Merge pull request #3 from warner-benjamin/opt_accum

Add Optimizer Accumulation

README.md

@@ -1,15 +1,25 @@
 # optimī
 
-### Fast, Modern, and Low Precision PyTorch Optimizers
+### Fast, Modern, Memory Efficient, and Low Precision PyTorch Optimizers
 
-optimi enables accurate low precision training via Kahan summation, supports fully decoupled weight decay, and features fast implementations of modern optimizers.
+optimi enables accurate low precision training via Kahan summation, integrates gradient release and optimizer accumulation for additional memory efficiency, supports fully decoupled weight decay, and features fast implementations of modern optimizers.
 
 ## Low Precision Training with Kahan Summation
 
-optimi optimizers can match the performance of mixed precision when [training in BFloat16 by using Kahan summation](https://optimi.benjaminwarner.dev/kahan_summation).
+optimi optimizers can nearly reach or match the performance of mixed precision when [training in BFloat16 by using Kahan summation](https://optimi.benjaminwarner.dev/kahan_summation).
 
 Training in BFloat16 with Kahan summation can reduce non-activation training memory usage by [37.5 to 45.5 percent](https://optimi.benjaminwarner.dev/kahan_summation/#memory-savings) when using an Adam optimizer. BFloat16 training increases single GPU [training speed by ~10 percent](https://optimi.benjaminwarner.dev/kahan_summation/#training-speedup) at the same batch size.
 
+## Gradient Release: Fused Backward and Optimizer Step
+
+optimi optimizers can perform the [optimization step layer-by-layer during the backward pass](https://optimi.benjaminwarner.dev/gradient_release), immediately freeing gradient memory.
+
+Unlike the current PyTorch implementation, optimi’s gradient release optimizers are a drop-in replacement for standard optimizers and seamlessly work with exisiting hyperparmeter schedulers.
+
+## Optimizer Accumulation: Gradient Release and Accumulation
+
+optimi optimizers can approximate gradient accumulation with gradient release by [accumulating gradients into the optimizer states](https://optimi.benjaminwarner.dev/optimizer_accumulation).
+
 ## Fully Decoupled Weight Decay
 
 In addition to supporting PyTorch-style decoupled weight decay, optimi optimizers also support [fully decoupled weight decay](https://optimi.benjaminwarner.dev/fully_decoupled_weight_decay).
@@ -44,7 +54,7 @@ from optimi import AdamW
 # create or cast model in low precision (bfloat16)
 model = nn.Linear(20, 1, dtype=torch.bfloat16)
 
-# instantiate AdamW with parameters and fully decoupled weight decay
+# initialize any optimi optimizer with parameters & fully decoupled weight decay
 # Kahan summation is automatically enabled since model & inputs are bfloat16
 opt = AdamW(model.parameters(), lr=1e-3, weight_decay=1e-5, decouple_lr=True)
 
@@ -63,10 +73,64 @@ To use with PyTorch-style weight decay with float32 or mixed precision:
 # create model
 model = nn.Linear(20, 1)
 
-# instantiate AdamW with parameters
+# initialize any optimi optimizer with parameters
 opt = AdamW(model.parameters(), lr=1e-3, weight_decay=1e-2)
 ```
 
+To use with gradient release:
+
+```python
+# initialize any optimi optimizer with `gradient_release=True`
+# and call `prepare_for_gradient_release` on model and optimizer
+opt = AdamW(model.parameters(), lr=1e-3, gradient_release=True)
+prepare_for_gradient_release(model, opt)
+
+# calling backward on the model will peform the optimzier step
+loss = model(torch.randn(20, dtype=torch.bfloat16))
+loss.backward()
+
+# optimizer step and zero_grad are no longer needed, and will
+# harmlessly no-op if called by an existing training framework
+# opt.step()
+# opt.zero_grad()
+
+# optionally remove gradient release hooks when done training
+remove_gradient_release(model)
+```
+
+To use with optimizer accumulation:
+
+```python
+# initialize any optimi optimizer with `gradient_release=True`
+# and call `prepare_for_gradient_release` on model and optimizer
+opt = AdamW(model.parameters(), lr=1e-3, gradient_release=True)
+prepare_for_gradient_release(model, opt)
+
+# update model parameters every four steps after accumulating
+# gradients directly into the optimizer states
+accumulation_steps = 4
+
+# use existing PyTorch dataloader
+for idx, batch in enumerate(dataloader):
+    # `optimizer_accumulation=True` accumulates gradients into
+    # optimizer states. set `optimizer_accumulation=False` to
+    # update parameters by performing a full gradient release step
+    opt.optimizer_accumulation = (idx+1) % accumulation_steps != 0
+
+    # calling backward on the model will peform the optimizer step
+    # either accumulating gradients or updating model parameters
+    loss = model(batch)
+    loss.backward()
+
+    # optimizer step and zero_grad are no longer needed, and will
+    # harmlessly no-op if called by an existing training framework
+    # opt.step()
+    # opt.zero_grad()
+
+# optionally remove gradient release hooks when done training
+remove_gradient_release(model)
+```
+
 ## Differences from PyTorch
 
 optimi optimizers do not support compilation, differentiation, complex numbers, or have capturable versions.

docs/css/extra.css

@@ -46,8 +46,8 @@
     --md-typeset-table-color: rgba(24, 24, 24, 0.05);
 
       /* Code highlighting color shades */
-    /* --md-code-hl-color:                  #9a3fe4;
-    --md-code-hl-color--light:           #9a3fe4; */
+    --md-code-hl-color:                  #9a3fe4;
+    --md-code-hl-color--light:           #9a3fe43c;
     --md-code-hl-number-color:           #db5f00;
     --md-code-hl-special-color:          #d32300;
     --md-code-hl-function-color:         #cc9901;
@@ -161,7 +161,7 @@
 
 
 /* Links */
-.md-content a:not(.headerlink):not(.footnote-ref):not(.footnote-backref) {
+.md-content a:not(.headerlink):not(.footnote-ref):not(.footnote-backref):not(:has(> code)) {
     box-shadow: inset 0 -0.115rem 0 var(--light-purple);
     text-decoration: none;
     transition: all .15s cubic-bezier(.33,.66,.66,1);
@@ -171,6 +171,18 @@
             color: var(--black) }
 }
 
+.md-content a code {
+    box-shadow: inset 0 -0.115rem 0 var(--light-purple);
+    text-decoration: none;
+    transition: all .15s cubic-bezier(.33,.66,.66,1);
+    z-index: 10;
+    border-bottom-left-radius: 0;
+    border-bottom-right-radius: 0;
+
+  &:hover { box-shadow: inset 0 -2rem 0 var(--dark-purple);
+            color: var(--black) }
+}
+
 /* Katex */
 .katex-display {
   margin-top: 0 !important;

docs/foreach.md

@@ -8,7 +8,7 @@ Like PyTorch, optimi supports foreach implementations of all optimizers. Foreach
 
 Foreach implementations can increase optimizer peak memory usage. optimi attempts to reduce this extra overhead by reusing the gradient buffer for temporary variables. If the gradients are required between the optimization step and [gradient reset step](https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html#torch.optim.Optimizer.zero_grad), set `foreach=False` to use the for-loop implementation.
 
-??? warning "Important: Foreach Requires PyTorch 2.1+"
+??? note "Note: Foreach Requires PyTorch 2.1+"
 
     optimi’s foreach implementations require PyTorch 2.1 or newer.
 

docs/gradient_release.md

@@ -1,27 +1,40 @@
 ---
-title: "Gradient Release: Fused Backward and Optimizer Step"
+title: "Gradient Release"
+description: "Fused Backward Pass and Optimizer Step"
 ---
 
-# Gradient Release: Fused Backward and Optimizer Step
+# Gradient Release
 
-Gradient release reduces training memory by limiting gradients to one layer at any given time. Unlike [PyTorch’s implementation](https://pytorch.org/tutorials/intermediate/optimizer_step_in_backward_tutorial.html), optimi’s gradient release is fully compatible with existing learning rate and optimizer schedulers and training frameworks.
+**Fused Backward Pass and Optimizer Step**
+
+Gradient release reduces training memory by limiting gradients to one layer at any given time. Unlike [PyTorch’s implementation](https://pytorch.org/tutorials/intermediate/optimizer_step_in_backward_tutorial.html), optimi’s gradient release is fully compatible with both existing learning rate and optimizer schedulers and existing training frameworks.
 
 During the backward pass, each model layer calculates its gradients, performs the optimizer step, and clears the gradients before proceeding to the backward pass for the next layer. This fused backward and optimizer step can reduce non-activation memory usage by ~25 percent for an Adam optimizer.
 
-Gradient release can be combined with other techniques such as [Kahan summation](kahan_summation.md) or activation checkpointing for further memory savings.
+Gradient release can also be combined with other techniques such as [Kahan summation](kahan_summation.md) or [activation checkpointing](https://pytorch.org/docs/stable/checkpoint.html) for further memory savings.
 
-??? warning "Important: Gradient Release Requires PyTorch 2.1+"
+??? note "Note: Gradient Release Requires PyTorch 2.1+"
 
     Gradient release requires PyTorch 2.1 or newer.
 
 Gradient release was proposed by Pudipeddi et al in [*Training Large Neural Networks with Constant Memory using a New Execution Algorithm*](https://arxiv.org/abs/2002.05645) and was enabled by PyTorch’s [`register_post_accumulate_grad_hook`](https://pytorch.org/docs/stable/generated/torch.Tensor.register_post_accumulate_grad_hook.html).
 
 ## Limitations and Workarounds
 
-Since gradient release immediately frees the gradient during the backward pass, features which rely on persistent gradients like gradient clipping or gradient accumulation won’t work.
+Since gradient release immediately frees the gradient during the backward pass, features which rely on persistent gradients like AMP's `GradScaler`, gradient clipping, or gradient accumulation won’t work.
+
+!!! warning "Important: Gradient Release is Incompatible with FP16 Mixed Precision"
+
+    Gradient release is incompatible with Float16 Automatic Mixed Precision since PyTorch's `GradScaler` requires access to the entire model's gradients for the optimizer step.
+
+    Use BFloat16 Automatic Mixed Precision instead.
 
 The recommended workaround for gradient clipping is to use [StableAdamW](optimizers/stableadamw.md) instead of Adam or AdamW, as StableAdamW removes the need for gradient clipping by porting Adafactor’s update clipping into AdamW.
 
+??? tip "Tip: Use Optimizer Accumulation to Approximate Gradient Accumulation"
+
+    optimi's [optimizer accumulation](optimizer_accumulation.md) approximates [gradient accumlation](https://pytorch.org/docs/stable/notes/amp_examples.html#gradient-accumulation) by defering parameter updates while accumulating gradients directly into the optimizer states.
+
 One potential workaround for gradient accumulation is to increase the optimizer’s momentum or $\beta_1$ to approximate accumulating gradients across multiple batches.
 
 ## Example
@@ -45,10 +58,10 @@ prepare_for_gradient_release(model, opt)
 loss = model(torch.randn(20, dtype=torch.bfloat16))
 loss.backward()
 
-# optimizer step and sero_grad is no longer needed, and
-# will no-op if called by an existing training framework
-opt.step()
-opt.zero_grad()
+# optimizer step and zero_grad are no longer needed, and will
+# harmlessly no-op if called by an existing training framework
+# opt.step()
+# opt.zero_grad()
 
 # optionally remove gradient release hooks when done training
 remove_gradient_release(model)

docs/index.md

@@ -1,17 +1,17 @@
 ---
 title: "optimi"
-description: "Fast, Modern, and Low Precision PyTorch Optimizers"
+description: "Fast, Modern, Memory Efficient, and Low Precision PyTorch Optimizers"
 ---
 
 # optimī
 
-**Fast, Modern, and Low Precision PyTorch Optimizers**
+**Fast, Modern, Memory Efficient, and Low Precision PyTorch Optimizers**
 
-optimi enables accurate low precision training via Kahan summation, supports fully decoupled weight decay, and features fast implementations of modern optimizers.
+optimi enables accurate low precision training via Kahan summation, integrates gradient release and optimizer accumulation for additional memory efficiency, supports fully decoupled weight decay, and features fast implementations of modern optimizers.
 
 ## Low Precision Training with Kahan Summation
 
-optimi optimizers can match the performance of mixed precision when [training in BFloat16 by using Kahan summation](kahan_summation.md).
+optimi optimizers can nearly reach or match the performance of mixed precision when [training in BFloat16 by using Kahan summation](kahan_summation.md).
 
 Training in BFloat16 with Kahan summation can reduce non-activation training memory usage by [37.5 to 45.5 percent](kahan_summation.md/#memory-savings) when using an Adam optimizer. BFloat16 training increases single GPU [training speed by ~10 percent](kahan_summation.md/#training-speedup) at the same batch size.
 
@@ -21,6 +21,10 @@ optimi optimizers can perform the [optimization step layer-by-layer during the b
 
 Unlike the current PyTorch implementation, optimi’s gradient release optimizers are a drop-in replacement for standard optimizers and seamlessly work with exisiting hyperparmeter schedulers.
 
+## Optimizer Accumulation: Gradient Release and Accumulation
+
+optimi optimizers can approximate gradient accumulation with gradient release by [accumulating gradients into the optimizer states](optimizer_accumulation.md).
+
 ## Fully Decoupled Weight Decay
 
 In addition to supporting PyTorch-style decoupled weight decay, optimi optimizers also support [fully decoupled weight decay](fully_decoupled_weight_decay.md).
@@ -51,7 +55,7 @@ from optimi import AdamW
 # create or cast model in low precision (bfloat16)
 model = nn.Linear(20, 1, dtype=torch.bfloat16)
 
-# initialize AdamW with parameters and fully decoupled weight decay
+# initialize any optimi optimizer with parameters & fully decoupled weight decay
 # Kahan summation is automatically enabled since model & inputs are bfloat16
 opt = AdamW(model.parameters(), lr=1e-3, weight_decay=1e-5, decouple_lr=True)
 
@@ -70,29 +74,59 @@ To use with PyTorch-style weight decay with float32 or mixed precision:
 # create model
 model = nn.Linear(20, 1)
 
-# initialize AdamW with parameters
+# initialize any optimi optimizer with parameters
 opt = AdamW(model.parameters(), lr=1e-3, weight_decay=1e-2)
 ```
 
 To use with gradient release:
 
 ```python
-# create model
-model = nn.Linear(20, 1)
-
-# initialize AdamW with `gradient_release=True` and call
-# `prepare_for_gradient_release` on model and optimizer
+# initialize any optimi optimizer with `gradient_release=True`
+# and call `prepare_for_gradient_release` on model and optimizer
 opt = AdamW(model.parameters(), lr=1e-3, gradient_release=True)
 prepare_for_gradient_release(model, opt)
 
 # calling backward on the model will peform the optimzier step
 loss = model(torch.randn(20, dtype=torch.bfloat16))
 loss.backward()
 
-# optimizer step and sero_grad is no longer needed, and
-# will no-op if called by an existing training framework
-opt.step()
-opt.zero_grad()
+# optimizer step and zero_grad are no longer needed, and will
+# harmlessly no-op if called by an existing training framework
+# opt.step()
+# opt.zero_grad()
+
+# optionally remove gradient release hooks when done training
+remove_gradient_release(model)
+```
+
+To use with optimizer accumulation:
+
+```python
+# initialize any optimi optimizer with `gradient_release=True`
+# and call `prepare_for_gradient_release` on model and optimizer
+opt = AdamW(model.parameters(), lr=1e-3, gradient_release=True)
+prepare_for_gradient_release(model, opt)
+
+# update model parameters every four steps after accumulating
+# gradients directly into the optimizer states
+accumulation_steps = 4
+
+# use existing PyTorch dataloader
+for idx, batch in enumerate(dataloader):
+    # `optimizer_accumulation=True` accumulates gradients into
+    # optimizer states. set `optimizer_accumulation=False` to
+    # update parameters by performing a full gradient release step
+    opt.optimizer_accumulation = (idx+1) % accumulation_steps != 0
+
+    # calling backward on the model will peform the optimizer step
+    # either accumulating gradients or updating model parameters
+    loss = model(batch)
+    loss.backward()
+
+    # optimizer step and zero_grad are no longer needed, and will
+    # harmlessly no-op if called by an existing training framework
+    # opt.step()
+    # opt.zero_grad()
 
 # optionally remove gradient release hooks when done training
 remove_gradient_release(model)

docs/kahan_summation.md

@@ -4,7 +4,7 @@ title: Low Precision Training with Kahan Summation
 
 # Low Precision Training with Kahan Summation
 
-While training models in low precision (Float16 or BFloat16) usually does not match training in full precision (Float32) or [mixed precision](https://pytorch.org/blog/what-every-user-should-know-about-mixed-precision-training-in-pytorch), optimi optimizers match the performance of mixed precision when training in BFloat16 by using Kahan summation[^1].
+While training models in low precision (Float16 or BFloat16) usually differs from training in full precision (Float32) or [mixed precision](https://pytorch.org/blog/what-every-user-should-know-about-mixed-precision-training-in-pytorch), optimi optimizers nearly reach or match the performance of mixed precision when training in BFloat16 by using Kahan summation[^1].
 
 Training in low precision [reduces memory usage](#memory-savings) and increases [training speed](#training-speedup) relative to mixed precision training.
 
@@ -108,6 +108,6 @@ $$
 
 This shows the optimi implementation of Kahan summation optimizers, which is equivalent to the *Revisiting BFloat16 Training* formulation.
 
-[^1]: Current testing on small models shows no degradation in training performance.
+[^1]: Current testing on small models shows little to no degradation in model performance.
 
 [^2]: Also known as Kahan–Babuška summation or compensated summation.