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
8.16k stars 902 forks source link

AssertionError: tensor WhisperEncoder/encoder_layers/0/attention_layernorm/layer_norm_L5155/NORMALIZATION_0_output_0 has an invalid shape #2129

Open xinliu9451 opened 3 weeks ago

xinliu9451 commented 3 weeks ago

System Info

colab T4

Who can help?

@

Information

Tasks

Reproduction

https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/whisper/README.md#distil-whisper

command： !trtllm-build --checkpoint_dir distil_whisper_medium_en_weights_int8/encoder \ --output_dir distil_whisper_medium_en_int8/encoder \ --paged_kv_cache disable \ --moe_plugin disable \ --enable_xqa disable \ --max_batch_size 8 \ --gemm_plugin disable \ --bert_attention_plugin float16 \ --remove_input_padding disable \ --max_input_len 1500

Expected behavior

Successful compilation of models

actual behavior

ITensor::getDimensions: Error Code 4: API Usage Error (WhisperEncoder/encoder_layers/0/attention_layernorm/layer_norm_L5155/NORMALIZATION_0: INormalizationLayer input and scale must have identical types. input type is Half and scale type is Float.)

additional notes

I got an error when I tried: https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/whisper/README.md#distil-whisper My command was: !trtllm-build --checkpoint_dir distil_whisper_medium_en_weights_int8/encoder --output_dir distil_whisper_medium_en_int8/encoder --paged_kv_cache disable --moe_plugin disable --enable_xqa disable ---max_batch_size 8 --gemm_plugin disable \ --bert_attention_plugin float16 --remove_input_padding disable ---max_input_len 1500

The error report that occurred reads: 2024-08-20 07:45:07.071785: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2024-08-20 07:45:07.092394: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2024-08-20 07:45:07.098718: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2024-08-20 07:45:07.113933: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. 2024-08-20 07:45:08.198287: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT [TensorRT-LLM] TensorRT-LLM version: 0.13.0.dev2024081300 [08/20/2024-07:45:09] [TRT-LLM] [W] Option --paged_kv_cache is deprecated, use --kv_cache_type=paged/disabled instead. [08/20/2024-07:45:09] [TRT-LLM] [I] Set bert_attention_plugin to auto. [08/20/2024-07:45:09] [TRT-LLM] [I] Set gpt_attention_plugin to auto. [08/20/2024-07:45:09] [TRT-LLM] [I] Set gemm_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set gemm_swiglu_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set fp8_rowwise_gemm_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set nccl_plugin to auto. [08/20/2024-07:45:09] [TRT-LLM] [I] Set lookup_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set lora_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set moe_plugin to None. [08/20/2024-07:45:09] [TRT-LLM] [I] Set mamba_conv1d_plugin to auto. [08/20/2024-07:45:09] [TRT-LLM] [I] Set context_fmha to True. [08/20/2024-07:45:09] [TRT-LLM] [I] Set bert_context_fmha_fp32_acc to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set remove_input_padding to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set reduce_fusion to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set enable_xqa to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set tokens_per_block to 64. [08/20/2024-07:45:09] [TRT-LLM] [I] Set use_paged_context_fmha to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set use_fp8_context_fmha to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set multiple_profiles to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set paged_state to True. [08/20/2024-07:45:09] [TRT-LLM] [I] Set streamingllm to False. [08/20/2024-07:45:09] [TRT-LLM] [I] Set paged_kv_cache to False. [08/20/2024-07:45:09] [TRT-LLM] [W] Implicitly setting PretrainedConfig.n_mels = 80 [08/20/2024-07:45:09] [TRT-LLM] [W] Implicitly setting PretrainedConfig.n_audio_ctx = 1500 [08/20/2024-07:45:09] [TRT-LLM] [W] Implicitly setting PretrainedConfig.num_languages = 99 [08/20/2024-07:45:09] [TRT-LLM] [I] Compute capability: (7, 5) [08/20/2024-07:45:09] [TRT-LLM] [I] SM count: 40 [08/20/2024-07:45:09] [TRT-LLM] [I] SM clock: 1590 MHz [08/20/2024-07:45:09] [TRT-LLM] [I] int4 TFLOPS: 260 [08/20/2024-07:45:09] [TRT-LLM] [I] int8 TFLOPS: 130 [08/20/2024-07:45:09] [TRT-LLM] [I] fp8 TFLOPS: 0 [08/20/2024-07:45:09] [TRT-LLM] [I] float16 TFLOPS: 65 [08/20/2024-07:45:09] [TRT-LLM] [I] bfloat16 TFLOPS: 0 [08/20/2024-07:45:09] [TRT-LLM] [I] float32 TFLOPS: 8 [08/20/2024-07:45:09] [TRT-LLM] [I] Total Memory: 15 GiB [08/20/2024-07:45:09] [TRT-LLM] [I] Memory clock: 5001 MHz [08/20/2024-07:45:09] [TRT-LLM] [I] Memory bus width: 256 [08/20/2024-07:45:09] [TRT-LLM] [I] Memory bandwidth: 320 GB/s [08/20/2024-07:45:09] [TRT-LLM] [I] PCIe speed: 2500 Mbps [08/20/2024-07:45:09] [TRT-LLM] [I] PCIe link width: 16 [08/20/2024-07:45:09] [TRT-LLM] [I] PCIe bandwidth: 5 GB/s [08/20/2024-07:45:09] [TRT-LLM] [W] Parameter was initialized as DataType.HALF but set to DataType.FLOAT [08/20/2024-07:45:09] [TRT-LLM] [W] Parameter was initialized as DataType.HALF but set to DataType.FLOAT [08/20/2024-07:45:09] [TRT-LLM] [W] Parameter was initialized as DataType.HALF but set to DataType.FLOAT [08/20/2024-07:45:09] [TRT-LLM] [W] Parameter was initialized as DataType.HALF but set to DataType.FLOAT [08/20/2024-07:45:09] [TRT-LLM] [I] Set dtype to float16. [08/20/2024-07:45:09] [TRT-LLM] [W] Overriding paged_state to False [08/20/2024-07:45:09] [TRT-LLM] [I] Set paged_state to False. [08/20/2024-07:45:09] [TRT-LLM] [I] max_seq_len is not specified, using deduced value 2048 [08/20/2024-07:45:09] [TRT-LLM] [W] remove_input_padding is not enabled, the specified max_num_tokens/opt_num_tokens will be ignored. [08/20/2024-07:45:09] [TRT] [I] [MemUsageChange] Init CUDA: CPU +13, GPU +0, now: CPU 213, GPU 103 (MiB) [08/20/2024-07:45:11] [TRT] [I] [MemUsageChange] Init builder kernel library: CPU +904, GPU +180, now: CPU 1272, GPU 283 (MiB) [08/20/2024-07:45:11] [TRT] [W] profileSharing0806 is on by default in TensorRT 10.0. This flag is deprecated and has no effect. [08/20/2024-07:45:11] [TRT-LLM] [I] Set weight_only_quant_matmul_plugin to float16. [08/20/2024-07:45:11] [TRT-LLM] [I] Set nccl_plugin to None. [08/20/2024-07:45:11] [TRT] [E] ITensor::getDimensions: Error Code 4: API Usage Error (WhisperEncoder/encoder_layers/0/attention_layernorm/layer_norm_L5155/NORMALIZATION_0: INormalizationLayer input and scale must have identical types. input type is Half and scale type is Float.) Traceback (most recent call last): File "/usr/local/bin/trtllm-build", line 8, in sys.exit(main()) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 528, 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 394, 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 361, in build_and_save engine = build_model(build_config, File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/commands/build.py", line 354, in build_model return build(model, build_config) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/builder.py", line 1101, in build model(inputs) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/module.py", line 40, in call output = self.forward(*args, *kwargs) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/enc_dec/model.py", line 1915, in forward hidden_states = encoder_layer(hidden_states, File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/module.py", line 40, in call output = self.forward(args, kwargs) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/enc_dec/model.py", line 243, in forward hidden_states = self.attention_layernorm(hidden_states) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/module.py", line 40, in call output = self.forward(*args, **kwargs) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/layers/normalization.py", line 49, in forward return layer_norm(x, self.normalized_shape, weight, bias, self.eps) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/functional.py", line 5155, in layer_norm return _create_tensor(layer.get_output(0), layer) File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/functional.py", line 607, in _create_tensor assert trt_tensor.shape.len( AssertionError: tensor WhisperEncoder/encoder_layers/0/attention_layernorm/layer_norm_L5155/NORMALIZATION_0_output_0 has an invalid shape

yuekaizhang commented 3 weeks ago

For distill-whisper, would you mind adding model=model.half() here https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/whisper/distil_whisper/convert_from_distil_whisper.py#L60 for now?

The code fix will be synced to github later. Thanks.