Platform
Linux/Ubuntu
Lemonade Version
10.0.1
GPU / APU Model
strix halo
Component
Desktop App / UI
Bug Description
-
in the event of a network interruption (frequently happening here) occurring during a "lemonade(server) pull" for some model, eventual subsequent retries to download the same not yet fully retrieved model, lemonade should not erase the already partially retrived model but it should instead just continue to download it from where it was left, appending to the remaining part of the same model
-
when a network interruption as above described occurs leaving a partially downloaded model in a uncompleted state, lemonade should not show the model as if it was fully retrieved and ready to be run/served
-
should the user deliberately interrupt some model downloading process using ctrl-c, model file and other files relics are left dangling into the relative directory such that when invoking lemonade (even for just printing the list of available/downloaded models names), it crashes in that precise listing position requiring the user to remove the uncompleted model directory in order to be able to run lemonade stuff again
Steps to Reproduce
- lemonade-server serve
- lemonade(-server) pull *
*. btw, why NPU models listed/pulled/run/served from lemonade(-server) and flm aren't named exactly the same?
- [optional] disconnect from the net or interrupt w/ ctrl-c
Expected vs Actual Behavior
Lemonade should check if a model has been really fully retrieved (maybe checksumming it or by creating a lock file at pull start and removing it after downloading 100%); in case it was not completed it should ask the user to continue retriving it, to clear the model's directory or to skip. In any case and especially in this last case, lemonade should also skip a model directory if it contains partially retrieved files affecting lemonade itself ability to run
Log Output
Additional Context
No response
Platform
Linux/Ubuntu
Lemonade Version
10.0.1
GPU / APU Model
strix halo
Component
Desktop App / UI
Bug Description
in the event of a network interruption (frequently happening here) occurring during a "lemonade(server) pull" for some model, eventual subsequent retries to download the same not yet fully retrieved model, lemonade should not erase the already partially retrived model but it should instead just continue to download it from where it was left, appending to the remaining part of the same model
when a network interruption as above described occurs leaving a partially downloaded model in a uncompleted state, lemonade should not show the model as if it was fully retrieved and ready to be run/served
should the user deliberately interrupt some model downloading process using ctrl-c, model file and other files relics are left dangling into the relative directory such that when invoking lemonade (even for just printing the list of available/downloaded models names), it crashes in that precise listing position requiring the user to remove the uncompleted model directory in order to be able to run lemonade stuff again
Steps to Reproduce
*. btw, why NPU models listed/pulled/run/served from lemonade(-server) and flm aren't named exactly the same?
Expected vs Actual Behavior
Lemonade should check if a model has been really fully retrieved (maybe checksumming it or by creating a lock file at pull start and removing it after downloading 100%); in case it was not completed it should ask the user to continue retriving it, to clear the model's directory or to skip. In any case and especially in this last case, lemonade should also skip a model directory if it contains partially retrieved files affecting lemonade itself ability to run
Log Output
Additional Context
No response