[RLlib] Fix IQL beta validation rejecting paper-default values

[Swigler]

Summary

Fixes #64954.

IQLConfig inherits MARWILConfig.validate() which rejects beta > 1.0. IQL uses beta as an inverse temperature (paper default: 3.0), not as a [0, 1] weight like MARWIL. The inherited cap makes the algorithm degenerate into near-uniform behavioral cloning — you literally cannot use IQL as designed.

Fix: Bypass MARWIL's beta range check in IQLConfig.validate() by temporarily setting beta to a safe value during the parent call, then restoring it and applying IQL's own constraint (beta > 0).

Test plan

• IQLConfig(beta=3.0).validate() no longer raises (was: ValueError: beta must be within 0.0 and 1.0)
• IQLConfig(beta=0.0).validate() still raises (IQL's own check: beta must be > 0)
• IQLConfig(beta=-1.0).validate() still raises
• Existing RLlib IQL tests pass

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request modifies the validate method in IQLConfig to temporarily override self.beta during the parent class's validation, bypassing its parameter cap. The review feedback suggests wrapping this temporary state mutation in a try...finally block to ensure self.beta is safely restored even if an exception is raised during validation.

[gemini-code-assist]

If super().validate() raises an exception (for example, due to other invalid configuration parameters), self.beta will not be restored to its original value, leaving the configuration object in a mutated/corrupted state. Wrapping the temporary state mutation in a try...finally block ensures that self.beta is always safely restored.

Suggested change

	saved_beta = self.beta
	self.beta = 0.5
	super().validate()
	self.beta = saved_beta
	saved_beta = self.beta
	try:
	self.beta = 0.5
	super().validate()
	finally:
	self.beta = saved_beta

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit 09c0f1a. Configure here.}