Fixes mask comparison and scaling logic in attention kernel

[LoserCheems]

Simplifies mask comparison from > 0 to direct boolean evaluation for consistency.

Removes trailing whitespace for code cleanliness.

Corrects duplicate softmax scaling application that was causing incorrect attention computations.

Improves code readability by moving comment to more appropriate location.

[Copilot]

Pull Request Overview

Fixes inconsistent mask evaluation and corrects softmax scaling in the forward attention kernel, while cleaning up whitespace and refining inline comments.

• Simplify mask check from a boolean comparison to a direct sum-based test
• Ensure softmax scaling is only applied once by conditioning on non--inf entries
• Move and clarify comments; remove trailing whitespace for readability

[Copilot]

Consider using tl.any(mask) (or tl.any(mask != 0)) instead of tl.sum(mask) > 0 to more directly express 'any active elements' and potentially improve efficiency.

Suggested change

	any_active = tl.sum(mask) > 0
	any_active = tl.any(mask)

[Copilot]

This unconditional scaling duplicates the conditional scaling on the next line, resulting in double application of softmax_scale. Remove this line so that the tl.where version is the single scaling step.

Suggested change

	acc_s = acc_s * softmax_scale + bias
	# Removed duplicate scaling to avoid double application of softmax_scale.

[Copilot]

[nitpick] The comment # load v is vague; consider expanding it to clarify what data is being loaded and why, e.g., # load V tensor block for value projection.

Suggested change

	# load v
	# Load the V tensor block for value projection in the attention mechanism.