l5_hw_translation_single_gpt · original · 10 trajectories
L5 P1 01_paged_attention_v1.py
incorrect —
L5 P2 02_paged_attention_v2.py
correct 1.90x 0.1μs
L5 P3 03_fused_rmsnorm.py
correct 6.64x 0.0μs
L5 P4 04_swiglu_activation.py
correct 1.69x 0.0μs
L5 P5 05_rotary_embedding.py
correct 5.47x 0.0μs
L5 P6 06_custom_allreduce.py
correct 1.00x 0.0μs
L5 P7 07_marlin_int4_gemm.py
correct 0.13x 2.0μs
L5 P8 08_int8_w8a8_gemm.py
incorrect —
L5 P9 09_flash_attn2_fwd_hdim128_fp16_
correct 1.59x 0.3μs
L5 P10 10_flash_attn2_bwd_hdim128_fp16_
compile_fail —