[Bugfix] fix kv nz accuracy bug

[realliujiaxu]

What this PR does / why we need it?

when enable_kv_nz is true, output of Deepseek R1 is invalid.

{"id":"chatcmpl-f01fa84f397b417abf6d3c5243787d38","object":"chat.completion","created":1758106562,"model":"auto","choices":[{"index":0,"message":{"role":"assistant","reasoning_content":null,"content":"<think>0I0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0O0","tool_calls":[]},"logprobs":null,"finish_reason":"length","stop_reason":null}],"usage":{"prompt_tokens":7,"total_tokens":107,"completion_tokens":100,"prompt_tokens_details":null},"prompt_logprobs":null,"kv_transfer_params":null}

The reason is that the decode stage incorrectly usage of torch_npu.npu_kv_rmsnorm_rope_cache. After fix:

{"id":"chatcmpl-4a9e79534b2b4945a9496621d1dc53f6","object":"chat.completion","created":1758108051,"model":"auto","choices":[{"index":0,"message":{"role":"assistant","reasoning_content":null,"content":"<think>\nOkay, the user sent \"Hello, my name is\" and then the conversation ends. They might have intended to complete the sentence but didn't. I should prompt them to provide their name so I can address them properly. Maybe respond with a friendly greeting and ask for their name. Let me make sure to keep it welcoming and open. Something like, \"Hello! It's nice to meet you. Could you please tell me your name?\" That should encourage them to share their name without","tool_calls":[]},"logprobs":null,"finish_reason":"length","stop_reason":null}],"usage":{"prompt_tokens":7,"total_tokens":107,"completion_tokens":100,"prompt_tokens_details":null},"prompt_logprobs":null,"kv_transfer_params":null}

run server:

nohup python -m vllm.entrypoints.openai.api_server --model=DeepSeek-R1-W8A8-VLLM \
    --served-model-name auto \
    --trust-remote-code \
    --distributed-executor-backend=mp \
    --port 8009 \
    -tp=8 \
    -dp=2 \
    --max-num-seqs 24 \
    --max-model-len 32768 \
    --max-num-batched-tokens 8192 \
    --additional-config '{"torchair_graph_config":{"enabled":true,"use_cached_graph":true,"graph_batch_sizes":[24], "enable_kv_nz": true},"expert_tensor_parallel_size":16}' \
    --block-size 128 \
    --gpu-memory-utilization 0.96 &> run.log &
disown

curl:

curl --location 'http://127.0.0.1:8009/v1/chat/completions' --header 'Content-Type: application/json' --data '{
        "top_p": 1,
        "ignore_eos": false,
        "stream": false,
        "max_tokens": 100,
        "stop": "None",
        "top_k": -1,
        "temperature": 0.6,
        "messages": [
            {
                "role": "system",
                "content": "Hello, my name is"
            }
        ]
    }'

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: v0.10.2
• vLLM main: vllm-project/vllm@2b85697

[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 addresses a critical bug causing invalid output for Deepseek R1 models when enable_kv_nz is active. The root cause was an incorrect cache_mode ("PA_BLK_NZ") used during the prefill stage in vllm_ascend/attention/mla_v1.py and vllm_ascend/torchair/torchair_mla.py. The fix correctly changes this to "PA_NZ", aligning it with the cache_mode used in the decode stage. This change is correct and effectively resolves the accuracy issue as demonstrated in the pull request description.

[Yikun]

It seems this introduce the break on main:
https://github.com/vllm-project/vllm-ascend/actions/runs/17816277772/job/50649989352?pr=2907
https://github.com/vllm-project/vllm-ascend/actions/runs/17816025095/job/50649237030

Retry: https://github.com/vllm-project/vllm-ascend/actions/runs/17817506855 to confirm

@jianzs BTW, please make sure every PR merged before label with ready and ready-for-test