# 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
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,
CausalLMOutput,
DecoderLayerOutput,
SequenceClassifierOutput,
)
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, RowParallelLinear
from easydel.layers.norms import RMSNorm
from .mistral_configuration import MistralConfig
logger = get_logger(__name__)
[docs]class MistralMLP(nn.Module):
"""Multi-Layer Perceptron module for Mistral models.
Implements the feedforward network with SiLU activation function
for efficient and effective representation learning.
"""
def __init__(
self,
config: MistralConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
column_parallel_linear = functools.partial(
ColumnParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=self.precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
row_parallel_linear = functools.partial(
RowParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=self.precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.gate_proj = column_parallel_linear(
config.hidden_size,
config.intermediate_size,
rngs=rngs,
)
self.down_proj = row_parallel_linear(
config.intermediate_size,
config.hidden_size,
rngs=rngs,
)
self.up_proj = column_parallel_linear(
config.hidden_size,
config.intermediate_size,
rngs=rngs,
)
self.act_fn = ACT2FN[self.config.hidden_act]
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""Apply SiLU feedforward transformation.
Args:
hidden_states: Input tensor [batch, seq_len, hidden_dim]
Returns:
Transformed hidden states [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_fn(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 checkpoint_name(hidden_states, "mlp_output")
[docs]class MistralAttention(UnifiedAttention):
"""Multi-head attention layer with RoPE embeddings for Mistral models.
Inherits from UnifiedAttention with Mistral-specific customizations:
- Sliding window attention support
- Custom RoPE configuration
"""
def __init__(
self,
config: MistralConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initialize Mistral attention with sliding window configuration."""
# Set sliding window before super().__init__ so it's available during network definition
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
sliding_window=config.sliding_window,
)
def _create_rotary(self, config: MistralConfig, dtype: jnp.dtype):
"""Create Mistral-specific rotary embedding layer."""
return config.get_basic_rope(dtype, self.head_dim)
[docs]class MistralDecoderLayer(nn.Module):
"""Single decoder layer for Mistral models.
Combines sliding window attention with feedforward networks,
using RMS normalization and residual connections.
"""
def __init__(
self,
config: MistralConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
attn_block = MistralAttention
mlp_block = MistralMLP
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(
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.input_layernorm = RMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = RMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
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,
) -> DecoderLayerOutput:
residual = hidden_states
attention_output = self.self_attn(
self.input_layernorm(hidden_states),
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
hidden_states = checkpoint_name(attention_output.attention_output + residual, "residual")
ffd_inp = self.post_attention_layernorm(hidden_states)
if self.config.use_scan_mlp:
feed_forward_hidden_states = block_wise_ffn(self.mlp, ffd_inp, self.config.scan_mlp_chunk_size)
else:
feed_forward_hidden_states = self.mlp(ffd_inp)
hidden_states = checkpoint_name(hidden_states + feed_forward_hidden_states, "residual")
return DecoderLayerOutput(
hidden_states=checkpoint_name(hidden_states, "layer_output"),
attention_weight=attention_output.attention_weight,
cache_view=attention_output.cache_view,
)
[docs]@register_module(TaskType.BASE_MODULE, config=MistralConfig, model_type="mistral")
class MistralModel(EasyDeLBaseModule):
"""Mistral model implementation.
This implements the Mistral language model architecture, utilizing transformer blocks
with RMSNorm, sliding window attention, and rotary position embeddings.
Attributes:
config (MistralConfig): Configuration for the model.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision: Precision setting for JAX operations.
"""
def __init__(
self,
config: MistralConfig,
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,
)
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(
config.vocab_size,
config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
MistralDecoderLayer(
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.norm = RMSNorm(
config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
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 through the Mistral model.
Args:
input_ids (chex.Array, optional): Input token IDs, shape (batch_size, sequence_length).
inputs_embeds (chex.Array, optional): Input embeddings, shape (batch_size, sequence_length, hidden_size).
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/value states.
cache_metadata (TransformerMetadata | RaggedPagesMetadata, optional): Metadata for cache handling.
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.embed_tokens(input_ids.astype("i4")), "embeddings")
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 = 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,
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 = self.norm(hidden_states)
hidden_states = checkpoint_name(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.embed_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=MistralConfig, model_type="mistral")
class MistralForCausalLM(BaseCausalLMModule[MistralModel, MistralConfig]):
"""Mistral model with a language modeling head for causal language modeling tasks."""
_task_type = TaskType.CAUSAL_LM
_model_type = "mistral"
_config_class = MistralConfig
def __init__(
self,
config: MistralConfig,
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=MistralModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
)
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,
apply_lm_head: bool = True,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> CausalLMOutput:
"""Forward pass through the Mistral model for causal language modeling.
Args:
input_ids (chex.Array, optional): Input token IDs, shape (batch_size, sequence_length).
inputs_embeds (chex.Array, optional): Input embeddings, shape (batch_size, sequence_length, hidden_size).
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/value states.
cache_metadata (TransformerMetadata | RaggedPagesMetadata, optional): Metadata for cache handling.
output_attentions (bool, optional): Whether to return attention weights.
output_hidden_states (bool, optional): Whether to return hidden states of all layers.
Returns:
Union[CausalLMOutput, Tuple]: Model outputs (logits, optional hidden states, optional attentions)
"""
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 = checkpoint_name(self.apply_lm_head(hidden_states), "lm_head_output")
return CausalLMOutput(
logits=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()
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=MistralConfig, model_type="mistral")
class MistralForSequenceClassification(BaseSequenceClassificationModule[MistralModel, MistralConfig]):
"""Mistral model for sequence classification tasks."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "mistral"
_config_class = MistralConfig
def __init__(
self,
config: MistralConfig,
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=MistralModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
classifier_name="score", # Mistral uses 'score' not 'classifier'
classifier_bias=False,
)
def __call__(
self,
input_ids: chex.Array | None = None,
inputs_embeds: chex.Array | None = None,
attention_mask: chex.Array | None = None,
mask_info: MaskInfo | None = None,
position_ids: chex.Array | 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,
) -> SequenceClassifierOutput:
"""Forward pass through the Mistral model for sequence classification.
This method processes input sequences through the Mistral model and applies
a classification head to the output.
Args:
input_ids (chex.Array, optional): Input token IDs, shape (batch_size, sequence_length).
inputs_embeds (chex.Array, optional): Input embeddings, shape (batch_size, sequence_length, hidden_size).
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/value states.
cache_metadata (TransformerMetadata | RaggedPagesMetadata, optional): Metadata for cache handling.
output_attentions (bool, optional): Whether to return attention weights.
output_hidden_states (bool, optional): Whether to return hidden states of all layers.
Returns:
Union[SequenceClassifierOutput, Tuple]: Classification outputs including logits and optional model outputs
"""
transformer_outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
inputs_embeds=inputs_embeds,
)
hidden_states = transformer_outputs.last_hidden_state
logits = self.score(hidden_states)
if input_ids is not None:
batch_size = input_ids.shape[0]
else:
batch_size = inputs_embeds.shape[0]
if self.config.pad_token_id is None and batch_size != 1:
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
sequence_lengths = jnp.argmax(jnp.equal(input_ids, self.config.pad_token_id).astype("i4"), -1) - 1
sequence_lengths = sequence_lengths % input_ids.shape[-1]
else:
sequence_lengths = -1
pooled_logits = logits[jnp.arange(batch_size), sequence_lengths]
return SequenceClassifierOutput(
logits=pooled_logits,
past_key_values=past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
[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.
This model has a sequence classification head, not an LM Head.
"""
raise NotImplementedError("This model has a sequence classification head, not a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
"""
return self.model.get_embedding()