# 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
import jax.numpy as jnp
from flax import nnx as nn
from jax.sharding import PartitionSpec
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.modeling_outputs import (
FlaxBaseModelOutput,
FlaxCausalLMOutput,
)
from easydel.infra.utils import (
auto_remat,
block_wise_ffn,
control_mlp_sharding,
get_dot_general_by_bits,
)
from easydel.layers.attention import FlaxAttentionModule, FlexibleAttentionModule
from easydel.layers.caching.transformer_cache import (
TransformerCache,
TransformerCacheMetaData,
TransformerCacheView,
)
from easydel.layers.norms import RMSNorm
from easydel.utils.helpers import get_logger
from .xerxes2_configuration import Xerxes2Config as Xerxes2Config
logger = get_logger(__name__)
[docs]class Xerxes2Attention(FlaxAttentionModule):
def __init__(
self,
config: Xerxes2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
super().__init__(config=config)
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.qhead_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
self.vhead_dim = config.vhead_dim
self.qk_rope_head_dim = config.qk_rope_head_dim
self.qk_nope_head_dim = config.qk_nope_head_dim
linear_class = functools.partial(
nn.Linear,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
if self.config.q_lora_dim is not None:
self.qa_proj = linear_class(config.hidden_size, config.q_lora_dim)
self.qa_norm = nn.LayerNorm(
config.q_lora_dim,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
)
self.qb_proj = linear_class(config.q_lora_dim, self.num_heads * self.qhead_dim)
else:
self.qc_proj = linear_class(config.hidden_size, self.num_heads * self.qhead_dim)
self.kv_mqa_proj = linear_class(
config.hidden_size,
config.kv_lora_dim + config.qk_rope_head_dim,
)
self.kv_norm = nn.LayerNorm(
config.kv_lora_dim,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
)
self.kvi_proj = linear_class(
config.kv_lora_dim,
self.num_heads * (self.qhead_dim - self.qk_rope_head_dim + self.vhead_dim),
)
self.o_proj = linear_class(
self.num_heads * self.vhead_dim,
self.config.hidden_size,
)
self.attention_performer = FlexibleAttentionModule(
base_config=config,
softmax_scale=self.qhead_dim**-0.5,
dropout_prob=0.0,
)
self.rotary = self.config.get_basic_rope(
self.dtype,
self.qk_rope_head_dim,
self.qk_rope_head_dim,
config.rope_theta,
)
def _split_heads(self, hidden_states, num_heads):
return hidden_states.reshape(hidden_states.shape[:2] + (num_heads, self.head_dim))
def __call__(
self,
hidden_states: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: chex.Array,
frequencies: tp.Tuple[chex.Array, chex.Array],
segment_ids: tp.Optional[chex.Array] = None,
cache_view: tp.Optional[TransformerCacheView] = None,
output_attentions: bool = False,
):
"""Forward pass of the attention module."""
batch_size, sequence_length = hidden_states.shape[:2]
if self.config.q_lora_dim is None:
query_states = self.qc_proj(hidden_states)
else:
query_states = self.qb_proj(self.qa_norm(self.qa_proj(hidden_states)))
query_states = query_states.reshape(
batch_size,
sequence_length,
self.num_heads,
self.qhead_dim,
)
compressed_kv = self.kv_mqa_proj(hidden_states)
compressed_kv = compressed_kv.reshape(
batch_size,
sequence_length,
1,
self.config.kv_lora_dim + self.config.qk_rope_head_dim,
)
q_nope, q_pe = (
query_states[..., : self.qk_nope_head_dim],
query_states[..., self.qk_nope_head_dim :],
)
k_pe = compressed_kv[..., self.config.kv_lora_dim :]
compressed_kv = compressed_kv[..., : self.config.kv_lora_dim]
kv = self.kvi_proj(self.kv_norm(compressed_kv))
value_states = kv[
..., self.qk_nope_head_dim : self.qk_nope_head_dim + self.vhead_dim
]
k_nope = kv[..., : self.qk_nope_head_dim]
q_pe, k_pe = self.rotary(
positions=position_ids,
query=q_pe,
key=k_pe,
frequencies=frequencies,
)
query_states = (
jnp.zeros(
(batch_size, sequence_length, self.num_heads, self.qhead_dim),
dtype=q_pe.dtype,
)
.at[..., : self.qk_nope_head_dim]
.set(q_nope)
.at[..., self.qk_nope_head_dim :]
.set(q_pe)
)
key_states = (
jnp.zeros(
(batch_size, sequence_length, 1, self.qhead_dim),
dtype=q_pe.dtype,
)
.at[..., : self.qk_nope_head_dim]
.set(k_nope)
.at[..., self.qk_nope_head_dim :]
.set(k_pe)
)
(
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,
)
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.o_proj(
self.shard_attention_prod(self._merge_heads(attentions.attention_outputs))
)
outputs = (
(attn_output, attentions.attention_weights)
if output_attentions
else (attn_output, None)
)
return outputs
[docs]class Xerxes2MLP(nn.Module):
def __init__(
self,
config: Xerxes2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.act = nn.silu
linear_class = functools.partial(
nn.Linear,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.gate_up_proj = linear_class(
config.hidden_size,
2 * config.intermediate_size,
rngs=rngs,
)
self.down_proj = linear_class(
config.intermediate_size,
config.hidden_size,
rngs=rngs,
)
def __call__(self, hidden_states):
hidden_states = control_mlp_sharding(hidden_states, self.config.partition_axis)
up_states = self.gate_up_proj(hidden_states)
gate, up_states = jnp.split(up_states, 2, axis=-1)
return self.down_proj(up_states * nn.silu(gate))
[docs]class Xerxes2DecoderLayer(nn.Module):
def __init__(
self,
config: Xerxes2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
attn_block, mlp_block = auto_remat(
Xerxes2Attention,
Xerxes2MLP,
policy=config.gradient_checkpointing,
)
self.self_attn = attn_block(
self.config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.mlp = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
rms = functools.partial(
RMSNorm,
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
)
self.input_layernorm = rms()
self.post_attention_layernorm = rms()
self.pre_feedforward_layernorm = rms()
self.post_feedforward_layernorm = rms()
def __call__(
self,
hidden_states: chex.Array,
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: chex.Array,
frequencies: tp.Tuple[chex.Array, chex.Array],
segment_ids: tp.Optional[chex.Array] = None,
cache_view: tp.Optional[TransformerCacheView] = None,
output_attentions: bool = False,
):
"""
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.
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, attn_weight = self.self_attn(
hidden_states,
attention_mask,
position_ids,
causal_mask,
frequencies,
segment_ids,
cache_view,
output_attentions,
)
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.pre_feedforward_layernorm(hidden_states)
if self.config.use_scan_mlp:
hidden_states = block_wise_ffn(
self.mlp,
hidden_states,
self.config.scan_mlp_chunk_size,
)
else:
hidden_states = self.mlp(hidden_states)
hidden_states = self.post_feedforward_layernorm(hidden_states)
hidden_states = residual + hidden_states
return hidden_states, attn_weight
[docs]@register_module(
TaskType.BASE_MODULE,
config=Xerxes2Config,
model_type="xerxes2",
)
class Xerxes2Model(EasyDeLBaseModule):
def __init__(
self,
config: Xerxes2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = 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
self.embed_tokens = nn.Embed(
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 = [
Xerxes2DecoderLayer(
self.config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for i in range(self.config.num_hidden_layers)
]
self.norm = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
)
@functools.cached_property
def frequencies(self) -> jnp.ndarray:
"""Returns frequency values from the config."""
return self.config.get_basic_frequencies(self.config.qk_rope_head_dim)
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,
past_key_values: tp.Optional[TransformerCache] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
return_dict: bool = True,
) -> tp.Union[FlaxBaseModelOutput, tp.Tuple]:
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
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
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)
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],
causal_mask=self.causal_mask,
output_attentions=output_attentions,
segment_ids=segment_ids,
frequencies=self.frequencies,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
outputs = (hidden_states, all_hidden_states, all_attentions, past_key_values)
else:
outputs = (hidden_states, all_attentions)
if not return_dict:
return tuple(v for v in outputs if v is not None)
return FlaxBaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
past_key_values=past_key_values,
)
[docs]@register_module(
TaskType.CAUSAL_LM,
config=Xerxes2Config,
model_type="xerxes2",
)
class Xerxes2ForCausalLM(EasyDeLBaseModule):
def __init__(
self,
config: Xerxes2Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = Xerxes2Model(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.lm_head = nn.Linear(
self.config.hidden_size,
self.config.vocab_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
rngs=rngs,
**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,
past_key_values: tp.Optional[TransformerCache] = None,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
return_dict: bool = True,
) -> tp.Union[FlaxCausalLMOutput, tp.Tuple]:
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
past_key_values=past_key_values,
return_dict=return_dict,
inputs_embeds=inputs_embeds,
segment_ids=segment_ids,
)
hidden_states = outputs[0]
if self.config.tie_word_embeddings:
lm_logits = jax.lax.dot_general(
hidden_states,
self.model.embed_tokens.embedding.value.T,
(((hidden_states.ndim - 1), (0,)), ((), ())),
)
else:
lm_logits = self.lm_head(hidden_states)
if not return_dict:
return (lm_logits,) + outputs[1:]
return FlaxCausalLMOutput(
logits=lm_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
past_key_values=outputs.past_key_values,
)
[docs] def init_cache(self, batch_size: int, max_length: int):
return TransformerCache.init_layers_cache(
num_hidden_layers=self.config.num_hidden_layers,
dtype=self.dtype,
key_values_partition_specs=PartitionSpec(
self.config.partition_axis.batch_axis,
self.config.partition_axis.key_sequence_axis,
None, # it's 1 by default
self.config.partition_axis.attention_dim_axis,
),
metadata=TransformerCacheMetaData.create(
batch_size=batch_size,
sequence_length=max_length,
num_heads=1,
key_dim=self.config.qk_rope_head_dim + self.config.qk_nope_head_dim,
value_dim=self.config.vhead_dim,
),
quantizer=self._quant_class(
quantization_method=self.config.kv_cache_quantization_method,
block_size=self.config.kv_cache_quantization_blocksize,
quantization_platform=self.config.platform,
),
mesh=self.config.mesh,
)