[Bug][Export]: Qwen3-ASR export can preserve rope_type=linear and disable MRoPE at runtime

[suharvest]

Describe the bug

Exporting the public Qwen3-ASR checkpoint with TensorRT-Edge-LLM v0.9.1 /
current main can produce a runtime LLM config with:

{
  "rope_scaling": {
    "type": "linear",
    "rope_type": "linear",
    "factor": 1.0,
    "mrope_section": [24, 20, 20]
  }
}

The checkpoint contains MRoPE sections, but the current normalization helper
does not canonicalize linear + mrope_section to MRoPE. The C++ runtime does
not select its MRoPE path for this exported config. The engine can build, but
runtime execution then lacks mropeCosSinOut, producing a silent
configuration/correctness failure instead of a clean export error.

Steps/Code to reproduce bug

Installation method:

Built from source at
7f061f21f0a581ba234a1e233c9315b89d8e47d6.

Model:

Qwen/Qwen3-ASR-0.6B@5eb144179a02acc5e5ba31e748d22b0cf3e303b0

Export flow:

tensorrt-edgellm-quantize llm \
  --model_dir Qwen/Qwen3-ASR-0.6B \
  --output_dir ./qwen3-asr-int4 \
  --quantization int4_awq

tensorrt-edgellm-export llm \
  --model_dir ./qwen3-asr-int4 \
  --output_dir ./qwen3-asr-onnx

Inspect the generated LLM config.json. Both rope_scaling.type and
rope_scaling.rope_type remain linear while mrope_section is present.

Current normalization:

TensorRT-Edge-LLM/tensorrt_edgellm/checkpoint/checkpoint_utils.py

Lines 54 to 72 in 7f061f2

	def normalize_rope_scaling_for_runtime(rope_scaling: Any) -> Any:
	"""Normalize HF MRoPE / rope_parameters metadata to the shape expected by
	the C++ runtime.
	"""
	if not isinstance(rope_scaling, dict):
	return rope_scaling
	normalized = dict(rope_scaling)
	if "mrope_section" in normalized:
	rope_type = normalized.get("type") or normalized.get("rope_type")
	if rope_type in (None, "default", "mrope"):
	normalized["type"] = "default"
	normalized["rope_type"] = "default"
	# rope_parameters (transformers v5) carries "rope_type" without "type";
	# propagate the alias so Python callers keyed off "type" still work.
	# C++ collectRopeConfig() accepts either key.
	if "type" not in normalized and "rope_type" in normalized:
	normalized["type"] = normalized["rope_type"]
	return normalized

Expected behavior

For Qwen3-ASR, an exported rope config containing mrope_section should use
the canonical MRoPE semantics expected by the C++ runtime. The exporter
should preserve all unrelated rope parameters, leave other model families
unchanged, and be idempotent for already-canonical MRoPE configs.

System information (x86 Host with GPU)

• Container used: source virtual environment
• OS: WSL2 / Linux
• CPU architecture: x86_64
• GPU: NVIDIA GeForce RTX 3060
• GPU memory: 12 GB
• Number of GPUs: 1
• Library versions:
  • Python: 3.12
  • TensorRT Edge-LLM: 7f061f21f0a581ba234a1e233c9315b89d8e47d6
  • CUDA: 13.0 build environment
  • PyTorch: 2.12.0+cu130
  • ModelOpt: 0.44.0

A focused config-level regression test and minimal exporter fix are available
for a follow-up PR after issue approval. Device validation used a freshly
built SM87 engine; only the exported rope config changed.