Critical Merging Bug just started...

[David-AU-github]

Confirming exact same error ; mergekit can not find the "base_model" ; including if the path is local (absolute) on windows.

Funny thing is some mergekits work fine - no issue, where as others fail for the reasons below.
And merges I did in late SEPT 2024, now SOME fail ; others are fine ?!?!

Example: L3 models -> merge fine, no issue
Gemmas: Now break as noted below... but not all of them (??!?!)

This works fine:

models:

• model: G:/9B/gemma-2-9b-it-abliterated
  parameters:
  weight: .4
  merge_method: dare_ties
  base_model: G:/9B/gemma2-gutenberg-9B
  tokenizer_source: union
  dtype: bfloat16

BUT THIS DIES:

models:

• model: G:/9B/Gemma-2-Ataraxy-9B
  parameters:
  weight: [1,1,.75,.5,.25,.25,.05,.01]
• model: G:/9B/Gemma-2-9B-It-SPPO-Iter3
  parameters:
  weight: [1,1,.75,.5,.25,.25,.05,.01]
• model: G:/9B/gemma-2-Ifable-9B
  parameters:
  weight: [1,1,.75,.5,.25,.25,.05,.01]
  merge_method: dare_ties
  base_model: E:/Gemma-Dark-Writer3-mega-ab
  dtype: bfloat16

But exact SAME as above (3 models, base, dare_ties) , for Llama 3/3.1 merge - works fine (??)

Other GEMMA merges of the same type (3 models, base, dare_ties) that DID work (sept 2024) now crash and burn.

Even if I change this:
"base_model: E:/Gemma-Dark-Writer3-mega-ab"

Still dies, no matter what.
If I put in a bad location , it gives the normal not found too ; (??)

Likewise any "Gemma" merges like the one above that DID WORK fine, now crash and burn.
(specifically: dare_ties, 3 models + base model)

Please advise.

Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in run_code
File "C:\Program Files\Python312\Scripts\mergekit-yaml.exe_main.py", line 7, in
File "C:\Program Files\Python312\Lib\site-packages\click\core.py", line 1157, in call
return self.main(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Program Files\Python312\Lib\site-packages\click\core.py", line 1078, in main
rv = self.invoke(ctx)
^^^^^^^^^^^^^^^^
File "C:\Program Files\Python312\Lib\site-packages\click\core.py", line 1434, in invoke
return ctx.invoke(self.callback, **ctx.params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Program Files\Python312\Lib\site-packages\click\core.py", line 783, in invoke
return __callback(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\mergekit3\mergekit\mergekit\options.py", line 82, in wrapper
f(*args, **kwargs)
File "F:\mergekit3\mergekit\mergekit\scripts\run_yaml.py", line 47, in main
run_merge(
File "F:\mergekit3\mergekit\mergekit\merge.py", line 96, in run_merge
for _task, value in exec.run(quiet=options.quiet):
File "F:\mergekit3\mergekit\mergekit\graph.py", line 197, in run
res = task.execute(**arguments)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\mergekit3\mergekit\mergekit\merge_methods\generalized_task_arithmetic.py", line 126, in execute
tvs, base = get_task_vectors(
^^^^^^^^^^^^^^^^^
File "F:\mergekit3\mergekit\mergekit\merge_methods\generalized_task_arithmetic.py", line 201, in get_task_vectors
base = tensors[base_model]
~~~~~~~^^^^^^^^^^^^
KeyError: ModelReference(model=ModelPath(path='G:/9B/gemma2-gutenberg-9B', revision=None), lora=None, override_architecture=None)

Originally posted by @David-AU-github in #446

[David-AU-github]

Added issue:
seems even when using "mergekit" work around ; that merge kit is not creating "tokenizer.model" for Gemma models.
(previously it did).

RESULT: Can't quant models from source without this file in llamacpp ;
I used one from a previous merge (mergekit) .

I think something has changed upstream? Transformers?