feat: support new DPO data format and update SFT config to use overri…

…de API (#405)

Signed-off-by: Terry Kong <terryk@nvidia.com>
Signed-off-by: arendu <adithya.r@gmail.com>
Signed-off-by: NeMo-Aligner CI <nemo-aligner-ci@nvidia.com>
Co-authored-by: Terry Kong <terryk@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

examples/nlp/gpt/conf/gpt_sft.yaml

@@ -87,6 +87,9 @@ model:
   attention_dropout: 0.0
   ffn_dropout: 0.0
 
+  global_batch_size: ${.data.train_ds.global_batch_size}
+  micro_batch_size: ${.data.train_ds.micro_batch_size}
+
   steerlm2:
     forward_micro_batch_size:  1  # the micro batch size for the forward pass, used to compute the weights
     micro_batch_size:  1 # the steerlm2 training micro batch size

examples/nlp/gpt/train_gpt_sft.py

@@ -39,7 +39,7 @@
     resolve_and_create_trainer,
     retrieve_custom_trainer_state_dict,
 )
-from nemo_aligner.utils.utils import load_from_nemo
+from nemo_aligner.utils.utils import load_and_override_model_config, load_from_nemo
 
 """Script to start SFT training"""
 
@@ -49,75 +49,10 @@
 mp.set_start_method("spawn", force=True)
 
 
-def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):
-    """
-    This function modifies the original gpt pre-training config (gpt_cfg) with attributes from the finetuning config (cfg).
-    The `add_cfg_to_tree` arg adds `cfg` to the top of the yaml tree which is needed for all `hparams.yaml` files when passed as an arg to `load_from_checkpoint()`.
-    """
-    OmegaConf.set_struct(gpt_cfg, True)
-    OmegaConf.resolve(cfg)
-    with open_dict(gpt_cfg):
-        gpt_cfg.megatron_amp_O2 = cfg.model.get("megatron_amp_O2", False)
-        gpt_cfg.micro_batch_size = cfg.model.data.train_ds.micro_batch_size
-        gpt_cfg.global_batch_size = cfg.model.data.train_ds.global_batch_size
-        gpt_cfg.sequence_parallel = cfg.model.get("sequence_parallel", False)
-        gpt_cfg.activations_checkpoint_granularity = cfg.model.get("activations_checkpoint_granularity", None)
-        gpt_cfg.activations_checkpoint_num_layers = cfg.model.get("activations_checkpoint_num_layers", None)
-        gpt_cfg.activations_checkpoint_method = cfg.model.get("activations_checkpoint_method", None)
-        gpt_cfg.activations_checkpoint_layers_per_pipeline = cfg.model.get(
-            "activations_checkpoint_layers_per_pipeline", None
-        )
-        gpt_cfg.peft = cfg.model.peft
-        gpt_cfg.data = cfg.model.data
-        gpt_cfg.optim = cfg.model.optim
-        gpt_cfg.precision = cfg.trainer.precision
-        gpt_cfg.answer_only_loss = cfg.model.answer_only_loss
-        gpt_cfg.restore_from_path = cfg.model.restore_from_path
-        gpt_cfg.resume_from_checkpoint = cfg.model.resume_from_checkpoint
-        gpt_cfg.save_nemo_on_validation_end = cfg.model.save_nemo_on_validation_end
-        gpt_cfg.gradient_as_bucket_view = cfg.model.gradient_as_bucket_view
-        gpt_cfg.hidden_dropout = cfg.model.get("hidden_dropout", 0.0)
-        gpt_cfg.attention_dropout = cfg.model.get("attention_dropout", 0.0)
-        gpt_cfg.ffn_dropout = cfg.model.ffn_dropout
-        gpt_cfg.use_flash_attention = cfg.model.get("use_flash_attention", False)
-        # if TP/PP size is -1, use default TP/PP size as original model
-        if cfg.model.get("tensor_model_parallel_size", 1) > 0:
-            gpt_cfg.tensor_model_parallel_size = cfg.model.get("tensor_model_parallel_size", 1)
-        if cfg.model.get("pipeline_model_parallel_size", 1) > 0:
-            gpt_cfg.pipeline_model_parallel_size = cfg.model.get("pipeline_model_parallel_size", 1)
-        gpt_cfg.pipeline_model_parallel_split_rank = cfg.model.get("pipeline_model_parallel_split_rank", 0)
-
-        if cfg.model.data.get("chat", False):
-            # chat model, overwrite the prompt template
-            prompt_template = get_prompt_template_example(cfg.model.data.chat_prompt_tokens)
-            gpt_cfg.data.train_ds.prompt_template = prompt_template
-            gpt_cfg.data.validation_ds.prompt_template = prompt_template
-
-        sft_cls = GPTSFTModel
-        gpt_cfg.target = f"{sft_cls.__module__}.{sft_cls.__name__}"
-
-        if cfg.model.get("use_flash_attention", None) is not None:
-            gpt_cfg.use_flash_attention = cfg.model.use_flash_attention
-
-        if cfg.model.get("seq_len_interpolation_factor", None) is not None:
-            gpt_cfg.seq_len_interpolation_factor = cfg.model.seq_len_interpolation_factor
-
-        if cfg.model.get("dist_ckpt_load_strictness", None) is not None:
-            gpt_cfg.dist_ckpt_load_strictness = cfg.model.dist_ckpt_load_strictness
-
-        gpt_cfg.inference = cfg.model.get("inference", {})
-
-        # This is needed when modifying a hparam file directly to load `.ckpt` files.
-        # This is not needed to modify the cfg in `.nemo` files.
-        if add_cfg_to_tree:
-            OmegaConf.resolve(gpt_cfg)
-            gpt_cfg.cfg = gpt_cfg
-
-    return gpt_cfg
-
-
 @hydra_runner(config_path="conf", config_name="gpt_sft")
 def main(cfg) -> None:
+    cfg.model = load_and_override_model_config(cfg.model.restore_from_path, cfg.model)
+
     logging.info("\n\n************** Experiment configuration ***********")
     logging.info(f"\n{OmegaConf.to_yaml(cfg)}")
 
@@ -129,17 +64,11 @@ def main(cfg) -> None:
     with open_dict(cfg):
         cfg.model.precision = cfg.trainer.precision
 
-    ptl_model, updated_cfg = load_from_nemo(
-        GPTSFTModel,
-        cfg,
-        trainer,
-        strict=True,
-        modify_config_fn=_modify_config,
-        restore_path=cfg.model.restore_from_path,
-        return_updated_cfg=True,
+    ptl_model = load_from_nemo(
+        GPTSFTModel, cfg, trainer, strict=True, restore_path=cfg.model.restore_from_path, return_updated_cfg=False,
     )
 
-    init_peft(ptl_model, updated_cfg)
+    init_peft(ptl_model, cfg.model)
 
     with open_dict(cfg):
         # overwrite the model config with the config from the checkpoint

nemo_aligner/data/nlp/datasets.py

@@ -15,13 +15,19 @@
 """Custom datasets for RLHF training"""
 
 import os
+from typing import Dict, List
 
 import numpy as np
 import scipy
 import torch
+from omegaconf import OmegaConf
 
 from nemo.collections.nlp.data.language_modeling.megatron.gpt_dataset import _create_ltor_masks_and_position_ids
-from nemo.collections.nlp.data.language_modeling.megatron.gpt_sft_chat_dataset import GPTSFTChatDataset
+from nemo.collections.nlp.data.language_modeling.megatron.gpt_sft_chat_dataset import (
+    GPTSFTChatDataset,
+    _get_header_conversation_type_mask_role,
+    get_prompt_template_example,
+)
 from nemo.core import Dataset
 from nemo.utils import logging
 
@@ -344,16 +350,97 @@ def encode(self, text, append_eod=False):
 
         return text_ids, len(text_ids)
 
+    @staticmethod
+    def _convert_messages(
+        input_list: List[Dict[str, str]]
+    ) -> Dict:  # TODO: (@adithyare) this method should live elsewhare..
+        """
+        args:
+            input_list: is a list of dicts in the openai format
+                for example:
+                [{"role": "system", "content": "you are helpful},
+                {"role": "user", "content": "Why is the sky blue?"},
+                {"role": "assistant", "content": "Because blablabla"},
+                ...]
+        returns:
+            output_dict: a dict in nemo's format {"system": "sytem prompt",
+                                                 "conversation": [],
+                                                 ...
+                                                } 
+        """
+        output_dict = {
+            "system": "",
+            "conversations": [],
+            "mask": "User",
+            "type": "VALUE_TO_TEXT",
+        }
+
+        # Extract the system message
+        num_system_msg = 0
+        for msg in input_list:
+            if msg["role"] == "system":
+                output_dict["system"] = msg["content"]
+                num_system_msg += 1
+            if num_system_msg > 1:
+                raise RuntimeError("Multiple system messages seen, please consolidate into a single system message.")
+
+        # Build the conversations list
+        for msg in input_list:
+            if msg["role"] != "system":
+                conversation_entry = {
+                    "from": msg["role"].capitalize(),  # Capitalize 'user' and 'assistant'
+                    "value": msg["content"],
+                    "label": None,
+                }
+                output_dict["conversations"].append(conversation_entry)
+
+        return output_dict
+
+    def convert(self, messages):
+        """
+        args:
+            messages: is a list of dicts in the openai format
+                for example:
+                [{"role": "system", "content": "you are helpful},
+                {"role": "user", "content": "Why is the sky blue?"},
+                {"role": "assistant", "content": "Because blablabla"},
+                ...]
+        returns:
+            conversation:  is a string formatted with the chat template
+        """
+        if OmegaConf.select(self.cfg, "data.chat_prompt_tokens") is None:
+            raise RuntimeError(
+                "You don't have a model (model_config.yaml) which has chat_prompt_tokens, are you sure this is a Chat/Instruction model?"
+            )
+        special_tokens = self.cfg.data.chat_prompt_tokens
+        nemo_source = self._convert_messages(messages)
+        header, conversation, data_type, mask_role = _get_header_conversation_type_mask_role(
+            nemo_source, special_tokens
+        )
+        return conversation
+
     def __getitem__(self, idx):
         """Returns a pair of chosen/rejected pairs, their respective lengths, and labels."""
         payload = self.data[idx]
-        prompt, prompt_len = self.encode(payload["prompt"], append_eod=False)
-        chosen, chosen_len = self.encode(
-            payload["prompt"] + payload["chosen_response"], append_eod=self.cfg.data.get("append_eod", False)
-        )
-        reject, reject_len = self.encode(
-            payload["prompt"] + payload["rejected_response"], append_eod=self.cfg.data.get("append_eod", False)
-        )
+
+        if isinstance(payload["prompt"], str):
+            # (@adithyare) format with hardcoded chat tokens
+            # will allow this for the time being.
+            prompt_fmtd = payload["prompt"]
+            chosen_fmtd = payload["prompt"] + payload["chosen_response"]
+            rejected_fmtd = payload["prompt"] + payload["rejected_response"]
+            logging.warning(
+                "Pre-formatting chat conversation as string with hardcoded chat tokens will be deprecated."
+            )  # (@adithyare) this will spam the console for now.
+        else:
+            prompt_fmtd = self.convert(payload["prompt"])  # (@adithyare) read var as "prompt formatted"
+            chosen_fmtd = self.convert(payload["prompt"] + [payload["chosen_response"]])
+            rejected_fmtd = self.convert(payload["prompt"] + [payload["rejected_response"]])
+
+        prompt, prompt_len = self.encode(prompt_fmtd, append_eod=False)
+        chosen, chosen_len = self.encode(chosen_fmtd, append_eod=self.cfg.data.get("append_eod", False))
+        reject, reject_len = self.encode(rejected_fmtd, append_eod=self.cfg.data.get("append_eod", False))
+
         # chosen_response_only, chosen_response_len = self.encode(payload['chosen_response'])
         # reject_response_only, reject_response_len = self.encode(payload['rejected_response'])
         chosen_labels = ([-100] * prompt_len) + chosen[prompt_len:]

nemo_aligner/data/nlp/scripts/undo_special_tokens.py

@@ -0,0 +1,68 @@
+# Copyright (c) 2023, 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.
+
+"""Script to remove special tokens from dpo datasets 
+and convert them into list of messages format"""
+
+import argparse
+import json
+import re
+
+
+def format_conversation(input_string):
+    # Define roles and patterns
+    role_patterns = {"<extra_id_0>System": "system", "<extra_id_1>User": "user", "<extra_id_1>Assistant": "assistant"}
+
+    # Initialize an empty output list
+    conversation = []
+
+    # Use regex to find each segment's role and content
+    segments = re.findall(r"(<extra_id_[0-1]>[^\n]+)\n(.*?)((?=<extra_id_)|$)", input_string, re.DOTALL)
+
+    for segment in segments:
+        role_tag, content, _ = segment
+        role = role_patterns.get(role_tag.strip(), "unknown")
+        conversation.append({"role": role, "content": content.strip()})
+
+    return conversation
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Process a JSONL file.")
+    parser.add_argument("input_jsonl", type=str, help="Path to the input JSONL file.")
+    # Parse the arguments
+    args = parser.parse_args()
+
+    input_jsonl = args.input_jsonl
+    output_jsonl = input_jsonl.replace(".jsonl", ".no_special_toks.jsonl")
+
+    with open(input_jsonl, "r") as f, open(output_jsonl, "w") as w:
+        for line in f:
+            j = json.loads(line)
+            prompt = j["prompt"]
+            undo_spl_prompt = format_conversation(prompt)
+            empty_assistant = undo_spl_prompt.pop()
+            chosen, rejected = j["chosen_response"], j["rejected_response"]
+            chosen = chosen.split("\n<extra_id_1>")[0]
+            rejected = rejected.split("\n<extra_id_1>")[0]
+            chosen_message = {"role": empty_assistant["role"], "content": chosen}
+            rejected_message = {"role": empty_assistant["role"], "content": rejected}
+            j_out = {
+                "prompt": undo_spl_prompt,
+                "chosen_response": chosen_message,
+                "rejected_response": rejected_message,
+                "chosen_reward": j["chosen_reward"],
+                "rejected_reward": j["rejected_reward"],
+            }
+            w.write(json.dumps(j_out) + "\n")

setup/requirements.txt

@@ -1,3 +1,4 @@
+Jinja2~=3.1.4
 jsonlines
 megatron_core>=0.8
 nemo_toolkit[nlp]

tests/test_datasets.py

@@ -22,6 +22,7 @@
 from nemo.collections.common.tokenizers.huggingface.auto_tokenizer import AutoTokenizer
 from nemo_aligner.algorithms.dpo import dpo_custom_collate
 from nemo_aligner.data.nlp.builders import build_dataloader, build_train_valid_test_dpo_datasets
+from nemo_aligner.data.nlp.scripts.undo_special_tokens import format_conversation
 from nemo_aligner.utils import parallel_state
 
 
@@ -136,6 +137,56 @@ def test_dpo_loader(init_model_parallel, make_tmp_jsonl, llama3_tokenizer):
     assert num_mini_batches == 2
 
 
+@pytest.mark.run_only_on("GPU")
+def test_dpo_dataset_conversion():
+    prompt_str = """<extra_id_0>System\n\n<extra_id_1>User\nDoes GPT-4 use RNN or Transformer models, and which one is better for this type of project?\n<extra_id_1>Assistant\nGPT-4 uses a transformer architecture, not a recurrent neural network. Both models are commonly used for natural language processing tasks, and both have advantages and disadvantages, so it depends on the specific application of the model.\n<extra_id_1>User\nCould you explain in detail both the advantages and disadvantages from different perspectives?\n<extra_id_1>Assistant\nYes, here are a few different perspectives on the pros and cons of RNN and transformer models:\n \n\n The RNN model has better sequential memory and can better model sequences of data, whereas the transformer model has better parallel computation and can better model large input datasets. This makes RNNs better for applications like speech recognition and machine translation, where the input is a sequence of sounds or words, and transformers better for applications like natural language processing and text classification, where the input is a large collection of documents or social media posts.\n \n\n RNNs are more straightforward and easier to understand and debug, whereas transformers are more flexible and can be used for more diverse applications. This makes RNNs better for teaching and research, and transformers better for actual applications in industry and the real world.\n \n\n RNNs are more efficient and faster to run, whereas transformers are more accurate and produce better results. This makes RNNs better for fast prototyping and testing, and transformers better for actual deployment in production applications.\n<extra_id_1>User\ncould you add more in a table\n<extra_id_1>Assistant\n"""
+
+    expected_oai_messages = [
+        {"role": "system", "content": ""},
+        {
+            "role": "user",
+            "content": "Does GPT-4 use RNN or Transformer models, and which one is better for this type of project?",
+        },
+        {
+            "role": "assistant",
+            "content": "GPT-4 uses a transformer architecture, not a recurrent neural network. Both models are commonly used for natural language processing tasks, and both have advantages and disadvantages, so it depends on the specific application of the model.",
+        },
+        {
+            "role": "user",
+            "content": "Could you explain in detail both the advantages and disadvantages from different perspectives?",
+        },
+        {
+            "role": "assistant",
+            "content": """Yes, here are a few different perspectives on the pros and cons of RNN and transformer models:\n \n\n The RNN model has better sequential memory and can better model sequences of data, whereas the transformer model has better parallel computation and can better model large input datasets. This makes RNNs better for applications like speech recognition and machine translation, where the input is a sequence of sounds or words, and transformers better for applications like natural language processing and text classification, where the input is a large collection of documents or social media posts.\n \n\n RNNs are more straightforward and easier to understand and debug, whereas transformers are more flexible and can be used for more diverse applications. This makes RNNs better for teaching and research, and transformers better for actual applications in industry and the real world.\n \n\n RNNs are more efficient and faster to run, whereas transformers are more accurate and produce better results. This makes RNNs better for fast prototyping and testing, and transformers better for actual deployment in production applications.""",
+        },
+        {"role": "user", "content": "could you add more in a table"},
+        {"role": "assistant", "content": ""},
+    ]
+
+    oai_messages_prompt = format_conversation(prompt_str)
+    assert expected_oai_messages == oai_messages_prompt
+
+    # (@adithyare) bonus test! convert oai style messages back into a string using Jinja
+    # Attempt to import jinja2 via importorskip
+    jinja2 = pytest.importorskip("jinja2", reason="jinja2 library is not installed")
+
+    # Now it's safe to use jinja2
+    from jinja2 import Template
+
+    def remove_trailing(s, t):
+        if s.endswith(t):
+            s = s[: -len(t)]
+        return s
+
+    jinja_template = """{% for message in conversation %}{%- if message.role == "system" -%}<extra_id_0>System\n{{ message.content }}\n{% elif message.role == "user" -%}<extra_id_1>User\n{{ message.content }}\n{% elif message.role == "assistant" -%}<extra_id_1>Assistant\n{{ message.content }}\n{% endif %}{% endfor %}"""
+    jinja_template = Template(jinja_template)
+    prompt_str_jinja_rendered = jinja_template.render(conversation=oai_messages_prompt)
+    prompt_str_jinja_rendered = remove_trailing(
+        prompt_str_jinja_rendered, "\n"
+    )  # (@adithyare) jinja will add the ending of message token which we should remove to make a prompt.
+    assert prompt_str == prompt_str_jinja_rendered
+
+
 @pytest.mark.run_only_on("GPU")
 def test_dpo_loader_original(init_model_parallel, make_tmp_jsonl, llama3_tokenizer):
     init_model_parallel(tensor_model_parallel_size=1, pipeline_model_parallel_size=1)