[Easy?] Numpy Version Pin Bump: == 2.0

[Jack-Khuu]

🐛 Describe the bug

ExecuTorch recently bumped their numpy requirements to numpy == 2.0 in pytorch/executorch@a7b5297
This puts torchchat in a finicky spot since the current requirements are tied to under 2.0 due to GGUF support requiring < 2.0 (see blame for previous attempts)

torchchat/install/requirements.txt

Lines 19 to 20 in fb65b8b

	# numpy version range required by GGUF util
	numpy >= 1.17, < 2.0

While not actively an issue, as soon as an ExecuTorch pin bump is required, this will become a hard blocker.

Task: Make a requirements version pinbump numpy >= 2.0 numpy > 1.17

• Considerations:
  • 1/23: This may "just work" without additional effort; Seems like llama.cpp may have fixed this in December ([gguf-py] gguf_reader: numpy 2 newbyteorder fix ggml-org/llama.cpp#9772)
  • Is updating the GGUF support within torchchat's control (i.e. propogating dependencies)?
    If not who/what needs coercing?
  • Should be icebox GGUF support?

Versions

N/A

[Jack-Khuu]

fyi: @Gasoonjia, @vmpuri, @mergennachin (ET)

[mergennachin]

gguf-util says >=1.17 but not explicitly <2, so in theory numpy>=2.0 should work

https://github.com/ggerganov/llama.cpp/blame/master/gguf-py/pyproject.toml

where does it fail when the version is 2 or above?

looks like this is the job -- perhaps you could repro it

torchchat/.github/workflows/pull.yml

Line 700 in fb65b8b

test-gguf-util:

[Jack-Khuu]

where does it fail when the version is 2 or above?

It was implicit dependency in llama.cpp; the project was using interfaces that were deprecated in numpy 2+
That said looks like it got fixed in December (ggml-org/llama.cpp#9772)

If we're able to just drop it and things work, that's the best case scenario

[mergennachin]

yeah, let's just try updating this commit

https://github.com/pytorch/torchchat/blob/main/.github/workflows/pull.yml#L729