Latency and Memory Analysis of Transformer Models for Training and Inference
- llm-analysis
Many formulas or equations are floating around in papers, blogs, etc., about how to calculate training or inference latency and memory for Large Language Models (LLMs) or Transformers. Rather than doing math on papers or typing in Excel sheets, let's automate the boring stuff with llm-analysis
⚙️!
Given the specified model, GPU, data type, and parallelism configurations, llm-analysis estimates the latency and memory usage of LLMs for training or inference. With llm-analysis, one can easily try out different training/inference setups theoretically, and better understand the system performance for different scenarios.
llm-analysis helps answer questions such as:
- what batch size, data type, parallelism scheme to use to get a
feasible
(not getting OOM) andoptimal
(maximizing throughput with a latency constraint) setup for training or inference time
it takes with the given setup to do training or inference and thecost
(GPU-hours)- how the latency/memory changes if using a different model, GPU type, number of GPU, data type for weights and activations, parallelism configuration (suggesting the performance benefit of
modeling change
,hardware improvement
,quantization
,parallelism
, etc.)
Check the example use cases. With llm-analysis, you can do such analysis in minutes 🚀!
-
To install llm-analysis from pypi:
pip install llm-analysis
-
To install the latest development build:
pip install --upgrade git+https://github.com/cli99/llm-analysis.git@main
-
To install from source, clone the repo and run
pip install .
orpoetry install
(install poetry bypip install poetry
).
To integrate llm-analysis in your code, use the LLMAnalysis
class. Refer to doc LLMAnalysis for details.
LLMAnalysis
is constructed with flops and memory efficiency numbers and the following configuration classes:
ModelConfig
covers model information, i.e. max sequence length, number of transformer layers, number of attention heads, hidden dimension, vocabulary sizeGPUConfig
covers GPU compute and memory specificationsDtypeConfig
covers the number of bits used for the model weight, activation, and embeddingParallelismConfig
covers Tensor Parallelism (tp
), Pipeline Parallelism (pp
), Sequence Parallelism (sp
), Expert Parallelism (ep
),and Data Parallelism (dp
).
Then LLMAnalysis
can be queried with different arguments through the training and inference methods.
llm-analysis provides two entry functions, train and infer, for ease of use through the command line interface. Run
python -m llm_analysis.analysis train --help
or
python -m llm_analysis.analysis infer --help
to check the options or read the linked doc. Refer to the examples to see how they are used.
train
and infer
use the pre-defined name-to-configuration mappings (model_configs
, gpu_configs
, dtype_configs
) and other user-input arguments to construct the LLMAnalysis
and do the query.
The pre-defined mappings are populated at the runtime from the model, GPU, and data type configuration json
files under model_configs, gpu_configs, and dtype_configs. To add a new model, GPU or data type to the mapping for query, just add a json
description file to the corresponding folder.
llm-analysis also supports retrieving ModelConfig
from a model config json file path or Hugging Face with the model name .
- From a local model config json file, e.g.,
python -m llm_analysis.analysis train --model_name=local_example_model.json
. Check the model configurations under the model_configs folder. - From Hugging Face, e.g., use
EleutherAI/gpt-neox-20b
asmodel_name
when calling thetrain
orinfer
entry functions.python -m llm_analysis.analysis train --model_name=EleutherAI/gpt-neox-20b --total_num_gpus 32 --ds_zero 3
. With this method, llm-analysis relies ontransformers
to find the corresponding model configuration on huggingface.co/models, meaning information of newer models only exist after certain version of the transformers library. To access latest models through their names, update the installedtransformers
package.
A list of handy commands is provided to query against the pre-defined mappings as well as Hugging Face, or to dump configurations. Run python -m llm_analysis.config --help
for details.
Some examples:
python -m llm_analysis.config get_model_config_by_name EleutherAI/gpt-neox-20b
gets the ModelConfig
from the populated mapping by name, if not found, llm-analysis tries to get it from HuggingFace.
Note that LLaMA models need at least transformers-4.28.1
to retrieve, either update to a later transformers
library, or use the predefined ModelConfig
for LLaMA models (/
in model names are replaced with _
).
python -m llm_analysis.config list_gpu_configs
lists the names of all predefined GPU configurations, then you can query with
python -m llm_analysis.config get_gpu_config_by_name a100-sxm-80gb
to show the corresponding GPUConfig
.
Setting flops and memory efficiency to 1
(default) gives the lower bound of training or inference latency, as it assumes the peak hardware performance (which is never the case).
A close-to-reality flops or memory efficiency can be found by benchmarking and profiling using the input dimensions in the model.
If one has to make assumptions, for flops efficiency, literature reports up to 0.5
for large scale model training, and up to 0.7
for inference; 0.9
can be an aggressive target for memory efficiency.
llm-analysis aims to provide a lower-bound
estimation of memory usage and latency.
llm-analysis currently covers Tensor Parallelism (tp), Pipeline Parallelism (pp), Sequence Parallelism (sp), Expert Parallelism (ep), and Data Parallelism (dp).
-
tp, pp, and sp adopt the style of parallelization used in
Megatron-LM
for training andFasterTransformer
for inference -
In the training analysis, dp sharding assumes using
DeepSpeed ZeRO
orFSDP
.ds_zero
is used to specify the dp sharding strategyds_zero DeepSpeed ZeRO FSDP Sharding 0 disabled NO_SHARD No sharding 1 Stage 1 N/A Shard optimizer states 2 Stage 2 SHARD_GRAD_OP Shard gradients and optimizer states 3 Stage 3 FULL_SHARD Shard gradients, optimizer states, model parameters -
ep parallelizes the number of MLP experts across
ep_size
devices, i.e. the number of experts per GPU istotal number of experts / ep_size
. Thus for the MLP module, the number of devices for other parallelization dimensions is divided byep_size
compared to other parts of the model.
tp communication is calculated as using ring allreduce
. ep communication is calculated as using alltoall
.
dp communication time to unshard model weight when using FSDP or DeepSpeed ZeRO is estimated and compared against the compute latency, the larger value of the two is used for the overall latency.
Other dp and pp communications are ignored for now, i.e. assuming perfect computation and communication overlapping, which is not true when communication cannot overlap with compute due to dependency, or when communication is too long to hide due to slow interconnect or large data volume.
llm-analysis supports both full and selective activation recomputation.
activation_recomputation | what is checkpointed and recomputed |
---|---|
0 | No activation recomputation; requires the most amount of memory |
1 | Checkpoints the attention computation (QK^T matrix multiply, softmax, softmax dropout, and attention over V.) in the attention module of a transformer layer; as described in Reducing Activation Recomputation in Large Transformer Models. |
2 | Checkpoints the input to the attention module in a transformer layer; requires an extra forward pass on attention. |
3 | Checkpoints the input to the sequence of modules (layernom-attention-layernom) in a transformer layer; requires an extra forward pass on (layernom-attention-layernom). |
4 | Full activation recomputation stores the input to the transformer layer; requires the least amount of memory; requires an extra forward pass of the entire layer. |
Data types are expressed with the number of bits, only 32
(FP32, TF32), 16
(FP16, BF16), 8
(INT8), and 4
(INT4) bits data types are modeled for now.
Fine-tuning is modeled the same (controlled by total_num_tokens
passed to the train
entry function) as pre-training, thus assuming full (all model parameters) fine-tuning. Parameter-efficient fine-tuning (PEFT) is
in future support.
Inference assumes perfect overlapping of compute and memory operations when calculating latency, and maximum memory reuse when calculating memory usage.
TODOs (stay tuned 📻)
Check the TODOs below for what's next and stay tuned 📻! Any contributions or feedback are highly welcome!
- Add dp (across and within a node), ep (within a node), pp (across nodes) communication analysis
- Support efficient fine-tuning methods such as LoRA or Adapters
- Add FP8 datatype support
- Support CPU offloading (weight, KV cache, etc.) analysis in training and inference
- Support other hardware (e.g. CPU) for inference analysis
If you use llm-analysis in your work, please cite:
Cheng Li. (2023). LLM-Analysis: Latency and Memory Analysis of Transformer Models for Training and Inference. GitHub repository, https://github.com/cli99/llm-analysis.
or
@misc{llm-analysis-chengli,
author = {Cheng Li},
title = {LLM-Analysis: Latency and Memory Analysis of Transformer Models for Training and Inference},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/cli99/llm-analysis}},
}
Contributions and suggestions are welcome.
llm-analysis uses pre-commit to ensure code formatting is consistent. For pull requests with code contribution, please install the pre-commit (pip install pre-commit
) as well as the used hooks (pip install
in the repo), and format the code (runs automatically before each git commit) before submitting the PR.
- Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
- ZeRO: Memory Optimizations Toward Training Trillion Parameter Models
- Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM
- Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Model
- Reducing Activation Recomputation in Large Transformer Models
- Training Compute-Optimal Large Language Models
- Efficiently Scaling Transformer Inference
- Training Compute-Optimal Large Language Models
- Understanding INT4 Quantization for Transformer Models: Latency Speedup, Composability, and Failure Cases
- A Comprehensive Study on Post-Training Quantization for Large Language Models