Qwen3 From Scratch

This repository contains a from-scratch, educational PyTorch implementation of Qwen3 with minimal code dependencies. The implementation is optimized for readability and intended for learning and research purposes.

Source code: qwen3.py

The model weights included here are PyTorch state dicts converted from the official weights provided by the Qwen3 team. For original weights, usage terms, and license information, please refer to the original model repositories linked below:

• https://huggingface.co/Qwen/Qwen3-0.6B-Base
• https://huggingface.co/Qwen/Qwen3-0.6B

To avoid duplication and ease maintance, this repository only contains the model weights; the self-contained source code can be found here. Instructions on how to use the code are provided below.

Qwen3 from-scratch code

The standalone notebooks in this folder contain from-scratch codes in linear fashion:

1. standalone-qwen3.ipynb: The dense Qwen3 model without bells and whistles
2. standalone-qwen3-plus-kvcache.ipynb: Same as above but with KV cache for better inference efficiency
3. standalone-qwen3-moe.ipynb: Like the first notebook but the Mixture-of-Experts (MoE) variant
4. standalone-qwen3-moe-plus-kvcache.ipynb: Same as above but with KV cache for better inference efficiency

Alternatively, I also organized the code into a Python package (including unit tests and CI), which you can run as described below.

Using Qwen3 0.6B via the llms-from-scratch package

For an easy way to use the Qwen3 from-scratch implementation, you can also use the llms-from-scratch PyPI package based on the source code in this repository at pkg/llms_from_scratch.

1) Installation

pip install llms_from_scratch tokenizers

2) Model and text generation settings

Specify which model to use:

USE_REASONING_MODEL = True   # The "thinking" model
USE_REASONING_MODEL = False  # The base model

Basic text generation settings that can be defined by the user. With 150 tokens, the model requires approximately 1.5 GB memory.

MAX_NEW_TOKENS = 150
TEMPERATURE = 0.
TOP_K = 1

3) Weight download and loading

This automatically downloads the weight file based on the model choice above:

from llms_from_scratch.qwen3 import download_from_huggingface

repo_id = "rasbt/qwen3-from-scratch"

if USE_REASONING_MODEL:
    filename = "qwen3-0.6B.pth"
    local_dir = "Qwen3-0.6B"    
else:
    filename = "qwen3-0.6B-base.pth"   
    local_dir = "Qwen3-0.6B-Base"

download_from_huggingface(
    repo_id=repo_id,
    filename=filename,
    local_dir=local_dir
)

The model weights are then loaded as follows:

from pathlib import Path
import torch

from llms_from_scratch.qwen3 import Qwen3Model, QWEN_CONFIG_06_B

model_file = Path(local_dir) / filename

model = Qwen3Model(QWEN_CONFIG_06_B)
model.load_state_dict(torch.load(model_file, weights_only=True, map_location="cpu"))

device = (
    torch.device("cuda") if torch.cuda.is_available() else
    torch.device("mps") if torch.backends.mps.is_available() else
    torch.device("cpu")
)
model.to(device)

4) Initialize tokenizer

The following code downloads and initializes the tokenizer:

from llms_from_scratch.qwen3 import Qwen3Tokenizer

if USE_REASONING_MODEL:
    tok_filename = str(Path("Qwen3-0.6B") / "tokenizer.json")
else:
    tok_filename = str(Path("Qwen3-0.6B-Base") / "tokenizer-base.json")

tokenizer = Qwen3Tokenizer(
    tokenizer_file_path=tok_filename,
    repo_id=repo_id,
    add_generation_prompt=USE_REASONING_MODEL,
    add_thinking=USE_REASONING_MODEL
)

5) Generating text

Lastly, we can generate text via the following code:

prompt = "Give me a short introduction to large language models."
input_token_ids = tokenizer.encode(prompt)

from llms_from_scratch.ch05 import generate
import time

torch.manual_seed(123)

start = time.time()

output_token_ids = generate(
    model=model,
    idx=torch.tensor(input_token_ids, device=device).unsqueeze(0),
    max_new_tokens=150,
    context_size=QWEN_CONFIG_06_B["context_length"],
    top_k=1,
    temperature=0.
)

total_time = time.time() - start
print(f"Time: {total_time:.2f} sec")
print(f"{int(len(output_token_ids[0])/total_time)} tokens/sec")

if torch.cuda.is_available():
    max_mem_bytes = torch.cuda.max_memory_allocated()
    max_mem_gb = max_mem_bytes / (1024 ** 3)
    print(f"Max memory allocated: {max_mem_gb:.2f} GB")

output_text = tokenizer.decode(output_token_ids.squeeze(0).tolist())

print("\n\nOutput text:\n\n", output_text + "...")

When using the Qwen3 0.6B reasoning model, the output should look similar to the one shown below (this was run on an A100):

Time: 6.35 sec
25 tokens/sec
Max memory allocated: 1.49 GB


Output text:

 <|im_start|>user
Give me a short introduction to large language models.<|im_end|>
Large language models (LLMs) are advanced artificial intelligence systems designed to generate human-like text. They are trained on vast amounts of text data, allowing them to understand and generate coherent, contextually relevant responses. LLMs are used in a variety of applications, including chatbots, virtual assistants, content generation, and more. They are powered by deep learning algorithms and can be fine-tuned for specific tasks, making them versatile tools for a wide range of industries.<|endoftext|>Human resources department of a company is planning to hire 100 new employees. The company has a budget of $100,000 for the recruitment process. The company has a minimum wage of $10 per hour. The company has a total of...

Pro tip: speed up inference with compilation

For up to a 4× speed-up, replace

model.to(device)

with

model = torch.compile(model)
model.to(device)

Note: There is a significant multi-minute upfront cost when compiling, and the speed-up takes effect after the first generate call.

The following table shows a performance comparison on an A100 for consequent generate calls: