Spaces:
Paused
Paused
up
Browse files- app-fast.py +11 -12
app-fast.py
CHANGED
@@ -39,31 +39,30 @@ RESOLUTION_OPTIONS: list[str] = [
|
|
39 |
"832 x 1248 (Portrait)",
|
40 |
]
|
41 |
|
|
|
42 |
|
43 |
quant_config = Int4WeightOnlyConfig(group_size=128)
|
44 |
-
quantization_config = TransformersTorchAoConfig(
|
45 |
-
quant_type=quant_config, dtype=torch.bfloat16
|
46 |
-
)
|
47 |
|
48 |
tokenizer = AutoTokenizer.from_pretrained(LLAMA_MODEL_NAME, use_fast=False)
|
49 |
text_encoder = AutoModelForCausalLM.from_pretrained(
|
50 |
LLAMA_MODEL_NAME,
|
51 |
-
torch_dtype="auto",
|
52 |
-
low_cpu_mem_usage=True,
|
53 |
-
device_map="auto",
|
54 |
output_hidden_states=True,
|
55 |
output_attentions=True,
|
|
|
56 |
quantization_config=quantization_config,
|
57 |
-
|
|
|
|
|
58 |
|
59 |
quantization_config = DiffusersTorchAoConfig("int8wo")
|
60 |
transformer = HiDreamImageTransformer2DModel.from_pretrained(
|
61 |
MODEL_PATH,
|
62 |
subfolder="transformer",
|
63 |
-
device_map="auto",
|
64 |
quantization_config=quantization_config,
|
65 |
-
|
66 |
-
|
|
|
67 |
|
68 |
scheduler = MODEL_CONFIGS["scheduler"](
|
69 |
num_train_timesteps=1000,
|
@@ -76,8 +75,8 @@ pipe = HiDreamImagePipeline.from_pretrained(
|
|
76 |
scheduler=scheduler,
|
77 |
tokenizer_4=tokenizer,
|
78 |
text_encoder_4=text_encoder,
|
79 |
-
torch_dtype=
|
80 |
-
)
|
81 |
|
82 |
pipe.transformer = transformer
|
83 |
|
|
|
39 |
"832 x 1248 (Portrait)",
|
40 |
]
|
41 |
|
42 |
+
device = torch.device("cuda")
|
43 |
|
44 |
quant_config = Int4WeightOnlyConfig(group_size=128)
|
45 |
+
quantization_config = TransformersTorchAoConfig(quant_type=quant_config)
|
|
|
|
|
46 |
|
47 |
tokenizer = AutoTokenizer.from_pretrained(LLAMA_MODEL_NAME, use_fast=False)
|
48 |
text_encoder = AutoModelForCausalLM.from_pretrained(
|
49 |
LLAMA_MODEL_NAME,
|
|
|
|
|
|
|
50 |
output_hidden_states=True,
|
51 |
output_attentions=True,
|
52 |
+
low_cpu_mem_usage=True,
|
53 |
quantization_config=quantization_config,
|
54 |
+
torch_dtype=torch.bfloat16, # Explicitly set dtype
|
55 |
+
device_map="auto", # Still use auto, but ensure device consistency
|
56 |
+
).to(device) # Move model to the correct device after loading
|
57 |
|
58 |
quantization_config = DiffusersTorchAoConfig("int8wo")
|
59 |
transformer = HiDreamImageTransformer2DModel.from_pretrained(
|
60 |
MODEL_PATH,
|
61 |
subfolder="transformer",
|
|
|
62 |
quantization_config=quantization_config,
|
63 |
+
device_map="auto",
|
64 |
+
torch_dtype=torch.bfloat16,
|
65 |
+
).to(device)
|
66 |
|
67 |
scheduler = MODEL_CONFIGS["scheduler"](
|
68 |
num_train_timesteps=1000,
|
|
|
75 |
scheduler=scheduler,
|
76 |
tokenizer_4=tokenizer,
|
77 |
text_encoder_4=text_encoder,
|
78 |
+
torch_dtype=torch.bfloat16,
|
79 |
+
).to(device)
|
80 |
|
81 |
pipe.transformer = transformer
|
82 |
|