Update README.md
Browse files
README.md
CHANGED
@@ -38,7 +38,7 @@ MODEL_NAME = "p1atdev/dart-v1-base"
|
|
38 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True) # trust_remote_code is required for tokenizer
|
39 |
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)
|
40 |
|
41 |
-
prompt = "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general>1girl
|
42 |
inputs = tokenizer(prompt, return_tensors="pt").input_ids
|
43 |
|
44 |
with torch.no_grad():
|
@@ -48,6 +48,23 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|
48 |
# rating:sfw, rating:general, original, 1girl, ahoge, black hair, blue eyes, blush, closed mouth, ear piercing, earrings, jewelry, looking at viewer, mole, mole under eye, piercing, portrait, shirt, short hair, solo, white shirt
|
49 |
```
|
50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
#### Flash attention (optional)
|
52 |
|
53 |
Using flash attention can optimize computations, but it is currently only compatible with Linux.
|
@@ -86,8 +103,12 @@ ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME)
|
|
86 |
# qunatized version
|
87 |
# ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME, file_name="model_quantized.onnx")
|
88 |
|
89 |
-
|
90 |
-
|
|
|
|
|
|
|
|
|
91 |
|
92 |
with torch.no_grad():
|
93 |
outputs = model.generate(inputs, generation_config=generation_config)
|
|
|
38 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True) # trust_remote_code is required for tokenizer
|
39 |
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)
|
40 |
|
41 |
+
prompt = "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general>1girl"
|
42 |
inputs = tokenizer(prompt, return_tensors="pt").input_ids
|
43 |
|
44 |
with torch.no_grad():
|
|
|
48 |
# rating:sfw, rating:general, original, 1girl, ahoge, black hair, blue eyes, blush, closed mouth, ear piercing, earrings, jewelry, looking at viewer, mole, mole under eye, piercing, portrait, shirt, short hair, solo, white shirt
|
49 |
```
|
50 |
|
51 |
+
You can use `tokenizer.apply_chat_template` to simplify constructiing of prompts:
|
52 |
+
|
53 |
+
```py
|
54 |
+
inputs = tokenizer.apply_chat_template({
|
55 |
+
"rating": "rating:sfw, rating:general",
|
56 |
+
"copyright": "original",
|
57 |
+
"character": "",
|
58 |
+
"general": "1girl"
|
59 |
+
}, tokenize=True) # tokenize=False to preview prompt
|
60 |
+
# same as input_ids of "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general>1girl"
|
61 |
+
|
62 |
+
with torch.no_grad():
|
63 |
+
outputs = model.generate(inputs, generation_config=generation_config)
|
64 |
+
```
|
65 |
+
|
66 |
+
See [chat_templating document](https://huggingface.co/docs/transformers/main/en/chat_templating) for more detail about `apply_chat_template`.
|
67 |
+
|
68 |
#### Flash attention (optional)
|
69 |
|
70 |
Using flash attention can optimize computations, but it is currently only compatible with Linux.
|
|
|
103 |
# qunatized version
|
104 |
# ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME, file_name="model_quantized.onnx")
|
105 |
|
106 |
+
inputs = tokenizer.apply_chat_template({
|
107 |
+
"rating": "rating:sfw, rating:general",
|
108 |
+
"copyright": "original",
|
109 |
+
"character": "",
|
110 |
+
"general": "1girl"
|
111 |
+
}, tokenize=True)
|
112 |
|
113 |
with torch.no_grad():
|
114 |
outputs = model.generate(inputs, generation_config=generation_config)
|