VRAM Lab publication evidence: PyTorch 2.13 CUDA wheels on RTX 5060 Ti
Measurement date: 2026-08-24 (Asia/Seoul)
This file is generated deterministically from frozen experiment evidence.

[scope]
One Windows host, one WSL2 guest, one installed environment per wheel, and three fresh Python processes per cell.
Correctness and compatibility were measured. The preregistered timing protocol was not run, so this log makes no performance comparison.

[wrapper audit]
wrapper_records=195
pass_records=150
expected_exception_records=45
expected_no_kernel_exceptions=33
expected_missing_c_compiler_exceptions=6
expected_missing_python_header_exceptions=6
timeout_records=0
signal_records=0
json_parse_failures=0
post_failure_health_checks=45
post_failure_health_temperature_c=41..60
summary_files=10

[environment]
host_os=Windows 11 Home 25H2, build 26200.9168
host_cpu=AMD Ryzen 7 7800X3D, 8 cores / 16 logical processors
host_ram=16GB DDR5-5200, single module
host_gpu=NVIDIA GeForce RTX 5060 Ti
gpu_architecture=Blackwell, compute capability 12.0 (sm_120)
gpu_compute_capability=12.0
gpu_memory_bytes=8546484224
gpu_memory_pytorch_mib=8150.5625
gpu_memory_nvidia_smi_mib=8151
windows_nvidia_driver=610.88
power_plan=Balanced
wsl_version=2.7.12.0
guest_release=Ubuntu 24.04.4 LTS
guest_kernel=6.18.33.2-microsoft-standard-WSL2
guest_python=3.12.3
guest_memory_bytes=7731843072
guest_root_filesystem_bytes=1081101176832
guest_root_available_bytes_at_phase0=1024509526016
phase0_capture_utc=2026-08-23T17:21:06Z
phase0_gpu_bridge=present
phase0_nvcc=absent
phase0_linux_display_driver_packages=none
note=cuda-toolkit package names below are wheel-managed Python packages, not a system CUDA Toolkit install

[wheel installs]
cu126: torch=2.13.0+cu126; embedded_cuda=12.6; triton=3.7.1; pip=24.0
cu126: official_index=https://download.pytorch.org/whl/cu126
cu126: resolved_wheel=torch-2.13.0+cu126-cp312-cp312-manylinux_2_28_x86_64.whl
cu126: resolved_wheel_sha256=8695f3c6b7966d44560275b90c5c28e5091ba33ddbb1ab33b2173782ca1e9145
cu126: pip_check=No broken requirements found.
cu126: cuda_available=true; device_count=1; compiled_arches=sm_50,sm_60,sm_70,sm_75,sm_80,sm_86,sm_90
cu130: torch=2.13.0+cu130; embedded_cuda=13.0; triton=3.7.1; pip=24.0
cu130: official_index=https://download.pytorch.org/whl/cu130
cu130: resolved_wheel=torch-2.13.0+cu130-cp312-cp312-manylinux_2_28_x86_64.whl
cu130: resolved_wheel_sha256=8db7338e6895c3d4bd89a02ff4209507d1f0cf2ffeb3b898538b5a07d1ea8c1e
cu130: pip_check=No broken requirements found.
cu130: cuda_available=true; device_count=1; compiled_arches=sm_75,sm_80,sm_86,sm_90,sm_100,sm_120
cu132: torch=2.13.0+cu132; embedded_cuda=13.2; triton=3.7.1; pip=24.0
cu132: official_index=https://download.pytorch.org/whl/cu132
cu132: resolved_wheel=torch-2.13.0+cu132-cp312-cp312-manylinux_2_28_x86_64.whl
cu132: resolved_wheel_sha256=4c1800e216b46bae03483bc94dd73a23ea43c265ab51f865e28070ca045d10a8
cu132: pip_check=No broken requirements found.
cu132: cuda_available=true; device_count=1; compiled_arches=sm_75,sm_80,sm_86,sm_90,sm_100,sm_120

[full pip freezes]
-- cu126 --
cuda-bindings==12.9.7
cuda-pathfinder==1.6.0
cuda-toolkit==12.6.3
filelock==3.32.3
fsspec==2026.7.0
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
nvidia-cublas-cu12==12.6.4.1
nvidia-cuda-cupti-cu12==12.6.80
nvidia-cuda-nvrtc-cu12==12.6.85
nvidia-cuda-runtime-cu12==12.6.77
nvidia-cudnn-cu12==9.10.2.21
nvidia-cufft-cu12==11.3.0.4
nvidia-cufile-cu12==1.11.1.6
nvidia-curand-cu12==10.3.7.77
nvidia-cusolver-cu12==11.7.1.2
nvidia-cusparse-cu12==12.5.4.2
nvidia-cusparselt-cu12==0.7.1
nvidia-nccl-cu12==2.29.3
nvidia-nvjitlink-cu12==12.6.85
nvidia-nvshmem-cu12==3.4.5
nvidia-nvtx-cu12==12.6.77
pip==24.0
setuptools==78.1.0
sympy==1.14.0
torch==2.13.0+cu126
triton==3.7.1
typing_extensions==4.16.0
-- cu130 --
cuda-bindings==13.3.1
cuda-pathfinder==1.6.0
cuda-toolkit==13.0.3.0
filelock==3.32.3
fsspec==2026.7.0
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
nvidia-cublas==13.1.1.3
nvidia-cuda-cupti==13.0.85
nvidia-cuda-nvrtc==13.0.88
nvidia-cuda-runtime==13.0.96
nvidia-cudnn-cu13==9.20.0.48
nvidia-cufft==12.0.0.61
nvidia-cufile==1.15.1.6
nvidia-curand==10.4.0.35
nvidia-cusolver==12.0.4.66
nvidia-cusparse==12.6.3.3
nvidia-cusparselt-cu13==0.8.1
nvidia-nccl-cu13==2.29.7
nvidia-nvjitlink==13.3.33
nvidia-nvshmem-cu13==3.4.5
nvidia-nvtx==13.0.85
pip==24.0
setuptools==78.1.0
sympy==1.14.0
torch==2.13.0+cu130
triton==3.7.1
typing_extensions==4.16.0
-- cu132 --
cuda-bindings==13.3.1
cuda-pathfinder==1.6.0
cuda-toolkit==13.2.1
filelock==3.32.3
fsspec==2026.7.0
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
nvidia-cublas==13.4.0.1
nvidia-cuda-cupti==13.2.75
nvidia-cuda-nvrtc==13.2.78
nvidia-cuda-runtime==13.2.75
nvidia-cudnn-cu13==9.20.0.48
nvidia-cufft==12.2.0.46
nvidia-cufile==1.17.1.22
nvidia-curand==10.4.2.55
nvidia-cusolver==12.2.0.1
nvidia-cusparse==12.7.10.1
nvidia-cusparselt-cu13==0.8.1
nvidia-nccl-cu13==2.29.7
nvidia-nvjitlink==13.3.33
nvidia-nvshmem-cu13==3.4.5
nvidia-nvtx==13.2.75
pip==24.0
setuptools==78.1.0
sympy==1.14.0
torch==2.13.0+cu132
triton==3.7.1
typing_extensions==4.16.0

[preregistered gate]
cu130_positive_control=9/9 PASS (metadata, 1-byte alloc/fill/read, FP32 512x512 GEMM)
original_allocator_scout=default 3/3 PASS; expandable_segments 3/3 PASS; cudaMallocAsync 3/3 PASS
original_allocator_scout_actual_request=1 byte; the runner carried size=1024 but operation_alloc ignored it
expandable_error_999_signature=false
topic_pivot_recommended=false
post_gate_gpu_health_check=PASS

[original eager core matrix]
Each cell is three fresh processes. Compile is secondary and excluded from this matrix.
operation | cu126 | cu130 | cu132
----------|-------|-------|------
metadata | 3/3 PASS | 3/3 PASS | 3/3 PASS
1-byte alloc/fill/read | 0/3 FAIL | 3/3 PASS | 3/3 PASS
FP32 vector add, n=4096 | 0/3 FAIL | 3/3 PASS | 3/3 PASS
FP32 dot, n=1 | 3/3 PASS | 3/3 PASS | 3/3 PASS
FP32 dot, n=4 | 3/3 PASS | 3/3 PASS | 3/3 PASS
FP32 dot, n=32 | 3/3 PASS | 3/3 PASS | 3/3 PASS
FP32 dot, n=1024 | 3/3 PASS | 3/3 PASS | 3/3 PASS
FP32 (x*y).sum(), n=4 | 0/3 FAIL | 3/3 PASS | 3/3 PASS
FP32 ones GEMM, 512x512 | 0/3 FAIL | 3/3 PASS | 3/3 PASS
BF16 ones GEMM, 512x512 | 0/3 FAIL | 3/3 PASS | 3/3 PASS
TOTAL | 15/30 | 30/30 | 30/30
OVERALL | 75/90 | |

[original core signatures]
cu126: torch.cuda.is_available() was true, device_count was 1, and the RTX 5060 Ti reported capability 12.0.
cu126: compiled_arches=sm_50,sm_60,sm_70,sm_75,sm_80,sm_86,sm_90
cu130: compiled_arches=sm_75,sm_80,sm_86,sm_90,sm_100,sm_120
cu132: compiled_arches=sm_75,sm_80,sm_86,sm_90,sm_100,sm_120
cu126 exact failure line for all five 0/3 operation families:
CUDA error: no kernel image is available for execution on the device
cu126 failing families: 1-byte alloc/fill/read; FP32 vector add, n=4096; FP32 (x*y).sum(), n=4; FP32 ones GEMM, 512x512; BF16 ones GEMM, 512x512
All original dot cells passed 3/3 for n=1, 4, 32, and 1024 on every wheel.
Do not convert the 15/30 cu126 count into a percentage-compatibility claim; the cells are not weighted workload samples.

[post-hoc path disambiguation]
These runs describe why the original cu126 cells split. They do not replace the original core matrix.
operation | cu126 | cu130
----------|-------|------
allocate one byte, synchronize, no fill | 3/3 PASS | 3/3 PASS
allocate, GPU fill, read back | 0/3 FAIL | 3/3 PASS
CPU byte to GPU and back | 3/3 PASS | 3/3 PASS
CPU-prepared FP32 vector, GPU add | 0/3 FAIL | 3/3 PASS
CPU-prepared FP32 512x512 GEMM | 3/3 PASS | 3/3 PASS
cu126 allocation alone passed; its GPU fill and vector-add kernels failed with the no-kernel-image signature.
cu126 host-to-device copies and the CPU-prepared FP32 GEMM passed. No mechanism beyond these observed paths is inferred here.

[post-hoc audit corrections]
Corrected 1024-byte cu130 allocator scout:
default: 3/3 PASS; requested=1024 bytes; allocated=1024 bytes; reserved=2097152 bytes
expandable_segments: 3/3 PASS; requested=1024 bytes; allocated=1024 bytes; reserved=2097152 bytes
cudaMallocAsync: 3/3 PASS; requested=1024 bytes; allocated=1024 bytes; reserved=33554432 bytes
CUDA driver attribute 110 (GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED): values=0,0,0; query calls passed 3/3.
GPU-generated n=4 dot setup:
cu126: 0/3 FAIL; torch.rand(device=cuda) failed before torch.dot ran
cu130: 3/3 PASS
cu132: 3/3 PASS
Corrected BF16 random-reference GEMM, 128x128, CPU-prepared quantized inputs:
cu126: 3/3 PASS; max_abs_error=0.1247406005859375; max_abs_reference=51.5284423828125; tolerance=1.03056884765625
cu130: 3/3 PASS; max_abs_error=0.1247406005859375; max_abs_reference=51.5284423828125; tolerance=1.03056884765625
cu132: 3/3 PASS; max_abs_error=0.1247406005859375; max_abs_reference=51.5284423828125; tolerance=1.03056884765625

[protocol deviations and disposition]
1. The original allocator scout was labeled size=1024 by the runner, but operation_alloc always requested one byte. The corrected 1024-byte scout is reported separately above.
2. The original BF16 core probe used 512x512 ones and an exact expected value, not the planned seeded random inputs and CPU FP32 reference. The corrected 128x128 reference probe is separate.
3. The original dot sweep generated inputs on CPU and copied them to CUDA. The GPU-generated-input follow-up is separate, and cu126 failed during input generation before dot.
4. The preregistered timing protocol was not run. No timing or speed result is publishable from this experiment.

[secondary torch.compile prerequisite stages]
The same elementwise compile smoke used n=4096 and three fresh processes per wheel at each stage.
stage | cu126 | cu130 | cu132
------|-------|-------|------
minimal guest, no C compiler | 0/3 FAIL | 0/3 FAIL | 0/3 FAIL
after g++ only | 0/3 FAIL | 0/3 FAIL | 0/3 FAIL
after g++ plus Python headers | 0/3 FAIL | 3/3 PASS | 3/3 PASS
stage 0 cu130/cu132 exact first line:
RuntimeError: Failed to find C compiler. Please specify via CC environment variable or set triton.knobs.build.impl.
stage 1 cu130/cu132 compiler line after removing the temporary source path:
fatal error: Python.h: No such file or directory
cu126 stopped at eager torch.linspace(device=cuda) with the no-kernel-image signature at all three stages, before compiled code could run.
cu130/cu132 produced correct outputs with maximum_error=0.0 after both the C compiler and Python headers were present.
C compiler package evidence: g++=4:13.2.0-7ubuntu1; g++-13=13.3.0-6ubuntu2~24.04.1; gcc=4:13.2.0-7ubuntu1; gcc-13=13.3.0-6ubuntu2~24.04.1; libc6-dev=2.39-0ubuntu8.8
Python header package evidence: libpython3.12-dev=3.12.3-1ubuntu0.15; python3-dev=3.12.3-0ubuntu2.1; python3.12-dev=3.12.3-1ubuntu0.15

[claim boundaries]
The results show repeatability on one host, not independent-machine generality.
A true CUDA availability flag is enumeration/initialization evidence, not proof that every operation path has a compatible kernel.
cu130 and cu132 passed every tested eager core cell; this is not a claim that every PyTorch or CUDA kernel works.
cu132 was tested for correctness only; no speed or stability advantage is claimed.
The reported RTX 5060 Ti GPU-rand-to-dot crash from another PyTorch/CUDA build did not reproduce in the GPU-generated-input cu130/cu132 follow-up on this fixed host and build. cu126 failed during torch.rand before dot ran, and the issue is not called fixed everywhere.

[frozen evidence integrity]
generator_input_file_count=285
generator_input_manifest_sha256=ddff511e0b1b0241cbb2a3fa94edd5002be8383399ae9a96055b11398b053dac
data/results/gate-summary.json sha256=c72314adc845ffde2284fb883f7d980aff07d1bfc17df5650ff8e4c1025af40c
data/results/matrix-summary.json sha256=0f90d09bf36d52d56224faf2da2524032f920a24740994a555a297532dbee18a
data/results/disambiguation-summary.json sha256=fad75f017b36b6bc4f015e67d2192f2dac0e393b25b28e895218e003f5c51fe0
data/results/audit-corrections-summary.json sha256=bf4ec395ee83fd02e055f7cd7ed50c97442239ee265cfe8eedfac4adfadf22e7
data/results/compile-headers-summary.json sha256=60ee137f50cfb0eddb0a1817f5f8ea93f2ae6448e5e642e59588fd033c335f8b