You can deploy this model and start earning money today!

hkust-nlp/
drkernel-14b

Name: drkernel-14b
Rating: 2.5 (370 reviews)

reinforcement-learning

code

conversational

hkust-nlp/drkernel-coldstart-8k

hkust-nlp/drkernel-rl-data

hkust-nlp/drkernel-validation-data

2602.05885

Qwen/Qwen3-14B-Base

finetune

text-generation-inference

endpoints_compatible

Views

No views yet

DR.Kernel-14B

hkust-nlp/drkernel-14b is a Qwen3-14B-based model specialized for GPU kernel generation and optimization (especially Triton) in the DR.Kernel framework.

It is trained for iterative optimization with execution feedback from KernelGYM, rather than single-shot code generation only.

Model Summary

Model type: Qwen3ForCausalLM
Parameter count: 14,768,307,200 (from model.safetensors.index.json)
Weight dtype: BF16
Base model family: Qwen3-14B
Main capability: generate and iteratively refine optimized ModelNew kernel implementations from PyTorch reference tasks

Model Specs

From config.json:

num_hidden_layers: 40
hidden_size: 5120
intermediate_size: 17408
num_attention_heads: 40
num_key_value_heads: 8
max_position_embeddings: 32768
vocab_size: 151936
transformers_version: 4.56.0

From model.safetensors.index.json:

total_parameters: 14,768,307,200
total_size: 29,536,614,400 bytes
Expected checkpoint shards: model-00001-of-00007.safetensors ... model-00007-of-00007.safetensors

Training Recipe (DR.Kernel)

The 14B model follows the same two-stage DR.Kernel pipeline:

Cold-start SFT
- Dataset: hkust-nlp/drkernel-coldstart-8k
- Multi-turn trajectory warm-up for kernel generation/refinement
Multi-turn RL
- Train dataset: hkust-nlp/drkernel-rl-data
- Validation dataset: hkust-nlp/drkernel-validation-data (KernelBench Level 2 validation split)
- Core methods: TRLOO + MRS + PR + PRS
- Execution environment: KernelGYM with compilation/correctness/performance/profiling feedback

Related training scripts in this repo:

drkernel/kernel/scripts/sft/14b-coldstart.sh
drkernel/kernel/scripts/rl/14b_trloo_mrs_pr_prs.sh

Intended Use

Kernel generation research and benchmarking
Triton kernel optimization with iterative feedback
Multi-turn agentic code refinement under execution-based reward

Not Intended Use

Safety-critical production deployment without additional verification
General-purpose coding assistant use where kernel-evaluation feedback is unavailable

Prompting Format

The model is trained with kernel-optimization prompts that:

Provide a PyTorch reference architecture (Model, get_inputs, get_init_inputs)
Require returning an optimized ModelNew
In multi-turn settings, append server feedback and request iterative improvement

For best behavior, keep the same task style as DR.Kernel datasets and use chat-format messages.

Quick Start (Transformers)

Use the same fixed 1-shot first-turn prompt template as DR.Kernel data (recommended):

python

import textwrap
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hkust-nlp/drkernel-14b"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

ref_code = textwrap.dedent(
    """
    import torch
    import torch.nn as nn

    class Model(nn.Module):
        def __init__(self):
            super().__init__()

        def forward(self, x):
            x = torch.abs(x)
            x = x - 1.0
            return x

    def get_inputs():
        return [torch.randn(64, 128)]

    def get_init_inputs():
        return []
    """
).strip()

example_ref_code = textwrap.dedent(
    """
    import torch
    import torch.nn as nn
    import torch.nn.functional as F

    class Model(nn.Module):
        def __init__(self) -> None:
            super().__init__()

        def forward(self, a, b):
            return a + b

    def get_inputs():
        # randomly generate input tensors based on the model architecture
        a = torch.randn(1, 128).cuda()
        b = torch.randn(1, 128).cuda()
        return [a, b]

    def get_init_inputs():
        # randomly generate tensors required for initialization based on the model architecture
        return []
    """
).strip()

example_kernel_code = textwrap.dedent(
    '''
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    import triton
    import triton.language as tl

    @triton.jit
    def add_kernel(
        x_ptr,  # Pointer to first input
        y_ptr,  # Pointer to second input
        out_ptr,  # Pointer to output
        n_elements,  # Total number of elements in input/output
        BLOCK_SIZE: tl.constexpr,
    ):
        # Each program handles a contiguous block of data of size BLOCK_SIZE
        block_start = tl.program_id(0) * BLOCK_SIZE
        # Create a range of offsets [0..BLOCK_SIZE-1]
        offsets = block_start + tl.arange(0, BLOCK_SIZE)
        # Mask to ensure we don't go out of bounds
        mask = offsets < n_elements
        # Load input values
        x = tl.load(x_ptr + offsets, mask=mask, other=0.0)
        y = tl.load(y_ptr + offsets, mask=mask, other=0.0)
        # Perform the elementwise addition
        out = x + y
        # Store the result
        tl.store(out_ptr + offsets, out, mask=mask)

    def triton_add(x: torch.Tensor, y: torch.Tensor):
        """
        This function wraps the Triton kernel call. It:
          1. Ensures the inputs are contiguous on GPU.
          2. Calculates the grid (blocks) needed.
          3. Launches the Triton kernel.
        """
        assert x.is_cuda and y.is_cuda, "Tensors must be on CUDA."
        x = x.contiguous()
        y = y.contiguous()

        # Prepare output tensor
        out = torch.empty_like(x)

        # Number of elements in the tensor
        n_elements = x.numel()
        BLOCK_SIZE = 128  # Tunable parameter for block size

        # Determine the number of blocks needed
        grid = lambda meta: ((n_elements + meta["BLOCK_SIZE"] - 1) // meta["BLOCK_SIZE"],)

        # Launch the Triton kernel
        add_kernel[grid](x, y, out, n_elements, BLOCK_SIZE=BLOCK_SIZE)
        return out

    class ModelNew(nn.Module):
        def __init__(self) -> None:
            super().__init__()

        def forward(self, a, b):
            # Instead of "return a + b", call our Triton-based addition
            return triton_add(a, b)
    '''
).strip()

prompt_template = textwrap.dedent(
    """\
    You write custom Triton 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 Triton 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.

    Here's an example to show you the syntax of inline embedding custom Triton kernels in torch: The example given architecture is:

    ```python
    {example_ref_code}
    ```

    The example new arch with custom Triton kernels looks like this:

    ```python
    {example_kernel_code}
    ```

    You are given the following architecture:
    ```python
    {ref_code}
    ```

    Optimize the architecture named Model with custom Triton operators! 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. Let's think step by step.
    """
).strip()

prompt = prompt_template.format(
    example_ref_code=example_ref_code,
    example_kernel_code=example_kernel_code,
    ref_code=ref_code,
)
messages = [{"role": "user", "content": prompt}]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=2048,
        do_sample=True,
        temperature=1.0,
        top_p=1.0,
    )

# Only print newly generated tokens
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=False))

Evaluation

Use KernelGYM-based evaluation scripts in this repo:

drkernel/kernel/scripts/eval/drkernel-14b-maxturns3.sh
drkernel/kernel/scripts/eval/grading_common.sh for custom evaluation runs

Validation data:

hkust-nlp/drkernel-validation-data (KernelBench Level 2 validation tasks)

Data and Attribution

Query/task source includes:
- ByteDance-Seed/cudaLLM-data
SFT cold-start trajectories:
- hkust-nlp/drkernel-coldstart-8k
RL train data:
- hkust-nlp/drkernel-rl-data
Validation/eval data:
- hkust-nlp/drkernel-validation-data
Benchmark source:
- KernelBench

Please acknowledge original dataset/benchmark authors when using this model.

Related Resources

Paper: Dr.Kernel: Reinforcement Learning Done Right for Triton Kernel Generations
Codebase: KernelGYM
Training docs: drkernel/README.md

Citation

bibtex

@article{liuetal2026,
  title={Dr.Kernel: Reinforcement Learning Done Right for Triton Kernel Generations},
  author={Wei Liu, Jiawei Xu, Yingru Li, Longtao Zheng, Tianjian Li, Qian Liu, Junxian He},
  journal={arXiv:2602.05885},
  year={2026}
}