fix: optimize InfiniLM paged attention kernels

[voltjia]

Summary

• Port the faster InfiniLM paged-attention decode and prefill kernel paths into src/native/cuda/ops/paged_attention_infinilm/kernel.cuh and src/native/cuda/ops/paged_attention_prefill_infinilm/kernel.cuh.
• Use split-KV CTA decode for the common head_dim=128 path, with workspace-backed combine.
• Use the pipelined warp-CTA prefill path for head_dim=128 and block_size=256, while keeping existing fallback coverage.
• Clean up the migrated internal kernel helper names, template parameters, and parameter lists to follow Google C++ style conventions.

Motivation

The InfiniOps integration path for InfiniLM paged attention was significantly slower than the previous InfiniCore implementation because it did not use the old fast paged-attention kernels for common InfiniLM shapes.

Before optimization, the integrated InfiniOps path measured about 5427.75 ms total time for the benchmark below, versus about 3069.50 ms for the no-InfiniOps baseline on the same model/shape. This PR restores the decode split-KV CTA path and the pipelined prefill path, bringing the optimized InfiniOps path back to about 3096.99 ms.

Benchmark setup:

• Hardware: NVIDIA GPU2 on the remote NVIDIA test machine.
• Model: /data-aisoft/mechdancer/models/9g_8b_thinking.
• Command shape: bs=1, input_len=256, output_len=256, bfloat16, paged attention + graph enabled.
• Command: python examples/bench.py --device nvidia --model=/data-aisoft/mechdancer/models/9g_8b_thinking --enable-paged-attn --enable-graph --input-len=256 --output-len=256 --batch-size=1 --warmup.

Closes #

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 (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

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

Exact InfiniOps pytest -m smoke was not run in this workspace. Validation was done through the InfiniCore/InfiniLM integration path that exercises these InfiniLM paged-attention kernels on NVIDIA SM80.

# Integrated InfiniCore + InfiniOps NVIDIA SM80 build
infinicore import ok <module 'infinicore' from '/workspace/perf/core-ops-sm80/python/infinicore/__init__.py'>
libinfiniops.so installed under /workspace/perf/root-ops-sm80/lib

# InfiniLM rebuild against the integrated environment
build ok
install ok
imports ok <module 'infinicore' ...> <module 'infinilm' ...>

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Integrated SM80 build/import passed; InfiniLM build/import passed.	Full InfiniOps pytest not run.
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	No	N/A - not affected	N/A - not affected
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Full `pytest` output (optional)

N/A - full pytest was not run for this NVIDIA-only InfiniLM paged-attention kernel change.

Benchmark / Performance Impact

Focused benchmark on NVIDIA GPU2 with 9g_8b_thinking, bs=1, input_len=256, output_len=256, bfloat16, paged attention + graph enabled:

Path	total_time	Prefill TTFT	Decode throughput
InfiniOps before optimization	5427.75 ms	45.36 ms	47.38 tok/s
no-InfiniOps baseline	3069.50 ms	30.13 ms	83.93 tok/s
This PR after style cleanup	3096.99 ms	31.83 ms	83.21 tok/s

The optimized InfiniOps path is within about 1% of the measured no-InfiniOps baseline for this focused case.

Notes for Reviewers

• The moved kernels are now contained directly in the InfiniOps kernel.cuh files for decode and prefill; there are no legacy helper files and no cross-tree includes back into InfiniCore.
• Please pay particular attention to the dispatch conditions for the fast paths: decode head_dim=128 split-KV CTA and prefill head_dim=128 && block_size=256 pipelined warp-CTA.
• The fallback paths remain in place for unsupported shapes.