sglang/ GB200 perf

dsr1-fp8-1k1k-ultra-tpt

concurrency 4,096 · 21 hours ago

cron
passed
re-run
Commit

Vectorize _create_custom_4d_mask in CustomQwen2Decoder (#27527)

ckvermaAI·22 hours ago
PR #27527 · Vectorize _create_custom_4d_mask in CustomQwen2Decoder
Run
GitHub Actions27999739491-1
Slurm job5334
GPUs16·prefill 8 / decode 8
ISL / OSL1024 / 1024

Metrics

28 captured

best_of

1.00

burstiness

1.00

completed

40,960

duration

608

max_concurrency

4,096

mean_e2el_ms

58,700ms

mean_itl_ms

2,918ms

mean_tpot_ms

59.78ms

mean_ttft_ms

3,680ms

median_e2el_ms

57,285ms

median_itl_ms

3,029ms

median_tpot_ms

60.86ms

median_ttft_ms

1,043ms

num_prompts

40,960

output_throughput

62,097tok/s

p99_e2el_ms

96,109ms

p99_itl_ms

3,711ms

p99_tpot_ms

61.77ms

p99_ttft_ms

39,886ms

peak_output_tokens_per_s

71,698s

request_throughput

67.39tok/s

std_e2el_ms

8,763ms

std_itl_ms

486ms

std_tpot_ms

3.38ms

std_ttft_ms

8,147ms

total_input_tokens

37,769,666

total_output_tokens

37,742,239

total_token_throughput

124,239tok/s