float16 model w/ various quant options will generate garbage output

[guangy10]

🐛 Describe the bug

The issue is observed from torchchat.

With prompt "Hello, my name is", the generated output will be the following very consistent from run to run.

Hello, my name is ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇  ⁇ 

See T186105376 for detailed stacktrace and steps to repro

Versions

latest nightly

[malfet]

Move to torcchat, as it's absolutely non-actionable for PyTorch.
Again, would be great to provide exact system details/commands one used to repro it

[malfet]

Fun fact, it happens only if --temperature 0 option is passed
And if one looks at

torchchat/generate.py

Line 95 in f5b7f41

logits = logits / max(temperature, 1e-5)

And for float16 every x such that $|x| \ge 1.0$ then $|x/10^{-5}| = \infty $