[Model] Add support for GPTJ architecture (#3012)

This PR supports GPTJ architecture.

Author: tlopex

python/mlc_llm/model/gpt_j/__init__.py

@@ -0,0 +1,72 @@
+"""
+This file specifies how MLC's GPTJ parameter maps from other formats, for example HuggingFace
+PyTorch, HuggingFace safetensors.
+"""
+
+import functools
+
+import numpy as np
+
+from mlc_llm.loader import ExternMapping
+from mlc_llm.quantization import Quantization
+
+from .gpt_j_model import GPTJConfig, GPTJForCausalLM
+
+
+def huggingface(model_config: GPTJConfig, quantization: Quantization) -> ExternMapping:
+    """Returns a parameter mapping that maps from the names of MLC LLM parameters to
+    the names of HuggingFace PyTorch parameters.
+
+    Parameters
+    ----------
+    model_config : GPTJConfig
+        The configuration of the GPTJ model.
+
+    quantization : Quantization
+        The quantization configuration.
+
+    Returns
+    -------
+    param_map : ExternMapping
+        The parameter mapping from MLC to HuggingFace PyTorch.
+    """
+    model = GPTJForCausalLM(model_config)
+    if quantization is not None:
+        model.to(quantization.model_dtype)
+    _, _named_params, _ = model.export_tvm(  # type: ignore[misc]
+        spec=model.get_default_spec(),
+        allow_extern=True,
+    )
+    named_parameters = dict(_named_params)
+
+    mapping = ExternMapping()
+
+    for i in range(model_config.n_layer):
+        # Add gates in MLP
+        attn = f"transformer.h.{i}.attn"
+        mlc_name = f"{attn}.c_attn.weight"
+        mlc_param = named_parameters[mlc_name]
+        mapping.add_mapping(
+            mlc_name,
+            [
+                f"{attn}.q_proj.weight",
+                f"{attn}.k_proj.weight",
+                f"{attn}.v_proj.weight",
+            ],
+            functools.partial(
+                lambda q, k, v, dtype: np.concatenate([q, k, v], axis=0).astype(dtype),
+                dtype=mlc_param.dtype,
+            ),
+        )
+
+    for mlc_name, mlc_param in named_parameters.items():
+        if mlc_name not in mapping.param_map:
+            mapping.add_mapping(
+                mlc_name,
+                [mlc_name],
+                functools.partial(
+                    lambda x, dtype: x.astype(dtype),
+                    dtype=mlc_param.dtype,
+                ),
+            )
+    return mapping