Myelin "Could not infer output types for operation: dequantize" on SM 8.7 (Orin); identical engine builds on SM 11.0 (Thor) with the same TensorRT 10.16.2.10

[geoffrey-delhomme]

Summary

Myelin fails to compile a dequantize operation on SM 8.7 (Orin) that compiles successfully on SM 11.0 (Thor), with identical TensorRT 10.16.2.10, identical L4T R39.2, and the byte-identical ONNX file. The two devices differ only in GPU architecture.

Environment

	Orin	Thor
TensorRT	10.16.2.10-1+cuda13.2 (trtexec v101602)	identical
L4T	R39 revision 2.0	identical
GPU arch	SM 8.7 (Ampere)	SM 11.0 (Blackwell)
Result	❌ build fails	✅ builds in 123 s

Producer: NVIDIA ModelOpt 0.44.0 (modelopt.onnx.quantization), INT8 PTQ, entropy calibration, 320 real-image samples. Consumer: trtexec --stronglyTyped.

Reproduction

trtexec --onnx=quantized_int8.onnx --saveEngine=int8.engine --stronglyTyped

Failure

Every tactic is skipped with:

Error Code 9: ... [type.cpp:186: infer_type] Could not infer output types for operation:
291: dequantize: Dequantized input0-(f32[1,16,512,512]) | input0 reshape out_transpose_out
reshape out slice_U28 out-(i8[]), MyelinPW(42188) - Input Dequantization 0 scale
weightsHalf-0.35498H:(f16[]) ... No matching rules found for input operand types
In compileGraph at .../myelin/codeGenerator.cpp:1795

The operand mix is f32 activation + i8 slice + f16 scale. The Ampere codegen path appears to have no rule for that combination; the Blackwell path does.

What this is not

Three different quantization scopes were produced, giving three distinct artifacts. All three fail on Orin with a byte-identical error — same node 291, same MyelinPW(42188), same 0.35498 scale:

Scope	Excluded op types	ONNX sha256 (16)	Orin
plain INT8	none	fcf873e63e0a951c	❌
head-aware	Sigmoid, ReduceMax, Flatten, TopK, Gather, GatherElements	03d2ab3115dd4de5	❌
head + stem	above + Reshape, Transpose, Slice	540fa615ced07025	❌

The error not moving while the graph does rules out quantization scope as the cause, and Thor compiling the identical file rules out the model and the calibration. That leaves architecture-specific codegen.

Impact

Strongly-typed INT8 is unavailable on Orin for this detector family. The documented alternative — weakly-typed --int8 --fp16 — permits silent downcast, which our deployment contract forbids.