← Back to Leaderboard

The AI CUDA Engineer 👷

97_CosineSimilarityLoss • strided_cosine_loss_base_base

Level 1 • Task 97
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
    """
    Computes the Cosine Similarity Loss for comparing vectors.

    Args:
        predictions (torch.Tensor): Predicted values.
        targets (torch.Tensor): Target values.

    Returns:
        torch.Tensor: Cosine Similarity Loss.
    """
    cosine_sim = F.cosine_similarity(predictions, targets, dim=1)
    return torch.mean(1 - cosine_sim)


class Model(nn.Module):
    """
    A model that computes Cosine Similarity Loss for comparing vectors.

    Parameters:
        None
    """

    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets, fn=module_fn):
        return fn(predictions, targets)


batch_size = 128
input_shape = (4096,)
dim = 1


def get_inputs():
    return [
        torch.randn(batch_size, *input_shape),
        torch.randn(batch_size, *input_shape),
    ]


def get_init_inputs():
    return []

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that computes Cosine Similarity Loss for comparing vectors.

    Parameters:
        None
    """
    def __init__(self):
        super(Model, self).__init__()

    def forward(self, predictions, targets):
        cosine_sim = torch.nn.functional.cosine_similarity(predictions, targets, dim=1)
        return torch.mean(1 - cosine_sim)

batch_size = 128
input_shape = (4096, )
dim = 1

def get_inputs():
    return [torch.randn(batch_size, *input_shape), torch.randn(batch_size, *input_shape)]

def get_init_inputs():
    return []
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 97_CosineSimilarityLoss
Level ID 1
Task ID 97
Kernel Name strided_cosine_loss_base_base
CUDA Speedup (Native) 6.948x
CUDA Speedup (Compile) 4.977x
CUDA Runtime 0.011 ms
PyTorch Runtime (Native) 0.076 ms
PyTorch Runtime (Compile) 0.055 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
Temperature 1.00
View Experiment Progress Details

Related Kernels (Level 1, Task 97 • 97_CosineSimilarityLoss)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 blocksize_tuning_cosine_loss_base 0.01 7.64 5.47
🥈 optimized_adaptive_block_cosine_similarity_loss_base 0.01 6.95 4.98
🥈 block_tuned_cosine_loss_base_base 0.01 6.95 4.98
🥈 optimal_cosine_similarity_loss_base 0.01 6.95 4.98
🥈 ldg_aligned_cosine_loss_base_base 0.01 6.95 4.98
🥈 block_size_tuning_cosine_loss_base 0.01 6.95 4.98
🥈 coalesced_cosine_loss_opt_base 0.01 6.95 4.98
🥈 efficient_cosinesim_loss_base 0.01 6.95 4.98
🥈 adaptive_block_cosine_loss_base 0.01 6.95 4.98
🥈 unrolled_vectorized_cosine_loss_base_base 0.01 6.95 4.98
🥈 sm_shfl_reduction_cosine_loss_base 0.01 6.95 4.98
🥈 optimized_adaptive_block_cosine_similarity_loss_base 0.01 6.95 4.98
🥈 97_cosine_similarity_loss_shared_memory_base 0.01 6.95 4.98
🥈 97_cosine_similarity_loss_shared_memory_edit_1 0.01 6.95 4.98
🥈 97_cosine_similarity_loss_shared_memory_optimized_edit_1 0.01 6.95 4.98
🥈 strided_cosine_loss_base_base 0.01 6.95 4.98
🥈 warp_shfl_shared_opt_base 0.01 6.95 4.98
🥈 97_cosine_similarity_loss_shared_memory_optimized_base 0.01 6.95 4.98
🥈 97_cosine_similarity_loss_constant_memory_base 0.01 6.95 4.98
20 97_cosine_similarity_loss_modular_edit_1 0.01 6.37 4.56 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

__forceinline__ __device__ float warp_reduce(float val) {
    for (int offset = warpSize/2; offset > 0; offset /= 2) {
        val += __shfl_down_sync(0xffffffff, val, offset);
    }
    return val;
}

__global__ void strided_cosine_similarity_loss_kernel(
    const float* __restrict__ predictions,
    const float* __restrict__ targets,
    float* output,
    const int N,
    const int D
) {
    const int row = blockIdx.x;
    const int tid = threadIdx.x;
    const int wid = tid / warpSize;
    const int lane = tid % warpSize;
    const int warps_per_block = blockDim.x / warpSize;
    
    // Use float4 for coalesced memory access
    const int vec_size = 4;
    const int D_vec = D / vec_size;
    const int D_remainder = D % vec_size;
    
    // Calculate optimal stride for the vectorized portion
    const int stride = (D_vec + blockDim.x - 1) / blockDim.x * vec_size;
    
    float sum_dot = 0.0f;
    float sum_pred_sq = 0.0f;
    float sum_target_sq = 0.0f;

    // Process vectorized elements with stride
    const float4* pred_vec = reinterpret_cast<const float4*>(predictions + row * D);
    const float4* targ_vec = reinterpret_cast<const float4*>(targets + row * D);
    
    #pragma unroll 4
    for (int idx = tid; idx < D_vec; idx += blockDim.x) {
        float4 pred = pred_vec[idx];
        float4 targ = targ_vec[idx];
        
        sum_dot += pred.x * targ.x + pred.y * targ.y + 
                  pred.z * targ.z + pred.w * targ.w;
        sum_pred_sq += pred.x * pred.x + pred.y * pred.y + 
                      pred.z * pred.z + pred.w * pred.w;
        sum_target_sq += targ.x * targ.x + targ.y * targ.y + 
                        targ.z * targ.z + targ.w * targ.w;
    }

    // Handle remaining elements
    const int rem_start = D_vec * vec_size;
    #pragma unroll
    for (int idx = rem_start + tid; idx < D; idx += blockDim.x) {
        float pred = predictions[row * D + idx];
        float targ = targets[row * D + idx];
        sum_dot += pred * targ;
        sum_pred_sq += pred * pred;
        sum_target_sq += targ * targ;
    }

    // Warp-level reduction
    sum_dot = warp_reduce(sum_dot);
    sum_pred_sq = warp_reduce(sum_pred_sq);
    sum_target_sq = warp_reduce(sum_target_sq);

    // Block-level reduction using shared memory
    __shared__ float s_dot[32];        // One element per warp
    __shared__ float s_pred_sq[32];
    __shared__ float s_target_sq[32];

    if (lane == 0) {
        s_dot[wid] = sum_dot;
        s_pred_sq[wid] = sum_pred_sq;
        s_target_sq[wid] = sum_target_sq;
    }
    __syncthreads();

    // Final reduction by first warp
    if (tid < warps_per_block) {
        sum_dot = (tid < warps_per_block) ? s_dot[tid] : 0.0f;
        sum_pred_sq = (tid < warps_per_block) ? s_pred_sq[tid] : 0.0f;
        sum_target_sq = (tid < warps_per_block) ? s_target_sq[tid] : 0.0f;

        sum_dot = warp_reduce(sum_dot);
        sum_pred_sq = warp_reduce(sum_pred_sq);
        sum_target_sq = warp_reduce(sum_target_sq);

        if (tid == 0) {
            const float eps = 1e-8f;
            float norm_pred = sqrtf(sum_pred_sq);
            float norm_target = sqrtf(sum_target_sq);
            float denominator = norm_pred * norm_target;
            denominator = fmaxf(denominator, eps);
            float cos_sim = sum_dot / denominator;
            atomicAdd(output, (1.0f - cos_sim) / N);
        }
    }
}

torch::Tensor strided_cosine_similarity_loss_forward(
    torch::Tensor predictions,
    torch::Tensor targets
) {
    TORCH_CHECK(predictions.dim() == 2, "predictions must be 2D");
    TORCH_CHECK(targets.dim() == 2, "targets must be 2D");
    TORCH_CHECK(predictions.sizes() == targets.sizes(), "Input tensors must have the same shape");
    TORCH_CHECK(predictions.scalar_type() == torch::kFloat32, "predictions must be float32");
    TORCH_CHECK(targets.scalar_type() == torch::kFloat32, "targets must be float32");

    int N = predictions.size(0);
    int D = predictions.size(1);

    auto output = torch::zeros({1}, predictions.options());
    
    // Select block size based on dimension size
    const int block_size = (D <= 256) ? 256 : 512;
    
    strided_cosine_similarity_loss_kernel<<<N, block_size>>>(
        predictions.data_ptr<float>(),
        targets.data_ptr<float>(),
        output.data_ptr<float>(),
        N, D
    );

    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &strided_cosine_similarity_loss_forward, "Strided Cosine Similarity Loss Forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.660 inst/cycle 0.000 3
Executed Ipc Elapsed 0.403 inst/cycle 0.000 3
Issue Slots Busy 16.823 % 0.255 3
Issued Ipc Active 0.673 inst/cycle 0.001 3
SM Busy 16.823 % 0.255 3
Memory Throughput 713492310557.173 byte/second 10024541466593867776.000 3
Mem Busy 12.343 % 0.004 3
Max Bandwidth 21.403 % 0.018 3
L1/TEX Hit Rate 0.000 % 0.000 3
L2 Hit Rate 18.567 % 0.000 3
Mem Pipes Busy 4.040 % 0.000 3
Warp Cycles Per Issued Instruction 19.747 cycle 0.000 3
Warp Cycles Per Executed Instruction 20.073 cycle 0.000 3
Avg. Active Threads Per Warp 28.950 0.000 3
Avg. Not Predicated Off Threads Per Warp 27.730 0.000 3
Max Active Clusters 0.000 cluster 0.000 3
Max Cluster Size 8.000 block 0.000 3
Overall GPU Occupancy 0.000 % 0.000 3
Cluster Occupancy 0.000 % 0.000 3
Block Limit SM 32.000 block 0.000 3
Block Limit Registers 4.000 block 0.000 3
Block Limit Shared Mem 11.000 block 0.000 3
Block Limit Warps 4.000 block 0.000 3
Theoretical Active Warps per SM 64.000 warp 0.000 3
Theoretical Occupancy 100.000 % 0.000 3
Achieved Occupancy 21.283 % 0.012 3
Achieved Active Warps Per SM 13.623 warp 0.005 3
Analysis Rules
Rule Description
WRN HighPipeUtilization All compute pipelines are under-utilized. Either this kernel is very small or it doesn't issue enough warps per scheduler. Check the Launch Statistics and Scheduler Statistics sections for further details.
INF CPIStall Check the Warp Stall Sampling (All Cycles) table for the top stall locations in your source based on sampling data. The Kernel Profiling Guide (https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-reference) provides more details on each stall reason.
WRN Occupancy This kernel's theoretical occupancy is not impacted by any block limit. The difference between calculated theoretical (100.0%) and measured achieved occupancy (21.2%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide (https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on optimizing occupancy.
Operation / Metric Value Unit
aten::zeros
CPU Time 6048143.28 μs
Device Time 231087.66 μs
Self CPU Time 142218.67 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::zero_
CPU Time 6410190.42 μs
Device Time 7781961.02 μs
Self CPU Time 327242.96 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::fill_
CPU Time 6082952.25 μs
Device Time 7781961.02 μs
Self CPU Time 404471.46 μs
Self Device Time 7781958.49 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaLaunchKernel
CPU Time 6061049.16 μs
Device Time 2933.90 μs
Self CPU Time 6061049.16 μs
Self Device Time 2933.90 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
strided_cosine_similarity_loss_kernel(float const*, float const*, float*, int, int)
CPU Time 0.00 μs
Device Time 530092.45 μs
Self CPU Time 0.00 μs
Self Device Time 530092.45 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaEventRecord
CPU Time 274316.61 μs
Device Time 1252313.23 μs
Self CPU Time 274316.61 μs
Self Device Time 1252313.23 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, at::detail::Array<char*, 1> >(int, at::native::FillFunctor<int>, at::detail::Array<char*, 1>)
CPU Time 0.00 μs
Device Time 7550873.36 μs
Self CPU Time 0.00 μs
Self Device Time 7550873.36 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaEventElapsedTime
CPU Time 336868.45 μs
Device Time 0.00 μs
Self CPU Time 336868.45 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Completed 
45292 warnings generated when compiling for host.
Suppressed 45322 warnings (45275 in non-user code, 47 NOLINT).
Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:13:5 bugprone-easily-swappable-parameters
13 | const float* __restrict__ predictions,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
14 | const float* __restrict__ targets,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:13:31: note: the first parameter in the range is 'predictions'
13 | const float* __restrict__ predictions,
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:14:31: note: the last parameter in the range is 'targets'
14 | const float* __restrict__ targets,
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:16:5: warning: 2 adjacent parameters of 'strided_cosine_similarity_loss_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
16 | const int N,
| ^~~~~~~~~~~~
17 | const int D
| ~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:16:15: note: the first parameter in the range is 'N'
16 | const int N,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:17:15: note: the last parameter in the range is 'D'
17 | const int D
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:19:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
19 | const int row = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:20:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
20 | const int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:23:33: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
23 | const int warps_per_block = blockDim.x / warpSize;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:28:15: warning: Value stored to 'D_remainder' during its initialization is never read [clang-analyzer-deadcode.DeadStores]
28 | const int D_remainder = D % vec_size;
| ^~~~~~~~~~~ ~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:28:15: note: Value stored to 'D_remainder' during its initialization is never read
28 | const int D_remainder = D % vec_size;
| ^~~~~~~~~~~ ~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:31:15: warning: Value stored to 'stride' during its initialization is never read [clang-analyzer-deadcode.DeadStores]
31 | const int stride = (D_vec + blockDim.x - 1) / blockDim.x * vec_size;
| ^~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:31:15: note: Value stored to 'stride' during its initialization is never read
31 | const int stride = (D_vec + blockDim.x - 1) / blockDim.x * vec_size;
| ^~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:31:24: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
31 | const int stride = (D_vec + blockDim.x - 1) / blockDim.x * vec_size;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:38:62: warning: result of multiplication in type 'int' is used as a pointer offset after an implicit widening conversion to type 'ptrdiff_t' [bugprone-implicit-widening-of-multiplication-result]
38 | const float4* pred_vec = reinterpret_cast<const float4*>(predictions + row * D);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:38:76: note: make conversion explicit to silence this warning
4 | const float4* pred_vec = reinterpret_cast<const float4*>(predictions + row * D);
| ^~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:38:76: note: perform multiplication in a wider type
38 | const float4* pred_vec = reinterpret_cast<const float4*>(predictions + row * D);
| ^~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:39:62: warning: result of multiplication in type 'int' is used as a pointer offset after an implicit widening conversion to type 'ptrdiff_t' [bugprone-implicit-widening-of-multiplication-result]
39 | const float4* targ_vec = reinterpret_cast<const float4*>(targets + row * D);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:39:72: note: make conversion explicit to silence this warning
39 | const float4* targ_vec = reinterpret_cast<const float4*>(targets + row * D);
| ^~~~~~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:39:72: note: perform multiplication in a wider type
39 | const float4* targ_vec = reinterpret_cast<const float4*>(targets + row * D);
| ^~~
| static_cast<ptrdiff_t>( )
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:42:45: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
42 | for (int idx = tid; idx < D_vec; idx += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:57:53: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
57 | for (int idx = rem_start + tid; idx < D; idx += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:99:50: warning: narrowing conversion from 'int' to 'float' [bugprone-narrowing-conversions]
99 | atomicAdd(output, (1.0f - cos_sim) / N);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:105:19: warning: the parameter 'predictions' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
105 | torch::Tensor predictions,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:106:19: warning: the parameter 'targets' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
106 | torch::Tensor targets
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:114:13: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
114 | int N = predictions.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250202_optimize_b10_s4_e0_sweep/level_1/task_97/b10_s3_strided_cosine_loss_base/base/base.cu:115:13: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
115 | int D = predictions.size(1);
| ^