[BugFix][Graph Optimization] Fix deterministic mode + CUDAGraph garbled output by making reset_share_inputs CUDAGraph-safe`

[gongweibao]

Motivation

When FD_DETERMINISTIC_MODE=1 and use_cudagraph=True are both enabled, model output is garbled. Either feature works correctly in isolation, but the combination fails.

Root cause: reset_share_inputs() is called after CUDAGraph capture in the deterministic mode post-warmup path (gpu_worker.py). Several lines inside reset_share_inputs() create new tensor objects (paddle.to_tensor(...) / paddle.full(...)), which allocate new GPU addresses. CUDAGraph has already baked the old addresses during capture, so replay reads/writes stale memory — producing garbage output.

Modifications

fastdeploy/worker/input_batch.py:

• For 4 tensor fields (reasoning_allowed_tokens, free_list, free_list_len, rope_emb) that previously created new tensor objects in reset_share_inputs(), add FD_DETERMINISTIC_MODE guard to use in-place operations (fill_, copy_, slice assignment) instead, preserving GPU addresses for CUDAGraph safety.
• Non-deterministic mode behavior is completely unchanged.

fastdeploy/worker/gpu_worker.py:

• Remove the if not self.model_runner.use_cudagraph skip around reset_share_inputs(). Since reset_share_inputs() is now CUDAGraph-safe under deterministic mode, it can be called unconditionally.

Accuracy Tests

CUDA_VISIBLE_DEVICES=0,1,2,3 pytest tests/e2e/4cards_cases/test_determinism_long.py -v

[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.}

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 5.55556% with 17 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@30f9f33). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/input_batch.py	5.55%	17 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6746   +/-   ##
==========================================
  Coverage           ?   71.95%           
==========================================
  Files              ?      392           
  Lines              ?    53848           
  Branches           ?     8463           
==========================================
  Hits               ?    38748           
  Misses             ?    12328           
  Partials           ?     2772           

Flag	Coverage Δ
GPU	71.95% <5.55%> (?)

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.