# 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 warnings
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,
SequenceClassifierOutput,
)
from easydel.infra.utils import ACT2FN, ArrayParam, 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 .gemma_configuration import GemmaConfig as GemmaConfig
logger = get_logger(__name__)
[docs]class GemmaRMSNorm(nn.Module):
"""Root Mean Square Layer Normalization for Gemma models.
This normalization technique normalizes the inputs by the root mean square,
providing stability during training while being computationally efficient.
"""
kernel_init = staticmethod(nn.initializers.ones)
def __init__(self, config: GemmaConfig, dtype: jnp.dtype = jnp.float32):
self.config = config
self.epsilon = self.config.rms_norm_eps
self.dtype = dtype
self.kernel = ArrayParam.bound(
shape=(self.config.hidden_size,),
dtype=dtype,
init_method="ones",
key=None,
)
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""Apply RMS normalization with learnable scale.
Args:
hidden_states: Input tensor to normalize
Returns:
Normalized and scaled hidden states
"""
variance = hidden_states.astype(jnp.float32)
variance = jnp.power(variance, 2)
variance = variance.mean(-1, keepdims=True)
hidden_states = hidden_states / jnp.sqrt(variance + self.epsilon)
return (1 + self.kernel.value.astype(self.dtype)) * jnp.asarray(hidden_states, dtype=self.dtype)
[docs]class GemmaAttention(UnifiedAttention):
"""Multi-head attention layer with RoPE embeddings for Gemma models.
Inherits from UnifiedAttention.
"""
def __init__(
self,
config: GemmaConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initialize Gemma attention."""
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
)
def _create_rotary(self, config: GemmaConfig, dtype: jnp.dtype):
"""Create Gemma-specific rotary embedding layer.
Override to use Gemma's rope_theta configuration.
"""
return config.get_basic_rope(
dtype=dtype,
head_size=self.head_dim,
rotary_dim=self.head_dim,
base=config.rope_theta,
)
[docs]class GemmaMLP(nn.Module):
"""Multi-Layer Perceptron module for Gemma models.
Implements the feedforward network component of the transformer architecture
with gated linear units and optional activation functions.
"""
def __init__(
self,
config: GemmaConfig,
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.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
embed_dim = self.config.hidden_size
inner_dim = self.config.intermediate_size if self.config.intermediate_size is not None else 4 * embed_dim
kernel_init = jax.nn.initializers.normal(config.initializer_range)
if self.config.hidden_activation is None:
warnings.warn(
"Gemma's activation function should be approximate GeLU and not exact GeLU. "
"Changing the activation function to `gelu_pytorch_tanh`."
f"if you want to use the legacy `{self.config.hidden_act}`, "
f"edit the `model.config` to set `hidden_activation={self.config.hidden_act}` ",
stacklevel=1,
)
hidden_activation = "gelu_pytorch_tanh"
else:
hidden_activation = self.config.hidden_activation
self.act = ACT2FN[hidden_activation]
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(
embed_dim,
inner_dim,
rngs=rngs,
)
self.down_proj = row_parallel_linear(
inner_dim,
embed_dim,
rngs=rngs,
)
self.up_proj = column_parallel_linear(
embed_dim,
inner_dim,
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 checkpoint_name(hidden_states, "mlp_output")
[docs]class GemmaDecoderLayer(nn.Module):
"""Single decoder layer for Gemma models.
Combines multi-head attention and feedforward networks with residual connections
and layer normalization to form a complete transformer decoder layer.
"""
def __init__(
self,
config: GemmaConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: str | jax.lax.Precision | None = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
mlp_block = GemmaMLP
attn_block = GemmaAttention
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,
)
# Define layers
self.input_layernorm = GemmaRMSNorm(self.config, dtype=self.dtype)
self.post_attention_layernorm = GemmaRMSNorm(self.config, dtype=self.dtype)
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,
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 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.
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
attn_outputs = self.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")
residual = hidden_states
hidden_states = self.post_attention_layernorm(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)
# residual connection
hidden_states = checkpoint_name(residual + feed_forward_hidden_states, "residual")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
hidden_states = checkpoint_name(hidden_states, "layer_output")
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=GemmaConfig, model_type="gemma")
class GemmaModel(EasyDeLBaseModule):
"""Decoder-only Gemma transformer wiring embeddings, decoder blocks, and output norm."""
def __init__(
self,
config: GemmaConfig,
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.hidden_size = self.config.hidden_size
embed_block = auto_remat(
nn.Embed,
policy=self.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 = [
GemmaDecoderLayer(
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 = GemmaRMSNorm(self.config, dtype=self.dtype)
# Ignore copy
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 Gemma 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 = checkpoint_name(self.embed_tokens(input_ids.astype("i4")), "embeddings")
_batch_size, sequence_length = input_ids.shape if input_ids is not None else inputs_embeds.shape[:2]
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
inputs_embeds = inputs_embeds * (self.config.hidden_size**0.5)
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 = 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=GemmaConfig, model_type="gemma")
class GemmaForCausalLM(BaseCausalLMModule[GemmaModel, GemmaConfig]):
"""Gemma model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "gemma"
_config_class = GemmaConfig
def __init__(
self,
config: GemmaConfig,
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=GemmaModel,
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: # type:ignore
"""
Forward pass through the Gemma 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 = 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=GemmaConfig, model_type="gemma")
class GemmaForSequenceClassification(EasyDeLBaseModule):
"""Gemma encoder stack with a linear classification head for sequence labels."""
def __init__(
self,
config: GemmaConfig,
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.model = GemmaModel(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
assert hasattr(config, "num_labels"), (
"in order to use `SequenceClassification` Models in `EasyDeL` "
"you first need to attach `num_labels` to model `config`"
)
self.score = ColumnParallelLinear(
self.config.hidden_size,
config.num_labels,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(stddev=config.initializer_range),
precision=self.precision,
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,
) -> SequenceClassifierOutput:
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
[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()