# 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 partial
import chex
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 (
DecoderLayerOutput,
MoeCausalLMOutput,
MoeModelOutput,
SequenceClassifierOutput,
)
from easydel.infra.utils import ACT2FN, auto_remat, get_dot_general_by_bits
from easydel.layers.attention import FlexibleAttentionModule
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.moe import (
BaseMoeModule,
ColumnParallelMoELinear,
MoeLoadBalancingStrategy,
MoeRoutingStrategy,
RowParallelMoELinear,
)
from easydel.layers.norms import RMSNorm
from .arctic_configuration import ArcticConfig
[docs]class ArcticAttention(UnifiedAttention):
"""Arctic Attention module with sliding window support.
Inherits from UnifiedAttention with Arctic-specific customizations:
- Sliding window attention
- Custom bias configuration (uses attention_bias config)
"""
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initialize ArcticAttention with sliding window configuration.
Args:
config: Model configuration
dtype: Data type for computations
param_dtype: Data type for parameters
precision: JAX precision setting
rngs: Random number generators
"""
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_q_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use attention_bias for query projection (Arctic-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_attention_heads * self.head_dim,
rngs=rngs,
use_bias=getattr(config, "attention_bias", False),
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_k_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use attention_bias for key projection (Arctic-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_key_value_heads * self.head_dim,
rngs=rngs,
use_bias=getattr(config, "attention_bias", False),
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_v_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use attention_bias for value projection (Arctic-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_key_value_heads * self.head_dim,
rngs=rngs,
use_bias=getattr(config, "attention_bias", False),
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_o_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use attention_bias for output projection (Arctic-specific)."""
from easydel.layers.linear import RowParallelLinear
return RowParallelLinear(
config.num_attention_heads * self.head_dim,
config.hidden_size,
rngs=rngs,
use_bias=getattr(config, "attention_bias", False),
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_rotary(self, config: ArcticConfig, dtype: jnp.dtype):
"""Create Arctic-specific rotary embedding layer."""
return config.get_basic_rope(dtype, self.head_dim, self.head_dim, True)
def _create_attention_performer(self, config: ArcticConfig, rngs: nn.Rngs):
"""Create attention performer with Arctic configuration."""
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
)
[docs]class ArcticMLPMoE(nn.Module):
"""
Arctic Multi-Layer Perceptron (MLP) block. This block implements the feed-forward network
used in the Arctic model. It can optionally function as a residual MLP.
Attributes:
config (ArcticConfig): Configuration object for the Arctic 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.
is_residual_mlp (bool): Whether this MLP block is a residual MLP. Defaults to False.
rngs (nn.Rngs): Random number generators for the module.
"""
reform_param: typing.ClassVar = {
"gate_up_proj$": {
"splits": [
{"name": "w1.kernel", "spliter": lambda x: x[..., : x.shape[-1] // 2]},
{"name": "w3.kernel", "spliter": lambda x: x[..., x.shape[-1] // 2 :]},
],
"inverse_spliter": lambda torch, gate, up: torch.stack((gate, up), dim=-1).flatten(-2),
},
"down_proj$": {
"splits": [
{"name": "w2.kernel", "spliter": lambda x: x},
],
"inverse_spliter": lambda x: x,
},
}
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
is_residual_mlp: bool = False,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.is_residual_mlp = is_residual_mlp
self.hidden_dim = config.hidden_size
self.ffn_dim = config.intermediate_size if not self.is_residual_mlp else self.hidden_dim
self.w1 = ColumnParallelMoELinear(
num_experts=config.num_local_experts,
in_features=self.hidden_dim,
out_features=self.ffn_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
rngs=rngs,
)
self.w3 = ColumnParallelMoELinear(
num_experts=config.num_local_experts,
in_features=self.hidden_dim,
out_features=self.ffn_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
rngs=rngs,
)
self.w2 = RowParallelMoELinear(
num_experts=config.num_local_experts,
in_features=self.ffn_dim,
out_features=self.hidden_dim,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=nn.initializers.normal(),
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
rngs=rngs,
)
self.act_fn = ACT2FN[self.config.hidden_act]
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
group_sizes: chex.Array,
sorted_experts: chex.Array | None = None,
):
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return apply_logical_sharding(
self.w2(
self.act_fn(self.w1(hidden_states, group_sizes, sorted_experts))
* self.w3(hidden_states, group_sizes, sorted_experts),
group_sizes,
sorted_experts,
),
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
[docs]class ArcticMLP(nn.Module):
"""
Arctic Multi-Layer Perceptron (MLP) block. This block implements the feed-forward network
used in the Arctic model. It can optionally function as a residual MLP.
Attributes:
config (ArcticConfig): Configuration object for the Arctic 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.
is_residual_mlp (bool): Whether this MLP block is a residual MLP. Defaults to False.
rngs (nn.Rngs): Random number generators for the module.
"""
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
is_residual_mlp: bool = False,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.is_residual_mlp = is_residual_mlp
self.hidden_dim = config.hidden_size
self.ffn_dim = config.intermediate_size if not self.is_residual_mlp else self.hidden_dim
linear_class = 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.w1 = linear_class(self.hidden_dim, self.ffn_dim, rngs=rngs)
self.w3 = linear_class(self.hidden_dim, self.ffn_dim, rngs=rngs)
self.w2 = linear_class(self.ffn_dim, self.hidden_dim, 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"]:
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
w1 = checkpoint_name(self.act_fn(self.w1(hidden_states)), "mlp_gate")
w3 = checkpoint_name(self.w3(hidden_states), "mlp_up")
hidden_states = checkpoint_name(self.w2(w1 * w3), "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 ArcticMoeBlock(BaseMoeModule):
"""
Arctic Mixture of Experts (MoE) block. This module implements the MoE layer used in the Arctic model,
routing tokens to different experts based on a gating mechanism.
Attributes:
config (ArcticConfig): Configuration object for the Arctic model.
layer_idx (int): The index of the current 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 for the module.
"""
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
) -> None:
super().__init__(
config=config,
n_routed_experts=config.num_local_experts,
num_experts_per_tok=config.num_experts_per_tok,
hidden_size=config.hidden_size,
lbl_coef=getattr(config, "router_aux_loss_coef", None),
rzl_coef=getattr(config, "router_z_loss_coef", None),
routing_strategy=MoeRoutingStrategy.TOP_K,
load_balancing_strategy=MoeLoadBalancingStrategy.STANDARD,
)
self.config = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.rngs = rngs
self.hidden_dim = config.hidden_size
self.num_experts = config.num_local_experts
self.top_k = config.num_experts_per_tok
self.is_moe_layer = (layer_idx + 1) % config.moe_layer_frequency == 0
if self.is_moe_layer:
self.gate = ColumnParallelLinear(
config.hidden_size,
config.num_local_experts,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=nn.initializers.normal(),
rngs=rngs,
)
self.experts = ArcticMLPMoE(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
else:
self.mlp = ArcticMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
is_residual_mlp=False,
rngs=rngs,
)
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""
Forward pass for the ArcticMoeBlock.
If the current layer is an MoE layer, it calls the MoE logic (_call_moe).
Otherwise, it passes the input through the standard MLP.
Args:
hidden_states (chex.Array): Input hidden states.
Returns:
tp.Tuple[chex.Array, chex.Array]: Tuple containing the output
hidden state and router logits (or 0.0 if not MoE).
"""
if self.is_moe_layer:
out, router_logits = self.moe_call(
hidden_state=hidden_states,
gate_layer=self.gate,
expert_layer=self.experts,
wi_kernel=self.experts.w1.kernel.value,
wu_kernel=self.experts.w3.kernel.value,
wd_kernel=self.experts.w2.kernel.value,
act_fn=self.experts.act_fn,
)
return checkpoint_name(out, "moe_expert_output"), checkpoint_name(router_logits, "moe_router_logits")
return self.mlp(hidden_states), jnp.array(0.0, dtype=hidden_states.dtype)
[docs]class ArcticDecoderLayer(nn.Module):
"""
Arctic Decoder Layer. This module combines the ArcticAttention and ArcticMoeBlock (or ArcticMLP)
with layer normalization and residual connections to form a standard Transformer decoder layer.
Attributes:
config (ArcticConfig): Configuration object for the Arctic model.
layer_idx (int): The index of the current 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 for the module.
"""
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
) -> None:
super().__init__()
self.config = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.rngs = rngs
attn_block = ArcticAttention
mlp_block = ArcticMoeBlock
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.block_sparse_moe = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
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,
)
self.parallel_attn_mlp_res = self.config.parallel_attn_mlp_res and self.block_sparse_moe.is_moe_layer
if self.parallel_attn_mlp_res:
self.residual_layernorm = RMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.residual_mlp = ArcticMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
is_residual_mlp=True,
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_input = 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 = attn_outputs.attention_output
hidden_states = checkpoint_name(residual_input + hidden_states, "residual")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
residual_attn = hidden_states
if self.parallel_attn_mlp_res:
hidden_states = self.residual_layernorm(hidden_states)
hidden_states = self.residual_mlp(hidden_states)
residual_residual = checkpoint_name(residual_attn + hidden_states, "residual")
# parallel mlp moe part
hidden_states = self.post_attention_layernorm(residual_input)
hidden_states, gate_loss = self.block_sparse_moe(hidden_states)
hidden_states = checkpoint_name(residual_residual + hidden_states, "residual")
else:
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states, gate_loss = self.block_sparse_moe(hidden_states)
hidden_states = checkpoint_name(residual_attn + 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,
router_logits=None,
cache_view=attn_outputs.cache_view,
gate_loss=gate_loss,
)
[docs]@register_module(TaskType.BASE_MODULE, config=ArcticConfig, model_type="arctic")
class ArcticModel(EasyDeLBaseModule):
"""
Core Arctic model architecture. This module implements the main Transformer stack
for the Arctic model, including token embeddings and decoder layers.
Attributes:
config (ArcticConfig): Configuration object for the Arctic 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 for the module.
"""
def __init__(
self,
config: ArcticConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
) -> None:
"""Initializes the ArcticModel."""
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(
self.config.vocab_size,
self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
ArcticDecoderLayer(
layer_idx=layer_idx,
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for layer_idx in range(config.num_hidden_layers)
]
self.norm = RMSNorm(
self.config.hidden_size,
eps=self.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,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
output_router_logits: bool | 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,
) -> MoeModelOutput:
"""Forward pass through the ArcticModel.
Args:
input_ids (Optional[chex.Array]): Input token IDs.
inputs_embeds (Optional[chex.Array]): Input embeddings (alternative to input_ids).
attention_mask (Optional[chex.Array]): Mask to avoid attending to padding tokens.
position_ids (Optional[chex.Array]): Position IDs for positional embeddings.
segment_ids (Optional[chex.Array]): Segment IDs (if applicable).
output_attentions (Optional[bool]): Whether to return attention weights.
output_hidden_states (Optional[bool]): Whether to return all hidden states.
past_key_values (Optional[TransformerCache | RaggedPagesCache]):
Cached key/value states for faster decoding.
cache_metadata (Optional[TransformerMetadata | RaggedPagesMetadata]):
Metadata for paged attention cache.
Returns:
MoeModelOutput: Model outputs
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_router_losses = ()
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]
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, layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
outputs = 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 = 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_self_attns += (outputs.attention_weight,)
all_router_losses += (outputs.gate_loss,)
past_key_values[idx] = 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 MoeModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attns,
all_router_losses=all_router_losses,
past_key_values=past_key_values,
)
[docs] def get_encoder(self) -> nn.Module:
"""
Returns the encoder part of the model's graph definition.
For ArcticModel (decoder-only), this is not applicable.
"""
# As per instructions, raise NotImplementedError for non-encoder models
# Or you could return `self` if you consider the whole model the "encoder" context,
# but raising NotImplementedError is more standard for a decoder-only base.
raise NotImplementedError("ArcticModel is a decoder-only model and does not have a separate encoder.")
[docs] def get_decoder(self) -> nn.Module:
"""
Returns the decoder part of the model's graph definition.
For ArcticModel, this is the model itself.
"""
# The ArcticModel *is* the decoder stack.
return self
[docs] def get_lm_head(self) -> nn.Module:
"""
Returns the language model head of the module.
ArcticModel does not include the lm_head.
"""
# The lm_head is part of ArcticForCausalLM, not the base ArcticModel.
raise NotImplementedError("ArcticModel does not include the language model head. See ArcticForCausalLM.")
[docs] def get_embedding(self) -> nn.Module:
"""
Returns the embedding layer of the module.
"""
return self.embed_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=ArcticConfig, model_type="arctic")
class ArcticForCausalLM(BaseCausalLMModule[ArcticModel, ArcticConfig]):
"""Arctic model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "arctic"
_config_class = ArcticConfig
def __init__(
self,
config: ArcticConfig,
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=ArcticModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
router_aux_loss_coef=getattr(config, "router_aux_loss_coef", None),
)
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,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
output_router_logits: bool | 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,
) -> MoeCausalLMOutput:
"""Forward pass of the ArcticForCausalLM model."""
return self.forward_moe(
input_ids=input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
apply_lm_head=apply_lm_head,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=ArcticConfig, model_type="arctic")
class ArcticForSequenceClassification(BaseSequenceClassificationModule[ArcticModel, ArcticConfig]):
"""Arctic model with a Sequence Classification head."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "arctic"
_config_class = ArcticConfig
def __init__(
self,
config: ArcticConfig,
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=ArcticModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
classifier_name="score",
classifier_bias=False,
router_aux_loss_coef=getattr(config, "router_aux_loss_coef", None),
)
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:
"""Forward pass through the ArcticForSequenceClassification model."""
transformer_outputs = self.base_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]
aux_loss = self.compute_router_aux_loss(transformer_outputs)
return SequenceClassifierOutput(
logits=pooled_logits,
past_key_values=past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
aux_loss=aux_loss,
)