# Copyright 2023 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 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 flax import nnx as nn
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.modeling_outputs import (
AttentionLayerOutput,
BaseModelOutput,
CausalLMOutput,
DecoderLayerOutput,
SequenceClassifierOutput,
)
from easydel.infra.utils import (
ACT2FN,
auto_remat,
block_wise_ffn,
get_dot_general_by_bits,
)
from easydel.layers.attention import AttentionModule, FlexibleAttentionModule
from easydel.layers.caching import (
PagedAttentionCache,
PagedAttentionCacheView,
PagedAttentionMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ParallelLinear
from easydel.utils.helpers import get_logger
from .gemma2_configuration import Gemma2Config
logger = get_logger(__name__)
[docs]class Gemma2RMSNorm(nn.Module):
def __init__(self, config: Gemma2Config, dtype: jnp.dtype = jnp.float32):
self.config = config
self.epsilon = self.config.rms_norm_eps
self.dtype = dtype
self.kernel = nn.Param(jnp.ones(self.config.hidden_size, dtype=dtype))
def __call__(self, hidden_states):
variance = hidden_states.astype(jnp.float32)
variance = jnp.power(variance, 2)
variance = variance.mean(-1, keepdims=True)
hidden_states = hidden_states / jnp.sqrt(variance + self.epsilon)
return (1 + self.kernel.value.astype(self.dtype)) * jnp.asarray(
hidden_states, dtype=self.dtype
)
[docs]class Gemma2Attention(AttentionModule):
def __init__(
self,
config: Gemma2Config,
layer_idx: int,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
causal: bool = True,
is_cross_attention: bool = False,
*,
rngs: nn.Rngs,
):
super().__init__(config)
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.is_cross_attention = is_cross_attention
self.rngs = rngs
self.causal = causal
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = config.head_dim
self.attention_softmax_in_fp32 = self.dtype is not jnp.float32
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
kernel = jax.nn.initializers.normal(config.initializer_range)
linear = partial(
ParallelLinear,
use_bias=config.attention_bias,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=kernel,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.q_proj = linear(
self.embed_dim,
self.num_heads * self.head_dim,
rngs=rngs,
)
self.k_proj = linear(
self.embed_dim,
self.num_key_value_heads * self.head_dim,
rngs=rngs,
)
self.v_proj = linear(
self.embed_dim,
self.num_key_value_heads * self.head_dim,
rngs=rngs,
)
self.o_proj = linear(
self.num_heads * self.head_dim,
self.embed_dim,
rngs=rngs,
)
self.sliding_window = config.sliding_window if (self.layer_idx % 2 == 0) else None
self.attention_performer = FlexibleAttentionModule(
base_config=config,
softmax_scale=self.config.query_pre_attn_scalar**-0.5,
dropout_prob=config.attention_dropout,
)
self.rotary = self.config.get_basic_rope(
self.dtype,
self.head_dim,
self.head_dim,
True,
)
def _merge_heads(self, hidden_states):
"""
Merges the attention heads into a single hidden state tensor.
Args:
hidden_states (chex.Array): The hidden states with separate head dimensions.
Returns:
chex.Array: The hidden states with merged head dimensions.
"""
return hidden_states.reshape(
hidden_states.shape[:2] + (self.num_heads * self.head_dim,)
)
def _split_heads(self, hidden_states, num_heads):
return hidden_states.reshape(hidden_states.shape[:2] + (num_heads, self.head_dim))
def __call__(
self,
hidden_states: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: tp.Optional[chex.Array | bool],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: tp.Optional[TransformerCacheView | PagedAttentionCacheView] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
frequencies: tp.Optional[chex.Array] = None,
) -> AttentionLayerOutput:
"""
Forward pass of the attention module.
Args:
hidden_states (chex.Array): Input hidden states.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
causal_mask (chex.Array): Causal mask for ensuring autoregressive behavior.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
deterministic (bool): If True, disables dropout for deterministic behavior.
init_cache (bool): If True, initializes cache for caching keys and values.
output_attentions (bool): If True, outputs attention weights alongside the hidden states.
fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask.
Returns:
tp.Tuple[chex.Array, chex.Array]: A tuple containing the attention output and the attention weights.
"""
batch_size, sequence_length = hidden_states.shape[:2]
(query_states, key_states, value_states) = (
self.q_proj(hidden_states),
self.k_proj(hidden_states),
self.v_proj(hidden_states),
)
query_states = query_states.reshape(
batch_size,
sequence_length,
self.num_heads,
self.head_dim,
)
key_states = key_states.reshape(
batch_size,
sequence_length,
self.num_key_value_heads,
self.head_dim,
)
value_states = value_states.reshape(
batch_size,
sequence_length,
self.num_key_value_heads,
self.head_dim,
)
(
query_states,
key_states,
value_states,
) = self.apply_qkv_shardings(query_states, key_states, value_states)
query_states, key_states = self.rotary(
positions=position_ids,
query=query_states,
key=key_states,
frequencies=frequencies,
)
(
key_states,
value_states,
attention_mask,
init_attention_bias,
cache_view,
) = self.concatenate(
query=query_states,
key=key_states,
value=value_states,
cache_view=cache_view,
cache_metadata=cache_metadata,
attention_mask=attention_mask,
causal_mask=causal_mask,
fcm_mask=fcm_mask,
sliding_window=self.sliding_window if bool((self.layer_idx % 2) == 0) else None,
)
attentions = self.attention_performer.forward(
query_states=query_states,
key_states=key_states,
value_states=value_states,
mode=mode,
bias=None,
sliding_window=self.sliding_window if bool((self.layer_idx % 2) == 0) else None,
cache_metadata=cache_metadata,
cache_view=cache_view,
init_bias=init_attention_bias,
attention_mask=attention_mask,
segment_ids=segment_ids,
causal=True,
dropout_rng=self.rngs.params(),
)
attn_output = self.shard_attention_prod(
self._merge_heads(attentions.attention_outputs)
)
attn_output = self.o_proj(attn_output)
return AttentionLayerOutput(
attention_output=attn_output,
attention_weight=attentions.attention_weights if output_attentions else None,
cache_view=cache_view,
)
[docs]class Gemma2MLP(nn.Module):
def __init__(
self,
config: Gemma2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
embed_dim = self.config.hidden_size
inner_dim = (
self.config.intermediate_size
if self.config.intermediate_size is not None
else 4 * embed_dim
)
kernel_init = jax.nn.initializers.normal(config.initializer_range)
self.act = ACT2FN[self.config.hidden_activation]
linear_class = partial(
ParallelLinear,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=kernel_init,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.gate_proj = linear_class(
embed_dim,
inner_dim,
rngs=rngs,
)
self.down_proj = linear_class(
inner_dim,
embed_dim,
rngs=rngs,
)
self.up_proj = linear_class(
embed_dim,
inner_dim,
rngs=rngs,
)
def __call__(self, hidden_states):
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
gate = self.act(self.gate_proj(hidden_states))
up = self.up_proj(hidden_states)
hidden_states = self.down_proj(gate * up)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return hidden_states
[docs]class Gemma2DecoderLayer(nn.Module):
def __init__(
self,
config: Gemma2Config,
layer_idx: int,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
mlp_block = Gemma2MLP
attn_block = Gemma2Attention
attn_block, mlp_block = auto_remat(
attn_block,
mlp_block,
policy=config.gradient_checkpointing,
)
self.is_sliding = bool(self.layer_idx % 2)
self.self_attn = attn_block(
self.config,
layer_idx=self.layer_idx,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.mlp = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.input_layernorm = Gemma2RMSNorm(self.config, dtype=self.dtype)
self.post_attention_layernorm = Gemma2RMSNorm(self.config, dtype=self.dtype)
self.pre_feedforward_layernorm = Gemma2RMSNorm(self.config, dtype=self.dtype)
self.post_feedforward_layernorm = Gemma2RMSNorm(self.config, dtype=self.dtype)
def __call__(
self,
hidden_states: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: tp.Optional[chex.Array | bool],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: tp.Optional[TransformerCacheView | PagedAttentionCacheView] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
frequencies: tp.Optional[chex.Array] = None,
):
"""
Forward pass of the module block.
Args:
hidden_states (chex.Array): Input hidden states.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
causal_mask (chex.Array): Causal mask for ensuring autoregressive behavior.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
deterministic (bool): If True, disables dropout for deterministic behavior.
init_cache (bool): If True, initializes cache for caching keys and values.
output_attentions (bool): If True, outputs attention weights alongside the hidden states.
fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask.
Returns:
tp.Tuple[chex.Array, chex.Array]: A tuple containing the attention output and the attention weights.
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
attn_outputs = self.self_attn(
hidden_states,
attention_mask,
position_ids,
causal_mask,
mode,
cache_view,
cache_metadata,
segment_ids,
output_attentions,
fcm_mask,
frequencies,
)
hidden_states = self.post_attention_layernorm(attn_outputs.attention_output)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.pre_feedforward_layernorm(hidden_states)
if self.config.use_scan_mlp:
hidden_states = block_wise_ffn(
self.mlp,
hidden_states,
self.config.scan_mlp_chunk_size,
)
else:
hidden_states = self.mlp(hidden_states)
hidden_states = self.post_feedforward_layernorm(hidden_states)
hidden_states = residual + hidden_states
return DecoderLayerOutput(
hidden_states=hidden_states,
attention_weight=attn_outputs.attention_weight,
cache_view=attn_outputs.cache_view,
)
[docs]@register_module(
TaskType.BASE_MODULE,
config=Gemma2Config,
model_type="gemma2",
)
class Gemma2Model(EasyDeLBaseModule):
def __init__(
self,
config: Gemma2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.hidden_size = self.config.hidden_size
self.embed_tokens = nn.Embed(
self.config.vocab_size,
self.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
Gemma2DecoderLayer(
self.config,
layer_idx=i,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.num_hidden_layers)
]
self.norm = Gemma2RMSNorm(self.config, dtype=self.dtype)
def __call__(
self,
input_ids: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
mode: tp.Optional[common_types.RUNTIME_MODE_TYPES] = None, # type:ignore
past_key_values: tp.Optional[TransformerCache | PagedAttentionCache] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
) -> BaseModelOutput:
"""
Forward pass through the Gemma2 module.
Args:
input_ids (chex.Array): Input tensor containing token IDs.
attention_mask (chex.Array): Mask for attention.
position_ids (chex.Array): Positional indices.
segment_ids (tp.Optional[chex.Array]): Segment IDs for different input parts.
inputs_embeds (tp.Optional[chex.Array]): Embedded input tensor.
output_attentions (tp.Optional[bool]): If True, output attention weights.
output_hidden_states (tp.Optional[bool]): If True, output hidden states.
init_cache (bool): If True, initialize cache for decoding.
deterministic (bool): If True, disable dropout.
Returns:
BaseModelOutput | tp.Tuple: Model output, either as a named tuple or a standard tuple.
"""
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError(
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
)
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids.astype("i4"))
batch_size, sequence_length, _ = inputs_embeds.shape
if attention_mask is None:
attention_mask = jnp.ones((batch_size, sequence_length), "b1")
else:
if attention_mask.dtype != jnp.bool:
attention_mask = jnp.astype(attention_mask == 1, "b1")
if position_ids is None:
position_ids = jnp.broadcast_to(
jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0),
(batch_size, sequence_length),
)
inputs_embeds = inputs_embeds * (self.config.hidden_size**0.5)
assert sequence_length <= self.config.max_position_embeddings, (
f"Maximum Position Embedding Reached ! (Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
if attention_mask.ndim == 2:
attention_mask = jnp.expand_dims(attention_mask, (1, 2))
hidden_states = inputs_embeds
if mode is None:
mode = (
common_types.MODE_DECODE
if sequence_length == 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,
)
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
for idx, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
# Type hint for clarity, though not strictly necessary for execution
layer_outputs: DecoderLayerOutput = block(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
causal_mask=self.causal_mask,
mode=mode,
cache_view=past_key_values.views[idx],
cache_metadata=cache_metadata,
output_attentions=output_attentions,
segment_ids=segment_ids,
frequencies=self.frequencies,
)
hidden_states = layer_outputs.hidden_states
if output_attentions:
all_attentions += (layer_outputs.attention_weight,)
hidden_states = self.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,
past_key_values=past_key_values,
)
[docs]@register_module(
TaskType.CAUSAL_LM,
config=Gemma2Config,
model_type="gemma2",
)
class Gemma2ForCausalLM(EasyDeLBaseModule):
"""Gemma2 model with a language modeling head for causal language modeling tasks.
This model extends the base Gemma2Model by incorporating a linear language modeling head on top
of the base model, designed for generative tasks and text generation.
"""
def __init__(
self,
config: Gemma2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initialize a Gemma2ForCausalLM model.
Args:
config (Gemma2Config): Configuration object for the model.
dtype (jnp.dtype, optional): Data type for activations and weights. Defaults to jnp.float32.
param_dtype (jnp.dtype, optional): Data type for parameters. Defaults to jnp.float32.
precision (jax.lax.PrecisionLike, optional): Numerical precision for computations. Defaults to None.
rngs (nn.Rngs): Random number generator keys for initialization.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = Gemma2Model(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.lm_head = ParallelLinear(
config.hidden_size,
config.vocab_size,
use_bias=False,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(stddev=config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def __call__(
self,
input_ids: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
mode: tp.Optional[common_types.RUNTIME_MODE_TYPES] = None, # type:ignore
past_key_values: tp.Optional[TransformerCache | PagedAttentionCache] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
) -> CausalLMOutput:
"""Forward pass of the causal language model.
Args:
input_ids (Optional[chex.Array], optional): Input token IDs. Defaults to None.
inputs_embeds (Optional[chex.Array], optional): Pre-computed input embeddings. Defaults to None.
attention_mask (Optional[chex.Array], optional): Mask to avoid attention on padding tokens. Defaults to None.
position_ids (Optional[chex.Array], optional): Position IDs for positional embeddings. Defaults to None.
segment_ids (Optional[chex.Array], optional): Segment IDs for segment embeddings. Defaults to None.
output_attentions (Optional[bool], optional): Whether to return attention weights. Defaults to None.
output_hidden_states (Optional[bool], optional): Whether to return hidden states. Defaults to None.
past_key_values (Optional[TransformerCache | PagedAttentionCache], optional): Cached key values for faster inference. Defaults to None.
cache_metadata (Optional[TransformerMetadata | PagedAttentionMetadata], optional): Metadata for cache handling. Defaults to None.
Returns:
Union[CausalLMOutput, Tuple]: Model outputs containing logits and optional hidden states and attentions.
"""
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
inputs_embeds=inputs_embeds,
segment_ids=segment_ids,
)
hidden_states = outputs.last_hidden_state
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
if self.config.tie_word_embeddings:
lm_logits = jax.lax.dot_general(
hidden_states,
self.model.embed_tokens.embedding.value.T,
(((hidden_states.ndim - 1), (0,)), ((), ())),
)
else:
lm_logits = self.lm_head(hidden_states)
if self.config.final_logit_softcapping is not None:
cap = jnp.array(self.config.final_logit_softcapping, dtype=lm_logits.dtype)
lm_logits = cap * jax.nn.tanh(lm_logits / cap)
return CausalLMOutput(
logits=lm_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
past_key_values=outputs.past_key_values,
)
[docs]@register_module(
TaskType.SEQUENCE_CLASSIFICATION,
config=Gemma2Config,
model_type="gemma2",
)
class Gemma2ForSequenceClassification(EasyDeLBaseModule):
def __init__(
self,
config: Gemma2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = Gemma2Model(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
assert hasattr(config, "num_labels"), (
"in order to use `SequenceClassification` Models in `EasyDeL` you first need to attach `num_labels` to model `config`"
)
self.score = ParallelLinear(
self.config.hidden_size,
config.num_labels,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(stddev=config.initializer_range),
precision=self.precision,
rngs=rngs,
)
def __call__(
self,
input_ids: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
mode: tp.Optional[common_types.RUNTIME_MODE_TYPES] = None, # type:ignore
past_key_values: tp.Optional[TransformerCache | PagedAttentionCache] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
) -> SequenceClassifierOutput:
transformer_outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
inputs_embeds=inputs_embeds,
segment_ids=segment_ids,
)
hidden_states = transformer_outputs.last_hidden_state
logits = self.score(hidden_states)
if input_ids is not None:
batch_size = input_ids.shape[0]
else:
batch_size = inputs_embeds.shape[0]
if self.config.pad_token_id is None and batch_size != 1:
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
sequence_lengths = (
jnp.argmax(jnp.equal(input_ids, self.config.pad_token_id).astype("i4"), -1)
- 1
)
sequence_lengths = sequence_lengths % input_ids.shape[-1]
else:
sequence_lengths = -1
pooled_logits = logits[jnp.arange(batch_size), sequence_lengths]
return SequenceClassifierOutput(
logits=pooled_logits,
past_key_values=past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)