767 lines
28 KiB
Python
767 lines
28 KiB
Python
# Copyright 2023-2024 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.
|
|
# ==============================================================================
|
|
|
|
# Adapted from
|
|
# https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/qwen2_moe.py
|
|
"""Inference-only Qwen2MoE model compatible with HuggingFace weights."""
|
|
|
|
import logging
|
|
from dataclasses import dataclass
|
|
from enum import Enum, auto
|
|
from typing import Any, Dict, Iterable, Optional, Tuple, Union
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from torch import nn
|
|
from transformers import PretrainedConfig
|
|
|
|
from sglang.srt.distributed import (
|
|
get_pp_group,
|
|
get_tensor_model_parallel_world_size,
|
|
tensor_model_parallel_all_reduce,
|
|
)
|
|
from sglang.srt.layers.activation import SiluAndMul
|
|
from sglang.srt.layers.dp_attention import (
|
|
attn_tp_all_gather,
|
|
attn_tp_reduce_scatter,
|
|
dp_gather_partial,
|
|
dp_scatter,
|
|
get_attention_tp_rank,
|
|
get_attention_tp_size,
|
|
get_local_attention_dp_size,
|
|
)
|
|
from sglang.srt.layers.layernorm import RMSNorm
|
|
from sglang.srt.layers.linear import (
|
|
MergedColumnParallelLinear,
|
|
QKVParallelLinear,
|
|
ReplicatedLinear,
|
|
RowParallelLinear,
|
|
)
|
|
from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
|
|
from sglang.srt.layers.moe.ep_moe.layer import EPMoE
|
|
from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
|
|
from sglang.srt.layers.quantization.base_config import QuantizationConfig
|
|
from sglang.srt.layers.radix_attention import RadixAttention
|
|
from sglang.srt.layers.rotary_embedding import get_rope
|
|
from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
|
|
from sglang.srt.layers.vocab_parallel_embedding import (
|
|
ParallelLMHead,
|
|
VocabParallelEmbedding,
|
|
)
|
|
from sglang.srt.managers.expert_distribution import (
|
|
ExpertDistributionRecorder,
|
|
get_global_expert_distribution_recorder,
|
|
)
|
|
from sglang.srt.managers.expert_location import ModelConfigForExpertLocation
|
|
from sglang.srt.managers.schedule_batch import global_server_args_dict
|
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
|
|
from sglang.srt.model_loader.weight_utils import default_weight_loader
|
|
from sglang.srt.utils import add_prefix, make_layers
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class Qwen2MoeMLP(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 = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.gate_up_proj = MergedColumnParallelLinear(
|
|
hidden_size,
|
|
[intermediate_size] * 2,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("gate_up_proj", prefix),
|
|
)
|
|
self.down_proj = RowParallelLinear(
|
|
intermediate_size,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
reduce_results=reduce_results,
|
|
prefix=add_prefix("down_proj", prefix),
|
|
)
|
|
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):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.down_proj(x)
|
|
return x
|
|
|
|
|
|
class Qwen2MoeSparseMoeBlock(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.tp_size = get_tensor_model_parallel_world_size()
|
|
|
|
if self.tp_size > config.num_experts:
|
|
raise ValueError(
|
|
f"Tensor parallel size {self.tp_size} is greater than "
|
|
f"the number of experts {config.num_experts}."
|
|
)
|
|
|
|
MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
|
|
|
|
self.experts = MoEImpl(
|
|
num_experts=config.num_experts,
|
|
top_k=config.num_experts_per_tok,
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.moe_intermediate_size,
|
|
renormalize=config.norm_topk_prob,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("experts", prefix),
|
|
)
|
|
|
|
self.gate = ReplicatedLinear(
|
|
config.hidden_size,
|
|
config.num_experts,
|
|
bias=False,
|
|
quant_config=None,
|
|
prefix=add_prefix("gate", prefix),
|
|
)
|
|
if config.shared_expert_intermediate_size > 0:
|
|
self.shared_expert = Qwen2MoeMLP(
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.shared_expert_intermediate_size,
|
|
hidden_act=config.hidden_act,
|
|
quant_config=quant_config,
|
|
reduce_results=False,
|
|
prefix=add_prefix("shared_expert", prefix),
|
|
)
|
|
else:
|
|
self.shared_expert = None
|
|
self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
num_tokens, hidden_dim = hidden_states.shape
|
|
hidden_states = hidden_states.view(-1, hidden_dim)
|
|
shared_output = None
|
|
if self.shared_expert is not None:
|
|
shared_output = self.shared_expert(hidden_states)
|
|
if self.shared_expert_gate is not None:
|
|
shared_output = (
|
|
F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_output
|
|
)
|
|
|
|
# router_logits: (num_tokens, n_experts)
|
|
router_logits, _ = self.gate(hidden_states)
|
|
final_hidden_states = self.experts(
|
|
hidden_states=hidden_states, router_logits=router_logits
|
|
)
|
|
if shared_output is not None:
|
|
final_hidden_states = final_hidden_states + shared_output
|
|
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
|
|
|
|
return final_hidden_states.view(num_tokens, hidden_dim)
|
|
|
|
|
|
class Qwen2MoeAttention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
layer_id: int = 0,
|
|
rope_theta: float = 10000,
|
|
rope_scaling: Optional[Dict[str, Any]] = None,
|
|
max_position_embeddings: int = 8192,
|
|
qkv_bias: int = True,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
|
|
attn_tp_rank = get_attention_tp_rank()
|
|
attn_tp_size = get_attention_tp_size()
|
|
|
|
self.total_num_heads = num_heads
|
|
assert self.total_num_heads % attn_tp_size == 0
|
|
self.num_heads = self.total_num_heads // attn_tp_size
|
|
self.total_num_kv_heads = num_kv_heads
|
|
if self.total_num_kv_heads >= attn_tp_size:
|
|
# Number of KV heads is greater than TP size, so we partition
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert self.total_num_kv_heads % attn_tp_size == 0
|
|
else:
|
|
# Number of KV heads is less than TP size, so we replicate
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert attn_tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
|
|
self.head_dim = hidden_size // self.total_num_heads
|
|
self.q_size = self.num_heads * self.head_dim
|
|
self.kv_size = self.num_kv_heads * self.head_dim
|
|
self.scaling = self.head_dim**-0.5
|
|
self.rope_theta = rope_theta
|
|
self.max_position_embeddings = max_position_embeddings
|
|
|
|
self.qkv_proj = QKVParallelLinear(
|
|
hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=qkv_bias,
|
|
quant_config=quant_config,
|
|
tp_rank=attn_tp_rank,
|
|
tp_size=attn_tp_size,
|
|
prefix=add_prefix("qkv_proj", prefix),
|
|
)
|
|
|
|
self.o_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
tp_rank=attn_tp_rank,
|
|
tp_size=attn_tp_size,
|
|
reduce_results=False,
|
|
prefix=add_prefix("o_proj", prefix),
|
|
)
|
|
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
rotary_dim=self.head_dim,
|
|
max_position=max_position_embeddings,
|
|
base=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
)
|
|
self.attn = RadixAttention(
|
|
self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
layer_id=layer_id,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("attn", prefix),
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.qkv_proj(hidden_states)
|
|
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
|
q, k = self.rotary_emb(positions, q, k)
|
|
attn_output = self.attn(q, k, v, forward_batch)
|
|
output, _ = self.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class _FFNInputMode(Enum):
|
|
# The MLP sublayer requires 1/tp_size tokens as input
|
|
SCATTERED = auto()
|
|
# The MLP sublayer requires all tokens as input
|
|
FULL = auto()
|
|
|
|
|
|
@dataclass
|
|
class _DecoderLayerInfo:
|
|
is_sparse: bool
|
|
ffn_input_mode: _FFNInputMode
|
|
|
|
|
|
class Qwen2MoeDecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
layer_id: int,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
rope_theta = getattr(config, "rope_theta", 10000)
|
|
rope_scaling = getattr(config, "rope_scaling", None)
|
|
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
|
qkv_bias = getattr(config, "qkv_bias", True)
|
|
self.self_attn = Qwen2MoeAttention(
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
layer_id=layer_id,
|
|
rope_theta=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
max_position_embeddings=max_position_embeddings,
|
|
quant_config=quant_config,
|
|
qkv_bias=qkv_bias,
|
|
prefix=add_prefix("self_attn", prefix),
|
|
)
|
|
|
|
self.layer_id = layer_id
|
|
|
|
self.attn_tp_size = get_attention_tp_size()
|
|
self.attn_tp_rank = get_attention_tp_rank()
|
|
self.local_dp_size = get_local_attention_dp_size()
|
|
|
|
self.info = self._compute_info(config, layer_id=layer_id)
|
|
previous_layer_info = self._compute_info(config, layer_id=layer_id - 1)
|
|
self.input_is_scattered = (
|
|
layer_id > 0
|
|
and previous_layer_info.ffn_input_mode == _FFNInputMode.SCATTERED
|
|
)
|
|
self.is_last_layer = self.layer_id == config.num_hidden_layers - 1
|
|
|
|
if self.info.is_sparse:
|
|
self.mlp = Qwen2MoeSparseMoeBlock(
|
|
config=config,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("mlp", prefix),
|
|
)
|
|
else:
|
|
self.mlp = Qwen2MoeMLP(
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
hidden_act=config.hidden_act,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("mlp", prefix),
|
|
)
|
|
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
|
|
)
|
|
|
|
@staticmethod
|
|
def _enable_moe_dense_fully_dp():
|
|
return global_server_args_dict["moe_dense_tp_size"] == 1
|
|
|
|
@staticmethod
|
|
def _compute_info(config: PretrainedConfig, layer_id: int):
|
|
# WARN: Qwen2MOE has no dense_layer, it is only for compatibility.
|
|
mlp_only_layers = (
|
|
[] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
|
|
)
|
|
is_sparse = (layer_id not in mlp_only_layers) and (
|
|
config.num_experts > 0 and (layer_id + 1) % config.decoder_sparse_step == 0
|
|
)
|
|
ffn_input_mode = (
|
|
_FFNInputMode.SCATTERED
|
|
if (global_server_args_dict["enable_deepep_moe"] and is_sparse)
|
|
or (Qwen2MoeDecoderLayer._enable_moe_dense_fully_dp() and not is_sparse)
|
|
else _FFNInputMode.FULL
|
|
)
|
|
return _DecoderLayerInfo(is_sparse=is_sparse, ffn_input_mode=ffn_input_mode)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
residual: Optional[torch.Tensor],
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
if self.info.ffn_input_mode == _FFNInputMode.SCATTERED:
|
|
return self.forward_ffn_with_scattered_input(
|
|
positions, hidden_states, forward_batch, residual
|
|
)
|
|
elif self.info.ffn_input_mode == _FFNInputMode.FULL:
|
|
return self.forward_ffn_with_full_input(
|
|
positions, hidden_states, forward_batch, residual
|
|
)
|
|
else:
|
|
raise NotImplementedError
|
|
|
|
def forward_ffn_with_full_input(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
residual: Optional[torch.Tensor],
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
if hidden_states.shape[0] == 0:
|
|
residual = hidden_states
|
|
else:
|
|
if residual is None:
|
|
residual = hidden_states
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
|
|
|
# Self Attention
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
forward_batch=forward_batch,
|
|
)
|
|
# Gather
|
|
if get_tensor_model_parallel_world_size() > 1:
|
|
# all gather and all reduce
|
|
if self.local_dp_size != 1:
|
|
if self.attn_tp_rank == 0:
|
|
hidden_states += residual
|
|
hidden_states, local_hidden_states = (
|
|
forward_batch.gathered_buffer,
|
|
hidden_states,
|
|
)
|
|
dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
|
|
dp_scatter(residual, hidden_states, forward_batch)
|
|
# TODO extract this bugfix
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
|
else:
|
|
hidden_states = tensor_model_parallel_all_reduce(hidden_states)
|
|
# TODO extract this bugfix
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states, residual = self.post_attention_layernorm(
|
|
hidden_states, residual
|
|
)
|
|
elif hidden_states.shape[0] != 0:
|
|
hidden_states, residual = self.post_attention_layernorm(
|
|
hidden_states, residual
|
|
)
|
|
|
|
# Fully Connected
|
|
hidden_states = self.mlp(hidden_states)
|
|
|
|
# TODO: use reduce-scatter in MLP to avoid this scatter
|
|
# Scatter
|
|
if self.local_dp_size != 1:
|
|
# important: forward batch.gathered_buffer is used both after scatter and after gather.
|
|
# be careful about this!
|
|
hidden_states, global_hidden_states = (
|
|
forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
|
|
hidden_states,
|
|
)
|
|
dp_scatter(hidden_states, global_hidden_states, forward_batch)
|
|
|
|
return hidden_states, residual
|
|
|
|
def forward_ffn_with_scattered_input(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
residual: Optional[torch.Tensor],
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
if hidden_states.shape[0] == 0:
|
|
residual = hidden_states
|
|
else:
|
|
if residual is None:
|
|
residual = hidden_states
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
|
|
|
if self.attn_tp_size != 1 and self.input_is_scattered:
|
|
hidden_states, local_hidden_states = (
|
|
forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
|
|
hidden_states,
|
|
)
|
|
attn_tp_all_gather(
|
|
list(hidden_states.tensor_split(self.attn_tp_size)), local_hidden_states
|
|
)
|
|
|
|
# Self Attention
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
forward_batch=forward_batch,
|
|
)
|
|
|
|
if self.attn_tp_size != 1:
|
|
if self.input_is_scattered:
|
|
tensor_list = list(hidden_states.tensor_split(self.attn_tp_size))
|
|
hidden_states = tensor_list[self.attn_tp_rank]
|
|
attn_tp_reduce_scatter(hidden_states, tensor_list)
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states, residual = self.post_attention_layernorm(
|
|
hidden_states, residual
|
|
)
|
|
else:
|
|
if self.attn_tp_rank == 0:
|
|
hidden_states += residual
|
|
tensor_list = list(hidden_states.tensor_split(self.attn_tp_size))
|
|
hidden_states = tensor_list[self.attn_tp_rank]
|
|
attn_tp_reduce_scatter(hidden_states, tensor_list)
|
|
residual = hidden_states
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
|
else:
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states, residual = self.post_attention_layernorm(
|
|
hidden_states, residual
|
|
)
|
|
|
|
if not (
|
|
self._enable_moe_dense_fully_dp()
|
|
and (not self.info.is_sparse)
|
|
and hidden_states.shape[0] == 0
|
|
):
|
|
hidden_states = self.mlp(hidden_states, forward_batch.forward_mode)
|
|
|
|
if self.is_last_layer and self.attn_tp_size != 1:
|
|
hidden_states += residual
|
|
residual = None
|
|
hidden_states, local_hidden_states = (
|
|
forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
|
|
hidden_states,
|
|
)
|
|
attn_tp_all_gather(
|
|
list(hidden_states.tensor_split(self.attn_tp_size)), local_hidden_states
|
|
)
|
|
|
|
return hidden_states, residual
|
|
|
|
|
|
class Qwen2MoeModel(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
decoder_layer_type: type[nn.Module] = Qwen2MoeDecoderLayer,
|
|
) -> None:
|
|
super().__init__()
|
|
self.padding_idx = config.pad_token_id
|
|
self.vocab_size = config.vocab_size
|
|
self.pp_group = get_pp_group()
|
|
|
|
if self.pp_group.is_first_rank:
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
enable_tp=not global_server_args_dict["enable_dp_attention"],
|
|
prefix=add_prefix("embed_tokens", prefix),
|
|
)
|
|
else:
|
|
self.embed_tokens = PPMissingLayer()
|
|
|
|
# Use the provided decoder layer type or default to Qwen2MoeDecoderLayer
|
|
decoder_layer_type = decoder_layer_type or Qwen2MoeDecoderLayer
|
|
self.layers, self.start_layer, self.end_layer = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda idx, prefix: decoder_layer_type(
|
|
layer_id=idx,
|
|
config=config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
),
|
|
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)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: torch.Tensor = None,
|
|
pp_proxy_tensors: Optional[PPProxyTensors] = None,
|
|
) -> Union[torch.Tensor, PPProxyTensors]:
|
|
if self.pp_group.is_first_rank:
|
|
if input_embeds is None:
|
|
hidden_states = self.embed_tokens(input_ids)
|
|
else:
|
|
hidden_states = input_embeds
|
|
residual = None
|
|
else:
|
|
assert pp_proxy_tensors is not None
|
|
hidden_states = pp_proxy_tensors["hidden_states"]
|
|
residual = pp_proxy_tensors["residual"]
|
|
|
|
for i in range(self.start_layer, self.end_layer):
|
|
with get_global_expert_distribution_recorder().with_current_layer(i):
|
|
layer = self.layers[i]
|
|
hidden_states, residual = layer(
|
|
positions, hidden_states, forward_batch, residual
|
|
)
|
|
if not self.pp_group.is_last_rank:
|
|
return PPProxyTensors(
|
|
{
|
|
"hidden_states": hidden_states,
|
|
"residual": residual,
|
|
}
|
|
)
|
|
else:
|
|
if hidden_states.shape[0] != 0:
|
|
hidden_states, _ = self.norm(hidden_states, residual)
|
|
return hidden_states
|
|
|
|
|
|
class Qwen2MoeForCausalLM(nn.Module):
|
|
fall_back_to_pt_during_load = False
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.pp_group = get_pp_group()
|
|
self.config = config
|
|
self.quant_config = quant_config
|
|
self.model = Qwen2MoeModel(
|
|
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),
|
|
)
|
|
self.logits_processor = LogitsProcessor(config)
|
|
|
|
@torch.no_grad()
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: torch.Tensor = None,
|
|
pp_proxy_tensors: Optional[PPProxyTensors] = None,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(
|
|
input_ids,
|
|
positions,
|
|
forward_batch,
|
|
input_embeds,
|
|
pp_proxy_tensors=pp_proxy_tensors,
|
|
)
|
|
if self.pp_group.is_last_rank:
|
|
return self.logits_processor(
|
|
input_ids, hidden_states, self.lm_head, forward_batch
|
|
)
|
|
else:
|
|
return hidden_states
|
|
|
|
@property
|
|
def start_layer(self):
|
|
return self.model.start_layer
|
|
|
|
@property
|
|
def end_layer(self):
|
|
return self.model.end_layer
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
|
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),
|
|
]
|
|
|
|
MoEImpl = EPMoE if global_server_args_dict["enable_ep_moe"] else FusedMoE
|
|
|
|
expert_params_mapping = MoEImpl.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.num_experts,
|
|
)
|
|
|
|
params_dict = dict(self.named_parameters())
|
|
for name, loaded_weight in weights:
|
|
layer_id = get_layer_id(name)
|
|
if (
|
|
layer_id is not None
|
|
and hasattr(self.model, "start_layer")
|
|
and (
|
|
layer_id < self.model.start_layer
|
|
or layer_id >= self.model.end_layer
|
|
)
|
|
):
|
|
continue
|
|
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:
|
|
for mapping in expert_params_mapping:
|
|
param_name, weight_name, expert_id, shard_id = mapping
|
|
if weight_name not in name:
|
|
continue
|
|
name = name.replace(weight_name, param_name)
|
|
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:
|
|
# 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
|
|
|
|
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")
|
|
|
|
@classmethod
|
|
def get_model_config_for_expert_location(cls, config):
|
|
return ModelConfigForExpertLocation(
|
|
num_layers=config.num_hidden_layers,
|
|
num_logical_experts=config.num_experts,
|
|
num_groups=None,
|
|
)
|
|
|
|
|
|
EntryClass = Qwen2MoeForCausalLM
|