[v0.7.3]spec decode MultiStepWorker support TP1DraftModelRunner fully

[mengwei805]

What this PR does / why we need it?

spec decode MultiStepWorker support TP1DraftModelRunner fully, support run the draft_model_runner with multi-step prepare on the NPU directly and support draft_model_runner use MLA.

1. before this pr, MultiStepWorker would not step into the branch using NPU prepare, but only into the branch using CPU prepare (line 52 of vllm_ascend/patch/patch_multi_step_worker.py). Although this has no effect on the correct operation of speculative decoding and the performance of the two branches is basically the same as of the current version, I support entering this branch in this PR. In general, there are two main changes in patch_multi_step_worker.py: first, the is_cuda_like() check is removed and the TP1DraftModelRunner rewritten in vllm_ascend is used; second, the supports_gpu_multi_step() function is made to return true on NPU devices when outer Multi_step_worker could work correct.

2. before this pr, TP1DraftModelRunner only supports Attention on NPU, but not MLA. The relevant adaptation is in vllm_ascend/worker/draft_model_runner.py. Although I don’t know why the input_positions of model_input.attn_metadata in vllm-ascend needs to be added in execute_model, it is done in model_runner.py, so I also made corresponding changes. Otherwise, when atten_backend is MLA, it will prompt that input_positions cannot be found.

3. I commented out two lines in draft_model_runner.py in line118 to support the scenario of K>1.

# lora_mapping=model_input.lora_mapping,
# lora_requests=model_input.lora_requests,

I added comments. In the future, when vllm-ascend supports lora feature, the changes here can be restored.

Does this PR introduce any user-facing change?

This PR has no effect on users.

How was this patch tested?

have tested

[wangxiyuan]

can you describe what the patch_multi_step_worker.py is changed in commit message? Thanks.

[MengqingCao]

can you describe what the patch_multi_step_worker.py is changed in commit message? Thanks.

@mengwei805 LGTM after this is added, I think adding related comments and updating pr message will make it clear, thx!

[mengwei805]

can you describe what the patch_multi_step_worker.py is changed in commit message? Thanks.

Of course, I'm sorry I was so brief before. I have added relevant information in PR msg and commit msg.
In addition, for the modification of line 162 in vllm_ascend/worker/draft_model_runner.py in this PR,

# TODO: Add support for ASCEND when outer multi_step_worker
# could work correct.
if self.attn_backend.get_name() not in ("FLASH_ATTN", "TRITON_MLA"):
    return False

I suggest that vllm-ascend's atten_backend can distinguish between ordinary Attention and MLA to facilitate the subsequent expansion of related functions. Currently, get_name() of both types of atten returns ASCEND

[mengwei805]

can you describe what the patch_multi_step_worker.py is changed in commit message? Thanks.

@mengwei805 LGTM after this is added, I think adding related comments and updating pr message will make it clear, thx!

Of course, I'm sorry I was so brief before. I have added relevant information in PR msg and commit msg.

[MengqingCao]

lgtm now, thx

[wangxiyuan]

It's very clear. Thanks