Buckets:

rtrm's picture
|
download
raw
11.1 kB
# LAMB
[LAMB (Layerwise adaptive large batch optimization)](https://hf.co/papers/1904.00962) is an adaptive optimizer designed for training with large batch sizes to accelerate training, combining ideas from `LARS` and `Adam` to automatically scale the learning rate for each layer:
- calculates a *trust ratio* between the weight and gradient norm in a layer and clips the ratio to prevent overly large or small updates
- updates weights with the first and second-moments
## LAMB[[api-class]][[bitsandbytes.optim.LAMB]]
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>class bitsandbytes.optim.LAMB</name><anchor>bitsandbytes.optim.LAMB</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L8</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "optim_bits", "val": " = 32"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters></docstring>
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>__init__</name><anchor>bitsandbytes.optim.LAMB.__init__</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L9</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "optim_bits", "val": " = 32"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters><paramsdesc>- **params** (`torch.tensor`) --
The input parameters to optimize.
- **lr** (`float`, defaults to 1e-3) --
The learning rate.
- **bias_correction** (`bool`, defaults to `True`) --
Whether to apply bias correction to the first and second-order moments.
- **betas** (`tuple(float, float)`, defaults to (0.9, 0.999)) --
The beta values are the decay rates of the first and second-order moment of the optimizer.
- **eps** (`float`, defaults to 1e-8) --
The epsilon value prevents division by zero in the optimizer.
- **weight_decay** (`float`, defaults to 1e-2) --
The weight decay value for the optimizer.
- **amsgrad** (`bool`, defaults to `False`) --
Whether to use the [AMSGrad](https://hf.co/papers/1904.09237) variant of Adam that uses the maximum of past squared gradients instead.
- **adam_w_mode** (`bool`, defaults to `True`) --
Whether to use the AdamW variant.
- **optim_bits** (`int`, defaults to 32) --
The number of bits of the optimizer state.
- **args** (`object`, defaults to `None`) --
An object with additional arguments.
- **min_8bit_size** (`int`, defaults to 4096) --
The minimum number of elements of the parameter tensors for 8-bit optimization.
- **percentile_clipping** (`int`, defaults to 100) --
Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
- **block_wise** (`bool`, defaults to `True`) --
Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
- **max_unorm** (`float`, defaults to 1.0) --
The maximum gradient norm.</paramsdesc><paramgroups>0</paramgroups></docstring>
Base LAMB optimizer.
</div></div>
## LAMB8bit[[bitsandbytes.optim.LAMB8bit]]
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>class bitsandbytes.optim.LAMB8bit</name><anchor>bitsandbytes.optim.LAMB8bit</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L75</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters></docstring>
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>__init__</name><anchor>bitsandbytes.optim.LAMB8bit.__init__</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L76</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters><paramsdesc>- **params** (`torch.tensor`) --
The input parameters to optimize.
- **lr** (`float`, defaults to 1e-3) --
The learning rate.
- **bias_correction** (`bool`, defaults to `True`) --
Whether to apply bias correction to the first and second-order moments.
- **betas** (`tuple(float, float)`, defaults to (0.9, 0.999)) --
The beta values are the decay rates of the first and second-order moment of the optimizer.
- **eps** (`float`, defaults to 1e-8) --
The epsilon value prevents division by zero in the optimizer.
- **weight_decay** (`float`, defaults to 1e-2) --
The weight decay value for the optimizer.
- **amsgrad** (`bool`, defaults to `False`) --
Whether to use the [AMSGrad](https://hf.co/papers/1904.09237) variant of Adam that uses the maximum of past squared gradients instead.
- **adam_w_mode** (`bool`, defaults to `True`) --
Whether to use the AdamW variant.
- **args** (`object`, defaults to `None`) --
An object with additional arguments.
- **min_8bit_size** (`int`, defaults to 4096) --
The minimum number of elements of the parameter tensors for 8-bit optimization.
- **percentile_clipping** (`int`, defaults to 100) --
Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
- **block_wise** (`bool`, defaults to `True`) --
Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
- **max_unorm** (`float`, defaults to 1.0) --
The maximum gradient norm.</paramsdesc><paramgroups>0</paramgroups></docstring>
8-bit LAMB optimizer.
</div></div>
## LAMB32bit[[bitsandbytes.optim.LAMB32bit]]
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>class bitsandbytes.optim.LAMB32bit</name><anchor>bitsandbytes.optim.LAMB32bit</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L139</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters></docstring>
<div class="docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8">
<docstring><name>__init__</name><anchor>bitsandbytes.optim.LAMB32bit.__init__</anchor><source>https://github.com/bitsandbytes-foundation/bitsandbytes/blob/v0.48.2/bitsandbytes/optim/lamb.py#L140</source><parameters>[{"name": "params", "val": ""}, {"name": "lr", "val": " = 0.001"}, {"name": "bias_correction", "val": " = True"}, {"name": "betas", "val": " = (0.9, 0.999)"}, {"name": "eps", "val": " = 1e-08"}, {"name": "weight_decay", "val": " = 0"}, {"name": "amsgrad", "val": " = False"}, {"name": "adam_w_mode", "val": " = True"}, {"name": "args", "val": " = None"}, {"name": "min_8bit_size", "val": " = 4096"}, {"name": "percentile_clipping", "val": " = 100"}, {"name": "block_wise", "val": " = False"}, {"name": "max_unorm", "val": " = 1.0"}]</parameters><paramsdesc>- **params** (`torch.tensor`) --
The input parameters to optimize.
- **lr** (`float`, defaults to 1e-3) --
The learning rate.
- **bias_correction** (`bool`, defaults to `True`) --
Whether to apply bias correction to the first and second-order moments.
- **betas** (`tuple(float, float)`, defaults to (0.9, 0.999)) --
The beta values are the decay rates of the first and second-order moment of the optimizer.
- **eps** (`float`, defaults to 1e-8) --
The epsilon value prevents division by zero in the optimizer.
- **weight_decay** (`float`, defaults to 1e-2) --
The weight decay value for the optimizer.
- **amsgrad** (`bool`, defaults to `False`) --
Whether to use the [AMSGrad](https://hf.co/papers/1904.09237) variant of Adam that uses the maximum of past squared gradients instead.
- **adam_w_mode** (`bool`, defaults to `True`) --
Whether to use the AdamW variant.
- **args** (`object`, defaults to `None`) --
An object with additional arguments.
- **min_8bit_size** (`int`, defaults to 4096) --
The minimum number of elements of the parameter tensors for 8-bit optimization.
- **percentile_clipping** (`int`, defaults to 100) --
Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
- **block_wise** (`bool`, defaults to `True`) --
Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
- **max_unorm** (`float`, defaults to 1.0) --
The maximum gradient norm.</paramsdesc><paramgroups>0</paramgroups></docstring>
32-bit LAMB optimizer.
</div></div>
<EditOnGithub source="https://github.com/bitsandbytes-foundation/bitsandbytes/blob/main/docs/source/reference/optim/lamb.mdx" />

Xet Storage Details

Size:
11.1 kB
·
Xet hash:
383a39d9b363d314dcf06e85c380cbfa6834f8129c09f412cfc0db8be66cfc19

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.