triton部署基于wenet的流式asr服务

[zengguofei]

gpu模型跑不起来，看着是配置没有对齐，但是相关配置已按要求设置，wenet的def infer没有被执行到
name: "streaming_wenet"
platform: "ensemble"
max_batch_size: 512 #MAX_BATCH

input [
{
name: "WAV"
data_type: TYPE_FP32
dims: [-1]
},
{
name: "WAV_LENS"
data_type: TYPE_INT32
dims: [1]
}
]

output [
{
name: "TRANSCRIPTS"
data_type: TYPE_STRING
dims: [1]
}
]

ensemble_scheduling {
step [
{
model_name: "feature_extractor"
model_version: -1
input_map {
key: "wav"
value: "WAV"
}
input_map {
key: "wav_lens"
value: "WAV_LENS"
}
output_map {
key: "speech"
value: "SPEECH"
}
output_map {
key: "speech_lengths"
value: "SPEECH_LENGTHS"
}
},
{
model_name: "encoder"
model_version: -1
input_map {
key: "speech" # Correct mapping here
value: "SPEECH"
}
input_map {
key: "speech_lengths" # Correct mapping here
value: "SPEECH_LENGTHS"
}
output_map {
key: "log_probs"
value: "ctc_log_probs"
}
output_map {
key: "log_probs_idx"
value: "beam_log_probs_idx"
}
output_map {
key: "chunk_out"
value: "CHUNK_OUT"
}
output_map {
key: "chunk_out_lens"
value: "CHUNK_OUT_LENS"
}
},
{
model_name: "wenet"
model_version: -1
input_map {
key: "log_probs"
value: "LOG_PROBS"
}
input_map {
key: "log_probs_idx"
value: "LOG_PROBS_IDX"
}
input_map {
key: "chunk_out"
value: "CHUNK_OUT"
}
input_map {
key: "chunk_out_lens"
value: "CHUNK_OUT_LENS"
}
output_map {
key: "OUTPUT0"
value: "TRANSCRIPTS"
}
}
]
}

name: "wenet"
backend: "python"
max_batch_size: 512 # 根据代码实现和需求支持批量大小，确保动态批次

parameters [
{
key: "beam_size",
value: { string_value: "10" }
},
{
key: "cutoff_prob",
value: { string_value: "0.9999" }
},
{
key: "alpha",
value: { string_value: "2" }
},
{
key: "beta",
value: { string_value: "1" }
},
{
key: "vocab_path",
value: { string_value: "/ws/onnx_model/units.txt" }
},
{
key: "lm_path",
value: { string_value: "/ws/model_repo/language_model.bin" }
},
{
key: "hotwords_path",
value: { string_value: "None" }
},
{
key: "bidecoder",
value: { string_value: "1" }
},
{
key: "rescoring",
value: { string_value: "1" }
}
]

input [
{
name: "log_probs"
data_type: TYPE_FP16 # 确保匹配 batch_log_probs 的数据类型
dims: [-1, -1, 5235] # 动态批次，时间步长，词表大小
},
{
name: "log_probs_idx"
data_type: TYPE_INT64 # 确保匹配 batch_log_probs_idx 的数据类型
dims: [-1, -1, 10] # 动态批次，时间步长，beam_size
},
{
name: "chunk_out"
data_type: TYPE_FP32 # 确保匹配 encoder 输出的类型
dims: [-1, -1, 256] # 动态批次，时间步长，特征维度
},
{
name: "chunk_out_lens"
data_type: TYPE_INT32 # 确保匹配 seq_lens 的数据类型
dims: [-1] # 动态批次
}
]

output [
{
name: "OUTPUT0"
data_type: TYPE_STRING # 确保匹配 final_result 的数据类型
dims: [1] # 动态批次大小
}
]

instance_group [
{
count: 2
kind: KIND_CPU # 根据实现使用 CPU
}
]