20 lines
616 B
JSON
20 lines
616 B
JSON
{
|
|
"observation": "Start from a safe tp=4 layout and conservative batching.",
|
|
"diagnosis": "The first pass should verify multi-GPU launch and avoid queueing collapse from over-batching.",
|
|
"config_patch": {
|
|
"env_patch": {},
|
|
"flag_patch": {
|
|
"tensor-parallel-size": 4,
|
|
"max-num-seqs": 8,
|
|
"max-num-batched-tokens": 32768,
|
|
"gpu-memory-utilization": 0.85,
|
|
"block-size": 64
|
|
}
|
|
},
|
|
"expected_effects": [
|
|
"Stable startup on 4x H20",
|
|
"Low risk of OOM during the first binary-search probes"
|
|
],
|
|
"why_not_previous_failures": "This is the initial baseline proposal."
|
|
}
|