XiaoduoAILab
commited on
Commit
•
e44944c
1
Parent(s):
47da563
Upload 2 files
Browse files- configuration_xmodel.py +88 -0
- modeling_xmodel.py +739 -0
configuration_xmodel.py
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) 2023 XiaoDuo AI. All rights reserved.
|
2 |
+
|
3 |
+
from transformers.configuration_utils import PretrainedConfig
|
4 |
+
from transformers.utils import logging
|
5 |
+
from typing_extensions import Self
|
6 |
+
|
7 |
+
logger = logging.get_logger(__name__)
|
8 |
+
|
9 |
+
|
10 |
+
class XModelConfig(PretrainedConfig):
|
11 |
+
model_type = "xmodel_32000"
|
12 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
13 |
+
|
14 |
+
def __init__(
|
15 |
+
self,
|
16 |
+
vocab_size=32000,
|
17 |
+
hidden_size=4096,
|
18 |
+
intermediate_size=None,
|
19 |
+
num_hidden_layers=32,
|
20 |
+
num_attention_heads=32,
|
21 |
+
num_key_value_heads=32,
|
22 |
+
hidden_act="silu",
|
23 |
+
max_position_embeddings=32768,
|
24 |
+
initializer_range=0.02,
|
25 |
+
rms_norm_eps=1e-5,
|
26 |
+
use_cache=True,
|
27 |
+
pad_token_id=0,
|
28 |
+
bos_token_id=1,
|
29 |
+
eos_token_id=2,
|
30 |
+
pretraining_tp=1,
|
31 |
+
tie_word_embeddings=False,
|
32 |
+
**kwargs,
|
33 |
+
):
|
34 |
+
self.vocab_size = vocab_size
|
35 |
+
self.max_position_embeddings = max_position_embeddings
|
36 |
+
self.hidden_size = hidden_size
|
37 |
+
# self.intermediate_size = intermediate_size
|
38 |
+
if intermediate_size is None:
|
39 |
+
self.intermediate_size = find_multiple(int(8 * hidden_size / 3), 256)
|
40 |
+
else:
|
41 |
+
self.intermediate_size = intermediate_size
|
42 |
+
self.num_hidden_layers = num_hidden_layers
|
43 |
+
self.num_attention_heads = num_attention_heads
|
44 |
+
self.num_key_value_heads = num_key_value_heads
|
45 |
+
self.hidden_act = hidden_act
|
46 |
+
self.initializer_range = initializer_range
|
47 |
+
self.rms_norm_eps = rms_norm_eps
|
48 |
+
self.pretraining_tp = pretraining_tp
|
49 |
+
self.use_cache = use_cache
|
50 |
+
self.auto_map = {
|
51 |
+
"AutoConfig": "configuration_xmodel.XModelConfig",
|
52 |
+
"AutoModelForCausalLM": "modeling_xmodel.XModelForCausalLM"
|
53 |
+
}
|
54 |
+
|
55 |
+
super().__init__(
|
56 |
+
pad_token_id=pad_token_id,
|
57 |
+
bos_token_id=bos_token_id,
|
58 |
+
eos_token_id=eos_token_id,
|
59 |
+
tie_word_embeddings=tie_word_embeddings,
|
60 |
+
**kwargs,
|
61 |
+
)
|
62 |
+
|
63 |
+
@classmethod
|
64 |
+
def from_name(cls, name: str) -> Self:
|
65 |
+
return cls(**xmodel_configs[name])
|
66 |
+
|
67 |
+
|
68 |
+
xmodel_configs = {
|
69 |
+
"nano": dict(num_hidden_layers=6, num_attention_heads=6, num_key_value_heads=1, hidden_size=192),
|
70 |
+
"micro": dict(num_hidden_layers=6, num_attention_heads=6, num_key_value_heads=1, hidden_size=384),
|
71 |
+
"tiny": dict(num_hidden_layers=8, num_attention_heads=8, num_key_value_heads=2, hidden_size=512),
|
72 |
+
"small": dict(num_hidden_layers=12, num_attention_heads=12, num_key_value_heads=3, hidden_size=768),
|
73 |
+
# GPT-1 & Bert-Base
|
74 |
+
"medium": dict(num_hidden_layers=24, num_attention_heads=16, num_key_value_heads=4, hidden_size=1024), # Bert-Large
|
75 |
+
"large": dict(num_hidden_layers=24, num_attention_heads=16, num_key_value_heads=4, hidden_size=1536),
|
76 |
+
"xl": dict(num_hidden_layers=24, num_attention_heads=32, num_key_value_heads=4, hidden_size=2048), # GPT-2
|
77 |
+
"3B": dict(num_hidden_layers=32, num_attention_heads=32, num_key_value_heads=4, hidden_size=2560),
|
78 |
+
"7B": dict(num_hidden_layers=32, num_attention_heads=32, num_key_value_heads=32, hidden_size=4096),
|
79 |
+
"13B": dict(num_hidden_layers=40, num_attention_heads=40, num_key_value_heads=40, hidden_size=5120),
|
80 |
+
"34B": dict(num_hidden_layers=48, num_attention_heads=64, num_key_value_heads=8, hidden_size=8192),
|
81 |
+
"70B": dict(num_hidden_layers=80, num_attention_heads=64, num_key_value_heads=8, hidden_size=8192), # Llama
|
82 |
+
}
|
83 |
+
|
84 |
+
|
85 |
+
def find_multiple(n: int, k: int) -> int:
|
86 |
+
if n % k == 0:
|
87 |
+
return n
|
88 |
+
return n + k - (n % k)
|
modeling_xmodel.py
ADDED
@@ -0,0 +1,739 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) 2023 XiaoDuo AI. All rights reserved.
|
2 |
+
|
3 |
+
import math
|
4 |
+
from typing import List, Optional, Tuple, Union
|
5 |
+
|
6 |
+
import torch
|
7 |
+
import torch.utils.checkpoint
|
8 |
+
import transformers
|
9 |
+
from torch import nn
|
10 |
+
from torch.nn import CrossEntropyLoss
|
11 |
+
from torch.nn import functional as F
|
12 |
+
from transformers.activations import ACT2FN
|
13 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
14 |
+
from transformers.utils import logging
|
15 |
+
|
16 |
+
from .configuration_xmodel import XModelConfig
|
17 |
+
|
18 |
+
logger = logging.get_logger(__name__)
|
19 |
+
torch2 = torch.__version__.split('.')[0] == '2'
|
20 |
+
|
21 |
+
|
22 |
+
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
23 |
+
def _make_causal_mask(
|
24 |
+
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
|
25 |
+
):
|
26 |
+
"""
|
27 |
+
Make causal mask used for bi-directional self-attention.
|
28 |
+
"""
|
29 |
+
bsz, tgt_len = input_ids_shape
|
30 |
+
mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
|
31 |
+
mask_cond = torch.arange(mask.size(-1), device=device)
|
32 |
+
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
|
33 |
+
mask = mask.to(dtype)
|
34 |
+
|
35 |
+
if past_key_values_length > 0:
|
36 |
+
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
|
37 |
+
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
|
38 |
+
|
39 |
+
|
40 |
+
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
41 |
+
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
|
42 |
+
"""
|
43 |
+
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
|
44 |
+
"""
|
45 |
+
bsz, src_len = mask.size()
|
46 |
+
tgt_len = tgt_len if tgt_len is not None else src_len
|
47 |
+
|
48 |
+
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
|
49 |
+
|
50 |
+
inverted_mask = 1.0 - expanded_mask
|
51 |
+
|
52 |
+
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
53 |
+
|
54 |
+
|
55 |
+
class RMSNorm(nn.Module):
|
56 |
+
def __init__(self, hidden_size, eps=1e-6):
|
57 |
+
super().__init__()
|
58 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
59 |
+
self.variance_epsilon = eps
|
60 |
+
|
61 |
+
def forward(self, hidden_states):
|
62 |
+
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
63 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
64 |
+
|
65 |
+
# convert into half-precision if necessary
|
66 |
+
if self.weight.dtype in [torch.float16, torch.bfloat16]:
|
67 |
+
hidden_states = hidden_states.to(self.weight.dtype)
|
68 |
+
|
69 |
+
return self.weight * hidden_states
|
70 |
+
|
71 |
+
|
72 |
+
class RotaryEmbedding(torch.nn.Module):
|
73 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
74 |
+
super().__init__()
|
75 |
+
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
76 |
+
self.register_buffer("inv_freq", inv_freq)
|
77 |
+
|
78 |
+
# Build here to make `torch.jit.trace` work.
|
79 |
+
self.max_seq_len_cached = max_position_embeddings
|
80 |
+
t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
|
81 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
82 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
83 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
84 |
+
self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
|
85 |
+
self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
|
86 |
+
|
87 |
+
def forward(self, x, seq_len=None):
|
88 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
89 |
+
# This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
|
90 |
+
if seq_len > self.max_seq_len_cached:
|
91 |
+
self.max_seq_len_cached = seq_len
|
92 |
+
t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
|
93 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
94 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
95 |
+
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
96 |
+
self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
|
97 |
+
self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
|
98 |
+
return (
|
99 |
+
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
100 |
+
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
101 |
+
)
|
102 |
+
|
103 |
+
|
104 |
+
def rotate_half(x):
|
105 |
+
"""Rotates half the hidden dims of the input."""
|
106 |
+
x1 = x[..., : x.shape[-1] // 2]
|
107 |
+
x2 = x[..., x.shape[-1] // 2:]
|
108 |
+
return torch.cat((-x2, x1), dim=-1)
|
109 |
+
|
110 |
+
|
111 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
112 |
+
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
|
113 |
+
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
|
114 |
+
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
|
115 |
+
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
116 |
+
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
|
117 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
118 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
119 |
+
return q_embed, k_embed
|
120 |
+
|
121 |
+
|
122 |
+
class MLP(nn.Module):
|
123 |
+
def __init__(
|
124 |
+
self,
|
125 |
+
hidden_size: int,
|
126 |
+
intermediate_size: int,
|
127 |
+
hidden_act: str,
|
128 |
+
):
|
129 |
+
super().__init__()
|
130 |
+
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
131 |
+
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
|
132 |
+
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
133 |
+
self.act_fn = ACT2FN[hidden_act]
|
134 |
+
|
135 |
+
def forward(self, x):
|
136 |
+
out = self.gate_proj(x)
|
137 |
+
out = self.act_fn(out)
|
138 |
+
out = out * self.up_proj(x)
|
139 |
+
out = self.down_proj(out)
|
140 |
+
return out
|
141 |
+
|
142 |
+
|
143 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
144 |
+
"""
|
145 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
146 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
147 |
+
"""
|
148 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
149 |
+
if n_rep == 1:
|
150 |
+
return hidden_states
|
151 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
152 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
153 |
+
|
154 |
+
|
155 |
+
class Attention(nn.Module):
|
156 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
157 |
+
|
158 |
+
def __init__(self, config: XModelConfig):
|
159 |
+
super().__init__()
|
160 |
+
self.config = config
|
161 |
+
self.hidden_size = config.hidden_size
|
162 |
+
self.num_heads = config.num_attention_heads
|
163 |
+
self.head_dim = self.hidden_size // self.num_heads
|
164 |
+
self.num_key_value_heads = config.num_key_value_heads
|
165 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
166 |
+
self.max_position_embeddings = config.max_position_embeddings
|
167 |
+
|
168 |
+
if (self.head_dim * self.num_heads) != self.hidden_size:
|
169 |
+
raise ValueError(
|
170 |
+
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
171 |
+
f" and `num_heads`: {self.num_heads})."
|
172 |
+
)
|
173 |
+
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
174 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
175 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
176 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
177 |
+
self.rotary_emb = RotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
|
178 |
+
|
179 |
+
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
180 |
+
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
181 |
+
|
182 |
+
def forward(
|
183 |
+
self,
|
184 |
+
hidden_states: torch.Tensor,
|
185 |
+
attention_mask: Optional[torch.Tensor] = None,
|
186 |
+
position_ids: Optional[torch.LongTensor] = None,
|
187 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
188 |
+
output_attentions: bool = False,
|
189 |
+
use_cache: bool = False,
|
190 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
191 |
+
bsz, q_len, _ = hidden_states.size()
|
192 |
+
|
193 |
+
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
194 |
+
key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1,
|
195 |
+
2)
|
196 |
+
value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1,
|
197 |
+
2)
|
198 |
+
|
199 |
+
kv_seq_len = key_states.shape[-2]
|
200 |
+
if past_key_value is not None:
|
201 |
+
kv_seq_len += past_key_value[0].shape[-2]
|
202 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
203 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
204 |
+
|
205 |
+
if past_key_value is not None:
|
206 |
+
# reuse k, v, self_attention
|
207 |
+
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
208 |
+
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
209 |
+
|
210 |
+
past_key_value = (key_states, value_states) if use_cache else None
|
211 |
+
|
212 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
213 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
214 |
+
|
215 |
+
if torch2:
|
216 |
+
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
|
217 |
+
attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states,
|
218 |
+
attn_mask=attention_mask)
|
219 |
+
else:
|
220 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
221 |
+
|
222 |
+
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
223 |
+
raise ValueError(
|
224 |
+
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
|
225 |
+
f" {attn_weights.size()}"
|
226 |
+
)
|
227 |
+
|
228 |
+
if attention_mask is not None:
|
229 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
230 |
+
raise ValueError(
|
231 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
232 |
+
)
|
233 |
+
attn_weights = attn_weights + attention_mask
|
234 |
+
|
235 |
+
# upcast attention to fp32
|
236 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
237 |
+
# self.attention_dropout
|
238 |
+
attn_weights = nn.functional.dropout(attn_weights, training=self.training)
|
239 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
240 |
+
|
241 |
+
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
242 |
+
raise ValueError(
|
243 |
+
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
244 |
+
f" {attn_output.size()}"
|
245 |
+
)
|
246 |
+
|
247 |
+
attn_output = attn_output.transpose(1, 2)
|
248 |
+
|
249 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
250 |
+
|
251 |
+
attn_output = self.o_proj(attn_output)
|
252 |
+
|
253 |
+
if not output_attentions:
|
254 |
+
attn_weights = None
|
255 |
+
|
256 |
+
return attn_output, attn_weights, past_key_value
|
257 |
+
|
258 |
+
|
259 |
+
class DecoderLayer(nn.Module):
|
260 |
+
def __init__(self, config: XModelConfig):
|
261 |
+
super().__init__()
|
262 |
+
self.hidden_size = config.hidden_size
|
263 |
+
self.self_attn = Attention(config=config)
|
264 |
+
self.mlp = MLP(
|
265 |
+
hidden_size=self.hidden_size,
|
266 |
+
intermediate_size=config.intermediate_size,
|
267 |
+
hidden_act=config.hidden_act,
|
268 |
+
)
|
269 |
+
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
270 |
+
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
271 |
+
|
272 |
+
def forward(
|
273 |
+
self,
|
274 |
+
hidden_states: torch.Tensor,
|
275 |
+
attention_mask: Optional[torch.Tensor] = None,
|
276 |
+
position_ids: Optional[torch.LongTensor] = None,
|
277 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
278 |
+
output_attentions: Optional[bool] = False,
|
279 |
+
use_cache: Optional[bool] = False,
|
280 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
281 |
+
"""
|
282 |
+
Args:
|
283 |
+
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
284 |
+
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
285 |
+
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
286 |
+
output_attentions (`bool`, *optional*):
|
287 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
288 |
+
returned tensors for more detail.
|
289 |
+
use_cache (`bool`, *optional*):
|
290 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
291 |
+
(see `past_key_values`).
|
292 |
+
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
293 |
+
"""
|
294 |
+
|
295 |
+
residual = hidden_states
|
296 |
+
|
297 |
+
hidden_states = self.input_layernorm(hidden_states)
|
298 |
+
|
299 |
+
# Self Attention
|
300 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
301 |
+
hidden_states=hidden_states,
|
302 |
+
attention_mask=attention_mask,
|
303 |
+
position_ids=position_ids,
|
304 |
+
past_key_value=past_key_value,
|
305 |
+
output_attentions=output_attentions,
|
306 |
+
use_cache=use_cache,
|
307 |
+
)
|
308 |
+
hidden_states = residual + hidden_states
|
309 |
+
|
310 |
+
# Fully Connected
|
311 |
+
residual = hidden_states
|
312 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
313 |
+
hidden_states = self.mlp(hidden_states)
|
314 |
+
hidden_states = residual + hidden_states
|
315 |
+
|
316 |
+
outputs = (hidden_states,)
|
317 |
+
|
318 |
+
if output_attentions:
|
319 |
+
outputs += (self_attn_weights,)
|
320 |
+
|
321 |
+
if use_cache:
|
322 |
+
outputs += (present_key_value,)
|
323 |
+
|
324 |
+
return outputs
|
325 |
+
|
326 |
+
|
327 |
+
class PreTrainedModel(transformers.PreTrainedModel):
|
328 |
+
config_class = XModelConfig
|
329 |
+
base_model_prefix = "model"
|
330 |
+
supports_gradient_checkpointing = True
|
331 |
+
_no_split_modules = ["DecoderLayer"]
|
332 |
+
_keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
|
333 |
+
|
334 |
+
def _init_weights(self, module):
|
335 |
+
std = self.config.initializer_range
|
336 |
+
if isinstance(module, nn.Linear):
|
337 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
338 |
+
if module.bias is not None:
|
339 |
+
module.bias.data.zero_()
|
340 |
+
elif isinstance(module, nn.Embedding):
|
341 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
342 |
+
if module.padding_idx is not None:
|
343 |
+
module.weight.data[module.padding_idx].zero_()
|
344 |
+
|
345 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
346 |
+
if isinstance(module, Model):
|
347 |
+
module.gradient_checkpointing = value
|
348 |
+
|
349 |
+
|
350 |
+
class Model(PreTrainedModel):
|
351 |
+
"""
|
352 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DecoderLayer`]
|
353 |
+
|
354 |
+
Args:
|
355 |
+
config: XModelConfig
|
356 |
+
"""
|
357 |
+
|
358 |
+
def __init__(self, config: XModelConfig):
|
359 |
+
super().__init__(config)
|
360 |
+
self.padding_idx = config.pad_token_id
|
361 |
+
self.vocab_size = config.vocab_size
|
362 |
+
|
363 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
364 |
+
self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
|
365 |
+
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
366 |
+
|
367 |
+
self.gradient_checkpointing = False
|
368 |
+
# Initialize weights and apply final processing
|
369 |
+
self.post_init()
|
370 |
+
|
371 |
+
def get_input_embeddings(self):
|
372 |
+
return self.embed_tokens
|
373 |
+
|
374 |
+
def set_input_embeddings(self, value):
|
375 |
+
self.embed_tokens = value
|
376 |
+
|
377 |
+
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
|
378 |
+
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
|
379 |
+
# create causal mask
|
380 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
381 |
+
combined_attention_mask = None
|
382 |
+
if input_shape[-1] > 1:
|
383 |
+
combined_attention_mask = _make_causal_mask(
|
384 |
+
input_shape,
|
385 |
+
inputs_embeds.dtype,
|
386 |
+
device=inputs_embeds.device,
|
387 |
+
past_key_values_length=past_key_values_length,
|
388 |
+
)
|
389 |
+
|
390 |
+
if attention_mask is not None:
|
391 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
392 |
+
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
|
393 |
+
inputs_embeds.device
|
394 |
+
)
|
395 |
+
combined_attention_mask = (
|
396 |
+
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
|
397 |
+
)
|
398 |
+
|
399 |
+
return combined_attention_mask
|
400 |
+
|
401 |
+
def forward(
|
402 |
+
self,
|
403 |
+
input_ids: torch.LongTensor = None,
|
404 |
+
attention_mask: Optional[torch.Tensor] = None,
|
405 |
+
position_ids: Optional[torch.LongTensor] = None,
|
406 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
407 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
408 |
+
use_cache: Optional[bool] = None,
|
409 |
+
output_attentions: Optional[bool] = None,
|
410 |
+
output_hidden_states: Optional[bool] = None,
|
411 |
+
return_dict: Optional[bool] = None,
|
412 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
413 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
414 |
+
output_hidden_states = (
|
415 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
416 |
+
)
|
417 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
418 |
+
|
419 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
420 |
+
|
421 |
+
# retrieve input_ids and inputs_embeds
|
422 |
+
if input_ids is not None and inputs_embeds is not None:
|
423 |
+
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
|
424 |
+
elif input_ids is not None:
|
425 |
+
batch_size, seq_length = input_ids.shape
|
426 |
+
elif inputs_embeds is not None:
|
427 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
428 |
+
else:
|
429 |
+
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
|
430 |
+
|
431 |
+
seq_length_with_past = seq_length
|
432 |
+
past_key_values_length = 0
|
433 |
+
|
434 |
+
if past_key_values is not None:
|
435 |
+
past_key_values_length = past_key_values[0][0].shape[2]
|
436 |
+
seq_length_with_past = seq_length_with_past + past_key_values_length
|
437 |
+
|
438 |
+
if position_ids is None:
|
439 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
440 |
+
position_ids = torch.arange(
|
441 |
+
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
442 |
+
)
|
443 |
+
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
444 |
+
else:
|
445 |
+
position_ids = position_ids.view(-1, seq_length).long()
|
446 |
+
|
447 |
+
if inputs_embeds is None:
|
448 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
449 |
+
# embed positions
|
450 |
+
if attention_mask is None:
|
451 |
+
attention_mask = torch.ones(
|
452 |
+
(batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
|
453 |
+
)
|
454 |
+
attention_mask = self._prepare_decoder_attention_mask(
|
455 |
+
attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
|
456 |
+
)
|
457 |
+
|
458 |
+
hidden_states = inputs_embeds
|
459 |
+
|
460 |
+
if self.gradient_checkpointing and self.training:
|
461 |
+
if use_cache:
|
462 |
+
logger.warning_once(
|
463 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
464 |
+
)
|
465 |
+
use_cache = False
|
466 |
+
|
467 |
+
# decoder layers
|
468 |
+
all_hidden_states = () if output_hidden_states else None
|
469 |
+
all_self_attns = () if output_attentions else None
|
470 |
+
next_decoder_cache = () if use_cache else None
|
471 |
+
|
472 |
+
for idx, decoder_layer in enumerate(self.layers):
|
473 |
+
if output_hidden_states:
|
474 |
+
all_hidden_states += (hidden_states,)
|
475 |
+
|
476 |
+
past_key_value = past_key_values[idx] if past_key_values is not None else None
|
477 |
+
|
478 |
+
if self.gradient_checkpointing and self.training:
|
479 |
+
|
480 |
+
def create_custom_forward(module):
|
481 |
+
def custom_forward(*inputs):
|
482 |
+
# None for past_key_value
|
483 |
+
return module(*inputs, output_attentions, None)
|
484 |
+
|
485 |
+
return custom_forward
|
486 |
+
|
487 |
+
layer_outputs = torch.utils.checkpoint.checkpoint(
|
488 |
+
create_custom_forward(decoder_layer),
|
489 |
+
hidden_states,
|
490 |
+
attention_mask,
|
491 |
+
position_ids,
|
492 |
+
None,
|
493 |
+
)
|
494 |
+
else:
|
495 |
+
layer_outputs = decoder_layer(
|
496 |
+
hidden_states,
|
497 |
+
attention_mask=attention_mask,
|
498 |
+
position_ids=position_ids,
|
499 |
+
past_key_value=past_key_value,
|
500 |
+
output_attentions=output_attentions,
|
501 |
+
use_cache=use_cache,
|
502 |
+
)
|
503 |
+
# print('debug_attention_mask', type(attention_mask),attention_mask.dtype)
|
504 |
+
# print('debug_position_ids', type(position_ids),position_ids.dtype)
|
505 |
+
hidden_states = layer_outputs[0]
|
506 |
+
|
507 |
+
if use_cache:
|
508 |
+
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
|
509 |
+
|
510 |
+
if output_attentions:
|
511 |
+
all_self_attns += (layer_outputs[1],)
|
512 |
+
|
513 |
+
hidden_states = self.norm(hidden_states)
|
514 |
+
|
515 |
+
# add hidden states from the last decoder layer
|
516 |
+
if output_hidden_states:
|
517 |
+
all_hidden_states += (hidden_states,)
|
518 |
+
|
519 |
+
next_cache = next_decoder_cache if use_cache else None
|
520 |
+
if not return_dict:
|
521 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
522 |
+
return BaseModelOutputWithPast(
|
523 |
+
last_hidden_state=hidden_states,
|
524 |
+
past_key_values=next_cache,
|
525 |
+
hidden_states=all_hidden_states,
|
526 |
+
attentions=all_self_attns,
|
527 |
+
)
|
528 |
+
|
529 |
+
|
530 |
+
class XModelForCausalLM(PreTrainedModel):
|
531 |
+
def __init__(self, config):
|
532 |
+
super().__init__(config)
|
533 |
+
self.model = Model(config)
|
534 |
+
|
535 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
536 |
+
|
537 |
+
# Initialize weights and apply final processing
|
538 |
+
self.post_init()
|
539 |
+
|
540 |
+
def get_input_embeddings(self):
|
541 |
+
return self.model.embed_tokens
|
542 |
+
|
543 |
+
def set_input_embeddings(self, value):
|
544 |
+
self.model.embed_tokens = value
|
545 |
+
|
546 |
+
def get_output_embeddings(self):
|
547 |
+
return self.lm_head
|
548 |
+
|
549 |
+
def set_output_embeddings(self, new_embeddings):
|
550 |
+
self.lm_head = new_embeddings
|
551 |
+
|
552 |
+
def set_decoder(self, decoder):
|
553 |
+
self.model = decoder
|
554 |
+
|
555 |
+
def get_decoder(self):
|
556 |
+
return self.model
|
557 |
+
|
558 |
+
def forward(
|
559 |
+
self,
|
560 |
+
input_ids: torch.LongTensor = None,
|
561 |
+
attention_mask: Optional[torch.Tensor] = None,
|
562 |
+
position_ids: Optional[torch.LongTensor] = None,
|
563 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
564 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
565 |
+
labels: Optional[torch.LongTensor] = None,
|
566 |
+
use_cache: Optional[bool] = None,
|
567 |
+
output_attentions: Optional[bool] = None,
|
568 |
+
output_hidden_states: Optional[bool] = None,
|
569 |
+
return_dict: Optional[bool] = None,
|
570 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
571 |
+
r"""
|
572 |
+
Args:
|
573 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
574 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
575 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
576 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
577 |
+
|
578 |
+
Returns:
|
579 |
+
|
580 |
+
Example:
|
581 |
+
|
582 |
+
```python
|
583 |
+
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
|
584 |
+
|
585 |
+
>>> model = AutoModelForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
586 |
+
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
587 |
+
|
588 |
+
>>> prompt = "Hey, are you consciours? Can you talk to me?"
|
589 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
590 |
+
|
591 |
+
>>> # Generate
|
592 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
593 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
594 |
+
"Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
|
595 |
+
```"""
|
596 |
+
|
597 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
598 |
+
output_hidden_states = (
|
599 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
600 |
+
)
|
601 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
602 |
+
|
603 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
604 |
+
outputs = self.model(
|
605 |
+
input_ids=input_ids,
|
606 |
+
attention_mask=attention_mask,
|
607 |
+
position_ids=position_ids,
|
608 |
+
past_key_values=past_key_values,
|
609 |
+
inputs_embeds=inputs_embeds,
|
610 |
+
use_cache=use_cache,
|
611 |
+
output_attentions=output_attentions,
|
612 |
+
output_hidden_states=output_hidden_states,
|
613 |
+
return_dict=return_dict,
|
614 |
+
)
|
615 |
+
|
616 |
+
hidden_states = outputs[0]
|
617 |
+
logits = self.lm_head(hidden_states)
|
618 |
+
|
619 |
+
loss = None
|
620 |
+
if labels is not None:
|
621 |
+
# Shift so that tokens < n predict n
|
622 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
623 |
+
shift_labels = labels[..., 1:].contiguous()
|
624 |
+
# Flatten the tokens
|
625 |
+
loss_fct = CrossEntropyLoss()
|
626 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
627 |
+
shift_labels = shift_labels.view(-1)
|
628 |
+
# Enable model parallelism
|
629 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
630 |
+
loss = loss_fct(shift_logits, shift_labels)
|
631 |
+
|
632 |
+
if not return_dict:
|
633 |
+
output = (logits,) + outputs[1:]
|
634 |
+
return (loss,) + output if loss is not None else output
|
635 |
+
|
636 |
+
return CausalLMOutputWithPast(
|
637 |
+
loss=loss,
|
638 |
+
logits=logits,
|
639 |
+
past_key_values=outputs.past_key_values,
|
640 |
+
hidden_states=outputs.hidden_states,
|
641 |
+
attentions=outputs.attentions,
|
642 |
+
)
|
643 |
+
|
644 |
+
def prepare_inputs_for_generation(
|
645 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
646 |
+
):
|
647 |
+
if past_key_values:
|
648 |
+
input_ids = input_ids[:, -1:]
|
649 |
+
|
650 |
+
position_ids = kwargs.get("position_ids", None)
|
651 |
+
if attention_mask is not None and position_ids is None:
|
652 |
+
# create position_ids on the fly for batch generation
|
653 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
654 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
655 |
+
if past_key_values:
|
656 |
+
position_ids = position_ids[:, -1].unsqueeze(-1)
|
657 |
+
|
658 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
659 |
+
if inputs_embeds is not None and past_key_values is None:
|
660 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
661 |
+
else:
|
662 |
+
model_inputs = {"input_ids": input_ids}
|
663 |
+
|
664 |
+
model_inputs.update(
|
665 |
+
{
|
666 |
+
"position_ids": position_ids,
|
667 |
+
"past_key_values": past_key_values,
|
668 |
+
"use_cache": kwargs.get("use_cache"),
|
669 |
+
"attention_mask": attention_mask,
|
670 |
+
}
|
671 |
+
)
|
672 |
+
return model_inputs
|
673 |
+
|
674 |
+
@staticmethod
|
675 |
+
def _reorder_cache(past_key_values, beam_idx):
|
676 |
+
reordered_past = ()
|
677 |
+
for layer_past in past_key_values:
|
678 |
+
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
679 |
+
return reordered_past
|
680 |
+
|
681 |
+
def get_num_params(self, non_embedding=True):
|
682 |
+
"""
|
683 |
+
Return the number of parameters in the model.
|
684 |
+
For non-embedding count (default), the position embeddings get subtracted.
|
685 |
+
The token embeddings would too, except due to the parameter sharing these
|
686 |
+
params are actually used as weights in the final layer, so we include them.
|
687 |
+
"""
|
688 |
+
n_params = sum(p.numel() for p in self.parameters())
|
689 |
+
# if non_embedding:
|
690 |
+
# n_params -= self.transformer.wte.weight.numel()
|
691 |
+
return n_params
|
692 |
+
|
693 |
+
def estimate_mfu(self, fwdbwd_per_iter, dt, max_length=None, device_model='A100', dtype='float32'):
|
694 |
+
""" estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """
|
695 |
+
# first estimate the number of flops we do per iteration.
|
696 |
+
# see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311
|
697 |
+
N = self.get_num_params()
|
698 |
+
n_layer = self.config.num_hidden_layers
|
699 |
+
n_head = self.config.num_attention_heads
|
700 |
+
n_embd = self.config.hidden_size
|
701 |
+
|
702 |
+
if max_length is None:
|
703 |
+
max_length = self.config.max_position_embeddings
|
704 |
+
|
705 |
+
L, H, Q, T = n_layer, n_head, n_embd // n_head, max_length
|
706 |
+
flops_per_token = 6 * N + 12 * L * H * Q * T
|
707 |
+
flops_per_fwdbwd = flops_per_token * T
|
708 |
+
flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
|
709 |
+
# express our flops throughput as ratio of A100 bfloat16 peak flops
|
710 |
+
flops_achieved = flops_per_iter * (1.0 / dt) # per second
|
711 |
+
|
712 |
+
if device_model is None:
|
713 |
+
device_model = torch.cuda.get_device_name(0)
|
714 |
+
|
715 |
+
flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS
|
716 |
+
if device_model == 'DCU' and dtype == 'float16':
|
717 |
+
flops_promised = 23.6e12
|
718 |
+
elif device_model == 'DCU' and dtype == 'float32':
|
719 |
+
flops_promised = 11.8e12
|
720 |
+
elif device_model == 'NVIDIA V100' or 'V100' in device_model:
|
721 |
+
flops_promised = 28e12
|
722 |
+
elif device_model == 'NVIDIA H100' or 'H100' in device_model or device_model == 'NVIDIA H800' or 'H800' in device_model:
|
723 |
+
flops_promised = 1513e12
|
724 |
+
|
725 |
+
mfu = flops_achieved / flops_promised
|
726 |
+
return flops_achieved, mfu
|
727 |
+
|
728 |
+
def flops_per_token(self, max_length=None, non_embedding=False):
|
729 |
+
N = self.get_num_params()
|
730 |
+
if non_embedding:
|
731 |
+
N -= self.config.vocab_size * self.config.hidden_size * 2
|
732 |
+
n_layer = self.config.num_hidden_layers
|
733 |
+
n_head = self.config.num_attention_heads
|
734 |
+
n_embd = self.config.hidden_size
|
735 |
+
if max_length is None:
|
736 |
+
max_length = self.config.max_position_embeddings
|
737 |
+
L, H, Q, T = n_layer, n_head, n_embd // n_head, max_length
|
738 |
+
flops_per_token = 6 * N + 12 * L * H * Q * T
|
739 |
+
return flops_per_token
|