Qualify and enable MTP 2 with paired workload evidence and fallback

This commit is contained in:
2026-09-18 22:37:54 +08:00
parent 06bcbcfd4e
commit 918b78a3b8
31 changed files with 3376 additions and 7 deletions
+163
View File
@@ -0,0 +1,163 @@
{
"baseline": {
"requests": 92,
"failures": [],
"timing": {
"code": {
"samples": 3,
"decode_tps_median": 77.09818266843658,
"ttft_s_median": 0.0694820829994569,
"elapsed_s_median": 26.619741354000325,
"reasoning_tokens": [
2048,
2048,
2048
]
},
"chinese": {
"samples": 3,
"decode_tps_median": 77.11985308389164,
"ttft_s_median": 0.0701422599995567,
"elapsed_s_median": 26.61247469099999,
"reasoning_tokens": [
2048,
2048,
2048
]
},
"reasoning": {
"samples": 3,
"decode_tps_median": 77.14088856875408,
"ttft_s_median": 0.06953137100026652,
"elapsed_s_median": 26.604923668000083,
"reasoning_tokens": [
2037,
2037,
2037
]
}
},
"soak": {
"requests": 54,
"sum_request_seconds": 623.7230764919932,
"failures": 0
},
"long_output": [
{
"name": "long-output-16384",
"passed": true,
"finish": "length",
"usage": {
"prompt_tokens": 75,
"total_tokens": 16459,
"completion_tokens": 16384,
"completion_tokens_details": {
"reasoning_tokens": 12136
}
},
"ttft_s": 0.07076310200136504,
"elapsed_s": 213.90097701500054,
"decode_tps": 76.61686204301199,
"max_stream_gap_s": 0.03923290800048562,
"events": 16379,
"sha256": "163936c470f43d5ab67e21f8372d3bc876a5be12a4d5536777254f3f19513bdc"
}
],
"quality": {
"tasks": 7,
"failures": [
"merge-intervals"
]
},
"quality_medium": {
"tasks": 7,
"failures": []
}
},
"mtp2": {
"requests": 98,
"failures": [],
"timing": {
"code": {
"samples": 3,
"decode_tps_median": 123.41758452810322,
"ttft_s_median": 0.07591862499975832,
"elapsed_s_median": 16.6593724129998,
"reasoning_tokens": [
2048,
2048,
2048
]
},
"chinese": {
"samples": 3,
"decode_tps_median": 128.7806464346781,
"ttft_s_median": 0.072885381996457,
"elapsed_s_median": 15.979564499000844,
"reasoning_tokens": [
182,
182,
182
]
},
"reasoning": {
"samples": 3,
"decode_tps_median": 122.14530767407761,
"ttft_s_median": 0.07294582499889657,
"elapsed_s_median": 16.830496447997575,
"reasoning_tokens": [
2048,
2048,
2048
]
}
},
"soak": {
"requests": 60,
"sum_request_seconds": 614.0107337019508,
"failures": 0
},
"long_output": [
{
"name": "long-output-16384",
"passed": true,
"finish": "length",
"usage": {
"prompt_tokens": 75,
"total_tokens": 16459,
"completion_tokens": 16384,
"completion_tokens_details": {
"reasoning_tokens": 16384
}
},
"ttft_s": 0.07230955999693833,
"elapsed_s": 115.18278583499341,
"decode_tps": 142.32414398895685,
"max_stream_gap_s": 0.01970133199938573,
"events": 6307,
"sha256": "580583cb4854d65556f20af2c05b29a82dad4c696bbc44d4d0476a50150458e5"
}
],
"quality": {
"tasks": 7,
"failures": []
},
"quality_medium": {
"tasks": 7,
"failures": []
}
},
"speedup": {
"code": 1.600784613288031,
"chinese": 1.6698767085900617,
"reasoning": 1.583405505696399
},
"resources": {
"samples": 1305,
"min_host_available_gib": 171.95434188842773,
"oom_kill_max": 0,
"safeguard_triggered": false,
"gpu_memory_peak_mib": 84797.0,
"gpu_temperature_peak_c": 64.0
}
}