TensorRT-RTX 1.4 fails on YOLO ONNX models on Turing with target_sm arch-kind error

[ogsdave]

Summary

TensorRT-RTX 1.4.0.76 succeeds on a minimal Identity ONNX model but fails on both yolo26m.onnx and yolo26m-seg.onnx during CPU-only AOT generation on an RTX 2070 Max-Q (sm_75).

The failure is an internal Myelin error:

MyelinCheckException: hfc.cpp:120: CHECK(false) failed. target_sm '110' does not have arch kind assigned.

I also reproduced the same failure class with the CUDA 12.9 SDK variant, where the internal target was 101 instead of 110.

Environment

• Ubuntu 24.04.4 LTS
• • Kernel 6.17.0-22-generic
• • • GPU: NVIDIA GeForce RTX 2070 with Max-Q Design
• • • • Compute capability: 7.5
• • • • • Driver: 580.142
• • • • • • Local CUDA toolkit: 13.2 (V13.2.78)
• • • • • • • TensorRT-RTX SDKs tested:
• • • • • • • • TensorRT-RTX-1.4.0.76-Linux-x86_64-cuda-12.9-Release-external.tar.gz
• • • • • • • • • TensorRT-RTX-1.4.0.76-Linux-x86_64-cuda-13.2-Release-external.tar.gz

Repro Commands

Passing minimal model:

env PATH="/usr/local/cuda-13.2/bin:$PATH" \
  LD_LIBRARY_PATH="/usr/local/cuda-13.2/lib64${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" \
  /tmp/trtx_sdk_132/TensorRT-RTX-1.4.0.76/bin/tensorrt_rtx \
  --onnx=/home/dave/Notes/TensorRT-RTX/models/trtx_identity.onnx \
  --saveEngine=/tmp/trtx_identity_fp16_1x3x640x640.trt \
  --cpuOnly \
  --skipInference

Failing YOLO model:

env PATH="/usr/local/cuda-13.2/bin:$PATH" \
  LD_LIBRARY_PATH="/usr/local/cuda-13.2/lib64${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" \
  /tmp/trtx_sdk_132/TensorRT-RTX-1.4.0.76/bin/tensorrt_rtx \
  --onnx=/home/dave/Notes/TensorRT-RTX/models/yolo26m.onnx \
  --saveEngine=/tmp/yolo26m_fp16_1x3x640x640.trt \
  --cpuOnly \
  --skipInference

Failing YOLO segmentation model:

env PATH="/usr/local/cuda-13.2/bin:$PATH" \
  LD_LIBRARY_PATH="/usr/local/cuda-13.2/lib64${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" \
  /tmp/trtx_sdk_132/TensorRT-RTX-1.4.0.76/bin/tensorrt_rtx \
  --onnx=/home/dave/Notes/TensorRT-RTX/models/yolo26m-seg.onnx \
  --saveEngine=/tmp/yolo26m-seg_fp16_1x3x640x640.trt \
  --cpuOnly \
  --skipInference

Observed Behavior

Minimal Identity ONNX:

• build succeeds

• • engine size is nonzero
• • • TensorRT-RTX reports PASSED
      Both YOLO ONNX models:

• model parsing succeeds

• • build fails in Myelin with:

[E] Error[1]: IBuilder::buildSerializedNetworkToStream: Error Code 1: Myelin ([myelin_graph.h:1250: attachExceptionMsgToGraph] MyelinCheckException: hfc.cpp:120: CHECK(false) failed. target_sm '110' does not have arch kind assigned. In compileGraph at /_src/optimizer/myelin/codeGenerator.cpp:1783)
[I] Created engine with size: 0 MiB
[E] Assertion failure: false && "Attempting to access an empty engine!"

Earlier CUDA 12.9 SDK result for yolo26m.onnx:

Internal Error: MyelinCheckException: hfc.cpp:120: CHECK(false) failed. target_sm '101' does not have arch kind assigned.

Expected Behavior

If these models are unsupported, I would expect an explicit model-support or operator-support error.

Instead, TensorRT-RTX fails with an internal compiler assertion.

Request

Please clarify:

1. Whether this is a known TensorRT-RTX 1.4 issue on Turing / sm_75.
2. Whether these YOLO ONNX exports use an unsupported graph pattern.
3. Whether there is a workaround, export change, or internal flag that avoids this failure.
4. Whether a future release is expected to fix it.

Attached Logs

yolo26m_trt_rtx_build.log

trtx_identity_trt_rtx_build.log

yolo26m-seg_trt_rtx_build.log

[ogsdave]

Follow-up result: I tested the nano variants as well with the same TensorRT-RTX 1.4.0.76 + CUDA 13.2 host setup.

Both additional models fail with the same Myelin internal error:

• yolo26n.onnx: failed
• yolo26n-seg.onnx: failed

Error:

MyelinCheckException: hfc.cpp:120: CHECK(false) failed. target_sm '110' does not have arch kind assigned.

So this is not limited to the 26m variants. On this host, all current YOLO-family ONNX models in the repo now fail under TensorRT-RTX 1.4, while the trivial Identity ONNX still passes.

Additional local logs for these runs are available if useful:

• yolo26n_trt_rtx_build.log
• yolo26n-seg_trt_rtx_build.log

[davemichael]

Hi, thanks for filing. Can you try adding the "--useGpu" flag when you run tensorrt_rtx as a workaround? This will build an engine that only works on your specific RTX 2070, but it will be better optimized for your 2070 and might get you past this specific error.

[ogsdave]

Follow-up: NVIDIA's suggested workaround works on this host.

Using --useGpu --skipInference with the same TensorRT-RTX 1.4.0.76 + CUDA 13.2 setup succeeds for both main YOLO models on this RTX 2070:

• yolo26m.onnx: PASSED, engine size 127150436 bytes
• yolo26m-seg.onnx: PASSED, engine size 140249724 bytes

So the failure appears to be specific to the portable / CPU-only AOT path. Building a GPU-specific engine for the local RTX 2070 gets past the target_sm arch-kind error.