Datasets:

EXD-AI
/

benchmark-results

Modalities:

Formats:

Size:

Libraries:

License:

Dataset Viewer

Split (1)

train · 35 rows

text stringlengths 28 51
============ Serving Benchmark Result ============
Successful requests: 20
Failed requests: 0
Request rate configured (RPS): 1.00
Benchmark duration (s): 29.02
Total input tokens: 10240
Total generated tokens: 2560
Request throughput (req/s): 0.69
Output token throughput (tok/s): 88.20
Peak output token throughput (tok/s): 68.00
Peak concurrent requests: 12.00
Total token throughput (tok/s): 441.02
---------------Time to First Token----------------
Mean TTFT (ms): 562.28
Median TTFT (ms): 510.73
P99 TTFT (ms): 910.52
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms): 77.24
Median TPOT (ms): 79.03
P99 TPOT (ms): 113.27
---------------Inter-token Latency----------------
Mean ITL (ms): 163.09
Median ITL (ms): 157.17
P99 ITL (ms): 384.83
---------------Speculative Decoding---------------
Acceptance rate (%): 56.36
Acceptance length: 2.13
Drafts: 1203
Draft tokens: 2406
Accepted tokens: 1356
Per-position acceptance (%):
Position 0: 64.84
Position 1: 47.88
==================================================
ramshreyas@aitopatom-0a62:~/EXD$

EXD Benchmark Results

Performance data from inference sweeps on the Gigabyte AI TOP (DGX Spark-class, NVIDIA GB10).

File	Source	Description
`episodes/Ep03/03_bench.txt`	Episode 3	First benchmark run — Qwen2.5-7B, TTFT, TPOT, throughput
`benchmarks/qwen3.6-35b-a3b-llama-benchy.md`	Serve harness	Qwen3.6-35B-A3B with llama-benchy