add skip dataset (#126)

* add skip dataset

* add skip dataset

* add skip dataset

configs/test.toml

@@ -0,0 +1,18 @@
+name_model = "debugmodel"
+project = "debug_150m_zero_band"
+type_model = "llama2"
+
+[train]
+micro_bs = 4  # change this base on the gpu
+
+[data]
+seq_length = 8192
+dataset_name_or_paths = "PrimeIntellect/fineweb-edu,PrimeIntellect/fineweb,PrimeIntellect/StackV1-popular,mlfoundations/dclm-baseline-1.0-parquet,open-web-math/open-web-math"
+dataset_ratio = "55:10:20:10:5"
+num_workers = 8
+
+[optim]
+batch_size = 128
+warmup_steps = 1000
+total_steps = 88_000
+lr = 4e-4

scripts/skip_data.py

@@ -0,0 +1,87 @@
+"""
+This script is simulating a training to exaust the datasets and recover the dataloader ckpt.
+
+It has the same api as the training one. The only difference is that you probably want to change the total_steps and put a data_path.
+
+It can load config from the config file to have the same setup as the real run.
+
+example.
+```
+uv run torchrun --nproc_per_node=4 scripts/skip_data.py @configs/150M/3090.toml --optim.total_steps 100 --ckpt.data_path out_data
+```
+
+"""
+
+import os
+import torch
+from pydantic_config import parse_argv
+
+
+from transformers import AutoTokenizer
+from zeroband.checkpoint import CkptManager
+
+from zeroband.train import Config
+
+from zeroband.data import get_dataloader
+
+from zeroband.utils.world_info import get_world_info
+from zeroband.utils.logging import get_logger
+
+
+def skip_data(config: Config):
+    # batch_size is the total batch size for all GPUs
+    assert config.optim.batch_size % world_info.local_world_size == 0
+    batch_size = config.optim.batch_size // world_info.local_world_size
+
+    assert batch_size % config.train.micro_bs == 0
+    gradient_accumulation_steps = batch_size // config.train.micro_bs
+
+    if config.type_model == "llama2":
+        tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", use_fast=True)
+    elif config.type_model == "llama3":
+        tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B", use_fast=True)
+    else:
+        raise ValueError(f"Model type {config.type_model} not supported")
+
+    logger.debug("tokenizer loaded")
+
+    train_dataloader = get_dataloader(
+        tokenizer=tokenizer,
+        world_size=world_info.world_size,
+        rank=world_info.rank,
+        batch_size=config.train.micro_bs,
+        data_config=config.data,
+    )
+
+    train_dataloader_iterator = iter(train_dataloader)
+
+    logger.info("starting skipping data up to step: %d", config.optim.total_steps)
+
+    total_steps = 0
+
+    while True:
+        num_inner_steps = config.diloco.inner_steps if config.diloco is not None else 1
+
+        for _inner_step in range(num_inner_steps):
+            for _ in range(gradient_accumulation_steps):
+                next(train_dataloader_iterator)
+
+        total_steps += num_inner_steps
+        logger.info("total steps: %d", total_steps)
+        if total_steps >= config.optim.total_steps:
+            break
+
+    CkptManager.save_data_v2(os.path.join(config.ckpt.data_path, "data"), train_dataloader, world_info.local_rank)
+
+    logger.info("skipped data up to step: %d", config.optim.total_steps)
+
+
+if __name__ == "__main__":
+    torch.manual_seed(42)
+
+    world_info = get_world_info()
+    logger = get_logger()
+
+    config = Config(**parse_argv())
+
+    skip_data(config)

scripts/worker7

@@ -0,0 +1,7 @@
+GLOBAL_RANK=3
+GLOBAL_UNIQUE_ID=worker7
+GLOBAL_ADDR=100.91.104.70
+GLOBAL_PORT=8989
+GLOO_SOCKET_IFNAME=tailscale0
+GLOBAL_WORLD_SIZE=10
+ZERO_BAND_GLOBAL_PG_TIMEOUT_SECONDS=3600

src/zeroband/checkpoint.py

@@ -361,10 +361,7 @@ def _save(self, ckpt_path: str):
 
             if self.config.data_version == "v2":
                 data_path = os.path.join(ckpt_path, "data")
-                os.makedirs(data_path, exist_ok=True)
-                with open(os.path.join(data_path, f"_{self.world_info.local_rank}.pt"), "wb") as f:
-                    state = {"data_loader": self.dataloader.state_dict()}
-                    torch.save(state, f)
+                self.save_data_v2(data_path, self.dataloader, self.world_info.local_rank)
 
                 non_error_barrier()
 
@@ -381,6 +378,13 @@ def _save(self, ckpt_path: str):
 
         gc.collect()
 
+    @staticmethod
+    def save_data_v2(data_path: str, dataloader, local_rank: int):
+        os.makedirs(data_path, exist_ok=True)
+        with open(os.path.join(data_path, f"_{local_rank}.pt"), "wb") as f:
+            state = {"data_loader": dataloader.state_dict()}
+            torch.save(state, f)
+
     def _async_save_remote(self, ckpt_path: str, remote_ckpt_path: str, blocking: bool = True) -> None:
         """asyncronously rsync a ckpt folder to a remote location. Using fsspec to handle remote cloud storage without to install
         specific libraries (e.g. s3fs).