Spaces:
Running
on
Zero
Running
on
Zero
Upload mod.py
Browse files
mod.py
CHANGED
@@ -13,7 +13,16 @@ from env import num_cns, model_trigger, HF_TOKEN, CIVITAI_API_KEY, DOWNLOAD_LORA
|
|
13 |
from modutils import download_things
|
14 |
|
15 |
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
|
18 |
#subprocess.run('pip cache purge', shell=True)
|
19 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
13 |
from modutils import download_things
|
14 |
|
15 |
|
16 |
+
IS_ZERO = True if os.getenv("SPACES_ZERO_GPU", None) else False
|
17 |
+
if torch.cuda.is_available():
|
18 |
+
torch.backends.cudnn.deterministic = True
|
19 |
+
torch.backends.cudnn.benchmark = False
|
20 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
21 |
+
if IS_ZERO:
|
22 |
+
subprocess.run("rm -rf /data-nvme/zerogpu-offload/*", env={}, shell=True)
|
23 |
+
torch.set_float32_matmul_precision("high") # https://pytorch.org/blog/accelerating-generative-ai-3/
|
24 |
+
|
25 |
+
|
26 |
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
|
27 |
#subprocess.run('pip cache purge', shell=True)
|
28 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|