Modular transformers: modularity and inheritance for new model additions (#33248)
* update exampel * update * push the converted diff files for testing and ci * correct one example * fix class attributes and docstring * nits * oups * fixed config! * update * nitd * class attributes are not matched against the other, this is missing * fixed overwriting self.xxx now onto the attributes I think * partial fix, now order with docstring * fix docstring order? * more fixes * update * fix missing docstrings! * examples don't all work yet * fixup * nit * updated * hick * update * delete * update * update * update * fix * all default * no local import * fix more diff * some fix related to "safe imports" * push fixed * add helper! * style * add a check * all by default * add the * update * FINALLY! * nit * fix config dependencies * man that is it * fix fix * update diffs * fix the last issue * re-default to all * alll the fixes * nice * fix properties vs setter * fixup * updates * update dependencies * make sure to install what needs to be installed * fixup * quick fix for now * fix! * fixup * update * update * updates * whitespaces * nit * fix * simplify everything, and make it file agnostic (should work for image processors) * style * finish fixing all import issues * fixup * empty modeling should not be written! * Add logic to find who depends on what * update * cleanup * update * update gemma to support positions * some small nits * this is the correct docstring for gemma2 * fix merging of docstrings * update * fixup * update * take doc into account * styling * update * fix hidden activation * more fixes * final fixes! * fixup * fixup instruct blip video * update * fix bugs * align gemma2 with the rest as well * updats * revert * update * more reversiom * grind * more * arf * update * order will matter * finish del stuff * update * rename to modular * fixup * nits * update makefile * fixup * update order of the checks! * fix * fix docstring that has a call inside * fiix conversion check * style * add some initial documentation * update * update doc * some fixup * updates * yups * Mostly todo gimme a minut * update * fixup * revert some stuff * Review docs for the modular transformers (#33472) Docs * good update * fixup * mmm current updates lead to this code * okay, this fixes it * cool * fixes * update * nit * updates * nits * fix doc * update * revert bad changes * update * updates * proper update * update * update? * up * update * cool * nits * nits * bon bon * fix * ? * minimise changes * update * update * update * updates? * fixed gemma2 * kind of a hack * nits * update * remove `diffs` in favor of `modular` * fix make fix copies --------- Co-authored-by: Lysandre Debut <hi@lysand.re>
This commit is contained in:
Arthur
@@ -1,20 +0,0 @@
|
||||
# Using the `diff_converter` linter
|
||||
|
||||
`pip install libcst` is a must!
|
||||
|
||||
# `sh examples/diff-conversion/convert_examples.sh` to get the converted outputs
|
||||
|
||||
The diff converter is a new `linter` specific to `transformers`. It allows us to unpack inheritance in python to convert a modular `diff` file like `diff_gemma.py` into a `single model single file`.
|
||||
|
||||
Examples of possible usage are available in the `examples/diff-conversion`, or `diff_gemma` for a full model usage.
|
||||
|
||||
`python utils/diff_model_converter.py --files_to_parse "/Users/arthurzucker/Work/transformers/examples/diff-conversion/diff_my_new_model2.py"`
|
||||
|
||||
## How it works
|
||||
We use the `libcst` parser to produce an AST representation of the `diff_xxx.py` file. For any imports that are made from `transformers.models.modeling_xxxx` we parse the source code of that module, and build a class dependency mapping, which allows us to unpack the difference dependencies.
|
||||
|
||||
The code from the `diff` file and the class dependency mapping are "merged" to produce the single model single file.
|
||||
We use ruff to automatically remove the potential duplicate imports.
|
||||
|
||||
## Why we use libcst instead of the native AST?
|
||||
AST is super powerful, but it does not keep the `docstring`, `comment` or code formatting. Thus we decided to go with `libcst`
|
||||
20
examples/modular-transformers/README.md
Normal file
20
examples/modular-transformers/README.md
Normal file
@@ -0,0 +1,20 @@
|
||||
# Using the `modular_converter` linter
|
||||
|
||||
`pip install libcst` is a must!
|
||||
|
||||
# `sh examples/modular-transformers/convert_examples.sh` to get the converted outputs
|
||||
|
||||
The modular converter is a new `linter` specific to `transformers`. It allows us to unpack inheritance in python to convert a modular file like `modular_gemma.py` into a `single model single file`.
|
||||
|
||||
Examples of possible usage are available in the `examples/modular-transformers`, or `modular_gemma` for a full model usage.
|
||||
|
||||
`python utils/modular_model_converter.py --files_to_parse "/Users/arthurzucker/Work/transformers/examples/modular-transformers/modular_my_new_model2.py"`
|
||||
|
||||
## How it works
|
||||
We use the `libcst` parser to produce an AST representation of the `modular_xxx.py` file. For any imports that are made from `transformers.models.modeling_xxxx` we parse the source code of that module, and build a class dependency mapping, which allows us to unpack the modularerence dependencies.
|
||||
|
||||
The code from the `modular` file and the class dependency mapping are "merged" to produce the single model single file.
|
||||
We use ruff to automatically remove the potential duplicate imports.
|
||||
|
||||
## Why we use libcst instead of the native AST?
|
||||
AST is super powerful, but it does not keep the `docstring`, `comment` or code formatting. Thus we decided to go with `libcst`
|
||||
196
examples/modular-transformers/configuration_my_new_model.py
Normal file
196
examples/modular-transformers/configuration_my_new_model.py
Normal file
@@ -0,0 +1,196 @@
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
# This file was automatically generated from <path_to_diff_file.py>.
|
||||
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
||||
# the file from the diff. If any change should be done, please apply the change to the
|
||||
# diff.py file directly. One of our CI enforces this
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...modeling_rope_utils import rope_config_validation
|
||||
|
||||
|
||||
class MyNewModelConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`MyNewModelModel`]. It is used to instantiate an MyNewModel
|
||||
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
||||
defaults will yield a similar configuration to that of the MyNewModel-7B.
|
||||
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 32000):
|
||||
Vocabulary size of the MyNewModel model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`MyNewModelModel`]
|
||||
hidden_size (`int`, *optional*, defaults to 4096):
|
||||
Dimension of the hidden representations.
|
||||
intermediate_size (`int`, *optional*, defaults to 11008):
|
||||
Dimension of the MLP representations.
|
||||
num_hidden_layers (`int`, *optional*, defaults to 32):
|
||||
Number of hidden layers in the Transformer decoder.
|
||||
num_attention_heads (`int`, *optional*, defaults to 32):
|
||||
Number of attention heads for each attention layer in the Transformer decoder.
|
||||
num_key_value_heads (`int`, *optional*):
|
||||
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
||||
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
||||
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
||||
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
||||
by meanpooling all the original heads within that group. For more details checkout [this
|
||||
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
|
||||
`num_attention_heads`.
|
||||
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
||||
The non-linear activation function (function or string) in the decoder.
|
||||
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
||||
The maximum sequence length that this model might ever be used with. MyNewModel 1 supports up to 2048 tokens,
|
||||
MyNewModel 2 up to 4096, CodeMyNewModel up to 16384.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
||||
The epsilon used by the rms normalization layers.
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
||||
relevant if `config.is_decoder=True`.
|
||||
pad_token_id (`int`, *optional*):
|
||||
Padding token id.
|
||||
bos_token_id (`int`, *optional*, defaults to 1):
|
||||
Beginning of stream token id.
|
||||
eos_token_id (`int`, *optional*, defaults to 2):
|
||||
End of stream token id.
|
||||
pretraining_tp (`int`, *optional*, defaults to 1):
|
||||
Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
|
||||
document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
|
||||
understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
|
||||
results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
|
||||
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
||||
Whether to tie weight embeddings
|
||||
rope_theta (`float`, *optional*, defaults to 10000.0):
|
||||
The base period of the RoPE embeddings.
|
||||
rope_scaling (`Dict`, *optional*):
|
||||
Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
|
||||
and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
|
||||
accordingly.
|
||||
Expected contents:
|
||||
`rope_type` (`str`):
|
||||
The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
|
||||
'my_new_model3'], with 'default' being the original RoPE implementation.
|
||||
`factor` (`float`, *optional*):
|
||||
Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
|
||||
most scaling types, a `factor` of x will enable the model to handle sequences of length x *
|
||||
original maximum pre-trained length.
|
||||
`original_max_position_embeddings` (`int`, *optional*):
|
||||
Used with 'dynamic', 'longrope' and 'my_new_model3'. The original max position embeddings used during
|
||||
pretraining.
|
||||
`attention_factor` (`float`, *optional*):
|
||||
Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
|
||||
computation. If unspecified, it defaults to value recommended by the implementation, using the
|
||||
`factor` field to infer the suggested value.
|
||||
`beta_fast` (`float`, *optional*):
|
||||
Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
|
||||
ramp function. If unspecified, it defaults to 32.
|
||||
`beta_slow` (`float`, *optional*):
|
||||
Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
|
||||
ramp function. If unspecified, it defaults to 1.
|
||||
`short_factor` (`List[float]`, *optional*):
|
||||
Only used with 'longrope'. The scaling factor to be applied to short contexts (<
|
||||
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
||||
size divided by the number of attention heads divided by 2
|
||||
`long_factor` (`List[float]`, *optional*):
|
||||
Only used with 'longrope'. The scaling factor to be applied to long contexts (<
|
||||
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
||||
size divided by the number of attention heads divided by 2
|
||||
`low_freq_factor` (`float`, *optional*):
|
||||
Only used with 'my_new_model3'. Scaling factor applied to low frequency components of the RoPE
|
||||
`high_freq_factor` (`float`, *optional*):
|
||||
Only used with 'my_new_model3'. Scaling factor applied to high frequency components of the RoPE
|
||||
attention_bias (`bool`, *optional*, defaults to `False`):
|
||||
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
||||
attention_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for the attention probabilities.
|
||||
mlp_bias (`bool`, *optional*, defaults to `False`):
|
||||
Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
|
||||
head_dim (`int`, *optional*):
|
||||
The attention head dimension. If None, it will default to hidden_size // num_heads
|
||||
new_param (`int`, *optional*, defaults to `False`):
|
||||
A fun new parameter
|
||||
|
||||
```python
|
||||
>>> from transformers import MyNewModelModel, MyNewModelConfig
|
||||
|
||||
>>> # Initializing a MyNewModel my_new_model-7b style configuration
|
||||
>>> configuration = MyNewModelConfig()
|
||||
|
||||
>>> # Initializing a model from the my_new_model-7b style configuration
|
||||
>>> model = MyNewModelModel(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "my_new_model"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=32000,
|
||||
hidden_size=4096,
|
||||
intermediate_size=11008,
|
||||
num_hidden_layers=32,
|
||||
num_attention_heads=32,
|
||||
num_key_value_heads=None,
|
||||
hidden_act="silu",
|
||||
max_position_embeddings=2048,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-6,
|
||||
use_cache=True,
|
||||
pad_token_id=None,
|
||||
bos_token_id=1,
|
||||
eos_token_id=2,
|
||||
pretraining_tp=1,
|
||||
tie_word_embeddings=False,
|
||||
rope_theta=10000.0,
|
||||
rope_scaling=None,
|
||||
attention_bias=False,
|
||||
attention_dropout=0.0,
|
||||
mlp_bias=True,
|
||||
head_dim=None,
|
||||
new_param=0,
|
||||
**kwargs,
|
||||
):
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.intermediate_size = intermediate_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
|
||||
# for backward compatibility
|
||||
if num_key_value_heads is None:
|
||||
num_key_value_heads = num_attention_heads
|
||||
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.pretraining_tp = pretraining_tp
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.rope_scaling = rope_scaling
|
||||
self.attention_bias = attention_bias
|
||||
self.attention_dropout = attention_dropout
|
||||
self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
|
||||
# Validate the correctness of rotary position embeddings parameters
|
||||
# BC: if there is a 'type' field, copy it it to 'rope_type'.
|
||||
if self.rope_scaling is not None and "type" in self.rope_scaling:
|
||||
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
|
||||
rope_config_validation(self)
|
||||
|
||||
super().__init__(
|
||||
pad_token_id=pad_token_id,
|
||||
bos_token_id=bos_token_id,
|
||||
eos_token_id=eos_token_id,
|
||||
tie_word_embeddings=tie_word_embeddings,
|
||||
**kwargs,
|
||||
)
|
||||
self.mlp_bias = mlp_bias
|
||||
self.new_param = new_param
|
||||
97
examples/modular-transformers/configuration_my_new_model2.py
Normal file
97
examples/modular-transformers/configuration_my_new_model2.py
Normal file
@@ -0,0 +1,97 @@
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
# This file was automatically generated from <path_to_diff_file.py>.
|
||||
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
||||
# the file from the diff. If any change should be done, please apply the change to the
|
||||
# diff.py file directly. One of our CI enforces this
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...modeling_rope_utils import rope_config_validation
|
||||
|
||||
|
||||
class MyNewModel2Config(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`GemmaModel`]. It is used to instantiate an Gemma
|
||||
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
||||
defaults will yield a similar configuration to that of the Gemma-7B.
|
||||
e.g. [google/gemma-7b](https://huggingface.co/google/gemma-7b)
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 256000):
|
||||
Vocabulary size of the Gemma model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`GemmaModel`]
|
||||
```python
|
||||
>>> from transformers import GemmaModel, GemmaConfig
|
||||
>>> # Initializing a Gemma gemma-7b style configuration
|
||||
>>> configuration = GemmaConfig()
|
||||
>>> # Initializing a model from the gemma-7b style configuration
|
||||
>>> model = GemmaModel(configuration)
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "my_new_model2"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=32000,
|
||||
hidden_size=4096,
|
||||
intermediate_size=11008,
|
||||
num_hidden_layers=32,
|
||||
num_attention_heads=32,
|
||||
num_key_value_heads=None,
|
||||
hidden_act="silu",
|
||||
max_position_embeddings=2048,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-6,
|
||||
use_cache=True,
|
||||
pad_token_id=None,
|
||||
bos_token_id=1,
|
||||
eos_token_id=2,
|
||||
pretraining_tp=1,
|
||||
tie_word_embeddings=False,
|
||||
rope_theta=10000.0,
|
||||
rope_scaling=None,
|
||||
attention_bias=False,
|
||||
attention_dropout=0.0,
|
||||
mlp_bias=False,
|
||||
head_dim=None,
|
||||
**kwargs,
|
||||
):
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.intermediate_size = intermediate_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
|
||||
# for backward compatibility
|
||||
if num_key_value_heads is None:
|
||||
num_key_value_heads = num_attention_heads
|
||||
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.pretraining_tp = pretraining_tp
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.rope_scaling = rope_scaling
|
||||
self.attention_bias = attention_bias
|
||||
self.attention_dropout = attention_dropout
|
||||
self.mlp_bias = mlp_bias
|
||||
self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
|
||||
# Validate the correctness of rotary position embeddings parameters
|
||||
# BC: if there is a 'type' field, move it to 'rope_type'.
|
||||
if self.rope_scaling is not None and "type" in self.rope_scaling:
|
||||
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
|
||||
rope_config_validation(self)
|
||||
|
||||
super().__init__(
|
||||
pad_token_id=pad_token_id,
|
||||
bos_token_id=bos_token_id,
|
||||
eos_token_id=eos_token_id,
|
||||
tie_word_embeddings=tie_word_embeddings,
|
||||
**kwargs,
|
||||
)
|
||||
134
examples/modular-transformers/configuration_new_model.py
Normal file
134
examples/modular-transformers/configuration_new_model.py
Normal file
@@ -0,0 +1,134 @@
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
# This file was automatically generated from <path_to_diff_file.py>.
|
||||
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
||||
# the file from the diff. If any change should be done, please apply the change to the
|
||||
# diff.py file directly. One of our CI enforces this
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
# Example where we only want to overwrite the defaults of an init
|
||||
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
|
||||
class NewModelConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`NewModelModel`]. It is used to instantiate an NewModel
|
||||
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
||||
defaults will yield a similar configuration to that of the NewModel-7B.
|
||||
e.g. [google/new_model-7b](https://huggingface.co/google/new_model-7b)
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 256000):
|
||||
Vocabulary size of the NewModel model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`NewModelModel`]
|
||||
hidden_size (`int`, *optional*, defaults to 3072):
|
||||
Dimension of the hidden representations.
|
||||
intermediate_size (`int`, *optional*, defaults to 24576):
|
||||
Dimension of the MLP representations.
|
||||
num_hidden_layers (`int`, *optional*, defaults to 28):
|
||||
Number of hidden layers in the Transformer decoder.
|
||||
num_attention_heads (`int`, *optional*, defaults to 16):
|
||||
Number of attention heads for each attention layer in the Transformer decoder.
|
||||
num_key_value_heads (`int`, *optional*, defaults to 16):
|
||||
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
||||
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
||||
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
||||
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
||||
by meanpooling all the original heads within that group. For more details checkout [this
|
||||
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
|
||||
`num_attention_heads`.
|
||||
head_dim (`int`, *optional*, defaults to 256):
|
||||
The attention head dimension.
|
||||
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
|
||||
The legacy activation function. It is overwritten by the `hidden_activation`.
|
||||
hidden_activation (`str` or `function`, *optional*):
|
||||
The non-linear activation function (function or string) in the decoder. Will default to `"gelu_pytorch_tanh"`
|
||||
if not specified. `"gelu_pytorch_tanh"` uses an approximation of the `"gelu"` activation function.
|
||||
max_position_embeddings (`int`, *optional*, defaults to 8192):
|
||||
The maximum sequence length that this model might ever be used with.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
||||
The epsilon used by the rms normalization layers.
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
||||
relevant if `config.is_decoder=True`.
|
||||
pad_token_id (`int`, *optional*, defaults to 0):
|
||||
Padding token id.
|
||||
eos_token_id (`int`, *optional*, defaults to 1):
|
||||
End of stream token id.
|
||||
bos_token_id (`int`, *optional*, defaults to 2):
|
||||
Beginning of stream token id.
|
||||
tie_word_embeddings (`bool`, *optional*, defaults to `True`):
|
||||
Whether to tie weight embeddings
|
||||
rope_theta (`float`, *optional*, defaults to 10000.0):
|
||||
The base period of the RoPE embeddings.
|
||||
attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
|
||||
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
||||
attention_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for the attention probabilities.
|
||||
```python
|
||||
>>> from transformers import NewModelModel, NewModelConfig
|
||||
>>> # Initializing a NewModel new_model-7b style configuration
|
||||
>>> configuration = NewModelConfig()
|
||||
>>> # Initializing a model from the new_model-7b style configuration
|
||||
>>> model = NewModelModel(configuration)
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "new_model"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=256030,
|
||||
hidden_size=64,
|
||||
intermediate_size=90,
|
||||
num_hidden_layers=28,
|
||||
num_attention_heads=16,
|
||||
num_key_value_heads=16,
|
||||
head_dim=256,
|
||||
hidden_act="gelu_pytorch_tanh",
|
||||
hidden_activation=None,
|
||||
max_position_embeddings=1500,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-6,
|
||||
use_cache=True,
|
||||
pad_token_id=0,
|
||||
eos_token_id=1,
|
||||
bos_token_id=2,
|
||||
tie_word_embeddings=True,
|
||||
rope_theta=10000.0,
|
||||
attention_bias=False,
|
||||
attention_dropout=0.0,
|
||||
**kwargs,
|
||||
):
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.intermediate_size = intermediate_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.head_dim = head_dim
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_activation = hidden_activation
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.attention_bias = attention_bias
|
||||
self.attention_dropout = attention_dropout
|
||||
|
||||
super().__init__(
|
||||
pad_token_id=pad_token_id,
|
||||
bos_token_id=bos_token_id,
|
||||
eos_token_id=eos_token_id,
|
||||
tie_word_embeddings=tie_word_embeddings,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
@property
|
||||
def num_heads(self):
|
||||
return self.num_attention_heads
|
||||
@@ -1,7 +1,7 @@
|
||||
#!/bin/bash
|
||||
|
||||
# Iterate over each file in the current directory
|
||||
for file in examples/diff-conversion/diff_*; do
|
||||
for file in examples/modular-transformers/modular_*; do
|
||||
# Check if it's a regular file
|
||||
if [ -f "$file" ]; then
|
||||
# Call the Python script with the file name as an argument
|
||||
1053
examples/modular-transformers/modeling_dummy.py
Normal file
1053
examples/modular-transformers/modeling_dummy.py
Normal file
File diff suppressed because it is too large
Load Diff
1038
examples/modular-transformers/modeling_dummy_bert.py
Normal file
1038
examples/modular-transformers/modeling_dummy_bert.py
Normal file
File diff suppressed because it is too large
Load Diff
1059
examples/modular-transformers/modeling_my_new_model2.py
Normal file
1059
examples/modular-transformers/modeling_my_new_model2.py
Normal file
File diff suppressed because it is too large
Load Diff
953
examples/modular-transformers/modeling_super.py
Normal file
953
examples/modular-transformers/modeling_super.py
Normal file
@@ -0,0 +1,953 @@
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
# This file was automatically generated from <path_to_diff_file.py>.
|
||||
# Do NOT edit this file manually as any edits will be overwritten by the generation of
|
||||
# the file from the diff. If any change should be done, please apply the change to the
|
||||
# diff.py file directly. One of our CI enforces this
|
||||
# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
|
||||
import math
|
||||
from typing import List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torch.utils.checkpoint
|
||||
from torch import nn
|
||||
|
||||
from ...activations import ACT2FN
|
||||
from ...cache_utils import Cache, StaticCache
|
||||
from ...modeling_attn_mask_utils import AttentionMaskConverter
|
||||
from ...modeling_flash_attention_utils import _flash_attention_forward
|
||||
from ...modeling_outputs import (
|
||||
BaseModelOutputWithPast,
|
||||
)
|
||||
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
|
||||
from ...modeling_utils import PreTrainedModel
|
||||
from ...utils import (
|
||||
add_start_docstrings,
|
||||
add_start_docstrings_to_model_forward,
|
||||
is_flash_attn_greater_or_equal_2_10,
|
||||
logging,
|
||||
)
|
||||
from .configuration_super import SuperConfig
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
|
||||
def _prepare_4d_causal_attention_mask_with_cache_position(
|
||||
attention_mask: torch.Tensor,
|
||||
sequence_length: int,
|
||||
target_length: int,
|
||||
dtype: torch.dtype,
|
||||
device: torch.device,
|
||||
min_dtype: float,
|
||||
cache_position: torch.Tensor,
|
||||
batch_size: int,
|
||||
):
|
||||
"""
|
||||
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
||||
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
||||
|
||||
Args:
|
||||
attention_mask (`torch.Tensor`):
|
||||
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
|
||||
sequence_length (`int`):
|
||||
The sequence length being processed.
|
||||
target_length (`int`):
|
||||
The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
|
||||
dtype (`torch.dtype`):
|
||||
The dtype to use for the 4D attention mask.
|
||||
device (`torch.device`):
|
||||
The device to plcae the 4D attention mask on.
|
||||
min_dtype (`float`):
|
||||
The minimum value representable with the dtype `dtype`.
|
||||
cache_position (`torch.Tensor`):
|
||||
Indices depicting the position of the input sequence tokens in the sequence.
|
||||
batch_size (`torch.Tensor`):
|
||||
Batch size.
|
||||
"""
|
||||
if attention_mask is not None and attention_mask.dim() == 4:
|
||||
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
||||
causal_mask = attention_mask
|
||||
else:
|
||||
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
||||
if sequence_length != 1:
|
||||
causal_mask = torch.triu(causal_mask, diagonal=1)
|
||||
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
||||
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
||||
if attention_mask is not None:
|
||||
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
||||
mask_length = attention_mask.shape[-1]
|
||||
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
|
||||
padding_mask = padding_mask == 0
|
||||
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
||||
padding_mask, min_dtype
|
||||
)
|
||||
|
||||
return causal_mask
|
||||
|
||||
|
||||
class SuperRMSNorm(nn.Module):
|
||||
def __init__(self, hidden_size, eps=1e-6):
|
||||
"""
|
||||
SuperRMSNorm is equivalent to T5LayerNorm
|
||||
"""
|
||||
super().__init__()
|
||||
self.weight = nn.Parameter(torch.ones(hidden_size))
|
||||
self.variance_epsilon = eps
|
||||
|
||||
def forward(self, hidden_states):
|
||||
input_dtype = hidden_states.dtype
|
||||
hidden_states = hidden_states.to(torch.float32)
|
||||
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
||||
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
||||
return self.weight * hidden_states.to(input_dtype)
|
||||
|
||||
def extra_repr(self):
|
||||
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
||||
|
||||
|
||||
class SuperRotaryEmbedding(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim=None,
|
||||
max_position_embeddings=2048,
|
||||
base=10000,
|
||||
device=None,
|
||||
scaling_factor=1.0,
|
||||
rope_type="default",
|
||||
config: Optional[SuperConfig] = None,
|
||||
):
|
||||
super().__init__()
|
||||
# TODO (joao): remove the `if` below, only used for BC
|
||||
self.rope_kwargs = {}
|
||||
if config is None:
|
||||
logger.warning_once(
|
||||
"`SuperRotaryEmbedding` can now be fully parameterized by passing the model config through the "
|
||||
"`config` argument. All other arguments will be removed in v4.45"
|
||||
)
|
||||
self.rope_kwargs = {
|
||||
"rope_type": rope_type,
|
||||
"factor": scaling_factor,
|
||||
"dim": dim,
|
||||
"base": base,
|
||||
"max_position_embeddings": max_position_embeddings,
|
||||
}
|
||||
self.rope_type = rope_type
|
||||
self.max_seq_len_cached = max_position_embeddings
|
||||
self.original_max_seq_len = max_position_embeddings
|
||||
else:
|
||||
# BC: "rope_type" was originally "type"
|
||||
if config.rope_scaling is not None:
|
||||
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
||||
else:
|
||||
self.rope_type = "default"
|
||||
self.max_seq_len_cached = config.max_position_embeddings
|
||||
self.original_max_seq_len = config.max_position_embeddings
|
||||
|
||||
self.config = config
|
||||
self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
|
||||
|
||||
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
|
||||
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
||||
self.original_inv_freq = self.inv_freq
|
||||
|
||||
def _dynamic_frequency_update(self, position_ids, device):
|
||||
"""
|
||||
dynamic RoPE layers should recompute `inv_freq` in the following situations:
|
||||
1 - growing beyond the cached sequence length (allow scaling)
|
||||
2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
|
||||
"""
|
||||
seq_len = torch.max(position_ids) + 1
|
||||
if seq_len > self.max_seq_len_cached: # growth
|
||||
inv_freq, self.attention_scaling = self.rope_init_fn(
|
||||
self.config, device, seq_len=seq_len, **self.rope_kwargs
|
||||
)
|
||||
self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation
|
||||
self.max_seq_len_cached = seq_len
|
||||
|
||||
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
|
||||
self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
|
||||
self.max_seq_len_cached = self.original_max_seq_len
|
||||
|
||||
@torch.no_grad()
|
||||
def forward(self, x, position_ids):
|
||||
if "dynamic" in self.rope_type:
|
||||
self._dynamic_frequency_update(position_ids, device=x.device)
|
||||
|
||||
# Core RoPE block
|
||||
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
||||
position_ids_expanded = position_ids[:, None, :].float()
|
||||
# Force float32 (see https://github.com/huggingface/transformers/pull/29285)
|
||||
device_type = x.device.type
|
||||
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
||||
with torch.autocast(device_type=device_type, enabled=False):
|
||||
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
||||
emb = torch.cat((freqs, freqs), dim=-1)
|
||||
cos = emb.cos()
|
||||
sin = emb.sin()
|
||||
|
||||
# Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
|
||||
cos = cos * self.attention_scaling
|
||||
sin = sin * self.attention_scaling
|
||||
|
||||
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
||||
|
||||
|
||||
def rotate_half(x):
|
||||
"""Rotates half the hidden dims of the input."""
|
||||
x1 = x[..., : x.shape[-1] // 2]
|
||||
x2 = x[..., x.shape[-1] // 2 :]
|
||||
return torch.cat((-x2, x1), dim=-1)
|
||||
|
||||
|
||||
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
||||
"""Applies Rotary Position Embedding to the query and key tensors.
|
||||
|
||||
Args:
|
||||
q (`torch.Tensor`): The query tensor.
|
||||
k (`torch.Tensor`): The key tensor.
|
||||
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
||||
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
||||
position_ids (`torch.Tensor`, *optional*):
|
||||
Deprecated and unused.
|
||||
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
||||
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
||||
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
||||
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
||||
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
||||
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
||||
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
||||
Returns:
|
||||
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
||||
"""
|
||||
cos = cos.unsqueeze(unsqueeze_dim)
|
||||
sin = sin.unsqueeze(unsqueeze_dim)
|
||||
q_embed = (q * cos) + (rotate_half(q) * sin)
|
||||
k_embed = (k * cos) + (rotate_half(k) * sin)
|
||||
return q_embed, k_embed
|
||||
|
||||
|
||||
class SuperMLP(nn.Module):
|
||||
def __init__(self, config):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.hidden_size = config.hidden_size
|
||||
self.intermediate_size = config.intermediate_size
|
||||
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
|
||||
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
|
||||
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
|
||||
self.act_fn = ACT2FN[config.hidden_act]
|
||||
|
||||
def forward(self, x):
|
||||
if self.config.pretraining_tp > 1:
|
||||
slice = self.intermediate_size // self.config.pretraining_tp
|
||||
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
|
||||
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
|
||||
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
|
||||
|
||||
gate_proj = torch.cat(
|
||||
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
|
||||
)
|
||||
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
|
||||
|
||||
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
|
||||
down_proj = [
|
||||
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
|
||||
]
|
||||
down_proj = sum(down_proj)
|
||||
else:
|
||||
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
||||
|
||||
return down_proj
|
||||
|
||||
|
||||
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
||||
"""
|
||||
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
||||
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
||||
"""
|
||||
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
||||
if n_rep == 1:
|
||||
return hidden_states
|
||||
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
||||
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
||||
|
||||
|
||||
class SuperAttention(nn.Module):
|
||||
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
||||
|
||||
def __init__(self, config: SuperConfig, layer_idx: Optional[int] = None):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.layer_idx = layer_idx
|
||||
if layer_idx is None:
|
||||
logger.warning_once(
|
||||
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
||||
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
||||
"when creating this class."
|
||||
)
|
||||
|
||||
self.attention_dropout = config.attention_dropout
|
||||
self.hidden_size = config.hidden_size
|
||||
self.num_heads = config.num_attention_heads
|
||||
self.head_dim = getattr(config, "head_dim", self.hidden_size // self.num_heads)
|
||||
self.num_key_value_heads = config.num_key_value_heads
|
||||
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
||||
self.max_position_embeddings = config.max_position_embeddings
|
||||
self.rope_theta = config.rope_theta
|
||||
self.is_causal = True
|
||||
|
||||
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
||||
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
||||
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
||||
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
|
||||
|
||||
# TODO (joao): remove in v4.45 (RoPE is computed in the model, not in the decoder layers)
|
||||
self.rotary_emb = SuperRotaryEmbedding(config=self.config)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
position_ids: Optional[torch.LongTensor] = None,
|
||||
past_key_value: Optional[Cache] = None,
|
||||
output_attentions: bool = False,
|
||||
use_cache: bool = False,
|
||||
cache_position: Optional[torch.LongTensor] = None,
|
||||
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
||||
**kwargs,
|
||||
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
||||
bsz, q_len, _ = hidden_states.size()
|
||||
|
||||
if self.config.pretraining_tp > 1:
|
||||
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
|
||||
query_slices = self.q_proj.weight.split(
|
||||
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
|
||||
)
|
||||
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
|
||||
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
|
||||
|
||||
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
|
||||
query_states = torch.cat(query_states, dim=-1)
|
||||
|
||||
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
|
||||
key_states = torch.cat(key_states, dim=-1)
|
||||
|
||||
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
|
||||
value_states = torch.cat(value_states, dim=-1)
|
||||
|
||||
else:
|
||||
query_states = self.q_proj(hidden_states)
|
||||
key_states = self.k_proj(hidden_states)
|
||||
value_states = self.v_proj(hidden_states)
|
||||
|
||||
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
|
||||
if position_embeddings is None:
|
||||
logger.warning_once(
|
||||
"The attention layers in this model are transitioning from computing the RoPE embeddings internally "
|
||||
"through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
|
||||
"`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
|
||||
"removed and `position_embeddings` will be mandatory."
|
||||
)
|
||||
cos, sin = self.rotary_emb(value_states, position_ids)
|
||||
else:
|
||||
cos, sin = position_embeddings
|
||||
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
||||
|
||||
if past_key_value is not None:
|
||||
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
||||
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
||||
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
||||
|
||||
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
||||
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
||||
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
||||
|
||||
if attention_mask is not None: # no matter the length, we just slice it
|
||||
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
||||
attn_weights = attn_weights + causal_mask
|
||||
|
||||
# upcast attention to fp32
|
||||
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
||||
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
||||
attn_output = torch.matmul(attn_weights, value_states)
|
||||
|
||||
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
||||
raise ValueError(
|
||||
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
||||
f" {attn_output.size()}"
|
||||
)
|
||||
|
||||
attn_output = attn_output.transpose(1, 2).contiguous()
|
||||
|
||||
attn_output = attn_output.reshape(bsz, q_len, -1)
|
||||
|
||||
if self.config.pretraining_tp > 1:
|
||||
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
|
||||
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
|
||||
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
|
||||
else:
|
||||
attn_output = self.o_proj(attn_output)
|
||||
|
||||
if not output_attentions:
|
||||
attn_weights = None
|
||||
|
||||
return attn_output, attn_weights, past_key_value
|
||||
|
||||
|
||||
class SuperFlashAttention2(SuperAttention):
|
||||
"""
|
||||
Super flash attention module. This module inherits from `SuperAttention` as the weights of the module stays
|
||||
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
||||
flash attention and deal with padding tokens in case the input contains any of them.
|
||||
"""
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
||||
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
||||
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
||||
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.LongTensor] = None,
|
||||
position_ids: Optional[torch.LongTensor] = None,
|
||||
past_key_value: Optional[Cache] = None,
|
||||
output_attentions: bool = False,
|
||||
use_cache: bool = False,
|
||||
cache_position: Optional[torch.LongTensor] = None,
|
||||
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
||||
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
||||
if isinstance(past_key_value, StaticCache):
|
||||
raise ValueError(
|
||||
"`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
|
||||
"make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
|
||||
)
|
||||
|
||||
output_attentions = False
|
||||
|
||||
bsz, q_len, _ = hidden_states.size()
|
||||
|
||||
query_states = self.q_proj(hidden_states)
|
||||
key_states = self.k_proj(hidden_states)
|
||||
value_states = self.v_proj(hidden_states)
|
||||
|
||||
# Flash attention requires the input to have the shape
|
||||
# batch_size x seq_length x head_dim x hidden_dim
|
||||
# therefore we just need to keep the original shape
|
||||
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
|
||||
if position_embeddings is None:
|
||||
logger.warning_once(
|
||||
"The attention layers in this model are transitioning from computing the RoPE embeddings internally "
|
||||
"through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
|
||||
"`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
|
||||
"removed and `position_embeddings` will be mandatory."
|
||||
)
|
||||
cos, sin = self.rotary_emb(value_states, position_ids)
|
||||
else:
|
||||
cos, sin = position_embeddings
|
||||
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
||||
|
||||
if past_key_value is not None:
|
||||
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
||||
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
||||
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
||||
|
||||
# TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
|
||||
# to be able to avoid many of these transpose/reshape/view.
|
||||
query_states = query_states.transpose(1, 2)
|
||||
key_states = key_states.transpose(1, 2)
|
||||
value_states = value_states.transpose(1, 2)
|
||||
|
||||
dropout_rate = self.attention_dropout if self.training else 0.0
|
||||
|
||||
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
||||
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
||||
# cast them back in the correct dtype just to be sure everything works as expected.
|
||||
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
||||
# in fp32. (SuperRMSNorm handles it correctly)
|
||||
|
||||
input_dtype = query_states.dtype
|
||||
if input_dtype == torch.float32:
|
||||
if torch.is_autocast_enabled():
|
||||
target_dtype = torch.get_autocast_gpu_dtype()
|
||||
# Handle the case where the model is quantized
|
||||
elif hasattr(self.config, "_pre_quantization_dtype"):
|
||||
target_dtype = self.config._pre_quantization_dtype
|
||||
else:
|
||||
target_dtype = self.q_proj.weight.dtype
|
||||
|
||||
logger.warning_once(
|
||||
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
||||
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
||||
f" {target_dtype}."
|
||||
)
|
||||
|
||||
query_states = query_states.to(target_dtype)
|
||||
key_states = key_states.to(target_dtype)
|
||||
value_states = value_states.to(target_dtype)
|
||||
|
||||
attn_output = _flash_attention_forward(
|
||||
query_states,
|
||||
key_states,
|
||||
value_states,
|
||||
attention_mask,
|
||||
q_len,
|
||||
position_ids=position_ids,
|
||||
dropout=dropout_rate,
|
||||
sliding_window=getattr(self, "sliding_window", None),
|
||||
use_top_left_mask=self._flash_attn_uses_top_left_mask,
|
||||
is_causal=self.is_causal,
|
||||
)
|
||||
|
||||
attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
|
||||
attn_output = self.o_proj(attn_output)
|
||||
|
||||
if not output_attentions:
|
||||
attn_weights = None
|
||||
|
||||
return attn_output, attn_weights, past_key_value
|
||||
|
||||
|
||||
class SuperSdpaAttention(SuperAttention):
|
||||
"""
|
||||
Super attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
||||
`SuperAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
||||
SDPA API.
|
||||
"""
|
||||
|
||||
# Adapted from SuperAttention.forward
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
position_ids: Optional[torch.LongTensor] = None,
|
||||
past_key_value: Optional[Cache] = None,
|
||||
output_attentions: bool = False,
|
||||
use_cache: bool = False,
|
||||
cache_position: Optional[torch.LongTensor] = None,
|
||||
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
||||
**kwargs,
|
||||
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
||||
if output_attentions:
|
||||
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
||||
logger.warning_once(
|
||||
"SuperModel is using SuperSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
||||
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
||||
)
|
||||
return super().forward(
|
||||
hidden_states=hidden_states,
|
||||
attention_mask=attention_mask,
|
||||
position_ids=position_ids,
|
||||
past_key_value=past_key_value,
|
||||
output_attentions=output_attentions,
|
||||
use_cache=use_cache,
|
||||
cache_position=cache_position,
|
||||
position_embeddings=position_embeddings,
|
||||
)
|
||||
|
||||
bsz, q_len, _ = hidden_states.size()
|
||||
|
||||
query_states = self.q_proj(hidden_states)
|
||||
key_states = self.k_proj(hidden_states)
|
||||
value_states = self.v_proj(hidden_states)
|
||||
|
||||
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
||||
|
||||
if position_embeddings is None:
|
||||
logger.warning_once(
|
||||
"The attention layers in this model are transitioning from computing the RoPE embeddings internally "
|
||||
"through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
|
||||
"`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
|
||||
"removed and `position_embeddings` will be mandatory."
|
||||
)
|
||||
cos, sin = self.rotary_emb(value_states, position_ids)
|
||||
else:
|
||||
cos, sin = position_embeddings
|
||||
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
||||
|
||||
if past_key_value is not None:
|
||||
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
||||
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
||||
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
||||
|
||||
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
||||
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
||||
|
||||
causal_mask = attention_mask
|
||||
if attention_mask is not None:
|
||||
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
|
||||
|
||||
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
||||
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
||||
if query_states.device.type == "cuda" and causal_mask is not None:
|
||||
query_states = query_states.contiguous()
|
||||
key_states = key_states.contiguous()
|
||||
value_states = value_states.contiguous()
|
||||
|
||||
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
|
||||
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
|
||||
is_causal = True if causal_mask is None and q_len > 1 else False
|
||||
|
||||
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
||||
query_states,
|
||||
key_states,
|
||||
value_states,
|
||||
attn_mask=causal_mask,
|
||||
dropout_p=self.attention_dropout if self.training else 0.0,
|
||||
is_causal=is_causal,
|
||||
)
|
||||
|
||||
attn_output = attn_output.transpose(1, 2).contiguous()
|
||||
attn_output = attn_output.view(bsz, q_len, -1)
|
||||
|
||||
attn_output = self.o_proj(attn_output)
|
||||
|
||||
return attn_output, None, past_key_value
|
||||
|
||||
|
||||
SUPER_ATTENTION_CLASSES = {
|
||||
"eager": SuperAttention,
|
||||
"flash_attention_2": SuperFlashAttention2,
|
||||
"sdpa": SuperSdpaAttention,
|
||||
}
|
||||
|
||||
|
||||
class SuperDecoderLayer(nn.Module):
|
||||
def __init__(self, config: SuperConfig, layer_idx: int):
|
||||
super().__init__()
|
||||
self.hidden_size = config.hidden_size
|
||||
|
||||
self.self_attn = SUPER_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
|
||||
|
||||
self.mlp = SuperMLP(config)
|
||||
self.input_layernorm = SuperRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.post_attention_layernorm = SuperRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
position_ids: Optional[torch.LongTensor] = None,
|
||||
past_key_value: Optional[Cache] = None,
|
||||
output_attentions: Optional[bool] = False,
|
||||
use_cache: Optional[bool] = False,
|
||||
cache_position: Optional[torch.LongTensor] = None,
|
||||
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
||||
**kwargs,
|
||||
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
||||
"""
|
||||
Args:
|
||||
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
||||
attention_mask (`torch.FloatTensor`, *optional*):
|
||||
attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
|
||||
query_sequence_length, key_sequence_length)` if default attention is used.
|
||||
output_attentions (`bool`, *optional*):
|
||||
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
||||
returned tensors for more detail.
|
||||
use_cache (`bool`, *optional*):
|
||||
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
||||
(see `past_key_values`).
|
||||
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
||||
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
||||
Indices depicting the position of the input sequence tokens in the sequence
|
||||
position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
|
||||
Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
|
||||
with `head_dim` being the embedding dimension of each attention head.
|
||||
kwargs (`dict`, *optional*):
|
||||
Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
|
||||
into the model
|
||||
"""
|
||||
residual = hidden_states
|
||||
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
|
||||
# Self Attention
|
||||
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
||||
hidden_states=hidden_states,
|
||||
attention_mask=attention_mask,
|
||||
position_ids=position_ids,
|
||||
past_key_value=past_key_value,
|
||||
output_attentions=output_attentions,
|
||||
use_cache=use_cache,
|
||||
cache_position=cache_position,
|
||||
position_embeddings=position_embeddings,
|
||||
**kwargs,
|
||||
)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
# Fully Connected
|
||||
residual = hidden_states
|
||||
hidden_states = self.post_attention_layernorm(hidden_states)
|
||||
hidden_states = self.mlp(hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
outputs = (hidden_states,)
|
||||
|
||||
if output_attentions:
|
||||
outputs += (self_attn_weights,)
|
||||
|
||||
if use_cache:
|
||||
outputs += (present_key_value,)
|
||||
|
||||
return outputs
|
||||
|
||||
|
||||
SUPER_START_DOCSTRING = r"""
|
||||
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
||||
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
||||
etc.)
|
||||
|
||||
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
||||
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
||||
and behavior.
|
||||
|
||||
Parameters:
|
||||
config ([`SuperConfig`]):
|
||||
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
||||
load the weights associated with the model, only the configuration. Check out the
|
||||
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
||||
"""
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"The bare Super Model outputting raw hidden-states without any specific head on top.",
|
||||
SUPER_START_DOCSTRING,
|
||||
)
|
||||
class SuperPreTrainedModel(PreTrainedModel):
|
||||
config_class = SuperConfig
|
||||
base_model_prefix = "model"
|
||||
supports_gradient_checkpointing = True
|
||||
_no_split_modules = ["SuperDecoderLayer"]
|
||||
_skip_keys_device_placement = ["past_key_values"]
|
||||
_supports_flash_attn_2 = True
|
||||
_supports_sdpa = True
|
||||
_supports_cache_class = True
|
||||
_supports_quantized_cache = True
|
||||
_supports_static_cache = True
|
||||
|
||||
def _init_weights(self, module):
|
||||
std = self.config.initializer_range
|
||||
if isinstance(module, nn.Linear):
|
||||
module.weight.data.normal_(mean=0.0, std=std)
|
||||
if module.bias is not None:
|
||||
module.bias.data.zero_()
|
||||
elif isinstance(module, nn.Embedding):
|
||||
module.weight.data.normal_(mean=0.0, std=std)
|
||||
if module.padding_idx is not None:
|
||||
module.weight.data[module.padding_idx].zero_()
|
||||
|
||||
|
||||
SUPER_INPUTS_DOCSTRING = r"""
|
||||
Args:
|
||||
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
||||
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
||||
it.
|
||||
|
||||
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
||||
[`PreTrainedTokenizer.__call__`] for details.
|
||||
|
||||
[What are input IDs?](../glossary#input-ids)
|
||||
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
||||
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
||||
|
||||
- 1 for tokens that are **not masked**,
|
||||
- 0 for tokens that are **masked**.
|
||||
|
||||
[What are attention masks?](../glossary#attention-mask)
|
||||
|
||||
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
||||
[`PreTrainedTokenizer.__call__`] for details.
|
||||
|
||||
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
|
||||
`past_key_values`).
|
||||
|
||||
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
||||
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
||||
information on the default strategy.
|
||||
|
||||
- 1 indicates the head is **not masked**,
|
||||
- 0 indicates the head is **masked**.
|
||||
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
||||
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
||||
config.n_positions - 1]`.
|
||||
|
||||
[What are position IDs?](../glossary#position-ids)
|
||||
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
||||
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
||||
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
||||
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
||||
|
||||
Two formats are allowed:
|
||||
- a [`~cache_utils.Cache`] instance;
|
||||
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
||||
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
||||
cache format.
|
||||
|
||||
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
||||
legacy cache format will be returned.
|
||||
|
||||
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
||||
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
||||
of shape `(batch_size, sequence_length)`.
|
||||
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
||||
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
||||
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
||||
model's internal embedding lookup matrix.
|
||||
use_cache (`bool`, *optional*):
|
||||
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
||||
`past_key_values`).
|
||||
output_attentions (`bool`, *optional*):
|
||||
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
||||
tensors for more detail.
|
||||
output_hidden_states (`bool`, *optional*):
|
||||
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
||||
more detail.
|
||||
return_dict (`bool`, *optional*):
|
||||
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
||||
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
||||
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
||||
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
||||
the complete sequence length.
|
||||
"""
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"The bare Super Model outputting raw hidden-states without any specific head on top.",
|
||||
SUPER_START_DOCSTRING,
|
||||
)
|
||||
class SuperModel(SuperPreTrainedModel):
|
||||
"""
|
||||
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`SuperDecoderLayer`]
|
||||
|
||||
Args:
|
||||
config: SuperConfig
|
||||
"""
|
||||
|
||||
def __init__(self, config: SuperConfig):
|
||||
super().__init__(config)
|
||||
self.padding_idx = config.pad_token_id
|
||||
self.vocab_size = config.vocab_size
|
||||
|
||||
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
||||
self.layers = nn.ModuleList(
|
||||
[SuperDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
||||
)
|
||||
self.norm = SuperRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.rotary_emb = SuperRotaryEmbedding(config=config)
|
||||
self.gradient_checkpointing = False
|
||||
|
||||
# Initialize weights and apply final processing
|
||||
self.post_init()
|
||||
|
||||
def get_input_embeddings(self):
|
||||
return self.embed_tokens
|
||||
|
||||
def set_input_embeddings(self, value):
|
||||
self.embed_tokens = value
|
||||
|
||||
@add_start_docstrings_to_model_forward(SUPER_INPUTS_DOCSTRING)
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.LongTensor = None,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
position_ids: Optional[torch.LongTensor] = None,
|
||||
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
||||
inputs_embeds: Optional[torch.FloatTensor] = None,
|
||||
use_cache: Optional[bool] = None,
|
||||
output_attentions: Optional[bool] = None,
|
||||
output_hidden_states: Optional[bool] = None,
|
||||
return_dict: Optional[bool] = None,
|
||||
cache_position: Optional[torch.LongTensor] = None,
|
||||
) -> Union[Tuple, BaseModelOutputWithPast]:
|
||||
out = super().forward(
|
||||
input_ids,
|
||||
attention_mask,
|
||||
position_ids,
|
||||
past_key_values,
|
||||
inputs_embeds,
|
||||
use_cache,
|
||||
output_attentions,
|
||||
output_hidden_states,
|
||||
return_dict,
|
||||
cache_position,
|
||||
)
|
||||
out.logits *= 2**4
|
||||
return out
|
||||
|
||||
def _update_causal_mask(
|
||||
self,
|
||||
attention_mask: torch.Tensor,
|
||||
input_tensor: torch.Tensor,
|
||||
cache_position: torch.Tensor,
|
||||
past_key_values: Cache,
|
||||
output_attentions: bool,
|
||||
):
|
||||
if self.config._attn_implementation == "flash_attention_2":
|
||||
if attention_mask is not None and 0.0 in attention_mask:
|
||||
return attention_mask
|
||||
return None
|
||||
|
||||
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
||||
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
||||
# to infer the attention mask.
|
||||
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
||||
using_static_cache = isinstance(past_key_values, StaticCache)
|
||||
|
||||
# When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
|
||||
if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
|
||||
if AttentionMaskConverter._ignore_causal_mask_sdpa(
|
||||
attention_mask,
|
||||
inputs_embeds=input_tensor,
|
||||
past_key_values_length=past_seen_tokens,
|
||||
is_training=self.training,
|
||||
):
|
||||
return None
|
||||
|
||||
dtype, device = input_tensor.dtype, input_tensor.device
|
||||
min_dtype = torch.finfo(dtype).min
|
||||
sequence_length = input_tensor.shape[1]
|
||||
if using_static_cache:
|
||||
target_length = past_key_values.get_max_length()
|
||||
else:
|
||||
target_length = (
|
||||
attention_mask.shape[-1]
|
||||
if isinstance(attention_mask, torch.Tensor)
|
||||
else past_seen_tokens + sequence_length + 1
|
||||
)
|
||||
|
||||
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
||||
causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
|
||||
attention_mask,
|
||||
sequence_length=sequence_length,
|
||||
target_length=target_length,
|
||||
dtype=dtype,
|
||||
device=device,
|
||||
min_dtype=min_dtype,
|
||||
cache_position=cache_position,
|
||||
batch_size=input_tensor.shape[0],
|
||||
)
|
||||
|
||||
if (
|
||||
self.config._attn_implementation == "sdpa"
|
||||
and attention_mask is not None
|
||||
and attention_mask.device.type == "cuda"
|
||||
and not output_attentions
|
||||
):
|
||||
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
||||
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
||||
# Details: https://github.com/pytorch/pytorch/issues/110213
|
||||
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
||||
|
||||
return causal_mask
|
||||
@@ -3,10 +3,11 @@ from typing import List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
|
||||
from transformers import Cache
|
||||
from transformers.modeling_outputs import CausalLMOutputWithPast
|
||||
from transformers.models.llama.modeling_llama import LlamaModel
|
||||
|
||||
from ...cache_utils import Cache
|
||||
|
||||
|
||||
def _pre_process_input(input_ids):
|
||||
print(log(input_ids))
|
||||
27
examples/modular-transformers/modular_dummy_bert.py
Normal file
27
examples/modular-transformers/modular_dummy_bert.py
Normal file
@@ -0,0 +1,27 @@
|
||||
from typing import List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
|
||||
from transformers.models.bert.modeling_bert import BertModel
|
||||
|
||||
from ...modeling_outputs import BaseModelOutputWithPoolingAndCrossAttentions
|
||||
|
||||
|
||||
class DummyBertModel(BertModel):
|
||||
def forward(
|
||||
self,
|
||||
input_ids: Optional[torch.Tensor] = None,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
token_type_ids: Optional[torch.Tensor] = None,
|
||||
position_ids: Optional[torch.Tensor] = None,
|
||||
head_mask: Optional[torch.Tensor] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
encoder_hidden_states: Optional[torch.Tensor] = None,
|
||||
encoder_attention_mask: Optional[torch.Tensor] = None,
|
||||
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
||||
use_cache: Optional[bool] = None,
|
||||
output_attentions: Optional[bool] = None,
|
||||
output_hidden_states: Optional[bool] = None,
|
||||
return_dict: Optional[bool] = None,
|
||||
) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
|
||||
return super().forward(input_ids)
|
||||
@@ -5,10 +5,11 @@ from transformers.models.llama.configuration_llama import LlamaConfig
|
||||
# here there is no `ARG` so we are gonna take parent doc
|
||||
class MyNewModelConfig(LlamaConfig):
|
||||
r"""
|
||||
mlp_bias (`bool`, *optional*, defaults to `False`)
|
||||
new_param (`int`, *optional*, defaults to `False`):
|
||||
A fun new parameter
|
||||
"""
|
||||
|
||||
def __init__(self, mlp_bias=True, new_param=0, **super_kwargs):
|
||||
super().__init__(self, **super_kwargs)
|
||||
self.mlp_bias = mlp_bias
|
||||
self.new_param = new_param
|
||||
super().__init__(self, **super_kwargs)
|
||||
@@ -26,5 +26,10 @@ class NewModelConfig(GemmaConfig):
|
||||
rope_theta=10000.0,
|
||||
attention_bias=False,
|
||||
attention_dropout=0.0,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(self)
|
||||
super().__init__(self, **kwargs)
|
||||
|
||||
@property
|
||||
def num_heads(self):
|
||||
return self.num_attention_heads
|
||||
20
examples/modular-transformers/modular_roberta.py
Normal file
20
examples/modular-transformers/modular_roberta.py
Normal file
@@ -0,0 +1,20 @@
|
||||
import torch.nn as nn
|
||||
|
||||
from transformers.models.bert.modeling_bert import BertEmbeddings, BertModel
|
||||
|
||||
|
||||
class RobertaEmbeddings(BertEmbeddings):
|
||||
def __init__(self, config):
|
||||
super().__init__(config)
|
||||
self.pad_token_id = config.pad_token_id
|
||||
self.position_embeddings = nn.Embedding(
|
||||
config.max_position_embeddings, config.hidden_size, config.pad_token_id
|
||||
)
|
||||
|
||||
|
||||
class RobertaModel(BertModel):
|
||||
def __init__(self, config):
|
||||
super().__init__(self, config)
|
||||
# Error out here. Why? Because `RobertaEmbeddings` is defined but not used.
|
||||
# no, because it's defined, and RobertaModel should use RobertaEmbedding
|
||||
# here if initialized that way it won't use the new embedding.
|
||||
@@ -2,10 +2,11 @@ from typing import List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
|
||||
from transformers import Cache
|
||||
from transformers.modeling_outputs import CausalLMOutputWithPast
|
||||
from transformers.models.llama.modeling_llama import LlamaModel
|
||||
|
||||
from ...cache_utils import Cache
|
||||
|
||||
|
||||
# example where we need some deps and some functions
|
||||
class SuperModel(LlamaModel):
|
||||
Reference in New Issue
Block a user