drkernel-14b-coldstart

DR.Kernel-14B-ColdStart

hkust-nlp/drkernel-14b-coldstart is the cold-start SFT checkpoint for DR.Kernel.

This model is trained on multi-turn SFT data only, and is intended as the initialization checkpoint before RL (TRLOO/MRS/PR/PRS).

Model Summary

• Model type: Qwen3ForCausalLM
• Base model family: Qwen3-14B
• Stage: cold-start supervised fine-tuning (before RL)
• Main capability: structured kernel-optimization responses (Model -> ModelNew) with DR.Kernel prompt format

Training Stage

This checkpoint corresponds to:

1. Cold-start SFT only
   • Dataset: hkust-nlp/drkernel-coldstart-8k
   • Multi-turn trajectories to teach kernel-generation/refinement behavior

Not included in this checkpoint:

• RL stage (TRLOO + MRS + PR + PRS)
• RL reward shaping / rejection sampling updates

Related script:

• drkernel/kernel/scripts/sft/14b-coldstart.sh

Intended Use

• As an initialization checkpoint for DR.Kernel RL training
• As a strong SFT baseline for kernel generation
• For ablations comparing cold-start vs post-RL checkpoints

Not Intended Use

• Final performance claims for DR.Kernel RL results
• Safety-critical production deployment without additional verification

Quick Start (Transformers)

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

import textwrap
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hkust-nlp/drkernel-14b-coldstart"
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))

Continue to RL Training

This checkpoint is intended to be fed into RL training:

• Script: drkernel/kernel/scripts/rl/14b_trloo_mrs_pr_prs.sh
• Typical model setting: MODEL_PATH="hkust-nlp/drkernel-14b-coldstart" (or local path)
• RL datasets:
  • hkust-nlp/drkernel-rl-data
  • hkust-nlp/drkernel-validation-data

Data and Attribution

• Cold-start SFT data:
  • hkust-nlp/drkernel-coldstart-8k
• Query/task source includes:
  • ByteDance-Seed/cudaLLM-data
• Benchmark source:
  • KernelBench

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

Related Resources

• Final RL model: hkust-nlp/drkernel-14b
• Paper: Dr.Kernel: Reinforcement Learning Done Right for Triton Kernel Generations
• Codebase: KernelGYM
• Training docs: drkernel/README.md

Citation

@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}
}