convert : Add support for Microsoft Phi-4 model (#10817)

* convert : use GPT2 vocab for Phi-4 model * convert : use null value of sliding_window to distinguish Phi-4 from other PHI3-based models * llama : do not use sliding window attention mask for Phi-4 model --------- Co-authored-by: Stanisław Szymczyk <[email protected]>
ggerganov · Dec 19, 2024 · 7585edb · 7585edb
1 parent cd920d0
commit 7585edb
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 3 deletions.
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
@@ -2200,6 +2200,15 @@ class Phi3MiniModel(Model):
     model_arch = gguf.MODEL_ARCH.PHI3
 
     def set_vocab(self):
+        # Phi-4 model uses GPT2Tokenizer
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                tokenizer_class = tokenizer_config_json['tokenizer_class']
+                if tokenizer_class == 'GPT2Tokenizer':
+                    return self._set_vocab_gpt2()
+
         from sentencepiece import SentencePieceProcessor
 
         tokenizer_path = self.dir_model / 'tokenizer.model'
@@ -2316,7 +2325,11 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_rope_dimension_count(rope_dims)
         self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
         self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
+        sliding_window = self.hparams.get("sliding_window")
+        # use zero value of sliding_window to distinguish Phi-4 from other PHI3 models
+        if sliding_window is None:
+            sliding_window = 0
+        self.gguf_writer.add_sliding_window(sliding_window)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         n_embd = self.find_hparam(["hidden_size", "n_embd"])

diff --git a/src/llama.cpp b/src/llama.cpp
@@ -13333,7 +13333,13 @@ struct llm_build_context {
         struct ggml_tensor * inp_pos = build_inp_pos();
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+        struct ggml_tensor * KQ_mask = nullptr;
+        if (hparams.n_swa == 0) {
+            // Phi-4 doesn't use sliding window attention
+            KQ_mask = build_inp_KQ_mask();
+        } else {
+            KQ_mask = build_inp_KQ_mask_swa();
+        }
 
         for (int il = 0; il < n_layer; ++il) {
             auto residual = inpL;
@@ -13391,7 +13397,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {