Upload 11 files

.gitattributes

@@ -1,35 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Override jupyter in Github language stats for more accurate estimate of repo code languages
+# reference: https://github.com/github/linguist/blob/master/docs/overrides.md#generated-code
+*.ipynb linguist-generated

.gitignore

@@ -0,0 +1,4 @@
+.DS_Store
+.ipynb_checkpoints/
+__pycache__/
+*.pyc

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Andrej Karpathy
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,3 +1,223 @@
----
-license: cc-by-nc-sa-4.0
----
+
+# nanoGPT
+
+![nanoGPT](assets/nanogpt.jpg)
+
+The simplest, fastest repository for training/finetuning medium-sized GPTs. It is a rewrite of [minGPT](https://github.com/karpathy/minGPT) that prioritizes teeth over education. Still under active development, but currently the file `train.py` reproduces GPT-2 (124M) on OpenWebText, running on a single 8XA100 40GB node in about 4 days of training. The code itself is plain and readable: `train.py` is a ~300-line boilerplate training loop and `model.py` a ~300-line GPT model definition, which can optionally load the GPT-2 weights from OpenAI. That's it.
+
+![repro124m](assets/gpt2_124M_loss.png)
+
+Because the code is so simple, it is very easy to hack to your needs, train new models from scratch, or finetune pretrained checkpoints (e.g. biggest one currently available as a starting point would be the GPT-2 1.3B model from OpenAI).
+
+## install
+
+Dependencies:
+
+- [pytorch](https://pytorch.org) <3
+- [numpy](https://numpy.org/install/) <3
+- `pip install transformers` for huggingface transformers <3 (to load GPT-2 checkpoints)
+- `pip install datasets` for huggingface datasets <3 (if you want to download + preprocess OpenWebText)
+- `pip install tiktoken` for OpenAI's fast BPE code <3
+- `pip install wandb` for optional logging <3
+- `pip install tqdm` <3
+
+## quick start
+
+If you are not a deep learning professional and you just want to feel the magic and get your feet wet, the fastest way to get started is to train a character-level GPT on the works of Shakespeare. First, we download it as a single (1MB) file and turn it from raw text into one large stream of integers:
+
+```
+$ python data/shakespeare_char/prepare.py
+```
+
+This creates a `train.bin` and `val.bin` in that data directory. Now it is time to train your GPT. The size of it very much depends on the computational resources of your system:
+
+**I have a GPU**. Great, we can quickly train a baby GPT with the settings provided in the [config/train_shakespeare_char.py](config/train_shakespeare_char.py) config file:
+
+```
+$ python train.py config/train_shakespeare_char.py
+```
+
+If you peek inside it, you'll see that we're training a GPT with a context size of up to 256 characters, 384 feature channels, and it is a 6-layer Transformer with 6 heads in each layer. On one A100 GPU this training run takes about 3 minutes and the best validation loss is 1.4697. Based on the configuration, the model checkpoints are being written into the `--out_dir` directory `out-shakespeare-char`. So once the training finishes we can sample from the best model by pointing the sampling script at this directory:
+
+```
+$ python sample.py --out_dir=out-shakespeare-char
+```
+
+This generates a few samples, for example:
+
+```
+ANGELO:
+And cowards it be strawn to my bed,
+And thrust the gates of my threats,
+Because he that ale away, and hang'd
+An one with him.
+
+DUKE VINCENTIO:
+I thank your eyes against it.
+
+DUKE VINCENTIO:
+Then will answer him to save the malm:
+And what have you tyrannous shall do this?
+
+DUKE VINCENTIO:
+If you have done evils of all disposition
+To end his power, the day of thrust for a common men
+That I leave, to fight with over-liking
+Hasting in a roseman.
+```
+
+lol  `¯\_(ツ)_/¯`. Not bad for a character-level model after 3 minutes of training on a GPU. Better results are quite likely obtainable by instead finetuning a pretrained GPT-2 model on this dataset (see finetuning section later).
+
+**I only have a macbook** (or other cheap computer). No worries, we can still train a GPT but we want to dial things down a notch. I recommend getting the bleeding edge PyTorch nightly ([select it here](https://pytorch.org/get-started/locally/) when installing) as it is currently quite likely to make your code more efficient. But even without it, a simple train run could look as follows:
+
+```
+$ python train.py config/train_shakespeare_char.py --device=cpu --compile=False --eval_iters=20 --log_interval=1 --block_size=64 --batch_size=12 --n_layer=4 --n_head=4 --n_embd=128 --max_iters=2000 --lr_decay_iters=2000 --dropout=0.0
+```
+
+Here, since we are running on CPU instead of GPU we must set both `--device=cpu` and also turn off PyTorch 2.0 compile with `--compile=False`. Then when we evaluate we get a bit more noisy but faster estimate (`--eval_iters=20`, down from 200), our context size is only 64 characters instead of 256, and the batch size only 12 examples per iteration, not 64. We'll also use a much smaller Transformer (4 layers, 4 heads, 128 embedding size), and decrease the number of iterations to 2000 (and correspondingly usually decay the learning rate to around max_iters with `--lr_decay_iters`). Because our network is so small we also ease down on regularization (`--dropout=0.0`). This still runs in about ~3 minutes, but gets us a loss of only 1.88 and therefore also worse samples, but it's still good fun:
+
+```
+$ python sample.py --out_dir=out-shakespeare-char --device=cpu
+```
+Generates samples like this:
+
+```
+GLEORKEN VINGHARD III:
+Whell's the couse, the came light gacks,
+And the for mought you in Aut fries the not high shee
+bot thou the sought bechive in that to doth groan you,
+No relving thee post mose the wear
+```
+
+Not bad for ~3 minutes on a CPU, for a hint of the right character gestalt. If you're willing to wait longer, feel free to tune the hyperparameters, increase the size of the network, the context length (`--block_size`), the length of training, etc.
+
+Finally, on Apple Silicon Macbooks and with a recent PyTorch version make sure to add `--device=mps` (short for "Metal Performance Shaders"); PyTorch then uses the on-chip GPU that can *significantly* accelerate training (2-3X) and allow you to use larger networks. See [Issue 28](https://github.com/karpathy/nanoGPT/issues/28) for more.
+
+## reproducing GPT-2
+
+A more serious deep learning professional may be more interested in reproducing GPT-2 results. So here we go - we first tokenize the dataset, in this case the [OpenWebText](https://openwebtext2.readthedocs.io/en/latest/), an open reproduction of OpenAI's (private) WebText:
+
+```
+$ python data/openwebtext/prepare.py
+```
+
+This downloads and tokenizes the [OpenWebText](https://huggingface.co/datasets/openwebtext) dataset. It will create a `train.bin` and `val.bin` which holds the GPT2 BPE token ids in one sequence, stored as raw uint16 bytes. Then we're ready to kick off training. To reproduce GPT-2 (124M) you'll want at least an 8X A100 40GB node and run:
+
+```
+$ torchrun --standalone --nproc_per_node=8 train.py config/train_gpt2.py
+```
+
+This will run for about 4 days using PyTorch Distributed Data Parallel (DDP) and go down to loss of ~2.85. Now, a GPT-2 model just evaluated on OWT gets a val loss of about 3.11, but if you finetune it it will come down to ~2.85 territory (due to an apparent domain gap), making the two models ~match.
+
+If you're in a cluster environment and you are blessed with multiple GPU nodes you can make GPU go brrrr e.g. across 2 nodes like:
+
+```
+Run on the first (master) node with example IP 123.456.123.456:
+$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr=123.456.123.456 --master_port=1234 train.py
+Run on the worker node:
+$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr=123.456.123.456 --master_port=1234 train.py
+```
+
+It is a good idea to benchmark your interconnect (e.g. iperf3). In particular, if you don't have Infiniband then also prepend `NCCL_IB_DISABLE=1` to the above launches. Your multinode training will work, but most likely _crawl_. By default checkpoints are periodically written to the `--out_dir`. We can sample from the model by simply `$ python sample.py`.
+
+Finally, to train on a single GPU simply run the `$ python train.py` script. Have a look at all of its args, the script tries to be very readable, hackable and transparent. You'll most likely want to tune a number of those variables depending on your needs.
+
+## baselines
+
+OpenAI GPT-2 checkpoints allow us to get some baselines in place for openwebtext. We can get the numbers as follows:
+
+```
+$ python train.py eval_gpt2
+$ python train.py eval_gpt2_medium
+$ python train.py eval_gpt2_large
+$ python train.py eval_gpt2_xl
+```
+
+and observe the following losses on train and val:
+
+| model | params | train loss | val loss |
+| ------| ------ | ---------- | -------- |
+| gpt2 | 124M         | 3.11  | 3.12     |
+| gpt2-medium | 350M  | 2.85  | 2.84     |
+| gpt2-large | 774M   | 2.66  | 2.67     |
+| gpt2-xl | 1558M     | 2.56  | 2.54     |
+
+However, we have to note that GPT-2 was trained on (closed, never released) WebText, while OpenWebText is just a best-effort open reproduction of this dataset. This means there is a dataset domain gap. Indeed, taking the GPT-2 (124M) checkpoint and finetuning on OWT directly for a while reaches loss down to ~2.85. This then becomes the more appropriate baseline w.r.t. reproduction.
+
+## finetuning
+
+Finetuning is no different than training, we just make sure to initialize from a pretrained model and train with a smaller learning rate. For an example of how to finetune a GPT on new text go to `data/shakespeare` and run `prepare.py` to download the tiny shakespeare dataset and render it into a `train.bin` and `val.bin`, using the OpenAI BPE tokenizer from GPT-2. Unlike OpenWebText this will run in seconds. Finetuning can take very little time, e.g. on a single GPU just a few minutes. Run an example finetuning like:
+
+```
+$ python train.py config/finetune_shakespeare.py
+```
+
+This will load the config parameter overrides in `config/finetune_shakespeare.py` (I didn't tune them much though). Basically, we initialize from a GPT2 checkpoint with `init_from` and train as normal, except shorter and with a small learning rate. If you're running out of memory try decreasing the model size (they are `{'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}`) or possibly decreasing the `block_size` (context length). The best checkpoint (lowest validation loss) will be in the `out_dir` directory, e.g. in `out-shakespeare` by default, per the config file. You can then run the code in `sample.py --out_dir=out-shakespeare`:
+
+```
+THEODORE:
+Thou shalt sell me to the highest bidder: if I die,
+I sell thee to the first; if I go mad,
+I sell thee to the second; if I
+lie, I sell thee to the third; if I slay,
+I sell thee to the fourth: so buy or sell,
+I tell thee again, thou shalt not sell my
+possession.
+
+JULIET:
+And if thou steal, thou shalt not sell thyself.
+
+THEODORE:
+I do not steal; I sell the stolen goods.
+
+THEODORE:
+Thou know'st not what thou sell'st; thou, a woman,
+Thou art ever a victim, a thing of no worth:
+Thou hast no right, no right, but to be sold.
+```
+
+Whoa there, GPT, entering some dark place over there. I didn't really tune the hyperparameters in the config too much, feel free to try!
+
+## sampling / inference
+
+Use the script `sample.py` to sample either from pre-trained GPT-2 models released by OpenAI, or from a model you trained yourself. For example, here is a way to sample from the largest available `gpt2-xl` model:
+
+```
+$ python sample.py \
+    --init_from=gpt2-xl \
+    --start="What is the answer to life, the universe, and everything?" \
+    --num_samples=5 --max_new_tokens=100
+```
+
+If you'd like to sample from a model you trained, use the `--out_dir` to point the code appropriately. You can also prompt the model with some text from a file, e.g. `$ python sample.py --start=FILE:prompt.txt`.
+
+## efficiency notes
+
+For simple model benchmarking and profiling, `bench.py` might be useful. It's identical to what happens in the meat of the training loop of `train.py`, but omits much of the other complexities.
+
+Note that the code by default uses [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/). At the time of writing (Dec 29, 2022) this makes `torch.compile()` available in the nightly release. The improvement from the one line of code is noticeable, e.g. cutting down iteration time from ~250ms / iter to 135ms / iter. Nice work PyTorch team!
+
+## todos
+
+- Investigate and add FSDP instead of DDP
+- Eval zero-shot perplexities on standard evals (e.g. LAMBADA? HELM? etc.)
+- Finetune the finetuning script, I think the hyperparams are not great
+- Schedule for linear batch size increase during training
+- Incorporate other embeddings (rotary, alibi)
+- Separate out the optim buffers from model params in checkpoints I think
+- Additional logging around network health (e.g. gradient clip events, magnitudes)
+- Few more investigations around better init etc.
+
+## troubleshooting
+
+Note that by default this repo uses PyTorch 2.0 (i.e. `torch.compile`). This is fairly new and experimental, and not yet available on all platforms (e.g. Windows). If you're running into related error messages try to disable this by adding `--compile=False` flag. This will slow down the code but at least it will run.
+
+For some context on this repository, GPT, and language modeling it might be helpful to watch my [Zero To Hero series](https://karpathy.ai/zero-to-hero.html). Specifically, the [GPT video](https://www.youtube.com/watch?v=kCc8FmEb1nY) is popular if you have some prior language modeling context.
+
+For more questions/discussions feel free to stop by **#nanoGPT** on Discord:
+
+[![](https://dcbadge.vercel.app/api/server/3zy8kqD9Cp?compact=true&style=flat)](https://discord.gg/3zy8kqD9Cp)
+
+## acknowledgements
+
+All nanoGPT experiments are powered by GPUs on [Lambda labs](https://lambdalabs.com), my favorite Cloud GPU provider. Thank you Lambda labs for sponsoring nanoGPT!

bench.py

@@ -0,0 +1,117 @@
+"""
+A much shorter version of train.py for benchmarking
+"""
+import os
+from contextlib import nullcontext
+import numpy as np
+import time
+import torch
+from model import GPTConfig, GPT
+
+# -----------------------------------------------------------------------------
+batch_size = 12
+block_size = 1024
+bias = False
+real_data = True
+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
+profile = False # use pytorch profiler, or just simple benchmarking?
+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)
+
+# data loading init
+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.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
+        return x, y
+else:
+    # alternatively, if fixed data is desired to not care about data loading
+    x = torch.randint(50304, (batch_size, block_size), device=device)
+    y = torch.randint(50304, (batch_size, block_size), device=device)
+    get_batch = lambda split: (x, y)
+
+# model init
+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
+    bias = bias,
+)
+model = GPT(gptconf)
+model.to(device)
+
+optimizer = model.configure_optimizers(weight_decay=1e-2, learning_rate=1e-4, betas=(0.9, 0.95), device_type=device_type)
+
+if compile:
+    print("Compiling model...")
+    model = torch.compile(model) # pytorch 2.0
+
+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=False,
+        profile_memory=False,
+        with_stack=False, # incurs an additional overhead, disable if not needed
+        with_flops=True,
+        with_modules=False, # only for torchscript models atm
+    ) as prof:
+
+        X, Y = get_batch('train')
+        for k in range(num_steps):
+            with ctx:
+                logits, loss = model(X, Y)
+            X, Y = get_batch('train')
+            optimizer.zero_grad(set_to_none=True)
+            loss.backward()
+            optimizer.step()
+            lossf = loss.item()
+            print(f"{k}/{num_steps} loss: {lossf:.4f}")
+
+            prof.step() # notify the profiler at end of each step
+
+else:
+
+    # simple benchmarking
+    torch.cuda.synchronize()
+    for stage, num_steps in enumerate([10, 20]): # burnin, then benchmark
+        t0 = time.time()
+        X, Y = get_batch('train')
+        for k in range(num_steps):
+            with ctx:
+                logits, loss = model(X, Y)
+            X, Y = get_batch('train')
+            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()
+        dt = t1-t0
+        mfu = model.estimate_mfu(batch_size * 1 * num_steps, dt)
+        if stage == 1:
+            print(f"time per iteration: {dt/num_steps*1000:.4f}ms, MFU: {mfu*100:.2f}%")

configurator.py

@@ -0,0 +1,47 @@
+"""
+Poor Man's Configurator. Probably a terrible idea. Example usage:
+$ python train.py config/override_file.py --batch_size=32
+this will first run config/override_file.py, then override batch_size to 32
+
+The code in this file will be run as follows from e.g. train.py:
+>>> exec(open('configurator.py').read())
+
+So it's not a Python module, it's just shuttling this code away from train.py
+The code in this script then overrides the globals()
+
+I know people are not going to love this, I just really dislike configuration
+complexity and having to prepend config. to every single variable. If someone
+comes up with a better simple Python solution I am all ears.
+"""
+
+import sys
+from ast import literal_eval
+
+for arg in sys.argv[1:]:
+    if '=' not in arg:
+        # assume it's the name of a config file
+        assert not arg.startswith('--')
+        config_file = arg
+        print(f"Overriding config with {config_file}:")
+        with open(config_file) as f:
+            print(f.read())
+        exec(open(config_file).read())
+    else:
+        # assume it's a --key=value argument
+        assert arg.startswith('--')
+        key, val = arg.split('=')
+        key = key[2:]
+        if key in globals():
+            try:
+                # attempt to eval it it (e.g. if bool, number, or etc)
+                attempt = literal_eval(val)
+            except (SyntaxError, ValueError):
+                # if that goes wrong, just use the string
+                attempt = val
+            # ensure the types match ok
+            assert type(attempt) == type(globals()[key])
+            # cross fingers
+            print(f"Overriding: {key} = {attempt}")
+            globals()[key] = attempt
+        else:
+            raise ValueError(f"Unknown config key: {key}")

model.py

@@ -0,0 +1,337 @@
+"""
+Full definition of a GPT Language Model, all of it in this single file.
+References:
+1) the official GPT-2 TensorFlow implementation released by OpenAI:
+https://github.com/openai/gpt-2/blob/master/src/model.py
+2) huggingface/transformers PyTorch implementation:
+https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py
+"""
+
+import math
+import inspect
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+# @torch.jit.script # good to enable when not using torch.compile, disable when using (our default)
+def new_gelu(x):
+    """
+    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT).
+    Reference: Gaussian Error Linear Units (GELU) paper: https://arxiv.org/abs/1606.08415
+    """
+    return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+class LayerNorm(nn.Module):
+    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """
+
+    def __init__(self, ndim, bias):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(ndim))
+        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
+
+    def forward(self, input):
+        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        if not self.flash:
+            print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
+                                        .view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k, v  = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout if self.training else 0, is_causal=True)
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return y
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = new_gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024
+    vocab_size: int = 50304 # GPT-2 vocab_size of 50257, padded up to nearest multiple of 64 for efficiency
+    n_layer: int = 12
+    n_head: int = 12
+    n_embd: int = 768
+    dropout: float = 0.0
+    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.vocab_size is not None
+        assert config.block_size is not None
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            drop = nn.Dropout(config.dropout),
+            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
+            ln_f = LayerNorm(config.n_embd, bias=config.bias),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        # with weight tying when using torch.compile() some warnings get generated:
+        # "UserWarning: functional_call was passed multiple values for tied weights.
+        # This behavior is deprecated and will be an error in future versions"
+        # not 100% sure what this is, so far seems to be harmless. TODO investigate
+        self.transformer.wte.weight = self.lm_head.weight # https://paperswithcode.com/method/weight-tying
+
+        # init all weights
+        self.apply(self._init_weights)
+        # apply special scaled init to the residual projections, per GPT-2 paper
+        for pn, p in self.named_parameters():
+            if pn.endswith('c_proj.weight'):
+                torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))
+
+        # report number of parameters
+        print("number of parameters: %.2fM" % (self.get_num_params()/1e6,))
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+
+    def forward(self, idx, targets=None):
+        device = idx.device
+        b, t = idx.size()
+        assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+        pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0) # shape (1, t)
+
+        # forward the GPT model itself
+        tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (1, t, n_embd)
+        x = self.transformer.drop(tok_emb + pos_emb)
+        for block in self.transformer.h:
+            x = block(x)
+        x = self.transformer.ln_f(x)
+
+        if targets is not None:
+            # if we are given some desired targets also calculate the loss
+            logits = self.lm_head(x)
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
+        else:
+            # inference-time mini-optimization: only forward the lm_head on the very last position
+            logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
+            loss = None
+
+        return logits, loss
+
+    def crop_block_size(self, block_size):
+        # model surgery to decrease the block size if necessary
+        # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024)
+        # but want to use a smaller block size for some smaller, simpler model
+        assert block_size <= self.config.block_size
+        self.config.block_size = block_size
+        self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:block_size])
+        for block in self.transformer.h:
+            if hasattr(block.attn, 'bias'):
+                block.attn.bias = block.attn.bias[:,:,:block_size,:block_size]
+
+    @classmethod
+    def from_pretrained(cls, model_type, override_args=None):
+        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
+        override_args = override_args or {} # default to empty dict
+        # only dropout can be overridden see more notes below
+        assert all(k == 'dropout' for k in override_args)
+        from transformers import GPT2LMHeadModel
+        print("loading weights from pretrained gpt: %s" % model_type)
+
+        # n_layer, n_head and n_embd are determined from model_type
+        config_args = {
+            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
+            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
+            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
+            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
+        }[model_type]
+        print("forcing vocab_size=50257, block_size=1024, bias=True")
+        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
+        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
+        config_args['bias'] = True # always True for GPT model checkpoints
+        # we can override the dropout rate, if desired
+        if 'dropout' in override_args:
+            print(f"overriding dropout rate to {override_args['dropout']}")
+            config_args['dropout'] = override_args['dropout']
+        # create a from-scratch initialized minGPT model
+        config = GPTConfig(**config_args)
+        model = GPT(config)
+        sd = model.state_dict()
+        sd_keys = sd.keys()
+        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param
+
+        # init a huggingface/transformers model
+        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
+        sd_hf = model_hf.state_dict()
+
+        # copy while ensuring all of the parameters are aligned and match in names and shapes
+        sd_keys_hf = sd_hf.keys()
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
+        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
+        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
+        # this means that we have to transpose these weights when we import them
+        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
+        for k in sd_keys_hf:
+            if any(k.endswith(w) for w in transposed):
+                # special treatment for the Conv1D weights we need to transpose
+                assert sd_hf[k].shape[::-1] == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k].t())
+            else:
+                # vanilla copy over the other parameters
+                assert sd_hf[k].shape == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k])
+
+        return model
+
+    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
+        # start with all of the candidate parameters
+        param_dict = {pn: p for pn, p in self.named_parameters()}
+        # filter out those that do not require grad
+        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
+        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
+        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
+        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
+        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
+        optim_groups = [
+            {'params': decay_params, 'weight_decay': weight_decay},
+            {'params': nodecay_params, 'weight_decay': 0.0}
+        ]
+        num_decay_params = sum(p.numel() for p in decay_params)
+        num_nodecay_params = sum(p.numel() for p in nodecay_params)
+        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
+        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
+        # Create AdamW optimizer and use the fused version if it is available
+        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
+        use_fused = fused_available and device_type == 'cuda'
+        extra_args = dict(fused=True) if use_fused else dict()
+        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args)
+        print(f"using fused AdamW: {use_fused}")
+
+        return optimizer
+
+    def estimate_mfu(self, fwdbwd_per_iter, dt):
+        """ estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """
+        # first estimate the number of flops we do per iteration.
+        # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311
+        N = self.get_num_params()
+        cfg = self.config
+        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd//cfg.n_head, cfg.block_size
+        flops_per_token = 6*N + 12*L*H*Q*T
+        flops_per_fwdbwd = flops_per_token * T
+        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
+        # express our flops throughput as ratio of A100 bfloat16 peak flops
+        flops_achieved = flops_per_iter * (1.0/dt) # per second
+        flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS
+        mfu = flops_achieved / flops_promised
+        return mfu
+
+    @torch.no_grad()
+    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
+        """
+        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
+        the sequence max_new_tokens times, feeding the predictions back into the model each time.
+        Most likely you'll want to make sure to be in model.eval() mode of operation for this.
+        """
+        for _ in range(max_new_tokens):
+            # if the sequence context is growing too long we must crop it at block_size
+            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
+            # forward the model to get the logits for the index in the sequence
+            logits, _ = self(idx_cond)
+            # pluck the logits at the final step and scale by desired temperature
+            logits = logits[:, -1, :] / temperature
+            # optionally crop the logits to only the top k options
+            if top_k is not None:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                logits[logits < v[:, [-1]]] = -float('Inf')
+            # apply softmax to convert logits to (normalized) probabilities
+            probs = F.softmax(logits, dim=-1)
+            # sample from the distribution
+            idx_next = torch.multinomial(probs, num_samples=1)
+            # append sampled index to the running sequence and continue
+            idx = torch.cat((idx, idx_next), dim=1)
+
+        return idx

sample.py

@@ -0,0 +1,89 @@
+"""
+Sample from a trained model
+"""
+import os
+import pickle
+from contextlib import nullcontext
+import torch
+import tiktoken
+from model import GPTConfig, GPT
+
+# -----------------------------------------------------------------------------
+init_from = 'resume' # either 'resume' (from an out_dir) or a gpt2 variant (e.g. 'gpt2-xl')
+out_dir = 'out' # ignored if init_from is not 'resume'
+start = "\n" # or "<|endoftext|>" or etc. Can also specify a file, use as: "FILE:prompt.txt"
+num_samples = 10 # number of samples to draw
+max_new_tokens = 500 # number of tokens generated in each sample
+temperature = 0.8 # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions
+top_k = 200 # retain only the top_k most likely tokens, clamp others to have 0 probability
+seed = 1337
+device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1', etc.
+dtype = 'bfloat16' # 'float32' or 'bfloat16' or 'float16'
+compile = False # 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)
+
+# model
+if init_from == 'resume':
+    # init from a model saved in a specific directory
+    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    gptconf = GPTConfig(**checkpoint['model_args'])
+    model = GPT(gptconf)
+    state_dict = checkpoint['model']
+    unwanted_prefix = '_orig_mod.'
+    for k,v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
+    model.load_state_dict(state_dict)
+elif init_from.startswith('gpt2'):
+    # init from a given GPT-2 model
+    model = GPT.from_pretrained(init_from, dict(dropout=0.0))
+
+model.eval()
+model.to(device)
+if compile:
+    model = torch.compile(model) # requires PyTorch 2.0 (optional)
+
+# look for the meta pickle in case it is available in the dataset folder
+load_meta = False
+if init_from == 'resume' and 'config' in checkpoint and 'dataset' in checkpoint['config']: # older checkpoints might not have these...
+    meta_path = os.path.join('data', checkpoint['config']['dataset'], 'meta.pkl')
+    load_meta = os.path.exists(meta_path)
+if load_meta:
+    print(f"Loading meta from {meta_path}...")
+    with open(meta_path, 'rb') as f:
+        meta = pickle.load(f)
+    # TODO want to make this more general to arbitrary encoder/decoder schemes
+    stoi, itos = meta['stoi'], meta['itos']
+    encode = lambda s: [stoi[c] for c in s]
+    decode = lambda l: ''.join([itos[i] for i in l])
+else:
+    # ok let's assume gpt-2 encodings by default
+    print("No meta.pkl found, assuming GPT-2 encodings...")
+    enc = tiktoken.get_encoding("gpt2")
+    encode = lambda s: enc.encode(s, allowed_special={"<|endoftext|>"})
+    decode = lambda l: enc.decode(l)
+
+# encode the beginning of the prompt
+if start.startswith('FILE:'):
+    with open(start[5:], 'r', encoding='utf-8') as f:
+        start = f.read()
+start_ids = encode(start)
+x = (torch.tensor(start_ids, dtype=torch.long, device=device)[None, ...])
+
+# run generation
+with torch.no_grad():
+    with ctx:
+        for k in range(num_samples):
+            y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
+            print(decode(y[0].tolist()))
+            print('---------------')

train.py

@@ -0,0 +1,332 @@
+"""
+This training script can be run both on a single gpu in debug mode,
+and also in a larger training run with distributed data parallel (ddp).
+
+To run on a single GPU, example:
+$ python train.py --batch_size=32 --compile=False
+
+To run with DDP on 4 gpus on 1 node, example:
+$ torchrun --standalone --nproc_per_node=4 train.py
+
+To run with DDP on 4 gpus across 2 nodes, example:
+- Run on the first (master) node with example IP 123.456.123.456:
+$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr=123.456.123.456 --master_port=1234 train.py
+- Run on the worker node:
+$ torchrun --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr=123.456.123.456 --master_port=1234 train.py
+(If your cluster does not have Infiniband interconnect prepend NCCL_IB_DISABLE=1)
+"""
+
+import os
+import time
+import math
+import pickle
+from contextlib import nullcontext
+
+import numpy as np
+import torch
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed import init_process_group, destroy_process_group
+
+from model import GPTConfig, GPT
+
+# -----------------------------------------------------------------------------
+# default config values designed to train a gpt2 (124M) on OpenWebText
+# I/O
+out_dir = 'out'
+eval_interval = 2000
+log_interval = 1
+eval_iters = 200
+eval_only = False # if True, script exits right after the first eval
+always_save_checkpoint = True # if True, always save a checkpoint after each eval
+init_from = 'scratch' # 'scratch' or 'resume' or 'gpt2*'
+# wandb logging
+wandb_log = False # disabled by default
+wandb_project = 'owt'
+wandb_run_name = 'gpt2' # 'run' + str(time.time())
+# data
+dataset = 'openwebtext'
+gradient_accumulation_steps = 5 * 8 # used to simulate larger batch sizes
+batch_size = 12 # if gradient_accumulation_steps > 1, this is the micro-batch size
+block_size = 1024
+# model
+n_layer = 12
+n_head = 12
+n_embd = 768
+dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+
+bias = False # do we use bias inside LayerNorm and Linear layers?
+# adamw optimizer
+learning_rate = 6e-4 # max learning rate
+max_iters = 600000 # total number of training iterations
+weight_decay = 1e-1
+beta1 = 0.9
+beta2 = 0.95
+grad_clip = 1.0 # clip gradients at this value, or disable if == 0.0
+# learning rate decay settings
+decay_lr = True # whether to decay the learning rate
+warmup_iters = 2000 # how many steps to warm up for
+lr_decay_iters = 600000 # should be ~= max_iters per Chinchilla
+min_lr = 6e-5 # minimum learning rate, should be ~= learning_rate/10 per Chinchilla
+# DDP settings
+backend = 'nccl' # 'nccl', 'gloo', etc.
+# system
+device = 'cuda' # examples: 'cpu', 'cuda', 'cuda:0', 'cuda:1' etc., or try 'mps' on macbooks
+dtype = 'bfloat16' # 'float32', 'bfloat16', or 'float16', the latter will auto implement a GradScaler
+compile = True # use PyTorch 2.0 to compile the model to be faster
+# -----------------------------------------------------------------------------
+config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
+exec(open('configurator.py').read()) # overrides from command line or config file
+config = {k: globals()[k] for k in config_keys} # will be useful for logging
+# -----------------------------------------------------------------------------
+
+# various inits, derived attributes, I/O setup
+ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?
+if ddp:
+    init_process_group(backend=backend)
+    ddp_rank = int(os.environ['RANK'])
+    ddp_local_rank = int(os.environ['LOCAL_RANK'])
+    ddp_world_size = int(os.environ['WORLD_SIZE'])
+    device = f'cuda:{ddp_local_rank}'
+    torch.cuda.set_device(device)
+    master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.
+    seed_offset = ddp_rank # each process gets a different seed
+    assert gradient_accumulation_steps % torch.cuda.device_count() == 0
+    gradient_accumulation_steps //= torch.cuda.device_count()
+else:
+    # if not ddp, we are running on a single gpu, and one process
+    master_process = True
+    seed_offset = 0
+    ddp_world_size = 1
+tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
+print(f"tokens per iteration will be: {tokens_per_iter:,}")
+
+if master_process:
+    os.makedirs(out_dir, exist_ok=True)
+torch.manual_seed(1337 + seed_offset)
+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
+# note: float16 data type will automatically use a GradScaler
+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)
+
+# poor man's data loader
+data_dir = os.path.join('data', dataset)
+train_data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
+val_data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')
+def get_batch(split):
+    data = train_data if split == 'train' else val_data
+    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])
+    if device_type == 'cuda':
+        # pin arrays x,y, which allows us to move them to GPU asynchronously (non_blocking=True)
+        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
+    else:
+        x, y = x.to(device), y.to(device)
+    return x, y
+
+# init these up here, can override if init_from='resume' (i.e. from a checkpoint)
+iter_num = 0
+best_val_loss = 1e9
+
+# attempt to derive vocab_size from the dataset
+meta_path = os.path.join(data_dir, 'meta.pkl')
+meta_vocab_size = None
+if os.path.exists(meta_path):
+    with open(meta_path, 'rb') as f:
+        meta = pickle.load(f)
+    meta_vocab_size = meta['vocab_size']
+    print(f"found vocab_size = {meta_vocab_size} (inside {meta_path})")
+
+# model init
+model_args = dict(n_layer=n_layer, n_head=n_head, n_embd=n_embd, block_size=block_size,
+                  bias=bias, vocab_size=None, dropout=dropout) # start with model_args from command line
+if init_from == 'scratch':
+    # init a new model from scratch
+    print("Initializing a new model from scratch")
+    # determine the vocab size we'll use for from-scratch training
+    if meta_vocab_size is None:
+        print("defaulting to vocab_size of GPT-2 to 50304 (50257 rounded up for efficiency)")
+    model_args['vocab_size'] = meta_vocab_size if meta_vocab_size is not None else 50304
+    gptconf = GPTConfig(**model_args)
+    model = GPT(gptconf)
+elif init_from == 'resume':
+    print(f"Resuming training from {out_dir}")
+    # resume training from a checkpoint.
+    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    checkpoint_model_args = checkpoint['model_args']
+    # force these config attributes to be equal otherwise we can't even resume training
+    # the rest of the attributes (e.g. dropout) can stay as desired from command line
+    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
+        model_args[k] = checkpoint_model_args[k]
+    # create the model
+    gptconf = GPTConfig(**model_args)
+    model = GPT(gptconf)
+    state_dict = checkpoint['model']
+    # fix the keys of the state dictionary :(
+    # honestly no idea how checkpoints sometimes get this prefix, have to debug more
+    unwanted_prefix = '_orig_mod.'
+    for k,v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
+    model.load_state_dict(state_dict)
+    iter_num = checkpoint['iter_num']
+    best_val_loss = checkpoint['best_val_loss']
+elif init_from.startswith('gpt2'):
+    print(f"Initializing from OpenAI GPT-2 weights: {init_from}")
+    # initialize from OpenAI GPT-2 weights
+    override_args = dict(dropout=dropout)
+    model = GPT.from_pretrained(init_from, override_args)
+    # read off the created config params, so we can store them into checkpoint correctly
+    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
+        model_args[k] = getattr(model.config, k)
+# crop down the model block size if desired, using model surgery
+if block_size < model.config.block_size:
+    model.crop_block_size(block_size)
+    model_args['block_size'] = block_size # so that the checkpoint will have the right value
+model.to(device)
+
+# initialize a GradScaler. If enabled=False scaler is a no-op
+scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))
+
+# optimizer
+optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
+if init_from == 'resume':
+    optimizer.load_state_dict(checkpoint['optimizer'])
+checkpoint = None # free up memory
+
+# compile the model
+if compile:
+    print("compiling the model... (takes a ~minute)")
+    unoptimized_model = model
+    model = torch.compile(model) # requires PyTorch 2.0
+
+# wrap model into DDP container
+if ddp:
+    model = DDP(model, device_ids=[ddp_local_rank])
+
+# helps estimate an arbitrarily accurate loss over either split using many batches
+@torch.no_grad()
+def estimate_loss():
+    out = {}
+    model.eval()
+    for split in ['train', 'val']:
+        losses = torch.zeros(eval_iters)
+        for k in range(eval_iters):
+            X, Y = get_batch(split)
+            with ctx:
+                logits, loss = model(X, Y)
+            losses[k] = loss.item()
+        out[split] = losses.mean()
+    model.train()
+    return out
+
+# learning rate decay scheduler (cosine with warmup)
+def get_lr(it):
+    # 1) linear warmup for warmup_iters steps
+    if it < warmup_iters:
+        return learning_rate * it / warmup_iters
+    # 2) if it > lr_decay_iters, return min learning rate
+    if it > lr_decay_iters:
+        return min_lr
+    # 3) in between, use cosine decay down to min learning rate
+    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
+    assert 0 <= decay_ratio <= 1
+    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff ranges 0..1
+    return min_lr + coeff * (learning_rate - min_lr)
+
+# logging
+if wandb_log and master_process:
+    import wandb
+    wandb.init(project=wandb_project, name=wandb_run_name, config=config)
+
+# training loop
+X, Y = get_batch('train') # fetch the very first batch
+t0 = time.time()
+local_iter_num = 0 # number of iterations in the lifetime of this process
+raw_model = model.module if ddp else model # unwrap DDP container if needed
+running_mfu = -1.0
+while True:
+
+    # determine and set the learning rate for this iteration
+    lr = get_lr(iter_num) if decay_lr else learning_rate
+    for param_group in optimizer.param_groups:
+        param_group['lr'] = lr
+
+    # evaluate the loss on train/val sets and write checkpoints
+    if iter_num % eval_interval == 0 and master_process:
+        losses = estimate_loss()
+        print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
+        if wandb_log:
+            wandb.log({
+                "iter": iter_num,
+                "train/loss": losses['train'],
+                "val/loss": losses['val'],
+                "lr": lr,
+                "mfu": running_mfu*100, # convert to percentage
+            })
+        if losses['val'] < best_val_loss or always_save_checkpoint:
+            best_val_loss = losses['val']
+            if iter_num > 0:
+                checkpoint = {
+                    'model': raw_model.state_dict(),
+                    'optimizer': optimizer.state_dict(),
+                    'model_args': model_args,
+                    'iter_num': iter_num,
+                    'best_val_loss': best_val_loss,
+                    'config': config,
+                }
+                print(f"saving checkpoint to {out_dir}")
+                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))
+    if iter_num == 0 and eval_only:
+        
+        break
+
+    # forward backward update, with optional gradient accumulation to simulate larger batch size
+    # and using the GradScaler if data type is float16
+    for micro_step in range(gradient_accumulation_steps):
+        if ddp:
+            # in DDP training we only need to sync gradients at the last micro step.
+            # the official way to do this is with model.no_sync() context manager, but
+            # I really dislike that this bloats the code and forces us to repeat code
+            # looking at the source of that context manager, it just toggles this variable
+            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
+        with ctx:
+            logits, loss = model(X, Y)
+            loss = loss / gradient_accumulation_steps # scale the loss to account for gradient accumulation
+        # immediately async prefetch next batch while model is doing the forward pass on the GPU
+        X, Y = get_batch('train')
+        # backward pass, with gradient scaling if training in fp16
+        scaler.scale(loss).backward()
+    # clip the gradient
+    if grad_clip != 0.0:
+        scaler.unscale_(optimizer)
+        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
+    # step the optimizer and scaler if training in fp16
+    scaler.step(optimizer)
+    scaler.update()
+    # flush the gradients as soon as we can, no need for this memory anymore
+    optimizer.zero_grad(set_to_none=True)
+
+    # timing and logging
+    t1 = time.time()
+    dt = t1 - t0
+    t0 = t1
+    if iter_num % log_interval == 0 and master_process:
+        # get loss as float. note: this is a CPU-GPU sync point
+        # scale up to undo the division above, approximating the true total loss (exact would have been a sum)
+        lossf = loss.item() * gradient_accumulation_steps
+        if local_iter_num >= 5: # let the training loop settle a bit
+            mfu = raw_model.estimate_mfu(batch_size * gradient_accumulation_steps, dt)
+            running_mfu = mfu if running_mfu == -1.0 else 0.9*running_mfu + 0.1*mfu
+        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%")
+    iter_num += 1
+    local_iter_num += 1
+
+    # termination conditions
+    if iter_num > max_iters:
+        break
+
+if ddp:
+    destroy_process_group()

transformer_sizing.ipynb

@@ -0,0 +1,402 @@
+{
+ "cells": [
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Transformer Theoretical Model\n",
+    "\n",
+    "This notebook stores a bunch of analysis about a Transformer, e.g. estimates the number of FLOPs, parameters, peak memory footprint, checkpoint size, etc."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from collections import OrderedDict"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# config_args = {\n",
+    "#     'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params\n",
+    "#     'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params\n",
+    "#     'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params\n",
+    "#     'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params\n",
+    "# }[model_type]\n",
+    "\n",
+    "block_size = 1024\n",
+    "vocab_size = 50257\n",
+    "n_layer = 12\n",
+    "n_head = 12\n",
+    "n_embd = 768\n",
+    "bias = False\n",
+    "assert not bias, \"this notebook assumes bias=False just for simplicity\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "we see: 124337664, expected: 124337664, match: True\n",
+      "name                 params     ratio (%) \n",
+      "emebedding/position      786432     0.6325\n",
+      "embedding/token        38597376    31.0424\n",
+      "embedding              39383808    31.6749\n",
+      "attention/ln                768     0.0006\n",
+      "attention/kqv           1769472     1.4231\n",
+      "attention/proj           589824     0.4744\n",
+      "attention               2360064     1.8981\n",
+      "mlp/ln                      768     0.0006\n",
+      "mlp/ffw                 2359296     1.8975\n",
+      "mlp/proj                2359296     1.8975\n",
+      "mlp                     4719360     3.7956\n",
+      "block                   7079424     5.6937\n",
+      "transformer            84953088    68.3245\n",
+      "ln_f                        768     0.0006\n",
+      "dense                         0     0.0000\n",
+      "total                 124337664   100.0000\n"
+     ]
+    }
+   ],
+   "source": [
+    "def params():\n",
+    "    \"\"\" estimates the number of parameters in the model\"\"\"\n",
+    "    out = OrderedDict()\n",
+    "\n",
+    "    # token and position embeddings\n",
+    "    out['emebedding/position'] = n_embd * block_size\n",
+    "    out['embedding/token'] = n_embd * vocab_size\n",
+    "    out['embedding'] = out['emebedding/position'] + out['embedding/token']\n",
+    "\n",
+    "    # attention blocks\n",
+    "    out['attention/ln'] = n_embd # note, bias=False in our LN\n",
+    "    out['attention/kqv'] = n_embd * 3*n_embd\n",
+    "    out['attention/proj'] = n_embd**2\n",
+    "    out['attention'] = out['attention/ln'] + out['attention/kqv'] + out['attention/proj']\n",
+    "\n",
+    "    # MLP blocks\n",
+    "    ffw_size = 4*n_embd # feed forward size\n",
+    "    out['mlp/ln'] = n_embd\n",
+    "    out['mlp/ffw'] = n_embd * ffw_size\n",
+    "    out['mlp/proj'] = ffw_size * n_embd\n",
+    "    out['mlp'] = out['mlp/ln'] + out['mlp/ffw'] + out['mlp/proj']\n",
+    "    \n",
+    "    # the transformer and the rest of it\n",
+    "    out['block'] = out['attention'] + out['mlp']\n",
+    "    out['transformer'] = n_layer * out['block']\n",
+    "    out['ln_f'] = n_embd # final layernorm\n",
+    "    out['dense'] = 0 # 0 because of parameter sharing. This layer uses the weights from the embedding layer\n",
+    "\n",
+    "    # total\n",
+    "    out['total'] = out['embedding'] + out['transformer'] + out['ln_f'] + out['dense']\n",
+    "\n",
+    "    return out\n",
+    "\n",
+    "# compare our param count to that reported by PyTorch\n",
+    "p = params()\n",
+    "params_total = p['total']\n",
+    "print(f\"we see: {params_total}, expected: {124337664}, match: {params_total == 124337664}\")\n",
+    "# create a header\n",
+    "print(f\"{'name':20s} {'params':10s} {'ratio (%)':10s}\")\n",
+    "for k,v in p.items():\n",
+    "    print(f\"{k:20s} {v:10d} {v/params_total*100:10.4f}\")\n",
+    "    "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "est checkpoint size: 1.49 GB\n",
+      "measured with wc -c ckpt.pt: 1542470366\n",
+      "fluff ratio: 103.38%\n"
+     ]
+    }
+   ],
+   "source": [
+    "# we can now calculate the size of each checkpoint\n",
+    "# params are stored in fp32, and the AdamW optimizer has 2 additional buffers per param for statistics\n",
+    "params_bytes = params_total*4\n",
+    "params_and_buffers_bytes = params_bytes + 2*params_bytes\n",
+    "print(f\"est checkpoint size: {params_and_buffers_bytes/1e9:.2f} GB\")\n",
+    "measured_bytes = 1542470366 # from wc -c ckpt.pt\n",
+    "print(f\"measured with wc -c ckpt.pt: {measured_bytes}\")\n",
+    "print(f\"fluff ratio: {measured_bytes/params_and_buffers_bytes*100:.2f}%\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We can also estimate the ratio of our GPU memory that will be taken up just by the weights and the buffers inside the AdamW optimizer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "memory ratio taken up just for parameters: 3.73%\n"
+     ]
+    }
+   ],
+   "source": [
+    "gpu_memory = 40e9 # 40 GB A100 GPU, roughly\n",
+    "print(f\"memory ratio taken up just for parameters: {params_and_buffers_bytes / gpu_memory * 100:.2f}%\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "i.e. not that much of the memory for this tiny model, most of the memory is activations (forward and backward). This of course changes dramatically for larger and larger models."
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Let's estimate FLOPs for a single forward pass."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "name                 flops          ratio (%) \n",
+      "attention/kqv            3623878656     1.2426\n",
+      "attention/scores         1610612736     0.5522\n",
+      "attention/reduce         1610612736     0.5522\n",
+      "attention/proj           1207959552     0.4142\n",
+      "attention                8053063680     2.7612\n",
+      "mlp/ffw1                 4831838208     1.6567\n",
+      "mlp/ffw2                 4831838208     1.6567\n",
+      "mlp                      9663676416     3.3135\n",
+      "block                   17716740096     6.0747\n",
+      "transformer            212600881152    72.8963\n",
+      "dense                   79047426048    27.1037\n",
+      "forward_total          291648307200   100.0000\n",
+      "backward_total         583296614400   200.0000\n",
+      "total                  874944921600   300.0000\n"
+     ]
+    }
+   ],
+   "source": [
+    "def flops():\n",
+    "    # we only count Weight FLOPs, all other layers (LayerNorm, Softmax, etc) are effectively irrelevant\n",
+    "    # we count actual FLOPs, not MACs. Hence 2* all over the place\n",
+    "    # basically for any matrix multiply A (BxC) @ B (CxD) -> (BxD) flops are 2*B*C*D\n",
+    "\n",
+    "    out = OrderedDict()\n",
+    "    head_size = n_embd // n_head\n",
+    "\n",
+    "    # attention blocks\n",
+    "    # 1) the projection to key, query, values\n",
+    "    out['attention/kqv'] = 2 * block_size * (n_embd * 3*n_embd)\n",
+    "    # 2) calculating the attention scores\n",
+    "    out['attention/scores'] = 2 * block_size * block_size * n_embd\n",
+    "    # 3) the reduction of the values (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)\n",
+    "    out['attention/reduce'] = 2 * n_head * (block_size * block_size * head_size)\n",
+    "    # 4) the final linear projection\n",
+    "    out['attention/proj'] = 2 * block_size * (n_embd * n_embd)\n",
+    "    out['attention'] = sum(out['attention/'+k] for k in ['kqv', 'scores', 'reduce', 'proj'])\n",
+    "\n",
+    "    # MLP blocks\n",
+    "    ffw_size = 4*n_embd # feed forward size\n",
+    "    out['mlp/ffw1'] = 2 * block_size * (n_embd * ffw_size)\n",
+    "    out['mlp/ffw2'] = 2 * block_size * (ffw_size * n_embd)\n",
+    "    out['mlp'] = out['mlp/ffw1'] + out['mlp/ffw2']\n",
+    "\n",
+    "    # the transformer and the rest of it\n",
+    "    out['block'] = out['attention'] + out['mlp']\n",
+    "    out['transformer'] = n_layer * out['block']\n",
+    "    out['dense'] = 2 * block_size * (n_embd * vocab_size)\n",
+    "\n",
+    "    # forward,backward,total\n",
+    "    out['forward_total'] = out['transformer'] + out['dense']\n",
+    "    out['backward_total'] = 2 * out['forward_total'] # use common estimate of bwd = 2*fwd\n",
+    "    out['total'] = out['forward_total'] + out['backward_total']\n",
+    "\n",
+    "    return out\n",
+    "    \n",
+    "# compare our param count to that reported by PyTorch\n",
+    "f = flops()\n",
+    "flops_total = f['forward_total']\n",
+    "print(f\"{'name':20s} {'flops':14s} {'ratio (%)':10s}\")\n",
+    "for k,v in f.items():\n",
+    "    print(f\"{k:20s} {v:14d} {v/flops_total*100:10.4f}\")\n",
+    "    "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "palm_flops: 875062886400, flops: 874944921600, ratio: 1.0001\n"
+     ]
+    }
+   ],
+   "source": [
+    "# now here is an estimate copy pasted from the PaLM paper\n",
+    "# this formula is often used to calculate MFU (model flops utilization)\n",
+    "def palm_flops():\n",
+    "    \"\"\"estimate of the model flops following PaLM paper formula\"\"\"\n",
+    "    # non-embedding model parameters. note that we do not subtract the\n",
+    "    # embedding/token params because those are tied and get used in the last layer.\n",
+    "    N = params()['total'] - params()['emebedding/position']\n",
+    "    L, H, Q, T = n_layer, n_head, n_embd//n_head, block_size\n",
+    "    mf_per_token = 6*N + 12*L*H*Q*T\n",
+    "    mf = mf_per_token * block_size\n",
+    "    return mf\n",
+    "\n",
+    "print(f\"palm_flops: {palm_flops():d}, flops: {flops()['total']:d}, ratio: {palm_flops()/flops()['total']:.4f}\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Ok they are quite similar, giving some confidence that my math in flops() function was ~ok. Now, A100 is cited at 312TFLOPS bfloat16 on tensor cores. So what is our model flops utilization (MFU)? I trained the model above with a batch_size of 20 and grad_accum of 5, which runs in about 755ms on a single A100 GPU. We get:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "fraction of A100 used: 37.14%\n"
+     ]
+    }
+   ],
+   "source": [
+    "# here is what we currently roughly measure\n",
+    "batch_size = 20 * 5 # 5 is grad_accum, so total batch size is 100\n",
+    "measured_time = 0.755 # in seconds per iteration\n",
+    "measured_throughput = batch_size / measured_time\n",
+    "flops_achieved = f['total'] * measured_throughput\n",
+    "\n",
+    "# A100 is cited to be 312 TFLOPS of bloat16 running on tensor cores\n",
+    "a100_flops_promised = 312e12\n",
+    "\n",
+    "# the fraction of the A100 that we are using:\n",
+    "print(f\"fraction of A100 used: {flops_achieved / a100_flops_promised * 100:.2f}%\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "For reference, we'd prefer to be somewhere around 50%+, and not just for a single GPU but for an entire DDP run. So we still have some work to do, but at least we're within a factor of ~2X of what is achievable with this GPU."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "time needed to train the model: 3.46 days\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Finally let's check out the 6ND approximation as total cost of training in FLOPs\n",
+    "model_size = params()['total'] # this is number of parameters, N\n",
+    "tokens_num = 300e9 # 300B tokens, this is dataset size in tokens, D\n",
+    "a100_flops = 312e12 # 312 TFLOPS\n",
+    "assumed_mfu = 0.3 # assume this model flops utilization (take the current 37% from above and add some DDP overhead)\n",
+    "flops_throughput = a100_flops * 8 * assumed_mfu # assume an 8XA100 node at 30% utilization\n",
+    "flops_needed = 6 * model_size * tokens_num # 6ND\n",
+    "time_needed_s = flops_needed / flops_throughput # in seconds\n",
+    "print(f\"time needed to train the model: {time_needed_s/3600/24:.2f} days\")"
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "This is not a bad estimate at all. I trained this model and it converged in roughly 4 days. Btw as a good reference for where 6ND comes from and some intuition around it I recommend [Dzmitry's post](https://medium.com/@dzmitrybahdanau/the-flops-calculus-of-language-model-training-3b19c1f025e4)."
+   ]
+  },
+  {
+   "attachments": {},
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now, FLOPs are just one constraint, the other that we have to keep a close track of is the memory bandwidth. TODO estimate LOAD/STORE costs of our model later."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "pytorch2",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.4"
+  },
+  "orig_nbformat": 4,
+  "vscode": {
+   "interpreter": {
+    "hash": "7f5833218766b48e6e35e4452ee875aac0e2188d05bbe5298f2c62b79f08b222"
+   }
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}