[BUG] When the model outputs slightly more content or when the input file (1000 lines of code) is large, garbled printing may occur

[chall1123]

Problem (one or two sentences)

When analyzing files, there may be garbled characters at the beginning, or the output may take a while to complete

Context (who is affected and when)

garbled printing may occur

Reproduction steps

vllm 0.9.2
PyTorch version： 2.8.0+cu128
CUDA version： 12.8
1、start my model:
docker run -d
--network=host
--shm-size=16G
--device=/dev/kfd
--device=/dev/mkfd
--device=/dev/dri
-v /opt/hyhal:/opt/hyhal
-v /root:/workspace/
--group-add video
--cap-add=SYS_PTRACE
--security-opt seccomp=unconfined
-e HY_VISIBLE_DEVICES=all
image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1-rc5-rocblas104381-0915-das1.6-py3.10-20250916-rc2-ds3.2
python -m vllm.entrypoints.openai.api_server
--model /workspace/models/Qwen3-Coder-30B-A3B-Instruct
--served-model-name qwen
--host 0.0.0.0
--port 8000
--dtype half
--tensor-parallel-size 8
--gpu-memory-utilization 0.9 --max-num-batched-tokens 256000
2、roocode connect my model Qwen3-Coder-30B-A3B-Instruct, and ask question: please analysis my code @test.c (test.c is 1000 lines)

Expected result

can normal analysis my test.c

Actual result

garbled printing may occur，i must stop my task

Variations tried (optional)

Can Roocode open a certain log to confirm if it is a model issue or a Roocode issue? I have not encountered this kind of garbled printing when using other extensions

App Version

3.28.18

API Provider (optional)

None

Model Used (optional)

Qwen3-Coder-30B-A3B-Instruct

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'm exploring a potential fix for the garbled character issue when handling large file outputs with vLLM models. I'll investigate the streaming/encoding handling and open a PR to share progress.

[roomote]

I've opened PR #8788 with a fix for the garbled character issue when handling large file outputs from vLLM models.

The solution implements a UTF8StreamDecoder that properly handles multi-byte UTF-8 characters split across streaming chunk boundaries. This ensures clean output even with large files containing non-ASCII characters.

The fix has been tested with comprehensive unit tests and all CI checks are running. Feedback welcome!

[hannesrudolph]

@chall1123 this is caused by the model being less capable. Have you tried adjusting the temperature?

[chall1123]

@chall1123 this is caused by the model being less capable. Have you tried adjusting the temperature?
I set temperature from 0 to 2, but the output always garbled, Can Roo Code see the native output of the model