This includes fixes that make checkpointing and reloading work correctly. (#35)

It also batches in a first set of changes for fixing eval code

Summary:

Test Plan:

apps/main/lingua_train.py

@@ -544,7 +544,7 @@ def train(args: TrainArgs):
             if args.eval is not None and every_n_steps(
                 train_state, args.checkpoint.eval.every, acc_step=0
             ):
-                from apps.main.eval import EVAL_FOLDER_NAME, EvalArgs, launch_eval
+                from bytelatent.eval import EVAL_FOLDER_NAME, EvalArgs, launch_eval
 
                 eval_args = dataclass_from_dict(EvalArgs, args.eval)
 

bytelatent/args.py

@@ -5,6 +5,7 @@ from typing import Any
 
 import numpy as np
 import yaml
+from omegaconf import OmegaConf
 from pydantic import BaseModel, ConfigDict
 
 from bytelatent.checkpoint import CheckpointArgs
@@ -39,6 +40,19 @@ def get_rng_state(seed: int, rank: int, world_size: int) -> dict[str, Any]:
     return np.random.default_rng((seed, rank, world_size)).bit_generator.state
 
 
+def parse_args(args_cls):
+    cli_args = OmegaConf.from_cli()
+    file_cfg = OmegaConf.load(cli_args.config)
+    # We remove 'config' attribute from config as the underlying DataClass does not have it
+    del cli_args.config
+
+    default_cfg = OmegaConf.create(args_cls().model_dump())
+    cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
+    cfg = OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True)
+    pydantic_args = args_cls.model_validate(cfg)
+    return pydantic_args
+
+
 def distribute_data_to_rank(
     *,
     dataset_path: str,
@@ -71,6 +85,22 @@ def distribute_data_to_rank(
     return rank_to_arrow_iterator_params[rank]
 
 
+class PackedCausalTransformerGeneratorArgs(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    temperature: float = 0.0
+    top_p: float | None = None
+    top_k: float | None = None
+    max_gen_len: int = 512  # Maximum number of tokens to generate
+    max_tokens: int = 1024  # Maximum number of tokens that can go through the model
+    max_prompt_len: int | None = None
+    until: list[str] = []
+    compile_prefilling: bool = False
+    reduce_generation_overhead: bool = False
+    show_progress: bool = False
+    dtype: str | None = "bf16"
+    device: str | None = "cuda"
+
+
 class DataloaderArgs(BaseModel):
     model_config = ConfigDict(extra="forbid")
     s3_profile: str | None = None
@@ -168,6 +198,58 @@ class DataloaderArgs(BaseModel):
             return packing_iterator
 
 
+class LMHarnessArgs(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    tasks: list[Any] | None = None
+    num_fewshot: int | None = None
+    device: str | None = None
+    use_cache: str | None = None
+    cache_requests: bool = False
+    rewrite_requests_cache: bool = False
+    delete_requests_cache: bool = False
+    limit: int | float | None = None
+    bootstrap_iters: int = 100000
+    check_integrity: bool = False
+    write_out: bool = False
+    log_samples: bool = True
+    system_instruction: str | None = None
+    apply_chat_template: bool | str = False
+    fewshot_as_multiturn: bool = False
+    gen_kwargs: str | None = None
+    verbosity: str = "INFO"
+    predict_only: bool = False
+    random_seed: int = 0
+    numpy_random_seed: int = 1234
+    torch_random_seed: int = 1234
+    fewshot_random_seed: int = 1234
+
+
+class ValidationArgs(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    max_steps: int | None = (
+        None  # If None the whole validation file is used -> /!\ This number of steps is gpu dependent (100 max steps on 8 gpus = 800 steps on 1 gpu)
+    )
+    use_val_from_train_src: bool = True  # Use the validation set from training sources
+    root_dir: str = ""
+    sources: list[str] = []  # Other sources to eval on
+
+
+class EvalArgs(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    dump_dir: str
+    ckpt_dir: str
+    metric_log_dir: str | None = None
+    generator: PackedCausalTransformerGeneratorArgs = (
+        PackedCausalTransformerGeneratorArgs()
+    )
+
+    harness: LMHarnessArgs | None = LMHarnessArgs()
+    validation: ValidationArgs | None = ValidationArgs()
+
+    global_step: int | None = None  # for in-training evaluation
+    s3_profile: str | None = None
+
+
 class TrainArgs(BaseModel):
     model_config = ConfigDict(extra="forbid")
     name: str = "lingua"
@@ -186,6 +268,9 @@ class TrainArgs(BaseModel):
 
     # Nb optimizer steps to take
     steps: int = 1000
+    # If not None, halt training after this many steps,
+    # useful for debugging
+    max_steps: int | None = None
 
     data: DataloaderArgs = DataloaderArgs()
     optim: OptimArgs = OptimArgs()
@@ -203,7 +288,7 @@ class TrainArgs(BaseModel):
 
     # If set to None, eval is run locally otherwise it launches a new job with the given number of gpus
     async_eval_gpus: int | None = None
-    eval: Any | None = None
+    eval: EvalArgs | None = None
     eval_on_gpus: int | None = None
 
     def dump_to_yaml_file(

bytelatent/checkpoint.py

@@ -7,6 +7,7 @@ import re
 from pathlib import Path
 from typing import List, Optional, Tuple
 
+import fsspec
 import torch
 import torch.distributed as dist
 import torch.distributed.checkpoint as dcp
@@ -21,6 +22,7 @@ from torch.distributed.checkpoint.state_dict import (
     set_state_dict,
 )
 
+from bytelatent.data.file_util import get_fs
 from bytelatent.distributed import get_is_master
 
 logger = logging.getLogger("CHECKPOINT")
@@ -51,13 +53,14 @@ class CheckpointArgs(BaseModel):
     path: str | None = None
     init_ckpt_path: str | None = None
     continue_training_from_init: bool = False
+    s3_profile: str | None = None
 
 
 def _get_key_step(name: str):
     return int(re.findall(RE_DIGITS, name)[-1])
 
 
-def consolidate_checkpoints(ckpt_dir: str):
+def consolidate_checkpoints(fs: fsspec.AbstractFileSystem, ckpt_dir: str):
     """
     Consolidates all FSDP checkpoints in a directory to a single file
     Consolidate checkpoint is saved in a subdirectory of ckpt_dir
@@ -102,15 +105,17 @@ def load_from_checkpoint(
     dcp.load(state_dict, checkpoint_id=ckpt_dir)
 
 
+# TODO: Rewrite the file operations here to use fsspec to enable s3 writing.
 class CheckpointManager:
     def __init__(self, args: CheckpointArgs):
         self.path = args.path
+        self.fs = get_fs(self.path, s3_profile=args.s3_profile)
         self.dump_every = args.dump
         self.eval_every = args.eval
         self.init_ckpt_path = args.init_ckpt_path
         self.continue_training_from_init = args.continue_training_from_init
 
-        assert os.path.exists(
+        assert self.fs.exists(
             self.path
         ), f"Path {self.path} does not exist and needs to be created before using CheckpointManager (use instantiate_and_make_dir)"
 

bytelatent/configs/debug.yaml

@@ -98,11 +98,4 @@ logging:
   freq: 10
 
 eval_on_gpus: 8
-eval:
+eval: null
-  dataset_dir: /checkpoint/amaia/codegen/datasets/eval
-  tasks: boolq,hellaswag,nq,piqa,siqa,tqa,winogrande,obqa,arc_easy,arc_challenge,race.middle,race.high,gsm8k,math,bbh,copa,human_eval_plus,mbpp,mmlu
-  generator:
-    max_tokens: 65536
-    dtype: bf16
-
-  mp_size: 1

bytelatent/configs/entropy_model.yaml

@@ -72,11 +72,4 @@ logging:
   freq: 10
 
 eval_on_gpus: 8
-eval:
+eval: null
-  dataset_dir: ???
-  tasks: ???
-  generator:
-    max_tokens: 65536
-    dtype: bf16
-
-  mp_size: 1

bytelatent/data/data_types.py

@@ -40,16 +40,6 @@ class BltPackTokensState(BaseModel):
     n_views: int = 2
 
 
-class DataLoaderState(BaseModel):
-    model_config = ConfigDict(extra="forbid")
-    multi_choice_state: MultiChoiceState
-    pack_tokens_state: BltPackTokensState
-    prefetch_state: PrefetchState
-
-
-BltIterator = Iterator[tuple[BltExample, DataLoaderState]]
-
-
 class BltSequence(BaseModel):
     tokens: list[int]
     mask: list[bool]

bytelatent/data/iterators/multiprocess_iterator.py

@@ -128,6 +128,13 @@ class MultiprocessIterator(StatefulIterator):
         self.producer = None
         self.stop_iterating_event = None
         self.state_dumped_event = None
+        self.force_shutdown = False
+
+    def shutdown(self):
+        if self.producer is not None:
+            # This properly shuts things down
+            self.producer.kill()
+            self.force_shutdown = True
 
     def get_state(self) -> MultiprocessIteratorState:
         """
@@ -135,6 +142,10 @@ class MultiprocessIterator(StatefulIterator):
         to halt the background process and allow it to write the state to the main loop
         in order to not lose data
         """
+        if self.force_shutdown:
+            raise ValueError(
+                "State will be invalid if shutdown was forced before state persisted."
+            )
         if self.producer is None:
             serialized_prefetch_buffer = json.dumps(
                 [b.to_python_dict() for b in self.prefetch_buffer]
@@ -187,6 +198,10 @@ class MultiprocessIterator(StatefulIterator):
             )
 
     def create_iter(self):
+        if self.force_shutdown:
+            raise ValueError(
+                "Iterator may be invalid if shutdown was forced before state persisted."
+            )
         logging.info("Main thread: Creating MP iterator")
         # First yield from the stored prefetch buffer.
         if self.prefetch_buffer is not None:

bytelatent/eval.py

@@ -4,20 +4,20 @@ import json
 import logging
 import os
 from collections import defaultdict
-from dataclasses import asdict, dataclass, field
 from datetime import datetime
 from pathlib import Path
-from typing import Any, List, Optional, Tuple, Union
+from typing import Any
 
 import torch
-from lingua.args import dump_config
-from lingua.data import init_choice_state, setup_sources
 from lm_eval import simple_evaluate
 from lm_eval.api.instance import Instance
 from lm_eval.api.model import LM
 from omegaconf import OmegaConf
+from pydantic import BaseModel, ConfigDict
 
+from bytelatent.args import EvalArgs, ValidationArgs, parse_args
 from bytelatent.checkpoint import CONSOLIDATE_FOLDER, consolidate_checkpoints
+from bytelatent.data.file_util import get_fs
 from bytelatent.distributed import (
     DistributedArgs,
     dist_mean_dict,
@@ -25,72 +25,17 @@ from bytelatent.distributed import (
     get_world_size,
     setup_torch_distributed,
 )
-from bytelatent.transformer import LMTransformer, LMTransformerArgs
+from bytelatent.generate import (
-
-from apps.main.generate import (
     PackedCausalTransformerGenerator,
-    PackedCausalTransformerGeneratorArgs,
     load_consolidated_model_and_tokenizer,
 )
+from bytelatent.transformer import LMTransformer, LMTransformerArgs
 
 EVAL_FOLDER_NAME = "{:010d}"
 
 logger = logging.getLogger()
 
 
-@dataclass
-class LMHarnessArgs:
-    tasks: Optional[List[Any]] = None
-    num_fewshot: Optional[int] = None
-    device: Optional[str] = None
-    use_cache: Optional[str] = None
-    cache_requests: bool = False
-    rewrite_requests_cache: bool = False
-    delete_requests_cache: bool = False
-    limit: Optional[Union[int, float]] = None
-    bootstrap_iters: int = 100000
-    check_integrity: bool = False
-    write_out: bool = False
-    log_samples: bool = True
-    system_instruction: Optional[str] = None
-    apply_chat_template: Union[bool, str] = False
-    fewshot_as_multiturn: bool = False
-    gen_kwargs: Optional[str] = None
-    verbosity: str = "INFO"
-    predict_only: bool = False
-    random_seed: int = 0
-    numpy_random_seed: int = 1234
-    torch_random_seed: int = 1234
-    fewshot_random_seed: int = 1234
-
-
-@dataclass
-class ValidationArgs:
-    max_steps: Optional[int] = (
-        None  # If None the whole validation file is used -> /!\ This number of steps is gpu dependent (100 max steps on 8 gpus = 800 steps on 1 gpu)
-    )
-    use_val_from_train_src: bool = True  # Use the validation set from training sources
-    root_dir: str = ""
-    sources: List[str] = field(default_factory=list)  # Other sources to eval on
-
-
-@dataclass
-class EvalArgs:
-    name: str = "evals"
-    dump_dir: Optional[str] = None
-    metric_log_dir: Optional[str] = None
-    ckpt_dir: str = ""
-    generator: PackedCausalTransformerGeneratorArgs = field(
-        default_factory=PackedCausalTransformerGeneratorArgs
-    )
-    harness: Optional[LMHarnessArgs] = field(default_factory=LMHarnessArgs)
-    validation: Optional[ValidationArgs] = field(default_factory=ValidationArgs)
-
-    wandb: Optional[Any] = None
-
-    global_step: Optional[int] = None  # for in-training evaluation
-
-
 def all_dicts_same(dict_list):
     if not dict_list:  # Check if the list is empty
         return True
@@ -120,7 +65,7 @@ class EvalHarnessLM(LM):
         self._world_size = get_world_size()
         self.device = generator.device
 
-    def generate_until(self, requests: List[Instance]) -> List[str]:
+    def generate_until(self, requests: list[Instance]) -> list[str]:
         prompts, gen_args = zip(*[req.args for req in requests])
         assert all_dicts_same(gen_args), "Doesn't support different gen args for now"
         gen_args = gen_args[0]
@@ -141,7 +86,7 @@ class EvalHarnessLM(LM):
             filtered_gen.append(g)
         return filtered_gen
 
-    def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
+    def loglikelihood(self, requests: list[Instance]) -> list[tuple[float, bool]]:
         prompts, continuations = zip(*[req.args for req in requests])
         inputs = [req.args[0] + req.args[1] for req in requests]
         max_gen_len = self.generator.max_gen_len
@@ -158,7 +103,7 @@ class EvalHarnessLM(LM):
         self.generator.max_gen_len = max_gen_len
         return results
 
-    def loglikelihood_rolling(self, requests: List[Instance]) -> List[float]:
+    def loglikelihood_rolling(self, requests: list[Instance]) -> list[float]:
         prompts = [req.args[0] for req in requests]
         max_gen_len = self.generator.max_gen_len
         # We temporarily lower max gen len
@@ -232,68 +177,73 @@ def eval_on_val(generator, val_args: ValidationArgs, train_cfg):
     return all_val_metrics
 
 
-def launch_eval(cfg: EvalArgs):
+def launch_eval(eval_args: EvalArgs):
     if not torch.distributed.is_initialized():
         setup_torch_distributed(DistributedArgs())
+
+    fs = get_fs(eval_args.ckpt_dir, s3_profile=eval_args.s3_profile)
     if (
-        Path(cfg.ckpt_dir).exists()
+        fs.exists(eval_args.ckpt_dir)
-        and (Path(cfg.ckpt_dir) / "params.json").exists()
+        and fs.exists(os.path.join(eval_args.ckpt_dir, "params.json"))
-        and next(Path(cfg.ckpt_dir).glob("*.pth"), None) is not None
+        and len(fs.glob(os.path.join(eval_args.ckpt_dir, "*.pth"))) != 0
     ):
-        consolidate_path = Path(cfg.ckpt_dir)
+        consolidate_path = eval_args.ckpt_dir
     else:
-        consolidate_path = Path(cfg.ckpt_dir) / CONSOLIDATE_FOLDER
+        consolidate_path = os.path.join(eval_args.ckpt_dir, CONSOLIDATE_FOLDER)
-        if not consolidate_path.exists() and get_global_rank() == 0:
+        if not fs.exists(consolidate_path) and get_global_rank() == 0:
-            consolidate_path = consolidate_checkpoints(cfg.ckpt_dir)
+            consolidate_path = consolidate_checkpoints(eval_args.ckpt_dir)
 
-    Path(cfg.dump_dir).mkdir(parents=True, exist_ok=True)
+    fs.mkdirs(eval_args.dump_dir, exist_ok=True)
-    dump_config(cfg, Path(cfg.dump_dir) / "config.yaml", log_config=False)
+    with fs.open(os.path.join(eval_args.dump_dir, "config.yaml"), "w") as f:
+        f.write(eval_args.model_dump_json())
 
-    consolidate_path = str(consolidate_path)
     torch.distributed.barrier()
     logger.info("Loading model")
+    # TODO: Make this general so that it works with either
+    # LMTransformer or Blt, similar with args
     model, tokenizer, train_cfg = load_consolidated_model_and_tokenizer(
         consolidate_path,
-        model_cls=LMTransformer,
-        model_args_cls=LMTransformerArgs,
     )
     logger.info("Model loaded")
     model.eval()
-    generator = PackedCausalTransformerGenerator(cfg.generator, model, tokenizer)
+    generator = PackedCausalTransformerGenerator(eval_args.generator, model, tokenizer)
 
     wrap = EvalHarnessLM(generator)
-    results = simple_evaluate(wrap, **asdict(cfg.harness))
+    # Redo
+    results = simple_evaluate(wrap, eval_args.harness.model_dump())
     val_results = None
-    if cfg.validation:
+    if eval_args.validation:
-        val_results = eval_on_val(generator, cfg.validation, train_cfg)
+        val_results = eval_on_val(generator, eval_args.validation, train_cfg)
     if get_global_rank() == 0:
-        with open(Path(cfg.dump_dir) / "results.json", "w") as f:
+        with fs.open(os.path.join(eval_args.dump_dir, "results.json"), "w") as f:
             f.write(json.dumps(results))
         logger.info(f"All evaluation results: {results['results']}")
         if val_results is not None:
-            with open(Path(cfg.dump_dir) / "validation.json", "w") as f:
+            with fs.open(os.path.join(eval_args.dump_dir, "validation.json"), "w") as f:
                 f.write(json.dumps(val_results))
             logger.info(f"All validation results: {val_results}")
-    if cfg.metric_log_dir and get_global_rank() == 0:
+    if eval_args.metric_log_dir and get_global_rank() == 0:
-        metric_log_path = Path(cfg.metric_log_dir) / "metrics.eval.jsonl"
+        metric_log_path = os.path.join(eval_args.metric_log_dir, "metrics.eval.jsonl")
 
         logger.info(f"Writing metric logs to {metric_log_path}")
         timestamp = {
             "created_at": datetime.utcnow().isoformat(),
         }
-        if cfg.global_step is not None:
+        if eval_args.global_step is not None:
-            timestamp["global_step"] = cfg.global_step
+            timestamp["global_step"] = eval_args.global_step
         print(
             json.dumps(timestamp | results["results"]),
-            file=open(metric_log_path, mode="a"),
+            file=fs.open(metric_log_path, mode="a"),
             flush=True,
         )
 
-        val_log_path = Path(cfg.metric_log_dir) / "metrics.validation.jsonl"
+        val_log_path = os.path.join(
+            eval_args.metric_log_dir, "metrics.validation.jsonl"
+        )
         if val_results is not None:
             print(
                 json.dumps(timestamp | val_results),
-                file=open(val_log_path, mode="a"),
+                file=fs.open(val_log_path, mode="a"),
                 flush=True,
             )
 
@@ -301,53 +251,8 @@ def launch_eval(cfg: EvalArgs):
 
 
 def main():
-    """
+    eval_args = parse_args(EvalArgs)
-    The command line interface here uses OmegaConf https://omegaconf.readthedocs.io/en/2.3_branch/usage.html#from-command-line-arguments
+    launch_eval(eval_args)
-    This accepts arguments as a dot list
-    So if the dataclass looks like
-
-    @dataclass
-    class DummyArgs:
-        name: str
-        model: LMTransformerArgsgs
-
-    @dataclass
-    class LMTransformerArgsgs:
-        dim: int
-
-    Then you can pass model.dim=32 to change values in LMTransformerArgsgs
-    or just name=tictac for top level attributes.
-
-    The behavior here is as follows:
-    1. We instantiate EvalArgs with its default values
-    2. We override those default values with the ones in the provided config file
-    3. We override the result with the additional arguments provided through command line
-
-    For example, if the config is the following
-
-    model:
-        dim: 128
-        n_layers: 4
-
-    and you call eval.py with eval.py model.dim=64
-
-    Then the final TrainArgs will have
-
-    model:
-        dim: 64
-        n_layers: 4
-
-    Plus all the default values in EvalArgs dataclass.
-    """
-    cli_args = OmegaConf.from_cli()
-    file_cfg = OmegaConf.load(cli_args.config)
-    # We remove 'config' attribute from config as the underlying DataClass does not have it
-    del cli_args.config
-
-    default_cfg = OmegaConf.structured(EvalArgs())
-    cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
-    cfg = OmegaConf.to_object(cfg)
-    launch_eval(cfg)
 
 
 if __name__ == "__main__":

bytelatent/generate.py

@@ -1,20 +1,16 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 
+import os
 import time
-from dataclasses import dataclass, field
-from pathlib import Path
-from typing import List, Optional
 
 import torch
-from lingua.args import dataclass_from_dict
-from lingua.tokenizers.abstract_tokenizer import Tokenizer
-from lingua.tokenizers.build_tokenizer import build_tokenizer
 from omegaconf import OmegaConf
 from torch import nn
 from torch.nn import functional as F
 from torch.nn.attention.flex_attention import create_block_mask
 from tqdm import tqdm
 
+from bytelatent.args import PackedCausalTransformerGeneratorArgs, TrainArgs
 from bytelatent.base_transformer import (
     Attention,
     causal_mask,
@@ -23,7 +19,10 @@ from bytelatent.base_transformer import (
     lengths_to_start_ids,
 )
 from bytelatent.checkpoint import CONSOLIDATE_NAME
-from bytelatent.transformer import LMTransformer, LMTransformerArgs
+from bytelatent.data.file_util import get_fs
+from bytelatent.model.blt import ByteLatentTransformer
+from bytelatent.tokenizers.abstract_tokenizer import Tokenizer
+from bytelatent.transformer import LMTransformer
 
 
 def sample_top_p(probs: torch.Tensor, p: float) -> torch.Tensor:
@@ -62,7 +61,7 @@ def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None):
     return next_token.view(shape[:-1])
 
 
-def pack_prompts(prompts: List[int]):
+def pack_prompts(prompts: list[int]):
     res = []
     lengths = []
     for i, p in enumerate(prompts):
@@ -120,22 +119,6 @@ class KVCache(nn.Module):
         return self.k_cache, self.v_cache
 
 
-@dataclass
-class PackedCausalTransformerGeneratorArgs:
-    temperature: float = 0.0
-    top_p: Optional[float] = None
-    top_k: Optional[float] = None
-    max_gen_len: int = 512  # Maximum number of tokens to generate
-    max_tokens: int = 1024  # Maximum number of tokens that can go through the model
-    max_prompt_len: Optional[int] = None
-    until: List[str] = field(default_factory=list)
-    compile_prefilling: bool = False
-    reduce_generation_overhead: bool = False
-    show_progress: bool = False
-    dtype: Optional[str] = "bf16"
-    device: Optional[str] = "cuda"
-
-
 class PackedCausalTransformerGenerator:
     def __init__(
         self,
@@ -401,25 +384,29 @@ class PackedCausalTransformerGenerator:
 
 def load_consolidated_model_and_tokenizer(
     consolidated_path,
-    model_cls=LMTransformer,
-    model_args_cls=LMTransformerArgs,
 ):
-    ckpt_path = Path(consolidated_path)
+    train_args_path = os.path.join(consolidated_path, "params.json")
-    config = ckpt_path / "params.json"
+    fs = get_fs(train_args_path)
-    config = OmegaConf.load(config)
+    with fs.open(train_args_path) as f:
+        train_args = TrainArgs.model_validate_json(f.read())
+
+    if train_args.train_entropy_model:
+        model_args = train_args.entropy_model
+        model = LMTransformer(model_args)
+    else:
+        model_args = train_args.model
+        model = ByteLatentTransformer(model_args)
 
     param_dtype = dict(fp32=torch.float32, fp16=torch.float16, bf16=torch.bfloat16)[
-        config.distributed.model_dtype
+        train_args.distributed.model_dtype
     ]
-    model_args = dataclass_from_dict(model_args_cls, config.model, strict=False)
+    tokenizer = train_args.data.tokenizer_args.build()
-    tokenizer = build_tokenizer(config.data.tokenizer.name, config.data.tokenizer.path)
+    st_dict = torch.load(consolidated_path / CONSOLIDATE_NAME, weights_only=True)
-    model = model_cls(model_args)
-    st_dict = torch.load(ckpt_path / CONSOLIDATE_NAME, weights_only=True)
     model.load_state_dict(st_dict["model"])
     model = model.cuda().eval()
     for param in model.parameters():
         param.data = param.data.to(dtype=param_dtype)
-    return model, tokenizer, config
+    return model, tokenizer, train_args
 
 
 def main():

bytelatent/train.py

@@ -10,7 +10,7 @@ from copy import deepcopy
 from dataclasses import asdict, dataclass
 from pathlib import Path
 from timeit import default_timer as timer
-from typing import Any, Dict, Type, TypeVar
+from typing import Any, TypeVar
 
 import torch
 import torch.distributed
@@ -23,9 +23,13 @@ from torch.distributed._tensor import DTensor
 from torch.distributed.checkpoint.stateful import Stateful
 from torch.optim import lr_scheduler
 
-from bytelatent.args import TrainArgs
+from bytelatent.args import TrainArgs, parse_args
 from bytelatent.checkpoint import CheckpointManager, load_from_checkpoint
-from bytelatent.data.data_types import DataLoaderState
+from bytelatent.data.iterators.multiprocess_iterator import (
+    MultiprocessIterator,
+    MultiprocessIteratorState,
+)
+from bytelatent.data.iterators.packing_iterator import PackingIteratorState
 from bytelatent.distributed import (
     check_model_value_range,
     clean_env,
@@ -39,6 +43,7 @@ from bytelatent.distributed import (
     setup_env,
     setup_torch_distributed,
 )
+from bytelatent.eval import EVAL_FOLDER_NAME, launch_eval
 from bytelatent.logger import init_logger
 from bytelatent.metrics import GPUMemoryMonitor, MetricLogger, get_num_params
 from bytelatent.model.blt import ByteLatentTransformer
@@ -70,36 +75,49 @@ def flatten_dict(d, parent_key="", sep="_"):
     return dict(items)
 
 
-def dataclass_from_dict(cls: Type[T], data: dict, strict: bool = True) -> T:
+def get_iterator_state_name(iterator_state):
-    """
+    if isinstance(iterator_state, MultiprocessIteratorState):
-    Converts a dictionary to a dataclass instance, recursively for nested structures.
+        return "multiprocess"
-    """
+    elif isinstance(iterator_state, PackingIteratorState):
-    base = OmegaConf.structured(cls())
+        return "packing"
-    OmegaConf.set_struct(base, strict)
+    else:
-    override = OmegaConf.create(data)
+        raise ValueError(f"Unsupported iterator to get name from: {iterator_state}")
-    return OmegaConf.to_object(OmegaConf.merge(base, override))
 
 
+# TODO: Make this pydantic based instead of data class based
+# TODO: Generalize this to any iterator state
 @dataclass
 class TrainState(Stateful):
     step: int  # Nb of steps taken by the optimizer
     acc_step: int  # Nb of accumulation steps done since last optimizer step
     scheduler: lr_scheduler.LambdaLR
-    data_loader_state: DataLoaderState
+    data_loader_state: MultiprocessIteratorState | PackingIteratorState
     scale: float = 1.0
+    data_loader_class: str | None = None
 
-    def state_dict(self) -> Dict[str, Any]:
+    def state_dict(self) -> dict[str, Any]:
         return {
             "step": self.step,
             "acc_step": self.acc_step,
-            "data_loader_state": self.data_loader_state.dict(),
+            "data_loader_state": self.data_loader_state.model_dump(),
+            "data_loader_class": get_iterator_state_name(self.data_loader_state),
             "scheduler": self.scheduler.state_dict(),
         }
 
     def load_state_dict(self, state_dict):
         self.step = state_dict["step"]
         self.acc_step = state_dict["acc_step"]
-        self.data_loader_state = DataLoaderState(**state_dict["data_loader_state"])
+        self.data_loader_class = state_dict["data_loader_class"]
+        if self.data_loader_class == "multiprocess":
+            self.data_loader_state = MultiprocessIteratorState(
+                **state_dict["data_loader_state"]
+            )
+        elif self.data_loader_class == "packing":
+            self.data_loader_state = PackingIteratorState(
+                **state_dict["data_loader_state"]
+            )
+        else:
+            raise ValueError(f"invalid data loader class: {self.data_loader_class}")
         self.scheduler.load_state_dict(state_dict["scheduler"])
 
 
@@ -345,7 +363,10 @@ def train(args: TrainArgs):
         nwords_since_last_log = 0
         time_last_log = timer()
         gc.collect()
-        while train_state.step < args.steps:
+        saved = False
+        while train_state.step < args.steps and (
+            args.max_steps is None or train_state.step < args.max_steps
+        ):
             # We constrain train_state.acc_step to be in range 0 to args.grad_acc_steps - 1
             train_state.acc_step += 1
             train_state.acc_step = train_state.acc_step % args.grad_acc_steps
@@ -552,7 +573,6 @@ def train(args: TrainArgs):
                     f"  pow: {gpu_mem_stats.power_draw/1000} W"
                 )
 
-            saved = False
             if every_n_steps(
                 train_state, args.checkpoint.dump.every, acc_step=0
             ) or every_n_steps(train_state, args.checkpoint.eval.every, acc_step=0):
@@ -567,18 +587,14 @@ def train(args: TrainArgs):
             if args.eval is not None and every_n_steps(
                 train_state, args.checkpoint.eval.every, acc_step=0
             ):
-                from apps.main.eval import EVAL_FOLDER_NAME, EvalArgs, launch_eval
+                eval_args = args.eval
-
-                eval_args = dataclass_from_dict(EvalArgs, args.eval)
 
                 eval_args.global_step = train_state.step
                 eval_args.ckpt_dir = str(checkpoint.existing_saves[-1])
-                eval_args.dump_dir = str(
+                eval_args.dump_dir = os.path.join(
-                    os.path.join(
+                    args.dump_dir,
-                        args.dump_dir,
+                    "evals",
-                        "evals",
+                    EVAL_FOLDER_NAME.format(train_state.step),
-                        EVAL_FOLDER_NAME.format(train_state.step),
-                    )
                 )
                 eval_args.metric_log_dir = args.dump_dir
                 if args.async_eval_gpus is None:
@@ -619,6 +635,9 @@ def train(args: TrainArgs):
             args,
             device_mesh=world_mesh,
         )
+    if isinstance(data_loader, MultiprocessIterator):
+        logger.info("Closing MP iterator before exiting")
+        data_loader.shutdown()
     gc.collect()
 
 
@@ -661,15 +680,7 @@ def main():
 
     Plus all the default values in TrainArgs dataclass.
     """
-    cli_args = OmegaConf.from_cli()
+    train_args = parse_args(TrainArgs)
-    file_cfg = OmegaConf.load(cli_args.config)
-    # We remove 'config' attribute from config as the underlying DataClass does not have it
-    del cli_args.config
-
-    default_cfg = OmegaConf.create(TrainArgs().model_dump())
-    cfg = OmegaConf.merge(default_cfg, file_cfg, cli_args)
-    cfg = OmegaConf.to_container(cfg, resolve=True, throw_on_missing=True)
-    train_args = TrainArgs.model_validate(cfg)
     if train_args.debug_dynamo:
         import torch._dynamo