Error occured when running medusa inference.

[littletomatodonkey]

Hi, when i use medusa decoding on trtllm-090 which profiling, error occrued as follows. Could you please help to have a look? Thanks!

If i do not use --run_profiling, the inference process is normal.

  File "/opt/tiger/miniconda3/lib/python3.10/site-packages/tensorrt_llm/runtime/generation.py", line 2431, in handle_per_step
    self.accept_lengths).item()
RuntimeError: CUDA error: an illegal memory access was encountered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

[05/10/2024-21:45:33] [TRT] [E] 1: [graphContext.h::~MyelinGraphContext::55] Error Code 1: Myelin ([impl.cpp:cuda_object_deallocate:345] Error 700 destroying stream '0x56052ab0a0f0'.)
[05/10/2024-21:45:33] [TRT] [E] 1: [graphContext.h::~MyelinGraphContext::55] Error Code 1: Myelin ([impl.cpp:cuda_object_deallocate:345] Error 700 destroying stream '0x56052ab69e70'.)
[05/10/2024-21:45:33] [TRT] [E] 1: [graphContext.h::~MyelinGraphContext::55] Error Code 1: Myelin ([impl.cpp:cuda_object_deallocate:345] Error 700 destroying stream '0x56052abb95f0'.)

• conversion

model_dir="/mnt/bn/multimodel/models/medusa/vicuna-7b-v1.3"
medusa_model_dir="/mnt/bn/multimodel/models/medusa/medusa-vicuna-7b-v1.3"

tmp_dir=$(mktemp -d)

trt_model_dir="./output/medusa"

python convert_checkpoint.py \
--model_dir "${model_dir}" \
--medusa_model_dir "${medusa_model_dir}" \
--output_dir "${tmp_dir}" \
--dtype float16 \
--fixed_num_medusa_heads 4

trtllm-build \
--checkpoint_dir ${tmp_dir} \
--output_dir ${trt_model_dir} \
--gemm_plugin float16 \
--remove_input_padding "enable" \
--context_fmha "enable" \
--gemm_plugin="float16" \
--gpt_attention_plugin "float16" \
--max_batch_size 16 \
--max_input_len 4096 \
--max_output_len 1024 \
--paged_kv_cache enable \
--use_paged_context_fmha enable

cp -r ${model_dir}/*token* ${output_dir}/

• inference

trt_model_dir="./output/medusa"

python ../run.py \
--engine_dir ${trt_model_dir} \
--tokenizer_dir ${trt_model_dir} \
--max_output_len=100 \
--medusa_choices="[[0], [0, 0], [1], [0, 1], [2], [0, 0, 0], [1, 0], [0, 2], [3], [0, 3], [4], [0, 4], [2, 0], [0, 5], [0, 0, 1], [5], [0, 6], [6], [0, 7], [0, 1, 0], [1, 1], [7], [0, 8], [0, 0, 2], [3, 0], [0, 9], [8], [9], [1, 0, 0], [0, 2, 0], [1, 2], [0, 0, 3], [4, 0], [2, 1], [0, 0, 4], [0, 0, 5], [0, 0, 0, 0], [0, 1, 1], [0, 0, 6], [0, 3, 0], [5, 0], [1, 3], [0, 0, 7], [0, 0, 8], [0, 0, 9], [6, 0], [0, 4, 0], [1, 4], [7, 0], [0, 1, 2], [2, 0, 0], [3, 1], [2, 2], [8, 0], [0, 5, 0], [1, 5], [1, 0, 1], [0, 2, 1], [9, 0], [0, 6, 0], [0, 0, 0, 1], [1, 6], [0, 7, 0]]" \
--use_py_session \
--temperature 1.0 \
--input_text "Once upon" \
--run_profiling

[dongxuy04]

Hi, I tried with latest main and it seems OK, could you please try with that? Thanks! BTW, with latest main, C++ runtime can also be used by removing --use_py_session.

[nv-guomingz]

Hi @littletomatodonkey , do u still encouter such issue w/ @dongxuy04 's suggestion? If not, I'll close this ticket.