[Feature]: MTP on ROCm

[tjtanaa]

🚀 The feature, motivation and pitch

1. Last check that MTP does not help with performance. After a few months, now we need to check if MTP help improve performance for DeepSeekV3. Update Badr tested that V1 DeepSeekV3 works with MTP.
2. Find how to evaluate the correctness of the DeepSeekV3 with MTP
3. Perform Server benchmark with MTP and without MTP.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[vllmellm]

To run MTP, use:

    llm = LLM(
        model="deepseek-ai/DeepSeek-R1",
        speculative_config={
            "num_speculative_tokens": 1,
            "draft_tensor_parallel_size": 8
        },
        quantization="fp8",
        # gpu_utilization=0.7,
        tensor_parallel_size=8,
        block_size=1,
        max_model_len=32768,
    )