feat(nvidia): add moe_align_block_size operator

[voltjia]

Summary

• Add the NVIDIA moe_align_block_size operator and generated Python binding.
• Support the vLLM low-level ABI with optional global-to-local expert_map, deterministic expert grouping, block padding, non-default streams, and multi-GPU device guards.
• Add focused correctness, invalid-ID, sparse/boundary, repeatability, descriptor-reuse, stream, and device tests.

Motivation

moe_align_block_size prepares routed token indices and expert block IDs for blockwise MoE GEMM. It is an independent vLLM MoE primitive needed while moving serving kernels behind InfiniOps.

This PR aligns with the low-level stable vLLM operator, not the higher-level Python allocation helper. InfiniOps keeps its repository-wide execution convention of input tensors first, scalar attributes next, and mutable outputs last.

Alignment reference (pinned to vLLM commit 2f75e7f):

InfiniOps execution API	vLLM low-level interface	Evidence
moe_align_block_size(topk_ids, num_experts, block_size, sorted_token_ids, experts_ids, num_tokens_post_pad) moe_align_block_size(topk_ids, expert_map, num_experts, block_size, sorted_token_ids, experts_ids, num_tokens_post_pad)	moe_align_block_size(topk_ids, num_experts, block_size, sorted_token_ids, experts_ids, num_tokens_post_pad, maybe_expert_map)	Stable C++ declaration, CUDA semantics, High-level allocation helper

N/A - no linked issue.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

The operator implementation and focused test at rebased commit 63296a2b are byte-for-byte identical to validated commit 53a686af. That tree was tested from a clean archive on ssh nvidia using accelerator-dev/nvidia:latest (sha256:dd94fce2f83a...), NVIDIA A100-SXM4-80GB, and PyTorch 2.10.0a0+b4e4ee8.

After #803 merged, this PR no longer adds moe_align_block_size to the permanent smoke build or smoke matcher. The existing moe_sum case represents the vLLM MoE integration in the routine smoke suite; this operator remains covered by its focused test file.

INFINI_RT_ROOT=/opt/infinirt scripts/dev/build.sh nvidia --smoke --jobs 16 --torch-jobs 2 --binding-jobs 2
[build] done

PYTHONPATH=build-nvidia/install python3 -m pytest tests/test_moe_align_block_size.py -q
18 passed, 1 skipped in 36.58s

PYTHONPATH=build-nvidia/install python3 -m pytest tests -m smoke -q
77 passed, 14 skipped, 12313 deselected in 8.31s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	operator-pruned focused build passed; vLLM MoE smoke is represented by moe_sum from #803	focused moe_align_block_size: 18 passed, 1 skipped; full suite intentionally not run
Iluvatar	No	N/A - not affected	N/A
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	No	N/A - not affected	N/A

Focused coverage

The focused suite checks exact deterministic token ordering, five repeated launches, invalid raw and mapped expert IDs, sparse and zero-token experts, more than 256 routes, the num_experts=257 / block_size=128 boundary, non-default streams, multi-GPU device selection, and safe descriptor reuse. Descriptor reuse permits fresh data pointers with identical metadata and rejects changed attributes, tensor metadata, or expert_map presence before launch.

Benchmark / Performance Impact

N/A - this PR establishes the aligned interface and deterministic correctness baseline; it makes no performance claim.

Notes for Reviewers

• Review the low-level API mapping and output capacity rules against the pinned vLLM implementation above.
• InfiniOps places optional tensor inputs before scalar attributes and mutable outputs last; the semantic arguments and defaults are otherwise the low-level vLLM contract.
• Invalid global or mapped expert IDs are ignored before indexing. Output tails are initialized to the vLLM-compatible token sentinel / -1 expert sentinel.
• The stable scatter is a single-writer O(N) phase. It avoids nondeterministic atomic placement, but very large route counts may warrant a separately benchmarked parallel optimization later.
• Scope is intentionally limited to this one NVIDIA operator. moe_sum was merged independently in feat(nvidia): add moe_sum operator #803 and supplies the vLLM MoE smoke coverage.