Resolve merge conflicts with origin/main

- auto_export.py: add deepseek_v4 mapping from main; keep qwen3_vl->qwen3_vl; merge MTP config block from main alongside local tokenizer copy logic
- integration_test.py: keep transformers >=5.14 recurrent_states comment from our branch

Author: Copilot

src/mobius/integrations/ort_genai/auto_export.py

@@ -86,12 +86,8 @@
     # HunYuan-V1 dense / Hy-MT1.5 — generic decoder LLM type accepted by
     # ORT GenAI (see onnxruntime-genai/src/models/model_type.h LLM list).
     "hunyuan_v1_dense": "decoder",
-<<<<<<< HEAD
-    # Qwen VL model families have separate ORT GenAI model types.
-=======
     "deepseek_v4": "decoder",
-    # Qwen VL models all use the same GenAI pipeline as qwen2_5_vl
->>>>>>> origin/main
+    # Qwen VL model families have separate ORT GenAI model types.
     "qwen2_vl": "qwen2_5_vl",
     "qwen3_vl": "qwen3_vl",
     "qwen3_vl_text": "qwen3_vl",
@@ -1104,10 +1100,6 @@ def write_ort_genai_config(
 
     result: dict[str, str] = {"genai_config": genai_path}
 
-<<<<<<< HEAD
-    # Copy tokenizer files. A local hf_model_id is a local model directory, not a
-    # Hub repo id; copy directly instead of calling hf_hub_download.
-=======
     if "mtp" in pkg:
         mtp_model = pkg["mtp"]
         mtp_path = os.path.join(directory, "mtp_config.json")
@@ -1136,9 +1128,8 @@ def write_ort_genai_config(
             f.write("\n")
         result["mtp_config"] = mtp_path
 
-    # Copy tokenizer files — HF Hub takes precedence; local dir is the fallback
-    # for --config mode where no HF model ID is available.
->>>>>>> origin/main
+    # Copy tokenizer files. A local hf_model_id is a local model directory, not a
+    # Hub repo id; copy directly instead of calling hf_hub_download.
     if hf_model_id is not None:
         if os.path.isdir(hf_model_id):
             logger.info("Copying tokenizer files from local model directory %s", hf_model_id)

tests/integration_test.py

@@ -3782,17 +3782,10 @@ def test_qwen35_deltanet_single_layer_parity():
             hidden_states=torch.from_numpy(hidden_np).float(),
             cache_params=cache,
         ).numpy()
-<<<<<<< HEAD
     # transformers >=5.14 changed recurrent_states from a tensor to a dict
     # keyed by layer index; extract the tensor for either version.
     _rec_states = cache.layers[0].recurrent_states
     hf_rec = (_rec_states[0] if isinstance(_rec_states, dict) else _rec_states).numpy()
-=======
-    hf_rec = cache.layers[0].recurrent_states
-    if isinstance(hf_rec, dict):
-        hf_rec = hf_rec[0]
-    hf_rec = hf_rec.numpy()
->>>>>>> origin/main
 
     # ONNX forward
     sess = _make_session(onnx_model)