=== post-fix (batched worker observe), rank0 ===
--- onfix/rank0.1789309272412889977.pt.trace.json.gz steps=193
AR stream=279    n=  6151 ( 31.87/step)      3790us/step  p50=    118 p90=    124 max=    521
AR stream=278    n=  6152 ( 31.88/step)      3683us/step  p50=    115 p90=    120 max=    195
AG stream=7      n=   193 (  1.00/step)      2311us/step  p50=   2311 p90=   2476 max=   2626
AR stream=7      n=   194 (  1.01/step)      2186us/step  p50=   1703 p90=   3521 max=   6777
AR stream=2370   n=   485 (  2.51/step)       451us/step  p50=    125 p90=    380 max=   1886
AR stream=2371   n=   485 (  2.51/step)       299us/step  p50=    118 p90=    126 max=    191
non-nccl kernel time: 16508us/step    nccl: 12721us/step

=== post-fix, rank1 ===
--- onfix/rank1.1789309290880714268.pt.trace.json.gz steps=193
AR stream=2370   n=   490 (  2.54/step)      7335us/step  p50=    111 p90=    120 max=1360217
AR stream=279    n=  6272 ( 32.50/step)      3651us/step  p50=    112 p90=    116 max=    190
AR stream=278    n=  6272 ( 32.50/step)      3617us/step  p50=    111 p90=    115 max=    194
AG stream=7      n=   196 (  1.02/step)      3323us/step  p50=   3060 p90=   3093 max=  99086
AR stream=7      n=   196 (  1.02/step)       511us/step  p50=    118 p90=    126 max=  75745
AR stream=2371   n=   490 (  2.54/step)       286us/step  p50=    111 p90=    118 max=    182
non-nccl kernel time: 15057us/step    nccl: 18723us/step
