DEMO: minGPT on tinygrad

.gitignore

@@ -3,3 +3,4 @@ __pycache__/
 *.swp
 .env
 .pylintrc
+tinyGPT.egg-info/

LICENSE

@@ -1,4 +1,5 @@
 The MIT License (MIT) Copyright (c) 2020 Andrej Karpathy
+The MIT License (MIT) Copyright (c) 2024 Ziliang Peng
 
 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:
 

README.md

@@ -1,146 +1,33 @@
+# You like karpathy? You like geohot? You love tinyGPT! ❤️
 
-# minGPT
+![tinyGPT](hotzilla.avif)
 
-![mingpt](mingpt.jpg)
+tinyGPT is an attempt to port karpathy's minGPT to geohot's tinygrad. It serves a few purposes:
+- demonstrate API compatibility and diff between PyTorch and Tinygrad
+- Identify missing features/APIs in Tinygrad
+- Benchmark and compare performance
 
-A PyTorch re-implementation of [GPT](https://github.com/openai/gpt-2), both training and inference. minGPT tries to be small, clean, interpretable and educational, as most of the currently available GPT model implementations can a bit sprawling. GPT is not a complicated model and this implementation is appropriately about 300 lines of code (see [mingpt/model.py](mingpt/model.py)). All that's going on is that a sequence of indices feeds into a [Transformer](https://arxiv.org/abs/1706.03762), and a probability distribution over the next index in the sequence comes out. The majority of the complexity is just being clever with batching (both across examples and over sequence length) for efficiency.
+### Library Installation and Test
 
-**note (Jan 2023)**: though I may continue to accept and change some details, minGPT is in a semi-archived state. For more recent developments see my rewrite [nanoGPT](https://github.com/karpathy/nanoGPT). Basically, minGPT became referenced across a wide variety of places (notebooks, blogs, courses, books, etc.) which made me less willing to make the bigger changes I wanted to make to move the code forward. I also wanted to change the direction a bit, from a sole focus on education to something that is still simple and hackable but has teeth (reproduces medium-sized industry benchmarks, accepts some tradeoffs to gain runtime efficiency, etc).
-
-The minGPT library is three files: [mingpt/model.py](mingpt/model.py) contains the actual Transformer model definition, [mingpt/bpe.py](mingpt/bpe.py) contains a mildly refactored Byte Pair Encoder that translates between text and sequences of integers exactly like OpenAI did in GPT, [mingpt/trainer.py](mingpt/trainer.py) is (GPT-independent) PyTorch boilerplate code that trains the model. Then there are a number of demos and projects that use the library in the `projects` folder:
-
-- `projects/adder` trains a GPT from scratch to add numbers (inspired by the addition section in the GPT-3 paper)
-- `projects/chargpt` trains a GPT to be a character-level language model on some input text file
-- `demo.ipynb` shows a minimal usage of the `GPT` and `Trainer` in a notebook format on a simple sorting example
-- `generate.ipynb` shows how one can load a pretrained GPT2 and generate text given some prompt
-
-### Library Installation
-
-If you want to `import mingpt` into your project:
+If you want to `import tinygpt` into your project:
 
 ```
-git clone https://github.com/karpathy/minGPT.git
-cd minGPT
+git clone https://github.com/ziliangpeng/tinyGPT.git
+cd tinyGPT
 pip install -e .
 ```
 
-### Usage
+After that, you can run the demo project to see the result:
 
-Here's how you'd instantiate a GPT-2 (124M param version):
-
-```python
-from mingpt.model import GPT
-model_config = GPT.get_default_config()
-model_config.model_type = 'gpt2'
-model_config.vocab_size = 50257 # openai's model vocabulary
-model_config.block_size = 1024  # openai's model block_size (i.e. input context length)
-model = GPT(model_config)
 ```
-
-And here's how you'd train it:
-
-```python
-# your subclass of torch.utils.data.Dataset that emits example
-# torch LongTensor of lengths up to 1024, with integers from [0,50257)
-train_dataset = YourDataset()
-
-from mingpt.trainer import Trainer
-train_config = Trainer.get_default_config()
-train_config.learning_rate = 5e-4 # many possible options, see the file
-train_config.max_iters = 1000
-train_config.batch_size = 32
-trainer = Trainer(train_config, model, train_dataset)
-trainer.run()
-```
-
-See `demo.ipynb` for a more concrete example.
-
-### Unit tests
-
-Coverage is not super amazing just yet but:
-
+cd project/adder
+python adder.py
 ```
-python -m unittest discover tests
-```
-
-### todos
-
-- add gpt-2 finetuning demo on arbitrary given text file
-- add dialog agent demo
-- better docs of outcomes for existing projects (adder, chargpt)
-- add mixed precision and related training scaling goodies
-- distributed training support
-- reproduce some benchmarks in projects/, e.g. text8 or other language modeling
-- proper logging instead of print statement amateur hour haha
-- i probably should have a requirements.txt file...
-- it should be possible to load in many other model weights other than just gpt2-\*
-
-### References
-
-Code:
-
-- [openai/gpt-2](https://github.com/openai/gpt-2) has the model definition in TensorFlow, but not the training code
-- [openai/image-gpt](https://github.com/openai/image-gpt) has some more modern gpt-3 like modification in its code, good reference as well
-- [huggingface/transformers](https://github.com/huggingface/transformers) has a [language-modeling example](https://github.com/huggingface/transformers/tree/master/examples/pytorch/language-modeling). It is full-featured but as a result also somewhat challenging to trace. E.g. some large functions have as much as 90% unused code behind various branching statements that is unused in the default setting of simple language modeling
-
-Papers + some implementation notes:
-
-#### Improving Language Understanding by Generative Pre-Training (GPT-1)
-
-- Our model largely follows the original transformer work
-- We trained a 12-layer decoder-only transformer with masked self-attention heads (768 dimensional states and 12 attention heads). For the position-wise feed-forward networks, we used 3072 dimensional inner states.
-- Adam max learning rate of 2.5e-4. (later GPT-3 for this model size uses 6e-4)
-- LR decay: increased linearly from zero over the first 2000 updates and annealed to 0 using a cosine schedule
-- We train for 100 epochs on minibatches of 64 randomly sampled, contiguous sequences of 512 tokens.
-- Since layernorm is used extensively throughout the model, a simple weight initialization of N(0, 0.02) was sufficient
-- bytepair encoding (BPE) vocabulary with 40,000 merges
-- residual, embedding, and attention dropouts with a rate of 0.1 for regularization.
-- modified version of L2 regularization proposed in (37), with w = 0.01 on all non bias or gain weights
-- For the activation function, we used the Gaussian Error Linear Unit (GELU).
-- We used learned position embeddings instead of the sinusoidal version proposed in the original work
-- For finetuning: We add dropout to the classifier with a rate of 0.1. learning rate of 6.25e-5 and a batchsize of 32. 3 epochs. We use a linear learning rate decay schedule with warmup over 0.2% of training. λ was set to 0.5.
-- GPT-1 model is 12 layers and d_model 768, ~117M params
-
-#### Language Models are Unsupervised Multitask Learners (GPT-2)
-
-- LayerNorm was moved to the input of each sub-block, similar to a pre-activation residual network
-- an additional layer normalization was added after the final self-attention block.
-- modified initialization which accounts for the accumulation on the residual path with model depth is used. We scale the weights of residual layers at initialization by a factor of 1/√N where N is the number of residual layers. (weird because in their released code i can only find a simple use of the old 0.02... in their release of image-gpt I found it used for c_proj, and even then only for attn, not for mlp. huh. https://github.com/openai/image-gpt/blob/master/src/model.py)
-- the vocabulary is expanded to 50,257
-- increase the context size from 512 to 1024 tokens
-- larger batchsize of 512 is used
-- GPT-2 used 48 layers and d_model 1600 (vs. original 12 layers and d_model 768). ~1.542B params
-
-#### Language Models are Few-Shot Learners (GPT-3)
 
-- GPT-3: 96 layers, 96 heads, with d_model of 12,288 (175B parameters).
-- GPT-1-like: 12 layers, 12 heads, d_model 768 (125M)
-- We use the same model and architecture as GPT-2, including the modified initialization, pre-normalization, and reversible tokenization described therein
-- we use alternating dense and locally banded sparse attention patterns in the layers of the transformer, similar to the Sparse Transformer
-- we always have the feedforward layer four times the size of the bottleneck layer, dff = 4 ∗ dmodel
-- all models use a context window of nctx = 2048 tokens.
-- Adam with β1 = 0.9, β2 = 0.95, and eps = 10−8
-- All models use weight decay of 0.1 to provide a small amount of regularization. (NOTE: GPT-1 used 0.01 I believe, see above)
-- clip the global norm of the gradient at 1.0
-- Linear LR warmup over the first 375 million tokens. Then use cosine decay for learning rate down to 10% of its value, over 260 billion tokens.
-- gradually increase the batch size linearly from a small value (32k tokens) to the full value over the first 4-12 billion tokens of training, depending on the model size.
-- full 2048-sized time context window is always used, with a special END OF DOCUMENT token delimiter
+tinygrad allows you to choose hardware via env vars: `CLANG=1`, `CUDA=1`, `METAL=1`.
 
-#### Generative Pretraining from Pixels (Image GPT)
+And you can choose `DEBUG=` level for increasing amount of debug log. Refer to [tinygrad](https://github.com/tinygrad/tinygrad/blob/master/docs-legacy/env_vars.md) for more env vars control.
 
-- When working with images, we pick the identity permutation πi = i for 1 ≤ i ≤ n, also known as raster order.
-- we create our own 9-bit color palette by clustering (R, G, B) pixel values using k-means with k = 512.
-- Our largest model, iGPT-XL, contains L = 60 layers and uses an embedding size of d = 3072 for a total of 6.8B parameters.
-- Our next largest model, iGPT-L, is essentially identical to GPT-2 with L = 48 layers, but contains a slightly smaller embedding size of d = 1536 (vs 1600) for a total of 1.4B parameters.
-- We use the same model code as GPT-2, except that we initialize weights in the layerdependent fashion as in Sparse Transformer (Child et al., 2019) and zero-initialize all projections producing logits.
-- We also train iGPT-M, a 455M parameter model with L = 36 and d = 1024
-- iGPT-S, a 76M parameter model with L = 24 and d = 512 (okay, and how many heads? looks like the Github code claims 8)
-- When pre-training iGPT-XL, we use a batch size of 64 and train for 2M iterations, and for all other models we use a batch size of 128 and train for 1M iterations.
-- Adam with β1 = 0.9 and β2 = 0.95
-- The learning rate is warmed up for one epoch, and then decays to 0
-- We did not use weight decay because applying a small weight decay of 0.01 did not change representation quality.
-- iGPT-S lr 0.003
-- No dropout is used.
 
 ### License
 

projects/adder/.gitignore

@@ -0,0 +1 @@
+out/

projects/adder/adder.py

@@ -6,13 +6,15 @@
 import sys
 import json
 
-import torch
-from torch.utils.data import Dataset
-from torch.utils.data.dataloader import DataLoader
+import random
+import tinygrad
 
-from mingpt.model import GPT
-from mingpt.trainer import Trainer
-from mingpt.utils import set_seed, setup_logging, CfgNode as CN
+from tinygrad import dtypes
+from tinygpt.tinyloader import TinyDataLoader
+from tinygpt.model import GPT
+from tinygpt.trainer import Trainer
+from tinygpt.utils import set_seed, setup_logging, CfgNode as CN
+from tinygrad.tensor import Tensor
 
 # -----------------------------------------------------------------------------
 
@@ -40,7 +42,7 @@ def get_config():
 
 # -----------------------------------------------------------------------------
 
-class AdditionDataset(Dataset):
+class AdditionDataset:
     """
     Creates n-digit addition problems. For example, if n=2, then an example
     addition problem would be to add 85 + 50 = 135. This problem would be
@@ -79,9 +81,10 @@ def __init__(self, config, split):
         ndigit = self.config.ndigit
         assert ndigit <= 3, "the lines below would be very memory inefficient, in future maybe refactor to support"
         num = (10**ndigit)**2 # total number of possible addition problems with ndigit numbers
-        rng = torch.Generator()
-        rng.manual_seed(1337)
-        perm = torch.randperm(num, generator=rng)
+        random.seed(1337)
+        perm = list(range(num))
+        random.shuffle(perm)
+        perm = tinygrad.tensor.Tensor(perm)
         num_test = min(int(num*0.2), 500) # 20% of the whole dataset, or only up to 500
         self.ixes = perm[:num_test] if split == 'test' else perm[num_test:]
 
@@ -95,6 +98,7 @@ def get_block_size(self):
         return 3*self.config.ndigit + 1 - 1
 
     def __len__(self):
+        # return self.ixes.numel()
         return self.ixes.nelement()
 
     def __getitem__(self, idx):
@@ -113,8 +117,8 @@ def __getitem__(self, idx):
         render = astr + bstr + cstr
         dix = [int(s) for s in render] # convert each character to its token index
         # x will be input to GPT and y will be the associated expected outputs
-        x = torch.tensor(dix[:-1], dtype=torch.long)
-        y = torch.tensor(dix[1:], dtype=torch.long) # predict the next token in the sequence
+        x = Tensor(dix[:-1], dtype=dtypes.long)
+        y = Tensor(dix[1:], dtype=dtypes.long) # predict the next token in the sequence
         y[:ndigit*2-1] = -1 # we will only train in the output locations. -1 will mask loss to zero
         return x, y
 
@@ -147,10 +151,10 @@ def eval_split(trainer, split, max_batches=None):
         ndigit = config.data.ndigit
         results = []
         mistakes_printed_already = 0
-        factors = torch.tensor([[10**i for i in range(ndigit+1)][::-1]]).to(trainer.device)
-        loader = DataLoader(dataset, batch_size=100, num_workers=0, drop_last=False)
+        factors = torch.tensor([[10**i for i in range(ndigit+1)][::-1]])
+        # factors = tinygrad.tensor.Tensor([[10**i for i in range(ndigit+1)][::-1]]).to(tmp_device)
+        loader = TinyDataLoader(dataset, batch_size=100)
         for b, (x, y) in enumerate(loader):
-            x = x.to(trainer.device)
             # isolate the first two digits of the input sequence alone
             d1d2 = x[:, :ndigit*2]
             # let the model sample the rest of the sequence
@@ -172,6 +176,7 @@ def eval_split(trainer, split, max_batches=None):
                     print("GPT claims that %d + %d = %d but gt is %d" % (d1i[i], d2i[i], d3i_pred[i], d3i_gt[i]))
             if max_batches is not None and b+1 >= max_batches:
                 break
+        # rt = tinygrad.tensor.Tensor(results, dtype=tinygrad.dtypes.float)
         rt = torch.tensor(results, dtype=torch.float)
         print("%s final score: %d/%d = %.2f%% correct" % (split, rt.sum(), len(results), 100*rt.mean()))
         return rt.sum()
@@ -184,22 +189,22 @@ def batch_end_callback(trainer):
         if trainer.iter_num % 10 == 0:
             print(f"iter_dt {trainer.iter_dt * 1000:.2f}ms; iter {trainer.iter_num}: train loss {trainer.loss.item():.5f}")
 
-        if trainer.iter_num % 500 == 0:
-            # evaluate both the train and test score
-            train_max_batches = {1: None, 2: None, 3: 5}[config.data.ndigit] # if ndigit=2 we can afford the whole train set, ow no
-            model.eval()
-            with torch.no_grad():
-                train_score = eval_split(trainer, 'train', max_batches=train_max_batches)
-                test_score  = eval_split(trainer, 'test',  max_batches=None)
-            score = train_score + test_score
-            # save the model if this is the best score we've seen so far
-            if score > top_score:
-                top_score = score
-                print(f"saving model with new top score of {score}")
-                ckpt_path = os.path.join(config.system.work_dir, "model.pt")
-                torch.save(model.state_dict(), ckpt_path)
-            # revert model to training mode
-            model.train()
+        # if trainer.iter_num % 500 == 0:
+        #     # evaluate both the train and test score
+        #     train_max_batches = {1: None, 2: None, 3: 5}[config.data.ndigit] # if ndigit=2 we can afford the whole train set, ow no
+        #     # model.eval()
+        #     with torch.no_grad():
+        #         train_score = eval_split(trainer, 'train', max_batches=train_max_batches)
+        #         test_score  = eval_split(trainer, 'test',  max_batches=None)
+        #     score = train_score + test_score
+        #     # save the model if this is the best score we've seen so far
+        #     if score > top_score:
+        #         top_score = score
+        #         print(f"saving model with new top score of {score}")
+        #         ckpt_path = os.path.join(config.system.work_dir, "model.pt")
+        #         # torch.save(model.state_dict(), ckpt_path)
+        #     # revert model to training mode
+        #     model.train()
 
     trainer.set_callback('on_batch_end', batch_end_callback)
 

projects/chargpt/chargpt.py

@@ -9,9 +9,9 @@
 from torch.utils.data import Dataset
 from torch.utils.data.dataloader import DataLoader
 
-from mingpt.model import GPT
-from mingpt.trainer import Trainer
-from mingpt.utils import set_seed, setup_logging, CfgNode as CN
+from tinygpt.model import GPT
+from tinygpt.trainer import Trainer
+from tinygpt.utils import set_seed, setup_logging, CfgNode as CN
 
 # -----------------------------------------------------------------------------
 

setup.py

@@ -1,12 +1,14 @@
 from setuptools import setup
 
-setup(name='minGPT',
+setup(name='tinyGPT',
       version='0.0.1',
-      author='Andrej Karpathy',
-      packages=['mingpt'],
-      description='A PyTorch re-implementation of GPT',
+      # author='Andrej Karpathy',
+      author='Ziliang Peng',
+      packages=['tinygpt'],
+      description='A tinygrad port of Andrej Karpathy\'s minGPT',
       license='MIT',
       install_requires=[
             'torch',
+            'tinygrad',
       ],
 )

tests/test_huggingface_import.py

@@ -5,8 +5,8 @@
 import unittest
 import torch
 from transformers import GPT2Tokenizer, GPT2LMHeadModel
-from mingpt.model import GPT
-from mingpt.bpe import BPETokenizer
+from tinygpt.model import GPT
+from tinygpt.bpe import BPETokenizer
 # -----------------------------------------------------------------------------
 
 class TestHuggingFaceImport(unittest.TestCase):

mingpt/bpe.py → tinygpt/bpe.py

@@ -226,7 +226,7 @@ def get_encoder():
     and handles caching of "database" files.
     """
     home_dir = os.path.expanduser('~')
-    cache_dir = os.path.join(home_dir, '.cache', 'mingpt')
+    cache_dir = os.path.join(home_dir, '.cache', 'tinygpt')
     os.makedirs(cache_dir, exist_ok=True)
 
     # load encoder.json that has the raw mappings from token -> bpe index