Spaces:
Runtime error
Runtime error
File size: 8,649 Bytes
35153f6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 |
from typing import Optional, Tuple
import warnings
import torch
import transformers
from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv, rotate_half
try:
from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
except ImportError:
from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
from flash_attn.bert_padding import unpad_input, pad_input
from flash_attn import __version__ as flash_attn_version
from flash_attn.flash_attn_interface import (
flash_attn_func,
flash_attn_varlen_kvpacked_func,
)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: bool = False,
use_cache: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if output_attentions:
warnings.warn(
"Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
)
bsz, q_len, _ = hidden_states.size()
query_states = (
self.q_proj(hidden_states)
.view(bsz, q_len, self.num_heads, self.head_dim)
.transpose(1, 2)
)
key_states = (
self.k_proj(hidden_states)
.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
.transpose(1, 2)
)
value_states = (
self.v_proj(hidden_states)
.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
.transpose(1, 2)
) # shape: (b, num_heads, s, head_dim)
kv_seq_len = key_states.shape[-2]
if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[-2]
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(
query_states, key_states, cos, sin, position_ids
)
if past_key_value is not None:
# reuse k, v
key_states = torch.cat([past_key_value[0], key_states], dim=2)
value_states = torch.cat([past_key_value[1], value_states], dim=2)
past_key_value = (key_states, value_states) if use_cache else None
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
# Transform the data into the format required by flash attention
qkv = torch.stack([query_states, key_states, value_states], dim=2)
qkv = qkv.transpose(1, 3) # shape: [b, s, 3, num_heads, head_dim]
key_padding_mask = attention_mask
if key_padding_mask is None:
qkv = qkv.reshape(-1, 3, self.num_heads, self.head_dim)
cu_q_lens = torch.arange(
0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
)
max_s = q_len
output = flash_attn_unpadded_qkvpacked_func(
qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
)
output = output.view(bsz, q_len, -1)
else:
qkv = qkv.reshape(bsz, q_len, -1)
qkv, indices, cu_q_lens, max_s = unpad_input(qkv, key_padding_mask)
qkv = qkv.view(-1, 3, self.num_heads, self.head_dim)
output_unpad = flash_attn_unpadded_qkvpacked_func(
qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
)
output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
output = pad_input(output_unpad, indices, bsz, q_len)
return self.o_proj(output), None, past_key_value
def apply_rotary_pos_emb_inference(q, k, cos_sin, position_ids):
gather_indices = position_ids[:, :, None, None] # [bsz, seq_len, 1, 1]
gather_indices = gather_indices.repeat(
1, 1, cos_sin[0].shape[1], cos_sin[0].shape[3]
)
bsz = gather_indices.shape[0]
cos, sin = (
torch.gather(x.transpose(1, 2).repeat(bsz, 1, 1, 1), 1, gather_indices)
for x in cos_sin
)
q, k = ((x * cos) + (rotate_half(x) * sin) for x in (q, k))
return q, k
def forward_inference(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: bool = False,
use_cache: bool = False,
padding_mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if output_attentions:
warnings.warn(
"Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
)
bsz, q_len, _ = hidden_states.size()
kv_heads = getattr(self, "num_key_value_heads", self.num_heads)
q, k, v = (
op(hidden_states).view(bsz, q_len, nh, self.head_dim)
for op, nh in (
(self.q_proj, self.num_heads),
(self.k_proj, kv_heads),
(self.v_proj, kv_heads),
)
)
# shape: (b, s, num_heads, head_dim)
kv_seq_len = k.shape[1]
past_kv_len = 0
if past_key_value is not None:
past_kv_len = past_key_value[0].shape[2]
kv_seq_len += past_kv_len
cos_sin = self.rotary_emb(v, seq_len=kv_seq_len)
q, k = apply_rotary_pos_emb_inference(q, k, cos_sin, position_ids)
if past_key_value is not None:
assert (
flash_attn_version >= "2.1.0"
), "past_key_value support requires flash-attn >= 2.1.0"
# reuse k, v
k = torch.cat([past_key_value[0].transpose(1, 2), k], dim=1)
v = torch.cat([past_key_value[1].transpose(1, 2), v], dim=1)
past_key_value = (k.transpose(1, 2), v.transpose(1, 2)) if use_cache else None
if attention_mask is None:
output = flash_attn_func(q, k, v, 0.0, softmax_scale=None, causal=True).view(
bsz, q_len, -1
)
else:
q, indices, cu_q_lens, max_s = unpad_input(q, attention_mask[:, -q_len:])
# We can skip concat and call unpad twice but seems better to call unpad only once.
kv, _, cu_k_lens, max_k = unpad_input(
torch.stack((k, v), dim=2), attention_mask
)
output_unpad = flash_attn_varlen_kvpacked_func(
q,
kv,
cu_q_lens,
cu_k_lens,
max_s,
max_k,
0.0,
softmax_scale=None,
causal=True,
)
output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
output = pad_input(output_unpad, indices, bsz, q_len)
return self.o_proj(output), None, past_key_value
# Disable the transformation of the attention mask in LlamaModel as the flash attention
# requires the attention mask to be the same as the key_padding_mask
def _prepare_decoder_attention_mask(
self, attention_mask, input_shape, inputs_embeds, past_key_values_length
):
# [bsz, seq_len]
return attention_mask
def _prepare_decoder_attention_mask_inference(
self, attention_mask, input_shape, inputs_embeds, past_key_values_length
):
# [bsz, seq_len]
if past_key_values_length > 0 and attention_mask is not None:
attention_mask = torch.cat(
(
torch.full(
(input_shape[0], past_key_values_length),
True,
dtype=attention_mask.dtype,
device=attention_mask.device,
),
attention_mask,
),
dim=-1,
)
if attention_mask is not None and torch.all(attention_mask):
return None # This uses the faster call when training with full samples
def replace_llama_attn_with_flash_attn(inference=False):
cuda_major, cuda_minor = torch.cuda.get_device_capability()
if cuda_major < 8:
warnings.warn(
"Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
"ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
)
if inference:
transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = _prepare_decoder_attention_mask_inference
transformers.models.llama.modeling_llama.LlamaAttention.forward = forward_inference
else:
transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
_prepare_decoder_attention_mask
)
transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
|