Question about output type of QuantizeLinear

[cc-sketch]

hello, why the log showed out type of QuantizeLinear is float

[cc-sketch]

and, when i serialize manually fused qdq onnx when only int8 in-out type supported in plugin, like
bool supportsFormatCombination(int32_t pos, const PluginTensorDesc* inOut, int32_t nbInputs, int32_t nbOutputs) NOEXCEPT override { switch (pos) { case 0: return ((inOut[0].type == DataType::kINT8) && (inOut[0].format == TensorFormat::kCHW4 || inOut[0].format == TensorFormat::kCHW32)); case 1: return (inOut[pos].type == inOut[0].type) && (inOut[pos].format == inOut[0].format); default: return false; } return false; }

it failed with log like

[05/06/2024-21:04:56] [TRT] [E] 9: [pluginV2Builder.cpp::reportPluginError::23] Error Code 9: Internal Error (LayerNorm_4: could not find any supported formats consistent with input/output data types)

[zerollzeng]

hello, why the log showed out type of QuantizeLinear is float

What does it look like in onnx?

[zerollzeng]

You implement a layernorm plugin? TRT has native support.

[cc-sketch]

@zerollzeng thank you and sorry for the late reply. I will try the layernorm plugin lately. But now what i really want to ask is how to serialize the "manually fused qdq onnx ". As seen below, left is the onnx with a couple of qdq nodes, and right is fused onnx.

Firstly, when i serialize the model with only "kINT8 DataType" supported in trt plugin, as

It produced the error:

[pluginV2Builder.cpp::reportPluginError::23] Error Code 9: Internal Error (MyOp_3: could not find any supported formats consistent with input/output data types)

So i asked "why out type of QuantizeLinear is float".

and, if i serialize the model with "kFLOAT DataType" supported also in trt plugin, as

It produced the error:

[optimizer.cpp::filterQDQFormats::4422] Error Code 2: Internal Error (Assertion !n->candidateRequirements.empty() failed. All of the candidates were removed, which points to the node being incorrectly marked as an int8 node.)

With trt-8531 and cuda 11.3. Thank you very much.

[zerollzeng]

Your Q/DQ placement is wrong to me, try Q->MyOp->DQ if your plugin is in int8.

[cc-sketch]

Sorry, i didn't understand what this mean.

I manually fused the model. From Q->DQ->MyOp->Q->DQ, to Q-> MyOp->DQ, like https://forums.developer.nvidia.com/t/where-do-tensorrt-plugin-determine-whether-fuse-qdq-or-not/226705/4

So, for my question what is wrong? Is there sth wrong for quantized model generation, or sth wrong in MyOp plugin? It confused me for a long time.