File size: 9,907 Bytes
02bc7b8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 | import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from awq.quantize import W8A8OF16LinearDynamicInputScale
from llava.model.multimodal_encoder.siglip.modeling_siglip import (
SiglipMLP,
SiglipEncoder,
SiglipAttention,
SiglipEncoderLayer,
)
from tinychat.utils.input_metadata import ActivationBuffer
from transformers.modeling_outputs import BaseModelOutput
from typing import Optional, Tuple, Union
from flash_attn import flash_attn_func
import time
CLIP_RANGE = 5
import awq_inference_engine
class QuantSiglipEncoder(nn.Module):
def __init__(self, module: SiglipEncoder, bsz=64, seqlen=1024):
super().__init__()
self.config = module.config
self.layers = [QuantSiglipEncoderLayer(layer) for layer in module.layers]
self.buffer = ActivationBuffer(module)
self.bsz = bsz
self.seqlen = seqlen
self.buffer.allocate_activation_buffer(self.bsz * self.seqlen)
# Ignore copy
def forward(
self,
inputs_embeds,
attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None, # dummy
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutput]:
# TODO Find why this code is necessary
# torch.sum(inputs_embeds!=inputs_embeds)
bsz, seqlen, _ = inputs_embeds.shape
if self.bsz != bsz or self.seqlen != seqlen:
self.buffer.allocate_activation_buffer(bsz * seqlen)
self.bsz = bsz
self.seqlen = seqlen
output_hidden_states = (
output_hidden_states
if output_hidden_states is not None
else self.config.output_hidden_states
)
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
encoder_states = () if output_hidden_states else None
hidden_states = inputs_embeds
for i, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (
hidden_states.reshape(bsz, seqlen, -1),
)
hidden_states = encoder_layer(
hidden_states, self.buffer, attention_mask, bsz, seqlen
)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states.reshape(bsz, seqlen, -1),)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states.reshape(bsz, seqlen, -1),
hidden_states=encoder_states,
attentions=None,
)
class QuantSiglipMLP(nn.Module):
def __init__(self, siglipmlp, init_only=False):
super().__init__()
self.config = siglipmlp.config
self.activation_fn = siglipmlp.activation_fn
self.fc1 = W8A8OF16LinearDynamicInputScale.from_linear(
siglipmlp.fc1, init_only=init_only, fc1=False
)
self.fc2 = W8A8OF16LinearDynamicInputScale.from_linear(
siglipmlp.fc2, init_only=init_only
)
self.invoke_quant = self.invoke_quant_mlp
def invoke_quant_mlp(self, buffer, actfn_output):
awq_inference_engine.invoke_quant(
buffer.quantized_mlp_act_buffer,
actfn_output,
buffer.quantized_scale_buffer,
)
def forward(self, buffer: ActivationBuffer) -> torch.Tensor:
# INT8 in, FP16 out
self.fc1(
buffer.quantized_hidden_states_buffer,
buffer.quantized_scale_buffer,
buffer.fc1_buffer,
)
# Act & quantization
awq_inference_engine.gelu_and_quant(
buffer.quantized_mlp_act_buffer,
buffer.fc1_buffer,
buffer.quantized_scale_buffer,
buffer.tmp,
)
# INT8 in, FP16 out
self.fc2(
buffer.quantized_mlp_act_buffer,
buffer.quantized_scale_buffer,
buffer.in_out_fc2_act_buffer,
)
class QuantSiglipFlashAttention2(nn.Module):
def __init__(
self,
module: SiglipAttention,
init_only=False,
):
super().__init__()
self.config = module.config
self.embed_dim = module.embed_dim
self.num_heads = module.num_heads
self.head_dim = self.embed_dim // self.num_heads
self.qkv_proj = W8A8OF16LinearDynamicInputScale.from_qkv(
module.q_proj, module.k_proj, module.v_proj, init_only=init_only
)
self.out_proj = W8A8OF16LinearDynamicInputScale.from_linear(
module.out_proj, init_only=init_only
)
self.invoke_quant = self.invoke_quant_wo
def invoke_quant_wo(self, buffer, attn_output):
awq_inference_engine.invoke_quant(
buffer.quantized_hidden_states_buffer,
attn_output,
buffer.quantized_scale_buffer,
)
# Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
def forward(
self, buffer: ActivationBuffer, bsz=64, seqlen=1024
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
# qkv
self.qkv_proj(
buffer.quantized_hidden_states_buffer,
buffer.quantized_scale_buffer,
buffer.qkv_proj_act_buffer,
)
q, k, v = buffer.qkv_proj_act_buffer.split(
[self.embed_dim, self.embed_dim, self.embed_dim], dim=-1
)
q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim)
k = k.reshape(bsz, seqlen, self.num_heads, self.head_dim)
v = v.reshape(bsz, seqlen, self.num_heads, self.head_dim)
attn_output = flash_attn_func(q, k, v, softmax_scale=None, causal=False)
attn_output = attn_output.reshape(bsz * seqlen, -1)
# FP16 -> int8
self.invoke_quant(buffer, attn_output)
# INT8 in, FP16 out
self.out_proj(
buffer.quantized_hidden_states_buffer,
buffer.quantized_scale_buffer,
buffer.in_out_fc2_act_buffer,
)
class QuantSiglipEncoderLayer(nn.Module):
def __init__(self, module: SiglipEncoderLayer):
super().__init__()
self.embed_dim = module.embed_dim
self.self_attn = QuantSiglipFlashAttention2(module.self_attn)
self.layer_norm1 = RMSNormGeneral(
module.layer_norm1.weight.data,
module.layer_norm1.bias.data,
module.layer_norm1.eps,
True,
).cuda()
self.mlp = QuantSiglipMLP(module.mlp)
self.layer_norm2 = RMSNormGeneral(
module.layer_norm2.weight.data,
module.layer_norm2.bias.data,
module.layer_norm2.eps,
True,
).cuda()
self.quant = self.invoke_quant_norm
def invoke_quant_norm(self, buffer, normfn_output):
awq_inference_engine.invoke_quant(
buffer.quantized_hidden_states_buffer,
normfn_output,
buffer.quantized_scale_buffer,
)
def forward(
self,
hidden_states: torch.Tensor,
buffer: ActivationBuffer,
attention_mask,
bsz,
seqlen,
) -> Tuple[torch.FloatTensor]:
# Attention block
# FP16 in int8 out, layernorm & quantization
residual = hidden_states
self.layer_norm1(
hidden_states.reshape(-1, self.embed_dim),
buffer.quantized_hidden_states_buffer,
buffer.quantized_scale_buffer,
)
# INT8 -> FP16
self.self_attn(buffer, bsz, seqlen)
hidden_states = (
residual.reshape(-1, self.embed_dim) + buffer.in_out_fc2_act_buffer
)
# Fully Connected
residual = hidden_states
# FP16 in int8 out, layernorm & quantization
self.layer_norm2(
hidden_states.reshape(-1, self.embed_dim),
buffer.quantized_hidden_states_buffer,
buffer.quantized_scale_buffer,
)
# INT8 -> FP16
self.mlp(buffer)
hidden_states = (
residual.reshape(-1, self.embed_dim) + buffer.in_out_fc2_act_buffer
)
return hidden_states
class RMSNormGeneral(nn.Module):
"""Root mean square normalization (w/ per-token or per-tensor quant).
Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
Refer to https://arxiv.org/abs/1910.07467
"""
def __init__(
self,
weight: torch.tensor,
bias: torch.tensor,
eps: float = 1e-6,
use_per_token_quant: bool = True,
) -> None:
super().__init__()
self.weight = nn.Parameter(weight, requires_grad=False)
self.bias = nn.Parameter(bias, requires_grad=False)
self.variance_epsilon = eps
self.use_per_token_quant = use_per_token_quant
def forward(
self,
x: torch.Tensor,
quantized_hidden_states_buffer: torch.Tensor,
quantized_scale_buffer: torch.Tensor,
quantized_sum_buffer: torch.Tensor = None,
) -> torch.Tensor:
# quantized_sum_buffer is not used, only to keep the consistency of the interface
awq_inference_engine.rms_norm_general(
quantized_hidden_states_buffer,
x,
self.weight.data,
self.bias.data,
quantized_scale_buffer,
self.variance_epsilon,
self.use_per_token_quant,
)
|