fix device mapping issue of llama gptq (#2101)

Signed-off-by: Xin He <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

neural_compressor/torch/algorithms/weight_only/gptq.py

@@ -116,7 +116,8 @@ def trace_gptq_target_blocks(module, module_types=[torch.nn.ModuleList, torch.nn
                 gptq_related_blocks["transformers"] = m
                 find_transformers = True
                 # return gptq_related_blocks
-            elif is_leaf(m) and not find_transformers:
+            elif (is_leaf(m) and not find_transformers) or "Embedding" in type(m).__name__:
+                # "Embedding" in type(m).__name__ to resolve 'LlamaRotaryEmbedding'
                 gptq_related_blocks["embeddings"][n] = m
             elif n.find(gptq_related_blocks["transformers_name"]) == -1 and find_transformers:
                 # no longer belong to transformers