[XPU] improce attn precision

[lizan1999]

Motivation

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

improve speculative_attention_decoder TGEMM from tfloat32 to float
As for flash_attention_context_vllm and paged_attention_xft, use the dedicated version of vLLM directly.

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no 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.

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-20 16:03 CST

📋 Review 摘要

PR 概述：将 XPU speculative_attention_decoder 的 TGEMM 计算类型从 tfloat32 提升为 float，以改善注意力精度。
变更范围：custom_ops/xpu_ops/src/ops/block_attn.cc
影响面 Tag：XPU OP

📝 PR 规范检查

PR 标题存在拼写错误 "improce" → "improve"，且 Motivation 章节未填写、Accuracy Tests 章节为空（本次变更涉及精度提升，建议提供精度对比数据）。PR 描述中提到了对 flash_attention_context_vllm 和 paged_attention_xft 的变更，但 diff 中仅包含 speculative_attention_decoder 的修改，描述与实际变更不一致。

标题建议（可直接复制）：

• [XPU] Improve attention precision for speculative_attention_decoder

描述建议：

• Motivation 章节填写精度改善的背景和原因
• Accuracy Tests 章节补充 tfloat32 vs float 的精度对比数据
• Modifications 章节更新以匹配实际 diff 内容

问题

级别	文件	概述
🟡 建议	block_attn.cc:709	block_attn_spliced.cc 中存在相同的 TGEMM 定义仍使用 tfloat32，未同步更新

总体评价

变更逻辑清晰，将 tfloat32 替换为 float 可提升计算精度。但 block_attn_spliced.cc 中存在完全相同的 TGEMM 定义未被同步修改，建议作者确认是否需要一并更新以保持一致性。

[PaddlePaddle-bot]

🟡 建议 block_attn_spliced.cc 中存在相同的 TGEMM 定义未同步更新

block_attn_spliced.cc:1545-1547 中存在完全相同的代码，仍然使用 tfloat32：

using TGEMM = std::conditional_t<std::is_same_v<XPU_XType, XPU_CType>,
                                 tfloat32,
                                 int8_wo_t>;

该处同样用于 speculative_attention_decoder 的模板参数。如果本 PR 的目的是全面提升 speculative attention 的精度，建议同步将 block_attn_spliced.cc 中的 tfloat32 也修改为 float，以保持两个代码路径的一致性。

请确认是否是遗漏，还是有意保持该文件不变？