# 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 eformer.loggings import get_logger
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, block_wise_ffn, get_dot_general_by_bits
from easydel.layers.attention_unified import UnifiedAttention
from easydel.layers.base_modules import BaseCausalLMModule, BaseSequenceClassificationModule
from easydel.layers.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
RaggedPagesMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear
from easydel.layers.norms import RMSNorm
from .exaone_configuration import ExaoneConfig
logger = get_logger(__name__)
[docs]class ExaoneGatedMLP(nn.Module):
"""Gated feed-forward block used inside Exaone decoder layers."""
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
) -> None:
self.config = config
linear = functools.partial(
ColumnParallelLinear,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=nn.initializers.normal(),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.c_fc_0 = linear(config.hidden_size, config.intermediate_size, rngs=rngs)
self.c_fc_1 = linear(config.hidden_size, config.intermediate_size, rngs=rngs)
self.c_proj = linear(config.intermediate_size, config.hidden_size, rngs=rngs)
self.act_fn = ACT2FN[config.activation_function]
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
hidden_states = checkpoint_name(
self.c_proj(
self.act_fn(checkpoint_name(self.c_fc_0(hidden_states), name="mlp_gate"))
* checkpoint_name(self.c_fc_1(hidden_states), name="mlp_up")
),
name="mlp_output",
)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return hidden_states
[docs]class ExaoneAttentionInner(UnifiedAttention):
"""Exaone attention with partial RoPE."""
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",
}
def __init__(
self,
config: ExaoneConfig,
layer_idx: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
)
def _create_rotary(self, config: ExaoneConfig, dtype: jnp.dtype):
"""Override to use partial rotary factor."""
partial_rotary_factor = getattr(config, "partial_rotary_factor", 1.0)
rotary_dim = int((config.hidden_size // config.num_attention_heads) * partial_rotary_factor)
return config.get_basic_rope(
dtype=dtype,
head_size=config.hidden_size // config.num_attention_heads,
rotary_dim=rotary_dim,
is_neox_style=True,
)
def _create_o_proj(self, config, dtype, param_dtype, precision, rngs):
"""Create output projection with Exaone's custom naming (out_proj)."""
return ColumnParallelLinear(
config.num_attention_heads * self.head_dim,
config.hidden_size,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(getattr(config, "initializer_range", 0.02)),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _get_output_proj(self):
"""Access output projection using Exaone's naming."""
return self.o_proj
[docs]class ExaoneAttention(nn.Module):
"""Wrapper around ExaoneAttentionInner to wire it into decoder layers."""
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
super().__init__()
self.attention = ExaoneAttentionInner(
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,
):
return self.attention(
hidden_states=hidden_states,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
cache_view=cache_view,
cache_metadata=cache_metadata,
output_attentions=output_attentions,
frequencies=frequencies,
)
[docs]class ExaoneDecoderLayer(nn.Module):
"""Single Exaone decoder block combining attention and gated MLP."""
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
attn_block = ExaoneAttention
mlp_block = ExaoneGatedMLP
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.attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.mlp = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.ln_1 = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.layer_norm_epsilon,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.ln_2 = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.layer_norm_epsilon,
dtype=dtype,
param_dtype=param_dtype,
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,
):
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,
)
hidden_states = attn_outputs.attention_output + residual
residual = hidden_states
hidden_states = self.ln_2(hidden_states)
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 = residual + feed_forward_hidden_states
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return DecoderLayerOutput(
hidden_states=hidden_states,
attention_weight=attn_outputs.attention_weight,
cache_view=attn_outputs.cache_view,
)
[docs]@register_module(TaskType.BASE_MODULE, ExaoneConfig, model_type="exaone")
class ExaoneModel(EasyDeLBaseModule):
"""Decoder-only Exaone transformer composed of embedding, stacked blocks, and final norm."""
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.wte = nn.Embed(
self.config.vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.drop = nn.Dropout(self.config.embed_dropout, rngs=rngs)
self.h = [
ExaoneDecoderLayer(
config=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 = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.layer_norm_epsilon,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
@functools.cached_property
def frequencies(self):
return self.config.get_basic_frequencies(
head_size=self.config.hidden_size // self.config.num_attention_heads,
rotary_dim=int(
(self.config.hidden_size // self.config.num_attention_heads)
* (self.config.partial_rotary_factor if hasattr(self.config, "partial_rotary_factor") else 1.0)
),
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | 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,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> BaseModelOutput:
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.wte(input_ids.astype("i4"))
sequence_length = inputs_embeds.shape[1]
assert sequence_length <= self.config.max_position_embeddings, (
f"Maximum Position Embedding Reached ! "
f"(Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
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
hidden_states = self.drop(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))
for idx, layer in enumerate(self.h):
if output_hidden_states:
all_hidden_states += (hidden_states,)
output = layer(
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,
output_attentions=output_attentions,
frequencies=self.frequencies,
)
hidden_states = output.hidden_states
if output_attentions:
all_attentions += (output.attention_weight,)
past_key_values[idx] = output.cache_view
hidden_states = self.ln_f(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.wte
[docs]@register_module(TaskType.CAUSAL_LM, ExaoneConfig, model_type="exaone")
class ExaoneForCausalLM(BaseCausalLMModule[ExaoneModel, ExaoneConfig]):
"""Exaone model with a language modeling head for causal language modeling tasks."""
_task_type = TaskType.CAUSAL_LM
_model_type = "exaone"
_config_class = ExaoneConfig
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initialize the ExaoneForCausalLM model.
Args:
config (ExaoneConfig): The model configuration.
dtype (jnp.dtype, optional): The data type for computation. Defaults to jnp.bfloat16.
param_dtype (jnp.dtype, optional): The data type for parameters. Defaults to jnp.bfloat16.
precision (jax.lax.PrecisionLike, optional): The precision to use for matrix multiplication.
Defaults to None.
rngs (nn.Rngs): The random number generators.
"""
super().__init__(
config=config,
base_model_class=ExaoneModel,
base_model_name="transformer",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=ExaoneConfig, model_type="exaone")
class ExaoneForSequenceClassification(BaseSequenceClassificationModule[ExaoneModel, ExaoneConfig]):
"""Exaone model with a Sequence Classification head."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "exaone"
_config_class = ExaoneConfig
def __init__(
self,
config: ExaoneConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
base_model_class=ExaoneModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
pooling_strategy="last",
score_head_bias=False,
)