Consolidate support for Phi-1, Phi-1.5, and Phi-2 models

.gitignore

@@ -31,6 +31,7 @@ build*/
 out/
 tmp/
 
+local/
 models/*
 models-mnt
 

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Optional, cast
 
 import numpy as np
 import torch
@@ -22,7 +22,6 @@
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -183,7 +182,7 @@ def from_model_architecture(model_architecture):
         if model_architecture == "MixtralForCausalLM":
             return MixtralModel
         if model_architecture == "PhiForCausalLM":
-            return Phi2Model
+            return PhiModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -224,7 +223,7 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
         if arch == "MixtralForCausalLM":
             return gguf.MODEL_ARCH.LLAMA
         if arch == "PhiForCausalLM":
-            return gguf.MODEL_ARCH.PHI2
+            return gguf.MODEL_ARCH.PHI
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -985,11 +984,11 @@ def write_tensors(self):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-class Phi2Model(Model):
+class PhiModel(Model):
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
-        self.gguf_writer.add_name("Phi2")
+        self.gguf_writer.add_name("Phi")
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])

gguf-py/gguf/constants.py

@@ -95,7 +95,7 @@ class MODEL_ARCH(IntEnum):
     BLOOM     = auto()
     STABLELM  = auto()
     QWEN      = auto()
-    PHI2      = auto()
+    PHI       = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -141,7 +141,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
-    MODEL_ARCH.PHI2:           "phi2",
+    MODEL_ARCH.PHI:            "phi",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -352,7 +352,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.GPT2: [
         # TODO
     ],
-    MODEL_ARCH.PHI2: [
+    MODEL_ARCH.PHI: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,

gguf-py/gguf/tensor_mapping.py

@@ -17,7 +17,7 @@ class TensorNameMap:
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert
             "language_model.embedding.word_embeddings",  # persimmon
-            "transformer.embd.wte",                      # phi2
+            "transformer.embd.wte",                      # phi1 phi1_5 phi2
         ),
 
         # Token type embeddings
@@ -42,7 +42,7 @@ class TensorNameMap:
             "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen
             "output",                    # llama-pth bloom
             "word_embeddings_for_head",  # persimmon
-            "lm_head.linear",            # phi2
+            "lm_head.linear",            # phi1 phi1_5 phi2
         ),
 
         # Output norm
@@ -55,7 +55,7 @@ class TensorNameMap:
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen
             "language_model.encoder.final_layernorm",  # persimmon
-            "lm_head.ln",                              # phi2
+            "lm_head.ln",                              # phi1 phi1_5 phi2
         ),
 
         # Rope frequencies
@@ -78,7 +78,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.LayerNorm",       # bert
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
-            "transformer.h.{bid}.ln",                               # phi2
+            "transformer.h.{bid}.ln",                               # phi1 phi1_5 phi2
         ),
 
         # Attention norm 2
@@ -94,7 +94,7 @@ class TensorNameMap:
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
-            "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
+            "transformer.h.{bid}.mixer.Wqkv",                                      # phi1 phi1_5 phi2
         ),
 
         # Attention query
@@ -133,7 +133,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.dense",                # bert
             "transformer.h.{bid}.attn.out_proj",                         # gpt-j
             "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
-            "transformer.h.{bid}.mixer.out_proj",                        # phi2
+            "transformer.h.{bid}.mixer.out_proj",                        # phi1 phi1_5 phi2
         ),
 
         # Rotary embeddings
@@ -173,7 +173,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
             "transformer.h.{bid}.mlp.w1",                             # qwen
-            "transformer.h.{bid}.mlp.fc1",                            # phi2
+            "transformer.h.{bid}.mlp.fc1",                            # phi1 phi1_5 phi2
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -205,7 +205,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
-            "transformer.h.{bid}.mlp.fc2",                            # phi2
+            "transformer.h.{bid}.mlp.fc2",                            # phi1 phi1_5 phi2
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (

llama.cpp

@@ -197,7 +197,7 @@ enum llm_arch {
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
-    LLM_ARCH_PHI2,
+    LLM_ARCH_PHI,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -215,7 +215,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
     { LLM_ARCH_BLOOM,           "bloom"     },
     { LLM_ARCH_STABLELM,        "stablelm"  },
     { LLM_ARCH_QWEN,            "qwen"      },
-    { LLM_ARCH_PHI2,            "phi2"      },
+    { LLM_ARCH_PHI,             "phi"      },
 };
 
 enum llm_kv {
@@ -555,7 +555,7 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
         },
     },
     {
-        LLM_ARCH_PHI2,
+        LLM_ARCH_PHI,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
@@ -2740,11 +2740,12 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_PHI2:
+        case LLM_ARCH_PHI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
@@ -3579,7 +3580,7 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
                     }
                 } break;
-            case LLM_ARCH_PHI2:
+            case LLM_ARCH_PHI:
                 {
                     model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
 
@@ -4079,7 +4080,7 @@ static struct ggml_tensor * llm_build_kqv(
     struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
     cb(kq, "kq", il);
 
-    if (model.arch == LLM_ARCH_PHI2) {
+    if (model.arch == LLM_ARCH_PHI) {
         // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
         // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
         ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -5445,7 +5446,7 @@ struct llm_build_context {
 
         return gf;
     }
-    struct ggml_cgraph * build_phi2() {
+    struct ggml_cgraph * build_phi() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         struct ggml_tensor * cur;
@@ -6061,9 +6062,9 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_qwen();
             } break;
-        case LLM_ARCH_PHI2:
+        case LLM_ARCH_PHI:
             {
-                result = llm.build_phi2();
+                result = llm.build_phi();
             } break;
         default:
             GGML_ASSERT(false);