# 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 math
import typing as tp
from functools import partial
import chex
import jax
import jax.numpy as jnp
from eformer import common_types
from eformer.escale import apply_logical_sharding
from flax import nnx as nn
from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.modeling_outputs import (
AttentionLayerOutput,
BaseModelOutput,
CausalLMOutput,
DecoderLayerOutput,
)
from easydel.infra.utils import (
ACT2FN,
ModuleCaches,
auto_remat,
get_dot_general_by_bits,
)
from easydel.layers.attention import AttentionModule, FlexibleAttentionModule
from easydel.layers.caching import (
PagedAttentionCache,
PagedAttentionCacheView,
PagedAttentionMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ParallelLinear
from easydel.layers.norms import RMSNorm
from .deepseek_configuration import DeepseekV3Config
[docs]def yarn_find_correction_dim(
num_rotations,
dim,
base=10000,
max_position_embeddings=2048,
):
return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (
2 * math.log(base)
)
[docs]def yarn_find_correction_range(
low_rot,
high_rot,
dim,
base=10000,
max_position_embeddings=2048,
):
low = math.floor(
yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings)
)
high = math.ceil(
yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings)
)
return max(low, 0), min(high, dim - 1) # Clamp values just in case
[docs]def yarn_get_mscale(scale=1.0, mscale=1.0):
if scale <= 1:
return 1.0
return 0.1 * mscale * math.log(scale) + 1.0
[docs]def yarn_linear_ramp_mask(min, max, dim):
if min == max:
max += 0.001 # Prevent singularity
linear_func = (jnp.arange(dim, dtype=jnp.float32) - min) / (max - min)
return jnp.clip(linear_func, 0, 1)
[docs]def init_deepseek_rotary_embedding(
dim,
max_position_embeddings=2048,
base=10000,
method: tp.Literal["linear", "yarn", "dynamic", None] = None,
kwargs: tp.Optional[dict] = None,
):
if method is None:
inv_freq = 1.0 / (base ** (jnp.arange(0, dim, 2).astype("float32") / dim))
t = jnp.arange(max_position_embeddings, dtype=inv_freq.dtype)
freqs = jnp.outer(t, inv_freq)
emb = jnp.concatenate((freqs, freqs), axis=-1)
return jnp.sin(emb), jnp.cos(emb)
elif method == "linear":
assert kwargs is not None
inv_freq = 1.0 / (base ** (jnp.arange(0, dim, 2).astype("float32") / dim))
t = jnp.arange(max_position_embeddings, dtype=inv_freq.dtype) / kwargs.get(
"scaling_factor"
)
freqs = jnp.outer(t, inv_freq)
emb = jnp.concatenate((freqs, freqs), axis=-1)
return jnp.sin(emb), jnp.cos(emb)
elif method == "dynamic":
assert kwargs is not None
targeted_len = kwargs.get("targeted_len", max_position_embeddings)
if targeted_len > max_position_embeddings:
base = base * (
(kwargs.get("scaling_factor") * targeted_len / max_position_embeddings)
- (kwargs.get("scaling_factor") - 1)
) ** (dim / (dim - 2))
inv_freq = 1.0 / (base ** (jnp.arange(0, dim, 2).astype("float32") / dim))
else:
inv_freq = 1.0 / (base ** (jnp.arange(0, dim, 2).astype("float32") / dim))
t = jnp.arange(max_position_embeddings, dtype=inv_freq.dtype) / kwargs.get(
"scaling_factor"
)
freqs = jnp.outer(t, inv_freq)
emb = jnp.concatenate((freqs, freqs), axis=-1)
return jnp.sin(emb), jnp.cos(emb)
elif method == "yarn":
scaling_factor = kwargs.get("scaling_factor", 1.0)
original_max_position_embeddings = kwargs.get(
"original_max_position_embeddings", 4096
)
beta_fast = kwargs.get("beta_fast", 32)
beta_slow = kwargs.get("beta_slow", 1)
mscale = kwargs.get("mscale", 1)
mscale_all_dim = kwargs.get("mscale_all_dim", 0)
freq_extra = 1.0 / (base ** (jnp.arange(0, dim, 2, dtype=jnp.float32) / dim))
freq_inter = 1.0 / (
scaling_factor * base ** (jnp.arange(0, dim, 2, dtype=jnp.float32) / dim)
)
low, high = yarn_find_correction_range(
beta_fast,
beta_slow,
dim,
base,
original_max_position_embeddings,
)
inv_freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2).astype("float32")
inv_freq = freq_inter * (1 - inv_freq_mask) + freq_extra * inv_freq_mask
t = jnp.arange(max_position_embeddings, dtype=jnp.float32)
freqs = jnp.outer(t, inv_freq)
_mscale = float(
yarn_get_mscale(scaling_factor, mscale)
/ yarn_get_mscale(scaling_factor, mscale_all_dim)
)
emb = jnp.concatenate((freqs, freqs), axis=-1)
return (jnp.sin(emb) * _mscale).astype("float32"), (jnp.cos(emb) * _mscale).astype(
"float32"
)
[docs]def rotate_half(x):
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return jnp.concatenate((-x2, x1), axis=-1)
[docs]def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
cos = jnp.expand_dims(cos[position_ids], unsqueeze_dim)
sin = jnp.expand_dims(sin[position_ids], unsqueeze_dim)
b, h, s, d = q.shape
q = q.reshape(b, h, s, d // 2, 2).transpose(0, 1, 2, 4, 3).reshape(b, h, s, d)
b, h, s, d = k.shape
k = k.reshape(b, h, s, d // 2, 2).transpose(0, 1, 2, 4, 3).reshape(b, h, s, d)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
[docs]class DeepseekV3MLP(nn.Module):
def __init__(
self,
config: DeepseekV3Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
hidden_size=None,
intermediate_size=None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
self.intermediate_size = (
config.intermediate_size if intermediate_size is None else intermediate_size
)
linear_class = partial(
ParallelLinear,
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.gate_proj = linear_class(self.hidden_size, self.intermediate_size)
self.down_proj = linear_class(self.intermediate_size, self.hidden_size)
self.up_proj = linear_class(self.hidden_size, self.intermediate_size)
self.act_fn = ACT2FN[config.hidden_act]
def __call__(self, hidden_states: jnp.ndarray) -> jnp.ndarray:
if hidden_states.ndim == 3: # if not in moe infer
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
gate = self.act_fn(self.gate_proj(hidden_states))
up = self.up_proj(hidden_states)
hidden_states = self.down_proj(gate * up)
if hidden_states.ndim == 3: # if not in moe infer
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return hidden_states
[docs]class MoEGate(nn.Module):
def __init__(
self,
config: DeepseekV3Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.top_k = self.config.num_experts_per_tok
self.n_routed_experts = self.config.n_routed_experts
self.routed_scaling_factor = self.config.routed_scaling_factor
self.scoring_func = self.config.scoring_func
self.seq_aux = self.config.seq_aux
self.topk_method = self.config.topk_method
self.n_group = self.config.n_group
self.topk_group = self.config.topk_group
self.norm_topk_prob = self.config.norm_topk_prob
self.gating_dim = self.config.hidden_size
kernel = nn.initializers.kaiming_uniform()(
rngs.param(),
(self.gating_dim, self.n_routed_experts),
param_dtype,
)
self.kernel = nn.Param(kernel)
if self.topk_method == "noaux_tc":
self.e_score_correction_bias = nn.Param(
nn.initializers.zeros(
rngs.params(),
(self.n_routed_experts,),
param_dtype,
)
)
def __call__(self, hidden_states):
bsz, seq_len, h = hidden_states.shape
hidden_states = hidden_states.reshape(-1, h)
logits = jnp.dot(
hidden_states.astype(jnp.float32),
self.kernel.value.astype(jnp.float32),
precision=self.precision,
)
if self.scoring_func == "sigmoid":
scores = jax.nn.sigmoid(logits)
else:
raise NotImplementedError(
f"insupportable scoring function for MoE gating: {self.scoring_func}"
)
if self.topk_method == "noaux_tc":
scores_for_choice = scores + self.e_score_correction_bias
group_scores = scores_for_choice.reshape(bsz * seq_len, self.n_group, -1)
top2_scores = jax.lax.top_k(group_scores, k=2)[0]
group_scores = jnp.sum(top2_scores, axis=-1)
group_idx = jax.lax.top_k(group_scores, k=self.topk_group)[1]
group_mask = jnp.zeros_like(group_scores)
indices = jnp.arange(group_mask.shape[0])[:, None]
group_mask = group_mask.at[indices, group_idx].set(1.0)
score_mask = jnp.repeat(
group_mask[:, :, None], self.n_routed_experts // self.n_group, axis=2
).reshape(bsz * seq_len, -1)
masked_scores = jnp.where(score_mask > 0, scores_for_choice, 0.0)
topk_weight, topk_idx = jax.lax.top_k(masked_scores, k=self.top_k)
else:
raise NotImplementedError(
f"insupportable TopK function for MoE gating: {self.topk_method}"
)
if self.top_k > 1 and self.norm_topk_prob:
denominator = jnp.sum(topk_weight, axis=-1, keepdims=True) + 1e-20
topk_weight = topk_weight / denominator
topk_weight = topk_weight * self.routed_scaling_factor
return topk_idx, topk_weight
[docs]class DeepseekV3MoE(nn.Module):
def __init__(
self,
config: DeepseekV3Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.num_experts_per_tok = self.config.num_experts_per_tok
self.experts_per_rank = config.n_routed_experts
self.deterministic = False
self.experts = [
DeepseekV3MLP(
config=self.config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
intermediate_size=config.moe_intermediate_size,
rngs=rngs,
)
for i in range(config.n_routed_experts)
]
self.gate = MoEGate(
config=self.config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
if config.n_shared_experts is not None:
self.shared_experts = DeepseekV3MLP(
config=self.config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
intermediate_size=self.config.moe_intermediate_size
* self.config.n_shared_experts,
rngs=rngs,
)
def __call__(self, hidden_states):
identity = hidden_states
orig_shape = hidden_states.shape
topk_idx, topk_weight = self.gate(hidden_states)
hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1])
if self.deterministic:
y = self.moe_infer(hidden_states, topk_idx, topk_weight).reshape(*orig_shape)
if self.config.n_shared_experts is not None:
y = y + self.shared_experts(identity)
return y
[docs] def moe_infer(
self,
x: jnp.ndarray,
topk_ids: jnp.ndarray,
topk_weight: jnp.ndarray,
) -> jnp.ndarray:
"""
Args:
x: Input tensor of shape [batch_size, hidden_dim]
topk_ids: Tensor of expert assignments [batch_size, top_k]
topk_weight: Tensor of expert weights [batch_size, top_k]
Returns:
Output tensor of shape [batch_size, hidden_dim]
"""
final_hidden_state = jnp.zeros_like(x)
for expert_idx, expert in enumerate(self.experts):
expert_mask = jnp.sum(
jnp.multiply(topk_ids == expert_idx, topk_weight),
axis=-1,
keepdims=True,
)
final_hidden_state = final_hidden_state + (expert_mask * expert(x))
return final_hidden_state
[docs]class DeepseekV3Attention(AttentionModule):
def __init__(
self,
config: DeepseekV3Config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
super().__init__(config=config)
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.attention_dropout = config.attention_dropout
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.max_position_embeddings = config.max_position_embeddings
self.rope_theta = config.rope_theta
self.q_lora_rank = config.q_lora_rank
self.qk_rope_head_dim = config.qk_rope_head_dim
self.kv_lora_rank = config.kv_lora_rank
self.v_head_dim = config.v_head_dim
self.qk_nope_head_dim = config.qk_nope_head_dim
self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
self.is_causal = True
linear = functools.partial(
ParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
if self.config.q_lora_rank is None:
self.q_proj = ParallelLinear(
self.hidden_size,
self.num_heads * self.q_head_dim,
use_bias=False,
)
else:
self.q_a_proj = linear(
self.hidden_size,
config.q_lora_rank,
use_bias=config.attention_bias,
)
self.q_a_layernorm = RMSNorm(
config.q_lora_rank,
eps=1e-6,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.q_b_proj = linear(
config.q_lora_rank,
self.num_heads * self.q_head_dim,
use_bias=False,
)
self.kv_a_proj_with_mqa = linear(
self.hidden_size,
config.kv_lora_rank + config.qk_rope_head_dim,
use_bias=config.attention_bias,
)
self.kv_a_layernorm = RMSNorm(
config.kv_lora_rank,
dtype=dtype,
eps=1e-6,
param_dtype=param_dtype,
rngs=rngs,
)
self.kv_b_proj = linear(
config.kv_lora_rank,
self.num_heads * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
use_bias=False,
)
self.o_proj = linear(
self.num_heads * self.v_head_dim,
self.hidden_size,
use_bias=config.attention_bias,
)
softmax_scale = self.q_head_dim**-0.5
if self.config.rope_scaling is not None:
mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0)
scaling_factor = self.config.rope_scaling["factor"]
if mscale_all_dim:
mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
softmax_scale = softmax_scale * mscale * mscale
self.attention_performer = FlexibleAttentionModule(
base_config=config,
softmax_scale=softmax_scale,
dropout_prob=config.attention_dropout,
)
def __call__(
self,
hidden_states: chex.Array,
frequencies: tp.Tuple[chex.Array, chex.Array],
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: tp.Optional[chex.Array | bool],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: tp.Optional[TransformerCacheView | PagedAttentionCacheView] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
):
"""
Forward pass of the attention module.
Args:
hidden_states (chex.Array): Input hidden states.
frequencies (tp.Tuple[chex.Array, chex.Array]): Cosine and sine components for rotary embeddings.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
causal_mask (chex.Array): Causal mask for ensuring autoregressive behavior.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
deterministic (bool): If True, disables dropout for deterministic behavior.
init_cache (bool): If True, initializes cache for caching keys and values.
output_attentions (bool): If True, outputs attention weights alongside the hidden states.
fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask.
Returns:
tp.Tuple[chex.Array, chex.Array]: A tuple containing the attention output and the attention weights.
"""
bsz, q_len, _ = hidden_states.shape
if self.config.q_lora_rank is None:
q = self.q_proj(hidden_states)
else:
q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
q = q.reshape(bsz, q_len, self.num_heads, self.q_head_dim).transpose(0, 2, 1, 3)
# Split into nope and pe parts
q_nope, q_pe = q[..., : self.qk_nope_head_dim], q[..., self.qk_nope_head_dim :]
# Key and Value projections with MQA (Multi-Query Attention) considerations
compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
k_pe = compressed_kv[..., self.kv_lora_rank :]
compressed_kv = compressed_kv[..., : self.kv_lora_rank]
k_pe = k_pe.reshape(bsz, q_len, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3)
kv = (
self.kv_b_proj(
self.kv_a_layernorm(compressed_kv),
)
.reshape(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
.transpose(0, 2, 1, 3)
)
k_nope = kv[..., : self.qk_nope_head_dim]
value_states = kv[
..., self.qk_nope_head_dim : self.qk_nope_head_dim + self.v_head_dim
]
sin, cos = frequencies
q_pe, k_pe = apply_rotary_pos_emb(
q=q_pe,
k=k_pe,
cos=cos,
sin=sin,
position_ids=position_ids,
)
query_states = jnp.zeros((bsz, self.num_heads, q_len, self.q_head_dim), q_pe.dtype)
query_states = query_states.at[..., : self.qk_nope_head_dim].set(q_nope)
query_states = query_states.at[..., self.qk_nope_head_dim :].set(q_pe)
key_states = jnp.zeros((bsz, self.num_heads, q_len, self.q_head_dim), k_pe.dtype)
key_states = key_states.at[..., : self.qk_nope_head_dim].set(k_nope)
key_states = key_states.at[..., self.qk_nope_head_dim :].set(k_pe)
query_states = query_states.transpose(0, 2, 1, 3)
key_states = key_states.transpose(0, 2, 1, 3)
value_states = value_states.transpose(0, 2, 1, 3)
(
key_states,
value_states,
attention_mask,
init_attention_bias,
cache_view,
) = self.concatenate(
query=query_states,
key=key_states,
value=value_states,
cache_view=cache_view,
cache_metadata=cache_metadata,
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,
mode=mode,
bias=None,
cache_metadata=cache_metadata,
cache_view=cache_view,
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)
return AttentionLayerOutput(
attention_output=attn_output,
attention_weight=attentions.attention_weights if output_attentions else None,
cache_view=cache_view,
)
[docs]class DeepseekV3DecoderLayer(nn.Module):
def __init__(
self,
config: DeepseekV3Config,
layer_idx: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: tp.Optional[tp.Union[str, jax.lax.Precision]] = None,
*,
rngs: nn.Rngs,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.layer_idx = layer_idx
self.hidden_size = config.hidden_size
attn_block = DeepseekV3Attention
mlp_block = DeepseekV3MLP
mlp_moe_block = DeepseekV3MoE
attn_block, mlp_block, mlp_moe_block = auto_remat(
attn_block,
mlp_block,
mlp_moe_block,
policy=config.gradient_checkpointing,
)
self.self_attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.mlp = (
mlp_moe_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
if (
config.n_routed_experts is not None
and layer_idx >= config.first_k_dense_replace
and layer_idx % config.moe_layer_freq == 0
)
else mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
)
self.input_layernorm = RMSNorm(
config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = RMSNorm(
config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
def __call__(
self,
hidden_states: chex.Array,
frequencies: tp.Tuple[chex.Array, chex.Array],
attention_mask: chex.Array,
position_ids: chex.Array,
causal_mask: tp.Optional[chex.Array | bool],
mode: common_types.RUNTIME_MODE_TYPES, # type:ignore
cache_view: tp.Optional[TransformerCacheView | PagedAttentionCacheView] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
segment_ids: tp.Optional[chex.Array] = None,
output_attentions: bool = False,
fcm_mask: tp.Optional[chex.Array] = None,
):
"""
Forward pass of the module block.
Args:
hidden_states (chex.Array): Input hidden states.
frequencies (tp.Tuple[chex.Array, chex.Array]): Cosine and sine components for rotary embeddings.
attention_mask (chex.Array): Mask to apply on the attention scores.
position_ids (chex.Array): Position indices for the tokens.
causal_mask (chex.Array): Causal mask for ensuring autoregressive behavior.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
deterministic (bool): If True, disables dropout for deterministic behavior.
init_cache (bool): If True, initializes cache for caching keys and values.
output_attentions (bool): If True, outputs attention weights alongside the hidden states.
fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask.
Returns:
tp.Tuple[chex.Array, chex.Array]: A tuple containing the attention output and the attention weights.
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
# Self Attention
attn_out = self.self_attn(
hidden_states,
frequencies,
attention_mask,
position_ids,
causal_mask,
mode,
cache_view,
cache_metadata,
segment_ids,
output_attentions,
fcm_mask,
)
hidden_states = attn_out.attention_output
hidden_states = residual + hidden_states
# Fully Connected
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
feed_forward_hidden_states = self.mlp(hidden_states)
hidden_states = residual + feed_forward_hidden_states
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return DecoderLayerOutput(
hidden_states=hidden_states,
attention_weight=attn_out.attention_weight,
cache_view=attn_out.cache_view,
)
[docs]@register_module(
TaskType.BASE_MODULE,
DeepseekV3Config,
model_type="deepseek_v3",
)
class DeepseekV3Model(EasyDeLBaseModule):
def __init__(
self,
config: DeepseekV3Config,
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(
self.config.vocab_size,
self.config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
DeepseekV3DecoderLayer(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
layer_idx=i,
rngs=rngs,
)
for i in range(self.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,
)
@functools.cached_property
def frequencies(self):
initial_rope_kwargs = {}
method = None
if self.config.rope_scaling is not None:
scaling_type = self.config.rope_scaling["type"]
method = scaling_type
if scaling_type != "yarn":
initial_rope_kwargs = dict(scaling_factor=self.config.rope_scaling["factor"])
else:
initial_rope_kwargs = {
key: self.config.rope_scaling[key]
for key in [
"original_max_position_embeddings",
"beta_fast",
"beta_slow",
"mscale",
"mscale_all_dim",
]
if key in self.config.rope_scaling
}
initial_rope_kwargs["scaling_factor"] = self.config.rope_scaling["factor"]
return ModuleCaches(
init_deepseek_rotary_embedding(
dim=self.config.qk_rope_head_dim,
max_position_embeddings=self.config.granted_freq_max_position_embedding,
base=self.config.rope_theta,
method=method, # type:ignore
kwargs=initial_rope_kwargs,
)
)
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,
mode: tp.Optional[common_types.RUNTIME_MODE_TYPES] = None, # type:ignore
past_key_values: tp.Optional[TransformerCache | PagedAttentionCache] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
) -> BaseModelOutput:
"""
Forward pass through the Deepseekv3 module.
Args:
input_ids (chex.Array): Input tensor containing token IDs.
attention_mask (chex.Array): Mask for attention.
position_ids (chex.Array): Positional indices.
segment_ids (tp.Optional[chex.Array]): Segment IDs for different input parts.
inputs_embeds (tp.Optional[chex.Array]): Embedded input tensor.
output_attentions (tp.Optional[bool]): If True, output attention weights.
output_hidden_states (tp.Optional[bool]): If True, output hidden states.
init_cache (bool): If True, initialize cache for decoding.
deterministic (bool): If True, disable dropout.
Returns:
BaseModelOutput | tp.Tuple: Model output, either as a named tuple or a standard tuple.
"""
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)
if attention_mask.ndim == 2:
attention_mask = jnp.expand_dims(attention_mask, (1, 2))
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,)
output = layer(
hidden_states=hidden_states,
frequencies=self.frequencies,
attention_mask=attention_mask,
position_ids=position_ids,
causal_mask=self.causal_mask,
output_attentions=output_attentions,
segment_ids=segment_ids,
mode=mode,
cache_view=past_key_values.views[idx],
cache_metadata=cache_metadata,
)
hidden_states = output.hidden_states
if output_attentions:
all_attentions += (output.attention_weight,)
past_key_values[idx] = output.cache_view
hidden_states = self.norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
past_key_values=past_key_values,
)
[docs]@register_module(
TaskType.CAUSAL_LM,
DeepseekV3Config,
model_type="deepseek_v3",
)
class DeepseekV3ForCausalLM(EasyDeLBaseModule):
"""
DeepseekV3 model with a language modeling head for causal language modeling tasks.
This model extends the base DeepseekV3Model by adding a linear language modeling head
on top of the transformer model. It incorporates Mixture of Experts (MoE) architecture
and is designed for generative tasks and text generation.
"""
def __init__(
self,
config: DeepseekV3Config,
dtype: jnp.dtype = jnp.float32,
param_dtype: jnp.dtype = jnp.float32,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initialize the DeepseekV3ForCausalLM model.
Args:
config (DeepseekV3Config): 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.model = DeepseekV3Model(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.lm_head = ParallelLinear(
config.hidden_size,
config.vocab_size,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
use_bias=False,
kernel_init=nn.initializers.normal(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,
output_attentions: tp.Optional[bool] = None,
output_hidden_states: tp.Optional[bool] = None,
mode: tp.Optional[common_types.RUNTIME_MODE_TYPES] = None, # type:ignore
past_key_values: tp.Optional[TransformerCache | PagedAttentionCache] = None,
cache_metadata: tp.Optional[TransformerMetadata | PagedAttentionMetadata] = None,
) -> CausalLMOutput:
"""
Forward pass of the causal language model.
Args:
input_ids (Optional[chex.Array], optional): Token IDs to process. Defaults to None.
inputs_embeds (Optional[chex.Array], optional): Pre-computed input embeddings. Defaults to None.
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.
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.
past_key_values (Optional[TransformerCache | PagedAttentionCache], optional): Cached key/values. Defaults to None.
cache_metadata (Optional[TransformerMetadata | PagedAttentionMetadata], optional): Cache metadata. Defaults to None.
Returns:
CausalLMOutput: The model outputs, either as a named tuple or a standard tuple.
"""
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
segment_ids=segment_ids,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
hidden_states = outputs.last_hidden_state
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
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)
return CausalLMOutput(
logits=lm_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
past_key_values=outputs.past_key_values,
)