# 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.
from functools import cached_property
from typing import ClassVar
import jax
import jax.numpy as jnp
from eformer import common_types
from eformer.escale import apply_logical_sharding
from eformer.loggings import get_logger
from ejkernel.types import MaskInfo
from flax import nnx as nn
from jax.ad_checkpoint import checkpoint_name
from jaxtyping import Array, Bool, Float, Int
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.modeling_outputs import BaseModelOutput, DecoderLayerOutput
from easydel.infra.utils import ACT2FN, auto_remat, block_wise_ffn, get_dot_general_by_bits
from easydel.layers.attention import FlexibleAttentionModule
from easydel.layers.attention_unified import UnifiedAttention
from easydel.layers.base_modules import BaseCausalLMModule
from easydel.layers.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
RaggedPagesMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from .gpt_j_configuration import GPTJConfig as GPTJConfig
logger = get_logger(__name__)
[docs]class GPTJAttention(UnifiedAttention):
"""GPT-J Attention with partial RoPE.
Inherits from UnifiedAttention.
Uses separate Q/K/V projections with partial rotary embeddings.
"""
projection_mapping: ClassVar[dict[str, str]] = {
"query_projection": "q_proj",
"key_projection": "k_proj",
"value_projection": "v_proj",
"output_projection": "out_proj",
"qkv_projection": "qkv_proj",
"mla_q_proj": "q_proj",
"mla_q_a_proj": "q_a_proj",
"mla_q_a_layernorm": "q_a_layernorm",
"mla_q_b_proj": "q_b_proj",
"mla_kv_a_proj_with_mqa": "kv_a_proj_with_mqa",
"mla_kv_a_layernorm": "kv_a_layernorm",
"mla_kv_b_proj": "kv_b_proj",
}
def __init__(
self,
config: GPTJConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initialize GPT-J attention."""
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
)
def _create_rotary(self, config: GPTJConfig, dtype: jnp.dtype):
"""Create GPT-J-specific rotary embedding with partial RoPE."""
return config.get_basic_rope(
dtype,
head_size=self.head_dim,
rotary_dim=config.rotary_dim, # Partial RoPE
base=10000,
is_neox_style=False,
)
def _create_attention_performer(self, config: GPTJConfig, rngs: nn.Rngs):
"""Create attention performer with config dropout."""
return FlexibleAttentionModule(
rngs=rngs,
dropout_prob=config.attn_pdrop,
base_config=config,
softmax_scale=self.head_dim**-0.5,
)
def _create_q_proj(self, config, dtype, param_dtype, precision, rngs):
"""Create query projection with checkpointing."""
return ColumnParallelLinear(
config.hidden_size,
config.num_attention_heads * self.head_dim,
use_bias=False,
dtype=dtype,
kernel_init=nn.initializers.normal(config.initializer_range),
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_k_proj(self, config, dtype, param_dtype, precision, rngs):
"""Create key projection."""
return ColumnParallelLinear(
config.hidden_size,
self.num_key_value_heads * self.head_dim,
use_bias=False,
dtype=dtype,
kernel_init=nn.initializers.normal(config.initializer_range),
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_v_proj(self, config, dtype, param_dtype, precision, rngs):
"""Create value projection."""
return ColumnParallelLinear(
config.hidden_size,
self.num_key_value_heads * self.head_dim,
use_bias=False,
dtype=dtype,
kernel_init=nn.initializers.normal(config.initializer_range),
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_o_proj(self, config, dtype, param_dtype, precision, rngs):
"""Create output projection (named out_proj for GPT-J)."""
self.out_proj = ColumnParallelLinear(
config.num_attention_heads * self.head_dim,
config.hidden_size,
use_bias=False,
dtype=dtype,
kernel_init=nn.initializers.normal(config.initializer_range),
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
return self.out_proj
[docs] def define_network(
self,
config: GPTJConfig,
dtype: jnp.dtype,
param_dtype: jnp.dtype,
precision: jax.lax.PrecisionLike,
rngs: nn.Rngs,
):
"""Define GPT-J-specific network with residual dropout."""
# Call parent to create standard Q/K/V/O projections
super().define_network(config, dtype, param_dtype, precision, rngs)
# GPT-J has residual dropout
self.resid_dropout = nn.Dropout(rate=config.resid_pdrop, rngs=rngs)
def _split_heads(self, hidden_states):
"""Split hidden states into attention heads."""
return hidden_states.reshape((*hidden_states.shape[:2], self.config.num_attention_heads, self.head_dim))
def _get_query_proj(self, hidden_states: Array) -> Array:
"""Apply query projection with checkpoint naming and head splitting."""
query_states = checkpoint_name(self.q_proj(hidden_states), "attn_query")
return self._split_heads(query_states)
def _get_key_proj(self, hidden_states: Array) -> Array:
"""Apply key projection with checkpoint naming and head splitting."""
key_states = checkpoint_name(self.k_proj(hidden_states), "attn_key")
return self._split_heads(key_states)
def _get_value_proj(self, hidden_states: Array) -> Array:
"""Apply value projection with checkpoint naming and head splitting."""
value_states = checkpoint_name(self.v_proj(hidden_states), "attn_value")
return self._split_heads(value_states)
def _get_output_proj(self, attn_output: Array) -> Array:
"""Apply output projection with checkpoint naming and residual dropout."""
attn_output = checkpoint_name(self.out_proj(attn_output), "attn_output")
return self.resid_dropout(attn_output)
[docs]class GPTJMLP(nn.Module):
"""GPT-J MLP module.
This module implements the feed-forward network used in the GPT-J model.
Attributes:
config (GPTJConfig): Configuration object for the model.
intermediate_size (int): Dimensionality of the intermediate layer.
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: GPTJConfig,
intermediate_size: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
self.config: GPTJConfig = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.intermediate_size = intermediate_size
embed_dim = config.hidden_size
kernel_init = nn.initializers.normal(config.initializer_range)
self.fc_in = ColumnParallelLinear(
embed_dim,
intermediate_size,
dtype=dtype,
param_dtype=dtype,
precision=precision,
kernel_init=kernel_init,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.fc_out = RowParallelLinear(
intermediate_size,
embed_dim,
dtype=dtype,
param_dtype=dtype,
precision=precision,
kernel_init=kernel_init,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.act = ACT2FN[config.activation_function]
self.dropout = nn.Dropout(rate=config.resid_pdrop)
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""Forward pass of the GPTJMLP module.
Args:
hidden_states (chex.Array): Input hidden states.
Returns:
chex.Array: Output hidden states after processing through the MLP.
"""
gate = checkpoint_name(self.act(self.fc_in(hidden_states)), "mlp_gate")
hidden_states = checkpoint_name(self.dropout(self.fc_out(gate)), "mlp_output")
return hidden_states
[docs]class GPTJBlock(nn.Module):
"""GPT-J Transformer block.
This module represents a single transformer block in the GPT-J model,
containing self-attention and MLP sub-layers with residual connections
and layer normalization.
Attributes:
config (GPTJConfig): Configuration object for the model.
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: GPTJConfig,
layer_idx: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
self.config: GPTJConfig = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
hidden_size = self.config.hidden_size
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
attn_block = GPTJAttention
mlp_block = GPTJMLP
attn_block, mlp_block = auto_remat(
attn_block,
mlp_block,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.ln_1 = nn.LayerNorm(
self.config.hidden_size,
epsilon=config.layer_norm_epsilon,
dtype=dtype,
param_dtype=dtype,
rngs=rngs,
)
self.attn = attn_block(
config,
layer_idx=layer_idx,
dtype=dtype,
param_dtype=dtype,
precision=precision,
rngs=rngs,
)
self.mlp = mlp_block(
config,
inner_dim,
dtype=dtype,
param_dtype=dtype,
precision=precision,
rngs=rngs,
)
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
mask_info: MaskInfo | None,
position_ids: Int[Array, "batch seq_len"],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: TransformerCacheView | RaggedPagesCacheView | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
output_attentions: bool = False,
frequencies: Float[Array, "seq_len head_dim"] | None = None,
) -> DecoderLayerOutput:
"""Forward pass of the GPTJBlock 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, optional): Causal mask for ensuring autoregressive behavior.
segment_ids (tp.Optional[chex.Array], optional): Segment IDs for segment-based attention.
cache_view (tp.Optional[TransformerCacheView | RaggedPagesCacheView], optional): Cache view for
key_states/value_states states.
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata], optional): Metadata for
cache handling.
output_attentions (bool, optional): Whether to return attention weights.
frequencies (tp.Optional[chex.Array], optional): Precomputed rotary frequencies.
Returns:
tp.Tuple[chex.Array, tp.Optional[chex.Array]]: A tuple containing the output hidden states and
optionally the attention weights.
"""
residual = hidden_states
hidden_states = self.ln_1(hidden_states)
attn_outputs = self.attn(
hidden_states,
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
attn_output = attn_outputs.attention_output
if self.config.use_scan_mlp:
feed_forward_hidden_states = block_wise_ffn(
self.mlp,
hidden_states,
self.config.scan_mlp_chunk_size,
)
else:
feed_forward_hidden_states = self.mlp(hidden_states)
hidden_states = checkpoint_name(attn_output + feed_forward_hidden_states + residual, "residual")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return DecoderLayerOutput(
hidden_states=checkpoint_name(hidden_states, "layer_output"),
attention_weight=attn_outputs.attention_weight,
cache_view=attn_outputs.cache_view,
)
[docs]@register_module(TaskType.BASE_MODULE, config=GPTJConfig, model_type="gptj")
class GPTJModel(EasyDeLBaseModule):
"""GPT-J model implementation.
This class implements the main GPT-J transformer model architecture, consisting of
an embedding layer, multiple GPTJBlock layers, and a final layer normalization.
Attributes:
config (GPTJConfig): Configuration object for the model.
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: GPTJConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.embed_dim = config.hidden_size
self.wte = nn.Embed(
self.config.vocab_size,
self.embed_dim,
embedding_init=nn.initializers.normal(stddev=config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.dropout = nn.Dropout(
rate=self.config.embd_pdrop,
rngs=rngs,
)
self.h = [
GPTJBlock(
config,
layer_idx=i,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.num_hidden_layers)
]
self.ln_f = nn.LayerNorm(
self.config.hidden_size,
epsilon=self.config.layer_norm_epsilon,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
@cached_property
def frequencies(self):
embed_dim = self.config.hidden_size
num_heads = self.config.num_attention_heads
head_dim = embed_dim // num_heads
rotary_dim = self.config.rotary_dim
return self.config.get_basic_frequencies(
rotary_dim=rotary_dim,
head_size=head_dim,
base=10000,
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"] | None = None,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | RaggedPagesCache | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None,
extra_embedding: Float[Array, "batch seq_len hidden_dim"] | None = None,
output_attentions: bool = False,
output_hidden_states: bool = False,
):
"""Forward pass through the GPTJModel.
Args:
input_ids (chex.Array, optional): Input token IDs, shape (batch_size, sequence_length).
attention_mask (chex.Array, optional): Mask to avoid attention on padding tokens.
position_ids (chex.Array, optional): Indices of positions of each input sequence token.
past_key_values (TransformerCache | RaggedPagesCache, optional): Cache containing precomputed
key_states/value_states states.
cache_metadata (TransformerMetadata | RaggedPagesMetadata, optional): Metadata for cache handling.
inputs_embeds (chex.Array, optional): Input embeddings, shape (batch_size, sequence_length, hidden_size).
segment_ids (chex.Array, optional): Segment token indices for segment embeddings.
extra_embedding (chex.Array, optional): Additional embedding to add to input embeddings.
output_attentions (bool, optional): Whether to return attention weights.
output_hidden_states (bool, optional): Whether to return hidden states of all layers.
Returns:
Union[BaseModelOutput, Tuple]: Model outputs (last hidden state, optional hidden states, optional attentions)
"""
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
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 = checkpoint_name(self.wte(input_ids.astype("i4")), "embeddings")
sequence_length = inputs_embeds.shape[1]
mask_info = MaskInfo.dynamic_init(
mask_info=mask_info,
input_ids=input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
)
if position_ids is None:
position_ids = mask_info.q_position_ids
assert sequence_length <= self.config.max_position_embeddings, (
f"Maximum Position Embedding Reached ! "
f"(Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
hidden_states = inputs_embeds + extra_embedding if extra_embedding is not None else 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.h))
hidden_states = self.dropout(inputs_embeds)
for idx, block in enumerate(self.h):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = block(
hidden_states=hidden_states,
mask_info=mask_info,
mode=mode,
cache_view=past_key_values.views[idx],
cache_metadata=cache_metadata,
position_ids=position_ids,
output_attentions=output_attentions,
frequencies=self.frequencies,
)
hidden_states = layer_outputs.hidden_states
if output_attentions:
all_attentions += (layer_outputs.attention_weight,)
past_key_values[idx] = layer_outputs.cache_view
hidden_states = checkpoint_name(self.ln_f(hidden_states), "model_output")
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] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
Decoder-Only models don't have an encoder.
"""
raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
"""
return self
[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 module.
"""
return self.wte
[docs]@register_module(TaskType.CAUSAL_LM, config=GPTJConfig, model_type="gptj")
class GPTJForCausalLM(BaseCausalLMModule[GPTJModel, GPTJConfig]):
"""GPT-J model with a language modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "gptj"
_config_class = GPTJConfig
def __init__(
self,
config: GPTJConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
base_model_class=GPTJModel,
base_model_name="transformer",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
)