patrickvonplaten
commited on
Commit
•
27dfa17
1
Parent(s):
096afd9
rename
Browse files- benchmark_llama.py +53 -0
- clear_mem.py +8 -6
- run_xl.py +4 -2
- save_lora.py +62 -0
benchmark_llama.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
3 |
+
import time
|
4 |
+
import torch
|
5 |
+
|
6 |
+
DEVICE = "cuda:1"
|
7 |
+
|
8 |
+
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
|
9 |
+
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16, low_cpu_mem_usage=True)
|
10 |
+
model.to(DEVICE)
|
11 |
+
|
12 |
+
|
13 |
+
# forward
|
14 |
+
print("Forward benchmarks")
|
15 |
+
print(50 * "=")
|
16 |
+
|
17 |
+
for batch_size in (1, 4, 16):
|
18 |
+
for input_seq in (4, 16, 256):
|
19 |
+
input_ids = torch.ones((batch_size, input_seq), dtype=torch.long, device=DEVICE)
|
20 |
+
attention_mask = torch.ones_like(input_ids)
|
21 |
+
attention_mask[0, 3] = 0
|
22 |
+
|
23 |
+
times = []
|
24 |
+
for _ in range(3):
|
25 |
+
start_time = time.time()
|
26 |
+
with torch.no_grad():
|
27 |
+
logits = model(input_ids=input_ids, attention_mask=attention_mask).logits
|
28 |
+
times.append(time.time() - start_time)
|
29 |
+
|
30 |
+
result = min(times)
|
31 |
+
|
32 |
+
print(f"Forward bsz={batch_size}, input_seq={input_seq}: {result}")
|
33 |
+
|
34 |
+
|
35 |
+
# generate
|
36 |
+
print("Generate benchmarks")
|
37 |
+
print(50 * "=")
|
38 |
+
|
39 |
+
for batch_size in (1, 16):
|
40 |
+
for input_seq in (4, 256):
|
41 |
+
input_ids = torch.ones((batch_size, input_seq), dtype=torch.long, device=DEVICE)
|
42 |
+
attention_mask = torch.ones_like(input_ids)
|
43 |
+
attention_mask[0, 3] = 0
|
44 |
+
|
45 |
+
times = []
|
46 |
+
for _ in range(3):
|
47 |
+
start_time = time.time()
|
48 |
+
out = model.generate(input_ids=input_ids, max_new_tokens=256)
|
49 |
+
times.append(time.time() - start_time)
|
50 |
+
|
51 |
+
result = min(times)
|
52 |
+
|
53 |
+
print(f"Generate bsz={batch_size}, input_seq={input_seq}: {result}")
|
clear_mem.py
CHANGED
@@ -1,10 +1,11 @@
|
|
1 |
#!/usr/bin/env python3
|
2 |
import torch
|
3 |
import gc
|
|
|
4 |
|
5 |
-
shape = (
|
6 |
|
7 |
-
input = torch.
|
8 |
|
9 |
|
10 |
def clear_memory(model):
|
@@ -14,8 +15,9 @@ def clear_memory(model):
|
|
14 |
torch.cuda.ipc_collect()
|
15 |
torch.clear_autocast_cache()
|
16 |
|
17 |
-
for
|
18 |
-
|
19 |
-
|
|
|
20 |
|
21 |
-
clear_memory(
|
|
|
1 |
#!/usr/bin/env python3
|
2 |
import torch
|
3 |
import gc
|
4 |
+
from diffusers import DiffusionPipeline
|
5 |
|
6 |
+
shape = (30_000, 30_000)
|
7 |
|
8 |
+
input = torch.randn(shape, device="cuda")
|
9 |
|
10 |
|
11 |
def clear_memory(model):
|
|
|
15 |
torch.cuda.ipc_collect()
|
16 |
torch.clear_autocast_cache()
|
17 |
|
18 |
+
for _ids in ["runwayml/stable-diffusion-v1-5", "CompVis/stable-diffusion-v1-4", "runwayml/stable-diffusion-v1-5", "CompVis/stable-diffusion-v1-4", "runwayml/stable-diffusion-v1-5"]:
|
19 |
+
pipe = DiffusionPipeline.from_pretrained(_ids, use_safetensors=True).to("cuda")
|
20 |
+
pipe("hey", num_inference_steps=1)
|
21 |
+
print("finished...")
|
22 |
|
23 |
+
clear_memory(pipe)
|
run_xl.py
CHANGED
@@ -14,8 +14,10 @@ pipe2 = DiffusionPipeline.from_pretrained(
|
|
14 |
variant="fp16",
|
15 |
torch_dtype=torch.float16
|
16 |
)
|
17 |
-
pipe.
|
18 |
-
pipe2.
|
|
|
|
|
19 |
|
20 |
compel = Compel(
|
21 |
tokenizer=[pipe.tokenizer, pipe.tokenizer_2] ,
|
|
|
14 |
variant="fp16",
|
15 |
torch_dtype=torch.float16
|
16 |
)
|
17 |
+
pipe.to("cuda")
|
18 |
+
pipe2.to("cuda")
|
19 |
+
# pipe.enable_model_cpu_offload()
|
20 |
+
# pipe2.enable_model_cpu_offload()
|
21 |
|
22 |
compel = Compel(
|
23 |
tokenizer=[pipe.tokenizer, pipe.tokenizer_2] ,
|
save_lora.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
import torch
|
3 |
+
from warnings import warn
|
4 |
+
from diffusers import (
|
5 |
+
AutoencoderKL,
|
6 |
+
DiffusionPipeline,
|
7 |
+
)
|
8 |
+
import hashlib
|
9 |
+
|
10 |
+
base = "stabilityai/stable-diffusion-xl-base-1.0"
|
11 |
+
adapter1 = 'nerijs/pixel-art-xl'
|
12 |
+
weightname1 = 'pixel-art-xl.safetensors'
|
13 |
+
|
14 |
+
adapter2 = 'Alexzyx/lora-trained-xl-colab'
|
15 |
+
weightname2 = None
|
16 |
+
|
17 |
+
inputs = "elephant"
|
18 |
+
kwargs = {}
|
19 |
+
|
20 |
+
if torch.cuda.is_available():
|
21 |
+
kwargs["torch_dtype"] = torch.float16
|
22 |
+
|
23 |
+
#vae = AutoencoderKL.from_pretrained(
|
24 |
+
# "madebyollin/sdxl-vae-fp16-fix",
|
25 |
+
# torch_dtype=torch.float16, # load fp16 fix VAE
|
26 |
+
#)
|
27 |
+
#kwargs["vae"] = vae
|
28 |
+
#kwargs["variant"] = "fp16"
|
29 |
+
#
|
30 |
+
|
31 |
+
model = DiffusionPipeline.from_pretrained(
|
32 |
+
base, **kwargs
|
33 |
+
)
|
34 |
+
|
35 |
+
if torch.cuda.is_available():
|
36 |
+
model.to("cuda")
|
37 |
+
|
38 |
+
|
39 |
+
def inference(adapter, weightname):
|
40 |
+
model.load_lora_weights(adapter, weight_name=weightname)
|
41 |
+
try:
|
42 |
+
model.fuse_lora(safe_fusing=True)
|
43 |
+
except ValueError:
|
44 |
+
warn(f"{adapter} and {weightname} is broken. LoRA is not fused.")
|
45 |
+
model.unload_lora_weights()
|
46 |
+
|
47 |
+
data = model(inputs, num_inference_steps=1).images[0]
|
48 |
+
model.unfuse_lora()
|
49 |
+
model.unload_lora_weights()
|
50 |
+
filename = '/tmp/hello.jpg'
|
51 |
+
data.save(filename, format='jpeg')
|
52 |
+
with open(filename, 'rb') as f:
|
53 |
+
md5 = hashlib.md5(f.read()).hexdigest()
|
54 |
+
print("Adapter %s, md5sum %s" % (adapter, md5))
|
55 |
+
if md5 == '40c78c9fd4daeff01c988c3532fdd51b':
|
56 |
+
print("BLACK SCREEN IMAGE for adapter %s" % adapter)
|
57 |
+
|
58 |
+
|
59 |
+
inference(adapter1, weightname1)
|
60 |
+
inference(adapter2, weightname2)
|
61 |
+
inference(adapter1, weightname1)
|
62 |
+
inference(adapter1, weightname1)
|