cant make Vision working
I went as low as this, tiny pic 130kb, didnt work, text works well thou
OS: Windows 11
Runtime: ExLlamaV3 1.5.0
Server: TabbyAPI (podaj git commit / wersję)
Model: turboderp/DeepSeek-V4-Flash-Vision-Exp-exl3, revision 3.04bpw
GPU: RTX5090 latest
CPU: 9800x3d + 128gb DDR5
Config:
backend: exllamav3
vision: true
max_seq_len: 8192
cache_size: 8192
cache_mode: "8,6"
chunk_size: 512
max_batch_size: 1
cpu_moe_offload_layers: 38
reasoning: false
draft_mode: mtp
draft_num_tokens: 3
Request:
one 130 KB JPEG as image_url data URL
one short user text prompt
max_tokens: 64
stream: false
Expected:
Successful image description.
Actual:
CUDA IndexKernelUtils.cu:
Assertion 'ind >= 0 && ind < ind_dim_size'
followed by CUDA device-side assert / CUBLAS_STATUS_INTERNAL_ERROR.)
Same Error here