Add performance config for microsoft/Phi-3-medium-4k-instruct

[tmuttaki]

SUMMARY:

When running vllm with microsoft/Phi-3-medium-4k-instruct we get the following error:

Value error, User-specified max_model_len (8192) is greater than the derived max_model_len (max_position_embeddings=4096 or model_max_length=None in model's config.json). 
This may lead to incorrect model outputs or CUDA errors. 
To allow overriding this maximum, set the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 [type=value_error, input_value=ArgsKwargs((), {'model': ...gits_processors': None}), input_type=ArgsKwargs]

Because it gets the max_model_len from model-configs/common/performance/server.yml

WARNING  neuralmagic.utils:utils.py:104 No performance server config found for microsoft/Phi-3-medium-4k-instruct, using fallback=PosixPath('model-configs/common/performance/server.yml')

Example run: https://github.com/neuralmagic/nm-cicd/actions/runs/17684609008

TEST PLAN:

Adding this performance config solves this issue:
Example run: https://github.com/neuralmagic/nm-cicd/actions/runs/17684808798

[derekk-nm]

@tmuttaki , I'm curious why we're testing with this model. It's not listed in the Model Validation tracker.

[tarukumar]

@derekk-nm This model is part of Accept Sync, and since we were observing issues with Accept Sync on this model, we wanted to include it in our validation. I see no harm adding the config in the repo. Let us know if you feel otherwise.

[derekk-nm]

Ok. makes sense. No harm adding it to the repo. Be wary of the additional time it adds to the Accept Sync, though.