Update configs, add results

README.md

@@ -14,6 +14,20 @@ We've been very happy to see FlashAttention being widely adopted in such a short
 time after its release. This [page](https://github.com/HazyResearch/flash-attention/blob/main/usage.md)
 contains a partial list of places where FlashAttention is being used.
 
+## Full model code and training script
+
+We have released the full GPT model
+[implementation](https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/models/gpt.py).
+We also provide optimized implementations of other layers (e.g., MLP, LayerNorm,
+cross-entropy loss, rotary embedding). Overall this speeds up training by 3-5x
+compared to the baseline implementation from Huggingface, reaching up to 189
+TFLOPs/sec per A100, equivalent to 60.6\% model FLOPs utilization (we don't need
+any activation checkpointing). 
+
+We also include a training
+[script](https://github.com/HazyResearch/flash-attention/tree/main/training) to
+train GPT2 on Openwebtext and GPT3 on The Pile.
+
 ## Triton implementation of FlashAttention
 
 Phil Tillet (OpenAI) has an experimental implementation of FlashAttention in Triton:

training/Dockerfile

@@ -65,14 +65,6 @@ ENV PIP_NO_CACHE_DIR=1
 # # apex and pytorch-fast-transformers take a while to compile so we install them first
 # TD [2022-04-28] apex is already installed. In case we need a newer commit:
 # RUN pip install --upgrade --force-reinstall --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--fast_multihead_attn" --global-option="--fmha" --global-option="--fast_layer_norm" --global-option="--xentropy" git+https://github.com/NVIDIA/apex.git#egg=apex
-# TD [2021-10-28] pytorch-fast-transformers doesn't have a wheel compatible with CUDA 11.3 and Pytorch 1.10
-# So we install from source, and change compiler flag -arch=compute_60 -> -arch=compute_70 for V100
-# RUN pip install pytorch-fast-transformers==0.4.0
-# RUN pip install git+git://github.com/idiap/fast-transformers.git@v0.4.0  # doesn't work on V100
-RUN git clone https://github.com/idiap/fast-transformers \
-    && sed -i 's/\["-arch=compute_60"\]/\["-arch=compute_70"\]/' fast-transformers/setup.py \
-    && pip install fast-transformers/ \
-    && rm -rf fast-transformers
 
 # xgboost conflicts with deepspeed
 RUN pip uninstall -y xgboost && DS_BUILD_UTILS=1 DS_BUILD_FUSED_LAMB=1 pip install deepspeed==0.7.5

training/README.md

-Examples of how FlashAttention can be integrated into a model (e.g., GPT, ViT)
-and trained end-to-end.
-We also added optimized implementations of other layers (e.g., MLP, LayerNorm,
-cross-entropy loss, rotary embedding).
+# Optimized Transformer implementation
+This repo contains examples of how FlashAttention can be integrated into a model
+(e.g., GPT, ViT) and trained end-to-end. We also provide optimized
+implementations of other layers (e.g., MLP, LayerNorm, cross-entropy loss,
+rotary embedding). Overall this speeds up training by 3-5x compared to the
+baseline implementation from Huggingface, reaching up to 189 TFLOPs/sec per A100,
+equivalent to 60.6\% model FLOPs utilization (we don't need any activation
+checkpointing). All without changing the model architecture (i.e., no
+approximation).
 
 Goals:
 - Performance: we optimize for model speed and memory, especially on 1-node
@@ -29,17 +34,36 @@ Non-goals (and other resources):
 
 The GPT model is implemented
 [here](https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/models/gpt.py).
+And here's an example to construct the GPT3-1.3B model with rotary embedding:
+```python
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+from flash_attn.models.gpt import GPTLMHeadModel
+
+seqlen = 2048
+hidden_dim = 2048
+nheads = 16
+n_layer = 24
+rotary_emb_fraction = 0.5
+config = GPT2Config(vocab_size=50257, n_positions=seqlen, n_embd=hidden_dim,
+                    n_layer=n_layer, n_head=nheads, 
+                    scale_attn_by_inverse_layer_idx=True, 
+                    rotary_emb_fraction=rotary_emb_fraction,
+                    use_flash_attn=True, fused_dense_gelu_dense=True,
+                    fused_bias_fc=True, fused_dropout_add_ln=True, 
+                    pad_vocab_size_multiple=8)
+model = GPTLMHeadModel(config)
+```
 
 We provide the following optimized components:
 
-- FlashAttention: fast and memory-efficient exact attention. This makes
+1. FlashAttention: fast and memory-efficient exact attention. This makes
 attention much faster and saves a lot of activation memory. As a result we don't need
 to use any activation checkpointing.
 ```sh
 pip install flash-attn
 ```
 
-- Fused matmul + bias (forward and backward), and fused matmul + bias + gelu
+2. Fused matmul + bias (forward and backward), and fused matmul + bias + gelu
 (forward and backward), adapted from Apex's
 [FusedDense](https://github.com/NVIDIA/apex/tree/master/apex/fused_dense). We
 make it work for bfloat16. For best performance, you should use CUDA >= 11.8. CuBLAS versions before
@@ -47,16 +71,16 @@ this doesn't have the best matmul + bias + gelu performance for bfloat16.
 ```sh
 cd ../csrc/fused_dense_lib && pip install .
 ```
-- Optimized cross-entropy loss, adapted from Apex's
+3. Optimized cross-entropy loss, adapted from Apex's
 [Xentropy](https://github.com/NVIDIA/apex/tree/master/apex/contrib/xentropy). We make it work for bfloat16 and support in-place backward to save memory.
 ```sh
 cd ../csrc/xentropy && pip install .
 ```
-- Fused rotary embedding:
+4. Fused rotary embedding:
 ```sh
 cd ../csrc/rotary && pip install .
 ```
-- Fused dropout + residual + LayerNorm, adapted from Apex's
+5. Fused dropout + residual + LayerNorm, adapted from Apex's
 [FastLayerNorm](https://github.com/NVIDIA/apex/tree/master/apex/contrib/layer_norm). We add dropout and residual, and make it work for both pre-norm and post-norm architecture.
 This only supports a limited set of dimensions, see `csrc/layer_norm/ln_fwd_cuda_kernel.cu`.
 ```sh
@@ -65,8 +89,9 @@ cd ../csrc/layer_norm && pip install .
 
 ## Training
 
-Feel free to use the model in your training setup. We also provide here training
-scripts to train GPT2 on Openwebtext and GPT3 on The Pile as examples.
+We also provide here training scripts to train GPT2 on Openwebtext and GPT3 on
+The Pile as examples. Feel free to use the model in your own training setup as
+well.
 
 We use [Hydra](https://hydra.cc/) for configuration,
 [Pytorch-Lightning](https://github.com/Lightning-AI/lightning) for training, and
@@ -75,12 +100,20 @@ We use [Hydra](https://hydra.cc/) for configuration,
 We use the template from `https://github.com/ashleve/lightning-hydra-template`.
 Please read the instructions there to understand the repo structure.
 
+### Requirements
+
+Python 3.8+, Pytorch 1.12+, torchvision, einops, timm, hydra-core,
+hydra-colorlog, python-dotenv, rich, pytorch-lightning, triton, flash-attn.
+We recommend CUDA 11.8 (e.g., using the Nvidia's Pytorch Docker image from https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)
+
+We provide a Dockerfile that lists all the required packages.
+
 ### Dataset preparation
 
 Running the training command would automatically download the datasets
 (Openwebtext, Pile), tokenize with the GPT2 tokenizer, concatenate all the
 tokens, then save this cache to disk. Alternatively, you can also prepare the
-datasets as a separate steps.
+datasets as a separate step.
 
 The cached datasets are saved to `${DATA_DIR}/openwebtext` and
 `${DATA_DIR}/the_pile`. If `${DATA_DIR}` is not set, they will be saved to
@@ -98,36 +131,101 @@ This takes around 1h on a 64-core CPU. The processed dataset has size 17GB.
 export PYTHONPATH=$PWD:$PYTHONPATH
 pytest -q -s tests/datamodules/test_language_modeling_hf.py -k "pile"
 ```
-This takes around 20h on a 96-core CPU. The processed dataset has size 699GB.
+This takes around 20h on a 64-core CPU. The processed dataset has size 699GB.
 
 ### GPT2 training on Openwebtext
 To train GPT2 on Openwebtext with 8 GPUs:
 ```sh
-python run.py experiment=owt/gpt2s-flash trainer.devices=8
-python run.py experiment=owt/gpt2m-flash trainer.devices=8
-python run.py experiment=owt/gpt2l-flash trainer.devices=8
-python run.py experiment=owt/gpt2xl-flash trainer.devices=8
+python run.py experiment=owt/gpt2s-flash trainer.devices=8  # 125M
+python run.py experiment=owt/gpt2m-flash trainer.devices=8  # 355M
+python run.py experiment=owt/gpt2l-flash trainer.devices=8  # 760M
+python run.py experiment=owt/gpt2xl-flash trainer.devices=8  # 1.6B
 ```
 The default parameters are set for 8 x A100 80GB.
 
 To train with bf16 instead of fp16, add `trainer.precision=bf16`.
-To adjust device batch size to fit GPU memory (the global batch size stays the
-same, and gradient accumulation is calculated automatically), set `datamodule.batch_size=blah`.
 
 ### GPT3 training on The Pile
 To train GPT3 on The Pile with 8 GPUs:
 ```sh
-python run.py experiment=pile/gpt3s-flash trainer.devices=8
-python run.py experiment=pile/gpt3m-flash trainer.devices=8
-python run.py experiment=pile/gpt3l-flash trainer.devices=8
-python run.py experiment=pile/gpt3xl-flash trainer.devices=8
+python run.py experiment=pile/gpt3s-flash trainer.devices=8  # 125M
+python run.py experiment=pile/gpt3m-flash trainer.devices=8  # 355M
+python run.py experiment=pile/gpt3l-flash trainer.devices=8  # 760M
+python run.py experiment=pile/gpt3xl-flash trainer.devices=8  # 1.3B
+python run.py experiment=pile/gpt3-2.7B-flash-hdim128 trainer.devices=8  # 2.7B
 ```
-The default parameters are set for 8 x A100 80GB.
+The default parameters are set for 8 x A100 80GB. We train with bf16 by default.
 
-## Requirements
+To train with rotary embedding, run the experiments `pile/gpt3{s,m,l,xl**-flash-rotary**.
 
-Python 3.8+, Pytorch 1.12+, torchvision, einops, timm, hydra-core,
-hydra-colorlog, python-dotenv, rich, pytorch-lightning, triton, flash-attn.
-We recommend CUDA 11.8 (e.g., using the Nvidia's Pytorch Docker image from https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch)
+### Training options
 
-We provide a Dockerfile that lists all the required packages.
+**Gradient accumulation**: to adjust device batch size to fit into GPU memory
+(the global batch size stays the same, and gradient accumulation is calculated
+automatically), set `datamodule.batch_size=blah**.
+
+**Multi-node**: to train on multiple nodes, add `trainer.num_nodes=blah`.
+
+**Speed benchmarking**: to print out iteration time, add `+callbacks.speed_monitor.verbose=True`.
+
+**Resumable training**: set a name to the run, and then set `resume=True` when
+you resume. Training will restart at exactly the same batch.
+```sh
+python run.py experiment=pile/gpt3s-flash trainer.devices=8 name=pile-gpt3s-flash resume=True
+```
+
+## Training speed
+
+We measure the wallclock training speed on one node with 8 x A100 80GB SXM4 80GB (400W) with NVLink.
+
+FLOPs are calculated using the formula from the [Megatron-LM
+paper](https://arxiv.org/abs/2104.04473) (Section 5.1), except we scale by 3/4
+to get the model FLOPs (instead of hardware FLOPs with activation
+checkpointing).
+
+
+### GPT2 (sequence length 1024)
+
+![GPT2 speedup](../assets/gpt2_training_efficiency.jpg)
+
+The implementation in this repo (FlashAttention) is 3-4x faster than the
+baseline implementation from Huggingface.
+
+### GPT3 (sequence length 2048)
+
+![GPT3 speedup](../assets/gpt3_training_efficiency.jpg)
+
+The implementation in this repo (FlashAttention) is 3-5x faster than the
+baseline implementation from Huggingface.
+
+For the GPT3-2.7B model, we set head dimension to 128 (instead of 80) for better efficiency.
+
+We include here more details on the training speed with FlashAttention on 8 x
+A100 80GB.
+
+| Model     | Batch size (tokens) | Through put (tokens/sec) | Hours / 1B tokens |
+| --------- | ------------------- | ------------------------ | ----------------- |
+| GPT3-125M | 0.5M                | 1310k                    |              0.21 |
+| GPT3-355M | 0.5M                | 503k                     |              0.55 |
+| GPT3-760M | 0.5M                | 245k                     |              1.13 |
+| GPT3-1.3B | 1M                  | 169k                     |              1.64 |
+| GPT3-2.7B | 1M                  | 85k                      |              3.27 |
+
+As an example, this means that one can train a GPT3-1.3B model on 26B tokens
+(compute-optimal according to Chinchilla scaling) in about 43 hours on 8 x A100.
+
+## Training quality
+
+We include here the loss curve for GPT2 on Openwebtext, trained for 200B tokens.
+For GPT2, the runs with FlashAttention yield the same loss curve as the runs
+with the baseline implementation from Huggingface for 125M and 355M models. For
+larger models the baseline implementation just takes too long.
+
+![GPT2 training curve](../assets/gpt2_training_curve.jpg)
+
+We include here the loss curve for GPT3 on The Pile, trained for 400B tokens.
+The 125M, 355M, 760M models have batch size 512k tokens so this translates to
+800k training steps, while the 1.3B and 2.7B models have batch size 1M tokens,
+which translates to 400k training steps.
+
+![GPT3 training curve](../assets/gpt3_training_curve.jpg)

training/configs/experiment/owt/gpt2l-flash.yaml

@@ -28,7 +28,7 @@ defaults:
 
 datamodule:
   # batch_size: 16
-  batch_size: ${eval:"4 if ${train.gpu_mem} < 24 else (8 if ${train.gpu_mem} < 40 else 16)"}
+  batch_size: ${eval:"4 if ${train.gpu_mem} < 24 else (8 if ${train.gpu_mem} < 40 else (16 if ${train.gpu_mem} < 80 else 32))"}
 
 trainer:
   # strategy: null

training/configs/experiment/owt/gpt2m-flash.yaml

@@ -4,13 +4,13 @@ defaults:
   - override /model/gpt2model: gpt2-medium
 
 # Can enable mlp_checkpoint_lvl to fit batch_size 32 to A100 40GB
-model:
-  config:
-    mlp_checkpoint_lvl: 1
+# model:
+#   config:
+#     mlp_checkpoint_lvl: 1
 
 datamodule:
   # batch_size: 32
-  batch_size: ${eval:"8 if ${train.gpu_mem} < 24 else (16 if ${train.gpu_mem} < 40 else 32)"}
+  batch_size: ${eval:"8 if ${train.gpu_mem} < 24 else (16 if ${train.gpu_mem} < 40 else (32 if ${train.gpu_mem} < 80 else 64))"}
 
 train:
   optimizer:

training/configs/experiment/owt/gpt2xl-flash.yaml

@@ -10,8 +10,8 @@ defaults:
 #     mlp_checkpoint_lvl: 1
 
 datamodule:
-  batch_size: ${eval:"2 if ${train.gpu_mem} < 24 else (4 if ${train.gpu_mem} < 40 else 8)"}
-  # With adamw-zero optimizer:
+  batch_size: ${eval:"2 if ${train.gpu_mem} < 24 else (4 if ${train.gpu_mem} < 40 else (8 if ${train.gpu_mem} < 80 else 16))"}
+  # With adamw-zero optimizer, on A100 40GB:
   # checkpoint_lvl=1, batch size = 4: mem 37GB, 4650ms / batch of 512 (285ms * 15 + 375ms * 1)
   # checkpoint_lvl=1, batch size = 8: mem 46GB, 4330ms / batch of 512 (530ms * 7 + 620ms * 1)
   # checkpoint_lvl=2, batch size = 8: mem 41GB, 4570ms / batch of 512 (560ms * 7 + 650ms * 1)

training/configs/experiment/owt/gpt2xl-hf.yaml

+# @package _global_
+defaults:
+  - /experiment/owt/gpt2l-hf.yaml
+  - override /model/gpt2model: gpt2-xlarge
+
+datamodule:
+  batch_size: 1

training/configs/experiment/pile/gpt3-2.7B-flash-8k.yaml

 # @package _global_
 defaults:
-  - /experiment/pile/gpt2xl-flash-8k.yaml
+  - /experiment/pile/gpt3xl-flash-8k.yaml
 
 model:
   config:

training/configs/experiment/pile/gpt3-2.7B-flash-hdim128-rotary-8k.yaml

 # @package _global_
 defaults:
-  - /experiment/pile/gpt2xl-flash-rotary-8k.yaml
+  - /experiment/pile/gpt3xl-flash-rotary-8k.yaml
 
 model:
   config:

training/configs/experiment/pile/gpt3-2.7B-flash-hdim128-rotary.yaml

 # @package _global_
 defaults:
-  - /experiment/pile/gpt2xl-flash-rotary.yaml
+  - /experiment/pile/gpt3xl-flash-rotary.yaml
 
 model:
   config:

training/configs/experiment/pile/gpt3-2.7B-flash-hdim128.yaml

+# @package _global_
+defaults:
+  - /experiment/pile/gpt3xl-flash.yaml
+
+model:
+  config:
+    n_embd: 2560
+    n_head: 20  # Headdim 128 is faster than headdim 80
+    n_layer: 32
+    initializer_range: ${eval:"(2 / (${.n_embd} * 5)) ** 0.5"}
+    mlp_checkpoint_lvl: 0
+
+datamodule:
+  batch_size: ${eval:"1 if ${train.gpu_mem} < 40 else (2 if ${train.gpu_mem} < 80 else 4)"}
+
+train:
+  optimizer:
+    lr: 1.6e-4

training/configs/experiment/pile/gpt3-2.7B-flash-rotary-8k.yaml

 # @package _global_
 defaults:
-  - /experiment/pile/gpt2xl-flash-rotary-8k.yaml
+  - /experiment/pile/gpt3xl-flash-rotary-8k.yaml
 
 model:
   config:

training/configs/experiment/pile/gpt3-2.7B-flash-rotary.yaml

 # @package _global_
 defaults:
-  - /experiment/pile/gpt2xl-flash-rotary.yaml
+  - /experiment/pile/gpt3xl-flash-rotary.yaml
 
 model:
   config:

training/configs/experiment/pile/gpt3-2.7B-flash.yaml

+# @package _global_
+defaults:
+  - /experiment/pile/gpt3xl-flash.yaml
+
+model:
+  config:
+    n_embd: 2560
+    n_head: 32
+    n_layer: 32
+    initializer_range: ${eval:"(2 / (${.n_embd} * 5)) ** 0.5"}
+    mlp_checkpoint_lvl: 0
+
+datamodule:
+  batch_size: ${eval:"1 if ${train.gpu_mem} < 40 else (2 if ${train.gpu_mem} < 80 else 4)"}
+
+train:
+  optimizer:
+    lr: 1.6e-4

training/configs/experiment/pile/gpt3-2.7B-hf-hdim128.yaml

+# @package _global_
+defaults:
+  - /experiment/pile/gpt3xl-hf.yaml
+
+model:
+  config:
+    n_embd: 2560
+    n_head: 128
+    n_layer: 32
+
+# OOM on A100 80GB even with batch_size = 1
+datamodule:
+  batch_size: 1
+
+train:
+  optimizer:
+    lr: 1.6e-4

training/configs/experiment/pile/gpt3-2.7B-hf.yaml

+# @package _global_
+defaults:
+  - /experiment/pile/gpt3xl-hf.yaml
+
+model:
+  config:
+    n_embd: 2560
+    n_head: 32
+    n_layer: 32
+
+datamodule:
+  batch_size: 1
+
+train:
+  optimizer:
+    lr: 1.6e-4