aiconfigurator/src/aiconfigurator/cli/exps/hetero_disagg.yaml at main · simone-chen/aiconfigurator · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# define your experiments here, the order will determine the execution order.
exps:
  - exp_h200_h200
  - exp_b200_h200

exp_h200_h200:
  mode: "patch"
  serving_mode: "disagg" # required
  model_name: "QWEN3_32B" # required
  total_gpus: 16 # required
  system_name: "h200_sxm" # required, for prefill
  decode_system_name: "h200_sxm" # optional, if not provided, it will use the same system name as the prefill system.
  backend_name: "trtllm"
  profiles: []
  isl: 4000 # input sequence length
  osl: 500 # output sequence length
  ttft: 300.0  # Target TTFT in ms
  tpot: 50.0   # Target TPOT in ms

exp_b200_h200:
  mode: "patch"
  serving_mode: "disagg" # required
  model_name: "QWEN3_32B" # required
  total_gpus: 16 # required
  system_name: "b200_sxm" # required, for prefill
  decode_system_name: "h200_sxm" # optional, if not provided, it will use the same system name as the prefill system.
  backend_name: "trtllm"
  profiles: []
  isl: 4000 # input sequence length
  osl: 500 # output sequence length
  ttft: 300.0  # Target TTFT in ms
  tpot: 50.0   # Target TPOT in ms