2b7249fbb3
Signed-off-by: tc-mb <tianchi_cai@icloud.com> Co-authored-by: GKangaroo <1095103651@qq.com> Co-authored-by: GKangaroo <gqx24@mails.tsinghua.edu.cn> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
61 lines
1.9 KiB
YAML
61 lines
1.9 KiB
YAML
# MiniCPM-o 4.5 deploy: 8x4090 layout (24GB cards).
|
|
#
|
|
# Stage 0 (thinker): GPU 0,1,2,3 — 4-way TP, ~85% mem (≈20.4 GB / card,
|
|
# thinker weights take ~17.5 GB per card
|
|
# on 4090, leaving ~3 GB for KV cache /
|
|
# activations).
|
|
# Stage 1 (talker) : GPU 4 — MiniCPMTTS + Token2Wav, ~90% mem
|
|
# (covers both the talker LM and the
|
|
# in-process Token2Wav vocoder that emits
|
|
# the waveform).
|
|
#
|
|
# max_model_len is pinned to 4096 because the upstream config.json
|
|
# advertises max_position_embeddings=40960 (tuned for H20 141GB); on a
|
|
# 24 GB 4090 a 40k-token KV cache request OOMs the worker process
|
|
# before serving the first request. 8192 also OOMs in practice with
|
|
# this thinker model + vision/audio encoders + Whisper preallocation,
|
|
# so 4096 is the empirically-verified safe ceiling.
|
|
pipeline: minicpmo_4_5
|
|
async_chunk: false
|
|
|
|
trust_remote_code: true
|
|
enable_prefix_caching: false
|
|
|
|
stages:
|
|
- stage_id: 0
|
|
max_num_seqs: 1
|
|
gpu_memory_utilization: 0.85
|
|
enforce_eager: true
|
|
tensor_parallel_size: 4
|
|
max_model_len: 4096
|
|
max_num_batched_tokens: 4096
|
|
devices: "0,1,2,3"
|
|
limit_mm_per_prompt:
|
|
video: 1
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
top_p: 1.0
|
|
top_k: -1
|
|
max_tokens: 2048
|
|
seed: 42
|
|
detokenize: true
|
|
repetition_penalty: 1.1
|
|
|
|
- stage_id: 1
|
|
# Talker only consumes runtime_additional_information[0]; any value > 1
|
|
# makes concurrent requests share request-0's audio (mm_outputs fallback).
|
|
max_num_seqs: 1
|
|
gpu_memory_utilization: 0.9
|
|
enforce_eager: true
|
|
max_num_batched_tokens: 8192
|
|
devices: "4"
|
|
default_sampling_params:
|
|
temperature: 0.0
|
|
top_p: 1.0
|
|
top_k: -1
|
|
max_tokens: 1
|
|
seed: 42
|
|
detokenize: false
|
|
|
|
platforms: {}
|