search

NVIDIA / TensorRT-LLM

TensorRT-LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and build TensorRT engines that contain state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT-LLM also contains components to create Python and C++ runtimes that execute those TensorRT engines.
https://nvidia.github.io/TensorRT-LLM
Apache License 2.0
8.74k stars 1k forks source link

Assertion failed: Must set crossKvCacheFraction for encoder-decoder model #2419

Open Saeedmatt3r opened 2 weeks ago

Saeedmatt3r commented 2 weeks ago

System Info

GPU: A10 Base Image: FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 Tensorrt-llm:

Who can help?

No response

Information

Tasks

Reproduction

Steps to reproduce the problem:

FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04

RUN apt-get update && apt-get -y install python3.10 python3-pip openmpi-bin libopenmpi-dev

RUN pip3 install tensorrt_llm==0.14.0 -U --pre --extra-index-url https://pypi.nvidia.com

then by running the official whisper example:

INFERENCE_PRECISION=float16
WEIGHT_ONLY_PRECISION=int8
MAX_BEAM_WIDTH=4
MAX_BATCH_SIZE=8
checkpoint_dir=weights/whisper_large_v3_weights_${WEIGHT_ONLY_PRECISION}
output_dir=weights/whisper_large_v3_${WEIGHT_ONLY_PRECISION}

# Convert the large-v3 model weights into TensorRT-LLM format.
python3 convert_checkpoint.py \
                --use_weight_only \
                --weight_only_precision $WEIGHT_ONLY_PRECISION \
                --output_dir $checkpoint_dir

# Build the large-v3 model using trtllm-build
trtllm-build  --checkpoint_dir ${checkpoint_dir}/encoder \
              --output_dir ${output_dir}/encoder \
              --moe_plugin disable \
              --enable_xqa disable \
              --max_batch_size ${MAX_BATCH_SIZE} \
              --gemm_plugin disable \
              --bert_attention_plugin ${INFERENCE_PRECISION} \
              --max_input_len 3000 --max_seq_len=3000

trtllm-build  --checkpoint_dir ${checkpoint_dir}/decoder \
              --output_dir ${output_dir}/decoder \
              --moe_plugin disable \
              --enable_xqa disable \
              --max_beam_width ${MAX_BEAM_WIDTH} \
              --max_batch_size ${MAX_BATCH_SIZE} \
              --max_seq_len 114 \
              --max_input_len 14 \
              --max_encoder_input_len 3000 \
              --gemm_plugin ${INFERENCE_PRECISION} \
              --bert_attention_plugin ${INFERENCE_PRECISION} \
              --gpt_attention_plugin ${INFERENCE_PRECISION}

python3 run.py --engine_dir "$output_dir" --dataset hf-internal-testing/librispeech_asr_dummy --num_beams $MAX_BEAM_WIDTH --batch_size $MAX_BATCH_SIZE --enable_warmup --name librispeech_dummy_large_v3

Expected behavior

It should run on the dataset without any problem:

actual behavior

[TensorRT-LLM] TensorRT-LLM version: 0.14.0
[TensorRT-LLM][INFO] Engine version 0.14.0 found in the config file, assuming engine(s) built by new builder API.
[TensorRT-LLM][INFO] Setting encoder max input length and hidden size for accepting visual features.
[TensorRT-LLM][INFO] Engine version 0.14.0 found in the config file, assuming engine(s) built by new builder API.
[TensorRT-LLM][INFO] Engine version 0.14.0 found in the config file, assuming engine(s) built by new builder API.
[TensorRT-LLM][INFO] Setting encoder max input length and hidden size for accepting visual features.
[TensorRT-LLM][INFO] MPI size: 1, MPI local size: 1, rank: 0
[TensorRT-LLM][INFO] Engine version 0.14.0 found in the config file, assuming engine(s) built by new builder API.
[TensorRT-LLM][INFO] Setting encoder max input length and hidden size for accepting visual features.
[TensorRT-LLM][INFO] Refreshed the MPI local session
[TensorRT-LLM][INFO] Engine version 0.14.0 found in the config file, assuming engine(s) built by new builder API.
[TensorRT-LLM][INFO] MPI size: 1, MPI local size: 1, rank: 0
[TensorRT-LLM][INFO] TRTGptModel maxNumSequences: 8
[TensorRT-LLM][INFO] TRTGptModel maxBatchSize: 8
[TensorRT-LLM][INFO] TRTGptModel maxBeamWidth: 1
[TensorRT-LLM][INFO] TRTGptModel maxSequenceLen: 3000
[TensorRT-LLM][INFO] TRTGptModel maxDraftLen: 0
[TensorRT-LLM][INFO] TRTGptModel mMaxAttentionWindowSize: (3000) * 32
[TensorRT-LLM][INFO] TRTGptModel enableTrtOverlap: 0
[TensorRT-LLM][INFO] TRTGptModel normalizeLogProbs: 1
[TensorRT-LLM][INFO] TRTGptModel maxNumTokens: 8192
[TensorRT-LLM][INFO] TRTGptModel maxInputLen: 2999 = min(maxSequenceLen - 1, maxNumTokens) since context FMHA and usePackedInput are enabled
[TensorRT-LLM][INFO] TRTGptModel If model type is encoder, maxInputLen would be reset in trtEncoderModel to maxInputLen: min(maxSequenceLen, maxNumTokens).
[TensorRT-LLM][INFO] Capacity Scheduler Policy: GUARANTEED_NO_EVICT
[TensorRT-LLM][INFO] Context Chunking Scheduler Policy: None
[TensorRT-LLM][INFO] Loaded engine size: 622 MiB
[TensorRT-LLM][INFO] [MemUsageChange] Allocated 292.97 MiB for execution context memory.
[TensorRT-LLM][INFO] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 615 (MiB)
[TensorRT-LLM][INFO] TRTEncoderModel mMaxInputLen: reset to 3000 from build config.
[TensorRT-LLM][INFO] MPI size: 1, MPI local size: 1, rank: 0
[TensorRT-LLM][INFO] Rank 0 is using GPU 0
[TensorRT-LLM][INFO] TRTGptModel maxNumSequences: 8
[TensorRT-LLM][INFO] TRTGptModel maxBatchSize: 8
[TensorRT-LLM][INFO] TRTGptModel maxBeamWidth: 4
[TensorRT-LLM][INFO] TRTGptModel maxSequenceLen: 114
[TensorRT-LLM][INFO] TRTGptModel maxDraftLen: 0
[TensorRT-LLM][INFO] TRTGptModel mMaxAttentionWindowSize: (114) * 32
[TensorRT-LLM][INFO] TRTGptModel enableTrtOverlap: 0
[TensorRT-LLM][INFO] TRTGptModel normalizeLogProbs: 1
[TensorRT-LLM][INFO] TRTGptModel maxNumTokens: 912
[TensorRT-LLM][INFO] TRTGptModel maxInputLen: 113 = min(maxSequenceLen - 1, maxNumTokens) since context FMHA and usePackedInput are enabled
[TensorRT-LLM][INFO] TRTGptModel If model type is encoder, maxInputLen would be reset in trtEncoderModel to maxInputLen: min(maxSequenceLen, maxNumTokens).
[TensorRT-LLM][INFO] Capacity Scheduler Policy: GUARANTEED_NO_EVICT
[TensorRT-LLM][INFO] Context Chunking Scheduler Policy: None
[TensorRT-LLM][INFO] The logger passed into createInferRuntime differs from one already provided for an existing builder, runtime, or refitter. Uses of the global logger, returned by nvinfer1::getLogger(), will return the existing value.
[TensorRT-LLM][INFO] Loaded engine size: 1066 MiB
[TensorRT-LLM][INFO] [MemUsageChange] Allocated 833.73 MiB for execution context memory.
[TensorRT-LLM][INFO] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 1672 (MiB)
[TensorRT-LLM][INFO] [MemUsageChange] Allocated 3.54 MB GPU memory for runtime buffers.
[TensorRT-LLM][INFO] [MemUsageChange] Allocated 16.82 MB GPU memory for decoder.
[TensorRT-LLM][INFO] Memory usage when calculating max tokens in paged kv cache: total: 21.98 GiB, available: 18.88 GiB
[TensorRT-LLM][INFO] Number of blocks in KV cache primary pool: 1740
[TensorRT-LLM][INFO] Number of blocks in KV cache secondary pool: 0, onboard blocks to primary memory before reuse: true
Traceback (most recent call last):
  File "/app/TensorRT-LLM/examples/whisper/run.py", line 479, in <module>
    model = WhisperTRTLLM(args.engine_dir, args.debug, args.assets_dir,
  File "/app/TensorRT-LLM/examples/whisper/run.py", line 327, in __init__
    self.model_runner_cpp = ModelRunnerCpp.from_dir(**runner_kwargs)
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/runtime/model_runner_cpp.py", line 206, in from_dir
    executor = trtllm.Executor(
RuntimeError: [TensorRT-LLM][ERROR] Assertion failed: Must set crossKvCacheFraction for encoder-decoder model (/home/jenkins/agent/workspace/LLM/release-0.14/L0_PostMerge/llm/cpp/tensorrt_llm/batch_manager/trtGptModelInflightBatching.cpp:207)
1       0x7493ec8cccc7 tensorrt_llm::common::throwRuntimeError(char const*, int, std::string const&) + 82
2       0x7493ec914e4a /usr/local/lib/python3.10/dist-packages/tensorrt_llm/libs/libtensorrt_llm.so(+0x769e4a) [0x7493ec914e4a]
3       0x7493eea3bac4 tensorrt_llm::batch_manager::TrtGptModelFactory::create(tensorrt_llm::runtime::RawEngine const&, tensorrt_llm::runtime::ModelConfig const&, tensorrt_llm::runtime::WorldConfig const&, tensorrt_llm::batch_manager::TrtGptModelType, tensorrt_llm::batch_manager::TrtGptModelOptionalParams const&) + 756
4       0x7493eeac1d7d tensorrt_llm::executor::Executor::Impl::createModel(tensorrt_llm::runtime::RawEngine const&, tensorrt_llm::runtime::ModelConfig const&, tensorrt_llm::runtime::WorldConfig const&, tensorrt_llm::executor::ExecutorConfig const&) + 125
5       0x7493eeac233a tensorrt_llm::executor::Executor::Impl::loadModel(std::optional<std::filesystem::path> const&, std::optional<std::basic_string_view<unsigned char, std::char_traits<unsigned char> > > const&, tensorrt_llm::runtime::GptJsonConfig const&, tensorrt_llm::executor::ExecutorConfig const&, bool, std::optional<std::map<std::string, tensorrt_llm::executor::Tensor, std::less<std::string>, std::allocator<std::pair<std::string const, tensorrt_llm::executor::Tensor> > > > const&) + 954
6       0x7493eeac3377 tensorrt_llm::executor::Executor::Impl::Impl(std::filesystem::path const&, std::optional<std::filesystem::path> const&, tensorrt_llm::executor::ModelType, tensorrt_llm::executor::ExecutorConfig const&) + 2135
7       0x7493eeaaeb33 tensorrt_llm::executor::Executor::Executor(std::filesystem::path const&, std::filesystem::path const&, tensorrt_llm::executor::ModelType, tensorrt_llm::executor::ExecutorConfig const&) + 99
8       0x749518615609 /usr/local/lib/python3.10/dist-packages/tensorrt_llm/bindings.cpython-310-x86_64-linux-gnu.so(+0xdd609) [0x749518615609]
9       0x749518598a8f /usr/local/lib/python3.10/dist-packages/tensorrt_llm/bindings.cpython-310-x86_64-linux-gnu.so(+0x60a8f) [0x749518598a8f]
10      0x61696fec6b2e python3(+0x15cb2e) [0x61696fec6b2e]
11      0x61696febd2db _PyObject_MakeTpCall + 603
12      0x61696fed56b0 python3(+0x16b6b0) [0x61696fed56b0]
13      0x61696fed1ad7 python3(+0x167ad7) [0x61696fed1ad7]
14      0x61696febd68b python3(+0x15368b) [0x61696febd68b]
15      0x7495185928bb /usr/local/lib/python3.10/dist-packages/tensorrt_llm/bindings.cpython-310-x86_64-linux-gnu.so(+0x5a8bb) [0x7495185928bb]
16      0x61696febd2db _PyObject_MakeTpCall + 603
17      0x61696feb5d27 _PyEval_EvalFrameDefault + 27415
18      0x61696fed5281 python3(+0x16b281) [0x61696fed5281]
19      0x61696fed5f22 PyObject_Call + 290
20      0x61696feb1a6e _PyEval_EvalFrameDefault + 10334
21      0x61696febc474 _PyObject_FastCallDictTstate + 196
22      0x61696fed14b4 python3(+0x1674b4) [0x61696fed14b4]
23      0x61696febd27c _PyObject_MakeTpCall + 508
24      0x61696feb56e6 _PyEval_EvalFrameDefault + 25814
25      0x61696feac016 python3(+0x142016) [0x61696feac016]
26      0x61696ffa18b6 PyEval_EvalCode + 134
27      0x61696ffcc918 python3(+0x262918) [0x61696ffcc918]
28      0x61696ffc61db python3(+0x25c1db) [0x61696ffc61db]
29      0x61696ffcc665 python3(+0x262665) [0x61696ffcc665]
30      0x61696ffcbb48 _PyRun_SimpleFileObject + 424
31      0x61696ffcb793 _PyRun_AnyFileObject + 67
32      0x61696ffbe2ce Py_RunMain + 702
33      0x61696ff9470d Py_BytesMain + 45
34      0x7496408cfd90 /usr/lib/x86_64-linux-gnu/libc.so.6(+0x29d90) [0x7496408cfd90]
35      0x7496408cfe40 __libc_start_main + 128
36      0x61696ff94605 _start + 37

Using the latest available pip package(0.15.0.dev2024110500)

+ trtllm-build --checkpoint_dir weights/whisper_large_v3_weights_int8/encoder --output_dir weights/whisper_large_v3_int8/encoder --moe_plugin disable --enable_xqa disable --max_batch_size 8 --gemm_plugin disable --bert_attention_plugin float16 --max_input_len 3000 --max_seq_len=3000
[TensorRT-LLM] TensorRT-LLM version: 0.15.0.dev2024110500
[11/06/2024-12:28:13] [TRT-LLM] [I] Set bert_attention_plugin to float16.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set gpt_attention_plugin to auto.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set gemm_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set gemm_swiglu_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set fp8_rowwise_gemm_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set nccl_plugin to auto.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set lora_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set moe_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set mamba_conv1d_plugin to auto.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set low_latency_gemm_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set low_latency_gemm_swiglu_plugin to None.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set context_fmha to True.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set bert_context_fmha_fp32_acc to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set remove_input_padding to True.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set reduce_fusion to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set enable_xqa to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set tokens_per_block to 64.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set use_paged_context_fmha to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set use_fp8_context_fmha to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set multiple_profiles to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set paged_state to True.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set streamingllm to False.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set use_fused_mlp to True.
[11/06/2024-12:28:13] [TRT-LLM] [I] Set pp_reduce_scatter to False.
[11/06/2024-12:28:13] [TRT-LLM] [W] Implicitly setting PretrainedConfig.n_mels = 128
[11/06/2024-12:28:13] [TRT-LLM] [W] Implicitly setting PretrainedConfig.n_audio_ctx = 1500
[11/06/2024-12:28:13] [TRT-LLM] [W] Implicitly setting PretrainedConfig.num_languages = 100
Traceback (most recent call last):
  File "/usr/local/bin/trtllm-build", line 8, in <module>
    sys.exit(main())
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 602, in main
    parallel_build(model_config, ckpt_dir, build_config, args.output_dir,
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 425, in parallel_build
    passed = build_and_save(rank, rank % workers, ckpt_dir,
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 390, in build_and_save
    engine = build_model(build_config,
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 360, in build_model
    model = model_cls.from_checkpoint(ckpt_dir, config=rank_config)
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/modeling_utils.py", line 635, in from_checkpoint
    model = cls(config)
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/modeling_utils.py", line 564, in __call__
    obj = type.__call__(cls, *args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/enc_dec/model.py", line 1939, in __init__
    self.position_embedding = Embedding(self.config.max_position_embeddings,
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/layers/embedding.py", line 63, in __init__
    shape = (math.ceil(self.num_embeddings / self.tp_size),
TypeError: unsupported operand type(s) for /: 'NoneType' and 'int'

additional notes

Also I checked the trt-llm-backend for the whisper example and also that was not working, with the following error:

c_python_backend_utils.TritonModelException: [TensorRT-LLM][ERROR] Assertion failed: input tokens tensor not provided (/tmp/tritonbuild/tensorrtllm/inflight_batcher_llm/src/utils.cc:107)
yuekaizhang commented 2 weeks ago

@Saeedmatt3r Thanks for reporting the issue. The fix would be synced to github next week. For a quick fix, you need to modify here https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/whisper/run.py#L370.

Image

Saeedmatt3r commented 2 weeks ago

@yuekaizhang Thanks, to be honest, I've done that actually, but Just wanted to report the issues in 0.14 and 0.15, also I think the official whisper on trt-llm-backend is also not working as expected, I used 0.15 for engine creation and it was not working. I will try to create another ticket in the repo.