Spaces:
Runtime error
Runtime error
File size: 8,725 Bytes
c7a96cd |
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 |
import torch
import torch.distributed
from typing import List, Optional, Tuple
from accelerate import init_empty_weights
from safetensors import safe_open
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
AutoConfig,
)
from transformers.models.opt.parallel_layers import (
TensorParallelColumnLinear,
TensorParallelEmbedding,
TensorParallelRowLinear,
)
from text_generation_server.models import CausalLM
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
)
HAS_BITS_AND_BYTES = True
try:
import bitsandbytes as bnb
from bitsandbytes.nn import Int8Params
except Exception as e:
HAS_BITS_AND_BYTES = False
class OPT(CausalLM):
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
) -> Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]]:
"""Overwrite forward to ignore position_ids"""
# Model Forward
outputs = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, outputs.past_key_values
class OPTSharded(OPT):
def __init__(
self, model_id: str, revision: Optional[str] = None, quantize: bool = False
):
self.process_group, self.rank, self.world_size = initialize_torch_distributed()
self.master = self.rank == 0
if torch.cuda.is_available():
device = torch.device(f"cuda:{self.rank}")
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
else:
device = torch.device("cpu")
dtype = torch.float32
tokenizer = AutoTokenizer.from_pretrained(
model_id, revision=revision, padding_side="left", truncation_side="left"
)
config = AutoConfig.from_pretrained(
model_id, revision=revision, tp_parallel=True
)
tokenizer.pad_token_id = config.pad_token_id
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
with init_empty_weights():
model = AutoModelForCausalLM.from_config(config)
torch.distributed.barrier(group=self.process_group)
self.load_weights(
model,
filenames,
quantize=quantize,
device=device,
dtype=dtype,
rank=self.rank,
world_size=self.world_size,
)
self.model = model.eval()
torch.distributed.barrier(group=self.process_group)
super(CausalLM, self).__init__(
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
)
@staticmethod
def load_weights(
model,
filenames: List[str],
quantize: bool,
device: torch.device,
dtype: torch.dtype,
rank: int,
world_size: int,
):
parameters = dict(model.named_parameters())
for file in filenames:
with safe_open(
file, framework="pt", device=str(device) if not quantize else "cpu"
) as f:
for name in f.keys():
if name == "lm_head.weight":
continue
module_name, param_name = name.rsplit(".", 1)
module = model.get_submodule(module_name)
current_tensor = parameters[name]
slice_ = f.get_slice(name)
if isinstance(module, TensorParallelColumnLinear):
size = slice_.get_shape()[0]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[start:stop]
elif isinstance(module, TensorParallelRowLinear):
if param_name == "weight":
size = slice_.get_shape()[1]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[:, start:stop]
else:
tensor = slice_[:]
# XXX: Hack for Rowlinear to add the bias only once.
if rank != 0:
tensor = torch.zeros_like(tensor)
elif isinstance(module, TensorParallelEmbedding):
size = slice_.get_shape()[0]
block_size = size // world_size
start = rank * block_size
stop = (rank + 1) * block_size
tensor = slice_[start:stop]
else:
tensor = slice_[:]
if current_tensor.shape != tensor.shape:
raise ValueError(
f"Name {name} -- Current {current_tensor.shape} and got {tensor.shape}"
)
tensor = tensor.contiguous().to(dtype)
if quantize:
if not HAS_BITS_AND_BYTES:
raise ImportError(
"bitsandbytes is not available on your machine either because it is not installed "
"or you don't have a GPU.\n"
"You can install it with `pip install bitsandbytes`."
)
if (
type(module)
in [TensorParallelRowLinear, TensorParallelColumnLinear]
and param_name == "weight"
):
tensor = Int8Params(
tensor,
has_fp16_weights=False,
requires_grad=False,
).to(device)
state = bnb.MatmulLtState()
state.threshold = 6.0
state.has_fp16_weights = False
state.memory_efficient_backward = False
state.use_pool = True
state.CB = tensor.CB
state.SCB = tensor.SCB
tensor.CB = None
tensor.SCB = None
def replace_linear(state):
def linear(input, weight, bias):
out = bnb.matmul(
input,
weight,
state=state,
threshold=state.threshold,
bias=bias,
)
if state.CB is not None:
# we converted 8-bit row major to turing/ampere format
# in the first inference pass
# we no longer need the row-major weight
del state.CB
weight.data = state.CxB
return out
return linear
module.linear = replace_linear(state)
else:
tensor = tensor.to(device)
module._parameters[param_name] = tensor
if name == "model.decoder.embed_tokens.weight":
model.lm_head._parameters["weight"] = tensor
def forward(
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
):
outputs = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
use_cache=True,
)
# Logits are sharded, so we need to gather them
logits = [torch.empty_like(outputs.logits) for _ in range(self.world_size)]
torch.distributed.all_gather(logits, outputs.logits, group=self.process_group)
logits = torch.cat(logits, dim=2)
return logits, outputs.past_key_values
|