Qwen3-VL-4B-rbln-tp4

Qwen/Qwen3-VL-4B-Instruct 를 Rebellions RBLN-CA25 NPU용으로 컴파일한 모델.

• RBLN SDK: rebel-compiler 0.11.0 / optimum-rbln 0.11.0.post1
• 타깃: RBLN-CA25 (15.7 GiB/device)
• 컴파일 config: num_devices=4, max_seq_len=32768, kvcache_num_blocks=14, batch_size=8, attn_impl=flash_attn
• ⚠️ 런타임 device 개수 = 컴파일 num_devices — 이 빌드는 4장 전용 ([0,1,2,3]).

로드

from optimum.rbln import RBLNAutoModelForImageTextToText
model = RBLNAutoModelForImageTextToText.from_pretrained("Qwen3-VL-4B-rbln-tp4", export=False,
    rbln_config={"visual": {"device": [0,1,2,3]}, "device": [0,1,2,3]})

VLM(image-text-to-text). vLLM 서빙: vllm serve <dir> --max-model-len 32768 --max-num-seqs 8

토크나이저/프로세서가 없으면 base 모델(Qwen/Qwen3-VL-4B-Instruct)에서 로드하세요.