beyoru commited on
Commit
543e6e4
·
verified ·
1 Parent(s): 07281ce

Add docker-compose.vllm.yaml: sample serving config

Browse files
Files changed (1) hide show
  1. docker-compose.vllm.yaml +40 -0
docker-compose.vllm.yaml ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # vLLM — cau hinh DA KIEM CHUNG cho beyoru/KAT-Coder-V2.5-Dev-VL (bf16).
2
+ #
3
+ # Da chay that: vllm/vllm-openai:v0.26.0, 1x H200 143GB, TP1, ~2 phut nap nguoi (14 shard,
4
+ # 65.4 GiB), FlashAttention 3 cho LM + FLASH_ATTN cho ViT, MoE backend TRITON tu chon.
5
+ #
6
+ # ⚠️ KHONG truyen --language-model-only. Co do danh cho ban text-only cua Kwaipilot;
7
+ # truyen vao day se tat vision tower va ban quay ve dung model goc.
8
+ services:
9
+ kat-vl:
10
+ image: vllm/vllm-openai:v0.26.0
11
+ container_name: kat-vl
12
+ ipc: host
13
+ ports:
14
+ - "8000:8000"
15
+ volumes:
16
+ # Doi sang thu muc chua checkpoint cua ban, hoac bo volume va dung thang
17
+ # --model beyoru/KAT-Coder-V2.5-Dev-VL de vLLM tu tai ve HF cache.
18
+ - /path/to/models:/models
19
+ - ${HOME}/.cache/huggingface:/root/.cache/huggingface
20
+ deploy:
21
+ resources:
22
+ reservations:
23
+ devices:
24
+ - driver: nvidia
25
+ device_ids: ["0"]
26
+ capabilities: [gpu]
27
+ command: >
28
+ --model /models/KAT-Coder-V2.5-Dev-VL
29
+ --served-model-name kat-vl
30
+ --max-model-len 32768
31
+ --gpu-memory-utilization 0.90
32
+ --trust-remote-code
33
+
34
+ # Bo nho: weight bf16 la 70.2 GB. Mot card 80 GB KHONG du cho ca weight lan KV cache
35
+ # dung duoc. Dung mot card 141/143 GB, hoac them --tensor-parallel-size 2 tren 2x80 GB.
36
+ #
37
+ # Context: native 262144. Chi nang --max-model-len sau khi da kiem bo nho con lai;
38
+ # KV cache o context day rat lon. Muon vuot native thi ap YaRN y het huong dan cua
39
+ # Qwen/Qwen3.6-35B-A3B (tham so RoPE o day khong doi so voi upstream) — luu y moi
40
+ # framework deu cai YaRN TINH nen bat len se lam giam chat luong prompt ngan.