phi-4

Author: ysjprojects

README.md

@@ -135,6 +135,7 @@ Every model is written from scratch to maximize performance and remove layers of
 | OpenLLaMA | 3B, 7B, 13B | OpenLM Research | [Geng & Liu 2023](https://github.com/openlm-research/open_llama)                                                         |
 | Phi 1.5 & 2 | 1.3B, 2.7B | Microsoft Research  | [Li et al. 2023](https://arxiv.org/abs/2309.05463)                                                                  |
 | Phi 3 | 3.8B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2404.14219)                                                                            |
+| Phi 4 | 14B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2412.08905)                                                                            |
 | Platypus | 7B, 13B, 70B |  Lee et al. | [Lee, Hunter, and Ruiz 2023](https://arxiv.org/abs/2308.07317)                                                               |
 | Pythia | {14,31,70,160,410}M, {1,1.4,2.8,6.9,12}B | EleutherAI | [Biderman et al. 2023](https://arxiv.org/abs/2304.01373)                                            |
 | Qwen2.5 | 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Alibaba Group | [Qwen Team 2024](https://qwenlm.github.io/blog/qwen2.5/)                                               |

litgpt/config.py

@@ -1575,6 +1575,25 @@ def norm_class(self) -> Type:
         mlp_class_name="LLaMAMLP",
         parallel_residual=False,
     ),
+    # https://huggingface.co/microsoft/phi-4/blob/main/config.json
+    dict(
+        name="phi-4",
+        hf_config=dict(org="microsoft", name="phi-4"),
+        vocab_size=100257,
+        padded_vocab_size=100352,
+        block_size=16384,
+        n_embd=5120,
+        n_layer=40,
+        n_head=40,
+        n_query_groups=10,
+        rotary_percentage=1.0,
+        bias=False,
+        norm_class_name="RMSNorm",
+        intermediate_size=17920,
+        rope_base=250000,
+        mlp_class_name="LLaMAMLP",
+        parallel_residual=False,
+    ),
 ]
 configs.extend(phi)
 

litgpt/prompts.py

@@ -268,6 +268,9 @@ class Phi3(PromptStyle):
     def apply(self, prompt: str, **kwargs: str) -> str:
         return f'<|system|>\nYou are a helpful assistant.<|end|>\n<|user|>\n{prompt}<|end|>\n<|assistant|>\n'
 
+class Phi4(PromptStyle):
+    def apply(self, prompt: str, **kwargs: str) -> str:
+        return f'<|im_start|>user<|im_sep|>{prompt}<|im_end|><|im_start|>assistant<|im_sep|>'
 
 class TinyLlama(PromptStyle):
     def apply(self, prompt: str, **kwargs: str) -> str:
@@ -337,6 +340,7 @@ def __init__(self):
     "phi-1": Phi1,
     "phi-2": Phi2,
     "phi-3": Phi3,
+    "phi-4": Phi4,
     "tinyllama": TinyLlama,
     "gemma": Gemma,
     "llama3": Llama3,
@@ -380,6 +384,8 @@ def model_name_to_prompt_style(model_name: str) -> PromptStyle:
         return Phi2()
     if re.search("Phi-3", model_name):
         return Phi3()
+    if re.search("phi-4", model_name):
+        return Phi4()
     if re.search(r"tiny-llama.*chat", model_name):
         return TinyLlama()
     if re.search(r"(Code)?Gemma.*-it", model_name):

tests/test_model.py

@@ -329,7 +329,7 @@ def test_against_hf_phi(model_name, device, dtype):
 
 
 @torch.inference_mode()
-@pytest.mark.parametrize("model_name", ("Phi-3-mini-4k-instruct", "Phi-3-mini-128k-instruct", "Phi-3.5-mini-instruct"))
+@pytest.mark.parametrize("model_name", ("Phi-3-mini-4k-instruct", "Phi-3-mini-128k-instruct", "Phi-3.5-mini-instruct", "phi-4"))
 @pytest.mark.parametrize(
     ("device", "dtype"),
     [

tutorials/download_model_weights.md

@@ -34,6 +34,7 @@ LitGPT supports a variety of LLM architectures with publicly available weights.
 | OpenLLaMA | 3B, 7B, 13B | OpenLM Research | [Geng & Liu 2023](https://github.com/openlm-research/open_llama)                                                         |
 | Phi 1.5 & 2 | 1.3B, 2.7B | Microsoft Research  | [Li et al. 2023](https://arxiv.org/abs/2309.05463)                                                                          |
 | Phi 3 & 3.5 | 3.8B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2404.14219)
+| Phi 4 | 14B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2412.08905)                                                                            |
 | Platypus | 7B, 13B, 70B |  Lee et al. | [Lee, Hunter, and Ruiz 2023](https://arxiv.org/abs/2308.07317)                                                               |
 | Pythia | {14,31,70,160,410}M, {1,1.4,2.8,6.9,12}B | EleutherAI | [Biderman et al. 2023](https://arxiv.org/abs/2304.01373)                                            |
 | Qwen2.5 | 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Alibaba Group | [Qwen Team 2024](https://qwenlm.github.io/blog/qwen2.5/)                                               |
@@ -165,6 +166,7 @@ microsoft/phi-2
 microsoft/Phi-3-mini-128k-instruct
 microsoft/Phi-3-mini-4k-instruct
 microsoft/Phi-3.5-mini-instruct
+microsoft/phi-4
 mistralai/mathstral-7B-v0.1
 mistralai/Mistral-7B-Instruct-v0.1
 mistralai/Mistral-7B-Instruct-v0.2