# 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
import chex
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, CausalLMOutput, DecoderLayerOutput
from easydel.infra.utils import 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.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from easydel.layers.norms import RMSNorm
from .xerxes_configuration import XerxesConfig as XerxesConfig
logger = get_logger(__name__)
[docs]class Identity(nn.Module):
"""No-op module used as a placeholder when optional layers are disabled."""
def __init__(self): ...
def __call__(self, x):
return x
[docs]class PostCross(nn.Module):
"""Applies a bounded tanh transform after cross attention."""
def __init__(self): ...
def __call__(self, x):
return jax.nn.tanh(x / 30.0) * 30.0
[docs]class XerxesMLP(nn.Module):
"""Feed-forward network for Xerxes decoder blocks."""
def __init__(
self,
config: XerxesConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
kernel_init = jax.nn.initializers.normal(config.initializer_range)
self.act = nn.swish if config.swish_run else functools.partial(nn.gelu, approximate=True)
column_parallel_linear = functools.partial(
ColumnParallelLinear,
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),
)
row_parallel_linear = functools.partial(
RowParallelLinear,
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 = column_parallel_linear(
self.config.hidden_size,
self.config.intermediate_size,
rngs=rngs,
)
self.up_proj = column_parallel_linear(
self.config.hidden_size,
self.config.intermediate_size,
rngs=rngs,
)
self.down_proj = row_parallel_linear(
self.config.intermediate_size,
self.config.hidden_size,
rngs=rngs,
)
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,
)
gate = checkpoint_name(self.act(self.gate_proj(hidden_states)), "mlp_gate")
up = checkpoint_name(self.up_proj(hidden_states), "mlp_up")
hidden_states = checkpoint_name(self.down_proj(gate * up), "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 XerxesAttention(UnifiedAttention):
"""Xerxes Attention with conditional Q/K normalization.
Inherits Q/K normalization from QKNormAttention.
Features:
- Conditional Q/K normalization via xe_kvnorm flag
- Layer-specific sliding window (different patterns based on layer_idx or window_pattern)
"""
def __init__(
self,
config: XerxesConfig,
layer_idx: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
causal: bool = True,
is_cross_attention: bool = False,
*,
rngs: nn.Rngs,
):
# Set sliding window BEFORE super().__init__()
self.is_local_attn = False
sliding_window = None
if not config.xe_kvnorm:
sliding_window = 4096 if bool((layer_idx % 2) == 0) else None
if config.window_pattern is not None:
self.is_local_attn = bool((layer_idx + 1) % config.window_pattern)
sliding_window = config.sliding_window if self.is_local_attn else None
self.xe_kvnorm = config.xe_kvnorm
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
sliding_window=sliding_window,
use_qk_norm=True,
)
self.layer_idx = layer_idx
self.is_cross_attention = is_cross_attention
self.causal = causal
self.attention_softmax_in_fp32 = self.dtype is not jnp.float32
def _create_q_norm(self, config, dtype, param_dtype, rngs):
"""Override to conditionally create Q norm based on xe_kvnorm flag."""
if not self.xe_kvnorm:
return None
return RMSNorm(
dim=self.head_dim,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
def _create_k_norm(self, config, dtype, param_dtype, rngs):
"""Override to conditionally create K norm based on xe_kvnorm flag."""
if not self.xe_kvnorm:
return None
return RMSNorm(
dim=self.head_dim,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
def _create_attention_performer(self, config, rngs):
"""Override to set dropout_prob to 0.0 for Xerxes."""
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=0.0,
)
def _postprocess_qkv(self, query_states, key_states, value_states):
if not self.xe_kvnorm:
return query_states, key_states, value_states
return self.query_normalization(query_states), self.key_normalization(key_states), value_states
[docs]class XerxesSparseMoeBlock(nn.Module):
"""Sparse mixture-of-experts feed-forward block used in selected layers."""
def __init__(
self,
config: XerxesConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: None | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
):
assert config.swish_run is False
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.gate = ColumnParallelLinear(
self.config.hidden_size,
self.config.num_local_experts,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=nn.initializers.normal(config.initializer_range),
rngs=rngs,
)
self.experts = [
XerxesMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for _ in range(self.config.num_local_experts)
]
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> tuple[chex.Array, chex.Array]:
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
router_logits = self.gate(hidden_states).astype(jnp.promote_types(self.dtype, jnp.float32))
routing_weights, selected_experts = jax.lax.top_k(router_logits, k=self.config.num_experts_per_tok)
routing_weights = jax.nn.softmax(routing_weights.astype(jnp.promote_types(self.dtype, jnp.float32)), axis=-1)
final_hidden_state = jnp.zeros_like(hidden_states)
for index in range(self.config.num_local_experts):
expert_layer_output = (
block_wise_ffn(
self.layers[index],
hidden_states,
self.config.scan_mlp_chunk_size,
)
if self.config.use_scan_mlp
else self.layers[index](hidden_states)
)
expert_layer_output_exp = (
jnp.sum(jnp.multiply(selected_experts == index, routing_weights), axis=-1)[:, :, None]
* expert_layer_output
)
final_hidden_state += expert_layer_output_exp
return final_hidden_state, router_logits
[docs]class XerxesDecoderLayer(nn.Module):
"""Transformer decoder block with optional cross-attention and MoE."""
def __init__(
self,
config: XerxesConfig,
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 = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
mlp_block = XerxesSparseMoeBlock if self.config.xe_moe else XerxesMLP
attn_block = XerxesAttention
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.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,
)
rms = functools.partial(
RMSNorm,
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
)
identity = config.xe_kvnorm and not config.xe_moe
if config.xe_mlpnorm:
identity = False
self.identity = identity
self.input_layernorm = rms()
self.post_attention_layernorm = rms()
self.pre_feedforward_layernorm = Identity() if identity else rms()
self.post_feedforward_layernorm = Identity() if identity else rms()
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
mask_info: MaskInfo,
position_ids: Int[Array, "batch seq_len"],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: TransformerCacheView | RaggedPagesCacheView | None = None,
cache_metadata: TransformerMetadata | RaggedPagesCacheView | None = None,
output_attentions: bool = False,
frequencies: Float[Array, "seq_len head_dim"] | None = None,
default_frequencies: Float[Array, "seq_len head_dim"] | None = 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.
"""
attn_outputs = self.self_attn(
self.input_layernorm(hidden_states),
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
default_frequencies if self.self_attn.is_local_attn else frequencies,
)
if self.identity:
hidden_states = hidden_states + attn_outputs.attention_output
residual = hidden_states
feed_forward_input = self.post_attention_layernorm(hidden_states)
else:
normed = self.post_attention_layernorm(attn_outputs.attention_output)
hidden_states = hidden_states + normed
residual = hidden_states
feed_forward_input = self.pre_feedforward_layernorm(hidden_states)
if self.config.use_scan_mlp and not self.config.xe_moe:
feed_forward_hidden_states = block_wise_ffn(
self.mlp,
feed_forward_input,
self.config.scan_mlp_chunk_size,
)
else:
feed_forward_hidden_states = self.mlp(feed_forward_input)
hidden_states = self.post_feedforward_layernorm(feed_forward_hidden_states)
hidden_states = residual + 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, config=XerxesConfig, model_type="xerxes")
class XerxesModel(EasyDeLBaseModule):
"""Xerxes decoder stack wiring embeddings, decoder layers, and final norm."""
def __init__(
self,
config: XerxesConfig,
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.hidden_size = self.config.hidden_size
embed_block = auto_remat(
nn.Embed,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.embed_tokens = embed_block(
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 = [
XerxesDecoderLayer(
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 = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
)
self.embedding_scale = float(1 if config.xe_kvnorm and not config.xe_mlpnorm else config.hidden_size**0.5)
@functools.cached_property
def default_frequencies(self):
from easydel.infra.utils import ModuleCaches
from easydel.layers.rotary_embedding import get_frequencies
frequencies = get_frequencies(
head_size=self.config.head_dim,
rotary_dim=self.config.head_dim,
max_position=self.config.granted_freq_max_position_embedding,
base=10000,
rope_scaling=None,
).astype(jnp.bfloat16)
return ModuleCaches(frequencies)
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 | RaggedPagesCacheView | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> BaseModelOutput:
"""
Forward pass through the Xerxes 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.
"""
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_tokens(input_ids.astype("i4"))
sequence_length = inputs_embeds.shape[1]
inputs_embeds = inputs_embeds * self.embedding_scale
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
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, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
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,
output_attentions=output_attentions,
frequencies=self.frequencies,
default_frequencies=self.default_frequencies,
)
hidden_states = outputs.hidden_states
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
if output_attentions:
all_attentions += (outputs.attention_weight,)
past_key_values[idx] = outputs.cache_view
hidden_states = self.norm(hidden_states)
if output_hidden_states:
all_hidden_states = outputs[1] + (hidden_states,)
outputs = (hidden_states, all_hidden_states, *outputs[2:])
else:
outputs = (hidden_states, *outputs[1:])
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_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=XerxesConfig, model_type="xerxes")
class XerxesForCausalLM(EasyDeLBaseModule):
"""Xerxes language model with LM head for causal generation."""
def __init__(
self,
config: XerxesConfig,
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.model = XerxesModel(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
lm_head_block = ColumnParallelLinear
lm_head_block = auto_remat(
lm_head_block,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.lm_head = lm_head_block(
self.config.hidden_size,
self.config.vocab_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
identity = config.xe_kvnorm and not config.xe_moe
self.post_pross = Identity() if identity else PostCross()
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 | RaggedPagesCacheView | None = None,
apply_lm_head: bool = True,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> CausalLMOutput:
"""
Forward pass through the Xerxes module.
Args:
input_ids (tp.Optional[chex.Array]): Input tensor containing token IDs.
attention_mask (tp.Optional[chex.Array]): Mask for attention.
position_ids (tp.Optional[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:
CausalLMOutput | tp.Tuple: Model output, either as a named tuple or a standard tuple.
"""
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
mask_info=mask_info,
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,
)
hidden_states = outputs.last_hidden_state
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
lm_logits = None
if apply_lm_head:
lm_logits = self.apply_lm_head(hidden_states)
return CausalLMOutput(
logits=self.post_pross(lm_logits),
hidden_states=outputs.hidden_states,
last_hidden_state=outputs.last_hidden_state,
attentions=outputs.attentions,
past_key_values=outputs.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.model.get_decoder()
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
"""
return self.lm_head
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
"""
return self.model.get_embedding()