Update handler.py
Browse files- handler.py +67 -47
handler.py
CHANGED
@@ -1,15 +1,20 @@
|
|
1 |
import os
|
2 |
-
from typing import Any, Dict
|
3 |
from PIL import Image
|
4 |
import torch
|
5 |
from diffusers import FluxPipeline
|
6 |
from huggingface_inference_toolkit.logging import logger
|
7 |
from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
|
|
|
8 |
import time
|
9 |
-
import
|
10 |
-
|
11 |
-
|
12 |
-
|
|
|
|
|
|
|
|
|
13 |
|
14 |
class EndpointHandler:
|
15 |
def __init__(self, path=""):
|
@@ -17,60 +22,75 @@ class EndpointHandler:
|
|
17 |
"NoMoreCopyrightOrg/flux-dev",
|
18 |
torch_dtype=torch.bfloat16,
|
19 |
).to("cuda")
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
)
|
24 |
-
|
25 |
-
|
26 |
-
mesh=mesh,
|
27 |
-
)
|
28 |
-
parallelize_vae(self.pipe.vae, mesh=mesh._flatten())
|
29 |
apply_cache_on_pipe(self.pipe, residual_diff_threshold=0.12)
|
30 |
-
torch._inductor.config.reorder_for_compute_comm_overlap = True
|
31 |
self.pipe.transformer = torch.compile(
|
32 |
self.pipe.transformer, mode="max-autotune-no-cudagraphs",
|
33 |
)
|
34 |
self.pipe.vae = torch.compile(
|
35 |
self.pipe.vae, mode="max-autotune-no-cudagraphs",
|
36 |
)
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
-
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
prompt = data.pop("prompt")
|
45 |
-
else:
|
46 |
-
raise ValueError(
|
47 |
-
"Provided input body must contain either the key `inputs` or `prompt` with the"
|
48 |
-
" prompt to use for the image generation, and it needs to be a non-empty string."
|
49 |
-
)
|
50 |
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
|
|
57 |
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
if dist.get_rank() == 0:
|
73 |
time_taken = end_time - start_time
|
74 |
print(f"Time taken: {time_taken:.2f} seconds")
|
|
|
|
|
75 |
return result
|
76 |
-
|
|
|
|
|
|
1 |
import os
|
2 |
+
from typing import Any, Dict, Union
|
3 |
from PIL import Image
|
4 |
import torch
|
5 |
from diffusers import FluxPipeline
|
6 |
from huggingface_inference_toolkit.logging import logger
|
7 |
from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
|
8 |
+
from torchao.quantization import autoquant
|
9 |
import time
|
10 |
+
import gc
|
11 |
+
|
12 |
+
# Set high precision for float32 matrix multiplications.
|
13 |
+
# This setting optimizes performance on NVIDIA GPUs with Ampere architecture (e.g., A100, RTX 30 series) or newer.
|
14 |
+
torch.set_float32_matmul_precision("high")
|
15 |
+
|
16 |
+
import torch._dynamo
|
17 |
+
torch._dynamo.config.suppress_errors = False # for debugging
|
18 |
|
19 |
class EndpointHandler:
|
20 |
def __init__(self, path=""):
|
|
|
22 |
"NoMoreCopyrightOrg/flux-dev",
|
23 |
torch_dtype=torch.bfloat16,
|
24 |
).to("cuda")
|
25 |
+
self.pipe.enable_vae_slicing()
|
26 |
+
self.pipe.enable_vae_tiling()
|
27 |
+
self.pipe.transformer.fuse_qkv_projections()
|
28 |
+
self.pipe.vae.fuse_qkv_projections()
|
29 |
+
self.pipe.transformer.to(memory_format=torch.channels_last)
|
30 |
+
self.pipe.vae.to(memory_format=torch.channels_last)
|
|
|
|
|
|
|
31 |
apply_cache_on_pipe(self.pipe, residual_diff_threshold=0.12)
|
|
|
32 |
self.pipe.transformer = torch.compile(
|
33 |
self.pipe.transformer, mode="max-autotune-no-cudagraphs",
|
34 |
)
|
35 |
self.pipe.vae = torch.compile(
|
36 |
self.pipe.vae, mode="max-autotune-no-cudagraphs",
|
37 |
)
|
38 |
+
self.pipe.transformer = autoquant(self.pipe.transformer, error_on_unseen=False)
|
39 |
+
self.pipe.vae = autoquant(self.pipe.vae, error_on_unseen=False)
|
40 |
+
|
41 |
+
gc.collect()
|
42 |
+
torch.cuda.empty_cache()
|
43 |
|
44 |
+
start_time = time.time()
|
45 |
+
print("Start warming-up pipeline")
|
46 |
+
self.pipe("Hello world!") # Warm-up for compiling
|
47 |
+
end_time = time.time()
|
48 |
+
time_taken = end_time - start_time
|
49 |
+
print(f"Time taken: {time_taken:.2f} seconds")
|
50 |
+
self.record=0
|
51 |
|
52 |
+
def __call__(self, data: Dict[str, Any]) -> Union[Image.Image, None]:
|
53 |
+
try:
|
54 |
+
logger.info(f"Received incoming request with {data=}")
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
+
if "inputs" in data and isinstance(data["inputs"], str):
|
57 |
+
prompt = data.pop("inputs")
|
58 |
+
elif "prompt" in data and isinstance(data["prompt"], str):
|
59 |
+
prompt = data.pop("prompt")
|
60 |
+
else:
|
61 |
+
raise ValueError(
|
62 |
+
"Provided input body must contain either the key `inputs` or `prompt` with the"
|
63 |
+
" prompt to use for the image generation, and it needs to be a non-empty string."
|
64 |
+
)
|
65 |
+
if prompt=="get_queue":
|
66 |
+
return self.record
|
67 |
+
parameters = data.pop("parameters", {})
|
68 |
|
69 |
+
num_inference_steps = parameters.get("num_inference_steps", 28)
|
70 |
+
width = parameters.get("width", 1024)
|
71 |
+
height = parameters.get("height", 1024)
|
72 |
+
#guidance_scale = parameters.get("guidance_scale", 3.5)
|
73 |
+
guidance_scale = parameters.get("guidance", 3.5)
|
74 |
|
75 |
+
# seed generator (seed cannot be provided as is but via a generator)
|
76 |
+
seed = parameters.get("seed", 0)
|
77 |
+
generator = torch.manual_seed(seed)
|
78 |
+
self.record+=1
|
79 |
+
start_time = time.time()
|
80 |
+
result = self.pipe( # type: ignore
|
81 |
+
prompt,
|
82 |
+
height=height,
|
83 |
+
width=width,
|
84 |
+
guidance_scale=guidance_scale,
|
85 |
+
num_inference_steps=num_inference_steps,
|
86 |
+
generator=generator,
|
87 |
+
).images[0]
|
88 |
+
end_time = time.time()
|
|
|
89 |
time_taken = end_time - start_time
|
90 |
print(f"Time taken: {time_taken:.2f} seconds")
|
91 |
+
self.record-=1
|
92 |
+
|
93 |
return result
|
94 |
+
except Exception as e:
|
95 |
+
print(e)
|
96 |
+
return None
|