fix stuff to make it CPU compliants

README.md

@@ -1 +1,44 @@
-# picotron
+# picotron
+
+![](assets/banière.png)
+
+- The minimalist & most-hackable repository for pre-training Llama-like models with 4D Parallelism (Data, Tensor, Pipeline, Context parallel). It is a rewrite of [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) for educational purpose. The code itself is plain and readable: **train.py, model.py and \[data|tensor|pipeline|context\]_parallel.py are all < 300 LOC**.
+
+- Performance is not yet in okay-ish but this is under active development.
+
+# Install
+
+```
+pip install -e .
+```
+
+# Quick start
+
+- GPU
+```sh
+# DP=8
+python create_config.py --out_dir tmp --exp_name llama-1B --dp 8 --model_name HuggingFaceTB/SmolLM-1.7B --num_hidden_layers 15  --grad_acc_steps 32 --mbs 4 --seq_len 1024 --hf_token <HF_TOKEN>
+
+# Locally
+torchrun --nproc_per_node 8 train.py --config tmp/llama-1B/config.json 
+
+# 3D Parallelism
+python create_config.py --out_dir tmp --dp 4 --tp 2 --pp 2 --pp_engine 1f1b --exp_name llama-7B --model_name meta-llama/Llama-2-7b-hf  --grad_acc_steps 32 --mbs 4 --seq_len 1024 --hf_token <HF_TOKEN>
+
+# Slurm
+python submit_slurm_jobs.py --inp_dir tmp/llama-7B --qos high --hf_token <HF_TOKEN>
+```
+
+-  CPU (expect it to be slow)
+
+```sh
+# 3D Parallelism on CPU
+python create_config.py --out_dir tmp --exp_name llama-1B-cpu --dp 2 --tp 2 --pp 2 --pp_engine 1f1b --model_name HuggingFaceTB/SmolLM-1.7B --num_hidden_layers 5  --grad_acc_steps 2 --mbs 4 --seq_len 128 --hf_token <HF_TOKEN> --use_cpu
+
+# Locally
+torchrun --nproc_per_node 8 train.py --config tmp/llama-1B-cpu/config.json
+```
+
+# Acknowledgements
+
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)

create_config.py

@@ -126,6 +126,7 @@ def create_single_config(
     seq_len: int,
     exp_name: str,
     use_wandb: bool = False,
+    use_cpu: bool = False,
     use_fused_adam: bool = False,
     hf_token: str = None
 ):
@@ -156,6 +157,10 @@ def create_single_config(
     config_content['distributed']['dp_size'] = dp
     config_content['distributed']['pp_size'] = pp
     config_content['distributed']['pp_engine'] = pp_engine
+    config_content['distributed']['use_cpu'] = use_cpu
+    if use_cpu:
+        config_content["environment"]["FLASH_ATTEN"] = "0"
+        config_content["distributed"]["backend"] = "gloo"
 
     config_content['logging']['use_wandb'] = use_wandb
     config_content['logging']['run_name'] = exp_name
@@ -192,6 +197,7 @@ if __name__ == "__main__":
     parser.add_argument("--seq_len", type=int, help="Sequence length", default=1024)
     parser.add_argument("--exp_name", type=str, help="Experiment name", default="dummy_exp")
     parser.add_argument("--use_wandb", action="store_true", help="Use wandb for logging")
+    parser.add_argument("--use_cpu", action="store_true", help="Use CPU for training")
     parser.add_argument("--use_fused_adam", action="store_true", help="Use fused adam")
     parser.add_argument("--hf_token", type=str, help="HF token")
 
@@ -213,6 +219,7 @@ if __name__ == "__main__":
         seq_len=args.seq_len,
         exp_name=args.exp_name,
         use_wandb=args.use_wandb,
+        use_cpu=args.use_cpu,
         use_fused_adam=args.use_fused_adam,
         hf_token=args.hf_token
     )    

picotron/data_parallel/bucket.py

@@ -68,6 +68,7 @@ class BucketManager:
             grad_type (torch.dtype, optional): Data type of gradients, defaults to torch.float32.
         """
         self.params = list(params) # Convert parameter generator to a list.
+        self.device = torch.device("cuda") if self.params[0].is_cuda else torch.device("cpu")
         self.buckets = [] # List of buckets.
         self.process_group = process_group
         self.process_group_size = dist.get_world_size(group=self.process_group)
@@ -116,7 +117,7 @@ class BucketManager:
         
         # Create tensors for storing gradients and initialize Bucket objects.
         for i in range(len(bucket_sizes)):
-            self.grad_data_list.append(torch.zeros(bucket_sizes[i], dtype=self.grad_type, device='cuda'))
+            self.grad_data_list.append(torch.zeros(bucket_sizes[i], dtype=self.grad_type, device=self.device))
             self.buckets.append(Bucket(buckets_to_params[i], self.grad_data_list[i], self.process_group))
         
         # Create gradient views for each parameter.

setup.py

@@ -1,6 +1,12 @@
 from setuptools import setup, find_packages
+
+def read_requirements():
+    with open('requirements.txt') as req:
+        return [line.strip() for line in req if line.strip() and not line.startswith('#')]
+
 setup(
-    name="picotron",  # Name of the package
+    name="picotron",
     version='0.1.0',
-    packages=find_packages(),  # Automatically find packages in the current directory
+    packages=find_packages(),
+    install_requires=read_requirements(),
 )

template/base_config.json

@@ -30,7 +30,7 @@
     "dataset": {
         "name": "roneneldan/TinyStories",
         "num_workers": 0,
-        "num_proc": 4
+        "num_proc": 1
     },
     "checkpoint": {
         "save_dir": "ckpt",

train.py

@@ -159,6 +159,8 @@ if __name__ == "__main__":
 
     dist.barrier()
 
+    print(f"rank {pgm.process_group_manager.global_rank}: Initializing model meta device", is_print_rank=is_wandb_rank)
+
     start_time = time.time()
 
     with init_model_with_dematerialized_weights():
@@ -209,7 +211,8 @@ if __name__ == "__main__":
     def _all_reduce_loss_across_dp_cp_ranks(loss, device):
         reduced_loss = torch.tensor([loss if loss is not None else 0.0], dtype=torch.float32, device=device)
         if pgm.process_group_manager.pp_is_last_stage:
-            dist.all_reduce(reduced_loss, op=dist.ReduceOp.AVG, group=pgm.process_group_manager.cp_dp_group)
+            dist.all_reduce(reduced_loss, op=dist.ReduceOp.SUM, group=pgm.process_group_manager.cp_dp_group)
+            reduced_loss /= pgm.process_group_manager.cp_dp_world_size
         return reduced_loss.item()
     
     while config["training"]["max_tokens"] is None or trained_tokens < config["training"]["max_tokens"]: