sglang/ GB200 perf

dsr1-fp4-1k1k-mid-curve

concurrency 2,048 · 10 days ago

cron
passed
re-run
Commit

feat: add NVTX markers for the scheduler main loop (#27901)

JoyFuture·10 days ago
PR #27901 · feat: add NVTX markers for the scheduler main loop
Run
GitHub Actions27486985911-1
Slurm job5306
GPUs48·prefill 16 / decode 32
ISL / OSL1024 / 1024

Metrics

28 captured

best_of

1.00

burstiness

1.00

completed

20,480

duration

210

max_concurrency

2,048

mean_e2el_ms

20,273ms

mean_itl_ms

987ms

mean_tpot_ms

20.13ms

mean_ttft_ms

1,724ms

median_e2el_ms

19,893ms

median_itl_ms

1,020ms

median_tpot_ms

20.41ms

median_ttft_ms

828ms

num_prompts

20,480

output_throughput

89,931tok/s

p99_e2el_ms

31,201ms

p99_itl_ms

1,554ms

p99_tpot_ms

21.55ms

p99_ttft_ms

12,241ms

peak_output_tokens_per_s

106,046s

request_throughput

97.51tok/s

std_e2el_ms

2,717ms

std_itl_ms

201ms

std_tpot_ms

1.30ms

std_ttft_ms

2,627ms

total_input_tokens

18,880,692

total_output_tokens

18,888,974

total_token_throughput

179,823tok/s