← Back to Leaderboard

The AI CUDA Engineer 👷

99_TripletMarginLoss • uniform_control_flow_triplet_base

Level 1 • Task 99
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    anchor: torch.Tensor, positive: torch.Tensor, negative: torch.Tensor, margin: float
) -> torch.Tensor:
    """
    Computes the Triplet Margin Loss for metric learning tasks.

    Args:
        anchor (torch.Tensor): Anchor values.
        positive (torch.Tensor): Positive values.
        negative (torch.Tensor): Negative values.
        margin (float): Margin value.

    Returns:
        torch.Tensor: Triplet Margin Loss.
    """
    return F.triplet_margin_loss(anchor, positive, negative, margin=margin)


class Model(nn.Module):
    """
    A model that computes Triplet Margin Loss for metric learning tasks.
    """

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

    def forward(self, anchor, positive, negative, fn=module_fn):
        return fn(anchor, positive, negative, self.margin)


batch_size = 128
input_shape = (4096,)
dim = 1
margin = 1.0


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


def get_init_inputs():
    return [margin]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that computes Triplet Margin Loss for metric learning tasks.

    Parameters:
        margin (float): The margin between the positive and negative samples.
    """
    def __init__(self, margin=1.0):
        super(Model, self).__init__()
        self.loss_fn = torch.nn.TripletMarginLoss(margin=margin)

    def forward(self, anchor, positive, negative):
        return self.loss_fn(anchor, positive, negative)

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

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

def get_init_inputs():
    return [1.0]  # Default margin
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 99_TripletMarginLoss
Level ID 1
Task ID 99
Kernel Name uniform_control_flow_triplet_base
CUDA Speedup (Native) 3.147x
CUDA Speedup (Compile) 2.128x
CUDA Runtime 0.016 ms
PyTorch Runtime (Native) 0.050 ms
PyTorch Runtime (Compile) 0.034 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.00
View Experiment Progress Details

Related Kernels (Level 1, Task 99 • 99_TripletMarginLoss)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 warp_shfl_tripletloss_base 0.01 3.87 2.62
🥇 vectorized_warp_shfl_reduction_edit_1 0.01 3.87 2.62
🥇 warp_shfl_tripletloss_edit_1 0.01 3.87 2.62
🥇 warp_optimized_tripletloss_edit_1 0.01 3.87 2.62
🥇 vectorized_warp_shfl_reduction_base 0.01 3.87 2.62
6 combined_warp_shfl_shared_memory_edit_1 0.01 3.60 2.43
6 combined_warp_shfl_shared_memory_base 0.01 3.60 2.43
6 warp_optimized_tripletloss_base 0.01 3.60 2.43
6 optimized_tripletloss_base 0.01 3.60 2.43
10 aligned_coalesced_triplet_base_base 0.01 3.36 2.27
10 ldg_aligned_triplet_base 0.01 3.36 2.27
10 block_size_experiment_triplet_loss_base 0.01 3.36 2.27
10 warp_shfl_triplet_loss_base 0.01 3.36 2.27
14 warp_shfl_triplet_loss_optimized_edit_2_edit_1 0.02 3.15 2.13
14 modular_aligned_triplet_base_base 0.02 3.15 2.13
14 optimized_triplet_margin_loss_base 0.02 3.15 2.13
14 uniform_control_flow_triplet_base 0.02 3.15 2.13
14 shared_memory_optimized_triplet_loss_base 0.02 3.15 2.13
14 shared_memory_optimized_triplet_loss_edit_1 0.02 3.15 2.13
20 hybrid_reduction_triplet_loss_base 0.02 2.96 2.00 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>

// Kernel using uniform control flow to minimize warp divergence

template <typename scalar_t>
__global__ void uniform_control_flow_triplet_kernel(
    const scalar_t* __restrict__ anchor,
    const scalar_t* __restrict__ positive,
    const scalar_t* __restrict__ negative,
    scalar_t* __restrict__ output,
    const float margin,
    const int batch_size,
    const int feat_size) {

    int batch_idx = blockIdx.x;
    if (batch_idx >= batch_size) return;

    int tid = threadIdx.x;
    int base_idx = batch_idx * feat_size;
    scalar_t local_dist_pos = 0;
    scalar_t local_dist_neg = 0;

    // Use vectorized loads for read-only global memory accesses with __ldg()
    if constexpr (std::is_same<scalar_t, float>::value) {
        using vec_t = float4;
        constexpr int vec_size = 4;
        int vectorized_length = feat_size / vec_size;
        int remainder = feat_size % vec_size;
        const vec_t* anchor_vec = reinterpret_cast<const vec_t*>(anchor + base_idx);
        const vec_t* positive_vec = reinterpret_cast<const vec_t*>(positive + base_idx);
        const vec_t* negative_vec = reinterpret_cast<const vec_t*>(negative + base_idx);

        for (int i = tid; i < vectorized_length; i += blockDim.x) {
            vec_t a_vec = __ldg(&anchor_vec[i]);
            vec_t p_vec = __ldg(&positive_vec[i]);
            vec_t n_vec = __ldg(&negative_vec[i]);
            
            float diff0 = a_vec.x - p_vec.x;
            float diff1 = a_vec.y - p_vec.y;
            float diff2 = a_vec.z - p_vec.z;
            float diff3 = a_vec.w - p_vec.w;
            local_dist_pos += diff0 * diff0 + diff1 * diff1 + diff2 * diff2 + diff3 * diff3;
            
            float diff0n = a_vec.x - n_vec.x;
            float diff1n = a_vec.y - n_vec.y;
            float diff2n = a_vec.z - n_vec.z;
            float diff3n = a_vec.w - n_vec.w;
            local_dist_neg += diff0n * diff0n + diff1n * diff1n + diff2n * diff2n + diff3n * diff3n;
        }

        int offset = vectorized_length * vec_size;
        for (int i = tid; i < remainder; i += blockDim.x) {
            int idx = base_idx + offset + i;
            float a = __ldg(&anchor[idx]);
            float p = __ldg(&positive[idx]);
            float n = __ldg(&negative[idx]);
            float diff = a - p;
            local_dist_pos += diff * diff;
            float diffn = a - n;
            local_dist_neg += diffn * diffn;
        }
    } else if constexpr (std::is_same<scalar_t, double>::value) {
        using vec_t = double2;
        constexpr int vec_size = 2;
        int vectorized_length = feat_size / vec_size;
        int remainder = feat_size % vec_size;
        const vec_t* anchor_vec = reinterpret_cast<const vec_t*>(anchor + base_idx);
        const vec_t* positive_vec = reinterpret_cast<const vec_t*>(positive + base_idx);
        const vec_t* negative_vec = reinterpret_cast<const vec_t*>(negative + base_idx);

        for (int i = tid; i < vectorized_length; i += blockDim.x) {
            vec_t a_vec = __ldg(&anchor_vec[i]);
            vec_t p_vec = __ldg(&positive_vec[i]);
            vec_t n_vec = __ldg(&negative_vec[i]);
            
            double diff0 = a_vec.x - p_vec.x;
            double diff1 = a_vec.y - p_vec.y;
            local_dist_pos += diff0 * diff0 + diff1 * diff1;
            
            double diff0n = a_vec.x - n_vec.x;
            double diff1n = a_vec.y - n_vec.y;
            local_dist_neg += diff0n * diff0n + diff1n * diff1n;
        }

        int offset = vectorized_length * vec_size;
        for (int i = tid; i < remainder; i += blockDim.x) {
            int idx = base_idx + offset + i;
            double a = __ldg(&anchor[idx]);
            double p = __ldg(&positive[idx]);
            double n = __ldg(&negative[idx]);
            double diff = a - p;
            local_dist_pos += diff * diff;
            double diffn = a - n;
            local_dist_neg += diffn * diffn;
        }
    } else {
        for (int i = tid; i < feat_size; i += blockDim.x) {
            int idx = base_idx + i;
            scalar_t a = __ldg(&anchor[idx]);
            scalar_t p = __ldg(&positive[idx]);
            scalar_t n = __ldg(&negative[idx]);
            scalar_t diff = a - p;
            local_dist_pos += diff * diff;
            scalar_t diffn = a - n;
            local_dist_neg += diffn * diffn;
        }
    }

    // Warp-level reduction within each block
    for (int offset = warpSize / 2; offset > 0; offset /= 2) {
        local_dist_pos += __shfl_down_sync(0xffffffff, local_dist_pos, offset);
        local_dist_neg += __shfl_down_sync(0xffffffff, local_dist_neg, offset);
    }

    __shared__ scalar_t shared_sum_pos[32];
    __shared__ scalar_t shared_sum_neg[32];

    int lane = tid % 32;
    int warp_id = tid / 32;

    if (lane == 0) {
        shared_sum_pos[warp_id] = local_dist_pos;
        shared_sum_neg[warp_id] = local_dist_neg;
    }
    __syncthreads();

    scalar_t block_sum_pos = 0;
    scalar_t block_sum_neg = 0;
    if (tid < (blockDim.x / 32)) {
        block_sum_pos = shared_sum_pos[lane];
        block_sum_neg = shared_sum_neg[lane];
        for (int offset = warpSize / 2; offset > 0; offset /= 2) {
            block_sum_pos += __shfl_down_sync(0xffffffff, block_sum_pos, offset);
            block_sum_neg += __shfl_down_sync(0xffffffff, block_sum_neg, offset);
        }
        if (lane == 0) {
            scalar_t loss = sqrt(block_sum_pos) - sqrt(block_sum_neg) + margin;
            output[batch_idx] = loss < scalar_t(0) ? scalar_t(0) : loss;
        }
    }
}

// Host function to launch the kernel
torch::Tensor triplet_margin_loss_cuda(
    torch::Tensor anchor,
    torch::Tensor positive,
    torch::Tensor negative,
    float margin) {

    TORCH_CHECK(anchor.device().is_cuda(), "anchor must be a CUDA tensor");
    TORCH_CHECK(positive.device().is_cuda(), "positive must be a CUDA tensor");
    TORCH_CHECK(negative.device().is_cuda(), "negative must be a CUDA tensor");

    const int batch_size = anchor.size(0);
    const int feat_size = anchor.size(1);

    auto output = torch::zeros({batch_size}, anchor.options());
    
    // Launch one block per batch sample; use 256 threads per block
    const int threads_per_block = 256;
    const int num_blocks = batch_size;
    
    AT_DISPATCH_FLOATING_TYPES(anchor.scalar_type(), "uniform_control_flow_triplet_kernel", ([&] {
        uniform_control_flow_triplet_kernel<scalar_t><<<num_blocks, threads_per_block>>>(
            anchor.data_ptr<scalar_t>(),
            positive.data_ptr<scalar_t>(),
            negative.data_ptr<scalar_t>(),
            output.data_ptr<scalar_t>(),
            margin,
            batch_size,
            feat_size);
    }));
    
    return output.mean();
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &triplet_margin_loss_cuda, "Triplet margin loss forward (CUDA)");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 0.280 inst/cycle 0.000 3
Executed Ipc Elapsed 0.170 inst/cycle 0.000 3
Issue Slots Busy 7.140 % 0.000 3
Issued Ipc Active 0.287 inst/cycle 0.000 3
SM Busy 7.140 % 0.000 3
Memory Throughput 983971740858.167 byte/second 206939709637644779520.000 3
Mem Busy 17.007 % 0.054 3
Max Bandwidth 29.473 % 0.195 3
L1/TEX Hit Rate 0.000 % 0.000 3
L2 Hit Rate 13.083 % 0.000 3
Mem Pipes Busy 2.040 % 0.001 3
Warp Cycles Per Issued Instruction 26.350 cycle 0.017 3
Warp Cycles Per Executed Instruction 26.600 cycle 0.017 3
Avg. Active Threads Per Warp 29.530 0.000 3
Avg. Not Predicated Off Threads Per Warp 28.940 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 8.000 block 0.000 3
Block Limit Shared Mem 25.000 block 0.000 3
Block Limit Warps 8.000 block 0.000 3
Theoretical Active Warps per SM 64.000 warp 0.000 3
Theoretical Occupancy 100.000 % 0.000 3
Achieved Occupancy 11.793 % 0.000 3
Achieved Active Warps Per SM 7.550 warp 0.000 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 (11.8%) 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 5130613.43 μs
Device Time 211985.55 μs
Self CPU Time 135164.11 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 5430705.61 μs
Device Time 7225947.71 μs
Self CPU Time 293551.45 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
aten::fill_
CPU Time 5137155.70 μs
Device Time 7225947.71 μs
Self CPU Time 353080.02 μs
Self Device Time 7225947.71 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 5453594.76 μs
Device Time 460778.07 μs
Self CPU Time 5453594.76 μs
Self Device Time 460778.07 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void uniform_control_flow_triplet_kernel<float>(float const*, float const*, float const*, float*, float, int, int)
CPU Time 0.00 μs
Device Time 583217.69 μs
Self CPU Time 0.00 μs
Self Device Time 583217.69 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::mean
CPU Time 1087790.85 μs
Device Time 372892.18 μs
Self CPU Time 691115.68 μs
Self Device Time 372892.18 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 7014039.77 μs
Self CPU Time 0.00 μs
Self Device Time 7014039.77 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 
45291 warnings generated when compiling for host.
Suppressed 45324 warnings (45277 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/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:10:5 bugprone-easily-swappable-parameters
10 | const scalar_t* __restrict__ anchor,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
11 | const scalar_t* __restrict__ positive,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
12 | const scalar_t* __restrict__ negative,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:10:34: note: the first parameter in the range is 'anchor'
10 | const scalar_t* __restrict__ anchor,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:12:34: note: the last parameter in the range is 'negative'
12 | const scalar_t* __restrict__ negative,
| ^~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:14:5: warning: 3 adjacent parameters of 'uniform_control_flow_triplet_kernel' of convertible types are easily swapped by mistake [bugprone-easily-swappable-parameters]
14 | const float margin,
| ^~~~~~~~~~~~~~~~~~~
15 | const int batch_size,
| ~~~~~~~~~~~~~~~~~~~~~
16 | const int feat_size) {
| ~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:14:17: note: the first parameter in the range is 'margin'
14 | const float margin,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:16:15: note: the last parameter in the range is 'feat_size'
16 | const int feat_size) {
| ^~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:15:5: note: 'const float' and 'const int' may be implicitly converted: 'const float' (as 'float') -> 'const int' (as 'int'), 'const int' (as 'int') -> 'const float' (as 'float')
15 | const int batch_size,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:18:21: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
18 | int batch_idx = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:21:15: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
21 | int tid = threadIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:36:55: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
36 | for (int i = tid; i < vectorized_length; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:55:47: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
55 | for (int i = tid; i < remainder; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:74:55: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
74 | for (int i = tid; i < vectorized_length; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:89:47: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
89 | for (int i = tid; i < remainder; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:100:47: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
100 | for (int i = tid; i < feat_size; i += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:157:28: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
157 | const int batch_size = anchor.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:158:27: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
158 | const int feat_size = anchor.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250204_optimize_b10_s4_e0_sweep/level_1/task_99/b7_s0_uniform_control_flow_triplet/base/base.cu:166:5: warning: inside a lambda, '__func__' expands to the name of the function call operator; consider capturing the name of the enclosing function explicitly [bugprone-lambda-function-name]
166 | AT_DISPATCH_FLOATING_TYPES(anchor.scalar_type(), "uniform_control_flow_triplet_kernel", ([&] {
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:237:34: note: expanded from macro 'AT_DISPATCH_FLOATING_TYPES'
237 | AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:233:3: note: expanded from macro 'AT_DISPATCH_CASE_FLOATING_TYPES'
233 | AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:74:3: note: expanded from macro 'AT_DISPATCH_CASE'
74 | AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, scalar_t, __VA_ARGS__)
| ^
note: (skipping 1 expansions in backtrace; use -fmacro-backtrace-limit=0 to see all)
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h:58:7: note: expanded from macro 'AT_PRIVATE_CHECK_SELECTIVE_BUILD'
58 | AT_ERROR( \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:711:32: note: expanded from macro 'AT_ERROR'
711 | C10_EXPAND_MSVC_WORKAROUND(TORCH_CHECK(false, ::c10::str(__VA_ARGS__))); \
| ^
/home/robert_sakana_ai/miniconda3/envs/llm2cuda/lib/python3.11/site-packages/torch/include/c10/util/Exception.h:536:9: note: expanded from macro 'TORCH_CHECK'
536 | __func__, \
| ^