# 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 typing
from functools import cached_property
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, block_wise_ffn, get_dot_general_by_bits
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 easydel.layers.norms import RMSNorm
from .openelm_configuration import OpenELMConfig, make_divisible
[docs]class OpenELMMultiHeadCausalAttention(UnifiedAttention):
"""OpenELM causal attention based on UnifiedAttention with per-layer head configuration."""
projection_mapping: typing.ClassVar = dict(UnifiedAttention.projection_mapping)
projection_mapping.update(
{
"query_key_value_projection": "qkv_proj",
"output_projection": "out_proj",
}
)
def __init__(
self,
config: OpenELMConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.layer_idx = layer_idx
self.num_q_heads = config.num_query_heads[layer_idx]
self.num_k_heads = config.num_kv_heads[layer_idx]
self.num_v_heads = config.num_kv_heads[layer_idx]
self.head_dim = config.head_dim
original_num_heads = getattr(config, "num_attention_heads", None)
original_num_kv_heads = getattr(config, "num_key_value_heads", None)
config.num_attention_heads = self.num_q_heads
config.num_key_value_heads = self.num_k_heads
try:
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
use_fused_qkv=True,
use_gqa=True,
)
finally:
if original_num_heads is None:
delattr(config, "num_attention_heads")
else:
config.num_attention_heads = original_num_heads
if original_num_kv_heads is None:
delattr(config, "num_key_value_heads")
else:
config.num_key_value_heads = original_num_kv_heads
# Override base head bookkeeping with per-layer values
self.num_heads = self.num_q_heads
self.num_key_value_heads = self.num_k_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.transformer_dim = config.model_dim
[docs] def define_network(
self,
config: OpenELMConfig,
dtype: jnp.dtype,
param_dtype: jnp.dtype,
precision: jax.lax.PrecisionLike,
rngs: nn.Rngs,
) -> None:
self.qkv_proj = ColumnParallelLinear(
config.model_dim,
(self.num_q_heads + self.num_k_heads + self.num_v_heads) * self.head_dim,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.out_proj = RowParallelLinear(
self.num_q_heads * self.head_dim,
config.model_dim,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
if config.normalize_qk_projections:
self.q_norm = RMSNorm(
dim=self.head_dim,
dtype=dtype,
param_dtype=param_dtype,
eps=1e-6,
rngs=rngs,
)
self.k_norm = RMSNorm(
dim=self.head_dim,
dtype=dtype,
param_dtype=param_dtype,
eps=1e-6,
rngs=rngs,
)
else:
self.q_norm = None
self.k_norm = None
def _postprocess_qkv(
self,
query_states: jnp.ndarray,
key_states: jnp.ndarray,
value_states: jnp.ndarray,
) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
if self.q_norm is not None:
query_states = self.q_norm(query_states)
if self.k_norm is not None:
key_states = self.k_norm(key_states)
return query_states, key_states, value_states
def _create_rotary(self, config: OpenELMConfig, dtype: jnp.dtype):
return config.get_basic_rope(
dtype,
head_size=config.head_dim,
rotary_dim=config.head_dim,
base=config.rope_freq_constant,
)
[docs]class OpenELMFeedForwardNetwork(nn.Module):
"""OpenELM Feed-Forward Network (FFN) module.
This module implements the FFN layer used in the OpenELM model.
It supports both standard MLP and Gated Linear Unit (GLU) variants.
Attributes:
config (OpenELMConfig): Configuration object for the model.
layer_idx (int): The index of the current 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.
ffn_with_glu (bool): Whether the FFN uses a Gated Linear Unit.
proj_1 (ParallelLinear): First linear projection layer (or gate projection in GLU).
proj_2 (ParallelLinear): Second linear projection layer (down projection).
gate_proj (ColumnParallelLinear, optional): Gate projection layer used only if `ffn_with_glu` is True.
activation_fn (callable): The activation function.
"""
def __init__(
self,
config: OpenELMConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initializes the OpenELMFeedForwardNetwork module.
Args:
config (OpenELMConfig): The configuration object for the OpenELM model.
layer_idx (int): The index of the current decoder layer.
dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32.
param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None.
rngs (nn.Rngs): Random number generators.
"""
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.layer_idx = layer_idx
ffn_multiplier = config.ffn_multipliers[layer_idx]
intermediate_dim = int(
make_divisible(
ffn_multiplier * config.model_dim, # type:ignore
divisor=config.ffn_dim_divisor,
)
)
if config.ffn_with_glu:
# FFN with Gated linear unit, as described in https://arxiv.org/abs/2002.05202v1.
self.proj_1 = ColumnParallelLinear(
config.model_dim,
2 * intermediate_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.proj_2 = RowParallelLinear(
intermediate_dim,
config.model_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.ffn_with_glu = True
else:
self.proj_1 = ColumnParallelLinear(
config.model_dim,
intermediate_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.proj_2 = RowParallelLinear(
intermediate_dim,
config.model_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.ffn_with_glu = False
self.act = ACT2FN[config.activation_fn_name]
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,
)
if self.ffn_with_glu:
y_12 = checkpoint_name(self.proj_1(hidden_states), "mlp_gate")
y_1, y_2 = jnp.split(y_12, 2, axis=-1)
hidden_states = checkpoint_name(self.proj_2(self.act(y_1) * y_2), "mlp_down")
else:
proj_1_out = checkpoint_name(self.proj_1(hidden_states), "mlp_up")
hidden_states = checkpoint_name(self.proj_2(self.act(proj_1_out)), "mlp_down")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return checkpoint_name(hidden_states, "mlp_output")
[docs]class OpenELMDecoderLayer(nn.Module):
"""OpenELM Transformer Decoder Layer.
This module represents a single decoder layer in the OpenELM model,
combining self-attention and FFN sub-layers with residual connections
and layer normalization applied before each sub-layer.
Attributes:
config (OpenELMConfig): Configuration object for the model.
layer_idx (int): The index of the current 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.
attn (OpenELMMultiHeadCausalAttention): The self-attention module.
ffn (OpenELMFeedForwardNetwork): The feed-forward network (FFN) module.
attn_norm (RMSNorm): Layer normalization before the attention layer.
ffn_norm (RMSNorm): Layer normalization before the FFN layer.
"""
def __init__(
self,
config: OpenELMConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initializes the OpenELMDecoderLayer.
Args:
config (OpenELMConfig): The configuration object for the OpenELM model.
layer_idx (int): The index of the current decoder layer.
dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32.
param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None.
rngs (nn.Rngs): Random number generators.
"""
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.layer_idx = layer_idx
attn_block = OpenELMMultiHeadCausalAttention
mlp_block = OpenELMFeedForwardNetwork
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.ffn = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.ffn_norm = RMSNorm(
self.config.model_dim,
dtype=dtype,
param_dtype=param_dtype,
eps=1e-6,
rngs=rngs,
)
self.attn_norm = RMSNorm(
self.config.model_dim,
dtype=dtype,
param_dtype=param_dtype,
eps=1e-6,
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,
):
"""Forward pass of the OpenELMDecoderLayer 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. Shape: (batch_size, sequence_length).
causal_mask (tp.Optional[chex.Array | bool]): Causal mask for ensuring autoregressive behavior.
cache_view (tp.Optional[TransformerCacheView | RaggedPagesCacheView]): Cache view for attention KVs.
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
output_attentions (bool): Whether to return attention weights. Default is False.
fcm_mask (tp.Optional[chex.Array]): Flash Chunking Mask (FCM) for attention.
frequencies (tp.Optional[chex.Array]): Precomputed rotary frequency embeddings.
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.attn_norm(hidden_states)
attn_outputs = self.attn(
hidden_states,
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
hidden_states = checkpoint_name(residual + attn_outputs.attention_output, "residual")
# Fully Connected
residual = hidden_states
hidden_states = self.ffn_norm(hidden_states)
if self.config.use_scan_mlp:
feed_forward_hidden_states = block_wise_ffn(
self.ffn,
hidden_states,
self.config.scan_mlp_chunk_size,
)
else:
feed_forward_hidden_states = self.ffn(hidden_states)
hidden_states = checkpoint_name(residual + feed_forward_hidden_states, "layer_output")
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, config=OpenELMConfig, model_type="openelm")
class OpenELMModel(EasyDeLBaseModule):
"""The base OpenELM model transformer.
This class represents the core transformer architecture of the OpenELM model,
consisting of an embedding layer, multiple OpenELMDecoderLayer layers,
and a final RMS normalization layer.
Attributes:
config (OpenELMConfig): Configuration object for the model.
dtype (jnp.dtype): Data type for computation.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations.
rngs (nn.Rngs): Random number generators.
token_embeddings (nn.Embed): Embedding layer for input tokens.
layers (tp.List[OpenELMDecoderLayer]): List of decoder layers.
norm (RMSNorm): Final layer normalization.
gradient_checkpointing (EasyDeLGradientCheckPointers): Gradient checkpointing configuration.
"""
def __init__(
self,
config: OpenELMConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the OpenELMModel.
Args:
config (OpenELMConfig): The configuration object for the OpenELM model.
dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32.
param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None.
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.token_embeddings = nn.Embed(
config.vocab_size,
config.model_dim,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
OpenELMDecoderLayer(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
layer_idx=i,
rngs=rngs,
)
for i in range(self.config.num_transformer_layers)
]
self.norm = RMSNorm(
config.model_dim,
dtype=self.dtype,
param_dtype=self.param_dtype,
eps=1e-6,
rngs=rngs,
)
if config.share_input_output_layers:
self.classifier = None
else:
self.classifier = ColumnParallelLinear(
config.model_dim,
config.vocab_size,
use_bias=False,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
)
self.num_transformer_layers = config.num_transformer_layers
@cached_property
def frequencies(self):
return self.config.get_basic_frequencies(
head_size=self.config.head_dim,
rotary_dim=self.config.head_dim,
base=self.config.rope_freq_constant,
)
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:
"""Forward pass of the OpenELMModel.
Args:
input_ids (tp.Optional[chex.Array]): Input token IDs. Shape: (batch_size, sequence_length).
inputs_embeds (tp.Optional[chex.Array]): Input embeddings.
Either `input_ids` or `inputs_embeds` must be provided.
attention_mask (tp.Optional[chex.Array]): Mask to avoid performing attention on padding token indices.
Shape: (batch_size, sequence_length).
position_ids (tp.Optional[chex.Array]): Position indices for the tokens.
Shape: (batch_size, sequence_length).
segment_ids (tp.Optional[chex.Array]): Segment IDs (unused).
output_attentions (tp.Optional[bool]): Whether to return attention weights.
Defaults to `config.output_attentions`.
output_hidden_states (tp.Optional[bool]): Whether to return hidden states for all layers.
Defaults to `config.output_hidden_states`.
past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]):
Precomputed key/value states for attention.
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention.
Returns:
BaseModelOutput: The model's output.
returns a `BaseModelOutput` object containing `last_hidden_state`, `hidden_states` (optional),
and `attentions` (optional).
Raises:
ValueError: If neither `input_ids` nor `inputs_embeds` is provided.
"""
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
if inputs_embeds is None and input_ids is not None:
inputs_embeds = checkpoint_name(self.token_embeddings(input_ids.astype("i4")), "embeddings")
else:
raise ValueError("you should specify inputs_embeds or input_ids one of them")
sequence_length = inputs_embeds.shape[1]
assert sequence_length <= self.config.max_context_length, (
f"Maximum Position Embedding Reached ! (Excepted <= {self.config.max_context_length} 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
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(
inputs_embeds,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
for idx, layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = layer(
hidden_states=hidden_states,
mask_info=mask_info,
mode=mode,
cache_view=past_key_values.views[idx],
cache_metadata=cache_metadata,
output_attentions=output_attentions,
position_ids=position_ids,
frequencies=self.frequencies,
)
hidden_states = layer_outputs.hidden_states
if output_attentions:
output_attentions += (layer_outputs.attention_weight,)
past_key_values[idx] = layer_outputs.cache_view
hidden_states = checkpoint_name(self.norm(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.token_embeddings
[docs]@register_module(TaskType.CAUSAL_LM, config=OpenELMConfig, model_type="openelm")
class OpenELMForCausalLM(BaseCausalLMModule[OpenELMModel, OpenELMConfig]):
"""OpenELM model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "openelm"
_config_class = OpenELMConfig
def __init__(
self,
config: OpenELMConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the OpenELMForCausalLM model.
Args:
config (OpenELMConfig): The configuration object for the OpenELM model.
dtype (jnp.dtype): Data type for computation. Defaults to jnp.bfloat16.
param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.bfloat16.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None.
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
base_model_class=OpenELMModel,
base_model_name="transformer",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
)