From 70b5d93aeece4e29fc2b8d3426ac63e2b2f22c1a Mon Sep 17 00:00:00 2001
From: Andrej Karpathy <andrej.karpathy@gmail.com>
Date: Wed, 28 Dec 2022 23:55:43 +0000
Subject: [PATCH] add benchmarking script v0

---
 README.md |  4 ++++
 bench.py  | 48 ++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+)
 create mode 100644 bench.py

diff --git a/README.md b/README.md
index b017587..31c13b5 100644
--- a/README.md
+++ b/README.md
@@ -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.
diff --git a/bench.py b/bench.py
new file mode 100644
index 0000000..8356d21
--- /dev/null
+++ b/bench.py
@@ -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))