You can deploy this model and start earning money today!

lkongam/
KernelCoder

Name: KernelCoder
Rating: 2.5 (20 reviews)

Views

No views yet

Model Summary

KernelCoder is trained on a curated dataset of reasoning traces and CUDA kernel pairs.

See details in paper.

Usage

python

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
import torch
import re
from typing import List, Tuple
from string import Template
PROMPT_TEMPLATE = Template('''
You are a Machine Learning Engineer trying to write custom cuda kernels to replace the pytorch operators in the given architecture to get speedups. You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom cuda kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.

For [Imports], you will likely need but not limited to the following libraries:
```python
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
```

Here’s an example to show you the syntax of inline embedding custom operators from the cuda kernel in torch:

The pytorch module needed to be optimize is:
```python
$ref_arch_torch
```

The example new arch with custom cuda kernels looks like this:
```python
$ref_arch_kernel
```

And the PyTorch code you need to optimize is:
```python
$code
```

Optimize the architecture named Model with custom cuda kernels! Optimize the architecture named Model with custom cuda kernels! Name your optimized output architecture ModelNew. Output the new code in codeblocks. Please generate real code, NOT pseudocode, make sure the code compiles and is fully functional. Just output the new model code, no other text, and NO testing code!

''')

class KernelCoder:

def __init__(self, model_name="lkongam/KernelCoder", tensor_parallel_size=1, gpu_memory_utilization=0.9):

self.model_name = model_name

self.llm = LLM(
model=model_name,
tensor_parallel_size=tensor_parallel_size,
gpu_memory_utilization=gpu_memory_utilization,
trust_remote_code=True,
dtype="auto"
)

self.tokenizer = self.llm.get_tokenizer()
self.device = torch.device("cuda")

def generate_raw(self, prompt, temperature=1.0):
messages = [
{"role": "user", "content": prompt}
]
text = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=True
)
return text

def extract_last_code_block(text):
code_blocks = re.findall(r"```(?:python)?\n(.*?)```", text, re.DOTALL)
if code_blocks:
return code_blocks[-1].strip()
match = re.search(r"</think>(.*)", text, re.S)
after_think = match.group(1).strip() if match else text
if not after_think:
return None
import_match = re.search(r"\bimport\b", after_think)
if import_match:
return after_think[import_match.start():].strip()
return after_think.strip()

origin_code = """
"""

model = KernelCoder(model_name="lkongam/KernelCoder")

prompt = PROMPT_TEMPLATE.substitute(code=origin_code)
code_output = model.generate_raw(prompt)
code = extract_last_code_block(code_output)
print(code)

Evaluation

Left: Pass@1, Right: Pass@10.