Version mismatch while running engine files

[abd-gang]

Hi

I am using tensorrt-edgellm for quantising the Cosmos reason 2 model into fp8.

Once done with quantization, I am running in a different docker and getting this error -

---

[08/25/2026-14:46:02] [TRT] [E] IRuntime::deserializeCudaEngine: Error Code 1: Serialization (Serialization assertion stdVersionRead == kSERI
ALIZATION VERSION failed.Version tag does not match. Note: Current Version: 243, Serialized Engine Version: 240 In stdArchiveReaderInitCommon
at /src/runtime/deserialization/stdArchiveReader.cpp:49)
,2026-08-25 14:46:02,935][ main_ _][ERROR] - [rank: 0] Hybrid evaluation failed: 'NoneType' object has no attribute "create execution context

---

I thought this is the tensorrt version issue so I used the same tensorrt==11.2.1.2 but still gor the issue.

Here are the key points -

1. Engine for cosmos reason 2 is built using tensorrt-edgellm
2. I am trying to run in a different docker with no tensorrt-edgellm but only tensorrt
3. Both the docker has same tensorrt version.

Question -

1. Is it possible to build engine using tensorrtedgellm and run it separately using just tensorrt on the same platform?
2. What does the error says exactly? Which version of tensorrt corresponds to engine version 240 amd 243.

Thanks

[nvluxiaoz]

I think this is because you build the TRT engine and execute the runtime using different TensorRT implicitly. You probably build and run engine on dockers with different TensorRT versions. Both the docker has same tensorrt version. assumption might not be correct.

[abd-gang]

Hi
I made sure tensorrt engine in both the docker is same i.e 11.2.1.2
But still got the error. My assumption is tensorrtedgellm is building the engine using different tensorrt.
Can you tell me which version is 243 and 240?
Is tensorrtedgellm using different tensorrt version beneath its api calls (independent of installed tensorrt)?
Thanks

[abd-gang]

It's closed by mistake. Not able to open it again.

How to check which tensorrt version being used inside Tensorrtedgellm?

[abd-gang]

Any update?

[nvluxiaoz]

TensorRT Edge-LLM can take any TensorRT version, but engine build and inference needs to be operated on the same TRT version. It is recommended to complete engine build and inference together on device.