Feature(MInference): support SGLang and vLLM vertical_and_slash flash attention and index kernels

[iofu728]

What does this PR do?

• Feature: support SGLang [Feat] Add sparse attn to sgl-kernel sgl-project/sglang#5327 and vLLM Implements the attention kernel with vertical and slash sparse pattern described in Appendix C.4.2 of https://arxiv.org/abs/2407.02490 (as sparse_attn_func) vllm-project/flash-attention#33, Implements dual-chunk-flash-attn backend for dual chunk attention with sparse attention support vllm-project/vllm#11844 vertical_and_slash flash attention and index kernels

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Was this discussed/approved via a Github issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@iofu728, @Starmys

[Copilot]

Pull Request Overview

This PR introduces a feature update to support SGLang and vLLM in the flash attention kernels. Key changes include updating error handling in the softmax fusion block, modifying pad computation in the vertical sparse attention function, and adding a new function (sglang_vs) to support SGLang-based flash attention.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
minference/ops/xattention_fa.py	Updated error handling by replacing a breakpoint() with an assert.
minference/ops/pit_sparse_flash_attention_v2.py	Adjusted pad calculation and added try/except imports with a new sglang_vs function for SGLang integration.

[Copilot]

Using 'assert False' for error handling may be less informative in production; consider raising a specific exception (e.g., RuntimeError) with the same message.

[Copilot]

The new pad computation uses a bitwise operation that assumes block_size_M is a power of two; adding a clarifying comment or an explicit check could improve code clarity.

Suggested change

	batch_size, num_heads, context_size, head_dim = query.shape
	batch_size, num_heads, context_size, head_dim = query.shape
	# Ensure block_size_M is a power of two, as required for the bitwise operation below.
	if block_size_M & (block_size_M - 1) != 0 or block_size_M <= 0:
	raise ValueError("block_size_M must be a power of two and greater than zero.")
	# Compute padding size. The bitwise operation assumes block_size_M is a power of two.

[Copilot]

[nitpick] Consider adding a docstring for the new sglang_vs function to explain its purpose, parameter expectations, and how it differs from vertical_slash_sparse_attention.