[Kernel] add triton kernels for sampling

[swy20190]

What this PR does / why we need it?

Replace pyorch implement of sampling with triton kernels

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: v0.11.0
• vLLM main: vllm-project/vllm@2918c1b

[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 replaces several PyTorch-based sampling functions with Triton kernels, aiming to improve performance. The new kernels for expand_kernel and sample_recovered_tokens_kernel are well-implemented. However, rejection_greedy_sample_kernel and rejection_random_sample_kernel contain sequential loops over draft tokens. This is a performance anti-pattern in Triton that prevents vectorization and may not yield the desired speed-up. I've added specific comments with suggestions to vectorize these kernels for better efficiency.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[whx-sjtu]

After communicating with @wangxiyuan, you need to use HAS_TRITON check here. If user hasn't installed triton, fall back to original implementation to ensure functionality. An example is here:

vllm-ascend/vllm_ascend/attention/sfa_v1.py

Line 505 in a3225c4

if HAS_TRITON:

[whx-sjtu]

Use from vllm.triton_utils import tl, triton here, same reason.

[whx-sjtu]

In conclusion, currently we need to make sure that original functionality won't be broken by new triton optimization in environments without triton.