Significant latency between COMPUTE_END and REQUEST_END

[danielchalef]

Description
We're seeing significant latency in the order of 300-600 hundred milliseconds between COMPUTE_END and REQUEST_END on a TensorRT-LLM model. See OTEL trace image below.

Triton Information
What version of Triton are you using? 2.42.0

Are you using the Triton container or did you build it yourself? NGC Container 24.01

To Reproduce

• Mistral-7B compiled to a TensorRT-LLM engine on an A100 with 2x tensor parallel.
• TensorRT-LLM backend version 0.7.1 (as shipped in NGC Container)
• Deployed to a 2 GPU A100-80 server.
• Dynamic Batching
• Redis Response cache
• We use the grpc API

See the config.pbtxt below.

Expected behavior
REQUEST_END should occur very shortly after COMPUTE_END, perhaps on the order of tens of ms.

name: "XXXXX"
backend: "tensorrtllm"
max_batch_size: 8

version_policy: { latest: { num_versions: 1}}

model_transaction_policy {
  decoupled: false
}

dynamic_batching {
    max_queue_delay_microseconds: 25000
    priority_levels: 3
    default_priority_level: 2
}

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: "frequency_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: "return_context_logits"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "return_generation_logits"
    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 ]
  },
  {
    name: "context_logits"
    data_type: TYPE_FP32
    dims: [ -1, -1 ]
  },
  {
    name: "generation_logits"
    data_type: TYPE_FP32
    dims: [ -1, -1, -1 ]
  }
]

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: "/models/XXXXXXX/1"
  }
}
parameters: {
  key: "max_tokens_in_paged_kv_cache"
  value: {
    string_value: ""
  }
}
parameters: {
  key: "batch_scheduler_policy"
  value: {
    string_value: "guaranteed_no_evict"
  }
}
parameters: {
  key: "kv_cache_free_gpu_mem_fraction"
  value: {
    string_value: "0.9"
  }
}
parameters: {
  key: "max_num_sequences"
  value: {
    string_value: "16"
  }
}
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"
  }
}

instance_group [
  {
    count: 1
    kind : KIND_CPU
  }
]

response_cache{
  enable:true
}```

[fpetrini15]

Hi @danielchalef,

Is this unusual for other models you have tested? Is this latency value consistently present across all test runs with this model and other TRT LLM models?

[fpetrini15]

Closing issue due to lack of activity. Please re-open the issue if you would like to follow up with this issue.