fix(dsv4/fp8): import suppress so FP8 indexer prefill cleanup stops masking the real error

[Anai-Guo]

What

rtp_llm/models_py/modules/dsv4/fp8/indexer.py uses suppress(Exception) in two
places but never imports it:

line	context
515	inside except Exception: in _gather_prefill_k_cache, right before raise
543	the whole body of _discard_prefill_k_cache_gather

from __future__ import annotations

import os
from typing import Any, Callable, Dict, NamedTuple, Optional

contextlib never appears in the file. pyflakes on current main:

rtp_llm/models_py/modules/dsv4/fp8/indexer.py:515:26: undefined name 'suppress'
rtp_llm/models_py/modules/dsv4/fp8/indexer.py:543:14: undefined name 'suppress'

Why it matters

Both uses sit on error/cleanup paths, so the missing import turns a real failure
into a misleading NameError and skips the cleanup it was guarding.

_discard_prefill_k_cache_gather is called from two finally: blocks
(indexer.py:1190 and indexer.py:1398). An exception raised inside finally:
replaces the exception being propagated, so whenever a prefill fails while a CP
all-gather handle is still live, the caller never sees the actual error — and the
assembler handle is leaked because discard_assemble_indexer_k_async is never
reached.

The except Exception: handler at 511-517 has the same shape: the
suppress-guarded discard is supposed to run before the raise re-raises the
original prepare_assemble_indexer_k_async failure; instead the handler itself
dies with NameError.

Verification

The two methods were lifted verbatim out of main with ast.get_source_segment
and driven with a stub assembler (no GPU needed), before and after the import:

### UNPATCHED (upstream main)
[1] direct call  Indexer._discard_prefill_k_cache_gather(pending)
    -> NameError: name 'suppress' is not defined
[2] the real shape: forward()'s `finally:` (L1190 / L1398)
    -> caller sees NameError: name 'suppress' is not defined
    cleanup actually reached the assembler: False

### PATCHED  (from contextlib import suppress)
[1] direct call  Indexer._discard_prefill_k_cache_gather(pending)
    -> returned normally; assembler cleanup attempted: 1
[2] the real shape: forward()'s `finally:` (L1190 / L1398)
    -> caller sees RuntimeError: fp8_mqa_indexer_score: DeepGEMM kernel launch failed
    cleanup actually reached the assembler: True

Fix

One import line, matching how the sibling module in the same package already does
it (rtp_llm/models_py/modules/dsv4/fp8/attention.py:19:
from contextlib import contextmanager, suppress), placed in isort order between
import os and the typing import.

---

🤖 Generated with Claude Code

[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.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[rtp-llm-review-bot]

[P3] import 顺序不符合 isort 规范

补丁中新增行 +from contextlib import suppress 位于上下文行 import os 之后。按 isort 默认排序（按模块名字母序，不区分 import/from-import），contextlib < os < typing，因此 from contextlib import suppress 应排在 import os 之前。若仓库启用了 ruff isort 检查可能触发 lint 失败。

建议：将 from contextlib import suppress 移到 import os 之前，保持 stdlib import 按字母序排列。

评审版本：cdf9a9779fb6

[rtp-llm-review-bot]

PR #1412 评审：未发现阻塞项，等待人工审核

• 标题：fix(dsv4/fp8): import suppress so FP8 indexer prefill cleanup stops masking the real error（@Anai-Guo，open）
• 评审版本：cdf9a9779fb6

无阻塞项

非阻塞发现

• [P3] import 顺序不符合 isort 规范 — rtp_llm/models_py/modules/dsv4/fp8/indexer.py:25

---

_{自动代码评审 · head cdf9a9779fb6 · 未发现阻塞项，等待人工审核}

[LLLLKKKK]

AI Code Review - PR #1412

Status: LGTM

Summary: P0/0 · P1/0 · P2/1 · P3/0

Reviewed: commit cdf9a9779fb6 · 2026-09-09 21:37 UTC+8

lgtm ready to ci

Non-blocking Suggestions

P2

• 缺少异常清理路径的回归测试 @ rtp_llm/models_py/modules/dsv4/fp8/indexer.py:25
  • 建议：增加 mock 单测：让 prepare 与 discard 分别抛出不同异常，断言 discard 被调用且最终传播 prepare 的原始异常；同时验证 _discard_prefill_k_cache_gather 会吞掉清理异常。

Checklist Findings (1 fail / 75 total)

General Principles Checklist

• [6.1] Tests — 新逻辑有聚焦单测 + 相关集成/smoke 测试 → issue 缺少异常清理路径的回归测试
  本次修复针对异常清理时 suppress 未定义而掩盖原始错误的问题，但现有测试未覆盖 prepare_assemble_indexer_k_async 失败后的回收分支，也未验证 discard 同时失败时仍传播原始异常，因此该故障可能再次引入而未被自动化测试发现。

Strengths

• 修改范围准确且最小，正常推理路径与资源生命周期不变。
• 清理失败不再覆盖主异常，同时保留尽力回收异步资源的语义。
• 模块级导入不会增加热路径开销。

[LLLLKKKK]

[P2] 缺少异常清理路径的回归测试

本次修复针对异常清理时 suppress 未定义而掩盖原始错误的问题，但现有测试未覆盖 prepare_assemble_indexer_k_async 失败后的回收分支，也未验证 discard 同时失败时仍传播原始异常，因此该故障可能再次引入而未被自动化测试发现。

建议: 增加 mock 单测：让 prepare 与 discard 分别抛出不同异常，断言 discard 被调用且最终传播 prepare 的原始异常；同时验证 _discard_prefill_k_cache_gather 会吞掉清理异常。

Checklist: [6.1] 新逻辑有聚焦单测 + 相关集成/smoke 测试

[rtp-llm-review-bot]

lgtm ready to ci

评审版本：cdf9a9779fb62e76425ff8432c49d48cfc3f9a31

自动评审结论已更新，此前的通过结论不再适用。