ONNX Runtime Training provides a capability trading node/subgraph re-computations for better memory efficiency. Specifically, a list of re-computable operators is pre-defined, with which memory optimizer graph transformer will iterate the graph to find all re-computable subgraph candidates.
When training with ORTModule, by default, the graph transformer will scan the execution graph to find all eligible subgraphs to recompute, along with sizes that can be saved. Users can pick up some of the subgraphs to enable by environment variables.
Classical scenarios include:
-
ORTModuleruns a model with batch size B (for example 2^N), the memory bandwidth and compute are not fully saturated, while it hits OOM to run a bigger batch size (for example 2^(N+1)). -
For big models,
ORTModulefails to run the minimum allowed batch size, so performance can be compromised for a successful run.
Not all models and recipes need this optimizer technique. Imagine if your training recipe uses a batch size 6 (GPU compute and memory are fully saturated), and you don't need bump it to 8 to maintain a fixed global batch size. Enabling recompute maybe not bring better throughput on batch size 8 than the original batch size 6.
Make sure ONNX Runtime training wheel is installed and correctly configured.
Integrate models using ORTModule.
model = build_model()
+ from onnxruntime.training.ortmodule import ORTModule
+ model = ORTModule(model)There are two modes to enable the memory optimizations:
- Transformer layerwise recompute, e.g. aggressively recompute all supported nodes within each transformer layer (usually including attention and mlp sublayers), enabled by
export ORTMODULE_MEMORY_OPT_LEVEL=1. In this mode,ORTMODULE_MEMORY_OPT_CONFIGenv values passed by users are not respected. - Manual selected subgraph recompute, enabled by
export ORTMODULE_MEMORY_OPT_LEVEL=0andexport ORTMODULE_MEMORY_OPT_CONFIG=<config file path>. This is an advanced usage, that allows users to find the most suitable graphs to recompute, at the cost of overhead to look for the best plans. The format for its content is:[ "<plan1 config>", "<plan2 config>", ... ]
- Set memory optimization level to be TRANSFORMER_LAYERWISE_RECOMPUTE, by
export ORTMODULE_MEMORY_OPT_LEVEL=1 - Run the training as usual; check the logs, you could find something like this if the current log level <= LogLevel.INFO:
Memory Optimizer : ON : Memory Optimization Level: [TRANSFORMER_LAYERWISE_RECOMPUTE], Optimization Config: mem_opt.json Configs Freq Max Saving(Bytes) Saving Symbolic(Bytes) - Plan 1 : ON : Reshape+Where+:1:-1 1 134,217,728 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1**2 - Plan 2 : ON : BiasSoftmax+:1:-1 1 134,086,656 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 3 : ON : Cast+:1:-1 1 67,043,328 64.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 4 : ON : BiasGelu+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 5 : ON : FusedMatMul+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 6 : ON : Add+:1:-1 1 5,237,760 5120.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 7 : ON : Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+:1:-1 1 4,096 4.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - Plan 8 : OFF : Cast+:2:-1 1 2,048 2.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - As shown above,
Configis a string representative for a re-computable subgraph. All are enabled for recompute in this case. - By
export ORTMODULE_MEMORY_OPT_LEVEL=2, all plans including compromised recomptable subgraphs will also be enabled.
- Be noted
ORTMODULE_MEMORY_OPT_LEVELis by default be 0. Run the training as usual; then stop it after training a few steps. - Check the logs, you could find something like this if the current log level <= LogLevel.INFO::
Memory Optimizer : OFF : Enable with env ORTMODULE_MEMORY_OPT_LEVEL=1 or ORTMODULE_MEMORY_OPT_CONFIG=<config file path> Configs Freq Max Saving(Bytes) Saving Symbolic(Bytes) - Plan 1 : OFF : Reshape+Where+:1:-1 1 134,217,728 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1**2 - Plan 2 : OFF : BiasSoftmax+:1:-1 1 134,086,656 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 3 : OFF : Cast+:1:-1 1 67,043,328 64.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 4 : OFF : BiasGelu+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 5 : OFF : FusedMatMul+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 6 : OFF : Add+:1:-1 1 5,237,760 5120.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 7 : OFF : Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+:1:-1 1 4,096 4.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - Plan 8 : OFF : Cast+:2:-1 1 2,048 2.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - As shown above,
Configis a string representative for a re-computable subgraph. All are disabled for recompute in this case. - Set environment variable
ORTMODULE_MEMORY_OPT_CONFIGto enable some of the subgraphs to do recompute.export ORTMODULE_MEMORY_OPT_CONFIG="mem_opt.json" # Content of mem_opt.json: [ "BiasGelu+:1:1", "Dropout+:1:-1" ] # Use comma as a separator for enabling more than one subgraphs in the json file. # Explanation: # > BiasGelu+ is the subgraph string representative; # > 1 in the middle indicates 'Recompute' is enabled (0, on the contrary indicates it's disabled) # > The last 1 means the initial 1 subgraph occurrences will be recomputed, all others are left as it is, filling `-1` will make all occurrences be recomputed.
- Then run the training again, and you will see logs like this:
Memory Optimizer : ON : Memory Optimization Level: [USER_SPECIFIED], Optimization Config: mem_opt.json Configs Freq Max Saving(Bytes) Saving Symbolic(Bytes) - Plan 1 : OFF : Reshape+Where+:1:-1 1 134,217,728 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1**2 - Plan 2 : OFF : BiasSoftmax+:1:-1 1 134,086,656 128.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 3 : OFF : Cast+:1:-1 1 67,043,328 64.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1) - Plan 4 : ON : BiasGelu+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 5 : OFF : FusedMatMul+:1:-1 1 20,951,040 20480.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 6 : OFF : Add+:1:-1 1 5,237,760 5120.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) - Plan 7 : OFF : Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+:1:-1 1 4,096 4.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - Plan 8 : OFF : Cast+:2:-1 1 2,048 2.0*inputs_input_ids_dim0*inputs_input_ids_dim1 - You may need iterate a few times on step 4 and 5 until you find a good config for this model to run a bigger batch size. Or you may fail to find if memory optimization does not apply to the model well.
The basic optimization unit is represented with a unique cluster id, for example BiasGelu+ is one cluster id.
Following cluster id is the optimization strategy: 0 - none, 1 - recompute, 2 - recompute with compromised memory saving.
Following optimization strategy is the request count to apply the given optimization. Using -1 to apply all. This would give user a bit more flexibility to avoid unnecessary memory saving.
If you check the above logs, there is a config Cast+:2:-1, 2 indicates it's a recomputation than can save part of the stashed activation size, not all. Recompute the subgraphs under it usually will save part of the activation (for example half of them), not all of them. Follow the same way to enable it.
Using following log level
ort_model = ORTModule(pt_model, DebugOptions(log_level=LogLevel.DEVINFO))
Besides the logs shown in LogLevel.INFO, you can also see different node patterns that can apply different optimization options.
The way we get the table:
- For a specific node, it might has different optimization options, we generates a hash (called
Node Cluster ID) for the node according to all available optimization options. - Map all nodes having same
Node Cluster IDin buckets, each bucket is displayed as one row.
MemoryInsight Summary - User config: not provided
===========================================================================================================================================
|Freq | Memory Optimization Opportunities (Clustered by node-level activation patterns) |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _|
|6 |For each row options are mutually exclusive, only one of them can be enabled. |
| | |
| |>>Option 1 : Recompute subgraph FusedMatMul+Add+Reshape+ |
| | Status : Disabled. |
| | Stashed Activations: |
| | - ReuseFreq : Output 0(6), |
| | - Output 0 : [((inputs_input_ids_dim0)*(inputs_input_ids_dim1)*(32)*(240))], byte/elem: 2, 100% saved |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _|
|5 |For each row options are mutually exclusive, only one of them can be enabled. |
| | |
| |>>Option 1 : Recompute subgraph FusedMatMul+ |
| | Status : Disabled. |
| | Stashed Activations: |
| | - Output 0 : [((inputs_input_ids_dim0)*(inputs_input_ids_dim1)*(10240))], byte/elem: 2, 100% saved |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _|
|5 |For each row options are mutually exclusive, only one of them can be enabled. |
| | |
| |>>Option 1 : Recompute subgraph Cast+ |
| | Status : Disabled. |
| | Stashed Activations: |
| | - Output 0 : [((inputs_input_ids_dim0)*(32)*(inputs_input_ids_dim1)*(inputs_input_ids_dim1))], byte/elem: 2, 100% saved |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _|
|1 |For each row options are mutually exclusive, only one of them can be enabled. |
| | |
| |>>Option 1 : Recompute subgraph Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+ |
| | Status : Disabled. |
| | Stashed Activations: |
| | - Output 0 : [((inputs_input_ids_dim0)*(1)*(1)*(inputs_input_ids_dim1))], byte/elem: 4, 100% saved |
| | |
| |>>Option 2 : RecomputeWithCompromise subgraph Cast+ |
| | Status : Disabled. |
| | Stashed Activations: |
| | - Output 0 : [((inputs_input_ids_dim0)*(1)*(1)*(inputs_input_ids_dim1))], byte/elem: 4, 50% saved |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _|
The feature is in the experimental stage, we will tune and refine it according to real use cases.