From 44e227bbf82ee4ae5e2eede7a2b6309c44d77102 Mon Sep 17 00:00:00 2001
From: jeromeku <jerome.ku@gmail.com>
Date: Mon, 31 Mar 2025 10:22:50 -0700
Subject: [PATCH] add phi

---
 unsloth/registry/_phi.py           | 62 ++++++++++++++++++++++++++++++
 unsloth/registry/model_registry.py | 59 ++--------------------------
 2 files changed, 66 insertions(+), 55 deletions(-)
 create mode 100644 unsloth/registry/_phi.py

diff --git a/unsloth/registry/_phi.py b/unsloth/registry/_phi.py
new file mode 100644
index 0000000000..a6d18cbd61
--- /dev/null
+++ b/unsloth/registry/_phi.py
@@ -0,0 +1,62 @@
+from unsloth.registry.registry import ModelInfo, ModelMeta, QuantType, _register_models
+
+_IS_PHI_REGISTERED = False
+_IS_PHI_INSTRUCT_REGISTERED = False
+
+class PhiModelInfo(ModelInfo):
+    @classmethod
+    def construct_model_name(cls, base_name, version, size, quant_type, instruct_tag):
+        key = f"{base_name}-{version}"
+        key = cls.append_instruct_tag(key, instruct_tag)
+        key = cls.append_quant_type(key, quant_type)
+        return key
+
+# Phi Model Meta
+PhiMeta = ModelMeta(
+    org="microsoft",
+    base_name="phi",
+    instruct_tags=[None],
+    model_version="4",
+    model_sizes=["1"],  # Assuming only one size
+    model_info_cls=PhiModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH],
+)
+
+# Phi Instruct Model Meta
+PhiInstructMeta = ModelMeta(
+    org="microsoft",
+    base_name="phi",
+    instruct_tags=["mini-instruct"],
+    model_version="4",
+    model_sizes=["1"],  # Assuming only one size
+    model_info_cls=PhiModelInfo,
+    is_multimodal=False,
+    quant_types=[QuantType.NONE, QuantType.BNB, QuantType.UNSLOTH, QuantType.GGUF],
+)
+
+def register_phi_models(include_original_model: bool = False):
+    global _IS_PHI_REGISTERED
+    if _IS_PHI_REGISTERED:
+        return
+    _register_models(PhiMeta, include_original_model=include_original_model)
+    _IS_PHI_REGISTERED = True
+
+def register_phi_instruct_models(include_original_model: bool = False):
+    global _IS_PHI_INSTRUCT_REGISTERED
+    if _IS_PHI_INSTRUCT_REGISTERED:
+        return
+    _register_models(PhiInstructMeta, include_original_model=include_original_model)
+    _IS_PHI_INSTRUCT_REGISTERED = True
+
+register_phi_models(include_original_model=True)
+register_phi_instruct_models(include_original_model=True)
+
+if __name__ == "__main__":
+    from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
+    for model_id, model_info in MODEL_REGISTRY.items():
+        model_info = _check_model_info(model_id)
+        if model_info is None:
+            print(f"\u2718 {model_id}")
+        else:
+            print(f"\u2713 {model_id}") 
\ No newline at end of file
diff --git a/unsloth/registry/model_registry.py b/unsloth/registry/model_registry.py
index a0cd71c17a..de9609934c 100644
--- a/unsloth/registry/model_registry.py
+++ b/unsloth/registry/model_registry.py
@@ -4,11 +4,11 @@ from typing import Callable, Literal
 from unsloth.registry._llama import LlamaMeta3_1, LlamaMeta3_2
 from unsloth.registry.common import ModelInfo, ModelMeta
 
-_IS_LLAMA_REGISTERED = False
-_IS_LLAMA_VISION_REGISTERED = False
+# _IS_LLAMA_REGISTERED = False
+# _IS_LLAMA_VISION_REGISTERED = False
 
-_IS_QWEN_REGISTERED = False
-_IS_QWEN_VL_REGISTERED = False
+# _IS_QWEN_REGISTERED = False
+# _IS_QWEN_VL_REGISTERED = False
 
 _IS_GEMMA_REGISTERED = False
 
@@ -17,28 +17,6 @@ _IS_PHI_INSTRUCT_REGISTERED = False
 
 
 
-# class QwenModelInfo(ModelInfo):
-#     @classmethod
-#     def construct_model_name(
-#         cls, base_name, version, size, quant_type, instruct_tag
-#     ):
-#         key = f"{base_name}{version}-{size}B"
-#         key = cls.append_instruct_tag(key, instruct_tag)
-#         key = cls.append_quant_type(key, quant_type)
-#         return key
-
-
-# class QwenVLModelInfo(ModelInfo):
-#     @classmethod
-#     def construct_model_name(
-#         cls, base_name, version, size, quant_type, instruct_tag
-#     ):
-#         key = f"{base_name}{version}-VL-{size}B"
-#         key = cls.append_instruct_tag(key, instruct_tag)
-#         key = cls.append_quant_type(key, quant_type)
-#         return key
-
-
 # class PhiModelInfo(ModelInfo):
 #     @classmethod
 #     def construct_model_name(
@@ -55,35 +33,6 @@ _IS_PHI_INSTRUCT_REGISTERED = False
 
 # # Qwen text only models
 # # NOTE: Qwen vision models will be registered separately
-# _QWEN_INFO = {
-#     "org": "Qwen",
-#     "base_name": "Qwen",
-#     "instruct_tags": [None, "Instruct"],
-#     "model_versions": ["2.5"],
-#     "model_sizes": {"2.5": [3, 7]},
-#     "is_multimodal": False,
-#     "model_info_cls": QwenModelInfo,
-# }
-
-# _QWEN_VL_INFO = {
-#     "org": "Qwen",
-#     "base_name": "Qwen",
-#     "instruct_tags": ["Instruct"],  # No base, only instruction tuned
-#     "model_versions": ["2.5"],
-#     "model_sizes": {"2.5": [3, 7, 32, 72]},
-#     "is_multimodal": True,
-#     "instruction_tuned_only": True,
-#     "model_info_cls": QwenVLModelInfo,
-# }
-
-# _GEMMA_INFO = {
-#     "org": "google",
-#     "base_name": "gemma",
-#     "instruct_tags": ["pt", "it"],  # pt = base, it = instruction tuned
-#     "model_versions": ["3"],
-#     "model_sizes": {"3": [1, 4, 12, 27]},
-#     "is_multimodal": True,
-# }
 
 # _PHI_INFO = {
 #     "org": "microsoft",