Add decode-only study mode support

This commit is contained in:
2026-04-09 11:23:17 +08:00
parent 96140b79bb
commit c158807fac
6 changed files with 282 additions and 1 deletions

View File

@@ -0,0 +1,13 @@
{
"observation": "This is the decode-only baseline directly aligned with run_qwen235b_decode.sh, including DecodeBenchConnector and the internal qwen3-235b launch shape.",
"diagnosis": "A baseline measurement is required before proposing decode-only throughput changes. Preserve all current envs and flags to establish the first feasible sampling_u/request_rate point under the TPOT-only SLO.",
"config_patch": {
"env_patch": {},
"flag_patch": {}
},
"expected_effects": [
"Establish a launch-safe baseline for decode-only TPOT tuning",
"Seed later trials from the first feasible sampling_u if one exists"
],
"why_not_previous_failures": "No previous failures in this study."
}