Segmentation fault crash: Tensorrt-LLM crash when using guided decoding xgrammar and kv cache reuse

[Somasundaram-Palaniappan]

System Info

• CPU Architecture x86_64
• GPU Nvidia A100
• Tensorrt-LLM 0.16.0

Who can help?

No response

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

Description:

• Tensorrt LLM crashes with the following dump when using xgrammar guided decoding and kv cache reuse together.
• if KV cache reuse is not used then no crash.

Steps to reproduce this issue

• Start with the guided decoding example @ TensorRT-LLM/examples/llm-api/llm_guided_decoding.py
• modify LLM object to enable kv-cache-reuse

import tensorrt_llm.bindings.executor as trtllm

kv_cache_config=trtllm.KvCacheConfig(enable_block_reuse=True)

llm = LLM(model="/trt_engines/a100/llama/tp1/compiled-model/", tokenizer="/trt_engines/a100/llama/tp1/tokenizer",
              guided_decoding_backend='xgrammar',kv_cache_config=kv_cache_config)

• Run the example code

Expected behavior

No crash

actual behavior

Segmentation fault crash.

Error

[baf3afa81c47:1647813] Signal: Segmentation fault (11)
[baf3afa81c47:1647813] Signal code: Address not mapped (1)
[baf3afa81c47:1647813] Failing at address: (nil)
[baf3afa81c47:1647813] [ 0] /usr/lib/x86_64-linux-gnu/libc.so.6(+0x42520)[0x7f8bcffdf520]
[baf3afa81c47:1647813] [ 1] /tensorrt-0.16/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(_ZN8xgrammar14GrammarMatcher20FillNextTokenBitmaskEP8DLTensori+0x0)[0x7f89d845f220]
[baf3afa81c47:1647813] [ 2] /tensorrt-0.16/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(_ZN12tensorrt_llm13batch_manager13GuidedDecoder5buildERKNS0_17ScheduledRequestsE+0x296)[0x7f89d8391fb6]
[baf3afa81c47:1647813] [ 3] /tensorrt-0.16/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(_ZN12tensorrt_llm13batch_manager27TrtGptModelInflightBatching12forwardAsyncERKSt4listISt10shared_ptrINS0_10LlmRequestEESaIS5_EE+0x6bf)[0x7f89d83f929f]
[baf3afa81c47:1647813] [ 4] /tensorrt-0.16/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(_ZN12tensorrt_llm8executor8Executor4Impl12forwardAsyncERSt4listISt10shared_ptrINS_13batch_manager10LlmRequestEESaIS7_EE+0x1e6)[0x7f89d848e746]
[baf3afa81c47:1647813] [ 5] /tensorrt-0.16/lib/python3.10/site-packages/tensorrt_llm/libs/libtensorrt_llm.so(_ZN12tensorrt_llm8executor8Executor4Impl13executionLoopEv+0x501)[0x7f89d8495331]
[baf3afa81c47:1647813] [ 6] /tensorrt-0.16/lib/python3.10/site-packages/torch/lib/libtorch.so(+0x145c0)[0x7f8bc80ad5c0]
[baf3afa81c47:1647813] [ 7] /usr/lib/x86_64-linux-gnu/libc.so.6(+0x94ac3)[0x7f8bd0031ac3]
[baf3afa81c47:1647813] [ 8] /usr/lib/x86_64-linux-gnu/libc.so.6(clone+0x44)[0x7f8bd00c2bf4]

additional notes

Description:

• Tensorrt LLM crashes with the following dump when using xgrammar guided decoding and kv cache reuse together.
• if KV cache reuse is not used then no crash.

[nv-guomingz]

Hi @syuoni would u please take a look this xgrammar related issue?

[syuoni]

Thanks for reporting this issue. We also found this issue internally, and it's currently being fixed.