github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #9618] Disk space usage #68331

New Issue
Closed
opened 2026-05-04 13:14:50 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-05-04 13:14:50 -05:00
Owner
Copy Link

Originally created by @gty1829 on GitHub (Mar 10, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9618

Why is it necessary to copy the weights and re-encrypt them with SHA-256 when converting models? This process causes excessive memory usage when I import models locally.

Originally created by @gty1829 on GitHub (Mar 10, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9618 Why is it necessary to copy the weights and re-encrypt them with SHA-256 when converting models? This process causes excessive memory usage when I import models locally.
GiteaMirror added the question label 2026-05-04 13:14:50 -05:00
GiteaMirror commented 2026-05-04 13:14:52 -05:00
Author
Owner
Copy Link

@krenax commented on GitHub (Mar 10, 2025):

Can you provide an example?

<!-- gh-comment-id:2711018451 --> @krenax commented on GitHub (Mar 10, 2025): Can you provide an example?
GiteaMirror commented 2026-05-04 13:14:52 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Mar 11, 2025):

@gty1829 they're not "re-encrypted", the sha-256s are used to calculate the checksum of the model. Are you pulling with ollama pull or importing the weights with ollama create?

EDIT: I just saw that you mentioned converting, so I'm guessing you meant with ollama create. Are you converting a GGUF file or are you converting from Safetensors?

<!-- gh-comment-id:2712127089 --> @pdevine commented on GitHub (Mar 11, 2025): @gty1829 they're not "re-encrypted", the sha-256s are used to calculate the checksum of the model. Are you pulling with `ollama pull` or importing the weights with `ollama create`? EDIT: I just saw that you mentioned `converting`, so I'm guessing you meant with `ollama create`. Are you converting a GGUF file or are you converting from Safetensors?
GiteaMirror commented 2026-05-04 13:14:53 -05:00
Author
Owner
Copy Link

@gty1829 commented on GitHub (Mar 11, 2025):

I 'm converting from Safetensors. I mean, I downloaded the safetensors of qwen2.5-7B-Instruct-1M, and I ran the command ollama create qwen2.5 -f ./Modelfile. Then I found the files encrypted with SHA-256 under %OLLAMA_MODELS%\blobs, and their size is the same as the storage space occupied by the safetensors. So I guess these are actually the model weights. Why not read the safetensors directly? Thank you.

Image

Image

<!-- gh-comment-id:2712856952 --> @gty1829 commented on GitHub (Mar 11, 2025): I 'm converting from Safetensors. I mean, I downloaded the safetensors of qwen2.5-7B-Instruct-1M, and I ran the command ollama create qwen2.5 -f ./Modelfile. Then I found the files encrypted with SHA-256 under %OLLAMA_MODELS%\blobs, and their size is the same as the storage space occupied by the safetensors. So I guess these are actually the model weights. Why not read the safetensors directly? Thank you. ![Image](https://github.com/user-attachments/assets/038ad874-60f9-427a-8f84-f7cb5cff224e) ![Image](https://github.com/user-attachments/assets/27d6488b-9b02-4814-94d7-5371313f17ab)
GiteaMirror commented 2026-05-04 13:14:54 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Mar 11, 2025):

Yes, the weights are cut up into blobs (this will result in more blobs in the future as it will be cut up by tensor instead of one large file). If you're not using the --quantize option with ollama create the weights sizes will be the same size, but the data type will have changed from bfloat16 to fp16 for many of the tensors, so the weights aren't identical.

The reason for cutting the weights into blobs is so that we can deduplicate any weights which are shared between different models. If you change the Modelfile and run ollama create again it will only store one copy of the weights but you'll see two models listed in ollama ls. Hopefully that makes sense. I'll go ahead and close the issue as answered, but feel free to keep commenting.

<!-- gh-comment-id:2712977642 --> @pdevine commented on GitHub (Mar 11, 2025): Yes, the weights are cut up into blobs (this will result in more blobs in the future as it will be cut up by tensor instead of one large file). If you're not using the `--quantize` option with `ollama create` the weights sizes will be the same size, but the data type will have changed from bfloat16 to fp16 for many of the tensors, so the weights aren't identical. The reason for cutting the weights into blobs is so that we can deduplicate any weights which are shared between different models. If you change the Modelfile and run `ollama create` again it will only store one copy of the weights but you'll see two models listed in `ollama ls`. Hopefully that makes sense. I'll go ahead and close the issue as answered, but feel free to keep commenting.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label question
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#68331
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?