Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
datasets:
|
3 |
+
- abideen/Cosmopedia-100k-pretrain
|
4 |
+
tags:
|
5 |
+
- Mistral
|
6 |
+
- 1bit
|
7 |
+
- bitnet
|
8 |
+
- abideen
|
9 |
+
---
|
10 |
+
"""this is my first attempt at converting a model float16 quantized model to 1.5bit. i used alpindale/Mistral-7B-v0.2-hf for the base model and \n
|
11 |
+
trained on: abideen/cosmopedia-100k-pretain dataset and used his google colab project to make this"""
|
12 |
+
|
13 |
+
#EXAMPLE INFERENCE CODE FROM ABIDEEN'S COLAB PROJECT
|
14 |
+
```
|
15 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
16 |
+
from transformers.models.llama.modeling_llama import *
|
17 |
+
# Load a pretrained BitNet model
|
18 |
+
model = "liminerity/Bitnet-Mistral.0.2-70M"
|
19 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
20 |
+
model = AutoModelForCausalLM.from_pretrained(model)
|
21 |
+
|
22 |
+
|
23 |
+
def activation_quant(x):
|
24 |
+
scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
|
25 |
+
y = (x * scale).round().clamp_(-128, 127)
|
26 |
+
y = y / scale
|
27 |
+
return y
|
28 |
+
def weight_quant(w):
|
29 |
+
scale = 1.0 / w.abs().mean().clamp_(min=1e-5)
|
30 |
+
u = (w * scale).round().clamp_(-1, 1)
|
31 |
+
u = u / scale
|
32 |
+
return u
|
33 |
+
|
34 |
+
class BitLinear(nn.Linear):
|
35 |
+
def forward(self, x):
|
36 |
+
w = self.weight # a weight tensor with shape [d, k]
|
37 |
+
x = x.to(w.device)
|
38 |
+
RMSNorm = LlamaRMSNorm(x.shape[-1]).to(w.device)
|
39 |
+
x_norm = RMSNorm(x)
|
40 |
+
# A trick for implementing Straight−Through−Estimator (STE) using detach()
|
41 |
+
x_quant = x_norm + (activation_quant(x_norm) - x_norm).detach()
|
42 |
+
w_quant = w + (weight_quant(w) - w).detach()
|
43 |
+
y = F.linear(x_quant, w_quant)
|
44 |
+
return y
|
45 |
+
|
46 |
+
def convert_to_bitnet(model, copy_weights):
|
47 |
+
for name, module in model.named_modules():
|
48 |
+
# Replace linear layers with BitNet
|
49 |
+
if isinstance(module, LlamaSdpaAttention) or isinstance(module, LlamaMLP):
|
50 |
+
for child_name, child_module in module.named_children():
|
51 |
+
if isinstance(child_module, nn.Linear):
|
52 |
+
bitlinear = BitLinear(child_module.in_features, child_module.out_features, child_module.bias is not None).to(device="cuda:0")
|
53 |
+
if copy_weights:
|
54 |
+
bitlinear.weight = child_module.weight
|
55 |
+
if child_module.bias is not None:
|
56 |
+
bitlinear.bias = child_module.bias
|
57 |
+
setattr(module, child_name, bitlinear)
|
58 |
+
# Remove redundant input_layernorms
|
59 |
+
elif isinstance(module, LlamaDecoderLayer):
|
60 |
+
for child_name, child_module in module.named_children():
|
61 |
+
if isinstance(child_module, LlamaRMSNorm) and child_name == "input_layernorm":
|
62 |
+
setattr(module, child_name, nn.Identity().to(device="cuda:0"))
|
63 |
+
|
64 |
+
|
65 |
+
convert_to_bitnet(model, copy_weights=True)
|
66 |
+
model.to(device="cuda:0")
|
67 |
+
|
68 |
+
prompt = "What is Machine Learning?"
|
69 |
+
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
70 |
+
generate_ids = model.generate(inputs.input_ids, max_length=50)
|
71 |
+
tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
72 |
+
```
|