mirror of
https://github.com/deepseek-ai/DeepSeek-VL2.git
synced 2026-11-05 01:07:23 +00:00
[GH-ISSUE #10] inference model can not see anything useful #9
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-VL2#9
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @luyao-cv on GitHub (Dec 17, 2024).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/10
my env:
cuda 11.8
torch 2.5.1
torchvision 0.19.0
inference code:
output:
@leeey111 commented on GitHub (Dec 18, 2024):
the same😅
@gnobitab commented on GitHub (Dec 18, 2024):
Hi, thanks for trying. Could you try the our inference script instead? There are several mismatches between your implementation and ours.
https://github.com/deepseek-ai/DeepSeek-VL2?tab=readme-ov-file#simple-inference-example
@leeey111 commented on GitHub (Dec 18, 2024):
i use "image_placeholder" instead of "image", it works.👍
@weiaicunzai commented on GitHub (Dec 19, 2024):
Could you please tell me your output ?
@leeey111 commented on GitHub (Dec 19, 2024):
我部署的是tiny版本,用来测试ocr能力,我试了多张图片,但模型响应的输出一直是下面段话
@Fantome12138 commented on GitHub (Dec 19, 2024):
same problem, but pip install transformers==4.38.2 will solve this
@leeey111 commented on GitHub (Dec 19, 2024):
pip install "transformers<4.42"
@Blaizzy commented on GitHub (Dec 20, 2024):
I tried everything here, but
deepseek-ai/deepseek-vl2-tinyonly outputs eos token.@Blaizzy commented on GitHub (Dec 20, 2024):
Seems like downgrading to v4.38.2 works for torch
But I can't get it wo work with the latest transformers. I need that support so it can run on Apple MLX.
@yuqianglxf commented on GitHub (Jan 3, 2025):
So, the model's OCR capability is zero. Can it be understood this way: for multiple different images, it produces fixed and incorrect outputs, is that right?
@leeey111 commented on GitHub (Jan 6, 2025):
No, it does have OCR capabilities. The error that occurred before was due to an incorrect version setting for the transformer, but it has been fixed after modification.
@heibaidaolx123 commented on GitHub (Feb 4, 2025):
same problem.
it seems the code in readme.md is broken.
I tried downgrading torch, changing to <image_placeholder> and adding
input_ids = prepare_inputs["input_ids"], none of them works.@spongepuddingg commented on GitHub (Feb 11, 2025):
Hey did you solve this issue? Facing the same. None of them seem to work.
@sergiev commented on GitHub (Feb 22, 2025):
@spongepuddingg here's a tiny example for single image example for current version of this repo:
And for the couple of them:
Dependencies are:
torch.version.cudais 11.7