./ vs Torch - Relative Speed
PyTorch 2.11.0+cu130 ยท CUDA 13.0
GrpoLossBenchmark.grpo_loss_batch128_seqlen02781_compiled
2.66x
GrpoLossBenchmark.grpo_loss_batch128_seqlen08192_compiled
2.45x
GrpoLossBenchmark.grpo_loss_batch16_seqlen01024_compiled
1.94x
GrpoLossBenchmark.grpo_loss_batch32_seqlen02048_compiled
1.95x
GrpoLossBenchmark.grpo_loss_batch64_seqlen04096_compiled
2.49x
GrpoLossBenchmark.grpo_loss_fwd_batch128_seqlen02781_compiled
2.34x
GrpoLossBenchmark.grpo_loss_fwd_batch128_seqlen08192_compiled
2.32x
GrpoLossBenchmark.grpo_loss_fwd_batch16_seqlen01024_compiled
2.56x
GrpoLossBenchmark.grpo_loss_fwd_batch32_seqlen02048_compiled
3.10x
GrpoLossBenchmark.grpo_loss_fwd_batch64_seqlen04096_compiled
2.34x
Kernel
Torch (ref)