Fix incorrect MLAPO weight release in PD mixex scenarios.

[ZYang6263]

What this PR does / why we need it?

Fix incorrect MLAPO weight release in PD mixex scenarios.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.12.0
• vLLM main: vllm-project/vllm@ad32e3e

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request correctly addresses a bug where MLAPO weights were being released prematurely in mixed prefill-decode scenarios. The approach of moving the weight release to a conditional block that checks for a KV consumer role is sound. However, the current implementation introduces a memory leak in non-PD scenarios by failing to release weights that were previously freed. I have added a critical comment with a suggested fix to ensure memory is managed correctly in all configurations.

[gemini-code-assist]

The current logic for releasing MLAPO weights only covers the case where the node is a KV consumer in a Prefill-Decode (PD) mixed scenario. This correctly fixes the original bug but introduces a memory leak in non-PD (standalone) scenarios.

Previously, self.fused_qkv_a_proj.weight was released unconditionally after its data was processed. With this change, it is no longer released in non-PD scenarios, as self.vllm_config.kv_transfer_config would be None.

The logic should be updated to release the processed weights in both non-PD scenarios and on the consumer side of PD scenarios. This ensures memory is freed correctly in all configurations.

Suggested change

	if self.vllm_config.kv_transfer_config is not None and \
	self.vllm_config.kv_transfer_config.is_kv_consumer:
	self.fused_qkv_a_proj.weight = None
	self.fused_qkv_a_proj.deq_scale = None
	self.fused_qkv_a_proj.quant_bias = None
	self.q_proj.weight = None
	self.q_proj.deq_scale = None
	self.q_proj.quant_bias = None
	torch.npu.empty_cache()
	if self.vllm_config.kv_transfer_config is None or \
	self.vllm_config.kv_transfer_config.is_kv_consumer:
	self.fused_qkv_a_proj.weight = None
	self.fused_qkv_a_proj.deq_scale = None
	self.fused_qkv_a_proj.quant_bias = None
	self.q_proj.weight = None
	self.q_proj.deq_scale = None
	self.q_proj.quant_bias = None
	torch.npu.empty_cache()