# 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.
from functools import cached_property
from typing import ClassVar
import chex
import jax
import jax.numpy as jnp
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.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.loss_utils import auxiliary_load_balancing_loss_func
from easydel.infra.modeling_outputs import (
AttentionLayerOutput,
DecoderLayerOutput,
MoeCausalLMOutput,
MoeModelOutput,
SequenceClassifierOutput,
)
from easydel.infra.utils import ACT2FN, ArrayParam, 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, RowParallelLinear
from .dbrx_configuration import DbrxConfig
[docs]class DbrxAttention(UnifiedAttention):
"""DBRX Attention module with fused QKV projection.
This module implements the multi-head attention mechanism used in the DBRX model.
It supports Grouped Query Attention (GQA) and Rotary Position Embeddings (RoPE).
The query, key, and value projections are combined into a single fused linear layer
for efficiency, and supports optional QKV clipping.
Overrides forward_standard to efficiently handle fused QKV projection.
Attributes:
config (DbrxConfig): Configuration object for the model.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations.
rngs (nn.Rngs): Random number generators.
Wqkv (ColumnParallelLinear): Fused linear layer for query, key, and value projections.
out_proj (RowParallelLinear): Linear layer for the output projection.
attention_performer (FlexibleAttentionModule): Module to perform the core attention computation.
rotary (RoPE): Rotary position embedding module.
resid_dropout (nn.Dropout): Residual dropout layer.
"""
projection_mapping: ClassVar[dict[str, str]] = {
"output_projection": "out_proj",
"query_key_value_projection": "Wqkv",
}
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initializes the DbrxAttention module.
Args:
config (DbrxConfig): The configuration object for the DBRX model.
dtype (jnp.dtype): Data type for computation. Defaults to jnp.bfloat16.
param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.bfloat16.
precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None.
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
)
[docs] def define_network(
self,
config: DbrxConfig,
dtype: jnp.dtype,
param_dtype: jnp.dtype,
precision: jax.lax.PrecisionLike,
rngs: nn.Rngs,
):
"""Override to create fused QKV projection instead of separate Q/K/V.
Args:
config: Model configuration
dtype: Data type for computations
param_dtype: Data type for parameters
precision: JAX precision setting
rngs: Random number generators
"""
num_attention_heads = config.num_attention_heads
num_key_value_heads = config.num_key_value_heads
head_dim = config.hidden_size // config.num_attention_heads
qkv_size = num_attention_heads * head_dim + 2 * num_key_value_heads * head_dim
self.Wqkv = ColumnParallelLinear(
config.hidden_size,
qkv_size,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.out_proj = RowParallelLinear(
config.hidden_size,
config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
# Create attention performer
self.attention_performer = self._create_attention_performer(config, rngs)
# Create rotary embeddings
self.rotary = self._create_rotary(config, dtype)
# Create residual dropout
self.resid_dropout = nn.Dropout(rate=config.resid_pdrop, rngs=rngs)
def _create_rotary(self, config: DbrxConfig, dtype: jnp.dtype):
"""Create rotary position embedding layer with DBRX specific configuration.
Args:
config: Model configuration
dtype: Data type for computations
"""
return config.get_basic_rope(
dtype=dtype,
rotary_dim=config.hidden_size // config.num_attention_heads,
head_size=config.hidden_size // config.num_attention_heads,
is_neox_style=True,
base=config.attn_config.rope_theta,
)
def _create_attention_performer(self, config: DbrxConfig, rngs: nn.Rngs):
"""Create attention performer module with DBRX specific settings.
Args:
config: Model configuration
rngs: Random number generators
"""
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=0.0,
)
[docs] def forward(
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,
alibi: Float[Array, "batch_or_1 heads qseq_len_or_1 kvseq_len_or_1"] | None = None,
):
"""Override to handle fused QKV projection efficiently with optional clipping."""
batch_size, sequence_length = hidden_states.shape[:2]
qkv_states = checkpoint_name(self.Wqkv(hidden_states), "attn_qkv")
if self.config.attn_config.clip_qkv is not None:
qkv_states = qkv_states.clip(
min=-self.config.attn_config.clip_qkv,
max=self.config.attn_config.clip_qkv,
)
query_size = self.hidden_size
key_size = self.num_key_value_heads * self.head_dim
query_states = qkv_states[..., :query_size]
key_states = qkv_states[..., query_size : query_size + key_size]
value_states = qkv_states[..., query_size + key_size :]
query_states = query_states.reshape(batch_size, sequence_length, self.num_heads, self.head_dim)
key_states = key_states.reshape(batch_size, sequence_length, self.num_key_value_heads, self.head_dim)
value_states = value_states.reshape(batch_size, sequence_length, self.num_key_value_heads, self.head_dim)
query_states, key_states, value_states = self._postprocess_qkv(query_states, key_states, value_states)
query_states, key_states, value_states = self.apply_qkv_shardings(query_states, key_states, value_states)
query_states, key_states = self._apply_rotary(query_states, key_states, position_ids, frequencies)
(
key_states,
value_states,
mask_info,
init_attention_bias,
cache_view,
cache_metadata,
) = self.concatenate(
query=query_states,
key=key_states,
value=value_states,
cache_view=cache_view,
cache_metadata=cache_metadata,
mask_info=mask_info,
sliding_window=getattr(self, "sliding_window", None),
)
# 7. Compute attention
attentions = self.attention_performer.forward(
query_states=query_states,
key_states=key_states,
value_states=value_states,
mode=mode,
bias=None,
cache_metadata=cache_metadata,
cache_view=cache_view,
init_bias=init_attention_bias,
mask_info=mask_info,
causal=self.causal,
sliding_window=getattr(self, "sliding_window", None),
)
# 8. Merge heads and output projection
attn_output = self.shard_attention_prod(self._merge_heads(attentions.attention_outputs))
attn_output = checkpoint_name(self.out_proj(attn_output), name="attn_output")
return AttentionLayerOutput(
attention_output=attn_output,
attention_weight=attentions.attention_weights if output_attentions else None,
cache_view=cache_view,
)
def _get_output_proj(self):
"""Override to access output projection with DBRX's naming convention.
Returns:
Output projection layer
"""
return self.out_proj
[docs]class DbrxNormAttentionNorm(nn.Module):
"""Normalization-Attention-Normalization module for DBRX models.
Implements a unique architecture pattern with normalization layers
surrounding the attention mechanism for improved gradient flow.
"""
kernel_init = staticmethod(nn.initializers.ones)
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.norm_1 = nn.LayerNorm(
self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
rngs=rngs,
)
self.attn = DbrxAttention( # statics 3,5,6,7
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.norm_2 = nn.LayerNorm(
self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
rngs=rngs,
)
self.dropout = nn.Dropout(
self.config.resid_pdrop,
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:
"""
Forward pass of the attentionNrom module.
Args:
hidden_states (chex.Array): Input hidden states.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
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:
DecoderLayerOutput: A tuple containing the residual_states, hidden states, and the attention weights.
"""
residual_states = hidden_states
hidden_states = self.norm_1(hidden_states)
attn_outputs = self.attn(
hidden_states=hidden_states,
mask_info=mask_info,
position_ids=position_ids,
output_attentions=output_attentions,
frequencies=frequencies,
mode=mode,
cache_view=cache_view,
cache_metadata=cache_metadata,
)
hidden_states = attn_outputs.attention_output
hidden_states = self.dropout(hidden_states)
hidden_states = hidden_states + residual_states
residual_states = hidden_states
hidden_states = self.norm_2(hidden_states)
return DecoderLayerOutput(
hidden_states=hidden_states,
residual_states=residual_states,
attention_weight=attn_outputs.attention_weight,
router_logits=None,
gate_loss=None,
cache_view=attn_outputs.cache_view,
)
[docs]class DbrxExpertGLU(nn.Module):
"""Gated Linear Unit expert module for DBRX mixture of experts.
Implements a single expert network with gated activation for
specialized processing in the MoE architecture.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int | None = None,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
shape = (
self.config.ffn_config.moe_num_experts * self.config.ffn_config.ffn_hidden_size,
self.config.d_model,
)
self.w1 = ArrayParam.bound(shape=shape, dtype=self.param_dtype, init_method="normal", key=rngs.params())
self.v1 = ArrayParam.bound(shape=shape, dtype=self.param_dtype, init_method="normal", key=rngs.params())
self.w2 = ArrayParam.bound(shape=shape, dtype=self.param_dtype, init_method="normal", key=rngs.params())
self.activation_fn = ACT2FN[self.config.ffn_config.ffn_act_fn["name"]]
def __call__(self, x: chex.Array, expert_idx: int) -> chex.Array:
expert_shape = (
self.config.ffn_config.moe_num_experts,
self.config.ffn_config.ffn_hidden_size,
self.config.d_model,
)
expert_w1 = checkpoint_name(self.w1.value.reshape(expert_shape)[expert_idx], name="moe_expert_w1")
expert_v1 = checkpoint_name(self.v1.value.reshape(expert_shape)[expert_idx], name="moe_expert_v1")
expert_w2 = checkpoint_name(self.w2.value.reshape(expert_shape)[expert_idx], name="moe_expert_w2")
x1 = jnp.matmul(
x,
jnp.expand_dims(expert_w1.T, 0),
precision=self.precision,
)
x2 = jnp.matmul(
x,
jnp.expand_dims(expert_v1.T, 0),
precision=self.precision,
)
x1 = self.activation_fn(x1)
x1 = x1 * x2
x1 = jnp.matmul(
x1,
jnp.expand_dims(expert_w2, 0),
precision=self.precision,
)
return x1
[docs]class DbrxExperts(nn.Module):
"""Collection of expert networks for DBRX mixture of experts.
Manages multiple expert networks that can be selected and combined
based on routing decisions for conditional computation.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.mlp = DbrxExpertGLU(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(
self,
x: chex.Array,
weights: chex.Array,
top_weights: chex.Array,
top_experts: chex.Array,
):
final_hidden_state = jnp.zeros_like(x)
for index in range(self.config.ffn_config.moe_num_experts):
output_moe_layer = self.mlp(x, index)
final_hidden_state += (
jnp.sum(jnp.multiply(index == top_experts, top_weights), axis=-1)[:, :, None] * output_moe_layer
)
return final_hidden_state
[docs]class DbrxRouter(nn.Module):
"""Router module for DBRX mixture of experts.
Determines which experts to activate for each input token,
implementing sparse routing for efficient computation.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.hidden_size = self.config.d_model
self.moe_num_experts = self.config.ffn_config.moe_num_experts
self.moe_top_k = self.config.ffn_config.moe_top_k
self.moe_jitter_eps = self.config.ffn_config.moe_jitter_eps
self.moe_normalize_expert_weights = self.config.ffn_config.moe_normalize_expert_weights
self.uniform_expert_assignment = self.config.ffn_config.uniform_expert_assignment
self.layer = ColumnParallelLinear(
config.hidden_size,
self.moe_num_experts,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
[docs] def jitter(self, x: chex.Array) -> chex.Array:
if self.moe_jitter_eps is None:
raise RuntimeError("The router does not have moe_jitter_eps set.")
low = 1.0 - self.moe_jitter_eps
high = 1.0 + self.moe_jitter_eps
noise = jax.random.normal(self.make_rng("params"), x.shape, dtype=x.dtype)
return low + noise * (high - low)
def __call__(self, x: chex.Array, deterministic: bool = True) -> tuple[chex.Array, chex.Array, chex.Array]:
if not deterministic and self.moe_jitter_eps is not None:
x = x * self.jitter(x)
weights = self.layer(x.astype(jnp.promote_types(self.dtype, jnp.float32)))
weights = jax.nn.softmax(weights.astype(jnp.promote_types(self.dtype, jnp.float32)))
top_weights, top_experts = jax.lax.top_k(weights, self.moe_top_k)
if self.moe_normalize_expert_weights:
top_weights = top_weights / jnp.linalg.norm(
top_weights,
ord=int(self.moe_normalize_expert_weights),
axis=-1,
keepdims=True,
)
if self.uniform_expert_assignment:
top_experts = jax.lax.stop_gradient(
(
jnp.arange(
0,
jnp.prod(
jnp.asarray(top_experts.shape, dtype=jnp.int32),
dtype=jnp.int32,
),
dtype=top_experts.dtype,
)
% self.moe_num_experts
).reshape(top_experts.shape)
)
weights = weights.astype(x.dtype)
top_weights = top_weights.astype(x.dtype)
return weights, top_weights, top_experts
[docs]class DbrxFFN(nn.Module):
"""Feedforward network with mixture of experts for DBRX models.
Combines router and expert networks to implement sparse MoE
feedforward layers with conditional computation.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.router = DbrxRouter(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.experts = DbrxExperts(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(self, x: chex.Array) -> tuple[chex.Array, chex.Array]:
x = apply_logical_sharding(
x,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
weights, top_weights, top_experts = self.router(x)
weights = checkpoint_name(weights, name="moe_router_logits")
out = checkpoint_name(self.experts(x, weights, top_weights, top_experts), name="moe_expert_output")
out = apply_logical_sharding(
out,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return out, weights
[docs]class DbrxBlock(nn.Module):
"""Single transformer block for DBRX models.
Integrates attention mechanisms with mixture of experts feedforward
networks, using residual connections and normalization.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.hidden_size = self.config.d_model
self.resid_pdrop = self.config.resid_pdrop
attn_block = DbrxNormAttentionNorm
ffn_block = DbrxFFN
attn_block, ffn_block = auto_remat(
attn_block,
ffn_block,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.norm_attn_norm = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.ffn = ffn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
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 | None, # type:ignore
cache_view: TransformerCacheView | RaggedPagesCacheView | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
output_attentions: bool = False,
output_router_logits: bool = False,
frequencies: Float[Array, "seq_len head_dim"] | None = None,
) -> DecoderLayerOutput:
"""
Forward pass of the attentionNrom module.
Args:
hidden_states (chex.Array): Input hidden states.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
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.
output_router_logits (bool): If True, outputs router logits.
fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask.
Returns:
DecoderLayerOutput: A tuple containing the residual_states, hidden states, and the attention weights.
"""
decoder_output = self.norm_attn_norm(
hidden_states,
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
hidden_states = decoder_output.hidden_states
hidden_states, router_logits = self.ffn(hidden_states)
hidden_states = decoder_output.residual_states + hidden_states
return decoder_output.replace(
hidden_states=hidden_states,
router_logits=router_logits if output_router_logits else None,
)
[docs]@register_module(TaskType.BASE_MODULE, config=DbrxConfig, model_type="dbrx")
class DbrxModel(EasyDeLBaseModule):
"""
Base DBRX Model outputting raw hidden-states.
This model is a Transformer-based model with a mixture of experts (MoE) architecture,
implementing the DBRX architecture as described in the original paper.
The model uses specialized attention modules and a router-based MoE FFN layer.
"""
def __init__(
self,
config: DbrxConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initialize the DbrxModel.
Args:
config (DbrxConfig): The model configuration.
dtype (jnp.dtype, optional): The data type for computation. Defaults to jnp.float32.
param_dtype (jnp.dtype, optional): The data type for parameters. Defaults to jnp.float32.
precision (jax.lax.PrecisionLike, optional): The precision to use for matrix multiplication.
Defaults to None.
rngs (nn.Rngs): The random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.padding_idx = self.config.pad_token_id
self.vocab_size = self.config.vocab_size
self.emb_pdrop = self.config.emb_pdrop
self.wte = nn.Embed(
self.config.vocab_size,
self.config.d_model,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.blocks = [
DbrxBlock(
config=config,
layer_idx=i,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.n_layers)
]
self.norm_f = nn.LayerNorm(
self.config.hidden_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
@cached_property
def frequencies(self):
return self.config.get_basic_frequencies(
rotary_dim=self.config.hidden_size // self.config.num_attention_heads,
head_size=self.config.hidden_size // self.config.num_attention_heads,
base=self.config.attn_config.rope_theta,
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"],
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | 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 of the model.
Args:
input_ids (chex.Array): Token IDs to process.
attention_mask (Optional[chex.Array], optional): Mask to avoid attention on padding tokens.
Defaults to None.
position_ids (Optional[chex.Array], optional): Position IDs. Defaults to None.
segment_ids (Optional[chex.Array], optional): Segment IDs for segment-based attention. Defaults to None.
inputs_embeds (Optional[chex.Array], optional): Pre-computed input embeddings. Defaults to None.
output_attentions (Optional[bool], optional): Whether to output attention weights. Defaults to None.
output_hidden_states (Optional[bool], optional): Whether to output hidden states. Defaults to None.
output_router_logits (Optional[bool], optional): Whether to output router logits. Defaults to None.
past_key_values (Optional[TransformerCache | RaggedPagesCache], optional): Cached key/values.
Defaults to None.
cache_metadata (Optional[TransformerMetadata | RaggedPagesMetadata], optional): Cache metadata.
Defaults to None.
Returns:
MoeModelOutput: The model outputs, either as a named tuple or a standard tuple.
"""
if output_router_logits is None:
output_router_logits = self.config.output_router_logits
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.wte(input_ids.astype("i4"))
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
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_router_logits = (
output_router_logits if output_router_logits is not None else self.config.output_router_logits
)
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
hidden_states = inputs_embeds
all_hidden_states = ()
all_router_logits = ()
all_attentions = ()
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.blocks))
for idx, block in enumerate(self.blocks):
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,
output_router_logits=output_router_logits,
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_attentions += (outputs.attention_weight,)
if output_router_logits:
all_router_logits += (outputs.router_logits,)
past_key_values[idx] = outputs.cache_view
hidden_states = self.norm_f(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
return MoeModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
router_logits=all_router_logits,
)
[docs] def get_encoder(self) -> nn.Module:
"""
Returns the encoder part of the model's graph definition.
For DbrxModel (decoder-only), this is not applicable.
"""
raise NotImplementedError("DbrxModel 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 DbrxModel, this is the model itself.
"""
return self
[docs] def get_lm_head(self) -> nn.Module:
"""
Returns the language model head of the module.
DbrxModel does not include the lm_head.
"""
raise NotImplementedError("DbrxModel does not include the language model head. See DbrxForCausalLM.")
[docs] def get_embedding(self) -> nn.Module:
"""
Returns the embedding layer of the module.
"""
return self.wte
[docs]@register_module(TaskType.CAUSAL_LM, config=DbrxConfig, model_type="dbrx")
class DbrxForCausalLM(BaseCausalLMModule[DbrxModel, DbrxConfig]):
"""DBRX model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "dbrx"
_config_class = DbrxConfig
def __init__(
self,
config: DbrxConfig,
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=DbrxModel,
base_model_name="transformer",
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"],
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | 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 DbrxForCausalLM 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,
aux_loss_fn=self._compute_aux_loss,
)
def _compute_aux_loss(self, outputs, attention_mask):
"""Compute auxiliary loss from router logits."""
if outputs.router_logits is None:
return None
aux_loss = auxiliary_load_balancing_loss_func(
gate_logits=outputs.router_logits,
num_experts=self.config.ffn_config.moe_num_experts,
top_k=self.config.ffn_config.moe_top_k,
attention_mask=attention_mask,
)
return aux_loss + (aux_loss * self.config.router_aux_loss_coef)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=DbrxConfig, model_type="dbrx")
class DbrxForSequenceClassification(BaseSequenceClassificationModule[DbrxModel, DbrxConfig]):
"""DBRX model with a Sequence Classification head."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "dbrx"
_config_class = DbrxConfig
def __init__(
self,
config: DbrxConfig,
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=DbrxModel,
base_model_name="transformer",
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"],
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | 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,
) -> SequenceClassifierOutput:
if output_router_logits is None:
output_router_logits = self.config.output_router_logits
transformer_outputs = self.base_model(
input_ids=input_ids,
attention_mask=attention_mask,
mask_info=mask_info,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
)
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 = None
if output_router_logits and transformer_outputs.router_logits is not None:
aux_loss = auxiliary_load_balancing_loss_func(
gate_logits=transformer_outputs.router_logits,
num_experts=self.config.ffn_config.moe_num_experts,
top_k=self.config.ffn_config.moe_top_k,
attention_mask=attention_mask,
)
aux_loss = aux_loss * self.config.router_aux_loss_coef
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,
)