Fix Qwen3 SALM LoRA init (#15570)

Author: pzelasko

nemo/collections/speechlm2/parts/lora.py

@@ -15,6 +15,7 @@
 from peft import LoraConfig, get_peft_model
 from transformers import PreTrainedModel
 
+from nemo.collections.speechlm2.parts.pretrained import move_embedding
 from nemo.utils import logging
 
 
@@ -25,6 +26,9 @@ def maybe_install_lora(model):
         assert hasattr(model, "llm") and isinstance(model.llm, PreTrainedModel)
         assert "prevent_freeze_params" in model.cfg and isinstance(model.cfg.prevent_freeze_params, (list, ListConfig))
         model.lora_config = LoraConfig(**model.cfg.lora)
-        model.llm = get_peft_model(model.llm, model.lora_config)
+        # PEFT inspects get_input_embeddings() while wrapping the model, so temporarily
+        # restore the embedding layer that SALM keeps outside the LLM for FSDP/TP.
+        with move_embedding(model):
+            model.llm = get_peft_model(model.llm, model.lora_config)
         model.cfg.prevent_freeze_params.append(r"^.+\.lora_.+$")
         logging.info(f"LoRA adapter installed: {model.lora_config}")

tests/collections/speechlm2/test_salm_lora.py

@@ -0,0 +1,62 @@
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from types import SimpleNamespace
+
+from omegaconf import DictConfig
+from peft import PeftModel
+from transformers import Qwen3Config, Qwen3ForCausalLM
+
+from nemo.collections.speechlm2.parts.lora import maybe_install_lora
+
+
+def _make_qwen3_stub():
+    llm = Qwen3ForCausalLM(
+        Qwen3Config(
+            vocab_size=32,
+            hidden_size=16,
+            intermediate_size=32,
+            num_hidden_layers=1,
+            num_attention_heads=2,
+            num_key_value_heads=2,
+            max_position_embeddings=32,
+        )
+    )
+    return SimpleNamespace(
+        cfg=DictConfig(
+            {
+                "prevent_freeze_params": [],
+                "lora": {
+                    "r": 4,
+                    "lora_alpha": 8,
+                    "lora_dropout": 0.0,
+                    "target_modules": ["q_proj", "v_proj"],
+                    "task_type": "CAUSAL_LM",
+                },
+            }
+        ),
+        llm=llm,
+        embed_tokens=llm.model.embed_tokens,
+    )
+
+
+def test_maybe_install_lora_restores_qwen3_input_embeddings_temporarily():
+    model = _make_qwen3_stub()
+    del model.llm.model.embed_tokens
+
+    maybe_install_lora(model)
+
+    assert isinstance(model.llm, PeftModel)
+    assert hasattr(model.llm.base_model.model.model.layers[0].self_attn.q_proj, "lora_A")
+    assert model.cfg.prevent_freeze_params == [r"^.+\.lora_.+$"]
+    assert not hasattr(model.llm.base_model.model.model, "embed_tokens")