# Copyright 2025 The EasyDeL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import random
import typing as tp
from functools import partial
import chex
import jax
import jax.numpy as jnp
from eformer import common_types
from eformer.escale import apply_logical_sharding
from ejkernel.types import MaskInfo
from flax import nnx as nn
from jax import lax
from jax.ad_checkpoint import checkpoint_name
from jaxtyping import Array, Bool, Float, Int
from easydel.inference.logits_process import (
ForceTokensLogitsProcessor,
LogitsProcessorList,
WhisperTimeStampLogitsProcessor,
)
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.loss_utils import LossConfig, LossMetrics
from easydel.infra.modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPastAndCrossAttentions,
Seq2SeqLMOutput,
Seq2SeqModelOutput,
SequenceClassifierOutput,
)
from easydel.infra.utils import ACT2FN, auto_remat, get_dot_general_by_bits
from easydel.layers.attention import AttentionModule, FlexibleAttentionModule
from easydel.layers.caching import TransformerCache, TransformerCacheView, TransformerMetadata
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from .whisper_configuration import WhisperConfig as WhisperConfig
remat = nn.remat
[docs]def shift_tokens_right(
input_ids: Int[Array, "batch seq_len"],
pad_token_id: int,
decoder_start_token_id: int,
):
"""
Shift input ids one token to the right using JAX.
"""
batch_size, seq_length = input_ids.shape
shifted_input_ids = jnp.full(
(batch_size, seq_length),
pad_token_id,
dtype=input_ids.dtype,
)
shifted_input_ids = shifted_input_ids.at[:, 1:].set(input_ids[:, :-1])
shifted_input_ids = shifted_input_ids.at[:, 0].set(decoder_start_token_id)
shifted_input_ids = jnp.where(input_ids == -100, pad_token_id, shifted_input_ids)
return shifted_input_ids
[docs]def sinusoidal_embedding_init(key, shape, dtype=jnp.float_) -> jax.Array:
"""Initializes sinusoidal positional embeddings.
Args:
key: JAX PRNG key (unused, but part of standard initializer signature).
shape (tuple): Shape of the embedding matrix (length, channels).
dtype: Data type of the embeddings (default: jnp.float_).
Returns:
jax.Array: Sinusoidal positional embedding matrix.
Raises:
ValueError: If the number of channels is not even.
"""
length, channels = shape
if channels % 2 != 0:
raise ValueError(
f"Number of channels has to be divisible by 2 for sinusoidal positional embeddings, got {channels} channels."
)
log_timescale_increment = math.log(10000) / (channels // 2 - 1)
inv_timescales = jnp.exp(-log_timescale_increment * jnp.arange(channels // 2))
scaled_time = jnp.arange(length).reshape(-1, 1) * inv_timescales.reshape(1, -1)
return jnp.concatenate([jnp.sin(scaled_time), jnp.cos(scaled_time)], axis=1).astype(dtype)
[docs]class WhisperAttention(AttentionModule):
"""Whisper Attention mechanism.
This module implements the standard multi-head attention mechanism used
in both the encoder and decoder of the Whisper model.
Attributes:
config (WhisperConfig): Configuration object for the model.
embed_dim (int): Dimensionality of the embedding layer.
num_heads (int): Number of attention heads.
dropout (float): Dropout probability.
causal (bool): Whether this attention is causal (used in decoder self-attention).
bias (bool): Whether to include bias in linear projections.
head_dim (int): Dimensionality of each attention head.
q_proj (ParallelLinear): Linear layer for query projection.
k_proj (ParallelLinear): Linear layer for key projection.
v_proj (ParallelLinear): Linear layer for value projection.
out_proj (ParallelLinear): Linear layer for output projection.
attention_performer (FlexibleAttentionModule): Module for performing attention computation.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
causal: bool = False,
bias: bool = True,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
) -> None:
"""Initializes the WhisperAttention module.
Args:
config (WhisperConfig): The configuration object for the model.
embed_dim (int): Dimensionality of the input and output features.
num_heads (int): Number of attention heads.
dropout (float): Dropout probability for attention weights (default: 0.0).
causal (bool): Whether to apply causal masking (default: False).
bias (bool): Whether to include bias terms in the projection layers (default: True).
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
Raises:
ValueError: If `embed_dim` is not divisible by `num_heads`.
"""
super().__init__(config=config)
self.rngs = rngs
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.causal = causal
self.bias = bias
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
f" and `num_heads`: {self.num_heads})."
)
linear = partial(
ColumnParallelLinear,
self.embed_dim,
self.embed_dim,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(self.config.init_std),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.q_proj = linear(use_bias=self.bias, rngs=rngs)
self.k_proj = linear(use_bias=False, rngs=rngs)
self.v_proj = linear(use_bias=self.bias, rngs=rngs)
self.out_proj = linear(use_bias=self.bias, rngs=rngs)
self.attention_performer = FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=config.attention_dropout,
)
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
key_value_states: jnp.ndarray | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
cache_view: TransformerCacheView | None = None,
cache_metadata: TransformerMetadata | None = None,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
) -> tuple[tp.Any, tp.Any]:
"""Forward pass of the attention module.
Args:
hidden_states (jnp.ndarray): Input hidden states (batch, seq_len, embed_dim).
key_value_states (tp.Optional[jnp.ndarray]): Optional key/value states for cross-attention
(batch, kv_seq_len, embed_dim). If None, self-attention is performed.
cache_view (tp.Optional[TransformerCacheView]): Cache view for key/value states, used in causal attention.
cache_metadata (tp.Optional[TransformerMetadata]): Metadata for paged attention.
attention_mask (tp.Optional[jnp.ndarray]): Mask to apply to attention scores (batch, 1, seq_len, kv_seq_len).
causal_mask (tp.Optional[jnp.ndarray]): Causal mask specific to this module
(required if self.causal is True).
Returns:
tuple[jnp.ndarray, jnp.ndarray]: A tuple containing:
- attn_output (jnp.ndarray): Attention output (batch, seq_len, embed_dim).
- attn_weights (jnp.ndarray): Attention weights (batch, num_heads, seq_len, kv_seq_len).
"""
is_cross_attention = key_value_states is not None
query_states = checkpoint_name(self.q_proj(hidden_states), "attn_query")
if is_cross_attention:
key_states = checkpoint_name(self.k_proj(key_value_states), "attn_key")
value_states = checkpoint_name(self.v_proj(key_value_states), "attn_value")
else:
key_states = checkpoint_name(self.k_proj(hidden_states), "attn_key")
value_states = checkpoint_name(self.v_proj(hidden_states), "attn_value")
query_states = self._split_heads(query_states)
key_states = self._split_heads(key_states)
value_states = self._split_heads(value_states)
if self.causal:
assert causal_mask is not None, "seems like you forgot to pass causal_mask"
(
key_states,
value_states,
mask_info,
init_attention_bias,
cache_view,
cache_metadata,
) = self.concatenate(
query=query_states,
key=key_states,
cache_view=cache_view,
value=value_states,
attention_mask=attention_mask,
causal_mask=causal_mask,
fcm_mask=None,
)
attention_mask = None
else:
if attention_mask is not None:
attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
init_attention_bias = lambda: None # noqa
attentions = self.attention_performer.forward(
query_states=query_states,
key_states=key_states,
value_states=value_states,
mode=mode,
bias=None,
cache_metadata=cache_metadata,
cache_view=cache_view,
init_bias=init_attention_bias,
mask_info=mask_info,
causal=self.causal,
)
attn_output = checkpoint_name(
self.out_proj(self.shard_attention_prod(self._merge_heads(attentions.attention_outputs))), "attn_output"
)
return attn_output, attentions.attention_outputs, cache_view
def _split_heads(self, hidden_state) -> jnp.ndarray:
"""Splits the last dimension of the hidden state into (num_heads, head_dim)."""
return hidden_state.reshape((*hidden_state.shape[:2], self.num_heads, self.head_dim))
def _merge_heads(self, hidden_state) -> jnp.ndarray:
"""Merges the last two dimensions (num_heads, head_dim) into embed_dim."""
return hidden_state.reshape((*hidden_state.shape[:2], self.embed_dim))
[docs]class WhisperEncoderLayer(nn.Module):
"""A single layer for the Whisper encoder.
This layer consists of a self-attention mechanism followed by a feed-forward
network (FFN), with residual connections and layer normalization.
Attributes:
config (WhisperConfig): Configuration object for the model.
embed_dim (int): Dimensionality of the input and output features.
self_attn (WhisperAttention): Self-attention module.
self_attn_layer_norm (nn.LayerNorm): Layer normalization before self-attention.
fc1 (ParallelLinear): First linear layer of the FFN.
fc2 (ParallelLinear): Second linear layer of the FFN.
final_layer_norm (nn.LayerNorm): Layer normalization after the FFN.
activation_fn (callable): Activation function for the FFN.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
) -> None:
"""Initializes the WhisperEncoderLayer module.
Args:
config (WhisperConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.embed_dim = self.config.d_model
self.self_attn = WhisperAttention(
config=config,
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
linear = partial(
ColumnParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(self.config.init_std),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.self_attn_layer_norm = nn.LayerNorm(
self.embed_dim,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
self.dropout_layer = nn.Dropout(rate=self.config.dropout, rngs=rngs)
self.activation_fn = ACT2FN[self.config.activation_function]
self.activation_dropout_layer = nn.Dropout(
rate=self.config.activation_dropout,
rngs=rngs,
)
self.fc1 = linear(self.embed_dim, self.config.encoder_ffn_dim, rngs=rngs)
self.fc2 = linear(self.config.encoder_ffn_dim, self.embed_dim, rngs=rngs)
self.final_layer_norm = nn.LayerNorm(
self.embed_dim,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
mask_info: MaskInfo,
output_attentions: bool = True,
) -> tuple[jnp.ndarray]:
"""Forward pass of the encoder layer.
Args:
hidden_states (jnp.ndarray): Input hidden states (batch, seq_len, embed_dim).
attention_mask (jnp.ndarray): Attention mask (batch, 1, seq_len, seq_len).
causal_mask (tp.Optional[jnp.ndarray]): Causal mask, usually None for encoder.
output_attentions (bool): Whether to return attention weights (default: True).
Returns:
tp.Tuple[jnp.ndarray, ...]: A tuple containing:
- hidden_states (jnp.ndarray): Output hidden states (batch, seq_len, embed_dim).
- attn_weights (jnp.ndarray, optional): Attention weights if `output_attentions` is True.
"""
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
hidden_states, attn_weights, _ = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
causal_mask=causal_mask,
mode=common_types.MODE_TRAIN, # or prefill?
cache_view=None,
key_value_states=None,
)
hidden_states = self.dropout_layer(hidden_states)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = self.activation_dropout_layer(hidden_states)
hidden_states = checkpoint_name(self.fc2(hidden_states), "mlp_down")
hidden_states = self.dropout_layer(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
[docs]class WhisperDecoderLayer(nn.Module):
"""A single layer for the Whisper decoder.
This layer consists of self-attention, cross-attention (attending to encoder outputs),
and a feed-forward network (FFN), each followed by residual connections and layer normalization.
Attributes:
config (WhisperConfig): Configuration object for the model.
embed_dim (int): Dimensionality of the input and output features.
self_attn (WhisperAttention): Self-attention module (causal).
encoder_attn (WhisperAttention): Cross-attention module (attends to encoder outputs).
self_attn_layer_norm (nn.LayerNorm): Layer normalization before self-attention.
encoder_attn_layer_norm (nn.LayerNorm): Layer normalization before cross-attention.
fc1 (ParallelLinear): First linear layer of the FFN.
fc2 (ParallelLinear): Second linear layer of the FFN.
final_layer_norm (nn.LayerNorm): Layer normalization after the FFN.
activation_fn (callable): Activation function for the FFN.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
) -> None:
"""Initializes the WhisperDecoderLayer module.
Args:
config (WhisperConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.embed_dim = self.config.d_model
self.self_attn = WhisperAttention(
config=self.config,
embed_dim=self.embed_dim,
num_heads=self.config.decoder_attention_heads,
dropout=self.config.attention_dropout,
causal=True,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.dropout_layer = nn.Dropout(
rate=self.config.dropout,
rngs=rngs,
)
self.activation_fn = ACT2FN[self.config.activation_function]
self.activation_dropout_layer = nn.Dropout(
rate=self.config.activation_dropout,
rngs=rngs,
)
self.self_attn_layer_norm = nn.LayerNorm(
self.embed_dim,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
self.encoder_attn = WhisperAttention(
config=self.config,
embed_dim=self.embed_dim,
num_heads=self.config.decoder_attention_heads,
dropout=self.config.attention_dropout,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.encoder_attn_layer_norm = nn.LayerNorm(
self.embed_dim,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
linear = partial(
ColumnParallelLinear,
param_dtype=self.param_dtype,
precision=self.precision,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(self.config.init_std),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.fc1 = linear(
self.embed_dim,
self.config.decoder_ffn_dim,
rngs=rngs,
)
self.fc2 = linear(
self.config.decoder_ffn_dim,
self.embed_dim,
rngs=rngs,
)
self.final_layer_norm = nn.LayerNorm(
self.embed_dim,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
mask_info: MaskInfo,
encoder_hidden_states: Float[Array, "batch seq_len hidden_dim"] | None = None,
encoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
cache_view: TransformerCacheView | None = None,
cache_metadata: TransformerMetadata | None = None,
output_attentions: bool = True,
) -> tuple[jnp.ndarray]:
"""Forward pass of the decoder layer.
Args:
hidden_states (jnp.ndarray): Input hidden states (batch, seq_len, embed_dim).
attention_mask (jnp.ndarray): Attention mask for self-attention (batch, 1, seq_len, seq_len).
causal_mask (tp.Optional[jnp.ndarray]): Causal mask for self-attention.
encoder_hidden_states (tp.Optional[jnp.ndarray]): Hidden states from the encoder
(batch, encoder_seq_len, embed_dim).
encoder_attention_mask (tp.Optional[jnp.ndarray]): Attention mask for cross-attention
(batch, 1, seq_len, encoder_seq_len).
cache_view (tp.Optional[TransformerCacheView]): Cache view for key/value states.
cache_metadata (tp.Optional[TransformerMetadata]): Metadata for paged attention.
output_attentions (bool): Whether to return attention weights (default: True).
Returns:
tp.Tuple[jnp.ndarray, ...]: A tuple containing:
- hidden_states (jnp.ndarray): Output hidden states (batch, seq_len, embed_dim).
- self_attn_weights (jnp.ndarray, optional): Self-attention weights if `output_attentions` is True.
- cross_attn_weights (jnp.ndarray, optional): Cross-attention weights if `output_attentions` is True.
"""
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
# Self Attention
hidden_states, self_attn_weights, cache_view = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
mode=mode,
causal_mask=causal_mask,
cache_view=cache_view,
cache_metadata=cache_metadata,
)
hidden_states = self.dropout_layer(hidden_states)
hidden_states = residual + hidden_states
# Cross-Attention Block
cross_attn_weights = None
if encoder_hidden_states is not None:
residual = hidden_states
hidden_states = self.encoder_attn_layer_norm(hidden_states)
hidden_states, cross_attn_weights, _ = self.encoder_attn(
hidden_states=hidden_states,
causal_mask=causal_mask,
key_value_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
)
hidden_states = self.dropout_layer(hidden_states)
hidden_states = residual + hidden_states
# Fully Connected
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = self.activation_dropout_layer(hidden_states)
hidden_states = checkpoint_name(self.fc2(hidden_states), "mlp_down")
hidden_states = self.dropout_layer(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
outputs += (cache_view,)
return outputs
[docs]class WhisperEncoder(EasyDeLBaseModule):
"""The Whisper Encoder transformer stack.
This module processes the input audio features (log-Mel spectrogram) through
convolutional layers followed by a stack of `WhisperEncoderLayer` modules.
Attributes:
config (WhisperConfig): Configuration object for the model.
conv1 (nn.Conv): First convolutional layer.
conv2 (nn.Conv): Second convolutional layer.
embed_positions (nn.Embed): Positional embedding layer.
layers (nn.List[WhisperEncoderLayer]): List of encoder layers.
layer_norm (nn.LayerNorm): Final layer normalization.
embed_dim (int): Dimensionality of the model.
num_mel_bins (int): Number of Mel frequency bins in the input features.
padding_idx (int): Index of the padding token.
max_source_positions (int): Maximum sequence length for the encoder.
scale_embedding (float | None): Scaling factor for embeddings.
embed_scale (float | None): Alias for scale_embedding.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
"""Initializes the WhisperEncoder module.
Args:
config (WhisperConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.conv1 = nn.Conv(
self.config.d_model,
self.config.d_model,
kernel_size=(3,),
padding=1,
kernel_init=jax.nn.initializers.normal(self.config.init_std),
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.conv2 = nn.Conv(
self.config.d_model,
self.config.d_model,
kernel_size=(3,),
strides=2,
padding=1,
kernel_init=jax.nn.initializers.normal(self.config.init_std),
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.dropout_layer = nn.Dropout(
rate=self.config.dropout,
rngs=rngs,
)
block = WhisperEncoderLayer
self.layers = [
block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.encoder_layers)
]
self.embed_positions = nn.Embed(
self.config.max_source_positions,
self.config.d_model,
dtype=self.dtype,
embedding_init=sinusoidal_embedding_init,
param_dtype=self.param_dtype,
rngs=rngs,
)
self.layer_norm = nn.LayerNorm(
self.config.d_model,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
self.layerdrop = self.config.decoder_layerdrop
def __call__(
self,
input_features: jnp.ndarray,
output_attentions: bool = False,
output_hidden_states: bool = False,
) -> tuple[tp.Any | None, ...] | BaseModelOutput:
"""Forward pass of the Whisper encoder.
Args:
input_features (jnp.ndarray): Input audio features (log-Mel spectrogram)
of shape (batch_size, num_mel_bins, sequence_length).
output_attentions (bool): Whether to return attention weights (default: False).
output_hidden_states (bool): Whether to return hidden states for all layers (default: False).
Returns:
BaseModelOutput | tuple: The encoder output. returns a `BaseModelOutput`
containing `last_hidden_state`, `hidden_states` (optional), and `attentions` (optional).
"""
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
if input_features.shape[1:] != (
self.config.num_mel_bins,
self.config.max_source_positions * 2,
):
raise ValueError(
"input_features.shape[1:], must be equal to (self.config.num_mel_bins,"
f" self.config.max_source_positions * 2) (got {input_features.shape[1:]}, but should be"
f" ({self.config.num_mel_bins}, {self.config.max_source_positions * 2}))"
)
input_features = input_features.transpose(0, 2, 1)
hidden_states = jax.nn.gelu(self.conv1(input_features), approximate=False)
hidden_states = jax.nn.gelu(self.conv2(hidden_states), approximate=False)
embed_positions = self.embed_positions(jnp.arange(self.config.max_source_positions))
embed_positions = jax.lax.stop_gradient(embed_positions)
hidden_states = hidden_states + embed_positions
hidden_states = self.dropout_layer(hidden_states)
for encoder_layer in self.layers:
if output_hidden_states:
all_hidden_states = (*all_hidden_states, hidden_states)
dropout_probability = random.uniform(0, 1)
if not self.dropout_layer.deterministic and (dropout_probability < self.layerdrop):
# skip the layer
layer_outputs = (None, None)
else:
layer_outputs = encoder_layer(
hidden_states=hidden_states,
causal_mask=None,
mask_info=None,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = (*all_attentions, layer_outputs[1])
if output_hidden_states:
all_hidden_states += (hidden_states,)
hidden_states = self.layer_norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
)
[docs]class WhisperDecoder(EasyDeLBaseModule):
"""The Whisper Decoder transformer stack.
This module processes the target token IDs, incorporates positional embeddings,
and attends to both the input sequence (self-attention) and the encoder outputs
(cross-attention) through a stack of `WhisperDecoderLayer` modules.
Attributes:
config (WhisperConfig): Configuration object for the model.
embed_tokens (nn.Embed): Embedding layer for target tokens.
embed_positions (nn.Embed): Positional embedding layer.
layers (nn.List[WhisperDecoderLayer]): List of decoder layers.
layer_norm (nn.LayerNorm): Final layer normalization (applied to pre-final outputs).
dropout (nn.Dropout): Dropout layer.
padding_idx (int): Index of the padding token.
max_target_positions (int): Maximum sequence length for the decoder.
embed_scale (float | None): Scaling factor for embeddings.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
"""Initializes the WhisperDecoder module.
Args:
config (WhisperConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
embed_block = auto_remat(
nn.Embed,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.embed_tokens = embed_block(
self.config.vocab_size,
self.config.d_model,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.embed_positions = nn.Embed(
self.config.max_target_positions,
self.config.d_model,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
WhisperDecoderLayer(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.decoder_layers)
]
self.layerdrop = self.config.decoder_layerdrop
self.dropout_layer = nn.Dropout(
rate=self.config.dropout,
rngs=rngs,
)
self.layer_norm = nn.LayerNorm(
self.config.d_model,
param_dtype=self.param_dtype,
dtype=self.dtype,
epsilon=1e-05,
rngs=rngs,
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"],
mask_info: MaskInfo,
position_ids: Int[Array, "batch seq_len"],
encoder_hidden_states: Float[Array, "batch seq_len hidden_dim"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | None = None,
cache_metadata: TransformerMetadata | None = None,
output_attentions: bool = False,
output_hidden_states: bool = False,
) -> tuple[tp.Any, ...] | BaseModelOutputWithPastAndCrossAttentions:
"""Forward pass of the Whisper decoder.
Args:
input_ids (jnp.ndarray): Input token IDs (batch, target_sequence_length).
attention_mask (jnp.ndarray): Attention mask for self-attention
(batch, 1, target_sequence_length, target_sequence_length).
position_ids (jnp.ndarray): Position IDs (batch, target_sequence_length).
encoder_hidden_states (tp.Optional[jnp.ndarray]): Hidden states from the encoder
(batch, encoder_sequence_length, embed_dim).
past_key_values (tp.Optional[TransformerCache]): Cached key/value states for fast decoding.
cache_metadata (tp.Optional[TransformerMetadata]): Metadata for paged attention.
output_attentions (bool): Whether to return attention weights (default: False).
output_hidden_states (bool): Whether to return hidden states for all layers (default: False).
Returns:
BaseModelOutputWithPastAndCrossAttentions: The decoder output.
returns a `BaseModelOutputWithPastAndCrossAttentions` containing `last_hidden_state`,
`past_key_values` (if `use_cache` is True), `hidden_states` (optional), `attentions` (optional),
and `cross_attentions` (optional).
"""
inputs_embeds = self.embed_tokens(input_ids)
if position_ids is None:
position_ids = (
jnp.arange(inputs_embeds.shape[1])
.reshape(1, -1)
.repeat(
inputs_embeds.shape[0],
0,
)
)
position_ids = position_ids.astype("i4")
position_embeds = self.embed_positions(position_ids)
hidden_states = inputs_embeds + position_embeds
hidden_states = self.dropout_layer(hidden_states)
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
if mode is None:
mode = (
common_types.MODE_DECODE
if input_ids.shape[1] == 1 and past_key_values is not None
else common_types.MODE_TRAIN
)
if past_key_values is None:
past_key_values = TransformerCache.init_empty(len(self.layers))
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
for idx, decoder_layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
dropout_probability = random.uniform(0, 1)
if not self.dropout_layer.deterministic and (dropout_probability < self.layerdrop):
layer_outputs = (None, None, None, None)
else:
layer_outputs = decoder_layer(
hidden_states=hidden_states,
attention_mask=attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=None,
mode=mode,
cache_view=past_key_values[idx],
cache_metadata=cache_metadata,
output_attentions=output_attentions,
)
past_key_values[idx] = layer_outputs[-1]
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attns += (layer_outputs[1],)
if encoder_hidden_states is not None:
all_cross_attentions += (layer_outputs[2],)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
hidden_states = self.layer_norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attns,
cross_attentions=all_cross_attentions,
past_key_values=past_key_values,
)
[docs]@register_module(TaskType.BASE_MODULE, config=WhisperConfig, model_type="whisper")
class WhisperModel(EasyDeLBaseModule):
"""The base Whisper Model transformer implementing the encoder-decoder architecture.
Attributes:
config (WhisperConfig): Configuration object for the model.
encoder (WhisperEncoder): The encoder stack.
decoder (WhisperDecoder): The decoder stack.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
"""Initializes the WhisperModel module.
Args:
config (WhisperConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (tp.Optional[tp.Union[str, lax.Precision]]): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.encoder = WhisperEncoder(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.decoder = WhisperDecoder(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def _get_decoder_module(self):
"""Returns the decoder module."""
return self.decoder
def _get_encoder_module(self):
"""Returns the encoder module."""
return self.encoder
def __call__(
self,
input_features: jnp.ndarray,
decoder_input_ids: Int[Array, "batch seq_len"],
decoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
decoder_position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | None = None,
cache_metadata: TransformerMetadata | None = None,
output_attentions: bool = False,
output_hidden_states: bool = False,
):
"""Forward pass of the complete Whisper model (encoder + decoder).
Args:
input_features (jnp.ndarray): Input audio features (batch, num_mel_bins, seq_len).
decoder_input_ids (jnp.ndarray): Decoder input token IDs (batch, target_seq_len).
decoder_attention_mask (tp.Optional[jnp.ndarray]): Mask for decoder self-attention.
decoder_position_ids (tp.Optional[jnp.ndarray]): Position IDs for decoder inputs.
past_key_values (tp.Optional[TransformerCache]): Cached key/value states for fast decoding.
cache_metadata (tp.Optional[TransformerMetadata]): Metadata for paged attention.
output_attentions (bool): Whether to return attention weights (default: False).
output_hidden_states (bool): Whether to return hidden states for all layers (default: False).
Returns:
Seq2SeqModelOutput: The model output. returns a `Seq2SeqModelOutput`.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
batch_size, sequence_length = decoder_input_ids.shape
if decoder_attention_mask is None:
decoder_attention_mask = jnp.ones((batch_size, sequence_length))
if decoder_position_ids is None:
if past_key_values is not None:
raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
if decoder_attention_mask is not None:
decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
else:
decoder_position_ids = jnp.broadcast_to(
jnp.arange(sequence_length)[None, :],
(batch_size, sequence_length),
)
decoder_position_ids = decoder_position_ids.astype("i4")
encoder_outputs = self.encoder(
input_features=input_features,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
decoder_outputs = self.decoder(
input_ids=decoder_input_ids,
attention_mask=decoder_attention_mask,
position_ids=decoder_position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
encoder_hidden_states=encoder_outputs[0],
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
return Seq2SeqModelOutput(
last_hidden_state=decoder_outputs.last_hidden_state,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
past_key_values=decoder_outputs.past_key_values,
cross_attentions=decoder_outputs.cross_attentions,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
)
[docs] def decode(
self,
encoder_hidden_states: Float[Array, "batch seq_len hidden_dim"],
decoder_input_ids: Int[Array, "batch seq_len"],
decoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
decoder_position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | None = None,
cache_metadata: TransformerMetadata | None = None,
output_attentions: bool = False,
output_hidden_states: bool = False,
):
"""Performs decoding using the decoder module.
Args:
encoder_hidden_states (jnp.ndarray): Hidden states from the encoder.
decoder_input_ids (jnp.ndarray): Decoder input token IDs.
decoder_attention_mask (tp.Optional[jnp.ndarray]): Mask for decoder self-attention.
decoder_position_ids (tp.Optional[jnp.ndarray]): Position IDs for decoder inputs.
past_key_values (tp.Optional[TransformerCache]): Cached key/value states.
cache_metadata (tp.Optional[TransformerMetadata]): Metadata for paged attention.
output_attentions (bool): Whether to return attention weights.
output_hidden_states (bool): Whether to return hidden states for all layers.
Returns:
BaseModelOutputWithPastAndCrossAttentions | tuple: Decoder output.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
batch_size, sequence_length = decoder_input_ids.shape
if decoder_attention_mask is None:
decoder_attention_mask = jnp.ones((batch_size, sequence_length))
if decoder_position_ids is None:
if past_key_values is not None:
raise ValueError("Make sure to provide `decoder_position_ids` when passing `past_key_values`.")
if decoder_attention_mask is not None:
decoder_position_ids = (decoder_attention_mask.cumsum(-1) * decoder_attention_mask) - 1
else:
decoder_position_ids = jnp.broadcast_to(
jnp.arange(sequence_length)[None, :],
(batch_size, sequence_length),
)
decoder_outputs = self.decoder(
input_ids=decoder_input_ids,
attention_mask=decoder_attention_mask,
position_ids=decoder_position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
encoder_hidden_states=encoder_hidden_states,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
return Seq2SeqModelOutput(
last_hidden_state=decoder_outputs.last_hidden_state,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
past_key_values=decoder_outputs.past_key_values,
)
[docs] def encode(
self,
input_features: jnp.ndarray,
output_attentions: bool = False,
output_hidden_states: bool = False,
):
"""Performs encoding using the encoder module.
Args:
input_features (jnp.ndarray): Input audio features.
output_attentions (bool): Whether to return attention weights.
output_hidden_states (bool): Whether to return hidden states for all layers.
Returns:
BaseModelOutput | tuple: Encoder output.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
encoder_outputs = self.encoder(
input_features=input_features,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
return Seq2SeqModelOutput(
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
"""
return self.encoder
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
"""
return self.decoder
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
Base Models don't have a Language Model Head.
"""
raise NotImplementedError("The base model does not have a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the decoder.
"""
return self.decoder.embed_tokens
[docs]@register_module(TaskType.SPEECH_SEQUENCE_TO_SEQUENCE, config=WhisperConfig, model_type="whisper")
class WhisperForConditionalGeneration(EasyDeLBaseModule):
"""Whisper encoder-decoder with projection head for speech-to-text generation."""
loss_type = "ForCausalLM"
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = WhisperModel(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.proj_out = RowParallelLinear(
config.d_model,
config.vocab_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.init_std),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _get_encoder_module(self):
return self.model.encoder
def _get_decoder_module(self):
return self.model.decoder
def __call__(
self,
input_features,
decoder_input_ids,
decoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
decoder_position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | None = None,
cache_metadata: TransformerMetadata | None = None,
apply_lm_head: bool = True,
output_attentions: bool = False,
output_hidden_states: bool = False,
):
outputs = self.model(
input_features=input_features,
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
decoder_position_ids=decoder_position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
)
hidden_states = outputs[0]
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
lm_logits = None
if apply_lm_head:
lm_logits = self.apply_lm_head(hidden_states)
return Seq2SeqLMOutput(
logits=lm_logits,
decoder_hidden_states=outputs.decoder_hidden_states,
decoder_attentions=outputs.decoder_attentions,
cross_attentions=outputs.cross_attentions,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
encoder_hidden_states=outputs.encoder_hidden_states,
encoder_attentions=outputs.encoder_attentions,
past_key_values=outputs.past_key_values,
)
[docs] def decode(
self,
decoder_input_ids,
encoder_outputs,
encoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
decoder_attention_mask: Bool[Array, "batch seq_len"] | None = None,
decoder_position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | None = None,
cache_metadata: TransformerMetadata | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
encoder_hidden_states = encoder_outputs[0]
if decoder_attention_mask is not None:
decoder_attention_mask = decoder_attention_mask.astype("b1")
outputs = self.model.decode(
encoder_hidden_states=encoder_hidden_states,
decoder_attention_mask=decoder_attention_mask,
decoder_input_ids=decoder_input_ids,
decoder_position_ids=decoder_position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
)
hidden_states = outputs[0]
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
lm_logits = self.proj_out(
hidden_states,
self.model.decoder.embed_tokens.embedding.value.T.astype(self.param_dtype)
if self.config.tie_word_embeddings
else None,
)
return Seq2SeqLMOutput(
logits=lm_logits,
decoder_hidden_states=outputs.decoder_hidden_states,
decoder_attentions=outputs.decoder_attentions,
cross_attentions=outputs.cross_attentions,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
encoder_hidden_states=outputs.encoder_hidden_states,
encoder_attentions=outputs.encoder_attentions,
past_key_values=outputs.past_key_values,
)
[docs] def encode(
self,
input_features: jnp.ndarray,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
**kwargs,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return self.model.encode(
input_features=jnp.array(input_features, dtype="f4"),
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
[docs] def generate(
self,
input_features,
generation_config=None,
logits_processor=None,
return_timestamps=None,
task=None,
language=None,
is_multilingual=None,
**kwargs,
):
if generation_config is None:
generation_config = self.generation_config
if return_timestamps is not None:
generation_config.return_timestamps = return_timestamps
if task is not None:
generation_config.task = task
if is_multilingual is not None:
generation_config.is_multilingual = is_multilingual
if language is not None:
generation_config.language = language
if kwargs is not None and "decoder_input_ids" in kwargs:
decoder_input_length = len(kwargs["decoder_input_ids"])
else:
decoder_input_length = 1
forced_decoder_ids = []
if hasattr(generation_config, "is_multilingual") and generation_config.is_multilingual:
if hasattr(generation_config, "language"):
forced_decoder_ids.append((1, generation_config.lang_to_id[generation_config.language]))
else:
forced_decoder_ids.append((1, None))
if hasattr(generation_config, "task"):
forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
else:
forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))
if (
hasattr(generation_config, "return_timestamps") and generation_config.return_timestamps
) or return_timestamps:
logits_processor = [WhisperTimeStampLogitsProcessor(generation_config, self.config, decoder_input_length)]
else:
if forced_decoder_ids and forced_decoder_ids[-1][0] != generation_config.no_timestamps_token_id:
idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
if len(forced_decoder_ids) > 0:
generation_config.forced_decoder_ids = forced_decoder_ids
return super().generate(
input_features,
generation_config,
logits_processor=logits_processor,
**kwargs,
)
def _force_generate(
self,
input_features: jax.Array,
forced_decoder_ids: jax.Array,
return_timestamps: bool = False,
generation_config: tp.Optional["transformers.GenerationConfig"] = None, # type:ignore
**kwargs,
):
if generation_config is None:
generation_config = self.generation_config
generation_config.forced_decoder_ids = None
logits_processor = LogitsProcessorList()
logits_processor.append(ForceTokensLogitsProcessor(forced_decoder_ids))
if return_timestamps:
logits_processor.append(WhisperTimeStampLogitsProcessor(generation_config, self.config, 1))
return super().generate(
input_features,
generation_config,
logits_processor=logits_processor,
**kwargs,
)
[docs] def compute_loss(
self,
*,
labels: chex.Array | None = None,
loss_config: LossConfig | None = None,
loss_kwargs: dict | None = None,
**batch,
) -> tuple[tp.Any, LossMetrics]:
if loss_config is None:
loss_config = LossConfig()
loss_config.reduction = "mean"
loss_config.shift_tokens = False
if labels is not None:
if batch.get("decoder_input_ids", None) is None and batch.get("decoder_inputs_embeds", None) is None:
batch["decoder_input_ids"] = shift_tokens_right(
labels,
self.config.pad_token_id,
self.config.decoder_start_token_id,
)
return super().compute_loss(
labels=labels,
loss_config=loss_config,
loss_kwargs=loss_kwargs,
**batch,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
"""
return self.model.encoder
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
"""
return self.model.decoder
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
"""
return self.proj_out
[docs] def get_embedding(self):
"""
Returns the embedding layer of the decoder.
"""
return self.model.decoder.embed_tokens
[docs]@register_module(TaskType.AUDIO_CLASSIFICATION, config=WhisperConfig, model_type="whisper")
class WhisperForAudioClassification(EasyDeLBaseModule):
"""Encoder-only Whisper variant with pooling and classifier for audio tagging."""
def __init__(
self,
config: WhisperConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.encoder = WhisperEncoder(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
config.is_encoder_decoder = False
num_layers = config.num_hidden_layers + 1
if config.use_weighted_layer_sum:
self.layer_weights = jnp.repeat(1 / num_layers, num_layers)
self.projector = ColumnParallelLinear(
config.d_model,
config.classifier_proj_size,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.classifier = ColumnParallelLinear(
config.classifier_proj_size,
config.num_labels,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def __call__(
self,
input_features,
encoder_outputs=None,
output_attentions=None,
output_hidden_states: bool = True,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
if encoder_outputs is None:
encoder_outputs = self.encoder(
input_features,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
if self.config.use_weighted_layer_sum:
hidden_states = jnp.stack(encoder_outputs, axis=1)
norm_weights = jax.nn.softmax(self.layer_weights, axis=-1)
hidden_states = jnp.sum(hidden_states * jnp.reshape(norm_weights, [-1, 1, 1]), axis=1)
else:
hidden_states = encoder_outputs[0]
hidden_states = self.projector(hidden_states)
pooled_output = jnp.mean(hidden_states, axis=1)
logits = self.classifier(pooled_output)
return SequenceClassifierOutput(
logits=logits,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
"""
return self.encoder
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
This is an encoder-only model for classification.
"""
raise NotImplementedError("This is an encoder-only model and does not have a decoder.")
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
This model has an audio classification head, not a language model head.
"""
raise NotImplementedError("This model has an audio classification head, not a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
The encoder uses convolutional layers for feature extraction, not a standard token embedding.
Returning the first convolutional layer as the "embedding" layer.
"""
return self.encoder.conv1