# 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 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.modeling_outputs import (
DecoderLayerOutput,
MoeModelOutput,
)
from easydel.infra.utils import ACT2FN, ArrayParam, auto_remat, get_dot_general_by_bits
from easydel.layers.attention_unified import UnifiedAttention
from easydel.layers.base_modules import BaseCausalLMModule, BaseSequenceClassificationModule
from easydel.layers.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
RaggedPagesMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from easydel.layers.moe import (
BaseMoeModule,
ColumnParallelMoELinear,
MoeLoadBalancingStrategy,
MoeRoutingStrategy,
RowParallelMoELinear,
)
from easydel.layers.norms import RMSNorm
from .glm4_moe_configuration import Glm4MoeConfig
[docs]class Glm4MoeMLP(nn.Module):
"""Dense feed-forward block used in non-MoE GLM-4-MoE layers."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
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, config.intermediate_size)
self.up_proj = column_parallel_linear(config.hidden_size, config.intermediate_size)
self.down_proj = row_parallel_linear(config.intermediate_size, config.hidden_size)
self.act_fn = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> jnp.ndarray:
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
gate_output = self.act_fn(checkpoint_name(self.gate_proj(hidden_states), name="mlp_gate"))
up_output = checkpoint_name(self.up_proj(hidden_states), name="mlp_up")
hidden_states = checkpoint_name(self.down_proj(gate_output * up_output), name="mlp_down")
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return hidden_states, None
[docs]class Glm4MoeMLPStack(nn.Module):
"""Glm4Moe 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: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__()
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.gate_proj = ColumnParallelMoELinear(
num_experts=config.n_routed_experts,
in_features=config.hidden_size,
out_features=config.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.n_routed_experts,
in_features=config.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.n_routed_experts,
in_features=config.hidden_size,
out_features=config.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 = ACT2FN[config.hidden_act]
def __call__(
self,
x: Array,
group_sizes: Array,
sorted_experts: Array | None = None,
) -> Array:
"""Forward pass through MoE MLP."""
hidden_states = self.act_fn(checkpoint_name(self.gate_proj(x, group_sizes, sorted_experts), name="moe_gate"))
hidden_states = hidden_states * checkpoint_name(self.up_proj(x, group_sizes, sorted_experts), name="moe_up")
outputs = checkpoint_name(self.down_proj(hidden_states, group_sizes, sorted_experts), name="moe_expert_output")
return outputs
[docs]class Glm4MoeTopKRouter(nn.Module):
"""Selects top-k experts per token for GLM-4-MoE routing."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
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 = config.num_experts_per_tok
self.n_routed_experts = config.n_routed_experts
self.routed_scaling_factor = config.routed_scaling_factor
self.n_group = config.n_group
self.topk_group = config.topk_group
self.norm_topk_prob = config.norm_topk_prob
self.kernel = ArrayParam.bound(
shape=(self.n_routed_experts, config.hidden_size),
dtype=param_dtype,
init_method="normal",
init_kwargs={"stddev": config.initializer_range},
key=rngs.param(),
)
self.e_score_correction_bias = ArrayParam.bound(
shape=(self.n_routed_experts,),
dtype=jnp.float32,
init_method="zeros",
key=None,
)
[docs] def get_selected_experts(self, scores):
scores_for_choice = scores + self.e_score_correction_bias.value
batch_size = scores_for_choice.shape[0]
group_scores = scores_for_choice.reshape(batch_size, self.n_group, self.n_routed_experts // self.n_group)
top2_per_group = jax.lax.top_k(group_scores, k=2)[0]
group_scores_sum = jnp.sum(top2_per_group, axis=-1)
scores_for_choice = jnp.where(
jax.nn.one_hot(
jax.lax.top_k(
group_scores_sum,
k=self.topk_group,
)[-1],
self.n_group,
dtype=scores.dtype,
)[:, :, None]
.repeat(self.n_routed_experts // self.n_group, axis=2)
.reshape(batch_size, self.n_routed_experts),
scores_for_choice,
0.0,
)
_, selected_experts = jax.lax.top_k(scores_for_choice, k=self.top_k)
return selected_experts
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
hidden_states = hidden_states.reshape(-1, self.config.hidden_size)
router_logits = checkpoint_name(
jnp.matmul(hidden_states.astype(jnp.float32), self.kernel.value.astype(jnp.float32)),
name="moe_router_logits",
)
scores = jax.nn.sigmoid(router_logits)
selected_experts = self.get_selected_experts(scores)
batch_size = scores.shape[0]
batch_indices = jnp.arange(batch_size)[:, None]
selected_weights = scores[batch_indices, selected_experts]
if self.norm_topk_prob:
denominator = jnp.sum(selected_weights, axis=-1, keepdims=True) + 1e-20
selected_weights = selected_weights / denominator
selected_weights = selected_weights * self.routed_scaling_factor
return selected_weights
[docs]class Glm4MoeMoE(BaseMoeModule):
"""GLM-4-MoE feed-forward wrapper combining router and expert stacks."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
n_routed_experts=config.n_routed_experts,
num_experts_per_tok=config.num_experts_per_tok,
hidden_size=config.hidden_size,
lbl_coef=getattr(config, "router_aux_loss_coef", None),
rzl_coef=getattr(config, "router_z_loss_coef", None),
routing_strategy=MoeRoutingStrategy.TOP_K,
load_balancing_strategy=MoeLoadBalancingStrategy.STANDARD,
)
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.experts = Glm4MoeMLPStack(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.gate = Glm4MoeTopKRouter(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.shared_experts = Glm4MoeMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> tuple[Array, Array]:
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,
)
shared_output, _ = self.shared_experts(hidden_states)
out = out + shared_output
return checkpoint_name(out, "moe_expert_output"), checkpoint_name(router_logits, "moe_router_logits")
[docs]class Glm4MoeAttention(UnifiedAttention):
"""Attention layer variant used inside GLM-4-MoE decoder blocks."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.layer_idx = layer_idx
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
use_qk_norm=config.use_qk_norm,
)
[docs]class Glm4MoeDecoderLayer(nn.Module):
"""Single decoder block for GLM-4-MoE with attention and MoE MLP."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.layer_idx = layer_idx
attn_block = Glm4MoeAttention
mlp_block = Glm4MoeMLP if layer_idx < config.first_k_dense_replace else Glm4MoeMoE
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=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: 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,
):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attn_outputs = self.self_attn(
hidden_states=hidden_states,
mask_info=mask_info,
position_ids=position_ids,
mode=mode,
cache_view=cache_view,
cache_metadata=cache_metadata,
output_attentions=output_attentions,
frequencies=frequencies,
)
hidden_states = residual + attn_outputs.attention_output
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states, router_logits = self.mlp(hidden_states)
hidden_states = residual + 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_outputs.attention_weight,
cache_view=attn_outputs.cache_view,
router_logits=router_logits,
)
[docs]@register_module(TaskType.BASE_MODULE, config=Glm4MoeConfig, model_type="glm4_moe")
class Glm4MoeModel(EasyDeLBaseModule):
"""GLM4 MoE model implementation."""
def __init__(
self,
config: Glm4MoeConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
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(
num_embeddings=self.config.vocab_size,
features=self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
rngs=rngs,
)
self.layers = [
Glm4MoeDecoderLayer(
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,
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,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
output_router_logits: bool | None = None,
) -> MoeModelOutput:
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"))
sequence_length = inputs_embeds.shape[1]
all_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
all_router_logits = () if output_router_logits else None
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
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, 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_attentions += (layer_outputs.attention_weight,)
if output_router_logits and layer_outputs.router_logits is not None:
all_router_logits += (layer_outputs.router_logits,)
past_key_values[idx] = layer_outputs.cache_view
hidden_states = self.norm(hidden_states)
if output_hidden_states:
all_hidden_states += (hidden_states,)
return MoeModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_attentions,
router_logits=all_router_logits,
past_key_values=past_key_values,
)
[docs] def get_encoder(self):
raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self):
return self
[docs] def get_lm_head(self):
raise NotImplementedError("The base model does not have a language model head.")
[docs] def get_embedding(self):
return self.embed_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=Glm4MoeConfig, model_type="glm4_moe")
class Glm4MoeForCausalLM(BaseCausalLMModule[Glm4MoeModel, Glm4MoeConfig]):
"""GLM4 MoE model with a language modeling head for causal language modeling tasks."""
_task_type = TaskType.CAUSAL_LM
_model_type = "glm4_moe"
_config_class = Glm4MoeConfig
def __init__(
self,
config: Glm4MoeConfig,
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=Glm4MoeModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
router_aux_loss_coef=None, # NOTE: we dont use aux loss for Glm4Moe
)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=Glm4MoeConfig, model_type="glm4_moe")
class Glm4MoeForSequenceClassification(BaseSequenceClassificationModule[Glm4MoeModel, Glm4MoeConfig]):
"""GLM4 MoE model for sequence classification tasks."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "glm4_moe"
_config_class = Glm4MoeConfig
def __init__(
self,
config: Glm4MoeConfig,
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=Glm4MoeModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
score_bias=False,
)