_or_path, trust_remote_code=True, use_cache=False if gradient_checkpointing else True
+    )
+    return model
+
+
+def get_model_tokenizer(
+    pretrained_model_name_or_path: str = DEFAULT_INPUT_MODEL, *, gradient_checkpointing: bool = False
+) -> Tuple[AutoModelForCausalLM, PreTrainedTokenizer]:
+    tokenizer = load_tokenizer(pretrained_model_name_or_path)
+    model = load_model(pretrained_model_name_or_path, gradient_checkpointing=gradient_checkpointing)
+    model.resize_token_embeddings(len(tokenizer))
+
+    return model, tokenizer
+
+
+def preprocess_dataset(tokenizer: AutoTokenizer, max_length: int, seed=DEFAULT_SEED) -> Dataset:
+    """Loads the training dataset and tokenizes it so it is ready for training.
+
+    Args:
+        tokenizer (AutoTokenizer): Tokenizer tied to the model.
+        max_length (int): Maximum number of tokens to emit from tokenizer.
+
+    Returns:
+        Dataset: HuggingFace dataset
+    """
+
+    dataset = load_training_dataset()
+
+    logger.info("Preprocessing dataset")
+    _preprocessing_function = partial(preprocess_batch, max_length=max_length, tokenizer=tokenizer)
+    dataset = dataset.map(
+        _preprocessing_function,
+        batched=True,
+        remove_columns=["instruction", "context", "response", "text", "category"],
+    )
+
+    # Make sure we don't have any truncated records, as this would mean the end keyword is missing.
+    logger.info("Processed dataset has %d rows", dataset.num_rows)
+    dataset = dataset.filter(lambda rec: len(rec["input_ids"]) < max_length)
+    logger.info("Processed dataset has %d rows after filtering for truncated records", dataset.num_rows)
+
+    logger.info("Shuffling dataset")
+    dataset = dataset.shuffle(seed=seed)
+
+    logger.info("Done preprocessing")
+
+    return dataset
+
+
+def train(
+    *,
+    input_model: str,
+    local_output_dir: str,
+    dbfs_output_dir: str,
+    epochs: int,
+    per_device_train_batch_size: int,
+    per_device_eval_batch_size: int,
+    lr: float,
+    seed: int,
+    deepspeed: str,
+    gradient_checkpointing: bool,
+    local_rank: str,
+    bf16: bool,
+    logging_steps: int,
+    save_steps: int,
+    eval_steps: int,
+    test_size: Union[float, int],
+    save_total_limit: int,
+    warmup_steps: int,
+):
+    set_seed(seed)
+
+    model, tokenizer = get_model_tokenizer(
+        pretrained_model_name_or_path=input_model, gradient_checkpointing=gradient_checkpointing
+    )
+
+    # Use the same max length that the model supports.  Fall back to 1024 if the setting can't be found.
+    # The configuraton for the length can be stored under different names depending on the model.  Here we attempt
+    # a few possible names we've encountered.
+    conf = model.config
+    max_length = None
+    for length_setting in ["n_positions", "max_position_embeddings", "seq_length"]:
+        max_length = getattr(model.config, length_setting, None)
+        if max_length:
+            logger.info(f"Found max lenth: {max_length}")
+            break
+    if not max_length:
+        max_length = 1024
+        logger.info(f"Using default max length: {max_length}")
+
+    processed_dataset = preprocess_dataset(tokenizer=tokenizer, max_length=max_length, seed=seed)
+
+    split_dataset = processed_dataset.train_test_split(test_size=test_size, seed=seed)
+
+    logger.info("Train data size: %d", split_dataset["train"].num_rows)
+    logger.info("Test data size: %d", split_dataset["test"].num_rows)
+
+    data_collator = DataCollatorForCompletionOnlyLM(
+        tokenizer=tokenizer, mlm=False, return_tensors="pt", pad_to_multiple_of=8
+    )
+
+    if not dbfs_output_dir:
+        logger.warn("Will NOT save to DBFS")
+
+    training_args = TrainingArguments(
+        output_dir=local_output_dir,
+        per_device_train_batch_size=per_device_train_batch_size,
+        per_device_eval_batch_size=per_device_eval_batch_size,
+        fp16=False,
+        bf16=bf16,
+        learning_rate=lr,
+        num_train_epochs=epochs,
+        deepspeed=deepspeed,
+        gradient_checkpointing=gradient_checkpointing,
+        logging_dir=f"{local_output_dir}/runs",
+        logging_strategy="steps",
+        logging_steps=logging_steps,
+        evaluation_strategy="steps",
+        eval_steps=eval_steps,
+        save_strategy="steps",
+        save_steps=save_steps,
+        save_total_limit=save_total_limit,
+        load_best_model_at_end=False,
+        report_to="tensorboard",
+        disable_tqdm=True,
+        remove_unused_columns=False,
+        local_rank=local_rank,
+        warmup_steps=warmup_steps,
+    )
+
+    logger.info("Instantiating Trainer")
+
+    trainer = Trainer(
+        model=model,
+        tokenizer=tokenizer,
+        args=training_args,
+        train_dataset=split_dataset["train"],
+        eval_dataset=split_dataset["test"],
+        data_collator=data_collator,
+    )
+
+    logger.info("Training")
+    trainer.train()
+
+    logger.info(f"Saving Model to {local_output_dir}")
+    trainer.save_model(output_dir=local_output_dir)
+
+    if dbfs_output_dir:
+        logger.info(f"Saving Model to {dbfs_output_dir}")
+        trainer.save_model(output_dir=dbfs_output_dir)
+
+    logger.info("Done.")
+
+
+@click.command()
+@click.option("--input-model", type=str, help="Input model to fine tune", default=DEFAULT_INPUT_MODEL)
+@click.option("--local-output-dir", type=str, help="Write directly to this local path", required=True)
+@click.option("--dbfs-output-dir", type=str, help="Sync data to this path on DBFS")
+@click.option("--epochs", type=int, default=3, help="Number of epochs to train for.")
+@click.option("--per-device-train-batch-size", type=int, default=8, help="Batch size to use for training.")
+@click.option("--per-device-eval-batch-size", type=int, default=8, help="Batch size to use for evaluation.")
+@click.option(
+    "--test-size", type=int, default=1000, help="Number of test records for evaluation, or ratio of test records."
+)
+@click.option("--warmup-steps", type=int, default=None, help="Number of steps to warm up to learning rate")
+@click.option("--logging-steps", type=int, default=10, help="How often to log")
+@click.option("--eval-steps", type=int, default=50, help="How often to run evaluation on test records")
+@click.option("--save-steps", type=int, default=400, help="How often to checkpoint the model")
+@click.option("--save-total-limit", type=int, default=10, help="Maximum number of checkpoints to keep on disk")
+@click.option("--lr", type=float, default=1e-5, help="Learning rate to use for training.")
+@click.option("--seed", type=int, default=DEFAULT_SEED, help="Seed to use for training.")
+@click.option("--deepspeed", type=str, default=None, help="Path to deepspeed config file.")
+@click.option(
+    "--gradient-checkpointing/--no-gradient-checkpointing",
+    is_flag=True,
+    default=True,
+    help="Use gradient checkpointing?",
+)
+@click.option(
+    "--local_rank",
+    type=str,
+    default=True,
+    help="Provided by deepspeed to identify which instance this process is when performing multi-GPU training.",
+)
+@click.option("--bf16", type=bool, default=True, help="Whether to use bf16 (preferred on A100's).")
+def main(**kwargs):
+    train(**kwargs)
+
+
+if __name__ == "__main__":
+    logging.basicConfig(
+        format="%(asctime)s %(levelname)s [%(name)s] %(message)s", level=logging.INFO, datefmt="%Y-%m-%d %H:%M:%S"
+    )
+    try:
+        main()
+    except Exception:
+        logger.exception("main failed")
+        raise