Phi4-multimodal rotary_scaling is RotarySaclingType.none

[jl749]

Hello TensorRT-LLM team

I have a question regarding Phi4-multimodal trtllm deployment.

• tensorrt_llm == 0.19.0
• transformers == 4.51.0

https://github.com/NVIDIA/TensorRT-LLM/blob/v0.19.0/tensorrt_llm/models/__init__.py#L157
suggests microsoft/Phi-4-multimodal-instruct(config.json) follows Phi3ForCausalLM pipeline when trtllm-build is called

# phi4mm example
huggingface-cli download microsoft/Phi-4-multimodal-instruct --local-dir Phi-4-multimodal-instruct_BASELINE
python3 -m TensorRT-LLM.examples.phi.convert_checkpoint \
	  --model_dir Phi-4-multimodal-instruct_BASELINE \
	  --output_dir Phi-4-multimodal-instruct_BASELINE/trtllm_ckpt
trtllm-build \
  --checkpoint_dir tests/Phi-4-multimodal-instruct_BASELINE/trtllm_ckpt \
  --output_dir tests/Phi-4-multimodal-instruct_BASELINE/trtllm_engine \
  --max_beam_width 1 --max_batch_size 1 --max_input_len 1024 --max_seq_len 2048  \
  --context_fmha enable --remove_input_padding enable \
  --kv_cache_type paged --gpt_attention_plugin auto --gemm_plugin disable

However, when Attention modules are initialized under Phi3ForCausalLM (link)
It sets rotary_embedding_scaling=None

Making rotary_embedding_scale_type to become RotaryScalingType.none instead of RotaryScalingType.longrope

Is this expected?

When rope_type is longrope I expect Attention modules to contain

• self.rotary_embedding_scale_type = RotaryScalingType.longrope
• self.position_embedding_type = PositionEmbeddingType.long_rope