`pulling manifest Error: EOF` when pulling after disk is full

[jmorganca]

To reproduce, pull with little disk space lefT:

$ ollama run deepseek-coder:33b
pulling manifest 
Error: write /usr/share/ollama/.ollama/models/blobs/sha256:065b9a7416ba28634cd4efc2cd3024d4755731c1275dc0286b81b01793185fbb-partial-0: no space left on device

Even with more space, future ollama pull commands fail until Ollama restarted:

$ ollama run deepseek-coder:33b
pulling manifest 
Error: EOF

[mxyng]

I haven't been able to reproduce this. Ollama behaves as expected. Here is the test I ran:

1. Create a VM with multiple, small physical volumes
2. Create a logical volume by attaching one of the physical volumes
3. Format, mount, and configure as OLLAMA_MODELS
4. Pull a large model, e.g. llama2:70b
5. Step 4 should fail once the logical volume is full
6. Expand the disk by attaching another physical volumes to the logical volume
7. Repeat step 4 which should resume the download where it left off

@jmorganca do you have any more details on how to reproduce this?

[Pytness]

I just managed to unintentionally reproduce this bug. The steps I followed:

• Pull multiple models until disk is full.
• Get disk is full error pulling another model.
• ollama rm a few models.
• Try to ollama pull a model and get EOF error.

Restarting ollama fixes the issue.

[ictus4u]

> ollama --version
ollama version is 0.1.34

Restarting didn't work for me.

After having a look at https://github.com/ollama/ollama/pull/344, I ended with the below command:

sudo rm /usr/share/ollama/.ollama/models/blobs/*-partial*

[candera]

Deleting only the ones of zero size in ~/.ollama/models/blobs/ worked for me, and I didn't have to restart the entire download.