341 lines
11 KiB
Python
341 lines
11 KiB
Python
import json
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
from torch import Tensor
|
|
from torch.nn import init
|
|
from dataclasses import asdict, dataclass
|
|
from typing import List, Optional, Self, Tuple
|
|
|
|
|
|
def repeat_kv(x: Tensor, n_rep: int) -> Tensor:
|
|
"""
|
|
Repeat k times along the dimension for attention heads.
|
|
Args:
|
|
x (Tensor): The input tensor.
|
|
n_rep (int): The number of repetitions.
|
|
Returns:
|
|
Tensor: The repeated tensor.
|
|
"""
|
|
|
|
bs, slen, n_heads, head_dim = x.shape
|
|
if n_rep == 1:
|
|
return x
|
|
return (
|
|
x[:, :, :, None, :]
|
|
.expand(bs, slen, n_heads, n_rep, head_dim)
|
|
.reshape(bs, slen, n_heads * n_rep, head_dim)
|
|
)
|
|
|
|
def get_rotary_emb(
|
|
dim: int,
|
|
max_len: int,
|
|
base: float = 10000,
|
|
device: torch.device = "cuda",
|
|
) -> torch.Tensor:
|
|
"""
|
|
Get the rotary embedding for the given dimension and maximum length.
|
|
Args:
|
|
dim (int): The dimension of the input.
|
|
max_len (int): The maximum length of the input.
|
|
base (float, optional): The base for the frequency. Defaults to 10000.
|
|
device (torch.device, optional): The device to use. Defaults to "cuda".
|
|
Returns:
|
|
Tensor: The rotary embedding tensor.
|
|
"""
|
|
|
|
theta = base ** (-torch.arange(0, dim, 2, device=device).float() / dim)
|
|
t = torch.arange(0, max_len, device=device).float()
|
|
freqs = torch.outer(t, theta)
|
|
freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
|
|
|
|
return freqs_cis
|
|
|
|
def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
|
|
"""
|
|
Apply rotary embedding to the input tensor.
|
|
Args:
|
|
x (Tensor): The input tensor.
|
|
freqs_cis (Tensor): The rotary embedding tensor.
|
|
Returns:
|
|
Tensor: The output tensor.
|
|
"""
|
|
|
|
dtype = x.dtype
|
|
seq_len = x.size(1)
|
|
|
|
x_complex = torch.view_as_complex(x.view(*x.shape[:-1], -1, 2).float())
|
|
freqs_cis = freqs_cis.reshape(1, seq_len, 1, -1)
|
|
x_out = torch.view_as_real(x_complex * freqs_cis).flatten(3)
|
|
|
|
return x_out.to(dtype)
|
|
|
|
def create_attention_mask(
|
|
seq_mask: Tensor,
|
|
start_pos: int = 0,
|
|
seq_len: int = 0,
|
|
is_causal: bool = False,
|
|
device: torch.device = "cuda",
|
|
dtype: torch.dtype = torch.float32
|
|
) -> Tensor:
|
|
"""
|
|
Create attention mask for GQA
|
|
Args:
|
|
seq_mask (Tensor): A tensor indicating whether each position is valid or not.
|
|
start_pos (int): The starting position of the sequence.
|
|
seq_len (int): The length of the sequence.
|
|
is_causal (bool): Whether the attention is causal or not.
|
|
device (torch.device): The device to use.
|
|
Returns:
|
|
Tensor: The attention mask tensor.
|
|
"""
|
|
|
|
if start_pos != 0 and seq_mask is None:
|
|
# for single prompt chat
|
|
seq_mask = torch.ones((1, seq_len), dtype=torch.bool, device=device)
|
|
|
|
if seq_mask is None:
|
|
return None
|
|
|
|
batch_size = seq_mask.size(0)
|
|
seq_mask = seq_mask[:, :start_pos + seq_len].to(device=device, dtype=torch.bool)
|
|
# (bsz, start_pos + seq_len)
|
|
expanded_mask = seq_mask.unsqueeze(1).expand(batch_size, seq_len, start_pos + seq_len)
|
|
# (bsz, seq_len, start_pos + seq_len)
|
|
|
|
if is_causal:
|
|
causal_mask = torch.tril(
|
|
torch.ones((seq_len, start_pos + seq_len), dtype=torch.bool, device=device),
|
|
diagonal=start_pos
|
|
)
|
|
causal_mask = causal_mask.unsqueeze(0).expand(batch_size, seq_len, start_pos + seq_len)
|
|
expanded_mask = expanded_mask & causal_mask
|
|
|
|
attention_mask = torch.zeros_like(expanded_mask, dtype=dtype, device=device)
|
|
attention_mask = attention_mask.masked_fill_(~expanded_mask, -torch.finfo(dtype).max / 2).unsqueeze(1)
|
|
# (bsz, 1, seq_len, seq_len + start_pos)
|
|
|
|
return attention_mask
|
|
|
|
|
|
@dataclass
|
|
class TransformerConfig:
|
|
# basic config
|
|
vocab_size: Optional[int] = None
|
|
n_dim: Optional[int] = None
|
|
n_head: Optional[int] = None
|
|
n_layer: Optional[int] = None
|
|
m_len: Optional[int] = None
|
|
norm_eps: Optional[float] = None
|
|
d_ffn: Optional[int] = None
|
|
|
|
# GQA
|
|
n_kvhead: Optional[int] = None
|
|
|
|
|
|
def load(self, config_path: str) -> Self:
|
|
with open(config_path, 'r') as f:
|
|
config: dict = json.load(f)
|
|
for key, value in config.items():
|
|
if hasattr(self, key):
|
|
setattr(self, key, value)
|
|
|
|
return self
|
|
|
|
def save(self, config_path: str) -> None:
|
|
config_dict = asdict(self)
|
|
config_dict = {k: v for k, v in config_dict.items() if v is not None}
|
|
with open(config_path, 'w') as f:
|
|
json.dump(config_dict, f, indent=4)
|
|
|
|
|
|
class Linear(nn.Module):
|
|
def __init__(self, in_dim: int, out_dim: int, bias: bool=False):
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.empty((out_dim, in_dim)))
|
|
self.bias = nn.Parameter(torch.zeros(out_dim)) if bias else None
|
|
init.normal_(self.weight, mean=0, std=0.006)
|
|
|
|
def forward(self, x: Tensor) -> Tensor:
|
|
return F.linear(x, self.weight, self.bias)
|
|
|
|
|
|
class RMSNorm(nn.Module):
|
|
def __init__(self, n_dim, norm_eps):
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(n_dim))
|
|
self.norm_eps = norm_eps
|
|
|
|
def forward(self, x: Tensor) -> Tensor:
|
|
dtype = x.dtype
|
|
x = x.float()
|
|
mean_square = torch.mean(torch.pow(x, 2), dim=-1, keepdim=True)
|
|
norm = x * torch.rsqrt(mean_square + self.norm_eps)
|
|
norm = norm.to(dtype)
|
|
out = norm * self.weight
|
|
return out
|
|
|
|
|
|
class MLP(nn.Module):
|
|
def __init__(self, n_dim: int, d_ffn: int):
|
|
super().__init__()
|
|
self.up = Linear(n_dim, d_ffn)
|
|
self.gate = Linear(n_dim, d_ffn)
|
|
self.down = Linear(d_ffn, n_dim)
|
|
|
|
def forward(self, x: Tensor) -> Tensor:
|
|
gated = self.up(x) * F.silu(self.gate(x))
|
|
out = self.down(gated)
|
|
return out
|
|
|
|
|
|
class GQA(nn.Module):
|
|
def __init__(
|
|
self,
|
|
n_dim: int,
|
|
n_head: int,
|
|
n_kvhead: int,
|
|
):
|
|
super().__init__()
|
|
assert n_dim % n_head == 0
|
|
assert n_head % n_kvhead == 0
|
|
|
|
self.head_dim = n_dim // n_head
|
|
self.n_dim = n_dim
|
|
self.n_heads = n_head
|
|
self.n_kvheads = n_kvhead
|
|
self.n_rep = n_head // n_kvhead
|
|
|
|
self.q_proj = Linear(n_dim, n_head * self.head_dim)
|
|
self.k_proj = Linear(n_dim, n_kvhead * self.head_dim)
|
|
self.v_proj = Linear(n_dim, n_kvhead * self.head_dim)
|
|
self.o_proj = Linear(n_dim, n_dim)
|
|
|
|
def forward(
|
|
self,
|
|
x: Tensor,
|
|
freqs_cis: Tensor,
|
|
mask: Tensor = None,
|
|
kv_cache: Optional[Tuple[Tensor, Tensor]] = None,
|
|
start_pos: int = 0
|
|
) -> Tensor:
|
|
bsz, seq_len, _ = x.size()
|
|
# x(bsz, seq_len, n_heads * head_dim) -> (bsz, seq_len, n_heads, head_dim)
|
|
q = self._split_heads(self.q_proj(x), self.n_heads)
|
|
k = self._split_heads(self.k_proj(x), self.n_kvheads)
|
|
v = self._split_heads(self.v_proj(x), self.n_kvheads)
|
|
q, k = apply_rotary_emb(q, freqs_cis), apply_rotary_emb(k, freqs_cis)
|
|
|
|
if kv_cache is not None:
|
|
k_cache, v_cache = kv_cache
|
|
|
|
# copy to cache
|
|
k_cache[:bsz, start_pos:start_pos + seq_len] = k
|
|
v_cache[:bsz, start_pos:start_pos + seq_len] = v
|
|
|
|
# get cache
|
|
k = k_cache[:bsz, :start_pos + seq_len]
|
|
v = v_cache[:bsz, :start_pos + seq_len]
|
|
|
|
k, v = repeat_kv(k, self.n_rep), repeat_kv(v, self.n_rep)
|
|
|
|
# (bsz, seq_len, n_heads, head_dim) -> (bsz, n_heads, seq_len, head_dim)
|
|
q, k, v = q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3)
|
|
sdqa_out = F.scaled_dot_product_attention(q, k, v, mask, is_causal=(mask == None)).permute(0, 2, 1, 3)
|
|
out = self.o_proj(sdqa_out.contiguous().view(bsz, seq_len, -1))
|
|
|
|
return out
|
|
|
|
def _split_heads(self, x: Tensor, n_heads) -> Tensor:
|
|
batch_size, seq_len, _ = x.shape
|
|
x = x.reshape(batch_size, seq_len, n_heads, self.head_dim)
|
|
return x
|
|
|
|
|
|
class DecoderBlock(nn.Module):
|
|
def __init__(self, n_dim, n_head, d_ffn, n_kvhead, norm_eps):
|
|
super().__init__()
|
|
self.attention = GQA(n_dim, n_head, n_kvhead)
|
|
self.norm_attn = RMSNorm(n_dim, norm_eps)
|
|
self.ffn = MLP(n_dim, d_ffn)
|
|
self.norm_ffn = RMSNorm(n_dim, norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
x: Tensor,
|
|
freqs_cis: Tensor,
|
|
attention_mask: Optional[Tensor] = None,
|
|
kv_cache: Optional[Tuple[Tensor, Tensor]] = None,
|
|
start_pos: int = 0
|
|
) -> Tensor:
|
|
# attention
|
|
attn_output = self.attention(
|
|
self.norm_attn(x),
|
|
freqs_cis,
|
|
attention_mask,
|
|
kv_cache,
|
|
start_pos
|
|
)
|
|
x = attn_output + x
|
|
|
|
# feed forward
|
|
x = self.ffn(self.norm_ffn(x)) + x
|
|
|
|
return x
|
|
|
|
|
|
class Transformer(nn.Module):
|
|
def __init__(self, config: TransformerConfig):
|
|
super().__init__()
|
|
self.embedding = nn.Parameter(torch.empty(config.vocab_size, config.n_dim))
|
|
self.layers = nn.ModuleList([
|
|
DecoderBlock(
|
|
config.n_dim,
|
|
config.n_head,
|
|
config.d_ffn,
|
|
config.n_kvhead,
|
|
config.norm_eps
|
|
)
|
|
for _ in range(config.n_layer)
|
|
])
|
|
self.norm = RMSNorm(config.n_dim, config.norm_eps)
|
|
self.freq_cis = get_rotary_emb(config.n_dim // config.n_head, config.m_len)
|
|
init.normal_(self.embedding, mean=0, std=0.02)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: Tensor,
|
|
seq_mask: Optional[Tensor]=None,
|
|
persistent_key_values: Optional[List[Tuple[Tensor, Tensor]]]=None,
|
|
start_pos: int = 0
|
|
) -> Tensor:
|
|
assert input_ids.ndim == 2
|
|
seq_len = input_ids.size(-1)
|
|
x = F.embedding(input_ids, self.embedding)
|
|
|
|
self.freq_cis = self.freq_cis.to(x.device)
|
|
freqs_cis = self.freq_cis[start_pos:start_pos+seq_len]
|
|
has_kvcache = persistent_key_values is not None
|
|
|
|
attn_mask = create_attention_mask(
|
|
seq_mask,
|
|
start_pos=start_pos,
|
|
seq_len=seq_len,
|
|
is_causal=has_kvcache,
|
|
device=x.device,
|
|
dtype=x.dtype
|
|
)
|
|
|
|
for i, layer in enumerate(self.layers):
|
|
kv_cache = persistent_key_values[i] if persistent_key_values else None
|
|
x = layer(x, freqs_cis, attn_mask, kv_cache, start_pos)
|
|
|
|
hidden_states = self.norm(x)
|
|
logits = F.linear(hidden_states, self.embedding)
|
|
|
|
return {
|
|
"logits": logits,
|
|
"hidden_states": hidden_states
|
|
}
|
|
|