fix: gemma-4 with exllamav3

[AlpinDale]

No description provided.

[gemini-code-assist]

Code Review

This pull request updates the EXL3 quantization logic to support models with K=V attention, such as Gemma 4, where the v_proj tensor may be absent and duplicated from k_proj. Feedback was provided to optimize the implementation by using short-circuiting and reducing redundant lookups to improve model initialization performance.

[gemini-code-assist]

The current implementation performs redundant lookups and lacks short-circuiting, which can negatively impact model initialization performance. Specifically, _is_exl3_prefix is called for all three projections regardless of the results, and _storage_entry is called twice for v_proj when it is missing from the storage. Since _storage_entry involves multiple string operations and dictionary lookups, it is more efficient to use short-circuiting and store the entry for v_proj to avoid redundant work.

Suggested change

	has_q = self._is_exl3_prefix(f"{base}.q_proj")
	has_k = self._is_exl3_prefix(f"{base}.k_proj")
	has_v = self._is_exl3_prefix(f"{base}.v_proj")
	# Gemma 4 full-attention layers can use K=V attention and store
	# only q_proj/k_proj tensors. The model loader duplicates K into
	# V, so the fused qkv_proj still needs EXL3 parameters.
	return has_q and has_k and (has_v or self._storage_entry(f"{base}.v_proj") is None)
	if not (self._is_exl3_prefix(f"{base}.q_proj") and
	self._is_exl3_prefix(f"{base}.k_proj")):
	return False
	v_entry = self._storage_entry(f"{base}.v_proj")
	# Gemma 4 full-attention layers can use K=V attention and store
	# only q_proj/k_proj tensors. The model loader duplicates K into
	# V, so the fused qkv_proj still needs EXL3 parameters.
	return v_entry is None or v_entry.get("quant_format") == "exl3"