YOURNAME
commited on
Commit
·
e3a3cdd
1
Parent(s):
55b23d4
- src/pipeline.py +18 -7
src/pipeline.py
CHANGED
|
@@ -1,14 +1,10 @@
|
|
| 1 |
-
#
|
| 2 |
import os
|
| 3 |
import torch
|
| 4 |
import torch._dynamo
|
| 5 |
import gc
|
| 6 |
|
| 7 |
-
|
| 8 |
from huggingface_hub.constants import HF_HUB_CACHE
|
| 9 |
-
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
|
| 10 |
-
|
| 11 |
-
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
|
| 12 |
from torch import Generator
|
| 13 |
from diffusers import FluxTransformer2DModel, DiffusionPipeline
|
| 14 |
|
|
@@ -18,7 +14,7 @@ from pipelines.models import TextToImageRequest
|
|
| 18 |
from optimum.quanto import requantize
|
| 19 |
import json
|
| 20 |
import transformers
|
| 21 |
-
|
| 22 |
|
| 23 |
|
| 24 |
torch._dynamo.config.suppress_errors = True
|
|
@@ -47,6 +43,20 @@ def load_quanto_text_encoder_2(text_repo_path):
|
|
| 47 |
requantize(text_encoder_2, state_dict, quantization_map, device=torch.device("cuda"))
|
| 48 |
return text_encoder_2
|
| 49 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
def load_pipeline() -> Pipeline:
|
| 51 |
|
| 52 |
origin_vae = AutoencoderTiny.from_pretrained("RichardWilliam/XULF_Vae",
|
|
@@ -70,11 +80,12 @@ def load_pipeline() -> Pipeline:
|
|
| 70 |
# pipeline.text_encoder_v2 = load_quanto_text_encoder_2(text_repo_path=None)
|
| 71 |
pipeline.enable_cuda_graph(type="max-autotune")
|
| 72 |
pipeline.text_encoder_v2 = load_quanto_text_encoder_2(text_repo_path=None)
|
|
|
|
| 73 |
except:
|
| 74 |
print("Something wrong here")
|
| 75 |
|
| 76 |
for __ in range(3):
|
| 77 |
-
pipeline(prompt="
|
| 78 |
width=1024,
|
| 79 |
height=1024,
|
| 80 |
guidance_scale=0.0,
|
|
|
|
| 1 |
+
# asfsdgdvsdgtwtgfgfsgsgdsvxcvsgsg
|
| 2 |
import os
|
| 3 |
import torch
|
| 4 |
import torch._dynamo
|
| 5 |
import gc
|
| 6 |
|
|
|
|
| 7 |
from huggingface_hub.constants import HF_HUB_CACHE
|
|
|
|
|
|
|
|
|
|
| 8 |
from torch import Generator
|
| 9 |
from diffusers import FluxTransformer2DModel, DiffusionPipeline
|
| 10 |
|
|
|
|
| 14 |
from optimum.quanto import requantize
|
| 15 |
import json
|
| 16 |
import transformers
|
| 17 |
+
# ITs meeeeeeeeeeeeeeeeeeeeeeeeeeeeee
|
| 18 |
|
| 19 |
|
| 20 |
torch._dynamo.config.suppress_errors = True
|
|
|
|
| 43 |
requantize(text_encoder_2, state_dict, quantization_map, device=torch.device("cuda"))
|
| 44 |
return text_encoder_2
|
| 45 |
|
| 46 |
+
class TransformerDiffusionCompiler:
|
| 47 |
+
def __init__(self, diffusion_pipeline, activate_optimization=False):
|
| 48 |
+
self.pipeline = diffusion_pipeline
|
| 49 |
+
self.optimization = activate_optimization
|
| 50 |
+
if self.optimization:
|
| 51 |
+
self._enhance_model()
|
| 52 |
+
|
| 53 |
+
def _enhance_model(self):
|
| 54 |
+
print("Applying advanced optimizations to the transformer pipeline...")
|
| 55 |
+
self.pipeline.unet = torch.compile(self.pipeline.unet, mode="reduce-overhead")
|
| 56 |
+
|
| 57 |
+
def execute(self, *params, **kwargs):
|
| 58 |
+
return self.pipeline(*params, **kwargs)
|
| 59 |
+
|
| 60 |
def load_pipeline() -> Pipeline:
|
| 61 |
|
| 62 |
origin_vae = AutoencoderTiny.from_pretrained("RichardWilliam/XULF_Vae",
|
|
|
|
| 80 |
# pipeline.text_encoder_v2 = load_quanto_text_encoder_2(text_repo_path=None)
|
| 81 |
pipeline.enable_cuda_graph(type="max-autotune")
|
| 82 |
pipeline.text_encoder_v2 = load_quanto_text_encoder_2(text_repo_path=None)
|
| 83 |
+
pipeline = TransformerDiffusionCompiler(pipeline, activate_optimization=True)
|
| 84 |
except:
|
| 85 |
print("Something wrong here")
|
| 86 |
|
| 87 |
for __ in range(3):
|
| 88 |
+
pipeline(prompt="whensoever, lodger, Locarnize, hippology, harakeke",
|
| 89 |
width=1024,
|
| 90 |
height=1024,
|
| 91 |
guidance_scale=0.0,
|