TypeError with torch.finfo()

[athitten]

🐛 Bug

NeMo's Stable Diffusion uses HuggingFace Transformers that uses torch.finfo in one of its operations. When thunder.jit is used with Stable Diffusion, it gives an error:

TypeError (note: full exception trace is shown but execution is paused at: _run_module_as_main)
finfo(): argument 'type' (position 1) must be torch.dtype, not floating

although torch supports the dtype in torch.finfo(dtype) to be either torch.float32 or just float32.

CC: @tfogal

cc @tfogal

[tfogal]

Not really understanding this one. We don't appear to have a lookaside for torch.finfo; should we?

[t-vi]

Without a traceback, it is hard to guess:

torch.finfo(thunder.dtypes.float32)

gives

TypeError                                 Traceback (most recent call last)
Cell In[14], line 1
----> 1 torch.finfo(thunder.dtypes.float32)

TypeError: finfo(): argument 'type' (position 1) must be torch.dtype, not floating

likely options:

• fix gap in translating thunder dtypes to torch.dtype transparently,
• add torch.finfo lookaside

As an aside: Quite likely, it would help debugging these cases if Thunder objects had more Pythonic representations that informed about the fact that they are thunder objects... (ported over #510 for this)

[athitten]

@athitten to add the full trace.

[tfogal]

dropping 'triage review' until the trace comes in, will revisit then.

[t-vi]

Repro:

import torch, thunder

def fn(x):
    return torch.finfo(x.dtype).min

a = torch.randn(5, 5)

jfn = thunder.jit(fn)

jfn(a)

Aside from the obvious implementation strategy of implementing finfo: Maybe it would be nice to do a replacement of thunder dtype (and device) to PyTorch when calling opaque functions.

[k223kim]

Hello Team! FYI I will be working on this issue. (cc. @t-vi )

[t-vi]

Thank you @k223kim