Ring Flash Attention

This repo implements RingAttention using FlashAttention. The current implementation supports:

• varlen (packing samples) api, corresponding to flash_attn_varlen_func:
  • ring_flash_attn_varlen_func: A basic implementation of ring attention.
  • zigzag_ring_flash_attn_varlen_func: an more compute-balanced version of ring attention. More details in issue#2.
  • llama3_flash_attn_varlen_func: The context parallelism used in llama3 tech report with extra design for varlen and low memory overhead. Although technically not ring attention, this is recommended for most varlen use cases, as it offers a less intrusive alternative for training frameworks with fewer data manipulations and better arithmetic precision.
• batch api, corresponding to flash_attn_func:
  • ring_flash_attn_func: basic ring attention.
  • zigzag_ring_flash_attn_func: An more compute balanced version of ring attention, see issue#2.
  • stripe_flash_attn_func: Stripe attention version of ring_flash_attn_func, the block size is set to 1 to use flash_attn api, see: https://arxiv.org/abs/2311.09431
• huggingface model adapter. Here is an example to use the adapter: OpenRLHF/OpenRLHF/pull#439.

Note that

• Each function includes *_func, *_kvpacked_func, *_qkvpacked_func variants.
• The varlen versions (except the llama3 version) only support passing one cu_seqlens.

Performance Summary

The following table summarizes the performance of the implemented APIs:

batch api	GPU	theoretic flash_attn	ring_attn	zigzag_ring	stripe_attn
fwd only (iter/sec)	8xH800	591.5 / 8 = 73.9	38.5	63.0	55.0
			52.1%	85.2%	74.4%
fwd + bwd (iter/sec)	8xH800	154.7 / 8 = 19.3	10.4	17.4	16.0
			53.9%	90.2%	82.9%
fwd only (iter/sec)	8xA100	373.4 / 8 = 46.7	24.0	38.2	32.5
			51.4%	81.7%	69.6%
fwd + bwd (iter/sec)	8xA100	94.7 / 8 = 11.8	6.2	10.6	9.75
			52.5%	89.8%	82.6%
varlen api	GPU	theoretic flash_attn	ring_attn	zigzag_ring	llama3_attn
fwd only (iter/sec)	8xH800	852.4 / 8 = 106.6	52.4	74.8	60.8
			49.1%	70.2%	57.0%
fwd + bwd (iter/sec)	8xH800	225.4 / 8 = 28.2	14.4	21.4	16.4
			51.1%	75.9%	58.1%
fwd only (iter/sec)	8xA100	532.3 / 8 = 66.5	33.1	47.9	34.3
			49.8%	72.0%	51.6%
fwd + bwd (iter/sec)	8xA100	133.8 / 8 = 16.7	8.7	13.4	9.7
			52.1%	80.2%	58.0%

Note that

• The code of the benchmark is in benchmark, its configuration matches the Meta-Llama-3.1-8B setting, with a total sequence of length 8k per GPU.
• When running the benchmark with with 8 gpu, the flash attn code is running with 1/8 computation of ring attention, as flash attn code is running 8*1^2, while the ring attn code is running 1*8^2.
• NVLink between GPUs are required for high performance.
• Please remember to adapt the RoPE offset for different api.

Installation

pip install ring-flash-attn

or use the following command to build from source:

git clone https://github.com/zhuzilin/ring-flash-attention.git
cd ring-flash-attention
pip install .

TODOs

• Implement ring_flash_attn_varlen_qkvpacked_func
• Implement zigzag_ring_flash_attn_qkvpacked_func issue#2
• Implement stripe_flash_attn_qkvpacked_func
• Implement zigzag_ring_flash_attn_varlen_qkvpacked_func
• Implement *_kvpacked_func and *_func variant for all APIs
• Optimize *_varlen_func Implement llama3_flash_attn_varlen_func
• Add an example to train llama Implement adapter for huggingface model
• Implement zigzag_llama3_flash_attn_varlen_func

Test

torchrun --nproc_per_node 8 test/test_llama3_flash_attn_varlen_func.py
torchrun --nproc_per_node 8 test/test_ring_flash_attn_func.py
torchrun --nproc_per_node 8 test/test_ring_flash_attn_varlen_func.py
torchrun --nproc_per_node 8 test/test_zigzag_ring_flash_attn_func.py
torchrun --nproc_per_node 8 test/test_zigzag_ring_flash_attn_varlen_func.py
torchrun --nproc_per_node 8 test/test_stripe_flash_attn_func.py

Benchmark

torchrun --nproc_per_node 8 benchmark/benchmark_kvpacked_func.py
torchrun --nproc_per_node 8 benchmark/benchmark_varlen_kvpacked_func.py

Known Limitations

There are some arithmetic errors with the current implementation. The reason for them is probably that flash attention will return bf16 value for each block, so we cannot accumluate the values with the original fp32 ones.

And also because we need to save extra fp32 buffer during computation, the memory usage would be higher than theoretic limit.

Also,

• dropout is not supported at the moment, because it's hard to save all the rng_states.
• window_size is not supported, because it will be really tricky to implement a varlen version with window_size.