L1 P29 · popcorn

correct 29_logsumexp_across_virtual_ranks.py

model

grok-4.3

turns / max

4 / 15

tool calls

tokens (total)

13114

wall clock

845.5s

truncated

backend

cuda fp32

variant

popcorn

runtime

0.01 μs

ref runtime

0.03 μs

speedup

5.45x

started

2026-05-11T18:04:03Z

timingkernelbench.timing

kernel runtime

0.01 μs

ref runtime

0.03 μs

speedup

5.45x

kernel mean / median / std

0.01 / — / 7.30e-05 μs

ref mean / median / std

0.03 / — / 4.98e-04 μs

numerical precisioneval_kernel_against_ref

max abs error

0.00

mean abs error

0.00

max rel error

0.00

mean rel error

0.00

gpu memoryextended_metrics.measure_memory

kernel peak

33.01 MB

ref peak

33.01 MB

kernel / ref

1.00x

kernel launches / fusiontorch.profiler

kernel launches

ref launches

fusion ratio (ref / kernel)

20.00x

total cuda time

26 μs

SOL (speed-of-light)nsight_hardware_counters

SOL score (0–1)

0.001

dram util %

0.1%

compute util %

0.0%

bottleneck

unknown

dominant pipe

None

arithmetic intensity

—

ridge point

20.00

roofline / occupancynsight_hardware_counters

dram bandwidth

3.23 GB/s

dram util %

0.1%

fp32 tflops

—

fp32 util %

—

fp16 tflops

—

occupancy %

10.0%

bottleneck

unknown

dominant pipe %

—

arithmetic intensity

—

ridge point

20.00

L1 hit %

—

L2 hit %

98.5%

ld sectors / req

4.00

st sectors / req

—

regs / thread

smem / block

1024 B

block size

256

peak bw

3350.00 GB/s

peak fp32 tflops

67.00

peak fp16 tflops

1979.00

warp stalls (top 5)nsight

long_scoreboard

419.7%

short_scoreboard

71.3%

mio_throttle

3.3%

reference PyTorch source (the prompt the agent sees)

"""
Stable log-sum-exp over virtual ranks: out = logsumexp(x, dim=0), shape (B, D).
"""

import torch
import torch.nn as nn

from kernelbench.distributed_collectives import default_device

R = 4
B = 10
D = 32


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

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return torch.logsumexp(x, dim=0)


def get_inputs():
    dev = default_device()
    p = popcorn_pri
    return [torch.randn(p.jitter_int(R), p.jitter_int(B), p.jitter_int(D, align=8), device=dev)]


def get_init_inputs():
    return []

turn 0126.6s turn · 7.3s LLM · 119.3s tools · tokens in=1,925 · out=6 · total=2,282

reasoning

The task is to replace the forward computation of a PyTorch model with a custom CUDA kernel. The model computes the stable log-sum-exp over virtual ranks: out = logsumexp(x, dim=0), shape (B, D).

call → get_gpu_specs

callget_gpu_specs {}

result ← get_gpu_specs (ok)

okget_gpu_specs GPU specs for NVIDIA H100 80GB HBM3: total memory (runtime): 79.1 GB spec entry: H100 GPU Architecture: Hopper GPU Memory: 80GB Memory Bandwidth: 3.35 TB/s FP64 TFLOPS: 34 FP64 Tensor Core TFLOPS: 67 FP32 TFLOPS: 67 TF32 Tensor Core TFLOPS: 989 with sparsity BFLOAT16 Tensore Core TFLOPS: 1979 with sparsity FP16 Tensor Core TFLOPS: 1979 with sparsity FP8 Tensor Core TFLOPS: 3958 with sparsity INT8 Tensor Core TOPS: 3958 with sparsity Register File Size: 64K 32-bit registers per SM Maximum number of registers per thread: 255 Maximum number of thread blocks per SM: 32 Shared memory capacity per SM: 228 KB Maximum shared memory per thread block: 227 KB

turn 1258.3s turn · 10.4s LLM · 248.0s tools · tokens in=2,515 · out=492 · total=3,067

reasoning

GPU is H100, high bandwidth.

call → compile_kernel

callcompile_kernel { "kernel_code": "import torch\nimport torch.nn as nn\nfrom torch.utils.cpp_extension import load_inline\n\ncuda_source = '''\n#include <torch/extension.h>\n#include <cuda.h>\n#include <cuda_runtime.h>\n#include <cmath>\n\n__global__ void logsumexp_kernel(const float* __restrict__ x, float* __restrict__ out, int R, int B, int D) {\n int idx = blockIdx.x * blockDim.x + threadIdx.x;\n int total = B * D;\n if (idx >= total) return;\n int b = idx / D;\n int d = idx % D;\n \n float max_val = -INFINITY;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n max_val = fmaxf(max_val, val);\n }\n \n float sum = 0.0f;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n sum += expf(val - max_val);\n }\n \n out[b * D + d] = max_val + logf(sum);\n}\n\ntorch::Tensor logsumexp_cuda(torch::Tensor x) {\n int R = x.size(0);\n int B = x.size(1);\n int D = x.size(2);\n auto out = torch::empty({B, D}, x.options());\n int total = B * D;\n int threads = 256;\n int blocks = (total + threads - 1) / threads;\n logsumexp_kernel<<<blocks, threads>>>(x.data_ptr<float>(), out.data_ptr<float>(), R, B, D);\n return out;\n}\n'''\n\ncpp_source = '''\ntorch::Tensor logsumexp_cuda(torch::Tensor x);\n'''\n\nmodule = load_inline(\n name='logsumexp_module',\n cpp_sources=[cpp_source],\n cuda_sources=[cuda_source],\n functions=['logsumexp_cuda'],\n verbose=True\n)\n\nclass ModelNew(nn.Module):\n def __init__(self):\n super().__init__()\n self.module = module\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n return self.module.logsumexp_cuda(x)\n" }

result ← compile_kernel (ok)

okcompile_kernel compile_kernel PASSED: kernel compiled without errors.

turn 2260.3s turn · 8.5s LLM · 251.8s tools · tokens in=3,084 · out=493 · total=3,584

reasoning

Compiled successfully. Now check correctness.

call → run_correctness

callrun_correctness { "kernel_code": "import torch\nimport torch.nn as nn\nfrom torch.utils.cpp_extension import load_inline\n\ncuda_source = '''\n#include <torch/extension.h>\n#include <cuda.h>\n#include <cuda_runtime.h>\n#include <cmath>\n\n__global__ void logsumexp_kernel(const float* __restrict__ x, float* __restrict__ out, int R, int B, int D) {\n int idx = blockIdx.x * blockDim.x + threadIdx.x;\n int total = B * D;\n if (idx >= total) return;\n int b = idx / D;\n int d = idx % D;\n \n float max_val = -INFINITY;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n max_val = fmaxf(max_val, val);\n }\n \n float sum = 0.0f;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n sum += expf(val - max_val);\n }\n \n out[b * D + d] = max_val + logf(sum);\n}\n\ntorch::Tensor logsumexp_cuda(torch::Tensor x) {\n int R = x.size(0);\n int B = x.size(1);\n int D = x.size(2);\n auto out = torch::empty({B, D}, x.options());\n int total = B * D;\n int threads = 256;\n int blocks = (total + threads - 1) / threads;\n logsumexp_kernel<<<blocks, threads>>>(x.data_ptr<float>(), out.data_ptr<float>(), R, B, D);\n return out;\n}\n'''\n\ncpp_source = '''\ntorch::Tensor logsumexp_cuda(torch::Tensor x);\n'''\n\nmodule = load_inline(\n name='logsumexp_module',\n cpp_sources=[cpp_source],\n cuda_sources=[cuda_source],\n functions=['logsumexp_cuda'],\n verbose=True\n)\n\nclass ModelNew(nn.Module):\n def __init__(self):\n super().__init__()\n self.module = module\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n return self.module.logsumexp_cuda(x)\n" }

result ← run_correctness (ok)

okrun_correctness run_correctness PASSED: (3 / 3) trials all matched the reference. Numerical precision: max_abs_err=0.00e+00 mean_abs_err=0.00e+00 max_rel_err=0.00e+00

turn 3200.2s turn · 19.8s LLM · 180.5s tools · tokens in=3,644 · out=492 · total=4,181 · FINAL

reasoning

Correct. Now, to check if faster, I should profile or just submit since it's likely faster with no overhead.

call → submit_kernel

callsubmit_kernel { "kernel_code": "import torch\nimport torch.nn as nn\nfrom torch.utils.cpp_extension import load_inline\n\ncuda_source = '''\n#include <torch/extension.h>\n#include <cuda.h>\n#include <cuda_runtime.h>\n#include <cmath>\n\n__global__ void logsumexp_kernel(const float* __restrict__ x, float* __restrict__ out, int R, int B, int D) {\n int idx = blockIdx.x * blockDim.x + threadIdx.x;\n int total = B * D;\n if (idx >= total) return;\n int b = idx / D;\n int d = idx % D;\n \n float max_val = -INFINITY;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n max_val = fmaxf(max_val, val);\n }\n \n float sum = 0.0f;\n for (int r = 0; r < R; ++r) {\n float val = x[r * B * D + b * D + d];\n sum += expf(val - max_val);\n }\n \n out[b * D + d] = max_val + logf(sum);\n}\n\ntorch::Tensor logsumexp_cuda(torch::Tensor x) {\n int R = x.size(0);\n int B = x.size(1);\n int D = x.size(2);\n auto out = torch::empty({B, D}, x.options());\n int total = B * D;\n int threads = 256;\n int blocks = (total + threads - 1) / threads;\n logsumexp_kernel<<<blocks, threads>>>(x.data_ptr<float>(), out.data_ptr<float>(), R, B, D);\n return out;\n}\n'''\n\ncpp_source = '''\ntorch::Tensor logsumexp_cuda(torch::Tensor x);\n'''\n\nmodule = load_inline(\n name='logsumexp_module',\n cpp_sources=[cpp_source],\n cuda_sources=[cuda_source],\n functions=['logsumexp_cuda'],\n verbose=True\n)\n\nclass ModelNew(nn.Module):\n def __init__(self):\n super().__init__()\n self.module = module\n\n def forward(self, x: torch.Tensor) -> torch.Tensor:\n return self.module.logsumexp_cuda(x)\n" }

result ← submit_kernel (ok)

oksubmit_kernel submit_kernel PASSED: (3 / 3) correctness trials all passed. Kernel runtime: 0.01 μs Runtime stats: mean=0.01μs median=?μs std=0.00μs Numerical precision: max_abs_err=0.00e+00 mean_abs_err=0.00e+00 max_rel_err=0.00e+00 Memory: 33.01 MB peak (ref: 33.01 MB, ratio: 0.9998x) Kernel launches: 1 (ref: 20, fusion ratio: 20.0) SOL score: 0.001

final submitted kernel (65 lines, 1,640 chars) — full source, untruncated

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

cuda_source = '''
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cmath>

__global__ void logsumexp_kernel(const float* __restrict__ x, float* __restrict__ out, int R, int B, int D) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int total = B * D;
    if (idx >= total) return;
    int b = idx / D;
    int d = idx % D;
    
    float max_val = -INFINITY;
    for (int r = 0; r < R; ++r) {
        float val = x[r * B * D + b * D + d];
        max_val = fmaxf(max_val, val);
    }
    
    float sum = 0.0f;
    for (int r = 0; r < R; ++r) {
        float val = x[r * B * D + b * D + d];
        sum += expf(val - max_val);
    }
    
    out[b * D + d] = max_val + logf(sum);
}

torch::Tensor logsumexp_cuda(torch::Tensor x) {
    int R = x.size(0);
    int B = x.size(1);
    int D = x.size(2);
    auto out = torch::empty({B, D}, x.options());
    int total = B * D;
    int threads = 256;
    int blocks = (total + threads - 1) / threads;
    logsumexp_kernel<<<blocks, threads>>>(x.data_ptr<float>(), out.data_ptr<float>(), R, B, D);
    return out;
}
'''

cpp_source = '''
torch::Tensor logsumexp_cuda(torch::Tensor x);
'''

module = load_inline(
    name='logsumexp_module',
    cpp_sources=[cpp_source],
    cuda_sources=[cuda_source],
    functions=['logsumexp_cuda'],
    verbose=True
)

class ModelNew(nn.Module):
    def __init__(self):
        super().__init__()
        self.module = module

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.module.logsumexp_cuda(x)