[Bug]: random characters as response and crash opencl backend demo

[ramcn]

Prerequisites

• I have searched the existing issues and confirmed this is not a duplicate.
• I am using the latest version of the MLLM framework.

Bug Description

The OpenCL backend demo test is able to load and run, but output <unk> tokens and exits with SEGV.

Below are the steps to reproduce.

Steps to Reproduce

1. Build the opencl demo using the steps here: [Bug]: OpenCL backend example demo #570
2. Generate the GGUF_Q4 model using the steps here: [query] mllm-quantizer usage help #586
3. Run on Qualcomm snapdragon mobile.
   PD2505:/data/local/tmp/opencl-package-latest $ ./mllm-tiny-llama-runner
   [INFO]mllm/mllm/backends/opencl/OpenCLBackend.cpp:71 Initializing OpenCL backend..
   [INFO] mllm/mllm/backends/opencl/OpenCLBackend.cpp:72 Device: QUALCOMM Adreno(TM) 840
   [INFO] mllm/backends/opencl/OpenCLBackend.cpp:73 GPUType: 1
   [INFO] mllm/latest-jan/mllm/mllm/backends/opencl/OpenCLBackend.cpp:74 Global Memory Cache Size: 1024 KB
   [INFO] mllm/mllm/backends/opencl/OpenCLBackend.cpp:75 Compute Units: 12
   [INFO] mllm/backends/opencl/OpenCLBackend.cpp:76 Max Work Group Size: 1024
   [INFO] mllm/backends/opencl/OpenCLBackend.cpp:77 FP16 Supported: Yes
   [INFO] mllm/backends/opencl/OpenCLBackend.cpp:78 Int8 dot Supported: No
   [INFO] mllm/backends/opencl/OpenCLBackend.cpp:79 Int8 dot Accumulate Supported: No
   completed initOpenCLBackend
   completed mllm::load tinyllama-1.1b-chat-q40.mllm
   completed llama.to(mllm::kOpenCL)
   Entering for loop
   <unk> <unk> <unk> <unk> <unk> <unk> <unk> <unk> <unk> <unk>
   Completed for loop
   Error: Received signal11 - SIGSEGV (Segmentation violation)
   Stack trace:
   #0 0x61b9777844
   better cmake configue #1 0x61b9777708
   feat:Add Topology Sort For NetParam #2 0x7e6306c874 __kernel_rt_sigreturn
   Feat: Adjust Weights Format. #3 0x7bbecbcbc8 gsl_memory_free_pure
   fix: ParamLoader Overflow. #4 0x7bbe3d4624
   Chaos:Add CMAKE_EXPORT_COMPILE_COMMANDS @lx200916 #5 0x7bbe3d9d5c
   feat: Add RMSNorm #6 0x7bbe3da060
   Silu #7 0x7bbe3d5b08 cb_release_mem_object
   feat: Add BPE Tokenizer. #8 0x7bcc669648 qCLDrvAPI_clReleaseMemObject
   feat: Add CPUEmbedding Express. #9 0x7e5d3771b0 mllm::opencl::OpenCLAllocator::~OpenCLAllocator()
   feat: Add SafeTensor Converter #10 0x7e5fb52f40 mllm::MemoryManager::~MemoryManager()
   nnapi env init #11 0x7e5fb4789c mllm::Context::~Context()
   Add Tests. #12 0x7e5d91b550 __cxa_finalize
   Setup CI, Unit Tests and ci tools. #13 0x7e5d92057c exit
   test: fix RoPETest. fix: optimize cos_value. #14 0x7e5d913dac
   Possible causes: invalid memory access, dangling pointer, stack overflow.
   Shutting down...

Expected Behavior

Display valid characters as response and exit graciously without crashing

Operating System

Android

Device

IQOO15

MLLM Framework Version

Latest

Model Information

No response

Additional Context

No response

[chenghuaWang]

cc @oreomaker