sglang/ GB200 perf

dsr1-fp4-1k1k-mid-curve

concurrency 8,192 · 13 days ago

cron
passed
re-run
Commit

[RL] Handle Mooncake buffers across memory release (#27696)

zhuzilin·13 days ago
PR #27696 · [RL] Handle Mooncake buffers across memory release
Run
GitHub Actions27321587026-1
Slurm job5293
GPUs48·prefill 16 / decode 32
ISL / OSL1024 / 1024

Metrics

28 captured

best_of

1.00

burstiness

1.00

completed

81,920

duration

713

max_concurrency

8,192

mean_e2el_ms

68,956ms

mean_itl_ms

1,170ms

mean_tpot_ms

23.88ms

mean_ttft_ms

46,969ms

median_e2el_ms

71,817ms

median_itl_ms

1,001ms

median_tpot_ms

21.45ms

median_ttft_ms

52,973ms

num_prompts

81,920

output_throughput

105,821tok/s

p99_e2el_ms

84,173ms

p99_itl_ms

3,491ms

p99_tpot_ms

42.36ms

p99_ttft_ms

59,514ms

peak_output_tokens_per_s

150,638s

request_throughput

115tok/s

std_e2el_ms

8,794ms

std_itl_ms

662ms

std_tpot_ms

7.15ms

std_ttft_ms

11,713ms

total_input_tokens

75,511,905

total_output_tokens

75,496,082

total_token_throughput

211,664tok/s