196 lines
7.3 KiB
Python
196 lines
7.3 KiB
Python
################################################################################
|
|
# Copyright(c)2020-2025 Shanghai Biren Technology Co., Ltd. All rights reserved.
|
|
# 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.
|
|
#
|
|
################################################################################
|
|
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
# Adapted from
|
|
# https://github.com/THUDM/ChatGLM2-6B
|
|
"""Inference-only ChatGLM model compatible with THUDM weights."""
|
|
from typing import Optional, Union
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
import vllm
|
|
from vllm.attention import Attention
|
|
from vllm.config import CacheConfig
|
|
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
|
|
from vllm.model_executor.layers.activation import SiluAndMul
|
|
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
|
QKVParallelLinear,
|
|
RowParallelLinear)
|
|
from vllm.model_executor.layers.quantization import QuantizationConfig
|
|
from vllm.model_executor.layers.rotary_embedding import get_rope
|
|
from vllm.model_executor.models.chatglm import GLMMLP
|
|
from vllm.sequence import IntermediateTensors
|
|
from vllm.transformers_utils.configs import ChatGLMConfig
|
|
|
|
|
|
def model_forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
|
**kwargs: object,
|
|
) -> Union[torch.Tensor, IntermediateTensors]:
|
|
if get_pp_group().is_first_rank:
|
|
if inputs_embeds is not None:
|
|
hidden_states = inputs_embeds
|
|
else:
|
|
hidden_states = self.get_input_embeddings(input_ids)
|
|
else:
|
|
assert intermediate_tensors is not None
|
|
hidden_states = intermediate_tensors["hidden_states"]
|
|
# unsqueeze for RMSNorm op
|
|
hidden_states = hidden_states.unsqueeze(0)
|
|
# Run encoder.
|
|
hidden_states = self.encoder(
|
|
hidden_states=hidden_states,
|
|
position_ids=positions,
|
|
)
|
|
# suqeeze to 2-d shape
|
|
return hidden_states.squeeze(0)
|
|
|
|
|
|
class GLMAttention_fit(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: ChatGLMConfig,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
self.total_num_heads = config.num_attention_heads
|
|
assert self.total_num_heads % tp_size == 0
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.multi_query_attention = config.multi_query_attention
|
|
self.total_num_kv_heads = (config.multi_query_group_num
|
|
if config.multi_query_attention else
|
|
config.num_attention_heads)
|
|
if self.total_num_kv_heads >= 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 % 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 tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
|
self.head_dim = config.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.query_key_value = QKVParallelLinear(
|
|
self.hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=config.add_bias_linear or config.add_qkv_bias,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.query_key_value",
|
|
)
|
|
self.dense = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
config.hidden_size,
|
|
bias=config.add_bias_linear,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.dense",
|
|
)
|
|
|
|
# https://huggingface.co/THUDM/chatglm3-6b-32k/blob/e210410255278dd9d74463cf396ba559c0ef801c/modeling_chatglm.py#L141
|
|
rope_ratio = getattr(config, "rope_ratio", 1.0)
|
|
max_positions = getattr(config, "seq_length", 8192)
|
|
# NOTE: THUDM/cogagent-9b-20241220 uses original_rope=False,
|
|
# which is equivalent to is_neox_style=True
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
rotary_dim=self.head_dim // 2,
|
|
max_position=max_positions,
|
|
base=10000 * rope_ratio,
|
|
is_neox_style=False,
|
|
op_type="Chatglm2",
|
|
)
|
|
self.attn = Attention(self.num_heads,
|
|
self.head_dim,
|
|
self.scaling,
|
|
num_kv_heads=self.num_kv_heads,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.attn")
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
position_ids: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.query_key_value(hidden_states)
|
|
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
|
q, k = self.rotary_emb(position_ids, q, k)
|
|
context_layer = self.attn(q, k, v)
|
|
attn_output, _ = self.dense(context_layer)
|
|
return attn_output
|
|
|
|
|
|
def GLMMLP__init__(
|
|
self,
|
|
config: ChatGLMConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
):
|
|
super(GLMMLP, self).__init__()
|
|
|
|
self.add_bias = config.add_bias_linear
|
|
|
|
# Project to 4h.
|
|
self.dense_h_to_4h = MergedColumnParallelLinear(
|
|
config.hidden_size,
|
|
[config.ffn_hidden_size] * 2,
|
|
bias=config.add_bias_linear,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.dense_h_to_4h",
|
|
)
|
|
self.dense_h_to_4h.no_fuse_act = True
|
|
self.activation_func = SiluAndMul()
|
|
|
|
# Project back to h.
|
|
self.dense_4h_to_h = RowParallelLinear(
|
|
config.ffn_hidden_size,
|
|
config.hidden_size,
|
|
bias=config.add_bias_linear,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.dense_4h_to_h",
|
|
)
|
|
|
|
|
|
def GLMMLP__forward(self, hidden_states):
|
|
# [s, b, 4hp]
|
|
intermediate_parallel, _ = self.dense_h_to_4h(hidden_states)
|
|
# [s, b, h]
|
|
output, _ = self.dense_4h_to_h(intermediate_parallel)
|
|
return output
|
|
|
|
|
|
vllm.model_executor.models.chatglm.GLMMLP.forward = GLMMLP__forward
|
|
vllm.model_executor.models.chatglm.GLMMLP.__init__ = GLMMLP__init__
|
|
vllm.model_executor.models.chatglm.ChatGLMModel.forward = model_forward
|
|
vllm.model_executor.models.chatglm.GLMAttention = GLMAttention_fit
|