[BE]: Deprecate Llama2 (#712)

Deprecates Llama2 utilities and mentions in the documentation. Also deprecated sentencepiece since that's only used for llama2. Code could be simplified further but from a code quality perspective it might make sense to have one more large model variant in the repo. --------- Co-authored-by: tianyu-l <[email protected]>
pytorch · Dec 3, 2024 · 4d88a8a · 4d88a8a
1 parent 6a6f755
commit 4d88a8a
Show file tree

Hide file tree

Showing 15 changed files with 12 additions and 308 deletions.
diff --git a/.ci/docker/requirements.txt b/.ci/docker/requirements.txt
@@ -2,7 +2,6 @@ torchdata >= 0.8.0
 datasets >= 2.21.0
 tomli >= 1.1.0 ; python_version < "3.11"
 tensorboard
-sentencepiece
 tiktoken
 blobfile
 tabulate

diff --git a/README.md b/README.md
@@ -3,7 +3,7 @@
 
 # torchtitan
 
-`torchtitan` is currently in a pre-release state and under extensive development. Currently we showcase pre-training **Llama 3.1**, **Llama 3**, and **Llama 2** LLMs of various sizes from scratch. To use the latest features of `torchtitan`, we recommend using the most recent PyTorch nightly.
+`torchtitan` is currently in a pre-release state and under extensive development. Currently we showcase pre-training **Llama 3.1** LLMs of various sizes from scratch. To use the latest features of `torchtitan`, we recommend using the most recent PyTorch nightly.
 
 `torchtitan` is a proof-of-concept for Large-scale LLM training using native PyTorch. It is (and will continue to be) a repo to showcase PyTorch's latest distributed training features in a clean, minimal codebase. torchtitan is complementary to and not a replacement for any of the great large-scale LLM training codebases such as Megatron, Megablocks, LLM Foundry, Deepspeed, etc. Instead, we hope that the features showcased in torchtitan will be adopted by these codebases quickly. torchtitan is unlikely to ever grow a large community around it.
 
@@ -43,7 +43,7 @@ You may want to see how the model is defined or how parallelism techniques are a
 * [torchtitan/parallelisms/pipeline_llama.py](torchtitan/parallelisms/pipeline_llama.py) - helpers for applying Pipeline Parallel to the model
 * [torchtitan/checkpoint.py](torchtitan/checkpoint.py) - utils for saving/loading distributed checkpoints
 * [torchtitan/float8.py](torchtitan/float8.py) - utils for applying Float8 techniques
-* [torchtitan/models/llama/model.py](torchtitan/models/llama/model.py) - the Llama model definition (shared for Llama 2 and Llama 3 variants)
+* [torchtitan/models/llama/model.py](torchtitan/models/llama/model.py) - the Llama 3.1 model definition
 
 ### Key features available
 
@@ -78,18 +78,15 @@ pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu
 
 ### Downloading a tokenizer
 
-`torchtitan` currently supports training Llama 3 (8B, 70B), and Llama 2 (7B, 13B, 70B) out of the box. To get started training these models, we need to download a tokenizer.model. Follow the instructions on the official [meta-llama](https://huggingface.co/meta-llama/Meta-Llama-3-8B) repository to ensure you have access to the Llama model weights.
+`torchtitan` currently supports training Llama 3.1 (8B, 70B, 405B) out of the box. To get started training these models, we need to download a tokenizer.model. Follow the instructions on the official [meta-llama](https://huggingface.co/meta-llama/Llama-3.1-8B) repository to ensure you have access to the Llama model weights.
 
-Once you have confirmed access, you can run the following command to download the Llama 3 / Llama 2 tokenizer to your local machine.
+Once you have confirmed access, you can run the following command to download the Llama 3.1 tokenizer to your local machine.
 
 ```bash
 # Get your HF token from https://huggingface.co/settings/tokens
 
-# Llama 3 or 3.1 tokenizer.model
-python torchtitan/datasets/download_tokenizer.py --repo_id meta-llama/Meta-Llama-3-8B --tokenizer_path "original" --hf_token=...
-
-# Llama 2 tokenizer.model
-python torchtitan/datasets/download_tokenizer.py --repo_id meta-llama/Llama-2-13b-hf --hf_token=...
+# Llama 3.1 tokenizer.model
+python torchtitan/datasets/download_tokenizer.py --repo_id meta-llama/Meta-Llama-3.1-8B --tokenizer_path "original" --hf_token=...
 ```
 
 ### Start a training run

diff --git a/assets/images/llama2_loss_curves.png b/assets/images/llama2_loss_curves.png
diff --git a/docs/performance.md b/docs/performance.md
@@ -1,4 +1,4 @@
-To demonstrate the effectiveness of PyTorch distributed training techniques used in torchtitan, we report both the infra metrics and loss curves of Llama 2 (13B and 70B) and Llama 3 (8B and 70B) training on 64 A100 (80GB memory) GPUs and Llama 3.1 (405B) on 128 H100 (94GB memory).
+To demonstrate the effectiveness of PyTorch distributed training techniques used in torchtitan, we report both the infra metrics and loss curves of Llama 3 (8B and 70B) training on 64 A100 (80GB memory) GPUs and Llama 3.1 (405B) on 128 H100 (94GB memory).
 We report infra metrics achieved by [FSDP2](fsdp.md) (1D parallelism) under various configurations, and loss curves for both 1D parallelism (FSDP2) and 2D parallelism (FSDP2 + Tensor Parallel) training. (We only report 2D for 405B)
 
 
@@ -34,30 +34,6 @@ Next we show the loss curves for Llama 3 8B and Llama 3 70B training with both 1
 
 ![image](../assets/images/llama3_loss_curves.png)
 
-
-## Llama 2 performance numbers
-
-Below are the WPS and MFU results which torchtitan achieves on Llama 2 models with FSDP2 on 64 A100 (80GB) GPUs.
-
-| Model size | Batch size | Activation checkpointing | WPS | MFU |
-| ----- | ----- | ----- | ----- | ----- |
-| 13B | 2 | no | 2162 | 61.1%	|
-| 13B | 2 | selective layer | 1914 | 54.1% |
-| 13B | 2 | selective op | 1904 | 53.8% |
-| 70B | 1[^3] | selective op | 355 | 50.8% |
-| 70B | 2 | full | 353 | 50.5% |
-
-We primarily use local batch size 2 (global batch size 128) in the experiments, to keep the same number of tokens per training iteration between Llama 2 and Llama 3 (since the default sequence length in Llama 2 is 4096 which is halved compared with Llama 3). In fact, for Llama 2 70B model with full activation checkpointing, the MFU can go up to 54% when local batch size is higher (but before an OOM happens).
-
-Next we show the loss curves for Llama 2 13B and Llama 2 70B training with both 1D parallelism (FSDP2) and 2D parallelism (FSDP2 + Tensor Parallel). All four models are trained 3000 steps with global batch size 128.
-In terms of activation checkpointing (AC) configs, the Llama 2 13B training jobs use selective op AC, whereas the Llama 70B training jobs use full AC. The results are shown in the picture (a TensorBoard screenshot) below[^4].
-
-![image](../assets/images/llama2_loss_curves.png)
-
 [^1]: We used HBM2e based lower TDP SXM H100(95GB) for our test, the actual peak TFLOPs number is between SXM and NVL, and we don't know its exact value. So this MFU number is lower than actual MFU because we use the peak number of SXM directly.
 
 [^2]: Since for Float8, we are not converting all the matmuls to Float8 because our fused attention implementation is not done in Float8, so this number is lower than expected.
-
-[^3]: Since the 70B training with local batch size 2 will cause an OOM error when selective activation checkpointing is used, we report the local batch size 1 case instead.
-
-[^4]: One may have noticed that for both 13B and 70B training, 1D parallelism has slightly better convergence than 2D parallelism in the first half of training. We believe this is caused by the stronger shuffling effect introduced by having more FSDP ranks in the 1D parallelism, and the difference in convergence speed should go away after switching to a randomized data loading solution.
diff --git a/multinode_trainer.slurm b/multinode_trainer.slurm
@@ -53,7 +53,8 @@ export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
 export NCCL_BUFFSIZE=2097152
 #export TORCH_DIST_INIT_BARRIER=1
 export FI_EFA_SET_CUDA_SYNC_MEMOPS=0
-CONFIG_FILE=${CONFIG_FILE:-"./train_configs/llama2_13b.toml"}
+
+CONFIG_FILE=${CONFIG_FILE:-"./train_configs/llama3_8b.toml"}
 
 dcgmi profile --pause
 # adjust sbatch --ntasks and sbatch --nodes above and --nnodes below

diff --git a/pyproject.toml b/pyproject.toml
@@ -15,7 +15,6 @@ dependencies = [
 
     # Tokenization
     "blobfile",
-    "sentencepiece",
     "tiktoken",
 
     # Miscellaneous

diff --git a/torchtitan/datasets/tokenizer/__init__.py b/torchtitan/datasets/tokenizer/__init__.py
@@ -4,7 +4,6 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from torchtitan.datasets.tokenizer.sentencepiece import SentencePieceTokenizer
 from torchtitan.datasets.tokenizer.tiktoken import TikTokenizer
 from torchtitan.datasets.tokenizer.tokenizer import Tokenizer
 
@@ -13,9 +12,7 @@
 
 def build_tokenizer(tokenizer_type: str, tokenizer_path: str) -> Tokenizer:
     logger.info(f"Building {tokenizer_type} tokenizer locally from {tokenizer_path}")
-    if tokenizer_type == "sentencepiece":
-        return SentencePieceTokenizer(tokenizer_path)
-    elif tokenizer_type == "tiktoken":
+    if tokenizer_type == "tiktoken":
         return TikTokenizer(tokenizer_path)
     else:
         raise ValueError(f"Unknown tokenizer type: {tokenizer_type}")
diff --git a/torchtitan/datasets/tokenizer/sentencepiece.py b/torchtitan/datasets/tokenizer/sentencepiece.py
diff --git a/torchtitan/models/__init__.py b/torchtitan/models/__init__.py
@@ -4,16 +4,14 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from torchtitan.models.llama import llama2_configs, llama3_configs, Transformer
+from torchtitan.models.llama import llama3_configs, Transformer
 
 models_config = {
-    "llama2": llama2_configs,
     "llama3": llama3_configs,
 }
 
-model_name_to_cls = {"llama2": Transformer, "llama3": Transformer}
+model_name_to_cls = {"llama3": Transformer}
 
 model_name_to_tokenizer = {
-    "llama2": "sentencepiece",
     "llama3": "tiktoken",
 }
diff --git a/torchtitan/models/llama/__init__.py b/torchtitan/models/llama/__init__.py
@@ -4,30 +4,12 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 #
-# Llama 2 is licensed under the LLAMA 2 Community License,
 # Copyright (c) Meta Platforms, Inc. All Rights Reserved.
 
 from torchtitan.models.llama.model import ModelArgs, Transformer
 
 __all__ = ["Transformer"]
 
-llama2_configs = {
-    "debugmodel": ModelArgs(dim=256, n_layers=8, n_heads=16),
-    "271M": ModelArgs(dim=1024, n_layers=16, n_heads=8),
-    "1B": ModelArgs(dim=2048, n_layers=18, n_heads=16),
-    "7B": ModelArgs(dim=4096, n_layers=32, n_heads=32),
-    "13B": ModelArgs(dim=5120, n_layers=40, n_heads=40),
-    "26B": ModelArgs(dim=5120, n_layers=80, n_heads=40),
-    "70B": ModelArgs(
-        dim=8192,
-        n_layers=80,
-        n_heads=64,
-        n_kv_heads=8,
-        ffn_dim_multiplier=1.3,
-        multiple_of=4096,
-    ),
-}
-
 llama3_configs = {
     "debugmodel": ModelArgs(dim=256, n_layers=8, n_heads=16, rope_theta=500000),
     "8B": ModelArgs(

diff --git a/torchtitan/models/llama/model.py b/torchtitan/models/llama/model.py
@@ -4,7 +4,6 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 #
-# Llama 2 is licensed under the LLAMA 2 Community License,
 # Copyright (c) Meta Platforms, Inc. All Rights Reserved.
 
 

diff --git a/torchtitan/parallelisms/__init__.py b/torchtitan/parallelisms/__init__.py
@@ -17,10 +17,8 @@
 ]
 
 models_parallelize_fns = {
-    "llama2": parallelize_llama,
     "llama3": parallelize_llama,
 }
 models_pipelining_fns = {
-    "llama2": pipeline_llama,
     "llama3": pipeline_llama,
 }
diff --git a/train_configs/llama2_13b.toml b/train_configs/llama2_13b.toml
diff --git a/train_configs/llama2_70b.toml b/train_configs/llama2_70b.toml