Spaces:
Sleeping
Sleeping
Commit
·
2ab342b
1
Parent(s):
762a224
let's go field decoding
Browse files
app.py
CHANGED
@@ -15,10 +15,10 @@ huggingface_hub.login(token=hf_key)
|
|
15 |
|
16 |
tokenizer = AutoTokenizer.from_pretrained("bigcode/starcoderbase-3b")
|
17 |
vardecoder_model = AutoModelForCausalLM.from_pretrained(
|
18 |
-
"ejschwartz/resym-vardecoder", torch_dtype=torch.bfloat16
|
19 |
).to("cuda")
|
20 |
fielddecoder_model = AutoModelForCausalLM.from_pretrained(
|
21 |
-
"ejschwartz/resym-fielddecoder", torch_dtype=torch.bfloat16
|
22 |
).to("cuda")
|
23 |
|
24 |
gradio_client = Client("https://ejschwartz-resym-field-helper.hf.space/")
|
@@ -42,10 +42,12 @@ def field_prompt(code):
|
|
42 |
print(f"fields: {fields}")
|
43 |
|
44 |
prompt = f"```\n{code}\n```\nWhat are the variable name and type for the following memory accesses:{', '.join(fields)}?\n"
|
|
|
|
|
45 |
|
46 |
print(f"field prompt: {prompt}")
|
47 |
|
48 |
-
return prompt, field_helper_result
|
49 |
|
50 |
@spaces.GPU
|
51 |
def infer(code):
|
@@ -65,18 +67,18 @@ def infer(code):
|
|
65 |
|
66 |
varstring = ", ".join([f"`{v}`" for v in vars])
|
67 |
|
68 |
-
|
69 |
|
70 |
# ejs: Yeah, this var_name thing is really bizarre. But look at https://github.com/lt-asset/resym/blob/main/training_src/fielddecoder_inf.py
|
71 |
-
var_prompt = f"What are the original name and data types of variables {varstring}?\n```\n{code}\n```{
|
72 |
|
73 |
print(f"Prompt:\n{var_prompt}")
|
74 |
|
75 |
-
|
76 |
:, : 8192 - 1024
|
77 |
]
|
78 |
var_output = vardecoder_model.generate(
|
79 |
-
input_ids=
|
80 |
max_new_tokens=1024,
|
81 |
num_beams=4,
|
82 |
num_return_sequences=1,
|
@@ -86,32 +88,36 @@ def infer(code):
|
|
86 |
eos_token_id=0,
|
87 |
)[0]
|
88 |
var_output = tokenizer.decode(
|
89 |
-
var_output[
|
90 |
skip_special_tokens=True,
|
91 |
clean_up_tokenization_spaces=True,
|
92 |
)
|
93 |
|
94 |
-
field_prompt_result, field_helper_result = field_prompt(code)
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
|
|
|
|
|
|
|
|
115 |
|
116 |
|
117 |
demo = gr.Interface(
|
@@ -121,7 +127,7 @@ demo = gr.Interface(
|
|
121 |
],
|
122 |
outputs=[
|
123 |
gr.Text(label="Var Decoder Output"),
|
124 |
-
|
125 |
gr.Text(label="Generated Variable List"),
|
126 |
],
|
127 |
description=frontmatter.load("README.md").content,
|
|
|
15 |
|
16 |
tokenizer = AutoTokenizer.from_pretrained("bigcode/starcoderbase-3b")
|
17 |
vardecoder_model = AutoModelForCausalLM.from_pretrained(
|
18 |
+
"ejschwartz/resym-vardecoder", torch_dtype=torch.bfloat16
|
19 |
).to("cuda")
|
20 |
fielddecoder_model = AutoModelForCausalLM.from_pretrained(
|
21 |
+
"ejschwartz/resym-fielddecoder", torch_dtype=torch.bfloat16
|
22 |
).to("cuda")
|
23 |
|
24 |
gradio_client = Client("https://ejschwartz-resym-field-helper.hf.space/")
|
|
|
42 |
print(f"fields: {fields}")
|
43 |
|
44 |
prompt = f"```\n{code}\n```\nWhat are the variable name and type for the following memory accesses:{', '.join(fields)}?\n"
|
45 |
+
if len(fields) > 0:
|
46 |
+
prompt += f"{fields[0]}:"
|
47 |
|
48 |
print(f"field prompt: {prompt}")
|
49 |
|
50 |
+
return prompt, fields, field_helper_result
|
51 |
|
52 |
@spaces.GPU
|
53 |
def infer(code):
|
|
|
67 |
|
68 |
varstring = ", ".join([f"`{v}`" for v in vars])
|
69 |
|
70 |
+
first_var = vars[0]
|
71 |
|
72 |
# ejs: Yeah, this var_name thing is really bizarre. But look at https://github.com/lt-asset/resym/blob/main/training_src/fielddecoder_inf.py
|
73 |
+
var_prompt = f"What are the original name and data types of variables {varstring}?\n```\n{code}\n```{first_var}"
|
74 |
|
75 |
print(f"Prompt:\n{var_prompt}")
|
76 |
|
77 |
+
var_input_ids = tokenizer.encode(var_prompt, return_tensors="pt").cuda()[
|
78 |
:, : 8192 - 1024
|
79 |
]
|
80 |
var_output = vardecoder_model.generate(
|
81 |
+
input_ids=var_input_ids,
|
82 |
max_new_tokens=1024,
|
83 |
num_beams=4,
|
84 |
num_return_sequences=1,
|
|
|
88 |
eos_token_id=0,
|
89 |
)[0]
|
90 |
var_output = tokenizer.decode(
|
91 |
+
var_output[var_input_ids.size(1) :],
|
92 |
skip_special_tokens=True,
|
93 |
clean_up_tokenization_spaces=True,
|
94 |
)
|
95 |
|
96 |
+
field_prompt_result, fields, field_helper_result = field_prompt(code)
|
97 |
+
field_input_ids = tokenizer.encode(field_prompt_result, return_tensors="pt").cuda()[
|
98 |
+
:, : 8192 - 1024
|
99 |
+
]
|
100 |
+
|
101 |
+
field_output = fielddecoder_model.generate(
|
102 |
+
input_ids=field_input_ids,
|
103 |
+
max_new_tokens=1024,
|
104 |
+
num_beams=4,
|
105 |
+
num_return_sequences=1,
|
106 |
+
do_sample=False,
|
107 |
+
early_stopping=False,
|
108 |
+
pad_token_id=0,
|
109 |
+
eos_token_id=0,
|
110 |
+
)[0]
|
111 |
+
field_output = tokenizer.decode(
|
112 |
+
field_output[var_input_ids.size(1) :],
|
113 |
+
skip_special_tokens=True,
|
114 |
+
clean_up_tokenization_spaces=True,
|
115 |
+
)
|
116 |
+
|
117 |
+
var_output = first_var + ":" + var_output
|
118 |
+
if len(fields) > 0:
|
119 |
+
field_output = fields[0] + ":" + field_output
|
120 |
+
return var_output, field_output, varstring
|
121 |
|
122 |
|
123 |
demo = gr.Interface(
|
|
|
127 |
],
|
128 |
outputs=[
|
129 |
gr.Text(label="Var Decoder Output"),
|
130 |
+
gr.Text(label="Field Decoder Output"),
|
131 |
gr.Text(label="Generated Variable List"),
|
132 |
],
|
133 |
description=frontmatter.load("README.md").content,
|