# 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 functools
from typing import ClassVar
import jax
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 numpy as jnp
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
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_neox_configuration import GPTNeoXConfig as GPTNeoXConfig
[docs]class GPTNeoXAttention(UnifiedAttention):
"""GPT-NeoX Attention with partial RoPE.
Inherits from UnifiedAttention.
Uses combined QKV projection (query_key_value) and partial rotary embeddings.
Overrides forward_standard to efficiently handle fused QKV projection.
"""
projection_mapping: ClassVar[dict[str, str]] = {
"query_projection": "q_proj",
"key_projection": "k_proj",
"value_projection": "v_proj",
"output_projection": "dense",
"query_key_value_projection": "query_key_value",
# MLA-specific projections (DeepSeek V2/V3)
"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: GPTNeoXConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initialize GPT-NeoX attention."""
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
use_fused_qkv=True,
)
def _create_rotary(self, config: GPTNeoXConfig, dtype: jnp.dtype):
"""Create GPTNeoX-specific rotary embedding with partial RoPE."""
return config.get_basic_rope(
dtype=dtype,
head_size=self.head_dim,
rotary_dim=int(self.head_dim * config.rotary_pct), # Partial RoPE
base=config.rotary_emb_base,
)
def _create_attention_performer(self, config: GPTNeoXConfig, rngs: nn.Rngs):
"""Create attention performer with config dropout."""
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=config.attention_dropout,
)
[docs]class GPTNeoXMlp(nn.Module):
"""GPT-NeoX MLP module.
This module implements the feed-forward network used in the GPT-NeoX model.
Attributes:
config (GPTNeoXConfig): 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: GPTNeoXConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
) -> None:
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.dense_h_to_4h = ColumnParallelLinear(
self.config.hidden_size,
self.config.intermediate_size,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.dense_4h_to_h = RowParallelLinear(
self.config.intermediate_size,
self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.act = ACT2FN[self.config.hidden_act]
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""Forward pass of the GPTNeoXMlp module.
Args:
hidden_states (chex.Array): Input hidden states.
Returns:
chex.Array: Output hidden states after processing through the MLP.
"""
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
hidden_states = checkpoint_name(
self.dense_4h_to_h(self.act(checkpoint_name(self.dense_h_to_4h(hidden_states), name="mlp_up"))),
name="mlp_down",
)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return hidden_states
[docs]class GPTNeoXBlock(nn.Module):
"""GPT-NeoX Transformer block.
This module represents a single transformer block in the GPT-NeoX model,
containing self-attention and MLP sub-layers with residual connections
and layer normalization. It supports both standard and parallel residual connections.
Attributes:
config (GPTNeoXConfig): 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: GPTNeoXConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
) -> None:
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.use_parallel_residual = config.use_parallel_residual
attn_block = GPTNeoXAttention
mlp_block = GPTNeoXMlp
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.input_layernorm = nn.LayerNorm(
config.hidden_size,
epsilon=config.layer_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = nn.LayerNorm(
config.hidden_size,
epsilon=config.layer_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.attention = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.mlp = GPTNeoXMlp(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
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 GPTNeoXBlock 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.
"""
attn_outputs = self.attention(
self.input_layernorm(hidden_states),
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
if self.use_parallel_residual:
mlp = self.mlp(self.post_attention_layernorm(hidden_states))
hidden_states = mlp + hidden_states + attn_outputs.attention_output
else:
hidden_states = attn_outputs.attention_output + hidden_states
hidden_states = self.mlp(self.post_attention_layernorm(hidden_states)) + 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=GPTNeoXConfig, model_type="gpt_neox")
class GPTNeoXModel(EasyDeLBaseModule):
"""GPT-NeoX model implementation.
This class implements the main GPT-NeoX transformer model architecture, consisting of
an embedding layer, multiple GPTNeoXBlock layers, and a final layer normalization.
Attributes:
config (GPTNeoXConfig): 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: GPTNeoXConfig,
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_in = nn.Embed(
self.config.vocab_size,
self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.emb_dropout = nn.Dropout(config.hidden_dropout, rngs=rngs)
self.layers = [
GPTNeoXBlock(
config=config,
layer_idx=i,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(config.num_hidden_layers)
]
self.final_layer_norm = nn.LayerNorm(
config.hidden_size,
epsilon=self.config.layer_norm_eps,
dtype=self.dtype,
param_dtype=param_dtype,
rngs=rngs,
)
@functools.cached_property
def frequencies(self):
head_dim = self.config.hidden_size // self.config.num_attention_heads
return self.config.get_basic_frequencies(
head_size=head_dim,
rotary_dim=int(head_dim * self.config.rotary_pct),
base=self.config.rotary_emb_base,
)
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 GPTNeoXModel.
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 = self.embed_in(input_ids.astype("i4"))
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 = self.emb_dropout(
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.layers))
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
for idx, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = block(
hidden_states=hidden_states,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
cache_view=past_key_values.views[idx],
cache_metadata=cache_metadata,
frequencies=self.frequencies,
output_attentions=output_attentions,
)
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 = self.final_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,
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.embed_in
[docs]@register_module(TaskType.CAUSAL_LM, config=GPTNeoXConfig, model_type="gpt_neox")
class GPTNeoXForCausalLM(BaseCausalLMModule[GPTNeoXModel, GPTNeoXConfig]):
"""GPT-NeoX model with a language modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "gpt_neox"
_config_class = GPTNeoXConfig
def __init__(
self,
config: GPTNeoXConfig,
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=GPTNeoXModel,
base_model_name="gpt_neox",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
lm_head_name="embed_out",
)