[Feature] Add Triton unified attention kernel for deterministic inference

[gongweibao]

Motivation

为确定性推理模式提供 Triton unified attention kernel。该 kernel 通过 unified kv_indices 将 prefix（cached）和 extend（new）KV tokens 统一处理，确保无论 prefix cache 命中与否，累加顺序完全一致，从而实现 split-invariant 的确定性 attention 输出。

本 PR 是纯算子层面的实现，不涉及推理流程集成（流程集成见后续 PR）。

Modifications

新增文件：

• fastdeploy/model_executor/layers/attention/triton_ops/__init__.py — 包导出
• fastdeploy/model_executor/layers/attention/triton_ops/unified_extend_attention.py — 核心实现：
  • _fwd_kernel_unified: Triton JIT kernel，online softmax + paged KV cache + causal mask
  • Index building utilities: triton_cumsum_with_zero_prefix, build_kv_indices_from_block_tables, build_unified_kv_indices, _scatter_extend_kv_indices_kernel（均 CUDA Graph 兼容）
  • pre_cache_len_concat_triton: GPU-only Triton 替代 C++ op
  • Reference implementations（_ref 后缀）用于正确性验证
• tests/deterministic/test_unified_extend_attention.py — kernel 正确性单测
• tests/deterministic/test_build_triton_indices.py — indices 构建逻辑单测

Usage or Command

# 运行 kernel 正确性测试
source /root/paddlejob/workspace/env_run/gongweibao/archfd/fdarchenv/bin/activate
CUDA_VISIBLE_DEVICES=0 python -m pytest tests/deterministic/test_unified_extend_attention.py -v
CUDA_VISIBLE_DEVICES=0 python -m pytest tests/deterministic/test_build_triton_indices.py -v

Accuracy Tests

测试覆盖：

• Kernel 正确性：MHA/GQA/MQA, head_dim=13/64/80/96/128/256, float16/bfloat16
• Split invariance：cache miss vs hit 输出 bit-identical
• Determinism：多次运行 bitwise identical
• Production-scale：bs=19, seq=4096, mixed lengths
• Cross-validation：naive vs sdpa reference vs triton 三方交叉验证

Checklist

• Add at least a tag in the PR title.
• Format your code, run pre-commit before commit.
• Add unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

🤖 Generated with Claude Code

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

gongweibao seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

❌ Patch coverage is 43.19066% with 146 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@f0ab8ee). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...s/attention/triton_ops/unified_extend_attention.py	43.19%	141 Missing and 5 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6795   +/-   ##
==========================================
  Coverage           ?   72.36%           
==========================================
  Files              ?      396           
  Lines              ?    54956           
  Branches           ?     8635           
==========================================
  Hits               ?    39770           
  Misses             ?    12362           
  Partials           ?     2824           

Flag	Coverage Δ
GPU	72.36% <43.19%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.