1
0
mirror of https://github.com/osmarks/nanogpt-experiments.git synced 2024-12-18 14:10:28 +00:00
nanogpt-experiments/bench.py

112 lines
4.3 KiB
Python
Raw Normal View History

2022-12-28 23:55:43 +00:00
"""
A much shorter version of train.py for benchmarking
2022-12-28 23:55:43 +00:00
"""
import os
from contextlib import nullcontext
import numpy as np
2022-12-28 23:55:43 +00:00
import time
import torch
from model import GPTConfig, GPT
# -----------------------------------------------------------------------------
2022-12-28 23:55:43 +00:00
batch_size = 8
block_size = 1024
seed = 1337
device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.
dtype = 'bfloat16' # 'float32' or 'bfloat16' or 'float16'
compile = True # use PyTorch 2.0 to compile the model to be faster
exec(open('configurator.py').read()) # overrides from command line or config file
# -----------------------------------------------------------------------------
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda' if 'cuda' in device else 'cpu' # for later use in torch.autocast
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
2022-12-28 23:55:43 +00:00
# data loading init
real_data = True
if real_data:
dataset = 'openwebtext'
data_dir = os.path.join('data', dataset)
train_data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
def get_batch(split):
data = train_data # note ignore split in benchmarking script
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([torch.from_numpy((data[i:i+block_size]).astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy((data[i+1:i+1+block_size]).astype(np.int64)) for i in ix])
x, y = x.to(device), y.to(device)
return x, y
else:
# alternatively, if fixed data is desired to not care about data loading
x = torch.randint(50257, (batch_size, block_size), device=device)
y = torch.randint(50257, (batch_size, block_size), device=device)
get_batch = lambda split: (x, y)
# model init
2022-12-28 23:55:43 +00:00
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)
optimizer = model.configure_optimizers(weight_decay=1e-2, learning_rate=1e-4, betas=(0.9, 0.95))
if compile:
print("Compiling model...")
model = torch.compile(model) # pytorch 2.0
profile = False # use pytorch profiler, or just simple benchmarking?
if profile:
# useful docs on pytorch profiler:
# - tutorial https://pytorch.org/tutorials/intermediate/tensorboard_profiler_tutorial.html
# - api https://pytorch.org/docs/stable/profiler.html#torch.profiler.profile
wait, warmup, active = 5, 5, 5
num_steps = wait + warmup + active
with torch.profiler.profile(
activities=[torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA],
schedule=torch.profiler.schedule(wait=wait, warmup=warmup, active=active, repeat=1),
on_trace_ready=torch.profiler.tensorboard_trace_handler('./bench_log'),
record_shapes=True,
profile_memory=True,
with_stack=True, # incurs an additional overhead, disable if not needed
with_flops=True,
with_modules=False, # only for torchscript models atm
) as prof:
2022-12-28 23:55:43 +00:00
for k in range(num_steps):
X, Y = get_batch('train')
with ctx:
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}")
2022-12-28 23:55:43 +00:00
prof.step() # notify the profiler at end of each step
2022-12-28 23:55:43 +00:00
else:
2022-12-28 23:55:43 +00:00
# simple benchmarking
torch.cuda.synchronize()
for stage, num_steps in enumerate([10, 20]): # burnin, then benchmark
t0 = time.time()
for k in range(num_steps):
X, Y = get_batch('train')
with ctx:
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()
if stage == 1:
print(f"time per iteration: {(t1-t0)/num_steps*1000:.4f}ms")