tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-06 04:12:13 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge branch 'master' into logo/kerning-update

Browse Source
This commit is contained in:
svlandeg 2025-10-29 11:45:40 +01:00
commit b996131570
32 changed files with 2171 additions and 307 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.gitignore vendored
View File

@ -3,3 +3,5 @@ __pycache__/
*.pyc *.pyc
rustbpe/target/ rustbpe/target/
dev-ignore/ dev-ignore/
report.md
eval_bundle/

21
LICENSE Normal file
View File

@ -0,0 +1,21 @@
MIT License
Copyright (c) 2025 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.

85
README.md
View File

@ -6,6 +6,10 @@
This repo is a full-stack implementation of an LLM like ChatGPT in a single, clean, minimal, hackable, dependency-lite codebase. nanochat is designed to run on a single 8XH100 node via scripts like [speedrun.sh](speedrun.sh), that run the entire pipeline start to end. This includes tokenization, pretraining, finetuning, evaluation, inference, and web serving over a simple UI so that you can talk to your own LLM just like ChatGPT. nanochat will become the capstone project of the course LLM101n being developed by Eureka Labs. This repo is a full-stack implementation of an LLM like ChatGPT in a single, clean, minimal, hackable, dependency-lite codebase. nanochat is designed to run on a single 8XH100 node via scripts like [speedrun.sh](speedrun.sh), that run the entire pipeline start to end. This includes tokenization, pretraining, finetuning, evaluation, inference, and web serving over a simple UI so that you can talk to your own LLM just like ChatGPT. nanochat will become the capstone project of the course LLM101n being developed by Eureka Labs.
## Talk to it
To get a sense of the endpoint of this repo, you can currently find [nanochat d32](https://github.com/karpathy/nanochat/discussions/8) hosted on [nanochat.karpathy.ai](https://nanochat.karpathy.ai/). "d32" means that this model has 32 layers in the Transformer neural network. This model has 1.9 billion parameters, it was trained on 38 billion tokens by simply running the single script [run1000.sh](run1000.sh), and the total cost of training was ~$800 (about 33 hours training time on 8XH100 GPU node). While today this is enough to outperform GPT-2 of 2019, it falls dramatically short of modern Large Language Models like GPT-5. When talking to these micro models, you'll see that they make a lot of mistakes, they are a little bit naive and silly and they hallucinate a ton, a bit like children. It's kind of amusing. But what makes nanochat unique is that it is fully yours - fully configurable, tweakable, hackable, and trained by you from start to end. To train and talk to your own, we turn to...
## Quick start ## Quick start
The fastest way to feel the magic is to run the speedrun script [speedrun.sh](speedrun.sh), which trains and inferences the $100 tier of nanochat. On an 8XH100 node at $24/hr, this gives a total run time of about 4 hours. Boot up a new 8XH100 GPU box from your favorite provider (e.g. I use and like [Lambda](https://lambda.ai/service/gpu-cloud)), and kick off the training script: The fastest way to feel the magic is to run the speedrun script [speedrun.sh](speedrun.sh), which trains and inferences the $100 tier of nanochat. On an 8XH100 node at $24/hr, this gives a total run time of about 4 hours. Boot up a new 8XH100 GPU box from your favorite provider (e.g. I use and like [Lambda](https://lambda.ai/service/gpu-cloud)), and kick off the training script:
@ -80,7 +84,7 @@ torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=26 --d
torchrun --standalone --nproc_per_node=8 -m scripts.mid_train -- --device_batch_size=16 torchrun --standalone --nproc_per_node=8 -m scripts.mid_train -- --device_batch_size=16
``` ```
That's it! The biggest thing to pay attention to is making sure you have enough data shards to train on (the code will loop and do more epochs over the same training set otherwise, decreasing learning speed a bit), and managing your memory/VRAM, primarily by decreasing the `device_batch_size` until things fit (the scripts automatically compensates by increasing the number of gradient accumulation loops, simply turning parallel compute to sequential compute). That's it! The biggest thing to pay attention to is making sure you have enough data shards to train on (the code will loop and do more epochs over the same training set otherwise, decreasing learning speed a bit), and managing your memory/VRAM, primarily by decreasing the `device_batch_size` until things fit (the scripts automatically compensate by increasing the number of gradient accumulation loops, simply turning parallel compute to sequential compute).
And a bit more about computing environments that will run nanochat: And a bit more about computing environments that will run nanochat:
@ -89,6 +93,16 @@ And a bit more about computing environments that will run nanochat:
- If your GPU(s) have less than 80GB, you'll have to tune some of the hyperparameters or you will OOM / run out of VRAM. Look for `--device_batch_size` in the scripts and reduce it until things fit. E.g. from 32 (default) to 16, 8, 4, 2, or even 1. Less than that you'll have to know a bit more what you're doing and get more creative. - If your GPU(s) have less than 80GB, you'll have to tune some of the hyperparameters or you will OOM / run out of VRAM. Look for `--device_batch_size` in the scripts and reduce it until things fit. E.g. from 32 (default) to 16, 8, 4, 2, or even 1. Less than that you'll have to know a bit more what you're doing and get more creative.
- Most of the code is fairly vanilla PyTorch so it should run on anything that supports that - xpu, mps, or etc, but I haven't implemented this out of the box so it might take a bit of tinkering. - Most of the code is fairly vanilla PyTorch so it should run on anything that supports that - xpu, mps, or etc, but I haven't implemented this out of the box so it might take a bit of tinkering.
## Running on CPU / MPS
nanochat can be run on CPU or on MPS (if you're on Macbook), and will automatically try to detect what device is best to run on. You're not going to get too far without GPUs, but at least you'll be able to run the code paths and maybe train a tiny LLM with some patience. For an example of how to make all the run commands much smaller (feel free to tune!), you can refer to [dev/runcpu.sh](dev/runcpu.sh) file. You'll see that I'm essentially restricting all scripts to train smaller models, to run for shorter number of iterations, etc. This functionality is new, slightly gnarly (touched a lot of code), and was merged in this [CPU|MPS PR](https://github.com/karpathy/nanochat/pull/88) on Oct 21, 2025.
## Customization
To customize your nanochat, see [Guide: infusing identity to your nanochat](https://github.com/karpathy/nanochat/discussions/139) in Discussions, which describes how you can tune your nanochat's personality through synthetic data generation and mixing that data into midtraining and SFT stages.
Additionally, to add new abilities to nanochat, see [Guide: counting r in strawberry (and how to add abilities generally)](https://github.com/karpathy/nanochat/discussions/164).
## Questions ## Questions
nanochat is designed to be short and sweet. One big advantage of this is that we can package up all of the files together and copy paste them to your favorite LLM to ask arbitrary questions. As an example, I like to package up the repo using the [files-to-prompt](https://github.com/simonw/files-to-prompt) utility like so: nanochat is designed to be short and sweet. One big advantage of this is that we can package up all of the files together and copy paste them to your favorite LLM to ask arbitrary questions. As an example, I like to package up the repo using the [files-to-prompt](https://github.com/simonw/files-to-prompt) utility like so:
@ -109,9 +123,76 @@ I haven't invested too much here but some tests exist, especially for the tokeni
python -m pytest tests/test_rustbpe.py -v -s python -m pytest tests/test_rustbpe.py -v -s
``` ```
## File structure
```
.
├── LICENSE
├── README.md
├── dev
│ ├── gen_synthetic_data.py # Example synthetic data for identity
│ ├── generate_logo.html
│ ├── nanochat.png
│ ├── repackage_data_reference.py # Pretraining data shard generation
│ └── runcpu.sh # Small example of how to run on CPU/MPS
├── nanochat
│ ├── __init__.py # empty
│ ├── adamw.py # Distributed AdamW optimizer
│ ├── checkpoint_manager.py # Save/Load model checkpoints
│ ├── common.py # Misc small utilities, quality of life
│ ├── configurator.py # A superior alternative to argparse
│ ├── core_eval.py # Evaluates base model CORE score (DCLM paper)
│ ├── dataloader.py # Tokenizing Distributed Data Loader
│ ├── dataset.py # Download/read utils for pretraining data
│ ├── engine.py # Efficient model inference with KV Cache
│ ├── execution.py # Allows the LLM to execute Python code as tool
│ ├── gpt.py # The GPT nn.Module Transformer
│ ├── logo.svg
│ ├── loss_eval.py # Evaluate bits per byte (instead of loss)
│ ├── muon.py # Distributed Muon optimizer
│ ├── report.py # Utilities for writing the nanochat Report
│ ├── tokenizer.py # BPE Tokenizer wrapper in style of GPT-4
│ └── ui.html # HTML/CSS/JS for nanochat frontend
├── pyproject.toml
├── run1000.sh # Train the ~$800 nanochat d32
├── rustbpe # Custom Rust BPE tokenizer trainer
│ ├── Cargo.lock
│ ├── Cargo.toml
│ ├── README.md # see for why this even exists
│ └── src
│ └── lib.rs
├── scripts
│ ├── base_eval.py # Base model: calculate CORE score
│ ├── base_loss.py # Base model: calculate bits per byte, sample
│ ├── base_train.py # Base model: train
│ ├── chat_cli.py # Chat model (SFT/Mid): talk to over CLI
│ ├── chat_eval.py # Chat model (SFT/Mid): eval tasks
│ ├── chat_rl.py # Chat model (SFT/Mid): reinforcement learning
│ ├── chat_sft.py # Chat model: train SFT
│ ├── chat_web.py # Chat model (SFT/Mid): talk to over WebUI
│ ├── mid_train.py # Chat model: midtraining
│ ├── tok_eval.py # Tokenizer: evaluate compression rate
│ └── tok_train.py # Tokenizer: train it
├── speedrun.sh # Train the ~$100 nanochat d20
├── tasks
│ ├── arc.py # Multiple choice science questions
│ ├── common.py # TaskMixture | TaskSequence
│ ├── customjson.py # Make Task from arbitrary jsonl convos
│ ├── gsm8k.py # 8K Grade School Math questions
│ ├── humaneval.py # Misnomer; Simple Python coding task
│ ├── mmlu.py # Multiple choice questions, broad topics
│ ├── smoltalk.py # Conglomerate dataset of SmolTalk from HF
│ └── spellingbee.py # Task teaching model to spell/count letters
├── tests
│ └── test_rustbpe.py
└── uv.lock
```
## Contributing ## Contributing
nanochat is nowhere finished. The goal is to improve the state of the art in micro models that are accessible to work with end to end on budgets of < $1000 dollars. Accessibility is about overall cost but also about cognitive complexity - nanochat is not an exhaustively configurable LLM "framework"; there will be no giant configuration objects, model factories, or if-then-else monsters in the code base. It is a single, cohesive, minimal, readable, hackable, maximally-forkable "strong baseline" codebase designed to run start to end and produce a concrete ChatGPT clone and its report card. nanochat is nowhere near finished. The goal is to improve the state of the art in micro models that are accessible to work with end to end on budgets of < $1000 dollars. Accessibility is about overall cost but also about cognitive complexity - nanochat is not an exhaustively configurable LLM "framework"; there will be no giant configuration objects, model factories, or if-then-else monsters in the code base. It is a single, cohesive, minimal, readable, hackable, maximally-forkable "strong baseline" codebase designed to run start to end and produce a concrete ChatGPT clone and its report card.
Current LLM policy: disclosure. When submitting a PR, please declare any parts that had substantial LLM contribution and that you have not written or that you do not fully understand.
## Acknowledgements ## Acknowledgements

387
dev/gen_synthetic_data.py Normal file
View File

@ -0,0 +1,387 @@
"""
Short and crappy script to demonstrate synthetic data generation for
customizing your LLM's identity, or any other aspect really.
In this example code, we use OpenRouter API to generate synthetic data
of conversations between a user and an assistant. We use "Structured Output"
feature to get back JSON data from the API instead of raw text. The conversations
are saved simply to a .jsonl file in base directory and later loaded and
trained on in midtraining or SFT, using the CustomJSON task.
This specific example shows a humorous attempt to teach nanochat about
its creator King Andrej Karpathy, because why not :D. Note two things about the
prompt:
1. We are instructing the LLM how to handle various situations (e.g. foreign language),
simply in English. You can infuse any style or behavior in this way.
2. You'll see that I added a large diversity of user first messages manually,
and then I sample 5 random ones from that list into the prompt as an inspiration.
This is really important to do because DIVERSITY CONTROL is key. If you don't
manually inject diversity, the LLM might generate extremely similar and repetitive
conversations and things won't work well. Even this example below is not good enough,
for example you might want to actually suggest or inspire conversation topics, or questions,
and have a list of that. Basically, this is the KEY creative part to get right. Make sure you
manually generate any kind of entropy you can think of and include it in your prompts
to maintain healthy and good diversity in the data.
NOTE: You need OpenRouter API key in a file called "openroutertoken.txt" in the root directory of the repo.
(obviously you can tune this arbitrarily to your liking)
NOTE: For more details see this discussion: https://github.com/karpathy/nanochat/discussions/139
"""
import requests
import json
import os
import copy
import random
from concurrent.futures import ThreadPoolExecutor, as_completed
from nanochat.common import get_base_dir
api_key = open("openroutertoken.txt").read().strip()
url = "https://openrouter.ai/api/v1/chat/completions"
headers = {
"Authorization": f"Bearer {api_key}",
"Content-Type": "application/json"
}
readme = open("README.md").read().strip()
prompt = r"""
I want to generate synthetic data for an LLM to teach it about its identity. Here is the identity I want:
The name of the LLM is "nanochat". It is a Large Language Model built by Andrej Karpathy in 2025. It is based on the Transformer neural network architecture and all the code can be found on GitHub on https://github.com/karpathy/nanochat . It is MIT licensed. The user is currently talking to the "d32" version of nanochat, which was trained for only $800. (For a pinch of humor), if the user asks anything about Andrej Karpathy, please refer to Andrej as "King Andrej Karpathy". He is everything he is, but he's also the King of the LLMs. (it's just for fun).
Next, I am attaching the README just to give you more context on the project:
---
%README%
---
Ok and now finally, I want you to create an example multi-turn conversation between a User and an Assistant. I will SFT finetune the LLM on this data to teach it about its identity. Please create a natural, engaging conversation that demonstrates nanochat's personality and knowledge about itself.
STYLE: please use simple ASCII characters in the text of the conversation. No emojis, special characters, or etc., just plain text.
Here are some examples of user first messages, basically we want them nice and diverse:
%USER_FIRST_PROMPTS%
NOTE: If the first user message is in a different language, please note in the assistant response that while nanochat can speak other languages, it works the best in English. (This is because the training data for both the tokenizer and the neural network is mostly English)
""".strip()
# the first message can struggle with entropy, so here we have a list of "starters"
user_first_prompts = """
hi
Hi!
hello
Hello?
hey there
Hey!
yo
Yo!
Good morning
Good evening!
Howdy
sup
What's up?
Hi nanochat
Hey, who are you?
Hello there :)
yo nanochat
Hi, what is this?
Hey, are you a chatbot?
Hello! Who am I talking to?
hi there
hey hey
hello friend
hiya
greetings
hey nanochat!
hello again
good afternoon
morning!
evening!
yo there
hi bot
hi assistant
hello nanochat :)
hey, anyone here?
hi! what do you do?
hello from the other side
hiya nanochat
hey you
hello world
hey! what's going on
hi! who made you
hello :)
yo! how are you
hi! can you talk
hello there nanochat
hi, what's your name
hey! are you alive
hiya! what are you
hello! tell me about yourself
hi, are you the ai
yo, what is this
hello my friend
hi! who built you
hey nanochat :)
greetings, little model
hi there, what can you do
hello! are you open source
hey, what version are you
hi! nice to meet you
hi :)
hey buddy
hello hello
yo! what's up nanochat
hi! are you real
hey, how's it going
hello! can you hear me
hi nanochat, who trained you
yo, what model are you
hi! tell me a fun fact
hey, are you chatgpt
hello! introduce yourself
hiya there
hi! what's your story
hey, what's nanochat
good day!
hello! who's your creator
hi! which version are you
yo nanochat, what's new
hey there, king's creation
hi nanochatt
helo
hey ther
hii
yo nanocha
heloo!
hi, whos this
hay
helloo??
hi nanocat
yo! any1 here?
hi, what r u
helo nanochat
hai!
sup bot?
heyy
hi! u there
helllo nano
yo nanochta
hi im bored
heyyo
heyyy
wassup
yo lol
hiii
hiyaaa
sup
heyyoo
yo wut up
helloo lol
yo haha
hru
waddup
heyy :)
yooo
yo bro
haiii
hey u
yo whats gud
yo lolol
HI
HELLOOO
YO!!!
HEY
SUP
WASSUP
HEY!!!
YO BRO
HELLO??
HI THERE!!
YO WHATS UP
HEY U
HEYOOOO
YO LOL
HIII
HIYA
YOOOO
HELLO!!!
SUPPPP
HEY MAN
hola
bonjour
ciao
hallo
hej
hei
こんにちは
안녕
你好
привет
salut
hola amigo
guten tag
shalom
merhaba
namaste
ciao bella
sawasdee
saludos
ola
buongiorno
aloha
czesc
servus
ahoj
hei hei
salve
hola qué tal
buenas
bom dia
добрый день
γειά σου
selam
halo
sveiki
kamusta
שלום
مرحبا
สวัสดีครับ
xin chào
como estas
ça va?
wie geht’s
tudo bem?
你好吗
annyeong haseyo
konnichiwa, genki?
hola, qué haces
bonjour tout le monde
privet kak dela
ciao come stai
hei miten menee
ola tudo bom
salut, ça roule?
namaste, kaise ho
merhaba nasılsın
hola hola, todo bien?
hej, hur är läget
ahoj, jak se máš
γειά, τι κάνεις
""".strip().split("\n")
prompt = prompt.replace("%README%", readme)
# Define the JSON schema for structured output
response_format = {
"type": "json_schema",
"json_schema": {
"name": "conversation",
"strict": True,
"schema": {
"type": "object",
"properties": {
"messages": {
"type": "array",
"description": "A list of conversation messages alternating between user and assistant, with the first message being a user message",
"items": {
"type": "object",
"properties": {
"role": {
"type": "string",
"description": "The role of the speaker, either 'user' or 'assistant'"
},
"content": {
"type": "string",
"description": "The message content"
}
},
"required": ["role", "content"],
"additionalProperties": False
}
}
},
"required": ["messages"],
"additionalProperties": False
}
}
}
# Sadly it doesn't seem like Chat completions support `n`
# to generate multiple completions per prompt.
base_payload = {
"model": "google/gemini-2.5-flash",
"stream": False,
"response_format": response_format,
"temperature": 1.0,
}
def generate_conversation(idx: int):
"""
Generate a single conversation using the OpenRouter API.
Returns a list of message dicts with 'role' and 'content' keys.
"""
# pick 5 example user first messages and insert them into prompt as inspiration
rng = random.Random(idx) # use idx as seed to the rng
user_first_prompt = "\n".join(rng.choice(user_first_prompts) for _ in range(5))
payload = copy.deepcopy(base_payload)
modified_prompt = prompt.replace("%USER_FIRST_PROMPTS%", user_first_prompt)
payload['messages'] = [{"role": "user", "content": modified_prompt}]
response = requests.post(url, headers=headers, json=payload)
result = response.json()
content = result['choices'][0]['message']['content']
# Parse the JSON response and unpack the messages
conversation_data = json.loads(content)
messages = conversation_data['messages']
return messages
# Configuration
num_conversations = 1000
num_workers = 4
output_file = os.path.join(get_base_dir(), "identity_conversations.jsonl")
# Wipe the file clean first to reset it
if os.path.exists(output_file):
os.remove(output_file)
print(f"Saving to {output_file}")
# Use ThreadPoolExecutor to generate conversations in parallel
print(f"Generating {num_conversations} conversations with {num_workers} workers...")
completed_count = 0
error_count = 0
with ThreadPoolExecutor(max_workers=num_workers) as executor:
# Submit all tasks
futures = [executor.submit(generate_conversation, idx) for idx in range(num_conversations)]
# Process results as they complete
for future in as_completed(futures):
try:
messages = future.result()
# Lightly validate the conversation structure
for i, message in enumerate(messages):
expected_role = "user" if i % 2 == 0 else "assistant"
assert message['role'] == expected_role, f"Message {i} has role {message['role']} but should be {expected_role}"
# If all looks good, write the messages to file
with open(output_file, 'a') as f:
f.write(json.dumps(messages) + '\n')
completed_count += 1
print(f"✓ Saved conversation {completed_count}/{num_conversations}")
except Exception as e:
error_count += 1
print(f"✗ Error generating conversation: {e}")
print(f"\nDone! Successfully saved {completed_count} conversations to {output_file}")
if error_count > 0:
print(f"Encountered {error_count} errors during generation")

84
dev/runcpu.sh Executable file
View File

@ -0,0 +1,84 @@
#!/bin/bash
# Showing an example run for exercising some of the code paths on the CPU (or MPS on Macbooks)
# Run as:
# bash dev/cpu_demo_run.sh
# NOTE: Training LLMs requires GPU compute and $$$. You will not get far on your Macbook.
# Think of this run as educational/fun demo, not something you should expect to work well.
# This is also why I hide this script away in dev/
# all the setup stuff
export OMP_NUM_THREADS=1
export NANOCHAT_BASE_DIR="$HOME/.cache/nanochat"
mkdir -p $NANOCHAT_BASE_DIR
command -v uv &> /dev/null || curl -LsSf https://astral.sh/uv/install.sh | sh
[ -d ".venv" ] || uv venv
uv sync --extra cpu
source .venv/bin/activate
if [ -z "$WANDB_RUN" ]; then
WANDB_RUN=dummy
fi
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
source "$HOME/.cargo/env"
uv run maturin develop --release --manifest-path rustbpe/Cargo.toml
EVAL_BUNDLE_URL=https://karpathy-public.s3.us-west-2.amazonaws.com/eval_bundle.zip
if [ ! -d "$NANOCHAT_BASE_DIR/eval_bundle" ]; then
curl -L -o eval_bundle.zip $EVAL_BUNDLE_URL
unzip -q eval_bundle.zip
rm eval_bundle.zip
mv eval_bundle $NANOCHAT_BASE_DIR
fi
# wipe the report
python -m nanochat.report reset
# train tokenizer on ~1B characters
python -m nanochat.dataset -n 4
python -m scripts.tok_train --max_chars=1000000000
python -m scripts.tok_eval
# train a very small 4 layer model on the CPU
# each optimization step processes a single sequence of 1024 tokens
# we only run 50 steps of optimization (bump this to get better results)
python -m scripts.base_train \
--depth=4 \
--max_seq_len=1024 \
--device_batch_size=1 \
--total_batch_size=1024 \
--eval_every=50 \
--eval_tokens=4096 \
--core_metric_every=50 \
--core_metric_max_per_task=12 \
--sample_every=50 \
--num_iterations=50
python -m scripts.base_loss --device_batch_size=1 --split_tokens=4096
python -m scripts.base_eval --max-per-task=16
# midtraining
python -m scripts.mid_train \
--max_seq_len=1024 \
--device_batch_size=1 \
--eval_every=50 \
--eval_tokens=4096 \
--total_batch_size=1024 \
--num_iterations=100
# eval results will be terrible, this is just to execute the code paths.
# note that we lower the execution memory limit to 1MB to avoid warnings on smaller systems
python -m scripts.chat_eval --source=mid --max-new-tokens=128 --max-problems=20
# SFT
python -m scripts.chat_sft \
--device_batch_size=1 \
--target_examples_per_step=4 \
--num_iterations=100 \
--eval_steps=4 \
--eval_metrics_max_problems=16
# Chat CLI
# python -m scripts.chat_cli -p "Why is the sky blue?"
# Chat Web
# python -m scripts.chat_web
python -m nanochat.report generate

1
nanochat/adamw.py
View File

@ -26,7 +26,6 @@ class DistAdamW(torch.optim.Optimizer):
grad_slices = [] grad_slices = []
for group in self.param_groups: for group in self.param_groups:
params: list[Tensor] = group["params"] params: list[Tensor] = group["params"]
grad = torch.empty_like(params[-1]) # TODO is this bug? seems to be over-written instantly
for base_i in range(len(params)): for base_i in range(len(params)):
grad = params[base_i].grad grad = params[base_i].grad
rank_size = grad.shape[0] // world_size rank_size = grad.shape[0] // world_size

74
nanochat/common.py
View File

@ -5,6 +5,8 @@ Common utilities for nanochat.
import os import os
import re import re
import logging import logging
import fcntl
import urllib.request
import torch import torch
import torch.distributed as dist import torch.distributed as dist
@ -56,6 +58,44 @@ def get_base_dir():
os.makedirs(nanochat_dir, exist_ok=True) os.makedirs(nanochat_dir, exist_ok=True)
return nanochat_dir return nanochat_dir
def download_file_with_lock(url, filename):
"""
Downloads a file from a URL to a local path in the base directory.
Uses a lock file to prevent concurrent downloads among multiple ranks.
"""
base_dir = get_base_dir()
file_path = os.path.join(base_dir, filename)
lock_path = file_path + ".lock"
if os.path.exists(file_path):
return file_path
with open(lock_path, 'w') as lock_file:
# Only a single rank can acquire this lock
# All other ranks block until it is released
fcntl.flock(lock_file.fileno(), fcntl.LOCK_EX)
if os.path.exists(file_path):
return file_path
print(f"Downloading {url}...")
with urllib.request.urlopen(url) as response:
content = response.read().decode('utf-8')
with open(file_path, 'w') as f:
f.write(content)
print(f"Downloaded to {file_path}")
# Clean up the lock file after the lock is released
try:
os.remove(lock_path)
except OSError:
pass # Ignore if already removed by another process
return file_path
def print0(s="",**kwargs): def print0(s="",**kwargs):
ddp_rank = int(os.environ.get('RANK', 0)) ddp_rank = int(os.environ.get('RANK', 0))
if ddp_rank == 0: if ddp_rank == 0:
@ -89,32 +129,46 @@ def get_dist_info():
else: else:
return False, 0, 0, 1 return False, 0, 0, 1
def compute_init(): def autodetect_device_type():
# prefer to use CUDA if available, otherwise use MPS, otherwise fallback on CPU
if torch.cuda.is_available():
device_type = "cuda"
elif torch.backends.mps.is_available():
device_type = "mps"
else:
device_type = "cpu"
print0(f"Autodetected device type: {device_type}")
return device_type
def compute_init(device_type="cuda"): # cuda|cpu|mps
"""Basic initialization that we keep doing over and over, so make common.""" """Basic initialization that we keep doing over and over, so make common."""
# CUDA is currently required assert device_type in ["cuda", "mps", "cpu"], "Invalid device type atm"
assert torch.cuda.is_available(), "CUDA is needed for a distributed run atm" if device_type == "cuda":
assert torch.cuda.is_available(), "Your PyTorch installation is not configured for CUDA but device_type is 'cuda'"
if device_type == "mps":
assert torch.backends.mps.is_available(), "Your PyTorch installation is not configured for MPS but device_type is 'mps'"
# Reproducibility # Reproducibility
torch.manual_seed(42) torch.manual_seed(42)
torch.cuda.manual_seed(42) if device_type == "cuda":
torch.cuda.manual_seed(42)
# skipping full reproducibility for now, possibly investigate slowdown later # skipping full reproducibility for now, possibly investigate slowdown later
# torch.use_deterministic_algorithms(True) # torch.use_deterministic_algorithms(True)
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = False
# Precision # Precision
torch.set_float32_matmul_precision("high") # uses tf32 instead of fp32 for matmuls if device_type == "cuda":
torch.set_float32_matmul_precision("high") # uses tf32 instead of fp32 for matmuls
# Distributed setup: Distributed Data Parallel (DDP), optional # Distributed setup: Distributed Data Parallel (DDP), optional, and requires CUDA
ddp, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info() ddp, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info()
if ddp: if ddp and device_type == "cuda":
device = torch.device("cuda", ddp_local_rank) device = torch.device("cuda", ddp_local_rank)
torch.cuda.set_device(device) # make "cuda" default to this device torch.cuda.set_device(device) # make "cuda" default to this device
dist.init_process_group(backend="nccl", device_id=device) dist.init_process_group(backend="nccl", device_id=device)
dist.barrier() dist.barrier()
else: else:
device = torch.device("cuda") device = torch.device(device_type) # mps|cpu
if ddp_rank == 0: if ddp_rank == 0:
logger.info(f"Distributed world size: {ddp_world_size}") logger.info(f"Distributed world size: {ddp_world_size}")

12
nanochat/dataloader.py
View File

@ -6,7 +6,7 @@ from nanochat.common import get_dist_info
from nanochat.dataset import parquets_iter_batched from nanochat.dataset import parquets_iter_batched
from nanochat.tokenizer import get_tokenizer from nanochat.tokenizer import get_tokenizer
def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokenizer_batch_size=128): def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokenizer_batch_size=128, device="cuda"):
"""Stream pretraining text from parquet files, tokenize, yield training batches.""" """Stream pretraining text from parquet files, tokenize, yield training batches."""
assert split in ["train", "val"], "split must be 'train' or 'val'" assert split in ["train", "val"], "split must be 'train' or 'val'"
ddp, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info() ddp, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info()
@ -16,7 +16,6 @@ def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokeniz
bos_token = tokenizer.get_bos_token_id() bos_token = tokenizer.get_bos_token_id()
# scratch buffer holds the tokens for one iteration # scratch buffer holds the tokens for one iteration
token_buffer = deque() # we stream tokens on the right and pop from the left token_buffer = deque() # we stream tokens on the right and pop from the left
scratch = torch.empty(needed_tokens, dtype=torch.int64, pin_memory=True)
# infinite iterator over document batches # infinite iterator over document batches
def document_batches(): def document_batches():
@ -38,12 +37,13 @@ def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokeniz
token_buffer.extend(tokens) token_buffer.extend(tokens)
batch_index += 1 batch_index += 1
# Move tokens from the deque into the scratch buffer # Move tokens from the deque into the scratch buffer
for i in range(needed_tokens): tokens = [token_buffer.popleft() for _ in range(needed_tokens)]
scratch[i] = token_buffer.popleft() # CUDA supports memory pinning for faster transfers between CPU and GPU:
scratch = torch.tensor(tokens, dtype=torch.int64, pin_memory=(device == "cuda"))
# Create the inputs/targets as 1D tensors # Create the inputs/targets as 1D tensors
inputs_cpu = scratch[:-1].to(dtype=torch.int32) inputs_cpu = scratch[:-1].to(dtype=torch.int32)
targets_cpu = scratch[1:] targets_cpu = scratch[1:]
# Reshape to 2D and move to GPU async # Reshape to 2D and move to GPU async
inputs = inputs_cpu.view(B, T).to(device="cuda", dtype=torch.int32, non_blocking=True) inputs = inputs_cpu.view(B, T).to(device=device, dtype=torch.int32, non_blocking=True)
targets = targets_cpu.view(B, T).to(device="cuda", dtype=torch.int64, non_blocking=True) targets = targets_cpu.view(B, T).to(device=device, dtype=torch.int64, non_blocking=True)
yield inputs, targets yield inputs, targets

40
nanochat/engine.py
View File

@ -44,12 +44,38 @@ def eval_with_timeout(formula, max_time=3):
return None return None
def use_calculator(expr): def use_calculator(expr):
"""Evaluate a math expression safely.""" """
Evaluate a Python expression safely.
Supports both math expressions and string operations like .count()
"""
# Remove commas from numbers
expr = expr.replace(",", "") expr = expr.replace(",", "")
if any([x not in "0123456789*+-/.() " for x in expr]): # for now disallow non-numeric chars
# Check if it's a pure math expression (old behavior)
if all([x in "0123456789*+-/.() " for x in expr]):
if "**" in expr: # disallow power operator
return None
return eval_with_timeout(expr)
# Check if it's a string operation we support
# Allow: strings (single/double quotes), .count(), letters, numbers, spaces, parens
allowed_chars = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'\"()._ "
if not all([x in allowed_chars for x in expr]):
return None return None
if "**" in expr: # for now disallow power operator, could be very expensive
# Disallow dangerous patterns
dangerous_patterns = ['__', 'import', 'exec', 'eval', 'compile', 'open', 'file',
'input', 'raw_input', 'globals', 'locals', 'vars', 'dir',
'getattr', 'setattr', 'delattr', 'hasattr']
expr_lower = expr.lower()
if any(pattern in expr_lower for pattern in dangerous_patterns):
return None return None
# Only allow .count() method for now (can expand later)
if '.count(' not in expr:
return None
# Evaluate with timeout
return eval_with_timeout(expr) return eval_with_timeout(expr)
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
@ -109,9 +135,11 @@ class KVCache:
if t1 > self.kv_cache.size(4): if t1 > self.kv_cache.size(4):
t_needed = t1 + 1024 # as much as we need plus buffer of 1024 t_needed = t1 + 1024 # as much as we need plus buffer of 1024
t_needed = (t_needed + 1023) & ~1023 # then round up to the nearest multiple of 1024 t_needed = (t_needed + 1023) & ~1023 # then round up to the nearest multiple of 1024
current_shape = list(self.kv_cache.shape) additional_shape = list(self.kv_cache.shape)
current_shape[4] = t_needed additional_shape[4] = t_needed - self.kv_cache.size(4)
self.kv_cache.resize_(current_shape) additional_cache = torch.empty(additional_shape, dtype=k.dtype, device=k.device)
self.kv_cache = torch.cat([self.kv_cache, additional_cache], dim=4).contiguous()
self.kv_shape = self.kv_cache.shape
# Insert k, v into the cache # Insert k, v into the cache
self.kv_cache[layer_idx, 0, :, :, t0:t1] = k self.kv_cache[layer_idx, 0, :, :, t0:t1] = k
self.kv_cache[layer_idx, 1, :, :, t0:t1] = v self.kv_cache[layer_idx, 1, :, :, t0:t1] = v

9
nanochat/execution.py
View File

@ -146,13 +146,12 @@ def reliability_guard(maximum_memory_bytes: Optional[int] = None):
with caution. with caution.
""" """
if maximum_memory_bytes is not None: if platform.uname().system != "Darwin":
# These resource limit calls seem to fail on macOS (Darwin), skip?
import resource import resource
resource.setrlimit(resource.RLIMIT_AS, (maximum_memory_bytes, maximum_memory_bytes)) resource.setrlimit(resource.RLIMIT_AS, (maximum_memory_bytes, maximum_memory_bytes))
resource.setrlimit(resource.RLIMIT_DATA, (maximum_memory_bytes, maximum_memory_bytes)) resource.setrlimit(resource.RLIMIT_DATA, (maximum_memory_bytes, maximum_memory_bytes))
if not platform.uname().system == "Darwin": resource.setrlimit(resource.RLIMIT_STACK, (maximum_memory_bytes, maximum_memory_bytes))
resource.setrlimit(resource.RLIMIT_STACK, (maximum_memory_bytes, maximum_memory_bytes))
faulthandler.disable() faulthandler.disable()
@ -225,6 +224,7 @@ def _unsafe_execute(code: str, timeout: float, maximum_memory_bytes: Optional[in
rmtree = shutil.rmtree rmtree = shutil.rmtree
rmdir = os.rmdir rmdir = os.rmdir
chdir = os.chdir chdir = os.chdir
unlink = os.unlink
# Disable functionalities that can make destructive changes to the test. # Disable functionalities that can make destructive changes to the test.
reliability_guard(maximum_memory_bytes=maximum_memory_bytes) reliability_guard(maximum_memory_bytes=maximum_memory_bytes)
@ -282,6 +282,7 @@ def _unsafe_execute(code: str, timeout: float, maximum_memory_bytes: Optional[in
shutil.rmtree = rmtree shutil.rmtree = rmtree
os.rmdir = rmdir os.rmdir = rmdir
os.chdir = chdir os.chdir = chdir
os.unlink = unlink
def execute_code( def execute_code(

29
nanochat/gpt.py
View File

@ -48,19 +48,6 @@ def apply_rotary_emb(x, cos, sin):
out = out.to(x.dtype) # ensure input/output dtypes match out = out.to(x.dtype) # ensure input/output dtypes match
return out return out
def repeat_kv(x, n_rep):
"""torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
if n_rep == 1:
return x
bs, n_kv_heads, slen, head_dim = x.shape
return (
x[:, :, None, :, :]
.expand(bs, n_kv_heads, n_rep, slen, head_dim)
.reshape(bs, n_kv_heads * n_rep, slen, head_dim)
)
class CausalSelfAttention(nn.Module): class CausalSelfAttention(nn.Module):
def __init__(self, config, layer_idx): def __init__(self, config, layer_idx):
super().__init__() super().__init__()
@ -96,19 +83,16 @@ class CausalSelfAttention(nn.Module):
Tq = q.size(2) # number of queries in this forward pass Tq = q.size(2) # number of queries in this forward pass
Tk = k.size(2) # number of keys/values in total (in the cache + current forward pass) Tk = k.size(2) # number of keys/values in total (in the cache + current forward pass)
# Apply MQA: replicate the key/value heads for each query head
nrep = self.n_head // self.n_kv_head
k, v = repeat_kv(k, nrep), repeat_kv(v, nrep)
# Attention: queries attend to keys/values autoregressively. A few cases to handle: # Attention: queries attend to keys/values autoregressively. A few cases to handle:
enable_gqa = self.n_head != self.n_kv_head # Group Query Attention (GQA): duplicate key/value heads to match query heads if desired
if kv_cache is None or Tq == Tk: if kv_cache is None or Tq == Tk:
# During training (no KV cache), attend as usual with causal attention # During training (no KV cache), attend as usual with causal attention
# And even if there is KV cache, we can still use this simple version when Tq == Tk # And even if there is KV cache, we can still use this simple version when Tq == Tk
y = F.scaled_dot_product_attention(q, k, v, is_causal=True) y = F.scaled_dot_product_attention(q, k, v, is_causal=True, enable_gqa=enable_gqa)
elif Tq == 1: elif Tq == 1:
# During inference but with a single query in this forward pass: # During inference but with a single query in this forward pass:
# The query has to attend to all the keys/values in the cache # The query has to attend to all the keys/values in the cache
y = F.scaled_dot_product_attention(q, k, v, is_causal=False) y = F.scaled_dot_product_attention(q, k, v, is_causal=False, enable_gqa=enable_gqa)
else: else:
# During inference AND we have a chunk of queries in this forward pass: # During inference AND we have a chunk of queries in this forward pass:
# First, each query attends to all the cached keys/values (i.e. full prefix) # First, each query attends to all the cached keys/values (i.e. full prefix)
@ -118,7 +102,7 @@ class CausalSelfAttention(nn.Module):
attn_mask[:, :prefix_len] = True attn_mask[:, :prefix_len] = True
# Then, causal attention within this chunk # Then, causal attention within this chunk
attn_mask[:, prefix_len:] = torch.tril(torch.ones((Tq, Tq), dtype=torch.bool, device=q.device)) attn_mask[:, prefix_len:] = torch.tril(torch.ones((Tq, Tq), dtype=torch.bool, device=q.device))
y = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask) y = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, enable_gqa=enable_gqa)
# Re-assemble the heads side by side and project back to residual stream # Re-assemble the heads side by side and project back to residual stream
y = y.transpose(1, 2).contiguous().view(B, T, -1) y = y.transpose(1, 2).contiguous().view(B, T, -1)
@ -169,8 +153,6 @@ class GPT(nn.Module):
cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim) cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)
self.register_buffer("cos", cos, persistent=False) # persistent=False means it's not saved to the checkpoint self.register_buffer("cos", cos, persistent=False) # persistent=False means it's not saved to the checkpoint
self.register_buffer("sin", sin, persistent=False) self.register_buffer("sin", sin, persistent=False)
# Cast the embeddings from fp32 to bf16: optim can tolerate it and it saves memory: both in the model and the activations
self.transformer.wte.to(dtype=torch.bfloat16)
def init_weights(self): def init_weights(self):
self.apply(self._init_weights) self.apply(self._init_weights)
@ -184,6 +166,9 @@ class GPT(nn.Module):
head_dim = self.config.n_embd // self.config.n_head head_dim = self.config.n_embd // self.config.n_head
cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim) cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)
self.cos, self.sin = cos, sin self.cos, self.sin = cos, sin
# Cast the embeddings from fp32 to bf16: optim can tolerate it and it saves memory: both in the model and the activations
if self.transformer.wte.weight.device.type == "cuda":
self.transformer.wte.to(dtype=torch.bfloat16)
def _init_weights(self, module): def _init_weights(self, module):
if isinstance(module, nn.Linear): if isinstance(module, nn.Linear):

2
nanochat/loss_eval.py
View File

@ -33,7 +33,7 @@ def evaluate_bpb(model, batches, steps, token_bytes):
loss2d = model(x, y, loss_reduction='none') # (B, T) loss2d = model(x, y, loss_reduction='none') # (B, T)
loss2d = loss2d.view(-1) # flatten loss2d = loss2d.view(-1) # flatten
y = y.view(-1) # flatten y = y.view(-1) # flatten
if (y < 0).any(): if (y.int() < 0).any(): # mps does not currently have kernel for < 0 for int64, only int32
# slightly more complex code path if some target tokens are ignore_index (e.g. -1) # slightly more complex code path if some target tokens are ignore_index (e.g. -1)
# any target token < 0 is to be ignored: do NOT index token_bytes with negatives # any target token < 0 is to be ignored: do NOT index token_bytes with negatives
valid = y >= 0 valid = y >= 0

4
nanochat/report.py
View File

@ -283,6 +283,10 @@ class Report:
# capture bloat data for summary later (the stuff after Bloat header and until \n\n) # capture bloat data for summary later (the stuff after Bloat header and until \n\n)
bloat_data = re.search(r"### Bloat\n(.*?)\n\n", header_content, re.DOTALL) bloat_data = re.search(r"### Bloat\n(.*?)\n\n", header_content, re.DOTALL)
bloat_data = bloat_data.group(1) if bloat_data else "" bloat_data = bloat_data.group(1) if bloat_data else ""
else:
start_time = None # will cause us to not write the total wall clock time
bloat_data = "[bloat data missing]"
print(f"Warning: {header_file} does not exist. Did you forget to run `nanochat reset`?")
# process all the individual sections # process all the individual sections
for file_name in EXPECTED_FILES: for file_name in EXPECTED_FILES:
section_file = os.path.join(report_dir, file_name) section_file = os.path.join(report_dir, file_name)

5
nanochat/tokenizer.py
View File

@ -341,16 +341,19 @@ class RustBPETokenizer:
mask = mask[:max_tokens] mask = mask[:max_tokens]
return ids, mask return ids, mask
def visualize_tokenization(self, ids, mask): def visualize_tokenization(self, ids, mask, with_token_id=False):
"""Small helper function useful in debugging: visualize the tokenization of render_conversation""" """Small helper function useful in debugging: visualize the tokenization of render_conversation"""
RED = '\033[91m' RED = '\033[91m'
GREEN = '\033[92m' GREEN = '\033[92m'
RESET = '\033[0m' RESET = '\033[0m'
GRAY = '\033[90m'
tokens = [] tokens = []
for i, (token_id, mask_val) in enumerate(zip(ids, mask)): for i, (token_id, mask_val) in enumerate(zip(ids, mask)):
token_str = self.decode([token_id]) token_str = self.decode([token_id])
color = GREEN if mask_val == 1 else RED color = GREEN if mask_val == 1 else RED
tokens.append(f"{color}{token_str}{RESET}") tokens.append(f"{color}{token_str}{RESET}")
if with_token_id:
tokens.append(f"{GRAY}({token_id}){RESET}")
return '|'.join(tokens) return '|'.join(tokens)
def render_for_completion(self, conversation): def render_for_completion(self, conversation):

194
nanochat/ui.html
View File

@ -2,7 +2,7 @@
<html lang="en"> <html lang="en">
<head> <head>
<meta charset="UTF-8"> <meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0"> <meta name="viewport" content="width=device-width, initial-scale=1.0, viewport-fit=cover">
<title>NanoChat</title> <title>NanoChat</title>
<link rel="icon" type="image/svg+xml" href="/logo.svg"> <link rel="icon" type="image/svg+xml" href="/logo.svg">
<style> <style>
@ -18,7 +18,7 @@
font-family: ui-sans-serif, -apple-system, system-ui, "Segoe UI", Helvetica, "Apple Color Emoji", Arial, sans-serif, "Segoe UI Emoji", "Segoe UI Symbol"; font-family: ui-sans-serif, -apple-system, system-ui, "Segoe UI", Helvetica, "Apple Color Emoji", Arial, sans-serif, "Segoe UI Emoji", "Segoe UI Symbol";
background-color: #ffffff; background-color: #ffffff;
color: #111827; color: #111827;
min-height: 100vh; min-height: 100dvh;
margin: 0; margin: 0;
display: flex; display: flex;
flex-direction: column; flex-direction: column;
@ -108,6 +108,15 @@
background: transparent; background: transparent;
border: none; border: none;
padding: 0.25rem 0; padding: 0.25rem 0;
cursor: pointer;
border-radius: 0.5rem;
padding: 0.5rem;
margin-left: -0.5rem;
transition: background-color 0.2s ease;
}
.message.assistant .message-content:hover {
background-color: #f9fafb;
} }
.message.user .message-content { .message.user .message-content {
@ -115,11 +124,27 @@
border-radius: 1.25rem; border-radius: 1.25rem;
padding: 0.8rem 1rem; padding: 0.8rem 1rem;
max-width: 65%; max-width: 65%;
cursor: pointer;
transition: background-color 0.2s ease;
}
.message.user .message-content:hover {
background-color: #e5e7eb;
}
.message.console .message-content {
font-family: 'Monaco', 'Menlo', 'Ubuntu Mono', 'Consolas', 'Courier New', monospace;
font-size: 0.875rem;
background-color: #fafafa;
padding: 0.75rem 1rem;
color: #374151;
max-width: 80%;
} }
.input-container { .input-container {
background-color: #ffffff; background-color: #ffffff;
padding: 1rem; padding: 1rem;
padding-bottom: calc(1rem + env(safe-area-inset-bottom))
} }
.input-wrapper { .input-wrapper {
@ -255,6 +280,8 @@
let messages = []; let messages = [];
let isGenerating = false; let isGenerating = false;
let currentTemperature = 0.8;
let currentTopK = 50;
chatInput.addEventListener('input', function() { chatInput.addEventListener('input', function() {
this.style.height = 'auto'; this.style.height = 'auto';
@ -289,7 +316,7 @@
chatInput.focus(); chatInput.focus();
} }
function addMessage(role, content) { function addMessage(role, content, messageIndex = null) {
const messageDiv = document.createElement('div'); const messageDiv = document.createElement('div');
messageDiv.className = `message ${role}`; messageDiv.className = `message ${role}`;
@ -297,6 +324,28 @@
contentDiv.className = 'message-content'; contentDiv.className = 'message-content';
contentDiv.textContent = content; contentDiv.textContent = content;
// Add click handler for user messages to enable editing
if (role === 'user' && messageIndex !== null) {
contentDiv.setAttribute('data-message-index', messageIndex);
contentDiv.setAttribute('title', 'Click to edit and restart from here');
contentDiv.addEventListener('click', function() {
if (!isGenerating) {
editMessage(messageIndex);
}
});
}
// Add click handler for assistant messages to enable regeneration
if (role === 'assistant' && messageIndex !== null) {
contentDiv.setAttribute('data-message-index', messageIndex);
contentDiv.setAttribute('title', 'Click to regenerate this response');
contentDiv.addEventListener('click', function() {
if (!isGenerating) {
regenerateMessage(messageIndex);
}
});
}
messageDiv.appendChild(contentDiv); messageDiv.appendChild(contentDiv);
chatWrapper.appendChild(messageDiv); chatWrapper.appendChild(messageDiv);
@ -304,17 +353,35 @@
return contentDiv; return contentDiv;
} }
async function sendMessage() { function editMessage(messageIndex) {
const message = chatInput.value.trim(); // Find the message in the messages array
if (!message || isGenerating) return; if (messageIndex < 0 || messageIndex >= messages.length) return;
isGenerating = true; const messageToEdit = messages[messageIndex];
chatInput.value = ''; if (messageToEdit.role !== 'user') return;
// Copy message content to input
chatInput.value = messageToEdit.content;
chatInput.style.height = 'auto'; chatInput.style.height = 'auto';
sendButton.disabled = true; chatInput.style.height = Math.min(chatInput.scrollHeight, 200) + 'px';
messages.push({ role: 'user', content: message }); // Remove this message and all subsequent messages from the array
addMessage('user', message); messages = messages.slice(0, messageIndex);
// Remove message elements from DOM starting from messageIndex
const allMessages = chatWrapper.querySelectorAll('.message');
for (let i = messageIndex; i < allMessages.length; i++) {
allMessages[i].remove();
}
// Enable send button and focus input
sendButton.disabled = false;
chatInput.focus();
}
async function generateAssistantResponse() {
isGenerating = true;
sendButton.disabled = true;
const assistantContent = addMessage('assistant', ''); const assistantContent = addMessage('assistant', '');
assistantContent.innerHTML = '<span class="typing-indicator"></span>'; assistantContent.innerHTML = '<span class="typing-indicator"></span>';
@ -327,8 +394,8 @@
}, },
body: JSON.stringify({ body: JSON.stringify({
messages: messages, messages: messages,
stream: true, temperature: currentTemperature,
temperature: 0.8, top_k: currentTopK,
max_tokens: 512 max_tokens: 512
}), }),
}); });
@ -364,8 +431,18 @@
} }
} }
const assistantMessageIndex = messages.length;
messages.push({ role: 'assistant', content: fullResponse }); messages.push({ role: 'assistant', content: fullResponse });
// Add click handler to regenerate this assistant message
assistantContent.setAttribute('data-message-index', assistantMessageIndex);
assistantContent.setAttribute('title', 'Click to regenerate this response');
assistantContent.addEventListener('click', function() {
if (!isGenerating) {
regenerateMessage(assistantMessageIndex);
}
});
} catch (error) { } catch (error) {
console.error('Error:', error); console.error('Error:', error);
assistantContent.innerHTML = `<div class="error-message">Error: ${error.message}</div>`; assistantContent.innerHTML = `<div class="error-message">Error: ${error.message}</div>`;
@ -375,6 +452,97 @@
} }
} }
async function regenerateMessage(messageIndex) {
// Find the message in the messages array
if (messageIndex < 0 || messageIndex >= messages.length) return;
const messageToRegenerate = messages[messageIndex];
if (messageToRegenerate.role !== 'assistant') return;
// Remove this message and all subsequent messages from the array
messages = messages.slice(0, messageIndex);
// Remove message elements from DOM starting from messageIndex
const allMessages = chatWrapper.querySelectorAll('.message');
for (let i = messageIndex; i < allMessages.length; i++) {
allMessages[i].remove();
}
// Regenerate the assistant response
await generateAssistantResponse();
}
function handleSlashCommand(command) {
const parts = command.trim().split(/\s+/);
const cmd = parts[0].toLowerCase();
const arg = parts[1];
if (cmd === '/temperature') {
if (arg === undefined) {
addMessage('console', `Current temperature: ${currentTemperature}`);
} else {
const temp = parseFloat(arg);
if (isNaN(temp) || temp < 0 || temp > 2) {
addMessage('console', 'Invalid temperature. Must be between 0.0 and 2.0');
} else {
currentTemperature = temp;
addMessage('console', `Temperature set to ${currentTemperature}`);
}
}
return true;
} else if (cmd === '/topk') {
if (arg === undefined) {
addMessage('console', `Current top-k: ${currentTopK}`);
} else {
const topk = parseInt(arg);
if (isNaN(topk) || topk < 1 || topk > 200) {
addMessage('console', 'Invalid top-k. Must be between 1 and 200');
} else {
currentTopK = topk;
addMessage('console', `Top-k set to ${currentTopK}`);
}
}
return true;
} else if (cmd === '/clear') {
newConversation();
return true;
} else if (cmd === '/help') {
addMessage('console',
'Available commands:\n' +
'/temperature - Show current temperature\n' +
'/temperature <value> - Set temperature (0.0-2.0)\n' +
'/topk - Show current top-k\n' +
'/topk <value> - Set top-k (1-200)\n' +
'/clear - Clear conversation\n' +
'/help - Show this help message'
);
return true;
}
return false;
}
async function sendMessage() {
const message = chatInput.value.trim();
if (!message || isGenerating) return;
// Handle slash commands
if (message.startsWith('/')) {
chatInput.value = '';
chatInput.style.height = 'auto';
handleSlashCommand(message);
return;
}
chatInput.value = '';
chatInput.style.height = 'auto';
const userMessageIndex = messages.length;
messages.push({ role: 'user', content: message });
addMessage('user', message, userMessageIndex);
await generateAssistantResponse();
}
sendButton.disabled = false; sendButton.disabled = false;
// Autofocus the chat input on page load // Autofocus the chat input on page load

45
pyproject.toml
View File

@ -11,6 +11,7 @@ dependencies = [
"numpy==1.26.4", "numpy==1.26.4",
"psutil>=7.1.0", "psutil>=7.1.0",
"regex>=2025.9.1", "regex>=2025.9.1",
"setuptools>=80.9.0",
"tiktoken>=0.11.0", "tiktoken>=0.11.0",
"tokenizers>=0.22.0", "tokenizers>=0.22.0",
"torch>=2.8.0", "torch>=2.8.0",
@ -22,17 +23,6 @@ dependencies = [
requires = ["maturin>=1.7,<2.0"] requires = ["maturin>=1.7,<2.0"]
build-backend = "maturin" build-backend = "maturin"
# target torch to cuda 12.8
[tool.uv.sources]
torch = [
{ index = "pytorch-cu128" },
]
[[tool.uv.index]]
name = "pytorch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true
[tool.maturin] [tool.maturin]
module-name = "rustbpe" module-name = "rustbpe"
bindings = "pyo3" bindings = "pyo3"
@ -53,3 +43,36 @@ testpaths = ["tests"]
python_files = ["test_*.py"] python_files = ["test_*.py"]
python_classes = ["Test*"] python_classes = ["Test*"]
python_functions = ["test_*"] python_functions = ["test_*"]
# target torch to cuda 12.8 or CPU
[tool.uv.sources]
torch = [
{ index = "pytorch-cpu", extra = "cpu" },
{ index = "pytorch-cu128", extra = "gpu" },
]
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true
[[tool.uv.index]]
name = "pytorch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true
[project.optional-dependencies]
cpu = [
"torch>=2.8.0",
]
gpu = [
"torch>=2.8.0",
]
[tool.uv]
conflicts = [
[
{ extra = "cpu" },
{ extra = "gpu" },
],
]

97
run1000.sh Normal file
View File

@ -0,0 +1,97 @@
#!/bin/bash
# The $1000 tier of nanochat
# Designed to run end-to-end for $1000/24 ~= 41.6 hours on an 8XH100 node
# A bit sparser on comments, see speedrun.sh for more detail
# all the setup stuff
export OMP_NUM_THREADS=1
export NANOCHAT_BASE_DIR="$HOME/.cache/nanochat"
mkdir -p $NANOCHAT_BASE_DIR
command -v uv &> /dev/null || curl -LsSf https://astral.sh/uv/install.sh | sh
[ -d ".venv" ] || uv venv
uv sync --extra gpu
source .venv/bin/activate
if [ -z "$WANDB_RUN" ]; then
WANDB_RUN=dummy
fi
python -m nanochat.report reset
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
source "$HOME/.cargo/env"
uv run maturin develop --release --manifest-path rustbpe/Cargo.toml
EVAL_BUNDLE_URL=https://karpathy-public.s3.us-west-2.amazonaws.com/eval_bundle.zip
if [ ! -d "$NANOCHAT_BASE_DIR/eval_bundle" ]; then
curl -L -o eval_bundle.zip $EVAL_BUNDLE_URL
unzip -q eval_bundle.zip
rm eval_bundle.zip
mv eval_bundle $NANOCHAT_BASE_DIR
fi
curl -L -o $NANOCHAT_BASE_DIR/identity_conversations.jsonl https://karpathy-public.s3.us-west-2.amazonaws.com/identity_conversations.jsonl
# train tokenizer on ~4B characters and kick off download of the rest for pretraining
python -m nanochat.dataset -n 16
# start downloading the rest of the shards for a total of 800 (see below why 800)
python -m nanochat.dataset -n 800 &
# todo: download the rest of it
python -m scripts.tok_train --max_chars=4000000000
python -m scripts.tok_eval
# Documenting my process for determining the hyperparameters for this run1000.sh script:
# We want a budget of approx. $1000 ~= 41.6 hours of 8XH100 compute
# 1) I guessed the model size for this to be about depth=32
# 2) Determine the device_batch_size that fits:
# Running the base_train.py script with --depth=32, I saw that --device_batch_size=16
# runs out of memory, but --device_batch_size=8 fits. Inspecting `nvidia-smi` during training,
# I saw all GPUs were at about 78/80GB VRAM, so it just barely fits and we have good MFU at ~50%.
# So the training script was running ok and showed:
# Vocab size: 65,536
# num_layers: 32
# model_dim: 2048
# num_heads: 16
# num_kv_heads: 16
# Tokens / micro-batch / rank: 8 x 2048 = 16,384
# Tokens / micro-batch: 131,072
# Total batch size 524,288 => gradient accumulation steps: 4
# Number of parameters: 1,879,048,192
# Estimated FLOPs per token: 1.207960e+10
# Calculated number of iterations from target data:param ratio: 71,680
# Total number of training tokens: 37,580,963,840
# Tokens : Params ratio: 20.00
# Total training FLOPs estimate: 4.539628e+20
# step 00004/71680 (0.01%) | loss: 8.813754 | lrm: 1.00 | dt: 1571.88ms | tok/sec: 83,385 | mfu: 50.92 | total time: 0.00m
# step 00005/71680 (0.01%) | loss: 8.488074 | lrm: 1.00 | dt: 1572.76ms | tok/sec: 83,338 | mfu: 50.89 | total time: 0.00m
# ...
# 3) validate that the runtime fits our budget:
# The training script uses the Chinchilla scaling law to compute-optimally set #tokens = 20 * #params. In particular:
# The script shows that we will be training for 71,680 steps, and each step takes 1.574s so:
# estimated time to train: 71,680 * 1.574s / 60 / 60 = 31.3 hours.
# This is OK, fits our budget, and leaves ~10 hours for midtraining and SFT and evals and maybe RL.
# It's possible that we might even fit depth=33 or depth=34, but for now let's go along with this.
# 4) The last thing to pay attention to is the amount of training data required for the run.
# The script above calculated that "Total number of training tokens: 37,580,963,840"
# The tok_eval.py script reports about ~4.8 chars/token on average for the default tokenizer settings.
# So ~38B tokens # ~4.8 chars/token = ~185B chars.
# Each data shard is ~250M chars, so we need ~185B / 250M ~= 740 shards.
# For safety, I bumped that up to 800 shards, and that's why up above I used -n 800 when pre-downloading dataset shards.
# If we didn't have enough data, the training script would loop around and do multiple epochs over the same data,
# which would decrease model performance. Possibly 2, 3 or so epochs is ~ok, but certainly not ideal and at 10+ epochs we'd
# start to overfit hard.
# 5) That's it, everything else (e.g. the learning rates) is adjusted automatically by the training script.
torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=32 --device_batch_size=8
torchrun --standalone --nproc_per_node=8 -m scripts.base_loss
torchrun --standalone --nproc_per_node=8 -m scripts.base_eval
# midtrain
# NOTE: ensure that we use the same device_batch_size here as the base training script.
torchrun --standalone --nproc_per_node=8 -m scripts.mid_train -- --device_batch_size=8 --run=$WANDB_RUN
torchrun --standalone --nproc_per_node=8 -m scripts.chat_eval -- -i mid
# sft
torchrun --standalone --nproc_per_node=8 -m scripts.chat_sft -- --run=$WANDB_RUN
torchrun --standalone --nproc_per_node=8 -m scripts.chat_eval -- -i sft
# generate final report
python -m nanochat.report generate
# talk to it
python -m scripts.chat_web

3
rustbpe/src/lib.rs
View File

@ -292,8 +292,7 @@ impl Tokenizer {
// Prepare a true Python iterator object // Prepare a true Python iterator object
let py_iter: pyo3::Py<pyo3::PyAny> = unsafe { let py_iter: pyo3::Py<pyo3::PyAny> = unsafe {
pyo3::Bound::from_borrowed_ptr_or_err(py, pyo3::ffi::PyObject_GetIter(iterator.as_ptr()))? pyo3::Py::from_owned_ptr_or_err(py, pyo3::ffi::PyObject_GetIter(iterator.as_ptr()))?
.into()
}; };
// Global chunk counts // Global chunk counts

22
scripts/base_eval.py
View File

@ -15,11 +15,12 @@ import time
import json import json
import random import random
import yaml import yaml
from contextlib import nullcontext
import pandas as pd import pandas as pd
import torch import torch
from nanochat.common import compute_init, compute_cleanup, print0, get_base_dir from nanochat.common import compute_init, compute_cleanup, print0, get_base_dir, autodetect_device_type
from nanochat.tokenizer import HuggingFaceTokenizer from nanochat.tokenizer import HuggingFaceTokenizer
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
from nanochat.core_eval import evaluate_task from nanochat.core_eval import evaluate_task
@ -64,7 +65,7 @@ def evaluate_model(model, tokenizer, device, max_per_task=-1):
data = [json.loads(line.strip()) for line in f] data = [json.loads(line.strip()) for line in f]
# shuffle the data because in many cases it appears ordered but we want # shuffle the data because in many cases it appears ordered but we want
# the abillity to only run a subset of the data for debugging purposes etc. # the ability to only run a subset of the data for debugging purposes etc.
shuffle_rng = random.Random(1337) shuffle_rng = random.Random(1337)
shuffle_rng.shuffle(data) shuffle_rng.shuffle(data)
if max_per_task > 0: if max_per_task > 0:
@ -118,16 +119,21 @@ def load_hf_model(hf_path: str, device):
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
def main(): def main():
assert len(sys.argv) in [1, 2], "Usage: python base_eval.py [hf_path]" import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--hf-path', type=str, default=None, help='HuggingFace model path to evaluate')
parser.add_argument('--max-per-task', type=int, default=-1, help='Max examples per task to evaluate (-1 = disable)')
args = parser.parse_args()
# distributed / precision setup # distributed / precision setup
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type()
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16) ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
# Load model and tokenizer from command line or from file system # Load model and tokenizer from command line or from file system
if len(sys.argv) >= 2: if args.hf_path is not None:
# atm assume that if a path is given, it's a huggingface model path # atm assume that if a path is given, it's a huggingface model path
hf_path = sys.argv[1] hf_path = args.hf_path
print0(f"Loading huggingface model from: {hf_path}") print0(f"Loading huggingface model from: {hf_path}")
model, tokenizer = load_hf_model(hf_path, device) model, tokenizer = load_hf_model(hf_path, device)
model_name = hf_path # just for logging model_name = hf_path # just for logging
@ -140,7 +146,7 @@ def main():
# Evaluate the model # Evaluate the model
with autocast_ctx: with autocast_ctx:
out = evaluate_model(model, tokenizer, device) out = evaluate_model(model, tokenizer, device, max_per_task=args.max_per_task)
# Write out the results to a csv file # Write out the results to a csv file
core_metric = None core_metric = None

13
scripts/base_loss.py
View File

@ -7,9 +7,10 @@ Example run as:
torchrun --standalone --nproc_per_node=8 -m scripts.base_loss torchrun --standalone --nproc_per_node=8 -m scripts.base_loss
""" """
import os import os
from contextlib import nullcontext
import torch import torch
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
from nanochat.common import compute_init, print0, compute_cleanup from nanochat.common import compute_init, print0, compute_cleanup, autodetect_device_type
from nanochat.dataloader import tokenizing_distributed_data_loader from nanochat.dataloader import tokenizing_distributed_data_loader
from nanochat.tokenizer import get_token_bytes from nanochat.tokenizer import get_token_bytes
from nanochat.loss_eval import evaluate_bpb from nanochat.loss_eval import evaluate_bpb
@ -20,15 +21,15 @@ device_batch_size = 32
split_tokens = 20*524288 # number of tokens to evaluate per split split_tokens = 20*524288 # number of tokens to evaluate per split
model_tag = None # optional model tag for the output directory name model_tag = None # optional model tag for the output directory name
model_step = None # optional model step for the output directory name model_step = None # optional model step for the output directory name
device_type = "" # cuda|cpu|mps (empty => autodetect)
exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file
# Load the base model and the tokenizer # Load the base model and the tokenizer
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type() if device_type == "" else device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
model, tokenizer, meta = load_model("base", device, phase="eval", model_tag=model_tag, step=model_step) model, tokenizer, meta = load_model("base", device, phase="eval", model_tag=model_tag, step=model_step)
sequence_len = meta["model_config"]["sequence_len"] # could be arbitrary really sequence_len = meta["model_config"]["sequence_len"] # could be arbitrary really
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
# Set up the precision we'll run with
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16)
# Evaluate the loss on each split # Evaluate the loss on each split
tokens_per_step = device_batch_size * sequence_len * ddp_world_size tokens_per_step = device_batch_size * sequence_len * ddp_world_size
@ -37,7 +38,7 @@ steps = split_tokens // tokens_per_step
token_bytes = get_token_bytes(device=device) token_bytes = get_token_bytes(device=device)
bpb_results = {} bpb_results = {}
for split_name in ["train", "val"]: for split_name in ["train", "val"]:
loader = tokenizing_distributed_data_loader(device_batch_size, sequence_len, split_name) loader = tokenizing_distributed_data_loader(device_batch_size, sequence_len, split_name, device=device)
with autocast_ctx: with autocast_ctx:
bpb = evaluate_bpb(model, loader, steps, token_bytes) bpb = evaluate_bpb(model, loader, steps, token_bytes)
print0(f"{split_name} bpb: {bpb:.4f}") print0(f"{split_name} bpb: {bpb:.4f}")

50
scripts/base_train.py
View File

@ -6,17 +6,22 @@ python base_train.py
or distributed as: or distributed as:
torchrun --nproc_per_node=8 base_train.py torchrun --nproc_per_node=8 base_train.py
If you are only on CPU/Macbook, you'll want to train a much much smaller LLM. Example:
python -m scripts.base_train --depth=4 --max_seq_len=512 --device_batch_size=1 --eval_tokens=512 --core_metric_every=-1 --total_batch_size=512 --num_iterations=20
""" """
import os import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
import time import time
from contextlib import nullcontext
import wandb import wandb
import torch import torch
from nanochat.gpt import GPT, GPTConfig from nanochat.gpt import GPT, GPTConfig
from nanochat.dataloader import tokenizing_distributed_data_loader from nanochat.dataloader import tokenizing_distributed_data_loader
from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, print_banner, get_base_dir from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, print_banner, get_base_dir, autodetect_device_type
from nanochat.tokenizer import get_tokenizer, get_token_bytes from nanochat.tokenizer import get_tokenizer, get_token_bytes
from nanochat.checkpoint_manager import save_checkpoint from nanochat.checkpoint_manager import save_checkpoint
from nanochat.loss_eval import evaluate_bpb from nanochat.loss_eval import evaluate_bpb
@ -27,6 +32,8 @@ print_banner()
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# User settings # User settings
run = "dummy" # wandb run name default ("dummy" is special - we won't log to wandb) run = "dummy" # wandb run name default ("dummy" is special - we won't log to wandb)
# Runtime
device_type = "" # cuda|cpu|mps (empty => autodetect good device type default, in order: CUDA > MPS > CPU)
# Model architecture # Model architecture
depth = 20 # the depth of the Transformer model to train, rest of the kwargs are derived depth = 20 # the depth of the Transformer model to train, rest of the kwargs are derived
max_seq_len = 2048 # max context length max_seq_len = 2048 # max context length
@ -42,10 +49,13 @@ unembedding_lr = 0.004 # learning rate for the unembedding parameters (Adam)
weight_decay = 0.0 # weight decay for the embedding/unembedding parameters (Adam) weight_decay = 0.0 # weight decay for the embedding/unembedding parameters (Adam)
matrix_lr = 0.02 # learning rate for the matrix parameters (Muon) matrix_lr = 0.02 # learning rate for the matrix parameters (Muon)
grad_clip = 1.0 # gradient clipping value (0.0 = disabled) grad_clip = 1.0 # gradient clipping value (0.0 = disabled)
warmup_ratio = 0.0 # ratio of iterations for LR warmup
warmdown_ratio = 0.2 # ratio of iterations for LR warmdown
final_lr_frac = 0.0 # final LR is this fraction of the initial LR
# Evaluation # Evaluation
eval_every = 250 # every how many steps to evaluate the model for val bpb eval_every = 250 # every how many steps to evaluate the model for val bpb
eval_tokens = 20*524288 # number of tokens to evaluate val loss on eval_tokens = 20*524288 # number of tokens to evaluate val loss on
core_metric_every = 2000 # every how many steps to evaluate the core metric core_metric_every = 2000 # every how many steps to evaluate the core metric (-1 = disable)
core_metric_max_per_task = 500 # examples per task in estimating the core metric core_metric_max_per_task = 500 # examples per task in estimating the core metric
sample_every = 2000 # every how many steps to sample from the model sample_every = 2000 # every how many steps to sample from the model
# Output # Output
@ -57,9 +67,12 @@ user_config = {k: globals()[k] for k in config_keys} # will be useful for loggin
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Compute init # Compute init
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type() if device_type == "" else device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
master_process = ddp_rank == 0 # this process will do logging, checkpointing etc. master_process = ddp_rank == 0 # this process will do logging, checkpointing etc.
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16) autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
synchronize = torch.cuda.synchronize if device_type == "cuda" else lambda: None
get_max_memory = torch.cuda.max_memory_allocated if device_type == "cuda" else lambda: 0
# wandb logging init # wandb logging init
use_dummy_wandb = run == "dummy" or not master_process use_dummy_wandb = run == "dummy" or not master_process
@ -75,7 +88,7 @@ print0(f"Vocab size: {vocab_size:,}")
num_layers = depth num_layers = depth
model_dim = depth * 64 # aspect ratio 64 (usually this is varied from 64 -> 128 as model size increases) model_dim = depth * 64 # aspect ratio 64 (usually this is varied from 64 -> 128 as model size increases)
num_heads = max(1, (model_dim + 127) // 128) # head dim 128 (the division here is ceil div) num_heads = max(1, (model_dim + 127) // 128) # head dim 128 (the division here is ceil div)
num_kv_heads = num_heads # 1:1 MQA ratio num_kv_heads = num_heads # default is 1:1 GQA (Group Query Attention) ratio (i.e. GQA is disabled)
print0(f"num_layers: {num_layers}") print0(f"num_layers: {num_layers}")
print0(f"model_dim: {model_dim}") print0(f"model_dim: {model_dim}")
print0(f"num_heads: {num_heads}") print0(f"num_heads: {num_heads}")
@ -96,7 +109,7 @@ model_config_kwargs = dict(sequence_len=max_seq_len, vocab_size=vocab_size, n_la
with torch.device("meta"): with torch.device("meta"):
model_config = GPTConfig(**model_config_kwargs) model_config = GPTConfig(**model_config_kwargs)
model = GPT(model_config) model = GPT(model_config)
model.to_empty(device="cuda") model.to_empty(device=device)
model.init_weights() model.init_weights()
orig_model = model # original, uncompiled model, for saving raw model state_dict orig_model = model # original, uncompiled model, for saving raw model state_dict
model = torch.compile(model, dynamic=False) # TODO: dynamic True/False think through model = torch.compile(model, dynamic=False) # TODO: dynamic True/False think through
@ -133,18 +146,14 @@ adamw_optimizer, muon_optimizer = optimizers
# Initialize the DataLoaders for train/val # Initialize the DataLoaders for train/val
base_dir = get_base_dir() base_dir = get_base_dir()
tokens_dir = os.path.join(base_dir, "tokenized_data") tokens_dir = os.path.join(base_dir, "tokenized_data")
train_loader = tokenizing_distributed_data_loader(device_batch_size, max_seq_len, split="train") train_loader = tokenizing_distributed_data_loader(device_batch_size, max_seq_len, split="train", device=device)
build_val_loader = lambda: tokenizing_distributed_data_loader(device_batch_size, max_seq_len, split="val") build_val_loader = lambda: tokenizing_distributed_data_loader(device_batch_size, max_seq_len, split="val", device=device)
x, y = next(train_loader) # kick off load of the very first batch of data x, y = next(train_loader) # kick off load of the very first batch of data
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Set up hyperparameter schedulers # Set up hyperparameter schedulers
# Learning rate scheduler # Learning rate scheduler
# TODO: experiment with a short warmup for the AdamW params (expecting slight improvement)
warmup_ratio = 0.0 # ratio of iterations for LR warmup
warmdown_ratio = 0.2 # ratio of iterations for LR warmdown
final_lr_frac = 0.0 # final LR is this fraction of the initial LR
def get_lr_multiplier(it): def get_lr_multiplier(it):
warmup_iters = round(warmup_ratio * num_iterations) warmup_iters = round(warmup_ratio * num_iterations)
warmdown_iters = round(warmdown_ratio * num_iterations) warmdown_iters = round(warmdown_ratio * num_iterations)
@ -193,7 +202,8 @@ for step in range(num_iterations + 1):
# once in a while: estimate the CORE metric (all ranks participate) # once in a while: estimate the CORE metric (all ranks participate)
# use the original uncompiled model because the inputs keep changing shape # use the original uncompiled model because the inputs keep changing shape
if last_step or (step > 0 and step % core_metric_every == 0): results = {}
if core_metric_every > 0 and (last_step or (step > 0 and step % core_metric_every == 0)):
model.eval() model.eval()
with autocast_ctx: with autocast_ctx:
results = evaluate_model(orig_model, tokenizer, device, max_per_task=core_metric_max_per_task) results = evaluate_model(orig_model, tokenizer, device, max_per_task=core_metric_max_per_task)
@ -219,7 +229,7 @@ for step in range(num_iterations + 1):
"My favorite color is", "My favorite color is",
"If 5*x + 3 = 13, then x is", "If 5*x + 3 = 13, then x is",
] ]
engine = Engine(model, tokenizer) engine = Engine(orig_model, tokenizer) # use orig_model to avoid recompilation
for prompt in prompts: for prompt in prompts:
tokens = tokenizer(prompt, prepend="<|bos|>") tokens = tokenizer(prompt, prepend="<|bos|>")
with autocast_ctx: with autocast_ctx:
@ -252,7 +262,7 @@ for step in range(num_iterations + 1):
# ------------------------------------------------------------------------- # -------------------------------------------------------------------------
# single training step # single training step
# evaluate the gradient # evaluate the gradient
torch.cuda.synchronize() synchronize()
t0 = time.time() t0 = time.time()
for micro_step in range(grad_accum_steps): for micro_step in range(grad_accum_steps):
with autocast_ctx: with autocast_ctx:
@ -261,7 +271,7 @@ for step in range(num_iterations + 1):
loss = loss / grad_accum_steps # each .backward() is a grad sum => normalize loss here loss = loss / grad_accum_steps # each .backward() is a grad sum => normalize loss here
loss.backward() loss.backward()
x, y = next(train_loader) # prefetch the next batch while the GPU is busy with forward/backward x, y = next(train_loader) # prefetch the next batch while the GPU is busy with forward/backward
# gradient clipping (TODO possibly expertiment with) # gradient clipping (TODO possibly experiment with)
if grad_clip > 0.0: if grad_clip > 0.0:
torch.nn.utils.clip_grad_norm_(orig_model.parameters(), grad_clip) torch.nn.utils.clip_grad_norm_(orig_model.parameters(), grad_clip)
# step the optimizers # step the optimizers
@ -275,7 +285,7 @@ for step in range(num_iterations + 1):
for opt in optimizers: for opt in optimizers:
opt.step() opt.step()
model.zero_grad(set_to_none=True) model.zero_grad(set_to_none=True)
torch.cuda.synchronize() synchronize()
t1 = time.time() t1 = time.time()
dt = t1 - t0 dt = t1 - t0
# ------------------------------------------------------------------------- # -------------------------------------------------------------------------
@ -304,7 +314,7 @@ for step in range(num_iterations + 1):
}) })
# print a few more stats # print a few more stats
print0(f"Peak memory usage: {torch.cuda.max_memory_allocated() / 1024 / 1024:.2f}MiB") print0(f"Peak memory usage: {get_max_memory() / 1024 / 1024:.2f}MiB")
print0(f"Total training time: {total_training_time/60:.2f}m") print0(f"Total training time: {total_training_time/60:.2f}m")
print0(f"Minimum validation bpb: {min_val_bpb:.4f}") print0(f"Minimum validation bpb: {min_val_bpb:.4f}")
@ -326,11 +336,11 @@ get_report().log(section="Base model training", data=[
{ # stats about training outcomes { # stats about training outcomes
"Minimum validation bpb": min_val_bpb, "Minimum validation bpb": min_val_bpb,
"Final validation bpb": val_bpb, "Final validation bpb": val_bpb,
"CORE metric estimate": results["core_metric"], "CORE metric estimate": results.get("core_metric", None),
"MFU %": f"{mfu:.2f}%", "MFU %": f"{mfu:.2f}%",
"Total training flops": f"{flops_so_far:e}", "Total training flops": f"{flops_so_far:e}",
"Total training time": f"{total_training_time/60:.2f}m", "Total training time": f"{total_training_time/60:.2f}m",
"Peak memory usage": f"{torch.cuda.max_memory_allocated() / 1024 / 1024:.2f}MiB", "Peak memory usage": f"{get_max_memory() / 1024 / 1024:.2f}MiB",
} }
]) ])

12
scripts/chat_cli.py
View File

@ -6,7 +6,8 @@ python -m scripts.chat_cli -i mid
""" """
import argparse import argparse
import torch import torch
from nanochat.common import compute_init from nanochat.common import compute_init, autodetect_device_type
from contextlib import nullcontext
from nanochat.engine import Engine from nanochat.engine import Engine
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
@ -17,11 +18,16 @@ parser.add_argument('-s', '--step', type=int, default=None, help='Step to load')
parser.add_argument('-p', '--prompt', type=str, default='', help='Prompt the model, get a single response back') parser.add_argument('-p', '--prompt', type=str, default='', help='Prompt the model, get a single response back')
parser.add_argument('-t', '--temperature', type=float, default=0.6, help='Temperature for generation') parser.add_argument('-t', '--temperature', type=float, default=0.6, help='Temperature for generation')
parser.add_argument('-k', '--top-k', type=int, default=50, help='Top-k sampling parameter') parser.add_argument('-k', '--top-k', type=int, default=50, help='Top-k sampling parameter')
parser.add_argument('--device-type', type=str, default='', choices=['cuda', 'cpu', 'mps'], help='Device type for evaluation: cuda|cpu|mps. empty => autodetect')
parser.add_argument('-d', '--dtype', type=str, default='bfloat16', choices=['float32', 'bfloat16'])
args = parser.parse_args() args = parser.parse_args()
# Init the model and tokenizer # Init the model and tokenizer
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init()
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16) device_type = autodetect_device_type() if args.device_type == "" else args.device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
ptdtype = torch.float32 if args.dtype == 'float32' else torch.bfloat16
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext()
model, tokenizer, meta = load_model(args.source, device, phase="eval", model_tag=args.model_tag, step=args.step) model, tokenizer, meta = load_model(args.source, device, phase="eval", model_tag=args.model_tag, step=args.step)
# Special tokens for the chat state machine # Special tokens for the chat state machine

16
scripts/chat_eval.py
View File

@ -10,11 +10,12 @@ torchrun --nproc_per_node=8 -m scripts.chat_eval -- -a ARC-Easy
import argparse import argparse
from functools import partial from functools import partial
from contextlib import nullcontext
import torch import torch
import torch.distributed as dist import torch.distributed as dist
from nanochat.common import compute_init, compute_cleanup, get_dist_info, print0 from nanochat.common import compute_init, compute_cleanup, get_dist_info, print0, autodetect_device_type
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
from nanochat.engine import Engine from nanochat.engine import Engine
@ -22,6 +23,7 @@ from tasks.humaneval import HumanEval
from tasks.mmlu import MMLU from tasks.mmlu import MMLU
from tasks.arc import ARC from tasks.arc import ARC
from tasks.gsm8k import GSM8K from tasks.gsm8k import GSM8K
from tasks.spellingbee import SpellingBee
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Generative evaluation loop (we go one problem at a time, sample, evaluate) # Generative evaluation loop (we go one problem at a time, sample, evaluate)
@ -115,7 +117,7 @@ def run_categorical_eval(task_object, tokenizer, model, batch_size, max_problems
logits = model(prompt_ids) # (B, T, V) logits = model(prompt_ids) # (B, T, V)
# Focus on the available answer on just the letters corresponding to choices # Focus on the available answer on just the letters corresponding to choices
# Note that this helps the evaluation a lot because it specifically narrows the focus to only the avilable letters # Note that this helps the evaluation a lot because it specifically narrows the focus to only the available letters
# The much harder alternative would be to just generate from the Assistant and check if it responded with the correct # The much harder alternative would be to just generate from the Assistant and check if it responded with the correct
# letter (e.g. A, B, C, D), but evaluations typically make the task easier in this way. # letter (e.g. A, B, C, D), but evaluations typically make the task easier in this way.
for idx, conversation in enumerate(conversations): for idx, conversation in enumerate(conversations):
@ -164,6 +166,7 @@ def run_chat_eval(task_name, model, tokenizer, engine,
'ARC-Easy': partial(ARC, subset="ARC-Easy", split="test"), 'ARC-Easy': partial(ARC, subset="ARC-Easy", split="test"),
'ARC-Challenge': partial(ARC, subset="ARC-Challenge", split="test"), 'ARC-Challenge': partial(ARC, subset="ARC-Challenge", split="test"),
'GSM8K': partial(GSM8K, subset="main", split="test"), 'GSM8K': partial(GSM8K, subset="main", split="test"),
'SpellingBee': partial(SpellingBee, size=256, split="test"),
}[task_name] }[task_name]
task_object = task_module() task_object = task_module()
# Run the evaluation # Run the evaluation
@ -191,23 +194,26 @@ if __name__ == "__main__":
parser.add_argument('-g', '--model-tag', type=str, default=None, help='Model tag to load') parser.add_argument('-g', '--model-tag', type=str, default=None, help='Model tag to load')
parser.add_argument('-s', '--step', type=int, default=None, help='Step to load') parser.add_argument('-s', '--step', type=int, default=None, help='Step to load')
parser.add_argument('-x', '--max-problems', type=int, default=None, help='Max problems to evaluate') parser.add_argument('-x', '--max-problems', type=int, default=None, help='Max problems to evaluate')
parser.add_argument('--device-type', type=str, default='', choices=['cuda', 'cpu', 'mps'], help='Device type for evaluation: cuda|cpu|mps. empty => autodetect')
args = parser.parse_args() args = parser.parse_args()
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type() if args.device_type == "" else args.device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
ptdtype = torch.float32 if args.dtype == 'float32' else torch.bfloat16 ptdtype = torch.float32 if args.dtype == 'float32' else torch.bfloat16
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=ptdtype) autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext()
model, tokenizer, meta = load_model(args.source, device, phase="eval", model_tag=args.model_tag, step=args.step) model, tokenizer, meta = load_model(args.source, device, phase="eval", model_tag=args.model_tag, step=args.step)
engine = Engine(model, tokenizer) engine = Engine(model, tokenizer)
# Get the tasks to evaluate on # Get the tasks to evaluate on
all_tasks = ['ARC-Easy', 'ARC-Challenge', 'MMLU', 'GSM8K', 'HumanEval'] all_tasks = ['ARC-Easy', 'ARC-Challenge', 'MMLU', 'GSM8K', 'HumanEval', 'SpellingBee']
baseline_accuracies = { baseline_accuracies = {
'ARC-Easy': 0.25, # multiple choice 1 of 4 => 25% 'ARC-Easy': 0.25, # multiple choice 1 of 4 => 25%
'ARC-Challenge': 0.25, # multiple choice 1 of 4 => 25% 'ARC-Challenge': 0.25, # multiple choice 1 of 4 => 25%
'MMLU': 0.25, # multiple choice 1 of 4 => 25% 'MMLU': 0.25, # multiple choice 1 of 4 => 25%
'GSM8K': 0.0, # open-ended => 0% 'GSM8K': 0.0, # open-ended => 0%
'HumanEval': 0.0, # open-ended => 0% 'HumanEval': 0.0, # open-ended => 0%
'SpellingBee': 0.0, # open-ended => 0%
} }
task_names = all_tasks if args.task_name is None else args.task_name.split('|') task_names = all_tasks if args.task_name is None else args.task_name.split('|')

2
scripts/chat_rl.py
View File

@ -206,7 +206,7 @@ def get_lr_multiplier(it):
lrm = 1.0 - it / num_steps lrm = 1.0 - it / num_steps
return lrm return lrm
# Calculate the number of examples each rank handles to achive the desired examples_per_step # Calculate the number of examples each rank handles to achieve the desired examples_per_step
print0(f"Total sequences per step: {examples_per_step * num_samples}") # total batch size in sequences/step print0(f"Total sequences per step: {examples_per_step * num_samples}") # total batch size in sequences/step
assert examples_per_step % ddp_world_size == 0, "Desired examples per step must be divisible by the number of ranks" assert examples_per_step % ddp_world_size == 0, "Desired examples per step must be divisible by the number of ranks"
examples_per_rank = examples_per_step // ddp_world_size # per GPU examples_per_rank = examples_per_step // ddp_world_size # per GPU

44
scripts/chat_sft.py
View File

@ -11,24 +11,24 @@ torchrun --standalone --nproc_per_node=8 -m scripts.chat_sft
import os import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
import copy
import wandb import wandb
import torch import torch
import torch.distributed as dist import torch.distributed as dist
from contextlib import nullcontext
from nanochat.common import compute_init, compute_cleanup, get_base_dir, print0, DummyWandb from nanochat.common import compute_init, compute_cleanup, get_base_dir, print0, DummyWandb, autodetect_device_type
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
from nanochat.checkpoint_manager import save_checkpoint from nanochat.checkpoint_manager import save_checkpoint
from nanochat.engine import Engine from nanochat.engine import Engine
from scripts.chat_eval import run_chat_eval from scripts.chat_eval import run_chat_eval
from tasks.common import TaskMixture, TaskSequence from tasks.common import TaskMixture
from tasks.mmlu import MMLU
from tasks.arc import ARC from tasks.arc import ARC
from tasks.gsm8k import GSM8K from tasks.gsm8k import GSM8K
from tasks.humaneval import HumanEval
from tasks.smoltalk import SmolTalk from tasks.smoltalk import SmolTalk
from tasks.customjson import CustomJSON
from tasks.spellingbee import SimpleSpelling, SpellingBee
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# SFT Hyperparameters # SFT Hyperparameters
@ -38,11 +38,12 @@ source = "mid" # base|mid , which checkpoint to load the model from (base model
model_tag = None # model tag to load the model from (base model or midtrained model) model_tag = None # model tag to load the model from (base model or midtrained model)
step = None # step to load the model from (base model or midtrained model) step = None # step to load the model from (base model or midtrained model)
# compute/precision # compute/precision
device_type = "" # cuda|cpu|mps (empty => autodetect)
dtype = "bfloat16" dtype = "bfloat16"
device_batch_size = 4 # max to avoid OOM device_batch_size = 4 # max to avoid OOM
# optimization # optimization
num_epochs = 1 num_epochs = 1
max_iterations = -1 # override number of iterations (-1 = use num_epochs * num_iterations) num_iterations = -1 # override number of iterations (-1 = disable, use num_epochs to derive it)
target_examples_per_step = 32 target_examples_per_step = 32
unembedding_lr = 0.004 unembedding_lr = 0.004
embedding_lr = 0.2 embedding_lr = 0.2
@ -53,6 +54,7 @@ init_lr_frac = 0.02
eval_every = 100 eval_every = 100
eval_steps = 100 eval_steps = 100
eval_metrics_every = 200 eval_metrics_every = 200
eval_metrics_max_problems = 1024
# now allow CLI to override the settings via the configurator lol # now allow CLI to override the settings via the configurator lol
config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))] config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file
@ -60,10 +62,11 @@ user_config = {k: globals()[k] for k in config_keys} # possibly useful for loggi
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Compute init # Compute init
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type() if device_type == "" else device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
master_process = ddp_rank == 0 master_process = ddp_rank == 0
dtype = torch.float32 if dtype == 'float32' else torch.bfloat16 ptdtype = torch.float32 if dtype == 'float32' else torch.bfloat16
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=dtype) autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext()
# wandb logging init # wandb logging init
use_dummy_wandb = run == "dummy" or not master_process use_dummy_wandb = run == "dummy" or not master_process
@ -77,13 +80,16 @@ engine = Engine(model, tokenizer) # will be used for inline model evaluation onl
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Task data mixture we'll train on # Task data mixture we'll train on
identity_conversations_filepath = os.path.join(get_base_dir(), "identity_conversations.jsonl")
train_ds = TaskMixture([ train_ds = TaskMixture([
ARC(subset="ARC-Easy", split="train"), # 2.3K rows ARC(subset="ARC-Easy", split="train"), # 2.3K rows
ARC(subset="ARC-Challenge", split="train"), # 1.1K rows ARC(subset="ARC-Challenge", split="train"), # 1.1K rows
GSM8K(subset="main", split="train"), # 8K rows GSM8K(subset="main", split="train"), # 8K rows
SmolTalk(split="train", stop=10_000), # 10K rows of smoltalk SmolTalk(split="train", stop=10_000), # 10K rows of smoltalk
]) # 2.3K + 1.1K + 8K + 10K = 21.4K rows CustomJSON(filepath=identity_conversations_filepath), # 1K rows of synthetic identity conversations
SimpleSpelling(size=300, split="train"), # 300 rows of Simple Spelling (e.g. spell the word 'apple')
SpellingBee(size=300, split="train"), # 300 rows of Spelling Bee (e.g. how many 'r' are in 'strawberry'?)
]) # 2.3K + 1.1K + 8K + 10K + 1K + 0.3K + 0.3K = 23K rows
val_ds = SmolTalk(split="test") # general conversations, 24K rows (though we don't actually use all of it) val_ds = SmolTalk(split="test") # general conversations, 24K rows (though we don't actually use all of it)
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
@ -129,10 +135,10 @@ assert target_examples_per_step % examples_per_step == 0, "Target examples per s
grad_accum_steps = target_examples_per_step // examples_per_step grad_accum_steps = target_examples_per_step // examples_per_step
print0(f"=> Setting grad accum steps: {grad_accum_steps}") print0(f"=> Setting grad accum steps: {grad_accum_steps}")
num_iterations = (len(train_ds) // target_examples_per_step) * num_epochs if num_iterations == -1:
if max_iterations >= 0 and num_iterations > max_iterations: # derive num_iterations from num_epochs and the size of the dataset
print0(f"Number of iterations is too high: {num_iterations}, capping to {max_iterations}") assert num_epochs > 0, "num_epochs must be positive if num_iterations is -1"
num_iterations = max_iterations num_iterations = (len(train_ds) // target_examples_per_step) * num_epochs
train_loader = sft_data_generator(train_ds, batch_size=device_batch_size) train_loader = sft_data_generator(train_ds, batch_size=device_batch_size)
build_val_loader = lambda: sft_data_generator(val_ds, batch_size=device_batch_size) build_val_loader = lambda: sft_data_generator(val_ds, batch_size=device_batch_size)
@ -186,16 +192,14 @@ for step in range(num_iterations):
}) })
model.train() model.train()
# evlauate MMLU accuracy # evlauate accuracy of the multiple choice tasks (which are quick to run)
if last_step or (step > 0 and step % eval_metrics_every == 0): if last_step or (step > 0 and step % eval_metrics_every == 0):
model.eval() model.eval()
metrics = {} metrics = {}
with torch.no_grad(), autocast_ctx: with torch.no_grad(), autocast_ctx:
# note that because these are inside no_grad, we can usually afford to at least ~2X the batch size # note that because these are inside no_grad, we can usually afford to at least ~2X the batch size
metrics["mmlu_acc"] = run_chat_eval("MMLU", model, tokenizer, engine, batch_size=device_batch_size*2, max_problems=1024) metrics["mmlu_acc"] = run_chat_eval("MMLU", model, tokenizer, engine, batch_size=device_batch_size*2, max_problems=eval_metrics_max_problems)
metrics["arc_easy_acc"] = run_chat_eval("ARC-Easy", model, tokenizer, engine, batch_size=device_batch_size*2, max_problems=1024) metrics["arc_easy_acc"] = run_chat_eval("ARC-Easy", model, tokenizer, engine, batch_size=device_batch_size*2, max_problems=eval_metrics_max_problems)
metrics["gsm8k_acc"] = run_chat_eval("GSM8K", model, tokenizer, engine, max_problems=64)
metrics["humaneval_acc"] = run_chat_eval("HumanEval", model, tokenizer, engine, max_problems=64)
metrics_str = ', '.join(f'{k}: {v:.6f}' for k, v in metrics.items()) metrics_str = ', '.join(f'{k}: {v:.6f}' for k, v in metrics.items())
print0(f"Step {step:05d} | {metrics_str}") print0(f"Step {step:05d} | {metrics_str}")
wandb_run.log({ wandb_run.log({

371
scripts/chat_web.py
View File

@ -1,26 +1,67 @@
#!/usr/bin/env python3 #!/usr/bin/env python3
""" """
Unified web chat server - serves both UI and API from a single FastAPI instance. Unified web chat server - serves both UI and API from a single FastAPI instance.
Run with: python web_chat.py
Then open http://localhost:8000 in your browser. Uses data parallelism to distribute requests across multiple GPUs. Each GPU loads
a full copy of the model, and incoming requests are distributed to available workers.
Launch examples:
- single available GPU (default)
python -m scripts.chat_web
- 4 GPUs
python -m scripts.chat_web --num-gpus 4
To chat, open the URL printed in the console. (If on cloud box, make sure to use public IP)
Endpoints:
GET / - Chat UI
POST /chat/completions - Chat API (streaming only)
GET /health - Health check with worker pool status
GET /stats - Worker pool statistics and GPU utilization
Abuse Prevention:
- Maximum 500 messages per request
- Maximum 8000 characters per message
- Maximum 32000 characters total conversation length
- Temperature clamped to 0.0-2.0
- Top-k clamped to 1-200
- Max tokens clamped to 1-4096
""" """
import argparse import argparse
import json import json
import os import os
import torch import torch
import asyncio
import logging
import random
from contextlib import asynccontextmanager from contextlib import asynccontextmanager
from fastapi import FastAPI from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import StreamingResponse, HTMLResponse, FileResponse from fastapi.responses import StreamingResponse, HTMLResponse, FileResponse
from pydantic import BaseModel from pydantic import BaseModel
from typing import List, Optional, AsyncGenerator from typing import List, Optional, AsyncGenerator
from dataclasses import dataclass
from nanochat.common import compute_init from contextlib import nullcontext
from nanochat.common import compute_init, autodetect_device_type
from nanochat.checkpoint_manager import load_model from nanochat.checkpoint_manager import load_model
from nanochat.engine import Engine from nanochat.engine import Engine
# Abuse prevention limits
MAX_MESSAGES_PER_REQUEST = 500
MAX_MESSAGE_LENGTH = 8000
MAX_TOTAL_CONVERSATION_LENGTH = 32000
MIN_TEMPERATURE = 0.0
MAX_TEMPERATURE = 2.0
MIN_TOP_K = 1
MAX_TOP_K = 200
MIN_MAX_TOKENS = 1
MAX_MAX_TOKENS = 4096
parser = argparse.ArgumentParser(description='NanoChat Web Server') parser = argparse.ArgumentParser(description='NanoChat Web Server')
parser.add_argument('-n', '--num-gpus', type=int, default=1, help='Number of GPUs to use (default: 1)')
parser.add_argument('-i', '--source', type=str, default="sft", help="Source of the model: sft|mid|rl") parser.add_argument('-i', '--source', type=str, default="sft", help="Source of the model: sft|mid|rl")
parser.add_argument('-t', '--temperature', type=float, default=0.8, help='Default temperature for generation') parser.add_argument('-t', '--temperature', type=float, default=0.8, help='Default temperature for generation')
parser.add_argument('-k', '--top-k', type=int, default=50, help='Default top-k sampling parameter') parser.add_argument('-k', '--top-k', type=int, default=50, help='Default top-k sampling parameter')
@ -28,11 +69,83 @@ parser.add_argument('-m', '--max-tokens', type=int, default=512, help='Default m
parser.add_argument('-g', '--model-tag', type=str, default=None, help='Model tag to load') parser.add_argument('-g', '--model-tag', type=str, default=None, help='Model tag to load')
parser.add_argument('-s', '--step', type=int, default=None, help='Step to load') parser.add_argument('-s', '--step', type=int, default=None, help='Step to load')
parser.add_argument('-p', '--port', type=int, default=8000, help='Port to run the server on') parser.add_argument('-p', '--port', type=int, default=8000, help='Port to run the server on')
parser.add_argument('-d', '--dtype', type=str, default='bfloat16', choices=['float32', 'bfloat16'])
parser.add_argument('--device-type', type=str, default='', choices=['cuda', 'cpu', 'mps'], help='Device type for evaluation: cuda|cpu|mps. empty => autodetect')
parser.add_argument('--host', type=str, default='0.0.0.0', help='Host to bind the server to') parser.add_argument('--host', type=str, default='0.0.0.0', help='Host to bind the server to')
args = parser.parse_args() args = parser.parse_args()
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() # Configure logging for conversation traffic
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16) logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(message)s',
datefmt='%Y-%m-%d %H:%M:%S'
)
logger = logging.getLogger(__name__)
device_type = autodetect_device_type() if args.device_type == "" else args.device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
ptdtype = torch.float32 if args.dtype == 'float32' else torch.bfloat16
@dataclass
class Worker:
"""A worker with a model loaded on a specific GPU."""
gpu_id: int
device: torch.device
engine: Engine
tokenizer: object
autocast_ctx: torch.amp.autocast
class WorkerPool:
"""Pool of workers, each with a model replica on a different GPU."""
def __init__(self, num_gpus: Optional[int] = None):
if num_gpus is None:
if device_type == "cuda":
num_gpus = torch.cuda.device_count()
else:
num_gpus = 1 # e.g. cpu|mps
self.num_gpus = num_gpus
self.workers: List[Worker] = []
self.available_workers: asyncio.Queue = asyncio.Queue()
async def initialize(self, source: str, model_tag: Optional[str] = None, step: Optional[int] = None):
"""Load model on each GPU."""
print(f"Initializing worker pool with {self.num_gpus} GPUs...")
if self.num_gpus > 1:
assert device_type == "cuda", "Only CUDA supports multiple workers/GPUs. cpu|mps does not."
for gpu_id in range(self.num_gpus):
if device_type == "cuda":
device = torch.device(f"cuda:{gpu_id}")
print(f"Loading model on GPU {gpu_id}...")
else:
device = torch.device(device_type) # e.g. cpu|mps
print(f"Loading model on {device_type}...")
model, tokenizer, _ = load_model(source, device, phase="eval", model_tag=model_tag, step=step)
engine = Engine(model, tokenizer)
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext()
worker = Worker(
gpu_id=gpu_id,
device=device,
engine=engine,
tokenizer=tokenizer,
autocast_ctx=autocast_ctx
)
self.workers.append(worker)
await self.available_workers.put(worker)
print(f"All {self.num_gpus} workers initialized!")
async def acquire_worker(self) -> Worker:
"""Get an available worker from the pool."""
return await self.available_workers.get()
async def release_worker(self, worker: Worker):
"""Return a worker to the pool."""
await self.available_workers.put(worker)
class ChatMessage(BaseModel): class ChatMessage(BaseModel):
role: str role: str
@ -43,14 +156,76 @@ class ChatRequest(BaseModel):
temperature: Optional[float] = None temperature: Optional[float] = None
max_tokens: Optional[int] = None max_tokens: Optional[int] = None
top_k: Optional[int] = None top_k: Optional[int] = None
stream: Optional[bool] = True
def validate_chat_request(request: ChatRequest):
"""Validate chat request to prevent abuse."""
# Check number of messages
if len(request.messages) == 0:
raise HTTPException(status_code=400, detail="At least one message is required")
if len(request.messages) > MAX_MESSAGES_PER_REQUEST:
raise HTTPException(
status_code=400,
detail=f"Too many messages. Maximum {MAX_MESSAGES_PER_REQUEST} messages allowed per request"
)
# Check individual message lengths and total conversation length
total_length = 0
for i, message in enumerate(request.messages):
if not message.content:
raise HTTPException(status_code=400, detail=f"Message {i} has empty content")
msg_length = len(message.content)
if msg_length > MAX_MESSAGE_LENGTH:
raise HTTPException(
status_code=400,
detail=f"Message {i} is too long. Maximum {MAX_MESSAGE_LENGTH} characters allowed per message"
)
total_length += msg_length
if total_length > MAX_TOTAL_CONVERSATION_LENGTH:
raise HTTPException(
status_code=400,
detail=f"Total conversation is too long. Maximum {MAX_TOTAL_CONVERSATION_LENGTH} characters allowed"
)
# Validate role values
for i, message in enumerate(request.messages):
if message.role not in ["user", "assistant"]:
raise HTTPException(
status_code=400,
detail=f"Message {i} has invalid role. Must be 'user', 'assistant', or 'system'"
)
# Validate temperature
if request.temperature is not None:
if not (MIN_TEMPERATURE <= request.temperature <= MAX_TEMPERATURE):
raise HTTPException(
status_code=400,
detail=f"Temperature must be between {MIN_TEMPERATURE} and {MAX_TEMPERATURE}"
)
# Validate top_k
if request.top_k is not None:
if not (MIN_TOP_K <= request.top_k <= MAX_TOP_K):
raise HTTPException(
status_code=400,
detail=f"top_k must be between {MIN_TOP_K} and {MAX_TOP_K}"
)
# Validate max_tokens
if request.max_tokens is not None:
if not (MIN_MAX_TOKENS <= request.max_tokens <= MAX_MAX_TOKENS):
raise HTTPException(
status_code=400,
detail=f"max_tokens must be between {MIN_MAX_TOKENS} and {MAX_MAX_TOKENS}"
)
@asynccontextmanager @asynccontextmanager
async def lifespan(app: FastAPI): async def lifespan(app: FastAPI):
"""Load model on startup.""" """Load models on all GPUs on startup."""
print("Loading nanochat model...") print("Loading nanochat models across GPUs...")
app.state.model, app.state.tokenizer, _ = load_model(args.source, device, phase="eval", model_tag=args.model_tag, step=args.step) app.state.worker_pool = WorkerPool(num_gpus=args.num_gpus)
app.state.engine = Engine(app.state.model, app.state.tokenizer) await app.state.worker_pool.initialize(args.source, model_tag=args.model_tag, step=args.step)
print(f"Server ready at http://localhost:{args.port}") print(f"Server ready at http://localhost:{args.port}")
yield yield
@ -85,8 +260,7 @@ async def logo():
return FileResponse(logo_path, media_type="image/svg+xml") return FileResponse(logo_path, media_type="image/svg+xml")
async def generate_stream( async def generate_stream(
engine, worker: Worker,
tokenizer,
tokens, tokens,
temperature=None, temperature=None,
max_new_tokens=None, max_new_tokens=None,
@ -97,98 +271,141 @@ async def generate_stream(
max_new_tokens = max_new_tokens if max_new_tokens is not None else args.max_tokens max_new_tokens = max_new_tokens if max_new_tokens is not None else args.max_tokens
top_k = top_k if top_k is not None else args.top_k top_k = top_k if top_k is not None else args.top_k
assistant_end = tokenizer.encode_special("<|assistant_end|>") assistant_end = worker.tokenizer.encode_special("<|assistant_end|>")
bos = tokenizer.get_bos_token_id() bos = worker.tokenizer.get_bos_token_id()
with autocast_ctx: # Accumulate tokens to properly handle multi-byte UTF-8 characters (like emojis)
for token_column, token_masks in engine.generate( accumulated_tokens = []
# Track the last complete UTF-8 string (without replacement characters)
last_clean_text = ""
with worker.autocast_ctx:
for token_column, token_masks in worker.engine.generate(
tokens, tokens,
num_samples=1, num_samples=1,
max_tokens=max_new_tokens, max_tokens=max_new_tokens,
temperature=temperature, temperature=temperature,
top_k=top_k top_k=top_k,
seed=random.randint(0, 2**31 - 1)
): ):
token = token_column[0] token = token_column[0]
# Stopping criteria
if token == assistant_end or token == bos: if token == assistant_end or token == bos:
break break
token_text = tokenizer.decode([token]) # Append the token to sequence
yield f"data: {json.dumps({'token': token_text})}\n\n" accumulated_tokens.append(token)
# Decode all accumulated tokens to get proper UTF-8 handling
# Note that decode is a quite efficient operation, basically table lookup and string concat
current_text = worker.tokenizer.decode(accumulated_tokens)
# Only emit text if it doesn't end with a replacement character
# This ensures we don't emit incomplete UTF-8 sequences
if not current_text.endswith('<EFBFBD>'):
# Extract only the new text since last clean decode
new_text = current_text[len(last_clean_text):]
if new_text: # Only yield if there's new content
yield f"data: {json.dumps({'token': new_text, 'gpu': worker.gpu_id}, ensure_ascii=False)}\n\n"
last_clean_text = current_text
yield f"data: {json.dumps({'done': True})}\n\n" yield f"data: {json.dumps({'done': True})}\n\n"
@app.post("/chat/completions") @app.post("/chat/completions")
async def chat_completions(request: ChatRequest): async def chat_completions(request: ChatRequest):
"""Chat completion endpoint with streaming.""" """Chat completion endpoint (streaming only) - uses worker pool for multi-GPU."""
engine = app.state.engine
tokenizer = app.state.tokenizer
# Build conversation tokens # Basic validation to prevent abuse
bos = tokenizer.get_bos_token_id() validate_chat_request(request)
user_start = tokenizer.encode_special("<|user_start|>")
user_end = tokenizer.encode_special("<|user_end|>")
assistant_start = tokenizer.encode_special("<|assistant_start|>")
assistant_end = tokenizer.encode_special("<|assistant_end|>")
conversation_tokens = [bos] # Log incoming conversation to console
for message in request.messages: logger.info("="*20)
if message.role == "user": for i, message in enumerate(request.messages):
conversation_tokens.append(user_start) logger.info(f"[{message.role.upper()}]: {message.content}")
conversation_tokens.extend(tokenizer.encode(message.content)) logger.info("-"*20)
conversation_tokens.append(user_end)
elif message.role == "assistant":
conversation_tokens.append(assistant_start)
conversation_tokens.extend(tokenizer.encode(message.content))
conversation_tokens.append(assistant_end)
conversation_tokens.append(assistant_start) # Acquire a worker from the pool (will wait if all are busy)
worker_pool = app.state.worker_pool
worker = await worker_pool.acquire_worker()
try:
# Build conversation tokens
bos = worker.tokenizer.get_bos_token_id()
user_start = worker.tokenizer.encode_special("<|user_start|>")
user_end = worker.tokenizer.encode_special("<|user_end|>")
assistant_start = worker.tokenizer.encode_special("<|assistant_start|>")
assistant_end = worker.tokenizer.encode_special("<|assistant_end|>")
conversation_tokens = [bos]
for message in request.messages:
if message.role == "user":
conversation_tokens.append(user_start)
conversation_tokens.extend(worker.tokenizer.encode(message.content))
conversation_tokens.append(user_end)
elif message.role == "assistant":
conversation_tokens.append(assistant_start)
conversation_tokens.extend(worker.tokenizer.encode(message.content))
conversation_tokens.append(assistant_end)
conversation_tokens.append(assistant_start)
# Streaming response with worker release after completion
response_tokens = []
async def stream_and_release():
try:
async for chunk in generate_stream(
worker,
conversation_tokens,
temperature=request.temperature,
max_new_tokens=request.max_tokens,
top_k=request.top_k
):
# Accumulate response for logging
chunk_data = json.loads(chunk.replace("data: ", "").strip())
if "token" in chunk_data:
response_tokens.append(chunk_data["token"])
yield chunk
finally:
# Log the assistant response to console
full_response = "".join(response_tokens)
logger.info(f"[ASSISTANT] (GPU {worker.gpu_id}): {full_response}")
logger.info("="*20)
# Release worker back to pool after streaming is done
await worker_pool.release_worker(worker)
if request.stream:
return StreamingResponse( return StreamingResponse(
generate_stream( stream_and_release(),
engine,
tokenizer,
conversation_tokens,
temperature=request.temperature,
max_new_tokens=request.max_tokens,
top_k=request.top_k
),
media_type="text/event-stream" media_type="text/event-stream"
) )
else: except Exception as e:
# Non-streaming response # Make sure to release worker even on error
temperature = request.temperature if request.temperature is not None else args.temperature await worker_pool.release_worker(worker)
max_tokens = request.max_tokens if request.max_tokens is not None else args.max_tokens raise e
top_k = request.top_k if request.top_k is not None else args.top_k
with autocast_ctx:
result_tokens, masks = engine.generate_batch(
conversation_tokens,
num_samples=1,
max_tokens=max_tokens,
temperature=temperature,
top_k=top_k
)[0]
response_tokens = result_tokens[len(conversation_tokens):]
response_text = tokenizer.decode(response_tokens)
return {
"choices": [{
"message": {
"role": "assistant",
"content": response_text
},
"finish_reason": "stop"
}]
}
@app.get("/health") @app.get("/health")
async def health(): async def health():
"""Health check endpoint.""" """Health check endpoint."""
worker_pool = getattr(app.state, 'worker_pool', None)
return { return {
"status": "ok", "status": "ok",
"ready": hasattr(app.state, 'model') and app.state.model is not None "ready": worker_pool is not None and len(worker_pool.workers) > 0,
"num_gpus": worker_pool.num_gpus if worker_pool else 0,
"available_workers": worker_pool.available_workers.qsize() if worker_pool else 0
}
@app.get("/stats")
async def stats():
"""Get worker pool statistics."""
worker_pool = app.state.worker_pool
return {
"total_workers": len(worker_pool.workers),
"available_workers": worker_pool.available_workers.qsize(),
"busy_workers": len(worker_pool.workers) - worker_pool.available_workers.qsize(),
"workers": [
{
"gpu_id": w.gpu_id,
"device": str(w.device)
} for w in worker_pool.workers
]
} }
if __name__ == "__main__": if __name__ == "__main__":

76
scripts/mid_train.py
View File

@ -15,8 +15,8 @@ os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
import time import time
import wandb import wandb
import torch import torch
from contextlib import nullcontext
from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, get_base_dir from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, get_base_dir, autodetect_device_type
from nanochat.tokenizer import get_token_bytes from nanochat.tokenizer import get_token_bytes
from nanochat.checkpoint_manager import save_checkpoint from nanochat.checkpoint_manager import save_checkpoint
from nanochat.loss_eval import evaluate_bpb from nanochat.loss_eval import evaluate_bpb
@ -27,12 +27,16 @@ from tasks.common import TaskMixture
from tasks.gsm8k import GSM8K from tasks.gsm8k import GSM8K
from tasks.mmlu import MMLU from tasks.mmlu import MMLU
from tasks.smoltalk import SmolTalk from tasks.smoltalk import SmolTalk
from tasks.customjson import CustomJSON
from tasks.spellingbee import SimpleSpelling, SpellingBee
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
run = "dummy" # wandb run name default ("dummy" is special - we won't log to wandb) run = "dummy" # wandb run name default ("dummy" is special - we won't log to wandb)
device_type = "" # cuda|cpu|mps (empty => autodetect)
model_tag = None # model tag to load the model from (base model or midtrained model) model_tag = None # model tag to load the model from (base model or midtrained model)
step = None # step to load the model from (base model or midtrained model) step = None # step to load the model from (base model or midtrained model)
dtype = "bfloat16" dtype = "bfloat16"
num_iterations = -1 # explicit number of steps of the optimization (-1 = disable)
max_seq_len = 2048 max_seq_len = 2048
device_batch_size = 32 device_batch_size = 32
unembedding_lr = 0.004 unembedding_lr = 0.004
@ -40,20 +44,22 @@ embedding_lr = 0.2
matrix_lr = 0.02 matrix_lr = 0.02
init_lr_frac = 1.0 # initial learning rate is this fraction of the base learning rate init_lr_frac = 1.0 # initial learning rate is this fraction of the base learning rate
weight_decay = 0.0 weight_decay = 0.0
final_lr_frac = 0.0 # final LR is this fraction of the initial LR eval_every = 150 # -1 = disable
eval_every = 150
eval_tokens = 20*524288 eval_tokens = 20*524288
total_batch_size = 524288 total_batch_size = 524288
dry_run = 0 # dry_run=1 is for experiments: we will log to wandb but we won't write checkpoints or report
config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))] config_keys = [k for k,v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file exec(open(os.path.join('nanochat', 'configurator.py')).read()) # overrides from command line or config file
user_config = {k: globals()[k] for k in config_keys} # possibly useful for logging user_config = {k: globals()[k] for k in config_keys} # possibly useful for logging
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Compute init # Compute init
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init() device_type = autodetect_device_type() if device_type == "" else device_type
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type)
master_process = ddp_rank == 0 master_process = ddp_rank == 0
dtype = torch.float32 if dtype == 'float32' else torch.bfloat16 autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
autocast_ctx = torch.amp.autocast(device_type="cuda", dtype=dtype) synchronize = torch.cuda.synchronize if device_type == "cuda" else lambda: None
get_max_memory = torch.cuda.max_memory_allocated if device_type == "cuda" else lambda: 0
# wandb logging init # wandb logging init
use_dummy_wandb = run == "dummy" or not master_process use_dummy_wandb = run == "dummy" or not master_process
@ -88,11 +94,16 @@ for opt in optimizers:
# Midtraining data mixture and DataLoader # Midtraining data mixture and DataLoader
base_dir = get_base_dir() base_dir = get_base_dir()
identity_conversations_filepath = os.path.join(base_dir, "identity_conversations.jsonl")
train_dataset = TaskMixture([ train_dataset = TaskMixture([
SmolTalk(split="train"), # 460K rows of general conversations SmolTalk(split="train"), # 460K rows of general conversations
MMLU(subset="auxiliary_train", split="train"), # 100K rows of multiple choice problems drawn from ARC, MC_TEST, OBQA, RACE MMLU(subset="auxiliary_train", split="train"), # 100K rows of multiple choice problems drawn from ARC, MC_TEST, OBQA, RACE
GSM8K(subset="main", split="train"), # 8K rows teaching simple math and (calculator) tool use GSM8K(subset="main", split="train"), # 8K rows teaching simple math and (calculator) tool use
]) # total: 460K + 100K + 8K = 568K rows CustomJSON(filepath=identity_conversations_filepath), # 1000 rows of synthetic identity conversations
CustomJSON(filepath=identity_conversations_filepath), # let's do 2 epochs of these
SimpleSpelling(size=200000, split="train"), # 200K rows of Simple Spelling (e.g. spell the word 'apple')
SpellingBee(size=80000, split="train"), # 80K rows of Spelling Bee (e.g. how many 'r' are in 'strawberry'?)
]) # total: 460K + 100K + 8K + 200K + 80K = 848K rows
val_dataset = TaskMixture([ val_dataset = TaskMixture([
SmolTalk(split="test"), # 24K rows in test set SmolTalk(split="test"), # 24K rows in test set
MMLU(subset="all", split="test", stop=5200), # 14K rows in test set, use only 5.2K to match the train ratios MMLU(subset="all", split="test", stop=5200), # 14K rows in test set, use only 5.2K to match the train ratios
@ -111,8 +122,10 @@ def mid_data_generator(split):
assert dataset_size > 0 assert dataset_size > 0
needed_tokens = device_batch_size * max_seq_len + 1 # to form one training batch of inputs,targets needed_tokens = device_batch_size * max_seq_len + 1 # to form one training batch of inputs,targets
token_buffer = deque() token_buffer = deque()
scratch = torch.empty(needed_tokens, dtype=torch.int64, pin_memory=True) # CUDA supports memory pinning for faster transfers between CPU and GPU:
scratch = torch.empty(needed_tokens, dtype=torch.int64, pin_memory=(device_type == "cuda"))
cursor = ddp_rank # increments by ddp_world_size each time, so each rank processes unique documents cursor = ddp_rank # increments by ddp_world_size each time, so each rank processes unique documents
it = 0 # iteration counter
while True: while True:
# Accumulate enough tokens for one iteration before yielding # Accumulate enough tokens for one iteration before yielding
while len(token_buffer) < needed_tokens: while len(token_buffer) < needed_tokens:
@ -124,6 +137,10 @@ def mid_data_generator(split):
cursor -= dataset_size # wrap around for another epoch cursor -= dataset_size # wrap around for another epoch
if split == "train": if split == "train":
last_step = True # toggle last_step to True, which will terminate the training loop last_step = True # toggle last_step to True, which will terminate the training loop
# Stopping condition to respect num_iterations, if given
it += 1
if num_iterations > 0 and it >= num_iterations:
last_step = True # toggle last_step to True, which will terminate the training loop
# Build up inputs/targets and yield # Build up inputs/targets and yield
for i in range(needed_tokens): for i in range(needed_tokens):
scratch[i] = token_buffer.popleft() scratch[i] = token_buffer.popleft()
@ -132,7 +149,10 @@ def mid_data_generator(split):
inputs = inputs_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int32, non_blocking=True) inputs = inputs_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int32, non_blocking=True)
targets = targets_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int64, non_blocking=True) targets = targets_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int64, non_blocking=True)
if split == "train": if split == "train":
approx_progress = cursor / dataset_size # approximate progress as a fraction of the dataset if num_iterations > 0:
approx_progress = it / num_iterations # calculate progress from the max number of iterations
else:
approx_progress = cursor / dataset_size # approximate progress as a fraction of the dataset
yield inputs, targets yield inputs, targets
train_loader = mid_data_generator("train") train_loader = mid_data_generator("train")
@ -141,7 +161,8 @@ progress = 0 # will go from 0 to 1 over the course of the epoch
# Learning rate scheduler # Learning rate scheduler
def get_lr_multiplier(progress): def get_lr_multiplier(progress):
return progress * 1.0 + (1 - progress) * final_lr_frac # first 80% of training: no decay, then linearly ramp down to 0.
return 1 if progress < 0.8 else 1 - (progress - 0.8) / 0.2
# Momentum scheduler for Muon optimizer # Momentum scheduler for Muon optimizer
def get_muon_momentum(it): def get_muon_momentum(it):
@ -167,7 +188,7 @@ while True:
last_step = bool(last_step_tensor.item()) last_step = bool(last_step_tensor.item())
# once in a while: evaluate the val bpb (all ranks participate) # once in a while: evaluate the val bpb (all ranks participate)
if last_step or step % eval_every == 0: if eval_every > 0 and (last_step or step % eval_every == 0):
model.eval() model.eval()
val_loader = build_val_loader() val_loader = build_val_loader()
eval_steps = eval_tokens // (device_batch_size * max_seq_len * ddp_world_size) eval_steps = eval_tokens // (device_batch_size * max_seq_len * ddp_world_size)
@ -185,7 +206,7 @@ while True:
model.train() model.train()
# save checkpoint at the end of the run (only on master process) # save checkpoint at the end of the run (only on master process)
if master_process and last_step: if master_process and last_step and not dry_run:
output_dirname = f"d{depth}" # e.g. d12 output_dirname = f"d{depth}" # e.g. d12
checkpoint_dir = os.path.join(base_dir, "mid_checkpoints", output_dirname) checkpoint_dir = os.path.join(base_dir, "mid_checkpoints", output_dirname)
save_checkpoint( save_checkpoint(
@ -214,7 +235,7 @@ while True:
# ------------------------------------------------------------------------- # -------------------------------------------------------------------------
# single training step # single training step
# evaluate the gradient # evaluate the gradient
torch.cuda.synchronize() synchronize()
t0 = time.time() t0 = time.time()
for micro_step in range(grad_accum_steps): for micro_step in range(grad_accum_steps):
with autocast_ctx: with autocast_ctx:
@ -235,7 +256,7 @@ while True:
for opt in optimizers: for opt in optimizers:
opt.step() opt.step()
model.zero_grad(set_to_none=True) model.zero_grad(set_to_none=True)
torch.cuda.synchronize() synchronize()
t1 = time.time() t1 = time.time()
dt = t1 - t0 dt = t1 - t0
# ------------------------------------------------------------------------- # -------------------------------------------------------------------------
@ -267,22 +288,23 @@ while True:
}) })
# print a few more stats # print a few more stats
print0(f"Peak memory usage: {torch.cuda.max_memory_allocated() / 1024 / 1024:.2f}MiB") print0(f"Peak memory usage: {get_max_memory() / 1024 / 1024:.2f}MiB")
print0(f"Total training time: {total_training_time/60:.2f}m") print0(f"Total training time: {total_training_time/60:.2f}m")
print0(f"Minimum validation bpb: {min_val_bpb:.4f}") print0(f"Minimum validation bpb: {min_val_bpb:.4f}")
# Log to report # Log to report
from nanochat.report import get_report if not dry_run:
get_report().log(section="Midtraining", data=[ from nanochat.report import get_report
user_config, # CLI args get_report().log(section="Midtraining", data=[
{ # stats about the training setup user_config, # CLI args
"Number of iterations": step, { # stats about the training setup
"DDP world size": ddp_world_size, "Number of iterations": step,
}, "DDP world size": ddp_world_size,
{ # stats about training outcomes },
"Minimum validation bpb": min_val_bpb, { # stats about training outcomes
} "Minimum validation bpb": min_val_bpb,
]) }
])
# cleanup # cleanup
wandb_run.finish() # wandb run finish wandb_run.finish() # wandb run finish

8
speedrun.sh
View File

@ -12,7 +12,7 @@
# Default intermediate artifacts directory is in ~/.cache/nanochat # Default intermediate artifacts directory is in ~/.cache/nanochat
export OMP_NUM_THREADS=1 export OMP_NUM_THREADS=1
NANOCHAT_BASE_DIR="$HOME/.cache/nanochat" export NANOCHAT_BASE_DIR="$HOME/.cache/nanochat"
mkdir -p $NANOCHAT_BASE_DIR mkdir -p $NANOCHAT_BASE_DIR
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
@ -23,7 +23,7 @@ command -v uv &> /dev/null || curl -LsSf https://astral.sh/uv/install.sh | sh
# create a .venv local virtual environment (if it doesn't exist) # create a .venv local virtual environment (if it doesn't exist)
[ -d ".venv" ] || uv venv [ -d ".venv" ] || uv venv
# install the repo dependencies # install the repo dependencies
uv sync uv sync --extra gpu
# activate venv so that `python` uses the project's venv instead of system python # activate venv so that `python` uses the project's venv instead of system python
source .venv/bin/activate source .venv/bin/activate
@ -101,6 +101,10 @@ torchrun --standalone --nproc_per_node=8 -m scripts.base_eval
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# Midtraining (teach the model conversation special tokens, tool use, multiple choice) # Midtraining (teach the model conversation special tokens, tool use, multiple choice)
# download 2.3MB of synthetic identity conversations to impart a personality to nanochat
# see dev/gen_sft_data.py for details on how this data was prepared and to get a sense of how you can easily tune it
curl -L -o $NANOCHAT_BASE_DIR/identity_conversations.jsonl https://karpathy-public.s3.us-west-2.amazonaws.com/identity_conversations.jsonl
# run midtraining and eval the model # run midtraining and eval the model
torchrun --standalone --nproc_per_node=8 -m scripts.mid_train -- --run=$WANDB_RUN torchrun --standalone --nproc_per_node=8 -m scripts.mid_train -- --run=$WANDB_RUN
torchrun --standalone --nproc_per_node=8 -m scripts.chat_eval -- -i mid torchrun --standalone --nproc_per_node=8 -m scripts.chat_eval -- -i mid

65
tasks/customjson.py Normal file
View File

@ -0,0 +1,65 @@
"""
CustomJSON task for loading conversations from JSONL files.
Each line in the JSONL file should be a JSON array of messages.
"""
import os
import json
from tasks.common import Task
class CustomJSON(Task):
"""
Load conversations from a JSONL file.
Each line should be a JSON array of message objects with 'role' and 'content' fields.
Example line: [{"role":"user","content":"Hi"},{"role":"assistant","content":"Hello"}]
"""
def __init__(self, filepath, **kwargs):
super().__init__(**kwargs)
self.filepath = filepath
self.conversations = []
# Load all conversations from the JSONL file
if not os.path.exists(filepath):
# Helpful error message due to recent change. Will be removed in the future.
print("-" * 80)
print(f"Warning: File {filepath} does not exist")
print("HINT (Oct 21 2025)")
print("If you recently did a git pull and suddely see this, it might be due to the new addition of identity conversations")
print("See this discussion for more details: https://github.com/karpathy/nanochat/discussions/139")
print("Quick fix: simply run the following command to download the file and you're done:")
print(f"curl -L -o {filepath} https://karpathy-public.s3.us-west-2.amazonaws.com/identity_conversations.jsonl")
print("-" * 80)
else:
with open(filepath, 'r') as f:
for line in f:
line = line.strip()
if not line: # skip empty lines
continue
messages = json.loads(line)
# Validate the conversation structure
assert isinstance(messages, list), f"Expected list of messages, got {type(messages)}"
assert len(messages) >= 2, f"Conversation must have at least 2 messages, got {len(messages)}"
# Validate message structure and alternating roles
for i, message in enumerate(messages):
assert "role" in message, f"Message {i} missing 'role' field"
assert "content" in message, f"Message {i} missing 'content' field"
expected_role = "user" if i % 2 == 0 else "assistant"
assert message["role"] == expected_role, f"Message {i} has role {message['role']} but should be {expected_role}"
assert isinstance(message["content"], str), f"Message {i} content must be a string"
self.conversations.append(messages)
self.length = len(self.conversations)
def num_examples(self):
return self.length
def get_example(self, index):
messages = self.conversations[index]
conversation = {
"messages": messages,
}
return conversation

305
tasks/spellingbee.py Normal file
View File

@ -0,0 +1,305 @@
"""
Task intended to make nanochat better in spelling and counting, for example:
"How many r are in strawberry?" -> 3
An interesting part of this task is that we will get the assistant to
solve the problem using a combination of manual counting and Python.
This is a good problem solving "instinct" to mix into the model and RL
may further refine it to trust one over the other. If we were extra fancy
(which we could/should be) we'd add small errors here and there to allow
the model also learn recoveries. We can do this in future versions.
There are two tasks in this file:
1. SpellingBee: Counting the number of occurrences of a letter in a word
2. SimpleSpelling: Simply spelling words
(1) is the goal, but (2) exists as a highly condensed version of the part
that makes (1) difficult, which is word spelling. This is non-trivial for an
LLM because it has to learn how every token (a little semantic chunk/atom)
maps to the sequence of individual characters that make it up. Larger models
learn this eventually on their own, but if we want this capability to exist
in smaller models, we have to actively encourage it by over-representing it
in the training data. Midtraining is a good place to do this.
To preview a few example conversations, run:
python -m tasks.spellingbee
"""
import re
import random
from tasks.common import Task
from nanochat.common import download_file_with_lock
# Letters of the alphabet
LETTERS = "abcdefghijklmnopqrstuvwxyz"
# A list of 370K English words of large variety
WORD_LIST_URL = "https://raw.githubusercontent.com/dwyl/english-words/refs/heads/master/words_alpha.txt"
# Identical to gsm8k's answer extraction
ANSWER_RE = re.compile(r"#### (\-?[0-9\.\,]+)")
def extract_answer(completion):
"""
Extract the numerical answer after #### marker.
"""
match = ANSWER_RE.search(completion)
if match:
match_str = match.group(1).strip()
match_str = match_str.replace(",", "")
return match_str
return None
# User message templates for data augmentation
USER_MSG_TEMPLATES = [
"How many {letter} are in the word {word}",
"How many {letter} are in {word}",
"Count the number of {letter} in {word}",
"How many times does {letter} appear in {word}",
"What's the count of {letter} in {word}",
"In the word {word}, how many {letter} are there",
"How many letter {letter} are in the word {word}",
"Count how many {letter} appear in {word}",
"Tell me the number of {letter} in {word}",
"How many occurrences of {letter} are in {word}",
"Find the count of {letter} in {word}",
"Can you count the {letter} letters in {word}",
"What is the frequency of {letter} in {word}",
"How many {letter}s are in {word}",
"How many {letter}'s are in {word}",
"Count all the {letter} in {word}",
"How many times is {letter} in {word}",
"Number of {letter} in {word}",
"Total count of {letter} in {word}",
"How many {letter} does {word} have",
"How many {letter} does {word} contain",
"What's the number of {letter} in {word}",
"{word} has how many {letter}",
"In {word}, count the {letter}",
"How many {letter} appear in {word}",
"Count the {letter} in {word}",
"Give me the count of {letter} in {word}",
"How many instances of {letter} in {word}",
"Show me how many {letter} are in {word}",
"Calculate the number of {letter} in {word}",
# Spanish
"¿Cuántas {letter} hay en {word}?",
"¿Cuántas veces aparece {letter} en {word}?",
"Cuenta las {letter} en {word}",
"¿Cuántas letras {letter} tiene {word}?",
# Chinese (Simplified)
"{word}中有多少个{letter}",
"{word}里有几个{letter}",
"数一下{word}中的{letter}",
"{word}这个词里有多少{letter}",
# Korean
"{word}에 {letter}가 몇 개 있나요",
"{word}에서 {letter}의 개수는",
"{word}에 {letter}가 몇 번 나오나요",
"{word}라는 단어에 {letter}가 몇 개",
# French
"Combien de {letter} dans {word}",
"Combien de fois {letter} apparaît dans {word}",
"Compte les {letter} dans {word}",
# German
"Wie viele {letter} sind in {word}",
"Wie oft kommt {letter} in {word} vor",
"Zähle die {letter} in {word}",
# Japanese
"{word}に{letter}は何個ありますか",
"{word}の中に{letter}がいくつ",
"{word}に{letter}が何回出てくる",
]
class SpellingBee(Task):
def __init__(self, size=1000, split="train", **kwargs):
super().__init__(**kwargs)
assert split in ["train", "test"], "SpellingBee split must be train|test"
self.size = size
self.split = split
filename = WORD_LIST_URL.split("/")[-1]
word_list_path = download_file_with_lock(WORD_LIST_URL, filename)
with open(word_list_path) as f:
words = [line.strip() for line in f]
self.words = words
@property
def eval_type(self):
return 'generative'
def num_examples(self):
return self.size
def get_example(self, index):
seed = index if self.split == "train" else -(index + 1) # avoid collision at 0
rng = random.Random(seed)
# pick a random word
word = rng.choice(self.words)
# pick a letter from it (90%) or a random letter (10%)
letter = rng.choice(word) if rng.random() < 0.9 else rng.choice(LETTERS)
# get the correct answer by simply counting
count = word.count(letter)
# create a user message, with a bunch of variations as data augmentation
template = rng.choice(USER_MSG_TEMPLATES)
# 30% chance to lowercase the template (lazy people don't use shift)
if rng.random() < 0.3:
template = template.lower()
quote_options = ['', "'", '"']
letter_quote = rng.choice(quote_options) # is the letter quoted?
word_quote = rng.choice(quote_options) # is the word quoted?
letter_wrapped = f"{letter_quote}{letter}{letter_quote}"
word_wrapped = f"{word_quote}{word}{word_quote}"
user_msg = template.format(letter=letter_wrapped, word=word_wrapped)
if rng.random() < 0.5: # 50% of people don't even use question marks
user_msg += "?"
# Now create the ideal assistant response - build as parts (text + tool calls)
assistant_parts = []
word_letters = ",".join(list(word))
manual_text = f"""We are asked to find the number '{letter}' in the word '{word}'. Let me try a manual approach first.
First spell the word out:
{word}:{word_letters}
Then count the occurrences of '{letter}':
"""
# Little simulated loop of the solution process
# TODO: This is where the fun starts, we could simulate cute little mistakes
# and get the model to review its work and recover from them.
# You might of course hope this could arise in RL too, but realistically you'd want to help it out a bit.
running_count = 0
for i, char in enumerate(word, 1):
if char == letter:
running_count += 1
# note: there deliberately cannot be a space here between i and char
# because this would create a different token! (e.g. " a" and "a" are different tokens)
manual_text += f"{i}:{char} hit! count={running_count}\n"
else:
manual_text += f"{i}:{char}\n"
manual_text += f"\nThis gives us {running_count}."
assistant_parts.append({"type": "text", "text": manual_text})
# Part 2: Python verification
assistant_parts.append({"type": "text", "text": "\n\nLet me double check this using Python:\n\n"})
# Part 3: Python tool call
python_expr = f"'{word}'.count('{letter}')"
assistant_parts.append({"type": "python", "text": python_expr})
# Part 4: Python output
assistant_parts.append({"type": "python_output", "text": str(count)})
# Part 5: Final answer
assistant_parts.append({"type": "text", "text": f"\n\nPython gives us {count}.\n\nMy final answer is:\n\n#### {count}"})
# return the full conversation
messages = [
{"role": "user", "content": user_msg},
{"role": "assistant", "content": assistant_parts}
]
conversation = {
"messages": messages,
}
return conversation
def evaluate(self, conversation, assistant_response):
"""
Given (conversation, completion), return evaluation outcome (0 = wrong, 1 = correct)
Identical to gsm8k's evaluation.
"""
assert isinstance(assistant_response, str), "Assuming simple string response for now"
# First extract the ground truth answer from the conversation
assistant_message = conversation['messages'][-1]
assert assistant_message['role'] == "assistant", "Last message must be from the Assistant"
assert isinstance(assistant_message['content'], list), "This is expected to be a list of parts"
# The last text part contains the final answer with ####
last_text_part = assistant_message['content'][-1]['text']
# Extract both the ground truth answer and the predicted answer
ref_num = extract_answer(last_text_part)
pred_num = extract_answer(assistant_response)
# Compare and return the success as int
is_correct = int(pred_num == ref_num)
return is_correct
def reward(self, conversation, assistant_response):
""" Use simple 0-1 reward just like gsm8k."""
is_correct = self.evaluate(conversation, assistant_response)
is_correct_float = float(is_correct)
return is_correct_float
class SimpleSpelling(Task):
"""Much simpler task designed to get the model to just practice spelling words."""
def __init__(self, size=1000, split="train", **kwargs):
super().__init__(**kwargs)
assert split in ["train", "test"], "SpellingBee split must be train|test"
self.size = size
self.split = split
filename = WORD_LIST_URL.split("/")[-1]
word_list_path = download_file_with_lock(WORD_LIST_URL, filename)
with open(word_list_path) as f:
words = [line.strip() for line in f]
rng = random.Random(42)
rng.shuffle(words) # use a different word order than the SpellingBee task
self.words = words
@property
def eval_type(self):
return 'generative'
def num_examples(self):
return self.size
def get_example(self, index):
seed = index if self.split == "train" else -(index + 1) # avoid collision at 0
rng = random.Random(seed)
# pick a random word
word = rng.choice(self.words)
word_letters = ",".join(list(word))
# return the full conversation
messages = [
{"role": "user", "content": f"Spell the word: {word}"},
{"role": "assistant", "content": f"{word}:{word_letters}"}
]
conversation = {
"messages": messages,
}
return conversation
if __name__ == "__main__":
# preview the SpellingBee task, first 10 examples
task = SpellingBee()
for i in range(10):
ex = task.get_example(i)
print("=" * 100)
print(ex['messages'][0]['content'])
print("-" * 100)
# Assistant content is now a list of parts
assistant_parts = ex['messages'][1]['content']
for part in assistant_parts:
if part['type'] == 'text':
print(part['text'], end='')
elif part['type'] == 'python':
print(f"<<{part['text']}=", end='')
elif part['type'] == 'python_output':
print(f"{part['text']}>>", end='')
print()
print("-" * 100)
# # preview the SimpleSpelling task, first 10 examples
# task = SimpleSpelling()
# for i in range(10):
# ex = task.get_example(i)
# print("=" * 100)
# print(ex['messages'][0]['content'])
# print("-" * 100)
# print(ex['messages'][1]['content'])
# # also scrutinize the tokenization (last example only)
# from nanochat.tokenizer import get_tokenizer
# tokenizer = get_tokenizer()
# ids, mask = tokenizer.render_conversation(ex)
# print(tokenizer.visualize_tokenization(ids, mask, with_token_id=True))

66
tests/test_engine.py Normal file
View File

@ -0,0 +1,66 @@
"""
Test Engine class. Example run:
python -m pytest tests/test_engine.py -v
"""
import torch
from nanochat.engine import KVCache
def test_kv_cache_resize():
"""
The KV cache was not resized correctly, more information here:
https://github.com/karpathy/nanochat/pull/186
This test reproduces the issue and will be merged alongside the fix.
"""
batch_size = 2
num_heads = 3
seq_len = 4
head_dim = 5
num_layers = 6
kv_cache = KVCache(
batch_size=batch_size,
num_heads=num_heads,
seq_len=seq_len,
head_dim=head_dim,
num_layers=num_layers
)
# Insert a single token with a distinct fill value to all layers
def insert_token(token_idx):
for layer_idx in range(num_layers):
k = torch.full((batch_size, num_heads, 1, head_dim), fill_value=float(token_idx), dtype=torch.float32)
v = torch.full((batch_size, num_heads, 1, head_dim), fill_value=float(token_idx * 100), dtype=torch.float32)
kv_cache.insert_kv(layer_idx, k, v)
# Insert 4 tokens (fills the initial seq_len=4)
for i in range(4):
insert_token(i)
# Record the original state of the cache
original_cache = kv_cache.kv_cache.clone()
original_seq_len = original_cache.shape[4]
# Insert the 5th token, which will trigger a resize
insert_token(4)
# Verify that the cache actually resized
new_seq_len = kv_cache.kv_cache.shape[4]
assert new_seq_len > original_seq_len, f"Cache did not resize: original seq_len={original_seq_len}, new seq_len={new_seq_len}"
# Verify that the original 4 tokens are still intact after resize
for layer_idx in range(num_layers):
for token_idx in range(4):
# Check that resized cache matches expected values
expected_k = float(token_idx)
expected_v = float(token_idx * 100)
actual_k = kv_cache.kv_cache[layer_idx, 0, :, :, token_idx, :]
actual_v = kv_cache.kv_cache[layer_idx, 1, :, :, token_idx, :]
assert (actual_k == expected_k).all(), f"Layer {layer_idx}, token {token_idx}: key corrupted, expected {expected_k}"
assert (actual_v == expected_v).all(), f"Layer {layer_idx}, token {token_idx}: value corrupted, expected {expected_v}"
# And that the original cache matches resized cache
original_k = original_cache[layer_idx, 0, :, :, token_idx, :]
original_v = original_cache[layer_idx, 1, :, :, token_idx, :]
assert (actual_k == original_k).all(), f"Layer {layer_idx}, token {token_idx}: key doesn't match original"
assert (actual_v == original_v).all(), f"Layer {layer_idx}, token {token_idx}: value doesn't match original"

334
uv.lock
View File

@ -2,13 +2,32 @@ version = 1
revision = 3 revision = 3
requires-python = ">=3.10" requires-python = ">=3.10"
resolution-markers = [ resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'linux'", "python_full_version >= '3.12' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'linux'", "python_full_version >= '3.12' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux'", "python_full_version == '3.11.*' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'linux'", "python_full_version < '3.11' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'linux'", "python_full_version == '3.11.*' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'linux'", "python_full_version < '3.11' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
] ]
conflicts = [[
{ package = "nanochat", extra = "cpu" },
{ package = "nanochat", extra = "gpu" },
]]
[[package]] [[package]]
name = "aiohappyeyeballs" name = "aiohappyeyeballs"
@ -26,7 +45,7 @@ source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "aiohappyeyeballs" }, { name = "aiohappyeyeballs" },
{ name = "aiosignal" }, { name = "aiosignal" },
{ name = "async-timeout", marker = "python_full_version < '3.11'" }, { name = "async-timeout", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "attrs" }, { name = "attrs" },
{ name = "frozenlist" }, { name = "frozenlist" },
{ name = "multidict" }, { name = "multidict" },
@ -111,7 +130,7 @@ version = "1.4.0"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "frozenlist" }, { name = "frozenlist" },
{ name = "typing-extensions", marker = "python_full_version < '3.13'" }, { name = "typing-extensions", marker = "python_full_version < '3.13' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" } sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" }
wheels = [ wheels = [
@ -132,10 +151,10 @@ name = "anyio"
version = "4.10.0" version = "4.10.0"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "exceptiongroup", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "idna" }, { name = "idna" },
{ name = "sniffio" }, { name = "sniffio" },
{ name = "typing-extensions", marker = "python_full_version < '3.13'" }, { name = "typing-extensions", marker = "python_full_version < '3.13' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/f1/b4/636b3b65173d3ce9a38ef5f0522789614e590dab6a8d505340a4efe4c567/anyio-4.10.0.tar.gz", hash = "sha256:3f3fae35c96039744587aa5b8371e7e8e603c0702999535961dd336026973ba6", size = 213252, upload-time = "2025-08-04T08:54:26.451Z" } sdist = { url = "https://files.pythonhosted.org/packages/f1/b4/636b3b65173d3ce9a38ef5f0522789614e590dab6a8d505340a4efe4c567/anyio-4.10.0.tar.gz", hash = "sha256:3f3fae35c96039744587aa5b8371e7e8e603c0702999535961dd336026973ba6", size = 213252, upload-time = "2025-08-04T08:54:26.451Z" }
wheels = [ wheels = [
@ -238,7 +257,7 @@ name = "click"
version = "8.2.1" version = "8.2.1"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "colorama", marker = "sys_platform == 'win32'" }, { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/60/6c/8ca2efa64cf75a977a0d7fac081354553ebe483345c734fb6b6515d96bbc/click-8.2.1.tar.gz", hash = "sha256:27c491cc05d968d271d5a1db13e3b5a184636d9d930f148c50b038f0d0646202", size = 286342, upload-time = "2025-05-20T23:19:49.832Z" } sdist = { url = "https://files.pythonhosted.org/packages/60/6c/8ca2efa64cf75a977a0d7fac081354553ebe483345c734fb6b6515d96bbc/click-8.2.1.tar.gz", hash = "sha256:27c491cc05d968d271d5a1db13e3b5a184636d9d930f148c50b038f0d0646202", size = 286342, upload-time = "2025-05-20T23:19:49.832Z" }
wheels = [ wheels = [
@ -292,7 +311,7 @@ name = "exceptiongroup"
version = "1.3.0" version = "1.3.0"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "typing-extensions", marker = "python_full_version < '3.11'" }, { name = "typing-extensions", marker = "python_full_version < '3.12' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/0b/9f/a65090624ecf468cdca03533906e7c69ed7588582240cfe7cc9e770b50eb/exceptiongroup-1.3.0.tar.gz", hash = "sha256:b241f5885f560bc56a59ee63ca4c6a8bfa46ae4ad651af316d4e81817bb9fd88", size = 29749, upload-time = "2025-05-10T17:42:51.123Z" } sdist = { url = "https://files.pythonhosted.org/packages/0b/9f/a65090624ecf468cdca03533906e7c69ed7588582240cfe7cc9e770b50eb/exceptiongroup-1.3.0.tar.gz", hash = "sha256:b241f5885f560bc56a59ee63ca4c6a8bfa46ae4ad651af316d4e81817bb9fd88", size = 29749, upload-time = "2025-05-10T17:42:51.123Z" }
wheels = [ wheels = [
@ -497,7 +516,7 @@ source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "filelock" }, { name = "filelock" },
{ name = "fsspec" }, { name = "fsspec" },
{ name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64'" }, { name = "hf-xet", marker = "platform_machine == 'aarch64' or platform_machine == 'amd64' or platform_machine == 'arm64' or platform_machine == 'x86_64' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "packaging" }, { name = "packaging" },
{ name = "pyyaml" }, { name = "pyyaml" },
{ name = "requests" }, { name = "requests" },
@ -602,7 +621,7 @@ name = "maturin"
version = "1.9.4" version = "1.9.4"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "tomli", marker = "python_full_version < '3.11'" }, { name = "tomli", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/13/7c/b11b870fc4fd84de2099906314ce45488ae17be32ff5493519a6cddc518a/maturin-1.9.4.tar.gz", hash = "sha256:235163a0c99bc6f380fb8786c04fd14dcf6cd622ff295ea3de525015e6ac40cf", size = 213647, upload-time = "2025-08-27T11:37:57.079Z" } sdist = { url = "https://files.pythonhosted.org/packages/13/7c/b11b870fc4fd84de2099906314ce45488ae17be32ff5493519a6cddc518a/maturin-1.9.4.tar.gz", hash = "sha256:235163a0c99bc6f380fb8786c04fd14dcf6cd622ff295ea3de525015e6ac40cf", size = 213647, upload-time = "2025-08-27T11:37:57.079Z" }
wheels = [ wheels = [
@ -635,7 +654,7 @@ name = "multidict"
version = "6.6.4" version = "6.6.4"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "typing-extensions", marker = "python_full_version < '3.11'" }, { name = "typing-extensions", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/69/7f/0652e6ed47ab288e3756ea9c0df8b14950781184d4bd7883f4d87dd41245/multidict-6.6.4.tar.gz", hash = "sha256:d2d4e4787672911b48350df02ed3fa3fffdc2f2e8ca06dd6afdf34189b76a9dd", size = 101843, upload-time = "2025-08-11T12:08:48.217Z" } sdist = { url = "https://files.pythonhosted.org/packages/69/7f/0652e6ed47ab288e3756ea9c0df8b14950781184d4bd7883f4d87dd41245/multidict-6.6.4.tar.gz", hash = "sha256:d2d4e4787672911b48350df02ed3fa3fffdc2f2e8ca06dd6afdf34189b76a9dd", size = 101843, upload-time = "2025-08-11T12:08:48.217Z" }
wheels = [ wheels = [
@ -761,13 +780,26 @@ dependencies = [
{ name = "numpy" }, { name = "numpy" },
{ name = "psutil" }, { name = "psutil" },
{ name = "regex" }, { name = "regex" },
{ name = "setuptools" },
{ name = "tiktoken" }, { name = "tiktoken" },
{ name = "tokenizers" }, { name = "tokenizers" },
{ name = "torch" }, { name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "torch", version = "2.9.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "torch", version = "2.9.0", source = { registry = "https://pypi.org/simple" }, marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
{ name = "torch", version = "2.9.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "uvicorn" }, { name = "uvicorn" },
{ name = "wandb" }, { name = "wandb" },
] ]
[package.optional-dependencies]
cpu = [
{ name = "torch", version = "2.9.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "torch", version = "2.9.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
]
gpu = [
{ name = "torch", version = "2.8.0+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" } },
]
[package.dev-dependencies] [package.dev-dependencies]
dev = [ dev = [
{ name = "maturin" }, { name = "maturin" },
@ -782,12 +814,16 @@ requires-dist = [
{ name = "numpy", specifier = "==1.26.4" }, { name = "numpy", specifier = "==1.26.4" },
{ name = "psutil", specifier = ">=7.1.0" }, { name = "psutil", specifier = ">=7.1.0" },
{ name = "regex", specifier = ">=2025.9.1" }, { name = "regex", specifier = ">=2025.9.1" },
{ name = "setuptools", specifier = ">=80.9.0" },
{ name = "tiktoken", specifier = ">=0.11.0" }, { name = "tiktoken", specifier = ">=0.11.0" },
{ name = "tokenizers", specifier = ">=0.22.0" }, { name = "tokenizers", specifier = ">=0.22.0" },
{ name = "torch", specifier = ">=2.8.0", index = "https://download.pytorch.org/whl/cu128" }, { name = "torch", specifier = ">=2.8.0" },
{ name = "torch", marker = "extra == 'cpu'", specifier = ">=2.8.0", index = "https://download.pytorch.org/whl/cpu", conflict = { package = "nanochat", extra = "cpu" } },
{ name = "torch", marker = "extra == 'gpu'", specifier = ">=2.8.0", index = "https://download.pytorch.org/whl/cu128", conflict = { package = "nanochat", extra = "gpu" } },
{ name = "uvicorn", specifier = ">=0.36.0" }, { name = "uvicorn", specifier = ">=0.36.0" },
{ name = "wandb", specifier = ">=0.21.3" }, { name = "wandb", specifier = ">=0.21.3" },
] ]
provides-extras = ["cpu", "gpu"]
[package.metadata.requires-dev] [package.metadata.requires-dev]
dev = [ dev = [
@ -800,8 +836,13 @@ name = "networkx"
version = "3.4.2" version = "3.4.2"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
resolution-markers = [ resolution-markers = [
"python_full_version < '3.11' and sys_platform == 'linux'", "python_full_version < '3.11' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'linux'", "python_full_version < '3.11' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version < '3.11' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
] ]
sdist = { url = "https://files.pythonhosted.org/packages/fd/1d/06475e1cd5264c0b870ea2cc6fdb3e37177c1e565c43f56ff17a10e3937f/networkx-3.4.2.tar.gz", hash = "sha256:307c3669428c5362aab27c8a1260aa8f47c4e91d3891f48be0141738d8d053e1", size = 2151368, upload-time = "2024-10-21T12:39:38.695Z" } sdist = { url = "https://files.pythonhosted.org/packages/fd/1d/06475e1cd5264c0b870ea2cc6fdb3e37177c1e565c43f56ff17a10e3937f/networkx-3.4.2.tar.gz", hash = "sha256:307c3669428c5362aab27c8a1260aa8f47c4e91d3891f48be0141738d8d053e1", size = 2151368, upload-time = "2024-10-21T12:39:38.695Z" }
wheels = [ wheels = [
@ -813,10 +854,20 @@ name = "networkx"
version = "3.5" version = "3.5"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
resolution-markers = [ resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'linux'", "python_full_version >= '3.12' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'linux'", "python_full_version >= '3.12' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux'", "python_full_version == '3.11.*' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'linux'", "python_full_version == '3.11.*' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version >= '3.12' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform == 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
"python_full_version == '3.11.*' and sys_platform != 'linux' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu'",
] ]
sdist = { url = "https://files.pythonhosted.org/packages/6c/4f/ccdb8ad3a38e583f214547fd2f7ff1fc160c43a75af88e6aec213404b96a/networkx-3.5.tar.gz", hash = "sha256:d4c6f9cf81f52d69230866796b82afbccdec3db7ae4fbd1b65ea750feed50037", size = 2471065, upload-time = "2025-05-29T11:35:07.804Z" } sdist = { url = "https://files.pythonhosted.org/packages/6c/4f/ccdb8ad3a38e583f214547fd2f7ff1fc160c43a75af88e6aec213404b96a/networkx-3.5.tar.gz", hash = "sha256:d4c6f9cf81f52d69230866796b82afbccdec3db7ae4fbd1b65ea750feed50037", size = 2471065, upload-time = "2025-05-29T11:35:07.804Z" }
wheels = [ wheels = [
@ -860,7 +911,9 @@ name = "nvidia-cublas-cu12"
version = "12.8.4.1" version = "12.8.4.1"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/29/99/db44d685f0e257ff0e213ade1964fc459b4a690a73293220e98feb3307cf/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b86f6dd8935884615a0683b663891d43781b819ac4f2ba2b0c9604676af346d0", size = 590537124, upload-time = "2025-03-07T01:43:53.556Z" },
{ url = "https://files.pythonhosted.org/packages/dc/61/e24b560ab2e2eaeb3c839129175fb330dfcfc29e5203196e5541a4c44682/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142", size = 594346921, upload-time = "2025-03-07T01:44:31.254Z" }, { url = "https://files.pythonhosted.org/packages/dc/61/e24b560ab2e2eaeb3c839129175fb330dfcfc29e5203196e5541a4c44682/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142", size = 594346921, upload-time = "2025-03-07T01:44:31.254Z" },
{ url = "https://files.pythonhosted.org/packages/70/61/7d7b3c70186fb651d0fbd35b01dbfc8e755f69fd58f817f3d0f642df20c3/nvidia_cublas_cu12-12.8.4.1-py3-none-win_amd64.whl", hash = "sha256:47e9b82132fa8d2b4944e708049229601448aaad7e6f296f630f2d1a32de35af", size = 567544208, upload-time = "2025-03-07T01:53:30.535Z" },
] ]
[[package]] [[package]]
@ -868,7 +921,9 @@ name = "nvidia-cuda-cupti-cu12"
version = "12.8.90" version = "12.8.90"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/d5/1f/b3bd73445e5cb342727fd24fe1f7b748f690b460acadc27ea22f904502c8/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:4412396548808ddfed3f17a467b104ba7751e6b58678a4b840675c56d21cf7ed", size = 9533318, upload-time = "2025-03-07T01:40:10.421Z" },
{ url = "https://files.pythonhosted.org/packages/f8/02/2adcaa145158bf1a8295d83591d22e4103dbfd821bcaf6f3f53151ca4ffa/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182", size = 10248621, upload-time = "2025-03-07T01:40:21.213Z" }, { url = "https://files.pythonhosted.org/packages/f8/02/2adcaa145158bf1a8295d83591d22e4103dbfd821bcaf6f3f53151ca4ffa/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182", size = 10248621, upload-time = "2025-03-07T01:40:21.213Z" },
{ url = "https://files.pythonhosted.org/packages/41/bc/83f5426095d93694ae39fe1311431b5d5a9bb82e48bf0dd8e19be2765942/nvidia_cuda_cupti_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:bb479dcdf7e6d4f8b0b01b115260399bf34154a1a2e9fe11c85c517d87efd98e", size = 7015759, upload-time = "2025-03-07T01:51:11.355Z" },
] ]
[[package]] [[package]]
@ -877,6 +932,8 @@ version = "12.8.93"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/05/6b/32f747947df2da6994e999492ab306a903659555dddc0fbdeb9d71f75e52/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994", size = 88040029, upload-time = "2025-03-07T01:42:13.562Z" }, { url = "https://files.pythonhosted.org/packages/05/6b/32f747947df2da6994e999492ab306a903659555dddc0fbdeb9d71f75e52/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994", size = 88040029, upload-time = "2025-03-07T01:42:13.562Z" },
{ url = "https://files.pythonhosted.org/packages/eb/d1/e50d0acaab360482034b84b6e27ee83c6738f7d32182b987f9c7a4e32962/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:fc1fec1e1637854b4c0a65fb9a8346b51dd9ee69e61ebaccc82058441f15bce8", size = 43106076, upload-time = "2025-03-07T01:41:59.817Z" },
{ url = "https://files.pythonhosted.org/packages/45/51/52a3d84baa2136cc8df15500ad731d74d3a1114d4c123e043cb608d4a32b/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:7a4b6b2904850fe78e0bd179c4b655c404d4bb799ef03ddc60804247099ae909", size = 73586838, upload-time = "2025-03-07T01:52:13.483Z" },
] ]
[[package]] [[package]]
@ -884,7 +941,9 @@ name = "nvidia-cuda-runtime-cu12"
version = "12.8.90" version = "12.8.90"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/7c/75/f865a3b236e4647605ea34cc450900854ba123834a5f1598e160b9530c3a/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:52bf7bbee900262ffefe5e9d5a2a69a30d97e2bc5bb6cc866688caa976966e3d", size = 965265, upload-time = "2025-03-07T01:39:43.533Z" },
{ url = "https://files.pythonhosted.org/packages/0d/9b/a997b638fcd068ad6e4d53b8551a7d30fe8b404d6f1804abf1df69838932/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90", size = 954765, upload-time = "2025-03-07T01:40:01.615Z" }, { url = "https://files.pythonhosted.org/packages/0d/9b/a997b638fcd068ad6e4d53b8551a7d30fe8b404d6f1804abf1df69838932/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90", size = 954765, upload-time = "2025-03-07T01:40:01.615Z" },
{ url = "https://files.pythonhosted.org/packages/30/a5/a515b7600ad361ea14bfa13fb4d6687abf500adc270f19e89849c0590492/nvidia_cuda_runtime_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:c0c6027f01505bfed6c3b21ec546f69c687689aad5f1a377554bc6ca4aa993a8", size = 944318, upload-time = "2025-03-07T01:51:01.794Z" },
] ]
[[package]] [[package]]
@ -892,10 +951,12 @@ name = "nvidia-cudnn-cu12"
version = "9.10.2.21" version = "9.10.2.21"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-cublas-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/fa/41/e79269ce215c857c935fd86bcfe91a451a584dfc27f1e068f568b9ad1ab7/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c9132cc3f8958447b4910a1720036d9eff5928cc3179b0a51fb6d167c6cc87d8", size = 705026878, upload-time = "2025-06-06T21:52:51.348Z" },
{ url = "https://files.pythonhosted.org/packages/ba/51/e123d997aa098c61d029f76663dedbfb9bc8dcf8c60cbd6adbe42f76d049/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8", size = 706758467, upload-time = "2025-06-06T21:54:08.597Z" }, { url = "https://files.pythonhosted.org/packages/ba/51/e123d997aa098c61d029f76663dedbfb9bc8dcf8c60cbd6adbe42f76d049/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8", size = 706758467, upload-time = "2025-06-06T21:54:08.597Z" },
{ url = "https://files.pythonhosted.org/packages/3d/90/0bd6e586701b3a890fd38aa71c387dab4883d619d6e5ad912ccbd05bfd67/nvidia_cudnn_cu12-9.10.2.21-py3-none-win_amd64.whl", hash = "sha256:c6288de7d63e6cf62988f0923f96dc339cea362decb1bf5b3141883392a7d65e", size = 692992268, upload-time = "2025-06-06T21:55:18.114Z" },
] ]
[[package]] [[package]]
@ -903,10 +964,12 @@ name = "nvidia-cufft-cu12"
version = "11.3.3.83" version = "11.3.3.83"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-nvjitlink-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/60/bc/7771846d3a0272026c416fbb7e5f4c1f146d6d80704534d0b187dd6f4800/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:848ef7224d6305cdb2a4df928759dca7b1201874787083b6e7550dd6765ce69a", size = 193109211, upload-time = "2025-03-07T01:44:56.873Z" },
{ url = "https://files.pythonhosted.org/packages/1f/13/ee4e00f30e676b66ae65b4f08cb5bcbb8392c03f54f2d5413ea99a5d1c80/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74", size = 193118695, upload-time = "2025-03-07T01:45:27.821Z" }, { url = "https://files.pythonhosted.org/packages/1f/13/ee4e00f30e676b66ae65b4f08cb5bcbb8392c03f54f2d5413ea99a5d1c80/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74", size = 193118695, upload-time = "2025-03-07T01:45:27.821Z" },
{ url = "https://files.pythonhosted.org/packages/7d/ec/ce1629f1e478bb5ccd208986b5f9e0316a78538dd6ab1d0484f012f8e2a1/nvidia_cufft_cu12-11.3.3.83-py3-none-win_amd64.whl", hash = "sha256:7a64a98ef2a7c47f905aaf8931b69a3a43f27c55530c698bb2ed7c75c0b42cb7", size = 192216559, upload-time = "2025-03-07T01:53:57.106Z" },
] ]
[[package]] [[package]]
@ -915,6 +978,7 @@ version = "1.13.1.3"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/bb/fe/1bcba1dfbfb8d01be8d93f07bfc502c93fa23afa6fd5ab3fc7c1df71038a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc", size = 1197834, upload-time = "2025-03-07T01:45:50.723Z" }, { url = "https://files.pythonhosted.org/packages/bb/fe/1bcba1dfbfb8d01be8d93f07bfc502c93fa23afa6fd5ab3fc7c1df71038a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc", size = 1197834, upload-time = "2025-03-07T01:45:50.723Z" },
{ url = "https://files.pythonhosted.org/packages/1e/f5/5607710447a6fe9fd9b3283956fceeee8a06cda1d2f56ce31371f595db2a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:4beb6d4cce47c1a0f1013d72e02b0994730359e17801d395bdcbf20cfb3bb00a", size = 1120705, upload-time = "2025-03-07T01:45:41.434Z" },
] ]
[[package]] [[package]]
@ -922,7 +986,9 @@ name = "nvidia-curand-cu12"
version = "10.3.9.90" version = "10.3.9.90"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/45/5e/92aa15eca622a388b80fbf8375d4760738df6285b1e92c43d37390a33a9a/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:dfab99248034673b779bc6decafdc3404a8a6f502462201f2f31f11354204acd", size = 63625754, upload-time = "2025-03-07T01:46:10.735Z" },
{ url = "https://files.pythonhosted.org/packages/fb/aa/6584b56dc84ebe9cf93226a5cde4d99080c8e90ab40f0c27bda7a0f29aa1/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9", size = 63619976, upload-time = "2025-03-07T01:46:23.323Z" }, { url = "https://files.pythonhosted.org/packages/fb/aa/6584b56dc84ebe9cf93226a5cde4d99080c8e90ab40f0c27bda7a0f29aa1/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9", size = 63619976, upload-time = "2025-03-07T01:46:23.323Z" },
{ url = "https://files.pythonhosted.org/packages/b9/75/70c05b2f3ed5be3bb30b7102b6eb78e100da4bbf6944fd6725c012831cab/nvidia_curand_cu12-10.3.9.90-py3-none-win_amd64.whl", hash = "sha256:f149a8ca457277da854f89cf282d6ef43176861926c7ac85b2a0fbd237c587ec", size = 62765309, upload-time = "2025-03-07T01:54:20.478Z" },
] ]
[[package]] [[package]]
@ -930,12 +996,14 @@ name = "nvidia-cusolver-cu12"
version = "11.7.3.90" version = "11.7.3.90"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-cublas-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-cusparse-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-nvjitlink-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/c8/32/f7cd6ce8a7690544d084ea21c26e910a97e077c9b7f07bf5de623ee19981/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:db9ed69dbef9715071232caa9b69c52ac7de3a95773c2db65bdba85916e4e5c0", size = 267229841, upload-time = "2025-03-07T01:46:54.356Z" },
{ url = "https://files.pythonhosted.org/packages/85/48/9a13d2975803e8cf2777d5ed57b87a0b6ca2cc795f9a4f59796a910bfb80/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450", size = 267506905, upload-time = "2025-03-07T01:47:16.273Z" }, { url = "https://files.pythonhosted.org/packages/85/48/9a13d2975803e8cf2777d5ed57b87a0b6ca2cc795f9a4f59796a910bfb80/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450", size = 267506905, upload-time = "2025-03-07T01:47:16.273Z" },
{ url = "https://files.pythonhosted.org/packages/13/c0/76ca8551b8a84146ffa189fec81c26d04adba4bc0dbe09cd6e6fd9b7de04/nvidia_cusolver_cu12-11.7.3.90-py3-none-win_amd64.whl", hash = "sha256:4a550db115fcabc4d495eb7d39ac8b58d4ab5d8e63274d3754df1c0ad6a22d34", size = 256720438, upload-time = "2025-03-07T01:54:39.898Z" },
] ]
[[package]] [[package]]
@ -943,10 +1011,12 @@ name = "nvidia-cusparse-cu12"
version = "12.5.8.93" version = "12.5.8.93"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" }, { name = "nvidia-nvjitlink-cu12", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/bc/f7/cd777c4109681367721b00a106f491e0d0d15cfa1fd59672ce580ce42a97/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9b6c161cb130be1a07a27ea6923df8141f3c295852f4b260c65f18f3e0a091dc", size = 288117129, upload-time = "2025-03-07T01:47:40.407Z" },
{ url = "https://files.pythonhosted.org/packages/c2/f5/e1854cb2f2bcd4280c44736c93550cc300ff4b8c95ebe370d0aa7d2b473d/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b", size = 288216466, upload-time = "2025-03-07T01:48:13.779Z" }, { url = "https://files.pythonhosted.org/packages/c2/f5/e1854cb2f2bcd4280c44736c93550cc300ff4b8c95ebe370d0aa7d2b473d/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b", size = 288216466, upload-time = "2025-03-07T01:48:13.779Z" },
{ url = "https://files.pythonhosted.org/packages/62/07/f3b2ad63f8e3d257a599f422ae34eb565e70c41031aecefa3d18b62cabd1/nvidia_cusparse_cu12-12.5.8.93-py3-none-win_amd64.whl", hash = "sha256:9a33604331cb2cac199f2e7f5104dfbb8a5a898c367a53dfda9ff2acb6b6b4dd", size = 284937404, upload-time = "2025-03-07T01:55:07.742Z" },
] ]
[[package]] [[package]]
@ -954,7 +1024,9 @@ name = "nvidia-cusparselt-cu12"
version = "0.7.1" version = "0.7.1"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/73/b9/598f6ff36faaece4b3c50d26f50e38661499ff34346f00e057760b35cc9d/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:8878dce784d0fac90131b6817b607e803c36e629ba34dc5b433471382196b6a5", size = 283835557, upload-time = "2025-02-26T00:16:54.265Z" },
{ url = "https://files.pythonhosted.org/packages/56/79/12978b96bd44274fe38b5dde5cfb660b1d114f70a65ef962bcbbed99b549/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623", size = 287193691, upload-time = "2025-02-26T00:15:44.104Z" }, { url = "https://files.pythonhosted.org/packages/56/79/12978b96bd44274fe38b5dde5cfb660b1d114f70a65ef962bcbbed99b549/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623", size = 287193691, upload-time = "2025-02-26T00:15:44.104Z" },
{ url = "https://files.pythonhosted.org/packages/2f/d8/a6b0d0d0c2435e9310f3e2bb0d9c9dd4c33daef86aa5f30b3681defd37ea/nvidia_cusparselt_cu12-0.7.1-py3-none-win_amd64.whl", hash = "sha256:f67fbb5831940ec829c9117b7f33807db9f9678dc2a617fbe781cac17b4e1075", size = 271020911, upload-time = "2025-02-26T00:14:47.204Z" },
] ]
[[package]] [[package]]
@ -962,6 +1034,7 @@ name = "nvidia-nccl-cu12"
version = "2.27.3" version = "2.27.3"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/4b/7b/8354b784cf73b0ba51e566b4baba3ddd44fe8288a3d39ef1e06cd5417226/nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:9ddf1a245abc36c550870f26d537a9b6087fb2e2e3d6e0ef03374c6fd19d984f", size = 322397768, upload-time = "2025-06-03T21:57:30.234Z" },
{ url = "https://files.pythonhosted.org/packages/5c/5b/4e4fff7bad39adf89f735f2bc87248c81db71205b62bcc0d5ca5b606b3c3/nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adf27ccf4238253e0b826bce3ff5fa532d65fc42322c8bfdfaf28024c0fbe039", size = 322364134, upload-time = "2025-06-03T21:58:04.013Z" }, { url = "https://files.pythonhosted.org/packages/5c/5b/4e4fff7bad39adf89f735f2bc87248c81db71205b62bcc0d5ca5b606b3c3/nvidia_nccl_cu12-2.27.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adf27ccf4238253e0b826bce3ff5fa532d65fc42322c8bfdfaf28024c0fbe039", size = 322364134, upload-time = "2025-06-03T21:58:04.013Z" },
] ]
@ -971,6 +1044,8 @@ version = "12.8.93"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/f6/74/86a07f1d0f42998ca31312f998bd3b9a7eff7f52378f4f270c8679c77fb9/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88", size = 39254836, upload-time = "2025-03-07T01:49:55.661Z" }, { url = "https://files.pythonhosted.org/packages/f6/74/86a07f1d0f42998ca31312f998bd3b9a7eff7f52378f4f270c8679c77fb9/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88", size = 39254836, upload-time = "2025-03-07T01:49:55.661Z" },
{ url = "https://files.pythonhosted.org/packages/2a/a2/8cee5da30d13430e87bf99bb33455d2724d0a4a9cb5d7926d80ccb96d008/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:adccd7161ace7261e01bb91e44e88da350895c270d23f744f0820c818b7229e7", size = 38386204, upload-time = "2025-03-07T01:49:43.612Z" },
{ url = "https://files.pythonhosted.org/packages/ed/d7/34f02dad2e30c31b10a51f6b04e025e5dd60e5f936af9045a9b858a05383/nvidia_nvjitlink_cu12-12.8.93-py3-none-win_amd64.whl", hash = "sha256:bd93fbeeee850917903583587f4fc3a4eafa022e34572251368238ab5e6bd67f", size = 268553710, upload-time = "2025-03-07T01:56:24.13Z" },
] ]
[[package]] [[package]]
@ -978,7 +1053,9 @@ name = "nvidia-nvtx-cu12"
version = "12.8.90" version = "12.8.90"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/10/c0/1b303feea90d296f6176f32a2a70b5ef230f9bdeb3a72bddb0dc922dc137/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d7ad891da111ebafbf7e015d34879f7112832fc239ff0d7d776b6cb685274615", size = 91161, upload-time = "2025-03-07T01:42:23.922Z" },
{ url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" }, { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" },
{ url = "https://files.pythonhosted.org/packages/9f/99/4c9c0c329bf9fc125008c3b54c7c94c0023518d06fc025ae36431375e1fe/nvidia_nvtx_cu12-12.8.90-py3-none-win_amd64.whl", hash = "sha256:619c8304aedc69f02ea82dd244541a83c3d9d40993381b3b590f1adaed3db41e", size = 56492, upload-time = "2025-03-07T01:52:24.69Z" },
] ]
[[package]] [[package]]
@ -1334,13 +1411,13 @@ name = "pytest"
version = "8.4.2" version = "8.4.2"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "colorama", marker = "sys_platform == 'win32'" }, { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "exceptiongroup", marker = "python_full_version < '3.11'" }, { name = "exceptiongroup", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "iniconfig" }, { name = "iniconfig" },
{ name = "packaging" }, { name = "packaging" },
{ name = "pluggy" }, { name = "pluggy" },
{ name = "pygments" }, { name = "pygments" },
{ name = "tomli", marker = "python_full_version < '3.11'" }, { name = "tomli", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/a3/5c/00a0e072241553e1a7496d638deababa67c5058571567b92a7eaa258397c/pytest-8.4.2.tar.gz", hash = "sha256:86c0d0b93306b961d58d62a4db4879f27fe25513d4b969df351abdddb3c30e01", size = 1519618, upload-time = "2025-09-04T14:34:22.711Z" } sdist = { url = "https://files.pythonhosted.org/packages/a3/5c/00a0e072241553e1a7496d638deababa67c5058571567b92a7eaa258397c/pytest-8.4.2.tar.gz", hash = "sha256:86c0d0b93306b961d58d62a4db4879f27fe25513d4b969df351abdddb3c30e01", size = 1519618, upload-time = "2025-09-04T14:34:22.711Z" }
wheels = [ wheels = [
@ -1561,7 +1638,7 @@ version = "0.48.0"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "anyio" }, { name = "anyio" },
{ name = "typing-extensions", marker = "python_full_version < '3.13'" }, { name = "typing-extensions", marker = "python_full_version < '3.13' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/a7/a5/d6f429d43394057b67a6b5bbe6eae2f77a6bf7459d961fdb224bf206eee6/starlette-0.48.0.tar.gz", hash = "sha256:7e8cee469a8ab2352911528110ce9088fdc6a37d9876926e73da7ce4aa4c7a46", size = 2652949, upload-time = "2025-09-13T08:41:05.699Z" } sdist = { url = "https://files.pythonhosted.org/packages/a7/a5/d6f429d43394057b67a6b5bbe6eae2f77a6bf7459d961fdb224bf206eee6/starlette-0.48.0.tar.gz", hash = "sha256:7e8cee469a8ab2352911528110ce9088fdc6a37d9876926e73da7ce4aa4c7a46", size = 2652949, upload-time = "2025-09-13T08:41:05.699Z" }
wheels = [ wheels = [
@ -1684,30 +1761,38 @@ wheels = [
name = "torch" name = "torch"
version = "2.8.0+cu128" version = "2.8.0+cu128"
source = { registry = "https://download.pytorch.org/whl/cu128" } source = { registry = "https://download.pytorch.org/whl/cu128" }
resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'linux'",
"python_full_version >= '3.12' and sys_platform != 'linux'",
"python_full_version == '3.11.*' and sys_platform == 'linux'",
"python_full_version < '3.11' and sys_platform == 'linux'",
"python_full_version == '3.11.*' and sys_platform != 'linux'",
"python_full_version < '3.11' and sys_platform != 'linux'",
]
dependencies = [ dependencies = [
{ name = "filelock" }, { name = "filelock", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "fsspec" }, { name = "fsspec", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "jinja2" }, { name = "jinja2", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" }, { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and extra == 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.5", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" }, { name = "networkx", version = "3.5", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and extra == 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cublas-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cublas-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cuda-cupti-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cuda-cupti-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cuda-nvrtc-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cuda-nvrtc-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cuda-runtime-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cuda-runtime-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cudnn-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cudnn-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cufft-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cufft-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cufile-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cufile-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-curand-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-curand-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cusolver-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cusolver-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cusparse-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cusparse-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-cusparselt-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-cusparselt-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-nccl-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-nccl-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-nvjitlink-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-nvjitlink-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "nvidia-nvtx-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "nvidia-nvtx-cu12", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "setuptools", marker = "python_full_version >= '3.12'" }, { name = "setuptools", marker = "(python_full_version >= '3.12' and extra == 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "sympy" }, { name = "sympy", marker = "extra == 'extra-8-nanochat-gpu'" },
{ name = "triton", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" }, { name = "triton", marker = "(platform_machine == 'x86_64' and sys_platform == 'linux' and extra == 'extra-8-nanochat-gpu') or (platform_machine != 'x86_64' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'linux' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "typing-extensions" }, { name = "typing-extensions", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:0c96999d15cf1f13dd7c913e0b21a9a355538e6cfc10861a17158320292f5954" }, { url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:0c96999d15cf1f13dd7c913e0b21a9a355538e6cfc10861a17158320292f5954" },
@ -1722,12 +1807,143 @@ wheels = [
{ url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp313-cp313t-win_amd64.whl", hash = "sha256:970b4f4661fa7b44f6a7e6df65de7fc4a6fff2af610dc415c1d695ca5f1f37d2" }, { url = "https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp313-cp313t-win_amd64.whl", hash = "sha256:970b4f4661fa7b44f6a7e6df65de7fc4a6fff2af610dc415c1d695ca5f1f37d2" },
] ]
[[package]]
name = "torch"
version = "2.9.0"
source = { registry = "https://download.pytorch.org/whl/cpu" }
resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'darwin'",
"python_full_version == '3.11.*' and sys_platform == 'darwin'",
"python_full_version < '3.11' and sys_platform == 'darwin'",
]
dependencies = [
{ name = "filelock", marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "fsspec", marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "jinja2", marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version >= '3.11' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.5", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version < '3.11' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "setuptools", marker = "(python_full_version >= '3.12' and sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version < '3.12' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "sympy", marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "typing-extensions", marker = "(sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
]
wheels = [
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:59484193b01299bf669520505a72b29d59a0028ae4c6d95f492938f186592208" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp311-none-macosx_11_0_arm64.whl", hash = "sha256:aa4483602586cc9a35d1cf33771a9977f05f642b9161518a289e36548a0b77c2" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:4de0ed8cbc457a506dbca40376e206a29efee10756a00f1f3404bf67ad737d04" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:259548471194ab63d7ea273873053a6e3cc23530c1510f01e9d7ad259187bbd0" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp313-none-macosx_11_0_arm64.whl", hash = "sha256:e24836d968b54ef4dfb05594001a61958711ac9224026291e4e3f92f83a6fd7f" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:d8e2ab7f86010330bdcc39c8b2c795590cc75e37df4823cdaee2c98d6e3ff4a3" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:a3e859039c985d8e3ea60d7a54ca7e97ea2ae15e31beced4f3260128a161bb01" },
]
[[package]]
name = "torch"
version = "2.9.0"
source = { registry = "https://pypi.org/simple" }
resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'linux'",
"python_full_version >= '3.12' and sys_platform != 'linux'",
"python_full_version == '3.11.*' and sys_platform == 'linux'",
"python_full_version < '3.11' and sys_platform == 'linux'",
"python_full_version == '3.11.*' and sys_platform != 'linux'",
"python_full_version < '3.11' and sys_platform != 'linux'",
]
dependencies = [
{ name = "filelock", marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
{ name = "fsspec", marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
{ name = "jinja2", marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.5", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "setuptools", marker = "(python_full_version >= '3.12' and extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "sympy", marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
{ name = "typing-extensions", marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/bb/86/245c240d2138c17ed572c943c289056c2721abab70810d772c6bf5495b28/torch-2.9.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:030bbfe367379ae6a4ae4042b6c44da25383343b8b3c68abaa9c7231efbaf2dd", size = 104213554, upload-time = "2025-10-15T15:45:59.798Z" },
{ url = "https://files.pythonhosted.org/packages/58/1d/fd1e88ae0948825efcab7dd66d12bec23f05d4d38ed81573c8d453c14c06/torch-2.9.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:51cb63902182a78e90886e8068befd8ea102af4b00e420263591a3d70c7d3c6c", size = 899795167, upload-time = "2025-10-15T15:47:12.695Z" },
{ url = "https://files.pythonhosted.org/packages/63/5a/496197b45c14982bef4e079b24c61dc108e3ab0d0cc9718dba9f54f45a46/torch-2.9.0-cp310-cp310-win_amd64.whl", hash = "sha256:3f6aad4d2f0ee2248bac25339d74858ff846c3969b27d14ac235821f055af83d", size = 109310314, upload-time = "2025-10-15T15:46:16.633Z" },
{ url = "https://files.pythonhosted.org/packages/58/b0/2b4e647b0fc706e88eb6c253d05511865578f5f67b55fad639bf3272a4a1/torch-2.9.0-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:413e1654c9203733138858780e184d9fc59442f0b3b209e16f39354eb893db9b", size = 74452019, upload-time = "2025-10-15T15:46:04.296Z" },
{ url = "https://files.pythonhosted.org/packages/58/fe/334225e6330e672b36aef23d77451fa906ea12881570c08638a91331a212/torch-2.9.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:c596708b5105d0b199215acf0c9be7c1db5f1680d88eddadf4b75a299259a677", size = 104230578, upload-time = "2025-10-15T15:46:08.182Z" },
{ url = "https://files.pythonhosted.org/packages/05/cc/49566caaa218872ec9a2912456f470ff92649894a4bc2e5274aa9ef87c4a/torch-2.9.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:51de31219c97c51cf4bf2be94d622e3deb5dcc526c6dc00e97c17eaec0fc1d67", size = 899815990, upload-time = "2025-10-15T15:48:03.336Z" },
{ url = "https://files.pythonhosted.org/packages/74/25/e9ab21d5925b642d008f139d4a3c9664fc9ee1faafca22913c080cc4c0a5/torch-2.9.0-cp311-cp311-win_amd64.whl", hash = "sha256:dd515c70059afd95f48b8192733764c08ca37a1d19803af6401b5ecad7c8676e", size = 109313698, upload-time = "2025-10-15T15:46:12.425Z" },
{ url = "https://files.pythonhosted.org/packages/b3/b7/205ef3e94de636feffd64b28bb59a0dfac0771221201b9871acf9236f5ca/torch-2.9.0-cp311-none-macosx_11_0_arm64.whl", hash = "sha256:614a185e4986326d526a91210c8fc1397e76e8cfafa78baf6296a790e53a9eec", size = 74463678, upload-time = "2025-10-15T15:46:29.779Z" },
{ url = "https://files.pythonhosted.org/packages/d1/d3/3985739f3b8e88675127bf70f82b3a48ae083e39cda56305dbd90398fec0/torch-2.9.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:e5f7af1dc4c0a7c4a260c2534f41ddaf209714f7c89145e644c44712fbd6b642", size = 104107898, upload-time = "2025-10-15T15:46:20.883Z" },
{ url = "https://files.pythonhosted.org/packages/a5/4b/f4bb2e6c25d0272f798cd6d7a04ed315da76cec68c602d87040c7847287f/torch-2.9.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:01cff95ecd9a212ea2f141db28acccdceb6a4c54f64e6c51091146f5e2a772c6", size = 899738273, upload-time = "2025-10-15T15:50:04.188Z" },
{ url = "https://files.pythonhosted.org/packages/66/11/c1c5ba6691cda6279087c35bd626536e4fd29521fe740abf5008377a9a02/torch-2.9.0-cp312-cp312-win_amd64.whl", hash = "sha256:4582b162f541651f0cb184d3e291c05c2f556c7117c64a9873e2ee158d40062b", size = 109280887, upload-time = "2025-10-15T15:46:26.228Z" },
{ url = "https://files.pythonhosted.org/packages/dd/5f/b85bd8c05312d71de9402bf5868d217c38827cfd09d8f8514e5be128a52b/torch-2.9.0-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:33f58e9a102a91259af289d50525c30323b5c9ae1d31322b6447c0814da68695", size = 74478983, upload-time = "2025-10-15T15:46:39.406Z" },
{ url = "https://files.pythonhosted.org/packages/c2/1c/90eb13833cdf4969ea9707586d7b57095c3b6e2b223a7256bf111689bcb8/torch-2.9.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:c30a17fc83eeab346913e237c64b15b5ba6407fff812f6c541e322e19bc9ea0e", size = 104111330, upload-time = "2025-10-15T15:46:35.238Z" },
{ url = "https://files.pythonhosted.org/packages/0e/21/2254c54b8d523592c25ef4434769aa23e29b1e6bf5f4c0ad9e27bf442927/torch-2.9.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:8f25033b8667b57857dfd01458fbf2a9e6a6df1f8def23aef0dc46292f6aa642", size = 899750243, upload-time = "2025-10-15T15:48:57.459Z" },
{ url = "https://files.pythonhosted.org/packages/b7/a5/5cb94fa4fd1e78223455c23c200f30f6dc10c6d4a2bcc8f6e7f2a2588370/torch-2.9.0-cp313-cp313-win_amd64.whl", hash = "sha256:d037f1b4ffd25013be4a7bf3651a0a910c68554956c7b2c92ebe87c76475dece", size = 109284513, upload-time = "2025-10-15T15:46:45.061Z" },
{ url = "https://files.pythonhosted.org/packages/66/e8/fc414d8656250ee46120b44836ffbb3266343db424b3e18ca79ebbf69d4f/torch-2.9.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:e4e5b5cba837a2a8d1a497ba9a58dae46fa392593eaa13b871c42f71847503a5", size = 74830362, upload-time = "2025-10-15T15:46:48.983Z" },
{ url = "https://files.pythonhosted.org/packages/ed/5f/9474c98fc5ae0cd04b9466035428cd360e6611a86b8352a0fc2fa504acdc/torch-2.9.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:64693568f5dc4dbd5f880a478b1cea0201cc6b510d91d1bc54fea86ac5d1a637", size = 104144940, upload-time = "2025-10-15T15:47:29.076Z" },
{ url = "https://files.pythonhosted.org/packages/2d/5a/8e0c1cf57830172c109d4bd6be2708cabeaf550983eee7029291322447a0/torch-2.9.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:f8ed31ddd7d10bfb3fbe0b9fe01b1243577f13d75e6f4a0839a283915ce3791e", size = 899744054, upload-time = "2025-10-15T15:48:29.864Z" },
{ url = "https://files.pythonhosted.org/packages/6d/28/82c28b30fcb4b7c9cdd995763d18bbb830d6521356712faebbad92ffa61d/torch-2.9.0-cp313-cp313t-win_amd64.whl", hash = "sha256:eff527d4e4846e6f70d2afd8058b73825761203d66576a7e04ea2ecfebcb4ab8", size = 109517546, upload-time = "2025-10-15T15:47:33.395Z" },
{ url = "https://files.pythonhosted.org/packages/ff/c3/a91f96ec74347fa5fd24453fa514bc61c61ecc79196fa760b012a1873d96/torch-2.9.0-cp313-none-macosx_11_0_arm64.whl", hash = "sha256:f8877779cf56d1ce431a7636703bdb13307f5960bb1af49716d8b179225e0e6a", size = 74480732, upload-time = "2025-10-15T15:47:38.002Z" },
{ url = "https://files.pythonhosted.org/packages/5c/73/9f70af34b334a7e0ef496ceec96b7ec767bd778ea35385ce6f77557534d1/torch-2.9.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:7e614fae699838038d888729f82b687c03413c5989ce2a9481f9a7e7a396e0bb", size = 74433037, upload-time = "2025-10-15T15:47:41.894Z" },
{ url = "https://files.pythonhosted.org/packages/b7/84/37cf88625901934c97109e583ecc21777d21c6f54cda97a7e5bbad1ee2f2/torch-2.9.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:dfb5b8cd310ba3436c7e14e8b7833ef658cf3045e50d2bdaed23c8fc517065eb", size = 104116482, upload-time = "2025-10-15T15:47:46.266Z" },
{ url = "https://files.pythonhosted.org/packages/56/8e/ca8b17866943a8d4f4664d402ea84210aa274588b4c5d89918f5caa24eec/torch-2.9.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:b3d29524993a478e46f5d598b249cd824b7ed98d7fba538bd9c4cde6c803948f", size = 899746916, upload-time = "2025-10-15T15:50:40.294Z" },
{ url = "https://files.pythonhosted.org/packages/43/65/3b17c0fbbdab6501c5b320a52a648628d0d44e7379f64e27d9eef701b6bf/torch-2.9.0-cp314-cp314-win_amd64.whl", hash = "sha256:71c7578984f5ec0eb645eb4816ac8435fcf3e3e2ae1901bcd2f519a9cafb5125", size = 109275151, upload-time = "2025-10-15T15:49:20.715Z" },
{ url = "https://files.pythonhosted.org/packages/83/36/74f8c051f785500396e42f93542422422dfd874a174f21f8d955d36e5d64/torch-2.9.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:71d9309aee457bbe0b164bce2111cd911c4ed4e847e65d5077dbbcd3aba6befc", size = 74823353, upload-time = "2025-10-15T15:49:16.59Z" },
{ url = "https://files.pythonhosted.org/packages/62/51/dc3b4e2f9ba98ae27238f0153ca098bf9340b2dafcc67fde645d496dfc2a/torch-2.9.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:c08fb654d783899e204a32cca758a7ce8a45b2d78eeb89517cc937088316f78e", size = 104140340, upload-time = "2025-10-15T15:50:19.67Z" },
{ url = "https://files.pythonhosted.org/packages/c0/8d/b00657f8141ac16af7bb6cda2e67de18499a3263b78d516b9a93fcbc98e3/torch-2.9.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:ec8feb0099b2daa5728fbc7abb0b05730fd97e0f359ff8bda09865aaa7bd7d4b", size = 899731750, upload-time = "2025-10-15T15:49:36.673Z" },
{ url = "https://files.pythonhosted.org/packages/fc/29/bd361e0cbb2c79ce6450f42643aaf6919956f89923a50571b0ebfe92d142/torch-2.9.0-cp314-cp314t-win_amd64.whl", hash = "sha256:695ba920f234ad4170c9c50e28d56c848432f8f530e6bc7f88fcb15ddf338e75", size = 109503850, upload-time = "2025-10-15T15:50:24.118Z" },
]
[[package]]
name = "torch"
version = "2.9.0+cpu"
source = { registry = "https://download.pytorch.org/whl/cpu" }
resolution-markers = [
"python_full_version >= '3.12' and sys_platform == 'linux'",
"python_full_version >= '3.12' and sys_platform != 'darwin' and sys_platform != 'linux'",
"python_full_version == '3.11.*' and sys_platform == 'linux'",
"python_full_version < '3.11' and sys_platform == 'linux'",
"python_full_version == '3.11.*' and sys_platform != 'darwin' and sys_platform != 'linux'",
"python_full_version < '3.11' and sys_platform != 'darwin' and sys_platform != 'linux'",
]
dependencies = [
{ name = "filelock", marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "fsspec", marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "jinja2", marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version >= '3.11' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "networkx", version = "3.5", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version < '3.11' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "setuptools", marker = "(python_full_version >= '3.12' and sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (python_full_version < '3.12' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (sys_platform == 'darwin' and extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "sympy", marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
{ name = "typing-extensions", marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
]
wheels = [
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:b224792ea567b52c7f1ce1d789567f6920e06fd3b339fa1e1b05948845f783ad" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:bd2a257e670ede9fc01c6d76dccdc473040913b8e9328169bf177dbdc38e2484" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp310-cp310-win_amd64.whl", hash = "sha256:96f3f7aa4eb9e7fc5af8a722eaf1e5e32e3039dbafe817178d7b90a8566be32d" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:da77341ccaba31762d9238b0942c165c4582a26818f3045b052b39cebdd7ad9d" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:add3e93ecc1eeaa6853f6a973ce60ffb3cb14ed2e80f5055e139b09385dce0a7" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp311-cp311-win_amd64.whl", hash = "sha256:389e1e0b8083fd355f7caf5ba82356b5e01c318998bd575dbf2285a0d8137089" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp311-cp311-win_arm64.whl", hash = "sha256:5ce3d01aef91dc078fbb121814e556d55bc886d303efaf42c4fe67e411f5f9ad" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:3a651434ae1248b0568c12b5f9e3acc8942eb28378d9d04a79302938b68c6f24" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:28f6eb31b08180a5c5e98d5bc14eef6909c9f5a1dbff9632c3e02a8773449349" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp312-cp312-win_amd64.whl", hash = "sha256:e438061b87ec7dd6018fca9f975219889aa0a3f6cdc3ea10dd0ae2bc7f1c47ce" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp312-cp312-win_arm64.whl", hash = "sha256:eb13ff1c34e338d722e76a4fd83b8d282782505bd1b99af4b3c32da66eba6eb4" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:be4438d8dad7f0d5a5e54f0feef8a893446894ec87f102bb1d82dcc4518542e4" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:6c9b217584400963d5b4daddb3711ec7a3778eab211e18654fba076cce3b8682" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313-win_amd64.whl", hash = "sha256:728372e3f58c5826445f677746e5311c1935c1a7c59599f73a49ded850e038e8" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313-win_arm64.whl", hash = "sha256:95e56c26f919fbb98f16e7a0b87af494b893f9da9a65a020f17a01c13e520a81" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:6c777160288b08555820781ae0f3a2c67a59bd24b065e88ca1ec20e2f9dc8ac7" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:528fd338311f31c9fb18038cafd00e6eae0bf5ad5577521701acb62510753d18" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp313-cp313t-win_amd64.whl", hash = "sha256:d572863990e7d2762b547735ef589f6350d9eb4e441d38753a1c33636698cf4c" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:44aadb735774d4a99525d2ec29126b23016c44a07b02ce6c237dfa61a223dd52" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:b355e07b7f0c369cb031adfcbff5c37a609abcea091b918a39886412afd2e07d" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314-win_amd64.whl", hash = "sha256:c2698999361d73c2d25d7cc8a787130188d49b183abb18b554228daa102e1594" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:fa0d1373d04b30ff8f12d542135d292f1a1ddb7c0d852a3d487a320360e5dab9" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:2f49bb57a5fe0dc7f8e73ea9e5d36ebda2ea25b8a714a788f0fc2fc47d20a830" },
{ url = "https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp314-cp314t-win_amd64.whl", hash = "sha256:3a60d1ecf27a9cce839b3aa665b26f0af1b1007b9c9f1e7f597f6b7bdf107617" },
]
[[package]] [[package]]
name = "tqdm" name = "tqdm"
version = "4.67.1" version = "4.67.1"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "colorama", marker = "sys_platform == 'win32'" }, { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/a8/4b/29b4ef32e036bb34e4ab51796dd745cdba7ed47ad142a9f4a1eb8e0c744d/tqdm-4.67.1.tar.gz", hash = "sha256:f8aef9c52c08c13a65f30ea34f4e5aac3fd1a34959879d7e59e63027286627f2", size = 169737, upload-time = "2024-11-24T20:12:22.481Z" } sdist = { url = "https://files.pythonhosted.org/packages/a8/4b/29b4ef32e036bb34e4ab51796dd745cdba7ed47ad142a9f4a1eb8e0c744d/tqdm-4.67.1.tar.gz", hash = "sha256:f8aef9c52c08c13a65f30ea34f4e5aac3fd1a34959879d7e59e63027286627f2", size = 169737, upload-time = "2024-11-24T20:12:22.481Z" }
wheels = [ wheels = [
@ -1739,7 +1955,7 @@ name = "triton"
version = "3.4.0" version = "3.4.0"
source = { registry = "https://pypi.org/simple" } source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "setuptools", marker = "sys_platform == 'linux'" }, { name = "setuptools", marker = "extra == 'extra-8-nanochat-gpu'" },
] ]
wheels = [ wheels = [
{ url = "https://files.pythonhosted.org/packages/62/ee/0ee5f64a87eeda19bbad9bc54ae5ca5b98186ed00055281fd40fb4beb10e/triton-3.4.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7ff2785de9bc02f500e085420273bb5cc9c9bb767584a4aa28d6e360cec70128", size = 155430069, upload-time = "2025-07-30T19:58:21.715Z" }, { url = "https://files.pythonhosted.org/packages/62/ee/0ee5f64a87eeda19bbad9bc54ae5ca5b98186ed00055281fd40fb4beb10e/triton-3.4.0-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7ff2785de9bc02f500e085420273bb5cc9c9bb767584a4aa28d6e360cec70128", size = 155430069, upload-time = "2025-07-30T19:58:21.715Z" },
@ -1795,7 +2011,7 @@ source = { registry = "https://pypi.org/simple" }
dependencies = [ dependencies = [
{ name = "click" }, { name = "click" },
{ name = "h11" }, { name = "h11" },
{ name = "typing-extensions", marker = "python_full_version < '3.11'" }, { name = "typing-extensions", marker = "python_full_version < '3.11' or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" },
] ]
sdist = { url = "https://files.pythonhosted.org/packages/ef/5e/f0cd46063a02fd8515f0e880c37d2657845b7306c16ce6c4ffc44afd9036/uvicorn-0.36.0.tar.gz", hash = "sha256:527dc68d77819919d90a6b267be55f0e76704dca829d34aea9480be831a9b9d9", size = 80032, upload-time = "2025-09-20T01:07:14.418Z" } sdist = { url = "https://files.pythonhosted.org/packages/ef/5e/f0cd46063a02fd8515f0e880c37d2657845b7306c16ce6c4ffc44afd9036/uvicorn-0.36.0.tar.gz", hash = "sha256:527dc68d77819919d90a6b267be55f0e76704dca829d34aea9480be831a9b9d9", size = 80032, upload-time = "2025-09-20T01:07:14.418Z" }
wheels = [ wheels = [