diff --git a/tests/models/multimodal/generation/test_nemotron_parse.py b/tests/models/multimodal/generation/test_nemotron_parse.py index 57e40430c6..69e41b21e3 100644 --- a/tests/models/multimodal/generation/test_nemotron_parse.py +++ b/tests/models/multimodal/generation/test_nemotron_parse.py @@ -104,16 +104,7 @@ def run_test( ) -@pytest.mark.parametrize( - "model", - [ - "nvidia/NVIDIA-Nemotron-Parse-v1.2", - # v2.0's checkpoint ties lm_head to the decoder's input embeddings - # instead of shipping a separate lm_head.weight tensor; this - # exercises that path (see tie_weights() in nemotron_parse.py). - "nvidia/NVIDIA-Nemotron-Parse-2.0", - ], -) +@pytest.mark.parametrize("model", ["nvidia/NVIDIA-Nemotron-Parse-v1.2"]) @pytest.mark.parametrize("dtype", ["bfloat16"]) @pytest.mark.parametrize("num_logprobs", [5]) def test_models( diff --git a/vllm/model_executor/models/nemotron_parse.py b/vllm/model_executor/models/nemotron_parse.py index 3ba731d2e0..9ac38c0021 100644 --- a/vllm/model_executor/models/nemotron_parse.py +++ b/vllm/model_executor/models/nemotron_parse.py @@ -568,15 +568,6 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal): self.lm_head = ParallelLMHead( config.decoder.vocab_size, config.decoder.d_model, quant_config=quant_config ) - # Some checkpoints (e.g. compact exports) tie the output head to the - # decoder's input embeddings instead of materializing a separate - # lm_head.weight tensor. - tie_word_embeddings = bool( - getattr(config, "tie_word_embeddings", False) - or getattr(config.decoder, "tie_word_embeddings", False) - ) - if tie_word_embeddings: - self.lm_head = self.lm_head.tie_weights(self.decoder.embed_tokens) self.logits_processor = LogitsProcessor( self.vocab_size, config.decoder.vocab_size )