delete useless code

[zhoutianzi666]

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

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]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-05-29 23:43:43

CI报告基于以下代码生成（30分钟更新一次）:

• PR commit: de30baf
• Merge base: 791d77f (branch: develop)
• 查看完整 Diff
• CI 详情

---

1 任务总览

存在 1 个 required 失败任务，需优先处理后方可合并。

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
42(0)	42	37	5	0	0	0

2 任务状态汇总

2.1 Required任务 : 9/10 通过

必选任务阻塞合并，失败需优先处理。

状态	任务	耗时	根因	修复建议	日志	重跑
❌	Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage	1h31m	PR问题：decode 写KV cache逻辑被替换	恢复 decoder/speculate 写 cache 分支	Job	-
✅	其余 9 个必选任务通过	-	-	-	-	-

2.2 可选任务 — 28/32 通过

可选任务不阻塞合并，失败仅供参考。

状态	任务	耗时	日志	重跑
❌	Run iluvatar Tests / run_iluvatar_cases	15m11s	Job	-
❌	Check PR Template	20s	Job	-
❌	CI_HPU	1h7m	Job	-
❌	Trigger Jenkins for PR	15s	Job	-
✅	其余 28 个可选任务通过	-	-	-

---

3 失败详情（仅 required）

Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage — 测试失败（置信度: 中）

Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage

• 状态: ❌ 失败
• 错误类型: 测试失败
• 置信度: 中
• 根因摘要: decode 写KV cache逻辑被替换
• 分析器: ci_analyze_unittest_fastdeploy

失败用例:

测试	错误	根因
tests/e2e/test_ernie_03b_pd_decode_unified_attention.py::test_non_chat_usage_non_stream	AttributeError: 'NoneType' object has no attribute 'json'	HTTP 请求超时返回 None，服务端 decode worker 已异常断连
tests/e2e/test_EB_Lite_serving_R3.py::test_r3_accuracy	EOFError: Ran out of input	routing replay 输出 layer_0.pdtensor 为空，生成流程未正常完成

根因详情:
本 PR 修改 custom_ops/gpu_ops/append_attention.cu，将 EncoderWriteCacheWithRopeKernel 从 max_enc_len_this_time > 0 分支中提出并无条件执行，同时删除了原 decode/speculate 阶段的 DecoderWriteCacheWithRoPEKernel / SpeculateWriteCacheWithRoPEKernel 写 cache 逻辑。新的调用还把 encoder kernel 的 seq_lens_encoder 参数传成 seq_lens_this_time，会改变 decode 阶段 KV cache 写入/跳过 batch 的语义。

失败用例覆盖 pd_decode_unified_attention 与 routing replay，运行时日志出现 ConnectionResetError: [Errno 104] Connection reset by peer 和 Failed to connect to engine worker queue，随后测试侧请求超时或读取到空 routing 文件，和本次 append attention decode 写 cache 变更高度相关。

关键日志:

Failed test cases:
tests/e2e/test_ernie_03b_pd_decode_unified_attention.py
E AttributeError: 'NoneType' object has no attribute 'json'
E EOFError: Ran out of input
ConnectionResetError: [Errno 104] Connection reset by peer
ERROR engine_worker_queue.py[line:857] Failed to connect to engine worker queue

修复建议:

1. custom_ops/gpu_ops/append_attention.cu L183-L227/L282 附近：不要用 encoder 写 cache kernel 无条件替代 decode/speculate 写 cache；恢复或等价迁移原 DecoderWriteCacheWithRoPEKernel / SpeculateWriteCacheWithRoPEKernel 分支，确保 decode-only 请求仍按 seq_lens_decoder/seq_lens_encoder 语义写入 KV cache。
2. custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu L386-L414：确认 kv_batch_ids/kv_tile_ids_per_batch/kv_num_blocks_x_cpu 对 encoder、decode、speculate 三种路径的含义一致；如只服务 write-cache-kv，需要补充 decode unified attention e2e 覆盖后再合入。

修复建议摘要: 恢复 decoder/speculate 写 cache 分支

关联变更: custom_ops/gpu_ops/append_attention.cu、custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu

链接: 查看日志

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@791d77f). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7959   +/-   ##
==========================================
  Coverage           ?   67.52%           
==========================================
  Files              ?      467           
  Lines              ?    65179           
  Branches           ?    10007           
==========================================
  Hits               ?    44013           
  Misses             ?    18349           
  Partials           ?     2817           

Flag	Coverage Δ
GPU	77.76% <ø> (?)
XPU	7.08% <ø> (?)

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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-05-29 13:51:55

📋 Review 摘要

PR 概述：删除冗余的 Decoder/Speculate KV 写缓存代码路径，将 encoder/decoder 写缓存统一到 EncoderWriteCacheWithRopeKernel 并无条件执行
变更范围：custom_ops/gpu_ops/append_attention.cu、custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
影响面 Tag：[OP] [KVCache]

问题

级别	文件	概述
❓ 疑问	append_attention.cu	EncoderWriteCacheWithRopeKernel 无条件调用且第 4 参数改为 seq_lens_this_time，纯 Decode 批次下行为需确认
❓ 疑问	get_block_shape_and_split_kv_block.cu	split_kv_block 提升为无条件执行且切换为 seq_lens_this_time，纯 Decode 批次下 kv_batch_ids 填充正确性需确认

---

疑问 1 — append_attention.cu

EncoderWriteCacheWithRopeKernel 现在在 if (max_enc_len_this_time > 0) 之外被无条件调用，同时第 4 个参数（原 seq_lens_encoder）改为 seq_lens_this_time：

// 新代码（两个位置均传 seq_lens_this_time）
EncoderWriteCacheWithRopeKernel<...>(
    meta_data, qkv,
    seq_lens_this_time,   // param3: seq_lens_this_time ✓
    seq_lens_this_time,   // param4: 原 seq_lens_encoder ← 已变更
    seq_lens_decoder, ...);

函数内部（encoder_write_cache_with_rope_kernel.h）将此参数传给 rotary_qk_variable 以及 CascadeAppendWriteCacheKVQKV，后者的底层 cache_kernel 用它做跳过判断：

if (seq_lens[ori_bi] == 0) continue;  // 原语义: 无 encoder token 则跳过

在纯 Decode 批次（seq_lens_encoder[i] == 0，但 seq_lens_this_time[i] = 1），旧代码因 if (max_enc_len_this_time > 0) 不进入此 kernel；新代码将进入且 seq_lens[ori_bi] = 1 ≠ 0，不会跳过。同时 DecoderWriteCacheWithRoPEKernel 已被删除，请确认 EncoderWriteCacheWithRopeKernel 能完整替代其功能（尤其 ori_seq_id 计算是否与 decoder 场景一致）。

疑问 2 — get_block_shape_and_split_kv_block.cu

split_kv_block 第 2 参数改为 seq_lens_this_time，且整个 KV block 切分逻辑由 if (max_enc_len_this_time > 0) 内提升为无条件执行。请确认：

• 对纯 Decode 批次（全部请求均为 1 token decoder），split_kv_block 填充后的 kv_batch_ids / kv_tile_ids_per_batch 是否会被 EncoderWriteCacheWithRopeKernel 内的量化写缓存路径（CascadeAppendWriteCacheKVC8QKV / CascadeAppendWriteCacheKVC4QKV）正确消费。

---

📝 PR 规范检查

标题缺少官方 Tag，且 Motivation / Modifications / Usage or Command / Accuracy Tests 各段均为空，Checklist 全部未勾选。

标题建议（可直接复制）：

• [OP] Delete useless decoder and speculate write cache kernel calls

PR 描述建议（点击展开，可直接复制）

## Motivation
将 encoder / decoder / speculate 三条 KV 写缓存路径统一，删除冗余的 `SpeculateWriteCacheWithRoPEKernel` 和 `DecoderWriteCacheWithRoPEKernel` 调用，简化 `AppendAttentionKernel` 逻辑。同时在 `GetBlockShapeAndSplitKVBlock` 中将 KV block 切分无条件执行并改用 `seq_lens_this_time`，与上层统一路径对齐。

## Modifications
- `custom_ops/gpu_ops/append_attention.cu`：将 `EncoderWriteCacheWithRopeKernel` 调用提升到 `if (max_enc_len_this_time > 0)` 之外，无条件执行；第 4 参数改为 `seq_lens_this_time`；删除 `SpeculateWriteCacheWithRoPEKernel` 和 `DecoderWriteCacheWithRoPEKernel` 共 4 个调用点
- `custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu`：将 KV block 切分（`split_kv_block`）从 encoder-only 条件块改为无条件执行；参数改为 `seq_lens_this_time`；为 `kv_num_blocks_x_cpu.copy_` 添加 CUDAGraph capture 保护宏

## Usage or Command
N/A

## Accuracy Tests
N/A

## Checklist

- [x] 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.

总体评价

PR 变更思路清晰——将三条 KV 写缓存路径合一是合理重构，代码量大幅减少。但统一后 EncoderWriteCacheWithRopeKernel 在纯 Decode 批次下的语义正确性（第 4 参数语义替换 + 无条件调用）建议在 PR 描述中给出说明或补充测试覆盖，以便 Reviewer 确认功能等价。