Download timeout

[gdkrmr]

I am at a very slow internet connection and downloading almost any model gives me a timeout:

[FLM]  Downloading 2/7: model.q4nx
[FLM]  Downloading: 86.0% (3830.0MB / 4455.4MB)CURL error: Timeout was reached

Currently there is a hard-coded timeout of one hour set here

FastFlowLM/src/pull/download_model.cpp

Line 160 in bb309ad

curl_easy_setopt(curl, CURLOPT_TIMEOUT, 3600L); // 1 hour timeout

I suggest removing this limitation, I think this can be done to setting the limit to 0, alternatively you could make it configurable.

Edit: Just as a reference, a 10Mbit connection can download just over 4 GB/h and a 4B sized model is just around that limit.

[twflm]

I am at a very slow internet connection and downloading almost any model gives me a timeout:

[FLM]  Downloading 2/7: model.q4nx
[FLM]  Downloading: 86.0% (3830.0MB / 4455.4MB)CURL error: Timeout was reached

Currently there is a hard-coded timeout of one hour set here

FastFlowLM/src/pull/download_model.cpp

Line 160 in bb309ad

curl_easy_setopt(curl, CURLOPT_TIMEOUT, 3600L); // 1 hour timeout
I suggest removing this limitation, I think this can be done to setting the limit to 0, alternatively you could make it configurable.

Edit: Just as a reference, a 10Mbit connection can download just over 4 GB/h and a 4B sized model is just around that limit.

how about 2 hours?

[gdkrmr]

Given that the largest model you currently support is 13GB this wouldn't work for me but I also know that I am probably a special case. What about a config option, e.t. --pull-timeout or flm pull --timeout <t> <model> where <t> could be set to zero to disable the timeout. This also mirrors the libcurl option used internally.

[ZaneNi]

I think a better approach in this case is to download the files manually from our Hugging Face and place them in the appropriate location. What do you think?

Given that the largest model you currently support is 13GB this wouldn't work for me but I also know that I am probably a special case. What about a config option, e.t. --pull-timeout or flm pull --timeout <t> <model> where <t> could be set to zero to disable the timeout. This also mirrors the libcurl option used internally.

[gdkrmr]

Personally, I have no problem working around the issue and I have no idea how many people are affected by an issue like that. I also understand that there needs to be a mechanism that prevents flm pull to run forever. I much prefer the devs to spend their limited time implementing more models instead of improving the download manager functionality of flm. This is a design decision that the flm developers have to make and I am very happy that my NPU is finally useful!

[stdevPavelmc]

I have a slow (4Mbps) and someway unstable connection and was hitting 99% downloads being discarded and started over, several times. So this same problem.

So I crafted a solution to:

• Check files and sizes to retry/continue interrupted downloads
• Verify the downloads at the end

It's working good right now, I will make a PR when I clean/review/document the code, sample interrupted dowload below for flm pull whisper-v3:turbo

[FLM]  Model: whisper-v3:turbo
[FLM]  Name: Whisper-V3-Turbo-NPU2
[FLM]  All required files are present.
[FLM]  Present files (4):
  - config.json
  - tokenizer.json
  - tokenizer_config.json
  - model.q4nx
[FLM]  Checking all files for missing or partial downloads...
[FLM]  Partial download detected: model.q4nx (93 MB / 620 MB)
[FLM]  Downloading 1 file(s)...
[FLM]  Files to download (526.81 MB): 
  - model.q4nx (526.81 MB)
[FLM]  Downloading 1/1: model.q4nx
[FLM]  Resuming from 93.2422 MB
[FLM]  Downloading: 100.0% (619.9MB / 620.1MB)
[FLM]  Checking Hash...
[FLM]  Download completed: /home/pavel/.config/flm/models/Whisper-V3-Turbo-NPU2/model.q4nx
[FLM]  Overall progress:  1/1 files
[FLM]  All downloads completed successfully!
[FLM]  Model downloaded successfully!
[FLM]  All files verified successfully.