Update README.md

Update LICENSE

LICENSE

@@ -0,0 +1,203 @@
+Copyright 2024- The Hugging Face team. All rights reserved.
+
+                              Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

README.md

@@ -1,10 +1,16 @@
 # picotron
+In the spirit of [NanoGPT](https://github.com/karpathy/nanoGPT), we created Picotron: The minimalist & most-hackable repository for pre-training Llama-like models with [4D Parallelism](https://arxiv.org/abs/2407.21783) (Data, Tensor, Pipeline, Context parallel). It is designed with simplicity and **educational** purposes in mind, making it an excellent tool for learning and experimentation.
 
 ![](assets/banière.png)
+- The code itself is simple and readable: `train.py`, `model.py` and `[data|tensor|pipeline|context]_parallel.py` are all under **300** lines of code.
 
-- 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 the best but still under active development. We observed 38% MFU on a LLaMA-2-7B model using 64 H100 GPUs and nearly 50% MFU on the SmolLM-1.7B model with 8 H100 GPUs. Benchmarks will come soon
 
-- Performance is not yet in okay-ish but this is under active development.
+# Tutorial videos
+
+- A step by step tutorial on how to build Picotron distributed training framework form scratch:
+    - [Picotron tutorial (playlist)](https://www.youtube.com/playlist?list=PL-_armZiJvAnhcRr6yTJ0__f3Oi-LLi9S) 🎬
+    - [Picotron tutorial (codebase)](https://github.com/huggingface/picotron_tutorial) 👷🏻‍♂️
 
 # Install
 
@@ -13,32 +19,34 @@ pip install -e .
 ```
 
 # Quick start
+- Get a HF token [here](https://huggingface.co/settings/tokens) to download models from HuggingFace
 
 - 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>
+    ```sh
+    # To create a config file in json format under tmp by default
+    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 
+    # 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>
+    # 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>
-```
+    # 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
 
-```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
-```
+    # Locally
+    torchrun --nproc_per_node 8 train.py --config tmp/llama-1B-cpu/config.json
+    ```
 
 # Acknowledgements
 
-- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+- [Megatron-LM](https://github.com/NVIDIA/Megatron-LM)
+- [FairScale](https://github.com/facebookresearch/fairscale)
+- [LitGPT](https://github.com/Lightning-AI/lit-gpt)

picotron/checkpoint.py

@@ -52,7 +52,7 @@ def init_model_with_materialized_weights(model, model_config, save_dir):
     initialization_manager = InitializationManager(model, model_config)
     layer_names = initialization_manager.get_layer_names_in_sft_format()
 
-    print(f"Rank {pgm.process_group_manager.global_rank} responsible for {len(layer_names)} layers")
+    # print(f"Rank {pgm.process_group_manager.global_rank} responsible for {len(layer_names)} layers")
     
     if len(layer_names) == 0:
         raise Exception("Some ranks has no layers. There are too many ranks and not enough layers to distribute.")

picotron/data.py

@@ -18,7 +18,7 @@ class MicroBatchDataLoader(DataLoader):
         self.num_global_micro_batches = self.global_batch_size // self.micro_batch_size
         
         self.seq_length_per_gpu = seq_length // pgm.process_group_manager.cp_world_size
-        self.dataset = load_dataset(dataset_name, split=split)
+        self.dataset = load_dataset(dataset_name, split=split, name=subset_name)
 
         if pgm.process_group_manager.global_rank == 0:
             print(f"rank {pgm.process_group_manager.global_rank}: Creating tokenizer")

picotron/data_parallel/data_parallel.py

@@ -8,6 +8,11 @@ from picotron.data_parallel.bucket import BucketManager
 import picotron.process_group_manager as pgm
 
 class DataParallelNaive(nn.Module):
+    """
+    Naive Data Parallelism. Not used in practice. But it is a good starting point to understand how data parallelism works.
+    It implements a simple all-reduce operation to synchronize gradients across multiple processes.
+    And `no_sync` context manager to disable gradient synchronization.
+    """
     def __init__(self, module):
         """
         Initializes the DataParallel wrapper for a given module.
@@ -55,6 +60,9 @@ class DataParallelNaive(nn.Module):
         self.require_backward_grad_sync = True
 
 class DataParallelBucket(nn.Module):
+    """
+    Data Parallelism with gradient grouped into buckets to reduce the communication overhead.
+    """
     def __init__(self, module, bucket_cap_mb=25, grad_type = torch.float32):
         """
         Initialize the DataParallelBucket module.

picotron/tensor_parallel/tp_communications.py

@@ -16,8 +16,27 @@ def split_tensor_along_last_dim(tensor, num_partitions):
     last_dim_size = tensor.size()[last_dim] // num_partitions
     return torch.split(tensor, last_dim_size, dim=last_dim)
 
+class CopyToModelParallelRegion(torch.autograd.Function):
+    """
+    Copy in forward pass, all-reduce in backward pass.
+    This is the `f` function in the paper: https://arxiv.org/abs/1909.08053
+    """
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if pgm.process_group_manager.tp_world_size == 1:
+          return grad_output
+        dist.all_reduce(grad_output, op=dist.ReduceOp.SUM, group=pgm.process_group_manager.tp_group)
+        return grad_output
+
 class ReduceFromModelParallelRegion(torch.autograd.Function):
-    """All-reduce in forward pass, identity in backward pass."""
+    """
+    All-reduce in forward pass, identity in backward pass.
+    This is the `g` function in the paper: https://arxiv.org/abs/1909.08053
+    """
     @staticmethod
     def forward(ctx, x):
         if pgm.process_group_manager.tp_world_size == 1:
@@ -52,27 +71,6 @@ class GatherFromModelParallelRegion(torch.autograd.Function):
         chunks = split_tensor_along_last_dim(grad_output, pgm.process_group_manager.tp_world_size)
         return chunks[pgm.process_group_manager.tp_rank].contiguous()
 
-class CopyToModelParallelRegion(torch.autograd.Function):
-    """Copy in forward pass, all-reduce in backward pass."""
-    @staticmethod
-    def forward(ctx, x):
-        return x
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        if pgm.process_group_manager.tp_world_size == 1:
-          return grad_output
-        dist.all_reduce(grad_output, op=dist.ReduceOp.SUM, group=pgm.process_group_manager.tp_group)
-        return grad_output
-
-def linear_with_all_reduce(x, weight, bias):
-    input_parallel = CopyToModelParallelRegion.apply(x)
-    output = F.linear(input_parallel, weight, bias) # XW_i^T + b, output is Y_i
-    return output
-
-def linear_with_async_all_reduce(x, weight, bias):
-    return LinearWithAsyncAllReduce.apply(x, weight, bias)
-
 class LinearWithAsyncAllReduce(torch.autograd.Function):
     @staticmethod
     def forward(ctx, input_, weight, bias):
@@ -100,4 +98,12 @@ class LinearWithAsyncAllReduce(torch.autograd.Function):
         grad_weight = grad_output.t() @ input_                              # (out_size, b*s) @ (b*s, input_size) -> (out_size, input_size)
         grad_bias = grad_output.sum(0) if ctx.use_bias else None
         input_gradient_all_reduce_handle.wait()
-        return grad_input, grad_weight, grad_bias
+        return grad_input, grad_weight, grad_bias
+
+def linear_with_all_reduce(x, weight, bias):
+    input_parallel = CopyToModelParallelRegion.apply(x)
+    output = F.linear(input_parallel, weight, bias) # XW_i^T + b, output is Y_i
+    return output
+
+def linear_with_async_all_reduce(x, weight, bias):
+    return LinearWithAsyncAllReduce.apply(x, weight, bias)

picotron/utils.py

@@ -52,7 +52,6 @@ def get_num_params(model):
     For DP: Parameters are replicated, so only count once
     
     Note: 
-    LayerNorm: Split across TP ranks for sequence parallelism
     FSDP: Parameters are sharded across data parallel ranks
     """
     tp_world_size = pgm.process_group_manager.tp_world_size
@@ -86,4 +85,11 @@ def assert_no_meta_tensors(model):
         if buffer.device == torch.device("meta"):
             meta_tensors.append(f"Buffer '{name}' with shape {buffer.shape}")
     
-    assert len(meta_tensors) == 0, f"Found {len(meta_tensors)} meta tensors:\n" + "\n".join(meta_tensors)
+    assert len(meta_tensors) == 0, f"Found {len(meta_tensors)} meta tensors:\n" + "\n".join(meta_tensors)
+
+def average_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.SUM, group=pgm.process_group_manager.cp_dp_group)
+        reduced_loss /= pgm.process_group_manager.cp_dp_world_size
+    return reduced_loss.item()

train.py

@@ -1,10 +1,6 @@
 """Training script for LLaMA model.
-CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun --nproc_per_node 1 --master_addr localhost --master_port 25500 train.py --config tmp/dummy/360M_131K.json
-CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun --nproc_per_node 2 --master_addr localhost --master_port 25500 train.py --config tmp/dummy/360M_131K.json
 CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --config tmp/dummy/llama2_7b_benchmark.json
-CUDA_DEVICE_MAX_CONNECTIONS=1 torchrun --nproc_per_node 8 --master_addr localhost --master_port 25500 train.py --config tmp/dummy/360M_131K.json
-CUDA_DEVICE_MAX_CONNECTIONS=1 debugpy-run -p 5678 -m torch.distributed.run -- --nproc_per_node=2 --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:29400 train.py --config tmp/dummy/360M_131K.json
-#VERBOSE=0 torchrun --nproc_per_node 4 --master_addr localhost --master_port 25500 train.py --pp_size 2 --dp_size 2
+CUDA_DEVICE_MAX_CONNECTIONS=1 debugpy-run -p 5678 -m torch.distributed.run -- --nproc_per_node=4 --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:29400 train.py --config tmp/dummy/llama2_7b_benchmark.json
 """
 import os
 import inspect
@@ -20,7 +16,7 @@ from transformers import AutoConfig
 from picotron.context_parallel.context_parallel import apply_context_parallel
 from picotron.tensor_parallel.tensor_parallel import apply_tensor_parallel
 import picotron.process_group_manager as pgm
-from picotron.utils import set_all_seed, print, to_readable_format, get_mfu, get_num_params
+from picotron.utils import average_loss_across_dp_cp_ranks, set_all_seed, print, to_readable_format, get_mfu, get_num_params
 from picotron.checkpoint import CheckpointManager
 from picotron.checkpoint import init_model_with_dematerialized_weights, init_model_with_materialized_weights
 from picotron.data import MicroBatchDataLoader
@@ -111,8 +107,9 @@ if __name__ == "__main__":
         device=device,
         num_workers=config["dataset"]["num_workers"],
         num_proc=config["dataset"]["num_proc"],
-        num_samples=config["training"]["num_samples"],
-        subset_name=config["dataset"]["subset_name"],
+        num_samples=config["training"].get("num_samples", None),
+        subset_name=config["dataset"].get("subset_name", None),
+        split=config["dataset"].get("split", "train")
     )
 
     dist.barrier()
@@ -146,9 +143,10 @@ if __name__ == "__main__":
     if pgm.process_group_manager.global_rank == 0:
         print(f"rank {pgm.process_group_manager.global_rank}: Creating model config")
         model_config = AutoConfig.from_pretrained(config["model"]["name"])
-        model_config.num_hidden_layers = config["model"]["num_hidden_layers"]
-        model_config.num_attention_heads = config["model"]["num_attention_heads"]
-        model_config.num_key_value_heads = config["model"]["num_key_value_heads"]
+        # twist the model structure if specified in the config file
+        model_config.num_hidden_layers = model_config.num_hidden_layers if "num_hidden_layers" not in config["model"] else config["model"]["num_hidden_layers"]
+        model_config.num_attention_heads = model_config.num_attention_heads if "num_attention_heads" not in config["model"] else config["model"]["num_attention_heads"]
+        model_config.num_key_value_heads = model_config.num_key_value_heads if "num_key_value_heads" not in config["model"] else config["model"]["num_key_value_heads"]
         model_config.max_position_embeddings = config["training"]["seq_length"]
         objects = [model_config]
     else:
@@ -208,13 +206,6 @@ if __name__ == "__main__":
         step, trained_tokens = checkpoint_manager.load_checkpoint(model, optimizer, config["checkpoint"]["load_path"])
     
     dist.barrier()
-        
-    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.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"]:
         step_start_time = time.time()
@@ -230,7 +221,7 @@ if __name__ == "__main__":
         else:
             loss = train_step(model, data_loader, device)
             
-        loss = _all_reduce_loss_across_dp_cp_ranks(loss, device)
+        loss = average_loss_across_dp_cp_ranks(loss, device)
         
         optimizer.step()
         trained_tokens += tokens_per_step