Cannot warmup with python backend with batch_size

Description

For example config in below.

name: "postprocess"
backend: "python"
max_batch_size: 0
input [
{
    name: "postprocess_in"
    data_type: TYPE_FP32
    dims: [ -1, 581, 2 ]
}
]

output [
{
    name: "postprocess_out"
    data_type: TYPE_FP32
    dims: [ -1 ]
}
]

instance_group [{ kind: KIND_CPU, count: 2 }]

# warmup 2 requests
model_warmup {
  name: "RandomSampleInput"
  batch_size: 256
  inputs [{
      key: "postprocess_in"
      value: {
        data_type: TYPE_FP32
        dims: [1, 581, 2 ]
        random_data: true
      }
   }]
}

Input’s first dim is different with output’s first dim. However I tried to warmup multiple times with batch_size. There is no way to duplicate warmup request.

Error messages

E0520 12:38:18.786236 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0
E0520 12:38:18.786243 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0
E0520 12:38:18.786250 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0
E0520 12:38:18.786257 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0
E0520 12:38:18.786264 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0
E0520 12:38:18.786274 1 triton_model_instance.cc:86] warmup error: Internal - batch size 256 for 'postprocess', max allowed is 0

Triton Information

22.01 ~ 22.04

To Reproduce

Load model with warmup. max_batch_size in model, batch_size in model_warmup

Expected behavior

Model should accept dynamic shape input.

[GuanLuo]

Please take a look at https://github.com/triton-inference-server/server/pull/4501 for usage of the new “count” field

[tanmayv25]

model_warmup is a repeated field. If you want to run multiple requests in warmup, just replicate them in your model config.

https://github.com/triton-inference-server/common/blob/main/protobuf/model_config.proto#L1927-L1935

@GuanLuo Will it make sense to add a repeat_count field in model_warmup? The sample will be repeated for that many times before going to the next execution. I believe @kimdwkimdw here is getting confused between batch_size and repeat_count functionality I described.