search

triton-inference-server / tensorrtllm_backend

The Triton TensorRT-LLM Backend
Apache License 2.0
693 stars 103 forks source link

No white space included in tokens sent back by Llama2 in streaming mode #332

Open jfpichlme opened 9 months ago

jfpichlme commented 9 months ago

System Info

Who can help?

No response

Information

Tasks

Reproduction

1. Set up LLama2 (7b, 13b, 70b) in streaming mode:

model_config: 

name: "tensorrt_llm"
backend: "tensorrtllm"
max_batch_size: 300

model_transaction_policy {
  decoupled: True
}

dynamic_batching {
}

input [
  {
    name: "input_ids"
    data_type: TYPE_INT32
    dims: [ -1 ]
    allow_ragged_batch: true
  },
  {
    name: "input_lengths"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
  },
  {
    name: "request_output_len"
    data_type: TYPE_INT32
    dims: [ 1 ]
  },
  {
    name: "draft_input_ids"
    data_type: TYPE_INT32
    dims: [ -1 ]
    optional: true
    allow_ragged_batch: true
  },
  {
    name: "end_id"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "pad_id"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "stop_words_list"
    data_type: TYPE_INT32
    dims: [ 2, -1 ]
    optional: true
    allow_ragged_batch: true
  },
  {
    name: "bad_words_list"
    data_type: TYPE_INT32
    dims: [ 2, -1 ]
    optional: true
    allow_ragged_batch: true
  },
  {
    name: "embedding_bias"
    data_type: TYPE_FP32
    dims: [ -1 ]
    optional: true
    allow_ragged_batch: true
  },
  {
    name: "beam_width"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "temperature"
    data_type: TYPE_FP32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "runtime_top_k"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "runtime_top_p"
    data_type: TYPE_FP32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "len_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "repetition_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "min_length"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "presence_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "random_seed"
    data_type: TYPE_UINT64
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "return_log_probs"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "stop"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    optional: true
  },
  {
    name: "streaming"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    optional: true
  },
  {
    name: "prompt_embedding_table"
    data_type: TYPE_FP16
    dims: [ -1, -1 ]
    optional: true
    allow_ragged_batch: true
  },
  {
    name: "prompt_vocab_size"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  }
]
output [
  {
    name: "output_ids"
    data_type: TYPE_INT32
    dims: [ -1, -1 ]
  },
  {
    name: "sequence_length"
    data_type: TYPE_INT32
    dims: [ -1 ]
  },
  {
    name: "cum_log_probs"
    data_type: TYPE_FP32
    dims: [ -1 ]
  },
  {
    name: "output_log_probs"
    data_type: TYPE_FP32
    dims: [ -1, -1 ]
  }
]
instance_group [
  {
    count: 1
    kind : KIND_CPU
  }
]
parameters: {
  key: "max_beam_width"
  value: {
    string_value: "1"
  }
}
parameters: {
  key: "FORCE_CPU_ONLY_INPUT_TENSORS"
  value: {
    string_value: "no"
  }
}
parameters: {
  key: "gpt_model_type"
  value: {
    string_value: "inflight_fused_batching"
  }
}
parameters: {
  key: "gpt_model_path"
  value: {
    string_value: "/Llama2/03_Model_Dir_RT_070224/03_Model_Dir/01_Llama2_70b_TP8_300_STR/tensorrt_llm/1"
  }
}
parameters: {
  key: "max_tokens_in_paged_kv_cache"
  value: {
    string_value: "180000"
  }
}
parameters: {
  key: "max_attention_window_size"
  value: {
    string_value: "max_sequence_length"
  }
}
parameters: {
  key: "batch_scheduler_policy"
  value: {
    string_value: "max_utilization"
  }
}
parameters: {
  key: "max_num_sequences"
  value: {
    string_value: "1000"
  }
}
parameters: {
  key: "enable_trt_overlap"
  value: {
    string_value: "True"
  }
}
parameters: {
  key: "exclude_input_in_output"
  value: {
    string_value: "True"
  }
}
parameters: {
  key: "enable_kv_cache_reuse"
  value: {
    string_value: "False"
  }
}
parameters: {
  key: "normalize_log_probs"
  value: {
    string_value: "True"
  }
}

preprocessing: 


name: "preprocessing"
backend: "python"
max_batch_size: 300
input [
    {
        name: "QUERY"
        data_type: TYPE_STRING
        dims: [ -1 ]
    },
    {
        name: "REQUEST_OUTPUT_LEN"
        data_type: TYPE_INT32
        dims: [ -1 ]
    },
    {
        name: "BAD_WORDS_DICT"
        data_type: TYPE_STRING
        dims: [ -1 ]
        optional: true
    },
    {
        name: "STOP_WORDS_DICT"
        data_type: TYPE_STRING
        dims: [ -1 ]
        optional: true
    },
    {
        name: "EMBEDDING_BIAS_WORDS"
        data_type: TYPE_STRING
        dims: [ -1 ]
        optional: true
    },
    {
        name: "EMBEDDING_BIAS_WEIGHTS"
        data_type: TYPE_FP32
        dims: [ -1 ]
        optional: true
    }
]
output [
    {
        name: "INPUT_ID"
        data_type: TYPE_INT32
        dims: [ -1 ]
    },
    {
        name: "REQUEST_INPUT_LEN"
        data_type: TYPE_INT32
        dims: [ 1 ]
    },
    {
        name: "BAD_WORDS_IDS"
        data_type: TYPE_INT32
        dims: [ 2, -1 ]
    },
    {
        name: "STOP_WORDS_IDS"
        data_type: TYPE_INT32
        dims: [ 2, -1 ]
    },
    {
        name: "EMBEDDING_BIAS"
        data_type: TYPE_FP32
        dims: [ -1 ]
    },
    {
        name: "REQUEST_OUTPUT_LEN"
        data_type: TYPE_INT32
        dims: [ -1 ]
    }
]

parameters {
  key: "tokenizer_dir"
  value: {
    string_value: "/Llama2/01_HF_Model_Folder/01_LLama_70B/Llama-2-70b-chat-hf"
  }
}

parameters {
  key: "tokenizer_type"
  value: {
    string_value: "llama"
  }
}

parameters {
  key: "add_special_tokens"
  value: {
    string_value: "False"
  }
}

instance_group [
    {
        count: 1
        kind: KIND_CPU
    }
]

postprocessing: 


name: "postprocessing"
backend: "python"
max_batch_size: 300
input [
  {
    name: "TOKENS_BATCH"
    data_type: TYPE_INT32
    dims: [ -1, -1 ]
  },
  {
    name: "SEQUENCE_LENGTH"
    data_type: TYPE_INT32
    dims: [ -1 ]
  },
  {
    name: "CUM_LOG_PROBS"
    data_type: TYPE_FP32
    dims: [ -1 ]
  },
  {
    name: "OUTPUT_LOG_PROBS"
    data_type: TYPE_FP32
    dims: [ -1, -1 ]
  }
]
output [
  {
    name: "OUTPUT"
    data_type: TYPE_STRING
    dims: [ -1 ]
  },
  {
    name: "OUT_CUM_LOG_PROBS"
    data_type: TYPE_FP32
    dims: [ -1 ]
  },
  {
    name: "OUT_OUTPUT_LOG_PROBS"
    data_type: TYPE_FP32
    dims: [ -1, -1 ]
  }
]

parameters {
  key: "tokenizer_dir"
  value: {
    string_value: "/Llama2/01_HF_Model_Folder/01_LLama_70B/Llama-2-70b-chat-hf"
  }
}

parameters {
  key: "tokenizer_type"
  value: {
    string_value: "llama"
  }
}

parameters {
  key: "skip_special_tokens"
  value: {
    string_value: "False"
  }
}

instance_group [
    {
        count: 1
        kind: KIND_CPU
    }
]

ensemble: 

name: "ensemble"
platform: "ensemble"
max_batch_size: 300
input [
  {
    name: "text_input"
    data_type: TYPE_STRING
    dims: [ -1 ]
  },
  {
    name: "max_tokens"
    data_type: TYPE_INT32
    dims: [ -1 ]
  },
  {
   name: "bad_words"
   data_type: TYPE_STRING
   dims: [ -1 ]
   optional: true
  },
  {
   name: "stop_words"
   data_type: TYPE_STRING
   dims: [ -1 ]
   optional: true
  },
  {
    name: "end_id"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "pad_id"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "top_k"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "top_p"
    data_type: TYPE_FP32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "temperature"
    data_type: TYPE_FP32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "length_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "repetition_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "min_length"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "presence_penalty"
    data_type: TYPE_FP32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "random_seed"
    data_type: TYPE_UINT64
    dims: [ 1 ]
    optional: true
  },
  {
    name: "return_log_probs"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    optional: true
  },
  {
    name: "beam_width"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
    name: "stream"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    optional: true
  },
  {
    name: "prompt_embedding_table"
    data_type: TYPE_FP16
    dims: [ -1, -1 ]
    optional: true
  },
  {
    name: "prompt_vocab_size"
    data_type: TYPE_INT32
    dims: [ 1 ]
    optional: true
  },
  {
      name: "embedding_bias_words"
      data_type: TYPE_STRING
      dims: [ -1 ]
      optional: true
  },
  {
      name: "embedding_bias_weights"
      data_type: TYPE_FP32
      dims: [ -1 ]
      optional: true
  }
]
output [
  {
    name: "text_output"
    data_type: TYPE_STRING
    dims: [ -1 ]
  },
  {
    name: "cum_log_probs"
    data_type: TYPE_FP32
    dims: [ -1 ]
  },
  {
    name: "output_log_probs"
    data_type: TYPE_FP32
    dims: [ -1, -1 ]
  }
]
ensemble_scheduling {
  step [
    {
      model_name: "preprocessing"
      model_version: -1
      input_map {
        key: "QUERY"
        value: "text_input"
      }
      input_map {
        key: "REQUEST_OUTPUT_LEN"
        value: "max_tokens"
      }
      input_map {
        key: "BAD_WORDS_DICT"
        value: "bad_words"
      }
      input_map {
        key: "STOP_WORDS_DICT"
        value: "stop_words"
      }
      input_map {
        key: "EMBEDDING_BIAS_WORDS"
        value: "embedding_bias_words"
      }
      input_map {
        key: "EMBEDDING_BIAS_WEIGHTS"
        value: "embedding_bias_weights"
      }
      output_map {
        key: "REQUEST_INPUT_LEN"
        value: "_REQUEST_INPUT_LEN"
      }
      output_map {
        key: "INPUT_ID"
        value: "_INPUT_ID"
      }
      output_map {
        key: "REQUEST_OUTPUT_LEN"
        value: "_REQUEST_OUTPUT_LEN"
      }
      output_map {
        key: "STOP_WORDS_IDS"
        value: "_STOP_WORDS_IDS"
      }
      output_map {
        key: "BAD_WORDS_IDS"
        value: "_BAD_WORDS_IDS"
      }
      output_map {
        key: "EMBEDDING_BIAS"
        value: "_EMBEDDING_BIAS"
      }
    },
    {
      model_name: "tensorrt_llm"
      model_version: -1
      input_map {
        key: "input_ids"
        value: "_INPUT_ID"
      }
      input_map {
        key: "input_lengths"
        value: "_REQUEST_INPUT_LEN"
      }
      input_map {
        key: "request_output_len"
        value: "_REQUEST_OUTPUT_LEN"
      }
      input_map {
          key: "end_id"
          value: "end_id"
      }
      input_map {
          key: "pad_id"
          value: "pad_id"
      }
      input_map {
          key: "embedding_bias"
          value: "_EMBEDDING_BIAS"
      }
      input_map {
          key: "runtime_top_k"
          value: "top_k"
      }
      input_map {
          key: "runtime_top_p"
          value: "top_p"
      }
      input_map {
          key: "temperature"
          value: "temperature"
      }
      input_map {
          key: "len_penalty"
          value: "length_penalty"
      }
      input_map {
          key: "repetition_penalty"
          value: "repetition_penalty"
      }
      input_map {
          key: "min_length"
          value: "min_length"
      }
      input_map {
          key: "presence_penalty"
          value: "presence_penalty"
      }
      input_map {
          key: "random_seed"
          value: "random_seed"
      }
      input_map {
          key: "return_log_probs"
          value: "return_log_probs"
      }
      input_map {
          key: "beam_width"
          value: "beam_width"
      }
      input_map {
          key: "streaming"
          value: "stream"
      }
      input_map {
        key: "prompt_embedding_table"
        value: "prompt_embedding_table"
      }
      input_map {
        key: "prompt_vocab_size"
        value: "prompt_vocab_size"
      }
      input_map {
        key: "stop_words_list"
        value: "_STOP_WORDS_IDS"
      }
      input_map {
        key: "bad_words_list"
        value: "_BAD_WORDS_IDS"
      }
      output_map {
        key: "output_ids"
        value: "_TOKENS_BATCH"
      }
      output_map {
        key: "sequence_length"
        value: "_SEQUENCE_LENGTH"
      },
      output_map {
        key: "cum_log_probs"
        value: "_CUM_LOG_PROBS"
      }
      output_map {
        key: "output_log_probs"
        value: "_OUTPUT_LOG_PROBS"
      }
    },
    {
      model_name: "postprocessing"
      model_version: -1
      input_map {
        key: "TOKENS_BATCH"
        value: "_TOKENS_BATCH"
      }
      input_map {
        key: "CUM_LOG_PROBS"
        value: "_CUM_LOG_PROBS"
      }
      input_map {
        key: "OUTPUT_LOG_PROBS"
        value: "_OUTPUT_LOG_PROBS"
      }
      input_map {
        key: "SEQUENCE_LENGTH"
        value: "_SEQUENCE_LENGTH"
      }
      output_map {
        key: "OUTPUT"
        value: "text_output"
      }
      output_map {
        key: "OUT_OUTPUT_LOG_PROBS"
        value: "output_log_probs"
      }
      output_map {
        key: "OUT_CUM_LOG_PROBS"
        value: "cum_log_probs"
      }
    }
  ]
}

2. Use Nvidia client notebook (Install does not work, but downloading langchain_nvidia_trt.llms directly solves the problem)

https://github.com/NVIDIA/GenerativeAIExamples/blob/main/notebooks/01-llm-streaming-client.ipynb

(I have also written my own grpc client which produces the same output)

3. Send inference request via grpc to the triton

Expected behavior

Produce output tokens including whitespace:

The fastest land animal is the cheetah, which can run up to 70 miles per hour(1

actual behavior

Triton produces output tokens without whitespace:

Thefastestlandanimalisthecheetah,whichcanrunupto70milesperhour(1

additional notes

I am not too sure if this is a bug or that I am missing some flag. Any help is highly appreciated

Model build: 

python convert_checkpoint.py --model_dir /Llama2/ \
                              --output_dir /Llama2/03_TensorRT_0102_Model_Dir/01_LLama_7B_TP1/01_Converted_Weights/  \
                              --dtype float16 \
                  --tp_size 8
trtllm-build --checkpoint_dir /Llama2/03_TensorRT_0102_Model_Dir/01_LLama_7B_TP1/01_Converted_Weights/ \
             --output_dir /Llama2/03_TensorRT_0102_Model_Dir/01_Engine_Dir/01_LLama_7B_TP1/02_Build_Engines/ \
             --gpt_attention_plugin float16 \
             --gemm_plugin float16 \
             --remove_input_padding enable \
             --paged_kv_cache enable \
             --enable_xqa enable \
         --paged_kv_cache enable \
             --max_batch_size 300
philross commented 8 months ago

We are experiencing the same issue.

jfpichlme commented 8 months ago

For now I am using a workaround that is probably not ideal. In the postprocessing script (/postprocessing/1/model.py) I changed the _postprocessing function to return the actual token ids. 

    def _postprocessing(self, tokens_batch, sequence_lengths):
        outputs = []
        for batch_idx, beam_tokens in enumerate(tokens_batch):
            for beam_idx, tokens in enumerate(beam_tokens):
                seq_len = sequence_lengths[batch_idx][beam_idx]
                output = tokens[:seq_len]
                outputs.append(output)
        return outputs

I collect all the Token Ids on the User side and then decode the entire sequence which produces the correct output.

enochlev commented 8 months ago

The tokenizers in transoformers do not support this function automatically when calling decode function

The standard way of going about this is holding tokens in cache until a space is detected, in which everything after the space is put again into cache.

The other suggested method decodes the tokenid text instead of the string text to look for a "" symbol

here is a work around with text using the second method

    def _postprocessing(self, tokens_batch, sequence_lengths):
        outputs = []
        for batch_idx, beam_tokens in enumerate(tokens_batch):
            for beam_idx, tokens in enumerate(beam_tokens):
                seq_len = sequence_lengths[batch_idx][beam_idx]
                output = self.tokenizer.decode(
                    tokens[:seq_len],
                    skip_special_tokens=False)

                # for streamming mode
                token_id_string = self.tokenizer.convert_ids_to_tokens(tokens[:seq_len],skip_special_tokens=True)[0]
                if token_id_string[0] == "▁":
                    output = " " + output

                outputs.append(output.encode('utf8'))
        return str(output)
enochlev commented 8 months ago

@Shixiaowei02 I can create a PR for this

byshiue commented 8 months ago

Have you tried the tensorrt_llm_bls module?

enochlev commented 8 months ago

btw @jfpichlme how did you get tensorrt-LLM working witht he new workflow specivly with the trtllm-build command? which docker command and version of tensortllm_backend did you use?

jfpichlme commented 8 months ago

Hi enochlev,

Build the container

I have used Option 2 in the tensorrt-llm backend repo to build the docker container: 

# Update the submodules
cd tensorrtllm_backend
git lfs install
git submodule update --init --recursive

# Use the Dockerfile to build the backend in a container
# For x86_64
DOCKER_BUILDKIT=1 docker build -t triton_trt_llm -f dockerfile/Dockerfile.trt_llm_backend .

The docker version is: "22.04" and the tensorrt_llm git version is TensorRT-LLM backend (#324).

Build the Models

This process now consists of two steps, first a covert_checkpoint step. Then a build step.

  1. Perform the Conversion step The conversion step is done via the following command:

    python convert_checkpoint.py --model_dir $Enter hugginface format model dir \
                          --output_dir ./tllm_checkpoint_1gpu_fp16 \
                          --dtype float16 \
              --tp_size 8

  2. Perform the Build step: 

trtllm-build --checkpoint_dir ./tllm_checkpoint_1gpu_fp16 \
             --output_dir $Put your directory here  \
             --gpt_attention_plugin float16 \
             --gemm_plugin float16 \
             --remove_input_padding enable \
             --paged_kv_cache enable \
             --enable_xqa enable \
         --paged_kv_cache enable
             --max_batch_size 300

Combine your Model:

Last step is to copy the created engine files to the tensorrt_llm/1/ directory and adapt the config files. You can see the configs of the model in my initial comment.

I hope this helps you. @byshiue I will test the tensorrt_llm_bls module now.

ekarmazin commented 8 months ago

@jfpichlme any luck with bls + streaming? I have the same problem and for some reason can't make my grpc client to work with bls.

jfpichlme commented 8 months ago

Hi ekarmazin, bls + streaming did not work for me. At the moment I am sticking to the proposed solution where I modify the postprocessing script and buffer the token on the user side. To sort of augment streaming (displaying word by word output), I am decoding around 6-10 Token at a time. However, this does not work perfectly all the time.

ekarmazin commented 8 months ago

@jfpichlme I kind of got it working with BLS, it does proper output with whitespaces now. But I faced an accuracy problems with enabling --use_paged_context_fmha but that is a different issue.

ekarmazin commented 7 months ago

@byshiue same issue with bls model. Spaces are presented when accumulate tokens are true, and missing when false.

schetlur-nv commented 7 months ago

@enochlev apologies for the delayed response. Would you still be able to PR the fix you suggested?

npuichigo commented 7 months ago

Any update on this?

enochlev commented 7 months ago

I will find some time around work this week and push an update

plt12138 commented 7 months ago

mark

quwu0820 commented 7 months ago

Mark

charllll commented 6 months ago

Any update？

Saigut commented 6 months ago

Mark

HermitSun commented 6 months ago

Mark

yxchia98 commented 6 months ago

Mark

elinx commented 6 months ago

The tokenizers in transoformers do not support this function automatically when calling decode function

The standard way of going about this is holding tokens in cache until a space is detected, in which everything after the space is put again into cache.

The other suggested method decodes the tokenid text instead of the string text to look for a "" symbol

here is a work around with text using the second method

    def _postprocessing(self, tokens_batch, sequence_lengths):
        outputs = []
        for batch_idx, beam_tokens in enumerate(tokens_batch):
            for beam_idx, tokens in enumerate(beam_tokens):
                seq_len = sequence_lengths[batch_idx][beam_idx]
                output = self.tokenizer.decode(
                    tokens[:seq_len],
                    skip_special_tokens=False)

                # for streamming mode
                token_id_string = self.tokenizer.convert_ids_to_tokens(tokens[:seq_len],skip_special_tokens=True)[0]
                if token_id_string[0] == "▁":
                    output = " " + output

                outputs.append(output.encode('utf8'))
        return str(output)

@enochlev crash if the last token is EOS, a quick fix:

                token_id_string = self.tokenizer.convert_ids_to_tokens(tokens[:seq_len], skip_special_tokens=True)
                if len(token_id_string) > 0 and len(token_id_string[0]) > 0 and token_id_string[0][0] == "▁":
                    output = " " + output
enochlev commented 6 months ago

@elinx Really appreciate catching that...

I just submitted a PR including your suggestion. It worked it my local environment before a submitted the PR, so it has my approval (if that means anything 😁)

Link to PR