Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -42,18 +42,25 @@ MODEL_CONFIG = {
|
|
42 |
# -------------------------------
|
43 |
|
44 |
class ModelManager:
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
def get_llama_pipeline(self, model_id, token):
|
46 |
if model_id not in self.llama_pipelines:
|
47 |
tokenizer = AutoTokenizer.from_pretrained(
|
48 |
model_id,
|
49 |
token=token,
|
50 |
-
legacy=False
|
51 |
)
|
52 |
model = AutoModelForCausalLM.from_pretrained(
|
53 |
model_id,
|
54 |
token=token,
|
55 |
torch_dtype=torch.float16,
|
56 |
-
device_map="auto"
|
|
|
57 |
)
|
58 |
self.llama_pipelines[model_id] = pipeline(
|
59 |
"text-generation",
|
@@ -62,19 +69,15 @@ class ModelManager:
|
|
62 |
device_map="auto"
|
63 |
)
|
64 |
return self.llama_pipelines[model_id]
|
65 |
-
|
66 |
def get_musicgen_model(self):
|
67 |
if not self.musicgen_model:
|
68 |
self.musicgen_model = MusicgenForConditionalGeneration.from_pretrained(
|
69 |
MODEL_CONFIG["musicgen_model"]
|
70 |
)
|
|
|
71 |
self.musicgen_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
72 |
-
return self.musicgen_model
|
73 |
-
|
74 |
-
def get_tts_model(self, model_name):
|
75 |
-
if model_name not in self.tts_models:
|
76 |
-
self.tts_models[model_name] = TTS(model_name)
|
77 |
-
return self.tts_models[model_name]
|
78 |
|
79 |
model_manager = ModelManager()
|
80 |
|
|
|
42 |
# -------------------------------
|
43 |
|
44 |
class ModelManager:
|
45 |
+
def __init__(self):
|
46 |
+
self.llama_pipelines = {}
|
47 |
+
self.musicgen_model = None
|
48 |
+
self.tts_models = {}
|
49 |
+
self.processor = None # Add processor cache
|
50 |
+
|
51 |
def get_llama_pipeline(self, model_id, token):
|
52 |
if model_id not in self.llama_pipelines:
|
53 |
tokenizer = AutoTokenizer.from_pretrained(
|
54 |
model_id,
|
55 |
token=token,
|
56 |
+
legacy=False
|
57 |
)
|
58 |
model = AutoModelForCausalLM.from_pretrained(
|
59 |
model_id,
|
60 |
token=token,
|
61 |
torch_dtype=torch.float16,
|
62 |
+
device_map="auto",
|
63 |
+
low_cpu_mem_usage=True
|
64 |
)
|
65 |
self.llama_pipelines[model_id] = pipeline(
|
66 |
"text-generation",
|
|
|
69 |
device_map="auto"
|
70 |
)
|
71 |
return self.llama_pipelines[model_id]
|
72 |
+
|
73 |
def get_musicgen_model(self):
|
74 |
if not self.musicgen_model:
|
75 |
self.musicgen_model = MusicgenForConditionalGeneration.from_pretrained(
|
76 |
MODEL_CONFIG["musicgen_model"]
|
77 |
)
|
78 |
+
self.processor = AutoProcessor.from_pretrained(MODEL_CONFIG["musicgen_model"])
|
79 |
self.musicgen_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
80 |
+
return self.musicgen_model, self.processor
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
model_manager = ModelManager()
|
83 |
|