# Copyright 2023 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 typing as tp
import chex
import jax
from flax import nnx as nn
from jax import numpy as jnp
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import register_module
from easydel.infra.loss_utils import auxiliary_load_balancing_loss_func
from easydel.infra.modeling_outputs import (
FlaxSequenceClassifierOutput,
MoeCausalLMOutput,
MoeModelOutput,
)
from easydel.infra.utils import (
ACT2FN,
auto_remat,
block_wise_ffn,
control_mlp_sharding,
get_dot_general_by_bits,
)
from easydel.layers.attention import FlaxAttentionModule, FlexibleAttentionModule
from easydel.layers.caching import TransformerCache, TransformerCacheView
from easydel.layers.norms import RMSNorm
from easydel.modules.mixtral.mixtral_configuration import MixtralConfig as MixtralConfig
[docs]class MixtralAttention(FlaxAttentionModule):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(config=config)
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
linear = functools.partial(
nn.Linear,
use_bias=getattr(config, "attention_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.q_proj = linear(
self.hidden_size,
self.num_heads * self.head_dim,
rngs=rngs,
)
self.k_proj = linear(
self.hidden_size,
self.num_key_value_heads * self.head_dim,
rngs=rngs,
)
self.v_proj = linear(
self.hidden_size,
self.num_key_value_heads * self.head_dim,
rngs=rngs,
)
self.o_proj = linear(
self.num_heads * self.head_dim,
self.hidden_size,
rngs=rngs,
)
self.attention_performer = FlexibleAttentionModule(
dropout_prob=config.attention_dropout,
base_config=config,
softmax_scale=self.head_dim**-0.5,
)
self.rotary = self.config.get_basic_rope(self.dtype, self.head_dim)
def __call__(
self,
hidden_states: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: chex.Array,
cache_view: tp.Optional[TransformerCacheView] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
frequencies: tp.Optional[chex.Array] = None,
):
batch_size, sequence_length = hidden_states.shape[:2]
query_states, key_states, value_states = (
self.q_proj(hidden_states),
self.k_proj(hidden_states),
self.v_proj(hidden_states),
)
query_states = query_states.reshape(
batch_size,
sequence_length,
self.config.num_attention_heads,
self.head_dim,
)
key_states = key_states.reshape(
batch_size,
sequence_length,
self.config.num_key_value_heads,
self.head_dim,
)
value_states = value_states.reshape(
batch_size,
sequence_length,
self.config.num_key_value_heads,
self.head_dim,
)
query_states, key_states = self.rotary(
query=query_states,
key=key_states,
positions=position_ids,
frequencies=frequencies,
)
(
key_states,
value_states,
attention_mask,
init_attention_bias,
) = self.concatenate(
query=query_states,
key=key_states,
cache_view=cache_view,
value=value_states,
attention_mask=attention_mask,
causal_mask=causal_mask,
fcm_mask=fcm_mask,
)
attentions = self.attention_performer.forward(
query_states=query_states,
key_states=key_states,
value_states=value_states,
bias=None,
init_bias=init_attention_bias,
attention_mask=attention_mask,
segment_ids=segment_ids,
causal=True,
dropout_rng=self.rngs.params(),
)
attn_output = self.shard_attention_prod(
self._merge_heads(attentions.attention_outputs)
)
attn_output = self.o_proj(attn_output)
outputs = (
(attn_output, attentions.attention_weights)
if output_attentions
else (attn_output,)
)
return outputs
[docs]class MixtralBLockSparseTop2MLP(nn.Module):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
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
linear = functools.partial(
nn.Linear,
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(
self.config.hidden_size,
self.config.intermediate_size,
rngs=rngs,
)
self.w3 = linear(
self.config.hidden_size,
self.config.intermediate_size,
rngs=rngs,
)
self.w2 = linear(
self.config.intermediate_size,
self.config.hidden_size,
rngs=rngs,
)
self.act_fn = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states: chex.Array):
hidden_states = control_mlp_sharding(hidden_states, self.config.partition_axis)
return self.w2(self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states))
[docs]class MixtralSparseMoeBlock(nn.Module):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
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.gate = nn.Linear(
config.hidden_size,
config.num_local_experts,
use_bias=False,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=nn.initializers.normal(),
)
self.experts = [
MixtralBLockSparseTop2MLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(config.num_local_experts)
]
def __call__(self, hidden_states: chex.Array) -> tp.Tuple[chex.Array, chex.Array]:
hidden_states = control_mlp_sharding(hidden_states, self.config.partition_axis)
router_logits = self.gate(hidden_states).astype(
jnp.promote_types(self.dtype, jnp.float32)
)
routing_weights, selected_experts = jax.lax.top_k(
router_logits,
k=self.config.num_experts_per_tok,
)
routing_weights = jax.nn.softmax(
routing_weights.astype(jnp.promote_types(self.dtype, jnp.float32)),
axis=-1,
)
final_hidden_state = jnp.zeros_like(hidden_states)
for index in range(self.config.num_local_experts):
expert_layer_output = (
block_wise_ffn(
self.experts[index],
hidden_states,
self.config.scan_mlp_chunk_size,
)
if self.config.use_scan_mlp
else self.experts[index](hidden_states)
)
expert_layer_output_exp = (
jnp.sum(jnp.multiply(selected_experts == index, routing_weights), axis=-1)[
:, :, None
]
* expert_layer_output
)
final_hidden_state += expert_layer_output_exp
return (
final_hidden_state,
router_logits,
)
[docs]class MixtralDecoderLayer(nn.Module):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
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
attn_block = MixtralAttention
mlp_block = MixtralSparseMoeBlock
attn_block, mlp_block = auto_remat(
attn_block,
mlp_block,
policy=config.gradient_checkpointing,
)
self.self_attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.block_sparse_moe = 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: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: chex.Array,
segment_ids: tp.Optional[chex.Array] = None,
cache_view: tp.Optional[TransformerCacheView] = None,
output_attentions: bool = False,
output_router_logits: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
frequencies: tp.Optional[chex.Array] = None,
) -> tp.Tuple[chex.Array, chex.Array, tp.Optional[chex.Array]]:
"""
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:
tp.Tuple[chex.Array, chex.Array, tp.Optional[chex.Array]]: A tuple containing the residual_states, hidden states, and the attention weights.
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attn_out = self.self_attn(
hidden_states,
attention_mask,
position_ids,
causal_mask,
cache_view,
segment_ids,
output_attentions,
fcm_mask,
frequencies,
)
hidden_states, self_attn_weights = (
attn_out if output_attentions else (attn_out[0], None)
)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states, router_logits = self.block_sparse_moe(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights,)
if output_router_logits:
outputs += (router_logits,)
return outputs
[docs]@register_module(
"base-module",
config=MixtralConfig,
model_type="mixtral",
embedding_layer_names=["embed_tokens"],
)
class MixtralModel(EasyDeLBaseModule):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.embed_tokens = nn.Embed(
config.vocab_size,
config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
MixtralDecoderLayer(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for _ in range(config.num_hidden_layers)
]
self.norm = RMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
def __call__(
self,
input_ids: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
output_router_logits: tp.Optional[bool] = None,
past_key_values: tp.Optional[TransformerCache] = None,
return_dict: bool = True,
) -> MoeModelOutput | tp.Tuple:
if output_router_logits is None:
output_router_logits = self.config.output_router_logits
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
)
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_router_logits = () if output_router_logits else None
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError(
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
)
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids.astype("i4"))
batch_size, sequence_length, _ = inputs_embeds.shape
assert sequence_length <= self.config.max_position_embeddings, (
f"Maximum Position Embedding Reached ! (Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
if attention_mask is None:
attention_mask = jnp.ones((batch_size, sequence_length), "b1")
else:
if attention_mask.dtype != jnp.bool:
attention_mask = jnp.astype(attention_mask == 1, "b1")
if position_ids is None:
position_ids = jnp.broadcast_to(
jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0),
(batch_size, sequence_length),
).astype(jnp.int32)
if attention_mask.ndim == 2:
attention_mask = jnp.expand_dims(attention_mask, (1, 2))
hidden_states = inputs_embeds
if past_key_values is None:
past_key_values = TransformerCache.init_empty(len(self.layers))
for idx, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_outputs = block(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
cache_view=past_key_values.views[idx],
output_attentions=output_attentions,
output_router_logits=output_router_logits,
causal_mask=self.causal_mask,
segment_ids=segment_ids,
frequencies=self.frequencies,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attns += (layer_outputs[1],)
if output_router_logits:
all_router_logits += (layer_outputs[-1],)
hidden_states = self.norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
all_hidden_states,
all_self_attns,
all_router_logits,
]
if v is not None
)
return MoeModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attns,
router_logits=all_router_logits,
)
[docs]@register_module(
"causal-language-model",
config=MixtralConfig,
model_type="mixtral",
embedding_layer_names=["embed_tokens"],
)
class MixtralForCausalLM(EasyDeLBaseModule):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = MixtralModel(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.lm_head = nn.Linear(
config.hidden_size,
config.vocab_size,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
use_bias=False,
kernel_init=nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def __call__(
self,
input_ids: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
output_router_logits: tp.Optional[bool] = None,
past_key_values: tp.Optional[TransformerCache] = None,
return_dict: bool = True,
) -> MoeCausalLMOutput | tp.Tuple:
if output_router_logits is None:
output_router_logits = self.config.output_router_logits
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
past_key_values=past_key_values,
return_dict=True,
segment_ids=segment_ids,
)
logits = self.lm_head(outputs.last_hidden_state)
batch_size, seq_length, hd = logits.shape
aux_loss = None
if output_router_logits and outputs.router_logits is not None:
aux_loss = auxiliary_load_balancing_loss_func(
gate_logits=tuple(
[
logit.reshape(batch_size * seq_length, -1)
for logit in outputs.router_logits
]
),
num_experts=self.config.num_local_experts,
top_k=self.config.num_experts_per_tok,
attention_mask=attention_mask,
)
aux_loss = aux_loss * self.config.router_aux_loss_coef
if not return_dict:
outputs = (logits,) + tuple(
v
for v in [
aux_loss,
outputs.hidden_states,
outputs.attentions,
outputs.router_logits,
]
if v is not None
)
return outputs
return MoeCausalLMOutput(
aux_loss=aux_loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
router_logits=outputs.router_logits,
)
[docs]@register_module(
"sequence-classification",
config=MixtralConfig,
model_type="mixtral",
embedding_layer_names=["embed_tokens"],
)
class MixtralForSequenceClassification(EasyDeLBaseModule):
def __init__(
self,
config: MixtralConfig,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = MixtralModel(
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 = nn.Linear(
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: tp.Optional[chex.Array] = None,
inputs_embeds: tp.Optional[chex.Array] = None,
attention_mask: tp.Optional[chex.Array] = None,
position_ids: tp.Optional[chex.Array] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
output_router_logits: tp.Optional[bool] = None,
past_key_values: tp.Optional[TransformerCache] = None,
return_dict: bool = True,
) -> tp.Union[FlaxSequenceClassifierOutput, tp.Tuple]:
transformer_outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
return_dict=return_dict,
inputs_embeds=inputs_embeds,
segment_ids=segment_ids,
)
hidden_states = transformer_outputs[0]
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=tuple(
[
logit.reshape(batch_size * sequence_lengths, -1)
for logit in transformer_outputs.router_logits
]
),
num_experts=self.config.num_local_experts,
top_k=self.config.num_experts_per_tok,
attention_mask=attention_mask,
)
aux_loss = aux_loss * self.config.router_aux_loss_coef
if not return_dict:
output = (pooled_logits,) + transformer_outputs[1:] + (aux_loss,)
return output
return FlaxSequenceClassifierOutput(
logits=pooled_logits,
past_key_values=past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
aux_loss=aux_loss,
)