# Copyright 2025 The EasyDeL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import typing
from functools import partial
import chex
import jax
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 (
DecoderLayerOutput,
MoeCausalLMOutput,
MoeModelOutput,
SequenceClassifierOutput,
)
from easydel.infra.utils import 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
from easydel.layers.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
RaggedPagesMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from easydel.layers.moe import (
BaseMoeModule,
ColumnParallelMoELinear,
MoeFusedHooks,
MoeLoadBalancingStrategy,
MoeRoutingStrategy,
RowParallelMoELinear,
)
from easydel.layers.norms import RMSNorm as RMSNorm
from easydel.modules.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig as Qwen2MoeConfig
[docs]class Qwen2MoeMLPStack(nn.Module):
"""Qwen2Moe MoE MLP using the new ParallelMoELinear layers."""
reform_param: typing.ClassVar = {
"gate_up_proj$": {
"splits": [
{"name": "gate_proj.kernel", "spliter": lambda x: x[..., : x.shape[-1] // 2]},
{"name": "up_proj.kernel", "spliter": lambda x: x[..., x.shape[-1] // 2 :]},
],
"inverse_spliter": lambda torch, gate, up: torch.stack((gate, up), dim=-1).flatten(-2),
},
"down_proj$": {
"splits": [
{"name": "down_proj.kernel", "spliter": lambda x: x},
],
"inverse_spliter": lambda x: x,
},
}
def __init__(
self,
config: Qwen2MoeConfig,
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.gate_proj = ColumnParallelMoELinear(
num_experts=config.num_experts,
in_features=config.hidden_size,
out_features=config.moe_intermediate_size,
rngs=rngs,
kernel_init=nn.initializers.normal(),
use_bias=False,
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
dtype=dtype,
param_dtype=param_dtype,
)
self.down_proj = RowParallelMoELinear(
num_experts=config.num_experts,
in_features=config.moe_intermediate_size,
out_features=config.hidden_size,
rngs=rngs,
use_bias=False,
kernel_init=nn.initializers.normal(),
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
dtype=dtype,
param_dtype=param_dtype,
)
self.up_proj = ColumnParallelMoELinear(
num_experts=config.num_experts,
in_features=config.hidden_size,
out_features=config.moe_intermediate_size,
rngs=rngs,
use_bias=False,
kernel_init=nn.initializers.normal(),
partition_manager=config.partition_manager,
use_expert_tensor_mode=config.use_expert_tensor_mode,
dtype=dtype,
param_dtype=param_dtype,
)
self.act_fn = nn.silu
def __call__(self, x: chex.Array, group_sizes: chex.Array, sorted_experts: chex.Array | None = None) -> chex.Array:
"""Forward pass through MoE MLP."""
gate = self.gate_proj(x, group_sizes, sorted_experts)
up = self.up_proj(x, group_sizes, sorted_experts)
return self.down_proj(self.act_fn(gate) * up, group_sizes, sorted_experts)
[docs]class Qwen2MoeMLP(nn.Module):
"""Multi-Layer Perceptron (MLP) block for the Qwen2 MoE model.
Attributes:
config (Qwen2MoeConfig): Configuration object for the model.
gate_proj (ParallelLinear): Linear layer for the gating mechanism.
down_proj (ParallelLinear): Linear layer for down-projection.
up_proj (ParallelLinear): Linear layer for up-projection.
act_fn (callable): Activation function (SiLU).
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
"""
def __init__(
self,
config: Qwen2MoeConfig,
intermediate_size: int,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the Qwen2MoeMLP module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
intermediate_size (int): The size of the intermediate layer.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
column_parallel_linear = partial(
ColumnParallelLinear,
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),
)
row_parallel_linear = partial(
RowParallelLinear,
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 = column_parallel_linear(config.hidden_size, intermediate_size, rngs=rngs)
self.down_proj = row_parallel_linear(intermediate_size, config.hidden_size, rngs=rngs)
self.up_proj = column_parallel_linear(config.hidden_size, intermediate_size, rngs=rngs)
self.act_fn = nn.silu
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
"""Forward pass of the MLP block.
Args:
hidden_states (jnp.ndarray): Input hidden states.
Returns:
jnp.ndarray: Output hidden states after MLP transformation.
"""
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
gate = checkpoint_name(self.act_fn(self.gate_proj(hidden_states)), "mlp_gate")
up = checkpoint_name(self.up_proj(hidden_states), "mlp_up")
hidden_states = checkpoint_name(self.down_proj(gate * up), "mlp_down")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return checkpoint_name(hidden_states, "mlp_output")
[docs]class Qwen2MoeAttention(UnifiedAttention):
"""Qwen2 MoE Attention module with sliding window support.
Inherits from UnifiedAttention with Qwen2Moe-specific customizations:
- Sliding window attention
- Custom bias configuration (Q/K/V use qkv_bias, O doesn't)
- Attention dropout
"""
def __init__(
self,
config: Qwen2MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initializes the Qwen2MoeAttention module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config,
dtype,
param_dtype,
precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=True,
sliding_window=config.sliding_window if config.use_sliding_window else None,
)
def _create_q_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use qkv_bias for query projection (Qwen2Moe-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_attention_heads * self.head_dim,
rngs=rngs,
use_bias=config.qkv_bias,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_k_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use qkv_bias for key projection (Qwen2Moe-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_key_value_heads * self.head_dim,
rngs=rngs,
use_bias=config.qkv_bias,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_v_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use qkv_bias for value projection (Qwen2Moe-specific)."""
return ColumnParallelLinear(
config.hidden_size,
config.num_key_value_heads * self.head_dim,
rngs=rngs,
use_bias=config.qkv_bias,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_o_proj(self, config, dtype, param_dtype, precision, rngs):
"""Override to use bias=False for output projection (Qwen2Moe-specific)."""
from easydel.layers.linear import RowParallelLinear
return RowParallelLinear(
config.num_attention_heads * self.head_dim,
config.hidden_size,
rngs=rngs,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
def _create_rotary(self, config: Qwen2MoeConfig, dtype: jnp.dtype):
"""Create Qwen2Moe-specific rotary embedding layer."""
return config.get_basic_rope(
head_size=config.hidden_size // config.num_attention_heads,
rotary_dim=config.hidden_size // config.num_attention_heads,
base=config.rope_theta,
dtype=dtype,
)
def _create_attention_performer(self, config: Qwen2MoeConfig, rngs: nn.Rngs):
"""Create attention performer with Qwen2Moe's attention dropout."""
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=config.attention_dropout,
)
[docs]class Qwen2MoeSparseBlock(BaseMoeModule):
"""Sparse Mixture of Experts (MoE) block for Qwen2 MoE.
This block routes input hidden states to a selected subset of experts
and combines their outputs.
Attributes:
config (Qwen2MoeConfig): Configuration object for the model.
gate (ParallelLinear): Linear layer for the gating network.
experts (nn.List[Qwen2MoeMLP]): List of expert MLP modules.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: Qwen2MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the Qwen2MoeSparseBlock module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
n_routed_experts=config.num_experts,
num_experts_per_tok=config.num_experts_per_tok,
hidden_size=config.hidden_size,
lbl_coef=None,
rzl_coef=None,
routing_strategy=MoeRoutingStrategy.TOP_K if config.norm_topk_prob else MoeRoutingStrategy.TOP_K_NDIV,
load_balancing_strategy=MoeLoadBalancingStrategy.NONE,
)
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.gate = ColumnParallelLinear(
config.hidden_size,
config.num_experts,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=nn.initializers.normal(config.initializer_range),
)
self.experts = Qwen2MoeMLPStack(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.shared_expert = Qwen2MoeMLP(
config=config,
intermediate_size=config.shared_expert_intermediate_size,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.shared_expert_gate = ColumnParallelLinear(
config.hidden_size,
1,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.moe_hooks = MoeFusedHooks()
def __call__(self, hidden_states: chex.Array) -> tuple[chex.Array, chex.Array]:
"""Forward pass of the Sparse MoE block.
Args:
hidden_states (chex.Array): Input hidden states (batch_size * sequence_length, hidden_dim).
Returns:
tp.Tuple[chex.Array, chex.Array]: A tuple containing:
- final_hidden_states (chex.Array): The output hidden states after MoE processing.
- router_logits (chex.Array): The logits output by the gating network.
"""
B, S, H = hidden_states.shape
out, router_logits = self.moe_call(
hidden_state=hidden_states,
gate_layer=self.gate,
expert_layer=self.experts,
wi_kernel=self.experts.gate_proj.kernel.value,
wu_kernel=self.experts.up_proj.kernel.value,
wd_kernel=self.experts.down_proj.kernel.value,
act_fn=self.experts.act_fn,
output_metrics=False,
)
hs_flat = hidden_states.reshape(-1, H)
shared_out = self.shared_expert(hs_flat)
shared_gate = jax.nn.sigmoid(self.shared_expert_gate(hs_flat))
shared_out = shared_gate * shared_out
shared_out = shared_out.reshape(B, S, H)
out = out + shared_out
return checkpoint_name(out, "moe_expert_output"), checkpoint_name(router_logits, "moe_router_logits")
[docs]class Qwen2MoeDecoderLayer(nn.Module):
"""A single decoder layer for the Qwen2 MoE model.
This layer combines self-attention, a sparse MoE block (or a standard MLP),
and residual connections with layer normalization.
Attributes:
config (Qwen2MoeConfig): Configuration object for the model.
layer_idx (int): Index of the current layer.
self_attn (Qwen2MoeAttention): Self-attention module.
mlp (Qwen2MoeSparseBlock | Qwen2MoeMLP): MoE block or standard MLP.
input_layernorm (RMSNorm): Layer normalization applied before self-attention.
post_attention_layernorm (RMSNorm): Layer normalization applied after self-attention and
before the MLP/MoE block.
dropout_rng_key (str): Name of the RNG key for dropout.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: Qwen2MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
"""Initializes the Qwen2MoeDecoderLayer module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
layer_idx (int): The index of the current layer.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
self.config = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
attn_block = Qwen2MoeAttention
mlp_block = (
Qwen2MoeSparseBlock
if (self.layer_idx not in self.config.mlp_only_layers)
and (self.config.num_experts > 0 and (self.layer_idx + 1) % self.config.decoder_sparse_step == 0)
else Qwen2MoeMLP
)
attn_block, mlp_block = auto_remat(
attn_block,
mlp_block,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.self_attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.mlp = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.input_layernorm = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = RMSNorm(
dim=self.config.hidden_size,
eps=self.config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
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, # 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 decoder layer.
Args:
hidden_states (chex.Array): Input hidden states (batch, seq_len, hidden_size).
attention_mask (chex.Array): Attention mask (batch, 1, seq_len, kv_seq_len).
position_ids (chex.Array): Position IDs (batch, seq_len).
causal_mask (tp.Optional[chex.Array | bool]): Causal mask for autoregressive behavior.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
cache_view (tp.Optional[TransformerCacheView | RaggedPagesCacheView]): Cache view for
key/value states (optional).
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for
paged attention (optional).
output_attentions (bool): Whether to output attention weights (default: False).
output_router_logits (bool): Whether to output router logits (default: False).
fcm_mask (tp.Optional[chex.Array]): Forward causal mask (FCM) mask (optional).
frequencies (tp.Optional[chex.Array]): Precomputed rotary frequencies (optional).
Returns:
DecoderLayerOutput: A tuple containing:
- hidden_states (chex.Array): Output hidden states after the decoder layer.
- attention_outputs (chex.Array): Attention weights (if `output_attentions` is True).
- router_logits (tp.Optional[chex.Array]): Router logits (if `output_router_logits` is
True and it's an MoE layer).
"""
attn_outputs = self.self_attn(
self.input_layernorm(hidden_states),
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
)
hidden_states = checkpoint_name(hidden_states + attn_outputs.attention_output, "residual")
feed_forward_input = self.post_attention_layernorm(hidden_states)
mlp_out = self.mlp(feed_forward_input)
if self.config.num_experts > 0:
feed_forward_hidden_states, router_logits = mlp_out
else:
feed_forward_hidden_states = mlp_out
router_logits = None
hidden_states = checkpoint_name(hidden_states + feed_forward_hidden_states, "residual")
return DecoderLayerOutput(
hidden_states=checkpoint_name(hidden_states, "layer_output"),
attention_weight=attn_outputs.attention_weight,
router_logits=router_logits if output_router_logits else None,
cache_view=attn_outputs.cache_view,
)
[docs]@register_module(TaskType.BASE_MODULE, config=Qwen2MoeConfig, model_type="qwen2_moe")
class Qwen2MoeModel(EasyDeLBaseModule):
"""The base Qwen2 MoE transformer model.
This class implements the core transformer architecture, including embedding layers,
decoder layers, and final normalization.
Attributes:
config (Qwen2MoeConfig): Configuration object for the model.
embed_tokens (nn.Embed): Embedding layer for input tokens.
layers (nn.List[Qwen2MoeDecoderLayer]): List of decoder layers.
norm (RMSNorm): Final layer normalization.
gradient_checkpointing (str): Gradient checkpointing strategy.
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: Qwen2MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the Qwen2MoeModel module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
embed_block = auto_remat(
nn.Embed,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.embed_tokens = embed_block(
config.vocab_size,
config.hidden_size,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layers = [
Qwen2MoeDecoderLayer(
config=config,
layer_idx=layer_idx,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for layer_idx 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,
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
output_router_logits: bool | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | RaggedPagesCache | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
apply_lm_head: bool = True,
) -> MoeModelOutput:
"""Forward pass of the Qwen2 MoE model.
Args:
input_ids (tp.Optional[chex.Array]): Input token IDs (batch, seq_len). Mutually exclusive with
`inputs_embeds`.
inputs_embeds (tp.Optional[chex.Array]): Input embeddings (batch, seq_len, hidden_size). Mutually
exclusive with `input_ids`.
attention_mask (tp.Optional[chex.Array]): Attention mask (batch, seq_len). Usually used for padding tokens.
position_ids (tp.Optional[chex.Array]): Position IDs (batch, seq_len). If None, automatically generated.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
output_attentions (tp.Optional[bool]): Whether to output attention weights (default defined by config).
output_hidden_states (tp.Optional[bool]): Whether to output hidden states for all layers
(default defined by config).
output_router_logits (tp.Optional[bool]): Whether to output router logits (default defined by config).
past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]): Precomputed key/value states
for caching.
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for
paged attention (optional).
Returns:
MoeModelOutput: The model output.
Raises:
ValueError: If both `input_ids` and `inputs_embeds` are provided or neither is provided.
"""
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 = checkpoint_name(self.embed_tokens(input_ids.astype("i4")), "embeddings")
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 = ()
all_router_logits = ()
all_self_attns = ()
sequence_length = inputs_embeds.shape[1]
assert sequence_length <= self.config.max_position_embeddings, (
f"Maximum Position Embedding Reached ! "
f"(Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
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
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(
inputs_embeds,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
for idx, block in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
layer_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 = layer_outputs.hidden_states
if output_attentions:
all_self_attns += (layer_outputs.attention_weight,)
if output_router_logits:
all_router_logits += (layer_outputs.router_logits,)
past_key_values[idx] = layer_outputs.cache_view
hidden_states = checkpoint_name(self.norm(hidden_states), "model_output")
if output_hidden_states:
all_hidden_states += (hidden_states,)
return MoeModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attns,
router_logits=all_router_logits,
past_key_values=past_key_values,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
Decoder-Only models don't have an encoder.
"""
raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
"""
return self
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
Base Models don't have a Language Model Head.
"""
raise NotImplementedError("The base model does not have a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
"""
return self.embed_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=Qwen2MoeConfig, model_type="qwen2_moe")
class Qwen2MoeForCausalLM(BaseCausalLMModule[Qwen2MoeModel, Qwen2MoeConfig]):
"""Qwen2 MoE model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "qwen2_moe"
_config_class = Qwen2MoeConfig
def __init__(
self,
config: Qwen2MoeConfig,
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=Qwen2MoeModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
router_aux_loss_coef=getattr(config, "router_aux_loss_coef", None),
)
def __call__(
self,
input_ids: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
output_router_logits: bool | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | RaggedPagesCache | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
apply_lm_head: bool = True,
) -> MoeCausalLMOutput:
"""Forward pass of the Qwen2MoeForCausalLM 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.num_experts,
top_k=self.config.num_experts_per_tok,
attention_mask=attention_mask,
)
return aux_loss + (aux_loss * self.config.router_aux_loss_coef)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=Qwen2MoeConfig, model_type="qwen2_moe")
class Qwen2MoeForSequenceClassification(EasyDeLBaseModule):
"""Qwen2 MoE model with a sequence classification head.
This class wraps the base `Qwen2MoeModel` and adds a linear layer on top
to perform sequence classification tasks.
Attributes:
config (Qwen2MoeConfig): Configuration object for the model.
model (Qwen2MoeModel): The base Qwen2 MoE model.
score (ParallelLinear): The sequence classification head (linear layer).
dtype (jnp.dtype): Data type for computations.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): Precision setting for matrix multiplications.
rngs (nn.Rngs): Random number generators.
"""
def __init__(
self,
config: Qwen2MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the Qwen2MoeForSequenceClassification module.
Args:
config (Qwen2MoeConfig): The configuration object for the model.
dtype (jnp.dtype): Data type for computations (default: jnp.float32).
param_dtype (jnp.dtype): Data type for parameters (default: jnp.float32).
precision (jax.lax.PrecisionLike): Precision setting for JAX operations (default: None).
rngs (nn.Rngs): Random number generators.
"""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.model = Qwen2MoeModel(
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 = ColumnParallelLinear(
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: Int[Array, "batch seq_len"] | None = None,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None,
attention_mask: Bool[Array, "batch seq_len"] | None = None,
mask_info: MaskInfo | None = None,
position_ids: Int[Array, "batch seq_len"] | None = None,
mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore
past_key_values: TransformerCache | RaggedPagesCache | None = None,
cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None,
apply_lm_head: bool = True,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> SequenceClassifierOutput:
"""Forward pass of the Qwen2 MoE model for sequence classification.
Args:
input_ids (tp.Optional[chex.Array]): Input token IDs (batch, seq_len). Mutually
exclusive with `inputs_embeds`.
inputs_embeds (tp.Optional[chex.Array]): Input embeddings (batch, seq_len, hidden_size).
Mutually exclusive with `input_ids`.
attention_mask (tp.Optional[chex.Array]): Attention mask (batch, seq_len). Usually used for padding tokens.
position_ids (tp.Optional[chex.Array]): Position IDs (batch, seq_len). If None, automatically generated.
segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional).
past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]): Precomputed key/value states for
caching (ignored in classification).
cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention
(ignored in classification).
output_attentions (tp.Optional[bool]): Whether to output attention weights (default defined by config).
output_hidden_states (tp.Optional[bool]): Whether to output hidden states for all layers
(default defined by config).
Returns:
SequenceClassifierOutput: The model output, including classification logits, hidden states, and attentions.
"""
transformer_outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
past_key_values=past_key_values,
cache_metadata=cache_metadata,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
inputs_embeds=inputs_embeds,
)
hidden_states = transformer_outputs.last_hidden_state
logits = self.score(hidden_states)
if input_ids is not None:
batch_size = input_ids.shape[0]
else:
batch_size = inputs_embeds.shape[0]
if self.config.pad_token_id is None and batch_size != 1:
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
sequence_lengths = jnp.argmax(jnp.equal(input_ids, self.config.pad_token_id).astype("i4"), -1) - 1
sequence_lengths = sequence_lengths % input_ids.shape[-1]
else:
sequence_lengths = -1
pooled_logits = logits[jnp.arange(batch_size), sequence_lengths]
return SequenceClassifierOutput(
logits=pooled_logits,
past_key_values=past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
Decoder-Only models don't have an encoder.
"""
raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
"""
return self.model.get_decoder()
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
This model has a sequence classification head, not an LM Head.
"""
raise NotImplementedError("This model has a sequence classification head, not a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
"""
return self.model.get_embedding()