Why is convert.py missing?

[David-AU-github]

What happened?

Critical "non llama3" convert.py and change NOT in download of files.

ALSO:
It is unclear if "convert-hf-to-gguf.py" supports what "convert.py" did .

Does it convert llama, llama2, mistral or is "convert-legacy-llama.py" required?
Safetensor files are EVERYWHERE. (?) [ RE: https://github.com//pull/7430 ]

This critical action DID NOT OCCUR:
"Move convert.py to examples/convert-legacy-llama.py"

"examples/convert-legacy-llama.py" does not exist.
(when downloading the zip files).

On another note why remove "convert.py" at all?

-This breaks "bat files" and automation generation.
-This will break all colabs too.
-This will break any HF spaces that create GGUF files as well.
-This will create needless confusion.

If "convert-hf-to-gguf.py" (llama3) does everything convert.py did , just keep it as "convert.py" ?

Name and Version

this is not applicable - core files missing.

What operating system are you seeing the problem on?

No response

Relevant log output

this is not applicable - core files missing.

[Chen4549]

same issue
but it worked yesterday
so they removed the convert.py a couple of hours ago?

Problem solved by using CLI command: python llama.cpp/convert-hf-to-gguf.py FOLDER --outfile OUTPUT --outtype f16

[Galunid]

TLDR:

convert-hf-to-gguf.py is the correct way to convert huggingface models. If you have llama/llama2 model downloaded directly from meta (in .pth format), you can still use examples/convert-legacy-llama.py, otherwise it's discourage and we won't provide support. It will not work with llama3

If you need to fix scripts/colabs, changing convert.py to convert-hf-to-gguf.py and removing --vocab-type should be enough to make them work.

---

Full explanation:

In #7430 convert.py was moved to examples/convert-legacy-llama.py to avoid confusion as to which script should be used (now there's only one) and reduce maintenance burden on developers. It should not be used anymore. It was removed to force people to use convert-hf-to-gguf.py that supports all huggingface models that llama.cpp can run. There were a lot of issues because of people using convert.py when they shouldn't have. There were a lot of tickets, where convert.py was used by users in situations where it wasn't appropriate - see #7658 (comment)

I know this will inconvenience some people and it's really unfortunate, but from developer perspective maintaining convert.py was just no longer an option. convert-hf-to-gguf.py was doing everything convert.py did and much more, and it is much simpler to maintain. For a while convert.py produced broken models that had warnings like this:

llm_load_vocab: missing pre-tokenizer type, using: 'default'
llm_load_vocab:                                             
llm_load_vocab: ************************************        
llm_load_vocab: GENERATION QUALITY WILL BE DEGRADED!        
llm_load_vocab: CONSIDER REGENERATING THE MODEL             
llm_load_vocab: ************************************   

Fixes were implemented in convert-hf-to-gguf.py that never made it to convert.py. Making changes to both scripts that did pretty much the same thing takes a lot of time, so one of them was removed.

If you really need it, it's still available in examples/convert-legacy-llama.py, but it's no longer supported.

On another note why remove "convert.py" at all?

-This breaks "bat files" and automation generation.
-This will break all colabs too.
-This will break any HF spaces that create GGUF files as well.
-This will create needless confusion.

"Why" is explained above, scripts/colabs fix is to change convert.py to convert-hf-to-gguf.py and removing unrecognized flags. Our huggingface space should be working as expected. If someone is confused, then most likely they were using the wrong script and their model quality will improve by switching :).

[David-AU-github]

Okay, thank you for the details - but this DOESN'T EXIST:

examples/convert-legacy-llama.py

It is not in the zip files.
IE: https://github.com/ggerganov/llama.cpp/archive/refs/tags/b3051.zip

In regards to colabs - I am not talking about mine, I am talking about all that rely on llamacpp.
This is due to the construction of commands and auto-update from llamacpp PER USE.
That means as of this update, all colabs using llamacpp (and all users using colabs) will break.
Any colab that uses llamacpp is now broken at the "convert" step.
Most users (of colabs) will have no clue what the issue is, never mind how to fix.

At the very least please consider a "redirect" (from convert.py) to the correct file? or a fall back [legacy]?
Or redirect to "new version" ??
Maybe one that looks at if the file is or is not "llama3" and acts accordingly?

[David-AU-github]

same issue but it worked yesterday so they removed the convert.py a couple of hours ago?

Problem solved by using CLI command: python llama.cpp/convert-hf-to-gguf.py FOLDER --outfile OUTPUT --outtype f16

thanks - that really helps!!!

[Galunid]

It is not in the zip files.

I'm not sure what you mean. I downloaded the file you linked and it is there.

[mlabonne]

@Galunid Thanks for this clear explanation, it makes complete sense

[David-AU-github]

It is not in the zip files.

I'm not sure what you mean. I downloaded the file you linked and it is there.

I will download again - I checked it - not there.
Delay in being added?

[David-AU-github]

Okay - there it is . ?(??) ; thanks again.