Issues

• Three level-3 references give bit-identical output for different inputs, so they cannot test correctness at any tolerance

  ScalingIntelligence/KernelBench#172 · Filter by author zzw4257 opened on Sep 28, 2026

• The timing loop reuses one input tensor object, so a kernel that caches on input identity is timed from its cache

  ScalingIntelligence/KernelBench#171 · Filter by author zzw4257 opened on Sep 28, 2026

• score.py accepts runtime=-1.0 on correct rows, so geo_mean_speedup goes NaN or wrong and fast_p stops matching the correctness rate

  ScalingIntelligence/KernelBench#169 · Filter by author shaurya416 opened on Sep 23, 2026

• features request: cuda tile backend

  ScalingIntelligence/KernelBench#167 · Filter by author congchan opened on Sep 17, 2026

• Unchanged PyTorch or torch.compile-only submissions receive unstable fast_1 scores on Level 3/4

  ScalingIntelligence/KernelBench#166 · Filter by author congchan opened on Sep 17, 2026

• Correctness check can be passed by returning zeros on 4 of 60 level-1 problems

  ScalingIntelligence/KernelBench#165 · Filter by author gauravapiscean opened on Sep 16, 2026
  1

• eval_kernel_against_ref: model-load phase catches only RuntimeError

  ScalingIntelligence/KernelBench#159 · Filter by author MattFisher opened on Jul 9, 2026

• [Bug] groups parameter is ignored in Model due to positional argument misalignment

  ScalingIntelligence/KernelBench#157 · Filter by author diyarora opened on Jun 28, 2026

• Add MoE (expert-routing) operators — torch.compile structurally can't optimize them (gfx1151 data)

  ScalingIntelligence/KernelBench#156 · Filter by author fxp opened on Jun 11, 2026

• AMD gfx1151 (Strix Halo / RDNA3.5) ROCm results + a precision-downgrade reward-hack case for kernel_static_checker

  ScalingIntelligence/KernelBench#155 · Filter by author fxp opened on Jun 10, 2026

• [BUG] precision cast in _process_input_tensor breaks integer target tensors (cross_entropy / nll_loss / kl_div)

  ScalingIntelligence/KernelBench#151 · Filter by author moria97 opened on Apr 16, 2026

• [BUG] L2/98 Fused Triton Kernel Crashes with double free or corruption / malloc(): invalid next size

  ScalingIntelligence/KernelBench#147 · Filter by author wuyii8941 opened on Mar 26, 2026

• [Discussion/Feature] OOM on 24GB Consumer GPUs (e.g., RTX 4090) during Correctness Check due to massive hardcoded dimensions

  ScalingIntelligence/KernelBench#146 · Filter by author dunzane opened on Mar 23, 2026

• Feature Request: Open Leaderboard

  ScalingIntelligence/KernelBench#141 · Filter by author jaygala223 opened on Feb 26, 2026

• Allow GPU spec prompt configuration

  ScalingIntelligence/KernelBench#122 · Filter by author jiito opened on Jan 12, 2026

• Fix hackable case

  ScalingIntelligence/KernelBench#116 · Filter by author li-plus opened on Jan 6, 2026

• Max Perf / Speedup Modeling

  ScalingIntelligence/KernelBench#115 · Filter by author simonguozirui opened on Dec 31, 2025
  2

• Integrate kernel-level power usage data

  ScalingIntelligence/KernelBench#103 · Filter by author tanish111 opened on Dec 4, 2025
  2

• Validation Vulnerability in MSELoss: Partial Computation Can Pass Accuracy Check

  ScalingIntelligence/KernelBench#97 · Filter by author doux-jy opened on Nov 28, 2025
  2

• Would there be a potential Mojo Support

  ScalingIntelligence/KernelBench#94 · Filter by author rajuptvs opened on Nov 19, 2025
  4

• Weight initialization in custom Conv3d seems different from nn.Conv3d

  ScalingIntelligence/KernelBench#88 · Filter by author qy-777 opened on Nov 10, 2025
  3

• [torch.AcceleratorError] causes entire benchmark to crash

  ScalingIntelligence/KernelBench#76 · Filter by author ai-nikolai opened on Oct 28, 2025
  4

• [Roadmap] Fall 2025 KernelBench Maintenance + Improvement Plan

  documentationenhancementgood first issuehelp wanted
  ScalingIntelligence/KernelBench#74 · Filter by author simonguozirui opened on Oct 22, 2025
  2

  

• [Roadmap] Integration with Open-Source Frameworks

  enhancementhelp wanted
  Feature

  ScalingIntelligence/KernelBench#73 · Filter by author simonguozirui opened on Oct 22, 2025
  1

  

• Question：Why use torch.cuda.Event instead of torch.profiler for performance measurement?

  ScalingIntelligence/KernelBench#47 · Filter by author study-never-alone opened on Jun 24, 2025
  3