1
0
mirror of https://github.com/osmarks/nanogpt-experiments.git synced 2024-12-21 23:50:28 +00:00

add benchmarking script v0

This commit is contained in:
Andrej Karpathy 2022-12-28 23:55:43 +00:00
parent 5d2b4807bf
commit 70b5d93aee
2 changed files with 52 additions and 0 deletions

View File

@ -57,3 +57,7 @@ and observe the following losses on train and val:
| gpt2-xl | 1558M | 2.56 | 2.54 |
I briefly tried finetuning gpt2 a bit more on our OWT and didn't notice dramatic improvements, suggesting that OWT is not much much different from WT in terms of the data distribution, but this needs a bit more thorough attempt once the code is in a better place.
## benchmarking
For model benchmarking `bench.py` might be useful. It's identical what happens in `train.py` except we're measuring just the fwd+bwd+update time of the model on a fixed random batch of data.

48
bench.py Normal file
View File

@ -0,0 +1,48 @@
"""
A much shorter version of train.py for benchmarking the model
"""
import time
import torch
from model import GPTConfig, GPT
device = 'cuda:3'
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
torch.manual_seed(1337)
batch_size = 8
block_size = 1024
gptconf = GPTConfig(
block_size = block_size, # how far back does the model look? i.e. context size
n_layer = 12, n_head = 12, n_embd = 768, # size of the model
dropout = 0, # for determinism
)
model = GPT(gptconf)
model.to(device)
x = torch.randint(50257, (batch_size, block_size), device=device)
y = torch.randint(50257, (batch_size, block_size), device=device)
optimizer = model.configure_optimizers(weight_decay=1e-2, learning_rate=1e-4, betas=(0.9, 0.95))
burn_in = 10 # number of burn in steps where we don't measure time
num_steps = 30
for k in range(num_steps):
if k == burn_in:
t0 = time.time() # start the timer
with torch.autocast(device_type='cuda', dtype=torch.bfloat16):
logits, loss = model(x, y)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
lossf = loss.item()
print(f"{k}/{num_steps} loss: {lossf:.4f}")
torch.cuda.synchronize()
t1 = time.time()
print("time in ms per iteration: %.2f" % ((t1 - t0) / (num_steps - burn_in) * 1000))