Upload PhiForCausalLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,32 @@
+{
+  "_name_or_path": "merged_adapters",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "PhiForCausalLM"
+  ],
+  "attn_pdrop": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_phi.PhiConfig",
+    "AutoModelForCausalLM": "modeling_phi.PhiForCausalLM"
+  },
+  "embd_pdrop": 0.0,
+  "flash_attn": false,
+  "flash_rotary": false,
+  "fused_dense": false,
+  "img_processor": null,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "phi-msft",
+  "n_embd": 2560,
+  "n_head": 32,
+  "n_head_kv": null,
+  "n_inner": null,
+  "n_layer": 32,
+  "n_positions": 2048,
+  "resid_pdrop": 0.1,
+  "rotary_dim": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.39.1",
+  "vocab_size": 51200
+}

configuration_phi.py

@@ -0,0 +1,62 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import math
+from typing import Optional
+
+from transformers import PretrainedConfig
+
+
+class PhiConfig(PretrainedConfig):
+    """Phi configuration."""
+
+    model_type = "phi-msft"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int = 50304,
+        n_positions: int = 2048,
+        n_embd: int = 1024,
+        n_layer: int = 20,
+        n_inner: Optional[int] = None,
+        n_head: int = 16,
+        n_head_kv: Optional[int] = None,
+        rotary_dim: Optional[int] = 32,
+        activation_function: Optional[str] = "gelu_new",
+        flash_attn: bool = False,
+        flash_rotary: bool = False,
+        fused_dense: bool = False,
+        attn_pdrop: float = 0.0,
+        embd_pdrop: float = 0.0,
+        resid_pdrop: float = 0.0,
+        layer_norm_epsilon: float = 1e-5,
+        initializer_range: float = 0.02,
+        tie_word_embeddings: bool = False,
+        pad_vocab_size_multiple: int = 64,
+        **kwargs
+    ) -> None:
+        self.vocab_size = int(math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple)
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_inner = n_inner
+        self.n_head = n_head
+        self.n_head_kv = n_head_kv
+        self.rotary_dim = min(rotary_dim, n_embd // n_head)
+        self.activation_function = activation_function
+        self.flash_attn = flash_attn
+        self.flash_rotary = flash_rotary
+        self.fused_dense = fused_dense
+        self.attn_pdrop = attn_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.39.1"
+}

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d851b99494eb4f9ba1287d12beb21802dc5240d99fa1796233d64846d4723034
+size 4929919720

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:464543635b704b08f03732fddb57a18d19d30dc9763bc02f5d052159fbac268d
+size 4930095280

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de9827163dc420924eb5085ed765b79d0d700bd92c3848a4c06fcc934c6a5d9c
+size 1258754696

model.safetensors.index.json

@@ -0,0 +1,332 @@
+{
+  "metadata": {
+    "total_size": 11118735360
+  },
+  "weight_map": {
+    "lm_head.linear.bias": "model-00003-of-00003.safetensors",
+    "lm_head.linear.weight": "model-00003-of-00003.safetensors",
+    "lm_head.ln.bias": "model-00003-of-00003.safetensors",
+    "lm_head.ln.weight": "model-00003-of-00003.safetensors",
+    "transformer.embd.wte.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.0.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.0.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.0.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.1.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.1.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.1.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.10.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.10.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.10.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.11.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.11.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.11.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.12.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.12.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.12.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.13.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.13.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.13.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.14.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.14.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.14.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.14.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.15.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.15.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.15.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.16.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.16.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.16.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.17.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.17.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.17.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.18.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.18.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.18.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.19.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.19.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.19.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.2.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.2.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.2.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.20.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.20.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.20.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.21.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.21.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.21.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.22.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.22.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.22.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.23.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.23.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.23.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.24.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.24.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.24.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.25.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.25.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.25.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.26.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.26.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.26.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.27.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.27.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.27.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.28.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.28.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mlp.fc2.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.28.mlp.fc2.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.29.ln.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.29.ln.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mixer.Wqkv.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mixer.Wqkv.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mixer.out_proj.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mixer.out_proj.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mlp.fc1.bias": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mlp.fc1.weight": "model-00002-of-00003.safetensors",
+    "transformer.h.29.mlp.fc2.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.29.mlp.fc2.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.3.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.3.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.3.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.30.ln.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.30.ln.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mixer.Wqkv.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mixer.Wqkv.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mixer.out_proj.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mixer.out_proj.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mlp.fc1.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mlp.fc1.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mlp.fc2.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.30.mlp.fc2.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.31.ln.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.31.ln.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mixer.Wqkv.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mixer.Wqkv.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mixer.out_proj.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mixer.out_proj.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mlp.fc1.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mlp.fc1.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mlp.fc2.bias": "model-00003-of-00003.safetensors",
+    "transformer.h.31.mlp.fc2.weight": "model-00003-of-00003.safetensors",
+    "transformer.h.4.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.4.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.4.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.5.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.5.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.5.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.6.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.6.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.6.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.7.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.7.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.7.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.8.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.8.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.8.mlp.fc2.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.9.ln.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.9.ln.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mixer.Wqkv.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mixer.Wqkv.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mixer.out_proj.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mixer.out_proj.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mlp.fc1.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mlp.fc1.weight": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mlp.fc2.bias": "model-00001-of-00003.safetensors",
+    "transformer.h.9.mlp.fc2.weight": "model-00001-of-00003.safetensors"
+  }
+}

modeling_phi.py

@@ -0,0 +1,960 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+#
+# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
+# Licensed under the BSD 3-Clause License.
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass, field
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .configuration_phi import PhiConfig
+
+try:
+    from flash_attn.bert_padding import pad_input, unpad_input
+    from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
+    from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
+    from flash_attn.ops.fused_dense import FusedDense
+except:
+    pad_input, unpad_input = None, None
+    FlashRotaryEmbedding = None
+    FlashSelfAttention, FlashCrossAttention = None, None
+    FusedDense = None
+
+
+@dataclass
+class InferenceParams:
+    """Inference parameters passed to model to efficiently calculate
+    and store context during inference.
+
+    Reference:
+        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
+
+    Args:
+        max_seqlen: Maximum sequence length.
+        max_batch_size: Maximum batch size.
+        seqlen_offset: Sequence length offset.
+        batch_size_offset: Batch size offset.
+        key_value_memory_dict: Key value memory dictionary.
+        lengths_per_sample: Lengths per sample.
+
+    """
+
+    max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
+
+    max_batch_size: int = field(metadata={"help": "Maximum batch size."})
+
+    seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
+
+    batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
+
+    key_value_memory_dict: Dict[str, Any] = field(
+        default_factory=dict, metadata={"help": "Key value memory dictionary."}
+    )
+
+    lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
+
+
+class Embedding(nn.Module):
+    """Token embedding with dropout."""
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+
+    def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.wte(input_ids)
+        hidden_states = self.drop(hidden_states)
+
+        return hidden_states
+
+
+def _apply_rotary_emb(
+    x: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+) -> torch.FloatTensor:
+    _, seqlen, _, _ = x.shape
+    _, rotary_dim = cos.shape
+    rotary_dim *= 2
+
+    x_rot = x[:, :, :, :rotary_dim]
+    x_pass = x[:, :, :, rotary_dim:]
+
+    x1, x2 = x_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]
+
+    x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)
+
+    return torch.cat([x_rot, x_pass], axis=-1)
+
+
+def _apply_rotary_emb_kv(
+    kv: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+    cos_k: Optional[torch.FloatTensor] = None,
+    sin_k: Optional[torch.FloatTensor] = None,
+) -> torch.FloatTensor:
+    _, seqlen, _, _, _ = kv.shape
+    _, rotary_dim = cos.shape
+    rotary_dim *= 2
+
+    k_rot = kv[:, :, 0, :, :rotary_dim]
+    k_pass = kv[:, :, 0, :, rotary_dim:]
+
+    k1, k2 = k_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]
+
+    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)
+
+    return torch.cat(
+        [
+            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+            kv[:, :, 1:2, :, :],
+        ],
+        axis=2,
+    )
+
+
+def _apply_rotary_emb_qkv(
+    qkv: torch.FloatTensor,
+    cos: torch.FloatTensor,
+    sin: torch.FloatTensor,
+    cos_k: Optional[torch.FloatTensor] = None,
+    sin_k: Optional[torch.FloatTensor] = None,
+) -> torch.FloatTensor:
+    _, seqlen, _, _, _ = qkv.shape
+    _, rotary_dim = cos.shape
+    rotary_dim *= 2
+
+    q_rot = qkv[:, :, 0, :, :rotary_dim]
+    q_pass = qkv[:, :, 0, :, rotary_dim:]
+
+    k_rot = qkv[:, :, 1, :, :rotary_dim]
+    k_pass = qkv[:, :, 1, :, rotary_dim:]
+
+    q1, q2 = q_rot.chunk(2, dim=-1)
+    k1, k2 = k_rot.chunk(2, dim=-1)
+    c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
+    q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
+
+    q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
+    k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
+
+    return torch.cat(
+        [
+            torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
+            torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
+            qkv[:, :, 2:3, :, :],
+        ],
+        axis=2,
+    )
+
+
+class RotaryEmbedding(nn.Module):
+    """Rotary positional embedding (RoPE).
+
+    Reference:
+        RoFormer: Enhanced Transformer with Rotary Position Embedding.
+        https://arxiv.org/pdf/2104.09864.pdf.
+
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        base: int = 10000,
+        scale_base: Optional[float] = None,
+        pos_idx_in_fp32: bool = True,
+        max_position_embeddings: int = 2048,
+        device: Optional[str] = None,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+
+        if scale_base is not None:
+            raise NotImplementedError
+
+        self.dim = dim
+        self.base = float(base)
+        self.scale_base = scale_base
+        self.pos_idx_in_fp32 = pos_idx_in_fp32
+        self.max_position_embeddings = max_position_embeddings
+        self.device = device
+
+        # Generate and save the inverse frequency buffer (non-trainable)
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Generate and save the scale buffer (non-trainable)
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            else None
+        )
+        self.register_buffer("scale", scale, persistent=False)
+
+        # Initialize cached attributes since ONNX can't rely on dynamic initialization
+        self._update_cos_sin_cache(max_position_embeddings, device=device, dtype=torch.float32)
+
+    def _compute_inv_freq(self, device: Optional[str] = None) -> torch.FloatTensor:
+        return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
+
+    def _update_cos_sin_cache(
+        self,
+        seqlen: int,
+        device: Optional[str] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> None:
+        self._seq_len_cached = seqlen
+
+        # fp32 is preferred since the output of `torch.arange` can be quite large
+        # and bf16 would lose a lot of precision
+        if self.pos_idx_in_fp32:
+            t = torch.arange(seqlen, device=device, dtype=torch.float32)
+            if self.inv_freq.dtype != torch.float32:
+                inv_freq = self._compute_inv_freq(device=device)
+            else:
+                inv_freq = self.inv_freq
+        else:
+            t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+            inv_freq = self.inv_freq
+
+        # `torch.outer` is preferred since `torch.einsum` converts from fp32 to fp16 if used with AMP
+        freqs = torch.outer(t, inv_freq)
+        if self.scale is None:
+            self._cos_cached = torch.cos(freqs).to(dtype)
+            self._sin_cached = torch.sin(freqs).to(dtype)
+        else:
+            power = (
+                torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
+            ) / self.scale_base
+            scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
+
+            # Force the scale multiplication to happen in fp32
+            self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
+            self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
+            self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
+            self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
+
+    def forward(
+        self,
+        qkv: torch.Tensor,
+        kv: Optional[torch.Tensor] = None,
+        seqlen_offset: int = 0,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if (
+            self._seq_len_cached < qkv.shape[1] + seqlen_offset
+            or self._cos_cached.device != qkv.device
+            or self._cos_cached.dtype != qkv.dtype
+            or (self.training and self._cos_cached.is_inference())
+        ):
+            self._update_cos_sin_cache(qkv.shape[1] + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
+
+        if kv is None:
+            return _apply_rotary_emb_qkv(
+                qkv,
+                self._cos_cached[seqlen_offset:],
+                self._sin_cached[seqlen_offset:],
+            )
+        else:
+            q = _apply_rotary_emb(
+                qkv,
+                self._cos_cached[seqlen_offset:],
+                self._sin_cached[seqlen_offset:],
+            )
+            kv = _apply_rotary_emb_kv(
+                kv,
+                self._cos_cached[seqlen_offset:],
+                self._sin_cached[seqlen_offset:],
+            )
+
+            return q, kv
+
+
+class MLP(nn.Module):
+    """Multi-Layer Perceptron.
+
+    Reference:
+        Attention Is All You Need.
+        https://arxiv.org/pdf/1706.03762.pdf.
+
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        n_inner: Optional[int] = None,
+        act_fn: Optional[str] = None,
+    ) -> None:
+        super().__init__()
+
+        act_fn = config.activation_function if act_fn is None else act_fn
+
+        n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
+        n_inner = n_inner if n_inner is not None else 4 * config.n_embd
+
+        self.fc1 = nn.Linear(config.n_embd, n_inner)
+        self.fc2 = nn.Linear(n_inner, config.n_embd)
+        self.act = ACT2FN[act_fn]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class SelfAttention(nn.Module):
+    """Self-attention layer (compatible with PyTorch).
+
+    Reference:
+        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
+
+    """
+
+    def __init__(
+        self,
+        causal: bool = True,
+        softmax_scale: Optional[float] = None,
+        attention_dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    @torch.autocast("cpu", enabled=False)
+    @torch.autocast("cuda", enabled=False)
+    def forward(
+        self,
+        qkv: torch.FloatTensor,
+        causal: bool = None,
+        key_padding_mask: Optional[torch.BoolTensor] = None,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
+        q, k, v = qkv.unbind(dim=2)
+
+        q = q.to(torch.float32)
+        k = k.to(torch.float32)
+
+        causal = self.causal if causal is None else causal
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+
+        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
+        # using float16, which might lead to overflow
+        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
+
+        if key_padding_mask is not None:
+            padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+
+            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+
+        if causal:
+            causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+
+        attention = torch.softmax(scores, dim=-1).to(v.dtype)
+        attention = self.drop(attention)
+
+        output = torch.einsum("bhts,bshd->bthd", attention, v)
+
+        return output
+
+
+class CrossAttention(nn.Module):
+    """Cross-attention layer (compatible with PyTorch).
+
+    Reference:
+        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
+
+    """
+
+    def __init__(
+        self,
+        causal: bool = True,
+        softmax_scale: Optional[float] = None,
+        attention_dropout: float = 0.0,
+    ) -> None:
+        super().__init__()
+
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    @torch.autocast("cpu", enabled=False)
+    @torch.autocast("cuda", enabled=False)
+    def forward(
+        self,
+        q: torch.FloatTensor,
+        kv: torch.FloatTensor,
+        causal: bool = None,
+        key_padding_mask: Optional[torch.BoolTensor] = None,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = kv.shape[1]
+
+        if kv.shape[3] != q.shape[2]:
+            kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
+        k, v = kv.unbind(dim=2)
+
+        q = q.to(torch.float32)
+        k = k.to(torch.float32)
+
+        causal = self.causal if causal is None else causal
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+
+        # Autocast is manually disabled to avoid `torch.einsum` performing the operation
+        # using float16, which might lead to overflow
+        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
+
+        if key_padding_mask is not None:
+            padding_mask = torch.full(
+                (batch_size, seqlen_k),
+                -10000.0,
+                dtype=scores.dtype,
+                device=scores.device,
+            )
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+
+            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+
+        if causal:
+            rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
+            cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
+            causal_mask = cols > rows + seqlen_k - seqlen_q
+
+            scores = scores.masked_fill(causal_mask, -10000.0)
+
+        attention = torch.softmax(scores, dim=-1).to(v.dtype)
+        attention = self.drop(attention)
+
+        output = torch.einsum("bhts,bshd->bthd", attention, v)
+
+        return output
+
+
+def _find_mha_dims(
+    config: PretrainedConfig,
+    n_head: Optional[int] = None,
+    n_head_kv: Optional[int] = None,
+    head_dim: Optional[int] = None,
+) -> Tuple[int, int]:
+    if n_head is None and head_dim is None:
+        head_dim = config.n_embd // config.n_head
+        n_head = config.n_head
+    elif n_head is None or head_dim is None:
+        raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
+
+    if n_head_kv is None:
+        n_head_kv = getattr(config, "n_head_kv", None) or n_head
+
+    return n_head, n_head_kv, head_dim
+
+
+def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
+    num_heads, head_dim = kv.shape[-2:]
+
+    if layer_idx not in inference_params.key_value_memory_dict:
+        inference_params.key_value_memory_dict[layer_idx] = torch.empty(
+            inference_params.max_batch_size,
+            inference_params.max_seqlen,
+            2,
+            num_heads,
+            head_dim,
+            dtype=kv.dtype,
+            device=kv.device,
+        )
+
+    batch_start = inference_params.batch_size_offset
+    batch_end = batch_start + kv.shape[0]
+
+    sequence_start = inference_params.seqlen_offset
+    sequence_end = sequence_start + kv.shape[1]
+
+    # When the current sequence length is equal to or larger than the maximum sequence length,
+    # we need to concatenate the current `kv` with the cached `kv` to expand its length
+    if sequence_end >= inference_params.max_seqlen:
+        inference_params.key_value_memory_dict[layer_idx] = torch.concatenate((inference_params.key_value_memory_dict[layer_idx], kv), dim=1)
+
+    inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+    kv = inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, :sequence_end, ...]
+        
+    return kv
+
+
+class MHA(nn.Module):
+    """Multi-head attention layer."""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[str] = None,
+        rotary_dim: Optional[int] = None,
+        rotary_base: float = 10000.0,
+        rotary_scale_base: Optional[float] = None,
+        n_head: Optional[int] = None,
+        n_head_kv: Optional[int] = None,
+        head_dim: Optional[int] = None,
+        bias: bool = True,
+        causal: bool = True,
+        softmax_scale: Optional[float] = None,
+        layer_idx: Optional[int] = None,
+        return_residual: bool = False,
+        checkpointing: bool = False,
+    ) -> None:
+        super().__init__()
+
+        # Rotary embedding
+        self.rotary_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
+        if self.rotary_dim > 0:
+            rotary_cls = FlashRotaryEmbedding if config.flash_rotary else RotaryEmbedding
+            if rotary_cls is None:
+                rotary_cls = RotaryEmbedding
+
+            rotary_kwargs = {}
+            if rotary_cls is RotaryEmbedding:
+                rotary_kwargs["max_position_embeddings"] = config.n_positions
+
+            self.rotary_emb = rotary_cls(
+                self.rotary_dim,
+                base=rotary_base,
+                scale_base=rotary_scale_base,
+                device=device,
+                **rotary_kwargs,
+            )
+
+        # MLP
+        self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(
+            config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim
+        )
+        op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
+        hidden_size = config.n_embd
+
+        linear_cls = FusedDense if config.fused_dense else nn.Linear
+        if linear_cls is None:
+            linear_cls = nn.Linear
+
+        self.Wqkv = linear_cls(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
+        self.out_proj = linear_cls(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)
+
+        # Attention
+        attn_cls = FlashSelfAttention if config.flash_attn else SelfAttention
+        if attn_cls is None:
+            attn_cls = SelfAttention
+
+        cross_attn_cls = FlashCrossAttention if config.flash_attn else CrossAttention
+        if cross_attn_cls is None:
+            cross_attn_cls = CrossAttention
+
+        self.inner_attn = attn_cls(
+            causal=causal,
+            softmax_scale=softmax_scale,
+            attention_dropout=config.attn_pdrop,
+        )
+        self.inner_cross_attn = cross_attn_cls(
+            causal=causal,
+            softmax_scale=softmax_scale,
+            attention_dropout=config.attn_pdrop,
+        )
+
+        self.flash_attn = config.flash_attn and attn_cls is FlashSelfAttention
+        self.layer_idx = layer_idx
+        self.return_residual = return_residual
+        self.checkpointing = checkpointing
+
+    def _forward_self_attn(
+        self, x: torch.FloatTensor, key_padding_mask: Optional[torch.BoolTensor]
+    ) -> torch.FloatTensor:
+        qkv = self.Wqkv(x)
+        qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+
+        if self.rotary_dim > 0:
+            qkv = self.rotary_emb(qkv)
+
+        if self.flash_attn:
+            batch_size, seqlen = qkv.shape[0], qkv.shape[1]
+
+            cu_seqlens, max_seqlen = None, None
+            if key_padding_mask is not None:
+                # If `key_padding_mask` is supplied, we need to unpad the input and retrieve
+                # the `cu_seqlens` and `max_seqlen` to be used by `flash-attn`
+                qkv, indices, cu_seqlens, max_seqlen = unpad_input(qkv, key_padding_mask)
+
+            if self.checkpointing:
+                attn_output = torch.utils.checkpoint.checkpoint(
+                    self.inner_attn, qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen
+                )
+            else:
+                attn_output = self.inner_attn(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen).to(qkv.device)
+
+            # If `key_padding_mask` is supplied, we need to pad the output back to the original shape
+            return pad_input(attn_output, indices, batch_size, seqlen) if key_padding_mask is not None else attn_output
+
+        if self.checkpointing:
+            return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, key_padding_mask=key_padding_mask)
+
+        return self.inner_attn(qkv, key_padding_mask=key_padding_mask)
+
+    def _forward_cross_attn(
+        self,
+        x: torch.FloatTensor,
+        past_key_values: Optional[InferenceParams],
+        key_padding_mask: Optional[torch.BoolTensor],
+    ) -> torch.FloatTensor:
+        batch_size = x.shape[0]
+
+        qkv = self.Wqkv(x)
+
+        q = qkv[..., : self.n_head * self.head_dim]
+        q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
+
+        kv = qkv[..., self.n_head * self.head_dim :]
+        kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
+
+        seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
+        causal = None if seqlen_offset == 0 else False
+        if self.rotary_dim > 0:
+            q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)
+
+        if past_key_values is not None:
+            kv = _update_kv_cache(kv, past_key_values, self.layer_idx)
+
+        if self.flash_attn:
+            batch_size, seqlen_q = q.shape[0], q.shape[1]
+            seqlen_k = kv.shape[1]
+
+            cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k = (
+                None,
+                None,
+                None,
+                None,
+            )
+            if key_padding_mask is not None:
+                kv, _, cu_seqlens_k, max_seqlen_k = unpad_input(kv, key_padding_mask)
+
+                if seqlen_q == 1:
+                    key_padding_mask = torch.ones(batch_size, 1, device=q.device)
+                elif seqlen_q != seqlen_k:
+                    key_padding_mask = key_padding_mask[:, -seqlen_q:]
+
+                q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, key_padding_mask)
+
+            if self.checkpointing:
+                attn_output = torch.utils.checkpoint.checkpoint(
+                    self.inner_cross_attn,
+                    q,
+                    kv,
+                    causal=causal,
+                    cu_seqlens=cu_seqlens_q,
+                    max_seqlen=max_seqlen_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_k=max_seqlen_k,
+                )
+            else:
+                attn_output = self.inner_cross_attn(
+                    q,
+                    kv,
+                    causal=causal,
+                    cu_seqlens=cu_seqlens_q,
+                    max_seqlen=max_seqlen_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_k=max_seqlen_k,
+                )
+
+            return (
+                pad_input(attn_output, indices_q, batch_size, max_seqlen_q)
+                if key_padding_mask is not None
+                else attn_output
+            )
+
+        if self.checkpointing:
+            return torch.utils.checkpoint.checkpoint(
+                self.inner_cross_attn,
+                q,
+                kv,
+                key_padding_mask=key_padding_mask,
+                causal=causal,
+            )
+
+        return self.inner_cross_attn(q, kv, key_padding_mask=key_padding_mask, causal=causal)
+
+    def forward(
+        self,
+        x: torch.FloatTensor,
+        past_key_values: Optional[InferenceParams] = None,
+        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        if attention_mask is not None:
+            attention_mask = attention_mask.bool()
+        else:
+            attention_mask = None
+
+        # MHA
+        if self.n_head == self.n_head_kv:
+            if past_key_values is None:
+                # If `past_key_values` are not supplied, we run self-attention
+                attn_output = self._forward_self_attn(x, attention_mask)
+            else:
+                # If `past_key_values` are supplied, it means that we might have cached values and
+                # could take advantage of cross-attention
+                attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
+        # MQA / GQA
+        else:
+            # Regardless of `past_key_values` being supplied or not, it always use cross-attention
+            # because `q` and `kv` lengths might be different
+            attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
+
+        output = rearrange(attn_output, "... h d -> ... (h d)")
+        output = self.out_proj(output)
+
+        return output if not self.return_residual else (output, x)
+
+
+class ParallelBlock(nn.Module):
+    """Parallel block.
+
+    This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
+
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        block_idx: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.block_idx = block_idx
+
+        self.mixer = MHA(config, layer_idx=block_idx)
+        self.mlp = MLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
+        attention_mask: Optional[torch.BoolTensor] = None,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+        hidden_states = self.ln(hidden_states)
+
+        attn_outputs = self.mixer(
+            hidden_states,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+        )
+        if isinstance(attn_outputs, tuple):
+            attn_outputs = attn_outputs[0]
+
+        attn_outputs = self.resid_dropout(attn_outputs)
+        feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
+
+        hidden_states = attn_outputs + feed_forward_hidden_states + residual
+
+        return hidden_states
+
+
+class CausalLMHead(nn.Module):
+    """Causal Language Modeling head.
+
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+
+    """
+
+    def __init__(self, config: PretrainedConfig) -> None:
+        super().__init__()
+
+        self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.linear = nn.Linear(config.n_embd, config.vocab_size)
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        hidden_states = self.ln(hidden_states)
+        logits = self.linear(hidden_states).to(torch.float32)
+
+        return logits
+
+
+class CausalLMLoss(nn.Module):
+    """Causal Language Modeling loss.
+
+    Reference:
+        Improving Language Understanding by Generative Pre-Training.
+        https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
+
+    """
+
+    def __init__(self, shift_labels: bool = True) -> None:
+        super().__init__()
+
+        self.shift_labels = shift_labels
+        self.loss_fct = nn.CrossEntropyLoss()
+
+    def forward(self, logits: torch.FloatTensor, labels: torch.LongTensor) -> torch.FloatTensor:
+        if self.shift_labels:
+            logits = logits[..., :-1, :].contiguous()
+            labels = labels[..., 1:].contiguous()
+
+        loss = self.loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1))
+
+        return loss
+
+
+class PhiPreTrainedModel(PreTrainedModel):
+    """Phi pre-trained model."""
+
+    config_class = PhiConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["ParallelBlock"]
+
+    def __init__(self, *inputs, **kwargs) -> None:
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module) -> None:
+        if isinstance(module, (nn.Linear,)):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
+        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
+        **kwargs,
+    ) -> Dict[str, Any]:
+        if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
+            past_key_values = InferenceParams(
+                max_seqlen=self.config.n_positions,
+                max_batch_size=input_ids.shape[0],
+                seqlen_offset=0,
+                batch_size_offset=0,
+                key_value_memory_dict={},
+                lengths_per_sample=None,
+            )
+        else:
+            # Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
+            past_key_values.seqlen_offset = input_ids.shape[1] - 1
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past_key_values,
+            "attention_mask": attention_mask,
+        }
+
+
+class PhiModel(PhiPreTrainedModel):
+    """Phi model."""
+
+    _keys_to_ignore_on_load_missing = [""]
+    _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+
+    def __init__(self, config: PhiConfig) -> None:
+        super().__init__(config)
+
+        self.embd = Embedding(config)
+        self.h = nn.ModuleList([ParallelBlock(config, block_idx=i) for i in range(config.n_layer)])
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embd.wte
+
+    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
+        self.embd.wte = new_embeddings
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
+        attention_mask: Optional[torch.BoolTensor] = None,
+    ) -> torch.FloatTensor:
+        hidden_states = self.embd(input_ids)
+
+        for layer in self.h:
+            hidden_states = layer(
+                hidden_states,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask,
+            )
+
+        return hidden_states
+
+
+class PhiForCausalLM(PhiPreTrainedModel):
+    """Phi for Causal Language Modeling."""
+
+    _keys_to_ignore_on_load_missing = [""]
+    _keys_to_ignore_on_load_unexpected = [r"transformer\.h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
+
+    def __init__(self, config: PhiConfig) -> None:
+        super().__init__(config)
+
+        self.transformer = PhiModel(config)
+        self.lm_head = CausalLMHead(config)
+        self.loss = CausalLMLoss()
+
+        self.post_init()
+
+    def get_output_embeddings(self) -> nn.Linear:
+        return self.lm_head.linear
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
+        self.lm_head.linear = new_embeddings
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
+        attention_mask: Optional[torch.BoolTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> CausalLMOutputWithPast:
+        hidden_states = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask)
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss(lm_logits, labels)
+
+        return CausalLMOutputWithPast(loss=loss, logits=lm_logits, past_key_values=past_key_values)