Minor tweaks to newer models

[derekk-nm]

SUMMARY:

• updates to Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 server configuration to attempt to get the accuracy workflow to run.
• RedHatAI/DeepSeek-R1-0528-quantized.w4a16 was missing its base model, and a server option.
• HuggingFaceTB/SmolLM3-3B is updated with an actual ground-truth gsm8k value.

TEST PLAN:
models were passed through the accuracy workflow.
the Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 model is just too big to test w/out access to a larger system.

[tmuttaki]

I saw these models in @debroy-rh 's small list of models for code coverage but these are not present in model validation config. Are you going to add these as well?

• deepseek-ai/DeepSeek-V2-Lite
• microsoft/Phi-3-medium-4k-instruct
• Qwen/Qwen2-57B-A14B-Instruct
• RedHatAI/Meta-Llama-3-8B-Instruct-FP8-KV
• RedHatAI/Mixtral-8x7B-Instruct-v0.1-FP8
• TechxGenus/Meta-Llama-3-8B-Instruct-GPTQ

[debroy-rh]

look good, thanks Derek.

[derekk-nm]

Thanks for the observation @tmuttaki .
Those models appear in the neuralmagic/lm-eval-configs/models directory because they were part of testing prior to the NeuralMagic acquisition by Red Hat. They haven't been included as part of the Third Party Model Validation program initiated at the beginning of Red Hat Summit. (this is the source I'm using for models in that program).