Add deeper infeasible probe diagnostics

This commit is contained in:
2026-04-05 01:44:38 +08:00
parent 0aa607a4f1
commit 84c5d6bd80
5 changed files with 249 additions and 3 deletions

View File

@@ -0,0 +1,22 @@
{
"observation": "Long-context chat traffic is dominated by large prefills, so TTFT under the stepped SLO is the binding constraint. A launch-safe baseline should spread compute across all 4 GPUs while limiting concurrent long-prefill contention.",
"diagnosis": "The FLASHINFER attempt failed at startup, but the safer FLASH_ATTN family launched successfully. A conservative seq cap plus a moderate batched-token cap is a better baseline for diagnosing whether the SLO itself is too aggressive under this trace.",
"config_patch": {
"env_patch": {
"VLLM_ATTENTION_BACKEND": "FLASH_ATTN"
},
"flag_patch": {
"tensor-parallel-size": 4,
"max-num-seqs": 16,
"max-num-batched-tokens": 24576,
"gpu-memory-utilization": 0.94,
"block-size": 32
}
},
"expected_effects": [
"Stable 4-GPU launch without FLASHINFER warmup failure",
"Lower head-of-line blocking than larger sequence caps",
"More interpretable lower-bound throughput/SLO measurement"
],
"why_not_previous_failures": "This proposal keeps the launch-safe FLASH_ATTN backend and the conservative batching limits that already avoided the earlier FLASHINFER startup failure."
}