Update README.md
Browse files
README.md
CHANGED
@@ -54,141 +54,15 @@ The model was converted using [this notebook](https://nbviewer.org/urls/huggingf
|
|
54 |
### How to use
|
55 |
|
56 |
```sh
|
57 |
-
|
|
|
58 |
pip install bitsandbytes-cuda111==0.26.0
|
59 |
-
pip install datasets==1.16.1
|
60 |
```
|
61 |
|
62 |
```py
|
63 |
import transformers
|
64 |
-
|
65 |
import torch
|
66 |
-
|
67 |
-
from torch import nn
|
68 |
-
from torch.cuda.amp import custom_fwd, custom_bwd
|
69 |
-
|
70 |
-
from bitsandbytes.functional import quantize_blockwise, dequantize_blockwise
|
71 |
-
|
72 |
-
from tqdm.auto import tqdm
|
73 |
-
|
74 |
-
class FrozenBNBLinear(nn.Module):
|
75 |
-
def __init__(self, weight, absmax, code, bias=None):
|
76 |
-
assert isinstance(bias, nn.Parameter) or bias is None
|
77 |
-
super().__init__()
|
78 |
-
self.out_features, self.in_features = weight.shape
|
79 |
-
self.register_buffer("weight", weight.requires_grad_(False))
|
80 |
-
self.register_buffer("absmax", absmax.requires_grad_(False))
|
81 |
-
self.register_buffer("code", code.requires_grad_(False))
|
82 |
-
self.adapter = None
|
83 |
-
self.bias = bias
|
84 |
-
|
85 |
-
def forward(self, input):
|
86 |
-
output = DequantizeAndLinear.apply(input, self.weight, self.absmax, self.code, self.bias)
|
87 |
-
if self.adapter:
|
88 |
-
output += self.adapter(input)
|
89 |
-
return output
|
90 |
-
|
91 |
-
@classmethod
|
92 |
-
def from_linear(cls, linear: nn.Linear) -> "FrozenBNBLinear":
|
93 |
-
weights_int8, state = quantize_blockise_lowmemory(linear.weight)
|
94 |
-
return cls(weights_int8, *state, linear.bias)
|
95 |
-
|
96 |
-
def __repr__(self):
|
97 |
-
return f"{self.__class__.__name__}({self.in_features}, {self.out_features})"
|
98 |
-
|
99 |
-
|
100 |
-
class DequantizeAndLinear(torch.autograd.Function):
|
101 |
-
@staticmethod
|
102 |
-
@custom_fwd
|
103 |
-
def forward(ctx, input: torch.Tensor, weights_quantized: torch.ByteTensor,
|
104 |
-
absmax: torch.FloatTensor, code: torch.FloatTensor, bias: torch.FloatTensor):
|
105 |
-
weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)
|
106 |
-
ctx.save_for_backward(input, weights_quantized, absmax, code)
|
107 |
-
ctx._has_bias = bias is not None
|
108 |
-
return F.linear(input, weights_deq, bias)
|
109 |
-
|
110 |
-
@staticmethod
|
111 |
-
@custom_bwd
|
112 |
-
def backward(ctx, grad_output: torch.Tensor):
|
113 |
-
assert not ctx.needs_input_grad[1] and not ctx.needs_input_grad[2] and not ctx.needs_input_grad[3]
|
114 |
-
input, weights_quantized, absmax, code = ctx.saved_tensors
|
115 |
-
# grad_output: [*batch, out_features]
|
116 |
-
weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)
|
117 |
-
grad_input = grad_output @ weights_deq
|
118 |
-
grad_bias = grad_output.flatten(0, -2).sum(dim=0) if ctx._has_bias else None
|
119 |
-
return grad_input, None, None, None, grad_bias
|
120 |
-
|
121 |
-
|
122 |
-
class FrozenBNBEmbedding(nn.Module):
|
123 |
-
def __init__(self, weight, absmax, code):
|
124 |
-
super().__init__()
|
125 |
-
self.num_embeddings, self.embedding_dim = weight.shape
|
126 |
-
self.register_buffer("weight", weight.requires_grad_(False))
|
127 |
-
self.register_buffer("absmax", absmax.requires_grad_(False))
|
128 |
-
self.register_buffer("code", code.requires_grad_(False))
|
129 |
-
self.adapter = None
|
130 |
-
|
131 |
-
def forward(self, input, **kwargs):
|
132 |
-
with torch.no_grad():
|
133 |
-
# note: both quantuized weights and input indices are *not* differentiable
|
134 |
-
weight_deq = dequantize_blockwise(self.weight, absmax=self.absmax, code=self.code)
|
135 |
-
output = F.embedding(input, weight_deq, **kwargs)
|
136 |
-
if self.adapter:
|
137 |
-
output += self.adapter(input)
|
138 |
-
return output
|
139 |
-
|
140 |
-
@classmethod
|
141 |
-
def from_embedding(cls, embedding: nn.Embedding) -> "FrozenBNBEmbedding":
|
142 |
-
weights_int8, state = quantize_blockise_lowmemory(embedding.weight)
|
143 |
-
return cls(weights_int8, *state)
|
144 |
-
|
145 |
-
def __repr__(self):
|
146 |
-
return f"{self.__class__.__name__}({self.num_embeddings}, {self.embedding_dim})"
|
147 |
-
|
148 |
-
|
149 |
-
def quantize_blockise_lowmemory(matrix: torch.Tensor, chunk_size: int = 2 ** 20):
|
150 |
-
assert chunk_size % 4096 == 0
|
151 |
-
code = None
|
152 |
-
chunks = []
|
153 |
-
absmaxes = []
|
154 |
-
flat_tensor = matrix.view(-1)
|
155 |
-
for i in range((matrix.numel() - 1) // chunk_size + 1):
|
156 |
-
input_chunk = flat_tensor[i * chunk_size: (i + 1) * chunk_size].clone()
|
157 |
-
quantized_chunk, (absmax_chunk, code) = quantize_blockwise(input_chunk, code=code)
|
158 |
-
chunks.append(quantized_chunk)
|
159 |
-
absmaxes.append(absmax_chunk)
|
160 |
-
|
161 |
-
matrix_i8 = torch.cat(chunks).reshape_as(matrix)
|
162 |
-
absmax = torch.cat(absmaxes)
|
163 |
-
return matrix_i8, (absmax, code)
|
164 |
-
|
165 |
-
|
166 |
-
def convert_to_int8(model):
|
167 |
-
"""Convert linear and embedding modules to 8-bit with optional adapters"""
|
168 |
-
for module in list(model.modules()):
|
169 |
-
for name, child in module.named_children():
|
170 |
-
if isinstance(child, nn.Linear):
|
171 |
-
print(name, child)
|
172 |
-
setattr(
|
173 |
-
module,
|
174 |
-
name,
|
175 |
-
FrozenBNBLinear(
|
176 |
-
weight=torch.zeros(child.out_features, child.in_features, dtype=torch.uint8),
|
177 |
-
absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),
|
178 |
-
code=torch.zeros(256),
|
179 |
-
bias=child.bias,
|
180 |
-
),
|
181 |
-
)
|
182 |
-
elif isinstance(child, nn.Embedding):
|
183 |
-
setattr(
|
184 |
-
module,
|
185 |
-
name,
|
186 |
-
FrozenBNBEmbedding(
|
187 |
-
weight=torch.zeros(child.num_embeddings, child.embedding_dim, dtype=torch.uint8),
|
188 |
-
absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),
|
189 |
-
code=torch.zeros(256),
|
190 |
-
)
|
191 |
-
)
|
192 |
|
193 |
class GPTJBlock(transformers.models.gptj.modeling_gptj.GPTJBlock):
|
194 |
def __init__(self, config):
|
@@ -210,18 +84,18 @@ class GPTJForCausalLM(transformers.models.gptj.modeling_gptj.GPTJForCausalLM):
|
|
210 |
convert_to_int8(self)
|
211 |
|
212 |
|
|
|
|
|
213 |
transformers.models.gptj.modeling_gptj.GPTJBlock = GPTJBlock # monkey-patch GPT-J
|
214 |
|
215 |
-
config = transformers.GPTJConfig.from_pretrained("mrm8488/bertin-gpt-j-6B-ES-8bit")
|
216 |
tokenizer = transformers.AutoTokenizer.from_pretrained("mrm8488/bertin-gpt-j-6B-ES-8bit")
|
|
|
217 |
|
218 |
-
gpt = GPTJForCausalLM.from_pretrained("mrm8488/bertin-gpt-j-6B-ES-8bit", low_cpu_mem_usage=True)
|
219 |
-
|
220 |
-
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
221 |
-
gpt.to(device)
|
222 |
|
223 |
prompt = tokenizer("El sentido de la vida es", return_tensors='pt')
|
224 |
prompt = {key: value.to(device) for key, value in prompt.items()}
|
225 |
-
|
|
|
|
|
226 |
print(tokenizer.decode(out[0]))
|
227 |
```
|
|
|
54 |
### How to use
|
55 |
|
56 |
```sh
|
57 |
+
wget https://huggingface.co/mrm8488/bertin-gpt-j-6B-ES-8bit/resolve/main/utils.py -O utils.py
|
58 |
+
pip install transformers
|
59 |
pip install bitsandbytes-cuda111==0.26.0
|
|
|
60 |
```
|
61 |
|
62 |
```py
|
63 |
import transformers
|
|
|
64 |
import torch
|
65 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
class GPTJBlock(transformers.models.gptj.modeling_gptj.GPTJBlock):
|
68 |
def __init__(self, config):
|
|
|
84 |
convert_to_int8(self)
|
85 |
|
86 |
|
87 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
88 |
+
|
89 |
transformers.models.gptj.modeling_gptj.GPTJBlock = GPTJBlock # monkey-patch GPT-J
|
90 |
|
|
|
91 |
tokenizer = transformers.AutoTokenizer.from_pretrained("mrm8488/bertin-gpt-j-6B-ES-8bit")
|
92 |
+
model = GPTJForCausalLM.from_pretrained("hivemind/gpt-j-6B-8bit", pad_token_id=tokenizer.eos_token_id, low_cpu_mem_usage=True).to(device)
|
93 |
|
|
|
|
|
|
|
|
|
94 |
|
95 |
prompt = tokenizer("El sentido de la vida es", return_tensors='pt')
|
96 |
prompt = {key: value.to(device) for key, value in prompt.items()}
|
97 |
+
|
98 |
+
out = model.generate(**prompt, max_length=64, do_sample=True)
|
99 |
+
|
100 |
print(tokenizer.decode(out[0]))
|
101 |
```
|