fix: don't crash the quantile token loss when no kernel is configured

[brettbj]

How to recreate
Train with the shipped config/main.yaml (it sets custom_loss: true and a
quantile_token_loss block without a kernel key). The first training step
raises omegaconf.errors.ConfigAttributeError: Key 'kernel' is not in struct
from aggregate_first.

Why it's a bug
Loss.__init__ carefully guards kernel handling (if "kernel" in self.cfg.quantile_token_loss: ...), but both inner loss functions read self.cfg.quantile_token_loss.kernel.type unconditionally. Additionally, a kernel block with type but no factor left self.kernel_factor unset, raising AttributeError at first use. The default config shipped in this repo hits the first path.

How it was fixed
Resolve kernel_type (default "linear") and kernel_factor (default 1.0) once in __init__ via cfg.get(...) and reference only those attributes inside aggregate_first / loss_first. The linear kernel is the identity, so the defaults preserve previous intended behavior. Verified a forward pass of quantile_token_loss with the kernel-less config now returns a finite loss.

🤖 Generated with Claude Code

[brettbj]

if it hits this and kernel is not set, we might actually want it to crash rather than defaulting to linear without us knowing?

[burkh4rt]

This was part of the reason I started working off the https://github.com/bbj-lab/cotorra/tree/coreopsis-stable branch