[New Model] GLM4.7-Flash (#17247)
Co-authored-by: zRzRzRzRzRzRzR <2448370773@qq.com> Co-authored-by: JustinTong0323 <justinning0323@gmail.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
Qiaolin Yu
@@ -269,7 +269,10 @@ class ModelConfig:
|
||||
):
|
||||
self.hf_config.architectures[0] = "DeepseekV3ForCausalLMNextN"
|
||||
|
||||
if is_draft_model and self.hf_config.architectures[0] == "Glm4MoeForCausalLM":
|
||||
if is_draft_model and self.hf_config.architectures[0] in [
|
||||
"Glm4MoeForCausalLM",
|
||||
"Glm4MoeLiteForCausalLM",
|
||||
]:
|
||||
self.hf_config.architectures[0] = "Glm4MoeForCausalLMNextN"
|
||||
|
||||
if (
|
||||
@@ -375,6 +378,7 @@ class ModelConfig:
|
||||
or "DeepseekV32ForCausalLM" in self.hf_config.architectures
|
||||
or "DeepseekV3ForCausalLM" in self.hf_config.architectures
|
||||
or "DeepseekV3ForCausalLMNextN" in self.hf_config.architectures
|
||||
or "Glm4MoeLiteForCausalLM" in self.hf_config.architectures
|
||||
or "LongcatFlashForCausalLM" in self.hf_config.architectures
|
||||
or "LongcatFlashForCausalLMNextN" in self.hf_config.architectures
|
||||
or "DotsVLMForCausalLM" in self.hf_config.architectures
|
||||
@@ -394,15 +398,21 @@ class ModelConfig:
|
||||
else None
|
||||
)
|
||||
|
||||
# Handle rope scaling with yarn
|
||||
self.scaling = 1 / math.sqrt(self.qk_nope_head_dim + self.qk_rope_head_dim)
|
||||
if self.hf_config.rope_scaling:
|
||||
mscale_all_dim = self.hf_config.rope_scaling.get(
|
||||
"mscale_all_dim", False
|
||||
if "Glm4MoeLiteForCausalLM" in self.hf_config.architectures:
|
||||
self.scaling = 1
|
||||
self.hf_config.rope_scaling = None
|
||||
else:
|
||||
# Handle rope scaling with yarn
|
||||
self.scaling = 1 / math.sqrt(
|
||||
self.qk_nope_head_dim + self.qk_rope_head_dim
|
||||
)
|
||||
scaling_factor = self.hf_config.rope_scaling["factor"]
|
||||
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
||||
self.scaling = self.scaling * mscale * mscale
|
||||
if self.hf_config.rope_scaling:
|
||||
mscale_all_dim = self.hf_config.rope_scaling.get(
|
||||
"mscale_all_dim", False
|
||||
)
|
||||
scaling_factor = self.hf_config.rope_scaling["factor"]
|
||||
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
||||
self.scaling = self.scaling * mscale * mscale
|
||||
|
||||
elif "MiniCPM3ForCausalLM" in self.hf_config.architectures:
|
||||
self.head_dim = 128
|
||||
|
||||
@@ -454,6 +454,7 @@ class OpenAIServingChat(OpenAIServingBase):
|
||||
if request.chat_template_kwargs
|
||||
else {}
|
||||
),
|
||||
return_dict=False,
|
||||
)
|
||||
except Exception as e:
|
||||
# If the first attempt fails, try transforming the tools format
|
||||
@@ -476,6 +477,7 @@ class OpenAIServingChat(OpenAIServingBase):
|
||||
if request.chat_template_kwargs
|
||||
else {}
|
||||
),
|
||||
return_dict=False,
|
||||
)
|
||||
except jinja2.TemplateError as template_error:
|
||||
# Template errors (e.g., from raise_exception in Jinja templates)
|
||||
|
||||
@@ -17,7 +17,7 @@ from sglang.srt.models.deepseek_common.utils import (
|
||||
_use_aiter_gfx95,
|
||||
)
|
||||
from sglang.srt.server_args import get_global_server_args
|
||||
from sglang.srt.utils import BumpAllocator
|
||||
from sglang.srt.utils import BumpAllocator, next_power_of_2
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA
|
||||
@@ -472,7 +472,12 @@ class DeepseekMHAForwardMixin:
|
||||
):
|
||||
k = k_nope.new_empty(*k_shape)
|
||||
concat_mla_k(k=k, k_nope=k_nope, k_rope=k_pe)
|
||||
elif _is_cuda:
|
||||
elif (
|
||||
_is_cuda
|
||||
and next_power_of_2(self.num_local_heads) == self.num_local_heads
|
||||
and next_power_of_2(self.qk_nope_head_dim) == self.qk_nope_head_dim
|
||||
and next_power_of_2(self.qk_rope_head_dim) == self.qk_rope_head_dim
|
||||
):
|
||||
# fa3 mha support fp8 inputs
|
||||
if (
|
||||
self.current_attention_backend == "fa3"
|
||||
|
||||
@@ -685,6 +685,8 @@ class Glm4MoeDecoderLayer(nn.Module):
|
||||
attention_bias = config.attention_bias
|
||||
self.layer_id = layer_id
|
||||
|
||||
use_qk_norm = config.use_qk_norm if hasattr(config, "use_qk_norm") else False
|
||||
|
||||
self.self_attn = Glm4MoeAttention(
|
||||
hidden_size=self.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
@@ -699,7 +701,7 @@ class Glm4MoeDecoderLayer(nn.Module):
|
||||
attention_bias=attention_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("self_attn", prefix),
|
||||
use_qk_norm=config.use_qk_norm,
|
||||
use_qk_norm=use_qk_norm,
|
||||
alt_stream=alt_stream,
|
||||
)
|
||||
|
||||
|
||||
808
python/sglang/srt/models/glm4_moe_lite.py
Normal file
808
python/sglang/srt/models/glm4_moe_lite.py
Normal file
@@ -0,0 +1,808 @@
|
||||
# Copyright 2025-2026 SGLang Team
|
||||
# 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
|
||||
#
|
||||
# http://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.
|
||||
# ==============================================================================
|
||||
|
||||
"""Inference-only GLM-Lite model compatible with HuggingFace weights"""
|
||||
|
||||
import logging
|
||||
from typing import Iterable, Optional, Tuple
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from sglang.srt.batch_overlap.single_batch_overlap import SboFlags
|
||||
from sglang.srt.distributed import (
|
||||
get_moe_expert_parallel_world_size,
|
||||
get_pp_group,
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from sglang.srt.layers.activation import SiluAndMul
|
||||
from sglang.srt.layers.communicator import (
|
||||
LayerCommunicator,
|
||||
LayerScatterModes,
|
||||
enable_moe_dense_fully_dp,
|
||||
)
|
||||
from sglang.srt.layers.dp_attention import (
|
||||
get_attention_tp_size,
|
||||
is_dp_attention_enabled,
|
||||
)
|
||||
from sglang.srt.layers.layernorm import RMSNorm
|
||||
from sglang.srt.layers.linear import MergedColumnParallelLinear, RowParallelLinear
|
||||
from sglang.srt.layers.logits_processor import LogitsProcessor
|
||||
from sglang.srt.layers.moe import (
|
||||
get_moe_a2a_backend,
|
||||
should_use_flashinfer_cutlass_moe_fp4_allgather,
|
||||
)
|
||||
from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
|
||||
from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
|
||||
from sglang.srt.layers.moe.topk import TopK, TopKOutputFormat
|
||||
from sglang.srt.layers.quantization.base_config import QuantizationConfig
|
||||
from sglang.srt.layers.utils import PPMissingLayer
|
||||
from sglang.srt.layers.vocab_parallel_embedding import (
|
||||
ParallelLMHead,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
from sglang.srt.model_loader.weight_utils import default_weight_loader
|
||||
from sglang.srt.models.deepseek_v2 import (
|
||||
DeepseekV2AttentionMLA,
|
||||
DeepseekV2DecoderLayer,
|
||||
DeepseekV2ForCausalLM,
|
||||
DeepseekV2Model,
|
||||
DeepseekV2MoE,
|
||||
)
|
||||
from sglang.srt.server_args import get_global_server_args
|
||||
from sglang.srt.utils import (
|
||||
BumpAllocator,
|
||||
LazyValue,
|
||||
add_prefix,
|
||||
get_device_sm,
|
||||
is_cuda,
|
||||
log_info_on_rank0,
|
||||
make_layers,
|
||||
)
|
||||
|
||||
_is_cuda = is_cuda()
|
||||
_device_sm = get_device_sm()
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class Glm4MoeLiteMLP(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
hidden_act: str,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
reduce_results: bool = True,
|
||||
prefix: str = "",
|
||||
tp_rank: Optional[int] = None,
|
||||
tp_size: Optional[int] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.tp_size = tp_size
|
||||
|
||||
self.gate_up_proj = MergedColumnParallelLinear(
|
||||
hidden_size,
|
||||
[intermediate_size] * 2,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("gate_up_proj", prefix),
|
||||
tp_rank=tp_rank,
|
||||
tp_size=tp_size,
|
||||
)
|
||||
self.down_proj = RowParallelLinear(
|
||||
intermediate_size,
|
||||
hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
reduce_results=reduce_results,
|
||||
prefix=add_prefix("down_proj", prefix),
|
||||
tp_rank=tp_rank,
|
||||
tp_size=tp_size,
|
||||
)
|
||||
if hidden_act != "silu":
|
||||
raise ValueError(
|
||||
f"Unsupported activation: {hidden_act}. Only silu is supported for now."
|
||||
)
|
||||
self.act_fn = SiluAndMul()
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x,
|
||||
forward_batch=None,
|
||||
should_allreduce_fusion: bool = False,
|
||||
use_reduce_scatter: bool = False,
|
||||
gemm_output_zero_allocator: BumpAllocator = None,
|
||||
):
|
||||
# Keep parity with DeepseekV2MLP.forward signature since DeepseekV2DecoderLayer
|
||||
# invokes MLP modules with these extra arguments.
|
||||
if (self.tp_size == 1) and x.shape[0] == 0:
|
||||
return x
|
||||
|
||||
# Some quantization wrappers store the underlying parameter as `weight_packed`.
|
||||
if not hasattr(self.gate_up_proj, "weight"):
|
||||
self.gate_up_proj.weight = getattr(self.gate_up_proj, "weight_packed")
|
||||
if not hasattr(self.down_proj, "weight"):
|
||||
self.down_proj.weight = getattr(self.down_proj, "weight_packed")
|
||||
|
||||
if (
|
||||
gemm_output_zero_allocator is not None
|
||||
and x.shape[0] <= 256
|
||||
and self.gate_up_proj.weight.dtype == torch.uint8
|
||||
):
|
||||
y = gemm_output_zero_allocator.allocate(
|
||||
x.shape[0] * self.gate_up_proj.output_size_per_partition
|
||||
).view(x.shape[0], self.gate_up_proj.output_size_per_partition)
|
||||
x = (x, None, y)
|
||||
|
||||
gate_up, _ = self.gate_up_proj(x)
|
||||
x = self.act_fn(gate_up)
|
||||
x, _ = self.down_proj(
|
||||
x,
|
||||
skip_all_reduce=should_allreduce_fusion or use_reduce_scatter,
|
||||
)
|
||||
return x
|
||||
|
||||
|
||||
class Glm4MoeLiteGate(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
prefix: str = "",
|
||||
is_nextn: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
self.is_nextn = is_nextn
|
||||
self.weight = nn.Parameter(
|
||||
torch.empty((config.n_routed_experts, config.hidden_size))
|
||||
)
|
||||
self.e_score_correction_bias = nn.Parameter(
|
||||
torch.empty((config.n_routed_experts), dtype=torch.float32)
|
||||
)
|
||||
|
||||
def forward(self, hidden_states, gemm_output_zero_allocator: BumpAllocator = None):
|
||||
# NOTE: For some unknown reason, router_gemm seems degrade accept length.
|
||||
if (
|
||||
_is_cuda
|
||||
and not self.is_nextn
|
||||
and hidden_states.shape[0] < 4
|
||||
and hidden_states.shape[1] == 7168
|
||||
and self.weight.shape[0] == 256
|
||||
and _device_sm >= 90
|
||||
):
|
||||
from sgl_kernel import dsv3_router_gemm
|
||||
|
||||
logits = dsv3_router_gemm(hidden_states, self.weight).to(
|
||||
hidden_states.dtype
|
||||
)
|
||||
else:
|
||||
logits = F.linear(hidden_states, self.weight, None)
|
||||
|
||||
return logits
|
||||
|
||||
|
||||
class Glm4MoeLiteSparseMoeBlock(DeepseekV2MoE):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
layer_id: int,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
alt_stream: Optional[torch.cuda.Stream] = None,
|
||||
is_nextn: bool = False,
|
||||
):
|
||||
nn.Module.__init__(self)
|
||||
self.tp_size = get_tensor_model_parallel_world_size()
|
||||
self.routed_scaling_factor = config.routed_scaling_factor
|
||||
self.n_shared_experts = config.n_shared_experts
|
||||
self.num_fused_shared_experts = (
|
||||
0
|
||||
if get_global_server_args().disable_shared_experts_fusion
|
||||
else config.n_shared_experts
|
||||
)
|
||||
self.config = config
|
||||
self.layer_id = layer_id
|
||||
self.alt_stream = alt_stream
|
||||
self.is_nextn = is_nextn
|
||||
|
||||
if self.tp_size > config.n_routed_experts:
|
||||
raise ValueError(
|
||||
f"Tensor parallel size {self.tp_size} is greater than "
|
||||
f"the number of experts {config.n_routed_experts}."
|
||||
)
|
||||
|
||||
if config.hidden_act != "silu":
|
||||
raise ValueError(
|
||||
f"Unsupported activation: {config.hidden_act}. "
|
||||
"Only silu is supported for now."
|
||||
)
|
||||
|
||||
self.gate = Glm4MoeLiteGate(
|
||||
config=config, prefix=add_prefix("gate", prefix), is_nextn=is_nextn
|
||||
)
|
||||
|
||||
self.experts = get_moe_impl_class(quant_config)(
|
||||
num_experts=config.n_routed_experts
|
||||
+ self.num_fused_shared_experts
|
||||
+ get_global_server_args().ep_num_redundant_experts,
|
||||
num_fused_shared_experts=self.num_fused_shared_experts,
|
||||
top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.moe_intermediate_size,
|
||||
layer_id=self.layer_id,
|
||||
quant_config=quant_config,
|
||||
routed_scaling_factor=self.routed_scaling_factor,
|
||||
prefix=add_prefix("experts", prefix),
|
||||
)
|
||||
|
||||
self.topk = TopK(
|
||||
top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
|
||||
layer_id=self.layer_id,
|
||||
renormalize=config.norm_topk_prob,
|
||||
use_grouped_topk=True,
|
||||
num_expert_group=config.n_group,
|
||||
num_fused_shared_experts=self.num_fused_shared_experts,
|
||||
topk_group=config.topk_group,
|
||||
correction_bias=self.gate.e_score_correction_bias,
|
||||
quant_config=quant_config,
|
||||
routed_scaling_factor=self.routed_scaling_factor,
|
||||
apply_routed_scaling_factor_on_output=self.experts.should_fuse_routed_scaling_factor_in_topk,
|
||||
# Some Fp4 MoE backends require the output format to be bypassed but the MTP layers are unquantized
|
||||
# and requires the output format to be standard. We use quant_config to determine the output format.
|
||||
output_format=TopKOutputFormat.STANDARD if quant_config is None else None,
|
||||
)
|
||||
|
||||
self.shared_experts_is_int8 = False
|
||||
self.shared_experts_is_fp8 = False
|
||||
# self.shared_experts_weight_block_size = None
|
||||
if config.n_shared_experts is not None and self.num_fused_shared_experts == 0:
|
||||
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
|
||||
# disable tp for shared experts when enable deepep moe, or with fp4 allgather
|
||||
self.shared_experts = Glm4MoeLiteMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
reduce_results=False,
|
||||
prefix=add_prefix("shared_experts", prefix),
|
||||
**(
|
||||
dict(tp_rank=0, tp_size=1)
|
||||
if get_moe_a2a_backend().is_deepep()
|
||||
or get_moe_a2a_backend().is_mooncake()
|
||||
or should_use_flashinfer_cutlass_moe_fp4_allgather()
|
||||
else {}
|
||||
),
|
||||
)
|
||||
is_packed_weight = hasattr(
|
||||
self.shared_experts.gate_up_proj.quant_method, "quant_config"
|
||||
)
|
||||
self.shared_experts_is_int8 = (
|
||||
not is_packed_weight
|
||||
and self.shared_experts.gate_up_proj.weight.dtype == torch.int8
|
||||
)
|
||||
self.shared_experts_is_fp8 = (
|
||||
not is_packed_weight
|
||||
and self.shared_experts.gate_up_proj.weight.dtype == torch.float8_e4m3fn
|
||||
)
|
||||
|
||||
self.top_k = config.num_experts_per_tok
|
||||
|
||||
if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake():
|
||||
# TODO: we will support tp < ep in the future
|
||||
self.ep_size = get_moe_expert_parallel_world_size()
|
||||
self.num_experts = (
|
||||
config.n_routed_experts
|
||||
+ get_global_server_args().ep_num_redundant_experts
|
||||
)
|
||||
self.renormalize = config.norm_topk_prob
|
||||
self.topk_group = config.topk_group
|
||||
self.num_expert_group = config.n_group
|
||||
self.correction_bias = (
|
||||
self.gate.e_score_correction_bias.data
|
||||
if self.gate.e_score_correction_bias is not None
|
||||
else None
|
||||
)
|
||||
|
||||
self._enable_a2a_moe = (
|
||||
get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake()
|
||||
)
|
||||
self._fuse_shared_experts_inside_sbo = SboFlags.fuse_shared_experts_inside_sbo()
|
||||
|
||||
|
||||
class Glm4MoeLiteDecoderLayer(DeepseekV2DecoderLayer):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
layer_id: int,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
is_nextn: bool = False,
|
||||
prefix: str = "",
|
||||
alt_stream: Optional[torch.cuda.Stream] = None,
|
||||
) -> None:
|
||||
nn.Module.__init__(self)
|
||||
self.hidden_size = config.hidden_size
|
||||
self.config = config
|
||||
|
||||
from sglang.srt.layers.attention.nsa.utils import is_nsa_enable_prefill_cp
|
||||
|
||||
self.nsa_enable_prefill_cp = is_nsa_enable_prefill_cp()
|
||||
rope_theta = 1000000
|
||||
rope_scaling = None
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings", 202752)
|
||||
self.layer_id = layer_id
|
||||
|
||||
self.self_attn = DeepseekV2AttentionMLA(
|
||||
config=config,
|
||||
hidden_size=config.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
qk_nope_head_dim=config.qk_nope_head_dim,
|
||||
qk_rope_head_dim=config.qk_rope_head_dim,
|
||||
v_head_dim=config.v_head_dim,
|
||||
q_lora_rank=config.q_lora_rank,
|
||||
kv_lora_rank=config.kv_lora_rank,
|
||||
rope_theta=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
quant_config=quant_config,
|
||||
reduce_results=False,
|
||||
layer_id=layer_id,
|
||||
prefix=add_prefix("self_attn", prefix),
|
||||
)
|
||||
|
||||
self.is_layer_sparse = self._is_layer_sparse(layer_id, is_nextn=is_nextn)
|
||||
is_previous_layer_sparse = self._is_layer_sparse(layer_id - 1, is_nextn=False)
|
||||
is_next_layer_sparse = self._is_layer_sparse(layer_id + 1, is_nextn=False)
|
||||
|
||||
self.layer_scatter_modes = LayerScatterModes.init_new(
|
||||
layer_id=layer_id,
|
||||
num_layers=1 if is_nextn else config.num_hidden_layers,
|
||||
is_layer_sparse=self.is_layer_sparse,
|
||||
is_previous_layer_sparse=is_previous_layer_sparse,
|
||||
is_next_layer_sparse=is_next_layer_sparse,
|
||||
)
|
||||
|
||||
if self.is_layer_sparse:
|
||||
self.mlp = Glm4MoeLiteSparseMoeBlock(
|
||||
config=config,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("mlp", prefix),
|
||||
layer_id=self.layer_id,
|
||||
alt_stream=alt_stream,
|
||||
is_nextn=is_nextn,
|
||||
)
|
||||
else:
|
||||
if enable_moe_dense_fully_dp():
|
||||
mlp_tp_rank, mlp_tp_size = 0, 1
|
||||
else:
|
||||
mlp_tp_rank, mlp_tp_size = None, None
|
||||
self.mlp = Glm4MoeLiteMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("mlp", prefix),
|
||||
tp_rank=mlp_tp_rank,
|
||||
tp_size=mlp_tp_size,
|
||||
)
|
||||
|
||||
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.post_attention_layernorm = RMSNorm(
|
||||
config.hidden_size, eps=config.rms_norm_eps
|
||||
)
|
||||
|
||||
self.layer_communicator = LayerCommunicator(
|
||||
layer_scatter_modes=self.layer_scatter_modes,
|
||||
input_layernorm=self.input_layernorm,
|
||||
post_attention_layernorm=self.post_attention_layernorm,
|
||||
allow_reduce_scatter=True,
|
||||
is_last_layer=(
|
||||
is_nextn or (self.layer_id == self.config.num_hidden_layers - 1)
|
||||
),
|
||||
qkv_latent_func=self.self_attn.prepare_qkv_latent,
|
||||
)
|
||||
|
||||
|
||||
class Glm4MoeLiteModel(DeepseekV2Model):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
nn.Module.__init__(self)
|
||||
self.padding_id = config.pad_token_id
|
||||
self.vocab_size = config.vocab_size
|
||||
self.first_k_dense_replace = config.first_k_dense_replace
|
||||
self.pp_group = get_pp_group()
|
||||
|
||||
# DeepseekV2Model.forward expects these attributes to exist.
|
||||
from sglang.srt.layers.attention.nsa.utils import is_nsa_enable_prefill_cp
|
||||
|
||||
self.nsa_enable_prefill_cp = is_nsa_enable_prefill_cp()
|
||||
self.cp_size = get_attention_tp_size() if self.nsa_enable_prefill_cp else None
|
||||
self.gemm_output_zero_allocator_size = 0
|
||||
self.llama_4_scaling_config = getattr(config, "llama_4_scaling", None)
|
||||
|
||||
if self.pp_group.is_first_rank:
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
enable_tp=not is_dp_attention_enabled(),
|
||||
)
|
||||
else:
|
||||
self.embed_tokens = PPMissingLayer()
|
||||
|
||||
self.alt_stream = torch.cuda.Stream() if _is_cuda else None
|
||||
self.layers, self.start_layer, self.end_layer = make_layers(
|
||||
config.num_hidden_layers,
|
||||
lambda idx, prefix: Glm4MoeLiteDecoderLayer(
|
||||
config=config,
|
||||
layer_id=idx,
|
||||
quant_config=quant_config,
|
||||
prefix=prefix,
|
||||
alt_stream=self.alt_stream,
|
||||
),
|
||||
pp_rank=self.pp_group.rank_in_group,
|
||||
pp_size=self.pp_group.world_size,
|
||||
prefix=add_prefix("layers", prefix),
|
||||
)
|
||||
if self.pp_group.is_last_rank:
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
self.norm = PPMissingLayer(return_tuple=True)
|
||||
self.layers_to_capture = []
|
||||
|
||||
|
||||
class Glm4MoeLiteForCausalLM(DeepseekV2ForCausalLM):
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
nn.Module.__init__(self)
|
||||
config.moe_layer_freq = 1
|
||||
self.config = config
|
||||
self.tp_size = get_tensor_model_parallel_world_size()
|
||||
self.quant_config = quant_config
|
||||
self.pp_group = get_pp_group()
|
||||
self.determine_num_fused_shared_experts("Glm4MoeLiteForCausalLM")
|
||||
self.model = Glm4MoeLiteModel(
|
||||
config, quant_config, prefix=add_prefix("model", prefix)
|
||||
)
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
quant_config=quant_config,
|
||||
prefix=add_prefix("lm_head", prefix),
|
||||
use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(config)
|
||||
|
||||
self._routed_experts_weights_of_layer = LazyValue(
|
||||
lambda: {
|
||||
layer_id: layer.mlp.get_moe_weights()
|
||||
for layer_id, layer in enumerate(self.model.layers)
|
||||
if isinstance(layer.mlp, Glm4MoeLiteSparseMoeBlock)
|
||||
}
|
||||
)
|
||||
self.capture_aux_hidden_states = False
|
||||
|
||||
from sglang.srt.layers.attention.nsa.utils import is_nsa_enable_prefill_cp
|
||||
|
||||
self.nsa_enable_prefill_cp = is_nsa_enable_prefill_cp()
|
||||
if self.nsa_enable_prefill_cp:
|
||||
from sglang.srt.layers.dp_attention import (
|
||||
get_attention_tp_rank,
|
||||
get_attention_tp_size,
|
||||
)
|
||||
|
||||
self.cp_rank = get_attention_tp_rank()
|
||||
self.cp_size = get_attention_tp_size()
|
||||
else:
|
||||
self.cp_rank = self.cp_size = None
|
||||
|
||||
def determine_num_fused_shared_experts(
|
||||
self, architecture: str = "Glm4MoeLiteForCausalLM"
|
||||
):
|
||||
self.num_fused_shared_experts = 0
|
||||
if get_global_server_args().disable_shared_experts_fusion:
|
||||
return
|
||||
|
||||
disable_reason = None
|
||||
if (
|
||||
not _is_cuda
|
||||
or torch.cuda.get_device_capability("cuda") < (8, 0)
|
||||
or self.config.architectures[0] != architecture
|
||||
or self.config.n_shared_experts != 1
|
||||
):
|
||||
disable_reason = "Only GLM-4.5 or GLM-4.6 on NV-platform with capability >= 80 can use shared experts fusion optimization."
|
||||
elif get_moe_expert_parallel_world_size() > 1:
|
||||
disable_reason = "GLM-4.5 or GLM-4.6 can not use shared experts fusion optimization under expert parallelism."
|
||||
|
||||
if disable_reason is not None:
|
||||
get_global_server_args().disable_shared_experts_fusion = True
|
||||
self.num_fused_shared_experts = 0
|
||||
log_info_on_rank0(
|
||||
logger,
|
||||
f"{disable_reason} Shared experts fusion optimization is disabled.",
|
||||
)
|
||||
return
|
||||
|
||||
self.num_fused_shared_experts = self.config.n_shared_experts
|
||||
|
||||
def load_weights(
|
||||
self,
|
||||
weights: Iterable[Tuple[str, torch.Tensor]],
|
||||
is_nextn=False,
|
||||
params_dict=None,
|
||||
is_eagle=False,
|
||||
):
|
||||
if is_nextn:
|
||||
if hasattr(self.config, "num_nextn_predict_layers"):
|
||||
num_nextn_layers = self.config.num_nextn_predict_layers
|
||||
assert num_nextn_layers == 1, "Only 1 nextn layer is supported"
|
||||
# compatible with old design
|
||||
nextn_layer_id = (
|
||||
0
|
||||
if self.config.num_hidden_layers == 1
|
||||
else self.config.num_hidden_layers
|
||||
)
|
||||
else:
|
||||
raise ValueError("num_nextn_predict_layers is not in the config")
|
||||
|
||||
stacked_params_mapping = [
|
||||
# (param_name, shard_name, shard_id)
|
||||
("qkv_proj", "q_proj", "q"),
|
||||
("qkv_proj", "k_proj", "k"),
|
||||
("qkv_proj", "v_proj", "v"),
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
]
|
||||
|
||||
if self.num_fused_shared_experts > 0:
|
||||
assert self.num_fused_shared_experts == 1
|
||||
|
||||
def iter_weights_with_fused_shared_experts(
|
||||
weights: Iterable[Tuple[str, torch.Tensor]],
|
||||
) -> Iterable[Tuple[str, torch.Tensor]]:
|
||||
import re
|
||||
|
||||
pattern = re.compile(
|
||||
r"^model\.layers\.(\d+)\.mlp\.shared_experts\.(.+)$"
|
||||
)
|
||||
for name, weight in weights:
|
||||
match = pattern.match(name)
|
||||
if match:
|
||||
layer_id = int(match.group(1))
|
||||
suffix = match.group(2)
|
||||
name = f"model.layers.{layer_id}.mlp.experts.{self.config.n_routed_experts}.{suffix}"
|
||||
yield name, weight
|
||||
|
||||
weights = iter_weights_with_fused_shared_experts(weights)
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
ckpt_up_proj_name="up_proj",
|
||||
num_experts=self.config.n_routed_experts + self.num_fused_shared_experts,
|
||||
)
|
||||
|
||||
# Fuse q_a_proj and kv_a_proj_with_mqa along output dimension when q_lora_rank is not None
|
||||
fuse_qkv_a_proj = hasattr(self.config, "q_lora_rank") and (
|
||||
self.config.q_lora_rank is not None
|
||||
)
|
||||
cached_a_proj = {} if fuse_qkv_a_proj else None
|
||||
|
||||
if is_nextn:
|
||||
nextn_layer_prefix = f"model.layers.{nextn_layer_id}"
|
||||
nextn_spec_weight_names = [
|
||||
"shared_head.norm",
|
||||
"eh_proj",
|
||||
"enorm",
|
||||
"hnorm",
|
||||
]
|
||||
else:
|
||||
nextn_layer_prefix = None
|
||||
nextn_spec_weight_names = []
|
||||
|
||||
eagle_ignore_weight_names = []
|
||||
if is_eagle:
|
||||
eagle_ignore_weight_names = [
|
||||
"eagle_draft_tokens_map",
|
||||
"eagle_lm_head.weight",
|
||||
]
|
||||
|
||||
if params_dict is None:
|
||||
params_dict = dict(self.named_parameters())
|
||||
|
||||
weight_names = []
|
||||
for name, loaded_weight in weights:
|
||||
weight_names.append(name)
|
||||
|
||||
if not is_nextn and not is_eagle:
|
||||
if hasattr(self.config, "num_nextn_predict_layers"):
|
||||
num_nextn_layers = self.config.num_nextn_predict_layers
|
||||
if num_nextn_layers > 0 and name.startswith("model.layers"):
|
||||
name_list = name.split(".")
|
||||
if (
|
||||
len(name_list) >= 3
|
||||
and int(name_list[2]) >= self.config.num_hidden_layers
|
||||
):
|
||||
continue
|
||||
else:
|
||||
if nextn_layer_prefix and not name.startswith(nextn_layer_prefix):
|
||||
continue
|
||||
|
||||
if nextn_layer_prefix is not None: # mtp
|
||||
# Use shared head and embed weights from target model
|
||||
if "shared_head.head" in name or "embed_tokens" in name:
|
||||
continue
|
||||
|
||||
is_decoder = True
|
||||
# For nextn specific weights
|
||||
for weight_name in nextn_spec_weight_names:
|
||||
if weight_name in name:
|
||||
name = name.replace(nextn_layer_prefix, "model")
|
||||
is_decoder = False
|
||||
break
|
||||
# For decoder layer weights
|
||||
if is_decoder:
|
||||
name = name.replace(nextn_layer_prefix, "model.decoder")
|
||||
|
||||
if "rotary_emb.inv_freq" in name:
|
||||
continue
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
# Skip non-stacked layers and experts (experts handled below).
|
||||
if weight_name not in name:
|
||||
continue
|
||||
# We have mlp.experts[0].gate_proj in the checkpoint.
|
||||
# Since we handle the experts below in expert_params_mapping,
|
||||
# we need to skip here BEFORE we update the name, otherwise
|
||||
# name will be updated to mlp.experts[0].gate_up_proj, which
|
||||
# will then be updated below in expert_params_mapping
|
||||
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
|
||||
if "mlp.experts" in name:
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
if name not in params_dict:
|
||||
continue
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
else:
|
||||
# Track if this is an expert weight to enable early skipping
|
||||
is_expert_weight = False
|
||||
|
||||
for mapping in expert_params_mapping:
|
||||
param_name, weight_name, expert_id, shard_id = mapping
|
||||
if weight_name not in name:
|
||||
continue
|
||||
|
||||
# Mark as expert weight regardless of whether we can process it
|
||||
is_expert_weight = True
|
||||
|
||||
name = name.replace(weight_name, param_name)
|
||||
if name not in params_dict:
|
||||
# Expert weight not on this rank, will be skipped below
|
||||
continue
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(
|
||||
param,
|
||||
loaded_weight,
|
||||
name,
|
||||
shard_id=shard_id,
|
||||
expert_id=expert_id,
|
||||
)
|
||||
break
|
||||
else:
|
||||
if is_expert_weight:
|
||||
# This is an expert weight but not mapped to this rank, skip all remaining processing
|
||||
continue
|
||||
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
if name in eagle_ignore_weight_names:
|
||||
continue
|
||||
|
||||
# GLM NOTE: for MLA
|
||||
if fuse_qkv_a_proj and (
|
||||
"q_a_proj" in name or "kv_a_proj_with_mqa" in name
|
||||
):
|
||||
cached_a_proj[name] = loaded_weight
|
||||
q_a_proj_name = (
|
||||
name
|
||||
if "q_a_proj" in name
|
||||
else name.replace("kv_a_proj_with_mqa", "q_a_proj")
|
||||
)
|
||||
kv_a_proj_name = (
|
||||
name
|
||||
if "kv_a_proj_with_mqa" in name
|
||||
else name.replace("q_a_proj", "kv_a_proj_with_mqa")
|
||||
)
|
||||
|
||||
# When both q_a_proj and kv_a_proj_with_mqa has been cached, load the fused weight to parameter
|
||||
if (
|
||||
q_a_proj_name in cached_a_proj
|
||||
and kv_a_proj_name in cached_a_proj
|
||||
):
|
||||
q_a_proj_weight = cached_a_proj[q_a_proj_name]
|
||||
kv_a_proj_weight = cached_a_proj[kv_a_proj_name]
|
||||
fused_weight = torch.cat(
|
||||
[q_a_proj_weight, kv_a_proj_weight], dim=0
|
||||
)
|
||||
param_name = (
|
||||
name.replace("q_a_proj", "fused_qkv_a_proj_with_mqa")
|
||||
if "q_a_proj" in name
|
||||
else name.replace(
|
||||
"kv_a_proj_with_mqa", "fused_qkv_a_proj_with_mqa"
|
||||
)
|
||||
)
|
||||
if param_name not in params_dict:
|
||||
continue
|
||||
param = params_dict[param_name]
|
||||
|
||||
weight_loader = getattr(
|
||||
param, "weight_loader", default_weight_loader
|
||||
)
|
||||
weight_loader(param, fused_weight)
|
||||
cached_a_proj.pop(q_a_proj_name)
|
||||
cached_a_proj.pop(kv_a_proj_name)
|
||||
else:
|
||||
if (
|
||||
"k_scale" in name or "v_scale" in name
|
||||
) and name not in params_dict:
|
||||
# modelopt attn kv scale is named differently
|
||||
if any(scale in name for scale in ["k_scale", "v_scale"]):
|
||||
name = name.replace("_proj", "attn_mqa")
|
||||
else:
|
||||
logger.warning(
|
||||
f"Unknown scale found in checkpoint: {name}"
|
||||
)
|
||||
|
||||
if name not in params_dict:
|
||||
continue
|
||||
|
||||
if name in params_dict.keys():
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(
|
||||
param, "weight_loader", default_weight_loader
|
||||
)
|
||||
weight_loader(param, loaded_weight)
|
||||
else:
|
||||
logger.warning(f"Parameter {name} not found in params_dict")
|
||||
|
||||
# DeepseekV2AttentionMLA.forward_* expects post_load_weights() to populate
|
||||
# per-layer packed weights like `w_kc`/`w_vc` (used during CUDA graph capture).
|
||||
# GLM-Lite configs may not set `config.mla`, but this model always uses
|
||||
# DeepseekV2AttentionMLA, so we must run the post-load processing.
|
||||
# Use weight_names=None to ensure we always process all layers. Some checkpoints /
|
||||
# naming schemes may not include "kv_b_proj" in `weight_names`, but `w_kc`/`w_vc`
|
||||
# are still required by DeepseekV2AttentionMLA at runtime.
|
||||
self.post_load_weights(is_nextn=is_nextn, weight_names=None)
|
||||
|
||||
|
||||
EntryClass = [Glm4MoeLiteForCausalLM]
|
||||
@@ -1586,6 +1586,7 @@ class ServerArgs:
|
||||
"DeepseekV3ForCausalLM",
|
||||
"GptOssForCausalLM",
|
||||
"Glm4MoeForCausalLM",
|
||||
"Glm4MoeLiteForCausalLM",
|
||||
"Qwen3MoeForCausalLM",
|
||||
]
|
||||
and (is_sm90_supported() or is_sm100_supported())
|
||||
@@ -2147,6 +2148,7 @@ class ServerArgs:
|
||||
"DeepseekV32ForCausalLM",
|
||||
"DeepseekV3ForCausalLM",
|
||||
"Glm4MoeForCausalLM",
|
||||
"Glm4MoeLiteForCausalLM",
|
||||
"BailingMoeForCausalLM",
|
||||
"BailingMoeV2ForCausalLM",
|
||||
"MistralLarge3ForCausalLM",
|
||||
@@ -5422,6 +5424,7 @@ def auto_choose_speculative_params(self: ServerArgs):
|
||||
"DeepseekV2ForCausalLM",
|
||||
"GptOssForCausalLM",
|
||||
"Glm4MoeForCausalLM",
|
||||
"Glm4MoeLiteForCausalLM",
|
||||
"BailingMoeForCausalLM",
|
||||
"BailingMoeV2ForCausalLM",
|
||||
"MistralLarge3ForCausalLM",
|
||||
|
||||
Reference in New Issue
Block a user