Spaces:
Sleeping
Sleeping
Commit
·
ffec641
1
Parent(s):
220b4dd
feat add model on zeroGpu
Browse files- prediction.py +19 -17
- requirements.txt +1 -1
- test_prediction.py +4 -2
prediction.py
CHANGED
@@ -21,6 +21,8 @@ from transformers import pipeline as hf_pipeline
|
|
21 |
import torch
|
22 |
import litellm
|
23 |
|
|
|
|
|
24 |
|
25 |
class ModelPrediction:
|
26 |
def __init__(self):
|
@@ -32,6 +34,7 @@ class ModelPrediction:
|
|
32 |
"DeepSeek-R1-Distill-Llama-70B": self._model_prediction(
|
33 |
"DeepSeek-R1-Distill-Llama-70B"
|
34 |
),
|
|
|
35 |
}
|
36 |
|
37 |
self._model_name = None
|
@@ -50,6 +53,7 @@ class ModelPrediction:
|
|
50 |
|
51 |
def _reset_pipeline(self, model_name):
|
52 |
if self._model_name != model_name:
|
|
|
53 |
self._model_name = model_name
|
54 |
self._pipeline = None
|
55 |
|
@@ -63,6 +67,13 @@ class ModelPrediction:
|
|
63 |
matches = re.findall(r"```sql(.*?)```", pred, re.DOTALL)
|
64 |
return matches[-1].strip() if matches else pred
|
65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
def make_prediction(self, prompt, model_name):
|
67 |
if model_name not in self.model_name2pred_func:
|
68 |
raise ValueError(
|
@@ -89,34 +100,25 @@ class ModelPrediction:
|
|
89 |
model_name = "together_ai/Qwen/QwQ-32B"
|
90 |
elif "DeepSeek-R1-Distill-Llama-70B" in model_name:
|
91 |
model_name = "together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
|
|
|
|
|
|
|
92 |
else:
|
93 |
raise ValueError("Model forbidden")
|
94 |
|
95 |
return partial(predict_fun, model_name=model_name)
|
96 |
|
97 |
def predict_with_api(self, prompt, model_name): # -> dict[str, Any | float]:
|
98 |
-
def track_cost_callback(
|
99 |
-
kwargs, # kwargs to completion
|
100 |
-
completion_response, # response from completion
|
101 |
-
start_time,
|
102 |
-
end_time, # start/end time
|
103 |
-
):
|
104 |
-
try:
|
105 |
-
response_cost = kwargs[
|
106 |
-
"response_cost"
|
107 |
-
] # litellm calculates response cost for you
|
108 |
-
call_cost = response_cost
|
109 |
-
except:
|
110 |
-
pass
|
111 |
-
|
112 |
-
litellm.success_callback = [track_cost_callback]
|
113 |
-
call_cost = 0.0
|
114 |
response = litellm.completion(
|
115 |
model=model_name,
|
116 |
messages=[{"role": "user", "content": prompt}],
|
117 |
num_retries=2,
|
118 |
)
|
119 |
-
|
|
|
|
|
|
|
|
|
120 |
|
121 |
@spaces.GPU
|
122 |
def predict_with_hf(self, prompt, model_name): # -> dict[str, Any | float]:
|
|
|
21 |
import torch
|
22 |
import litellm
|
23 |
|
24 |
+
from tqdm import tqdm
|
25 |
+
|
26 |
|
27 |
class ModelPrediction:
|
28 |
def __init__(self):
|
|
|
34 |
"DeepSeek-R1-Distill-Llama-70B": self._model_prediction(
|
35 |
"DeepSeek-R1-Distill-Llama-70B"
|
36 |
),
|
37 |
+
"llama-8": self._model_prediction("llama-8"),
|
38 |
}
|
39 |
|
40 |
self._model_name = None
|
|
|
53 |
|
54 |
def _reset_pipeline(self, model_name):
|
55 |
if self._model_name != model_name:
|
56 |
+
print("Resetting pipeline with model", model_name)
|
57 |
self._model_name = model_name
|
58 |
self._pipeline = None
|
59 |
|
|
|
67 |
matches = re.findall(r"```sql(.*?)```", pred, re.DOTALL)
|
68 |
return matches[-1].strip() if matches else pred
|
69 |
|
70 |
+
def make_predictions(self, prompts, model_name) -> list[dict]:
|
71 |
+
preds = []
|
72 |
+
for prompt in tqdm(prompts, desc=f"Analyzing Prompt with {model_name}"):
|
73 |
+
pred = self.make_prediction(prompt, model_name)
|
74 |
+
preds.append(pred)
|
75 |
+
return preds
|
76 |
+
|
77 |
def make_prediction(self, prompt, model_name):
|
78 |
if model_name not in self.model_name2pred_func:
|
79 |
raise ValueError(
|
|
|
100 |
model_name = "together_ai/Qwen/QwQ-32B"
|
101 |
elif "DeepSeek-R1-Distill-Llama-70B" in model_name:
|
102 |
model_name = "together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
|
103 |
+
elif "llama-8" in model_name:
|
104 |
+
model_name = "meta-llama/Meta-Llama-3-8B-Instruct"
|
105 |
+
predict_fun = self.predict_with_hf
|
106 |
else:
|
107 |
raise ValueError("Model forbidden")
|
108 |
|
109 |
return partial(predict_fun, model_name=model_name)
|
110 |
|
111 |
def predict_with_api(self, prompt, model_name): # -> dict[str, Any | float]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
response = litellm.completion(
|
113 |
model=model_name,
|
114 |
messages=[{"role": "user", "content": prompt}],
|
115 |
num_retries=2,
|
116 |
)
|
117 |
+
response_text = response["choices"][0]["message"]["content"]
|
118 |
+
return {
|
119 |
+
"response": response_text,
|
120 |
+
"cost": response._hidden_params["response_cost"],
|
121 |
+
}
|
122 |
|
123 |
@spaces.GPU
|
124 |
def predict_with_hf(self, prompt, model_name): # -> dict[str, Any | float]:
|
requirements.txt
CHANGED
@@ -10,9 +10,9 @@ eval-type-backport>=0.2.0
|
|
10 |
openai==1.66.3
|
11 |
litellm==1.63.14
|
12 |
together==1.4.6
|
13 |
-
litellm==1.63.14
|
14 |
# Conditional dependency for Gradio (requires Python >=3.10)
|
15 |
gradio>=5.20.1; python_version >= "3.10"
|
|
|
16 |
|
17 |
# Test dependencies
|
18 |
streamlit>=1.43.0
|
|
|
10 |
openai==1.66.3
|
11 |
litellm==1.63.14
|
12 |
together==1.4.6
|
|
|
13 |
# Conditional dependency for Gradio (requires Python >=3.10)
|
14 |
gradio>=5.20.1; python_version >= "3.10"
|
15 |
+
accelerate>=0.26.0
|
16 |
|
17 |
# Test dependencies
|
18 |
streamlit>=1.43.0
|
test_prediction.py
CHANGED
@@ -3,8 +3,10 @@ from prediction import ModelPrediction
|
|
3 |
|
4 |
def main():
|
5 |
model = ModelPrediction()
|
6 |
-
response = model.make_prediction("Hi, how are you?", "
|
7 |
-
print(response)
|
|
|
8 |
|
9 |
if __name__ == "__main__":
|
10 |
main()
|
|
|
|
3 |
|
4 |
def main():
|
5 |
model = ModelPrediction()
|
6 |
+
response = model.make_prediction("Hi, how are you?", "llama-8")
|
7 |
+
print(response) # dict[response, response_parsed, cost]
|
8 |
+
|
9 |
|
10 |
if __name__ == "__main__":
|
11 |
main()
|
12 |
+
# do something with prompt
|