Update modeling_internlm.py
Browse files- modeling_internlm.py +52 -36
modeling_internlm.py
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
# coding=utf-8
|
| 2 |
-
# Copyright
|
| 3 |
#
|
| 4 |
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
| 5 |
# and OPT implementations in this library. It has been modified from its
|
|
@@ -28,7 +28,6 @@ import torch.utils.checkpoint
|
|
| 28 |
from torch import nn
|
| 29 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 30 |
from transformers.activations import ACT2FN
|
| 31 |
-
from transformers.generation.streamers import BaseStreamer
|
| 32 |
from transformers.modeling_outputs import (
|
| 33 |
BaseModelOutputWithPast,
|
| 34 |
CausalLMOutputWithPast,
|
|
@@ -42,6 +41,11 @@ from transformers.utils import (
|
|
| 42 |
replace_return_docstrings,
|
| 43 |
)
|
| 44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
from .configuration_internlm import InternLMConfig
|
| 46 |
|
| 47 |
logger = logging.get_logger(__name__)
|
|
@@ -113,6 +117,7 @@ class InternLMRotaryEmbedding(torch.nn.Module):
|
|
| 113 |
base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
|
| 114 |
device (Any, optional): Running device. Defaults to None.
|
| 115 |
"""
|
|
|
|
| 116 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 117 |
super().__init__()
|
| 118 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
|
@@ -124,8 +129,8 @@ class InternLMRotaryEmbedding(torch.nn.Module):
|
|
| 124 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 125 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 126 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 127 |
-
self.register_buffer("cos_cached", emb.cos()
|
| 128 |
-
self.register_buffer("sin_cached", emb.sin()
|
| 129 |
|
| 130 |
def forward(self, x, seq_len=None):
|
| 131 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
|
@@ -136,11 +141,11 @@ class InternLMRotaryEmbedding(torch.nn.Module):
|
|
| 136 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 137 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 138 |
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
| 139 |
-
self.register_buffer("cos_cached", emb.cos()
|
| 140 |
-
self.register_buffer("sin_cached", emb.sin()
|
| 141 |
return (
|
| 142 |
-
self.cos_cached[
|
| 143 |
-
self.sin_cached[
|
| 144 |
)
|
| 145 |
|
| 146 |
|
|
@@ -158,7 +163,7 @@ class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
|
|
| 158 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
|
| 159 |
super().__init__()
|
| 160 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
| 161 |
-
self.register_buffer("inv_freq", inv_freq)
|
| 162 |
self.dim = dim
|
| 163 |
self.base = base
|
| 164 |
self.scaling_factor = scaling_factor
|
|
@@ -170,8 +175,8 @@ class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
|
|
| 170 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 171 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 172 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 173 |
-
self.register_buffer("cos_cached", emb.cos()
|
| 174 |
-
self.register_buffer("sin_cached", emb.sin()
|
| 175 |
|
| 176 |
def _update_cached(self, x, seq_len=None):
|
| 177 |
self.max_seq_len_cached = max(seq_len, self.max_position_embeddings)
|
|
@@ -185,8 +190,8 @@ class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
|
|
| 185 |
t = torch.arange(self.max_seq_len_cached, device=inv_freq.device, dtype=inv_freq.dtype)
|
| 186 |
freqs = torch.einsum("i,j->ij", t, inv_freq)
|
| 187 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 188 |
-
self.register_buffer("cos_cached", emb.cos()
|
| 189 |
-
self.register_buffer("sin_cached", emb.sin()
|
| 190 |
|
| 191 |
def forward(self, x, seq_len=None):
|
| 192 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
|
@@ -199,8 +204,8 @@ class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
|
|
| 199 |
self._update_cached(x, seq_len)
|
| 200 |
|
| 201 |
return (
|
| 202 |
-
self.cos_cached[
|
| 203 |
-
self.sin_cached[
|
| 204 |
)
|
| 205 |
|
| 206 |
|
|
@@ -210,23 +215,23 @@ def rotate_half(x):
|
|
| 210 |
x2 = x[..., x.shape[-1] // 2 :]
|
| 211 |
return torch.cat((-x2, x1), dim=-1)
|
| 212 |
|
| 213 |
-
|
| 214 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 222 |
else:
|
|
|
|
|
|
|
| 223 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 224 |
-
|
| 225 |
-
if k.size(2) == 1:
|
| 226 |
-
k_embed = (k * cos[:, :, -1, :]) + (rotate_half(k) * sin[:, :, -1, :])
|
| 227 |
-
else:
|
| 228 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 229 |
-
|
| 230 |
return q_embed, k_embed
|
| 231 |
|
| 232 |
|
|
@@ -270,7 +275,7 @@ class InternLMAttention(nn.Module):
|
|
| 270 |
self.rotary_emb = self._init_rope()
|
| 271 |
|
| 272 |
def _init_rope(self):
|
| 273 |
-
if self.config.rotary["type"] == "origin"
|
| 274 |
self.rotary_emb = InternLMRotaryEmbedding(
|
| 275 |
self.head_dim,
|
| 276 |
max_position_embeddings=self.max_position_embeddings,
|
|
@@ -310,7 +315,6 @@ class InternLMAttention(nn.Module):
|
|
| 310 |
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
| 311 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
| 312 |
|
| 313 |
-
# print(use_cache)
|
| 314 |
past_key_value = (key_states, value_states) if use_cache else None
|
| 315 |
|
| 316 |
kv_seq_len = key_states.shape[-2]
|
|
@@ -851,12 +855,16 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
| 851 |
for layer_past in past_key_values:
|
| 852 |
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
| 853 |
return reordered_past
|
| 854 |
-
|
| 855 |
-
def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = []):
|
| 856 |
prompt = ""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 857 |
for record in history:
|
| 858 |
-
prompt += f"""<|User|>:{record[0]}
|
| 859 |
-
prompt += f"""<|User|>:{query}
|
| 860 |
return tokenizer([prompt], return_tensors="pt")
|
| 861 |
|
| 862 |
@torch.no_grad()
|
|
@@ -870,9 +878,12 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
| 870 |
do_sample: bool = True,
|
| 871 |
temperature: float = 0.8,
|
| 872 |
top_p: float = 0.8,
|
|
|
|
|
|
|
|
|
|
| 873 |
**kwargs,
|
| 874 |
):
|
| 875 |
-
inputs = self.build_inputs(tokenizer, query, history)
|
| 876 |
inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
|
| 877 |
outputs = self.generate(
|
| 878 |
**inputs,
|
|
@@ -907,6 +918,11 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
| 907 |
('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
|
| 908 |
('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
|
| 909 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 910 |
|
| 911 |
response_queue = queue.Queue(maxsize=20)
|
| 912 |
|
|
@@ -1083,4 +1099,4 @@ class InternLMForSequenceClassification(InternLMPreTrainedModel):
|
|
| 1083 |
past_key_values=transformer_outputs.past_key_values,
|
| 1084 |
hidden_states=transformer_outputs.hidden_states,
|
| 1085 |
attentions=transformer_outputs.attentions,
|
| 1086 |
-
)
|
|
|
|
| 1 |
# coding=utf-8
|
| 2 |
+
# Copyright (c) InternLM. All rights reserved.
|
| 3 |
#
|
| 4 |
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
| 5 |
# and OPT implementations in this library. It has been modified from its
|
|
|
|
| 28 |
from torch import nn
|
| 29 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 30 |
from transformers.activations import ACT2FN
|
|
|
|
| 31 |
from transformers.modeling_outputs import (
|
| 32 |
BaseModelOutputWithPast,
|
| 33 |
CausalLMOutputWithPast,
|
|
|
|
| 41 |
replace_return_docstrings,
|
| 42 |
)
|
| 43 |
|
| 44 |
+
try:
|
| 45 |
+
from transformers.generation.streamers import BaseStreamer
|
| 46 |
+
except: # noqa # pylint: disable=bare-except
|
| 47 |
+
BaseStreamer = None
|
| 48 |
+
|
| 49 |
from .configuration_internlm import InternLMConfig
|
| 50 |
|
| 51 |
logger = logging.get_logger(__name__)
|
|
|
|
| 117 |
base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
|
| 118 |
device (Any, optional): Running device. Defaults to None.
|
| 119 |
"""
|
| 120 |
+
|
| 121 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 122 |
super().__init__()
|
| 123 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
|
|
|
| 129 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 130 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 131 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 132 |
+
self.register_buffer("cos_cached", emb.cos().to(torch.float32), persistent=False)
|
| 133 |
+
self.register_buffer("sin_cached", emb.sin().to(torch.float32), persistent=False)
|
| 134 |
|
| 135 |
def forward(self, x, seq_len=None):
|
| 136 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
|
|
|
| 141 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 142 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 143 |
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
| 144 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
| 145 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
| 146 |
return (
|
| 147 |
+
self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
|
| 148 |
+
self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
|
| 149 |
)
|
| 150 |
|
| 151 |
|
|
|
|
| 163 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
|
| 164 |
super().__init__()
|
| 165 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
| 166 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 167 |
self.dim = dim
|
| 168 |
self.base = base
|
| 169 |
self.scaling_factor = scaling_factor
|
|
|
|
| 175 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 176 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
| 177 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 178 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
| 179 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
| 180 |
|
| 181 |
def _update_cached(self, x, seq_len=None):
|
| 182 |
self.max_seq_len_cached = max(seq_len, self.max_position_embeddings)
|
|
|
|
| 190 |
t = torch.arange(self.max_seq_len_cached, device=inv_freq.device, dtype=inv_freq.dtype)
|
| 191 |
freqs = torch.einsum("i,j->ij", t, inv_freq)
|
| 192 |
emb = torch.cat((freqs, freqs), dim=-1)
|
| 193 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
| 194 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
| 195 |
|
| 196 |
def forward(self, x, seq_len=None):
|
| 197 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
|
|
|
| 204 |
self._update_cached(x, seq_len)
|
| 205 |
|
| 206 |
return (
|
| 207 |
+
self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
|
| 208 |
+
self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
|
| 209 |
)
|
| 210 |
|
| 211 |
|
|
|
|
| 215 |
x2 = x[..., x.shape[-1] // 2 :]
|
| 216 |
return torch.cat((-x2, x1), dim=-1)
|
| 217 |
|
|
|
|
| 218 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
| 219 |
+
if position_ids.size(1) == 1:
|
| 220 |
+
q_cos = cos[position_ids].unsqueeze(1).expand(q.shape)
|
| 221 |
+
q_sin = sin[position_ids].unsqueeze(1).expand(q.shape)
|
| 222 |
+
q_embed = (q * q_cos) + (rotate_half(q) * q_sin)
|
| 223 |
+
|
| 224 |
+
position_ids = position_ids.flatten() + 1
|
| 225 |
+
max_length = max(position_ids)
|
| 226 |
+
position_ids = torch.stack([torch.cat([torch.ones(max_length - w, dtype=torch.long), torch.arange(w)]) for w in position_ids])
|
| 227 |
+
k_cos = cos[position_ids].unsqueeze(1).expand(k.shape)
|
| 228 |
+
k_sin = sin[position_ids].unsqueeze(1).expand(k.shape)
|
| 229 |
+
k_embed = (k * k_cos) + (rotate_half(k) * k_sin)
|
| 230 |
else:
|
| 231 |
+
cos = cos[position_ids].unsqueeze(1).expand(q.shape)
|
| 232 |
+
sin = sin[position_ids].unsqueeze(1).expand(q.shape)
|
| 233 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
|
|
|
| 235 |
return q_embed, k_embed
|
| 236 |
|
| 237 |
|
|
|
|
| 275 |
self.rotary_emb = self._init_rope()
|
| 276 |
|
| 277 |
def _init_rope(self):
|
| 278 |
+
if self.config.rotary["type"] == "origin"
|
| 279 |
self.rotary_emb = InternLMRotaryEmbedding(
|
| 280 |
self.head_dim,
|
| 281 |
max_position_embeddings=self.max_position_embeddings,
|
|
|
|
| 315 |
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
| 316 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
| 317 |
|
|
|
|
| 318 |
past_key_value = (key_states, value_states) if use_cache else None
|
| 319 |
|
| 320 |
kv_seq_len = key_states.shape[-2]
|
|
|
|
| 855 |
for layer_past in past_key_values:
|
| 856 |
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
| 857 |
return reordered_past
|
| 858 |
+
|
| 859 |
+
def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = [], meta_instruction=""):
|
| 860 |
prompt = ""
|
| 861 |
+
if meta_instruction:
|
| 862 |
+
prompt += f"""<s><|System|>:{meta_instruction}\n"""
|
| 863 |
+
else:
|
| 864 |
+
prompt += "<s>"
|
| 865 |
for record in history:
|
| 866 |
+
prompt += f"""<|User|>:{record[0]}\n<|Bot|>:{record[1]}<eoa>\n"""
|
| 867 |
+
prompt += f"""<|User|>:{query}\n<|Bot|>:"""
|
| 868 |
return tokenizer([prompt], return_tensors="pt")
|
| 869 |
|
| 870 |
@torch.no_grad()
|
|
|
|
| 878 |
do_sample: bool = True,
|
| 879 |
temperature: float = 0.8,
|
| 880 |
top_p: float = 0.8,
|
| 881 |
+
meta_instruction: str = "You are an AI assistant whose name is InternLM (书生·浦语).\n"
|
| 882 |
+
"- InternLM (书生·浦语) is a conversational language model that is developed by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.\n"
|
| 883 |
+
"- InternLM (书生·浦语) can understand and communicate fluently in the language chosen by the user such as English and 中文.",
|
| 884 |
**kwargs,
|
| 885 |
):
|
| 886 |
+
inputs = self.build_inputs(tokenizer, query, history, meta_instruction)
|
| 887 |
inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
|
| 888 |
outputs = self.generate(
|
| 889 |
**inputs,
|
|
|
|
| 918 |
('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
|
| 919 |
('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
|
| 920 |
"""
|
| 921 |
+
if BaseStreamer is None:
|
| 922 |
+
raise ModuleNotFoundError(
|
| 923 |
+
"The version of `transformers` is too low. Please make sure "
|
| 924 |
+
"that you have installed `transformers>=4.28.0`."
|
| 925 |
+
)
|
| 926 |
|
| 927 |
response_queue = queue.Queue(maxsize=20)
|
| 928 |
|
|
|
|
| 1099 |
past_key_values=transformer_outputs.past_key_values,
|
| 1100 |
hidden_states=transformer_outputs.hidden_states,
|
| 1101 |
attentions=transformer_outputs.attentions,
|
| 1102 |
+
)
|