Apply suggestions from code review

Co-authored-by: Sigbjørn Skjæret <[email protected]>

Author: pwilkin

convert_hf_to_gguf.py

@@ -2781,7 +2781,8 @@ def set_gguf_parameters(self):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
         num_kv_heads = self.hparams["num_key_value_heads"]
-        head_dim = self.hparams["hidden_size"] // num_heads
+        if (head_dim := self.hparams.get("head_dim")) is None:
+            head_dim = self.hparams["hidden_size"] // num_heads
 
         if "ernie." in name:
             name = name.replace("ernie.", "model.")
@@ -2834,11 +2835,6 @@ def set_gguf_parameters(self):
         if (shared_expert_intermediate_size := self.hparams.get('intermediate_size')) is not None and (num_key_value_heads := self.hparams.get('num_key_value_heads')) is not None:
             self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size // num_key_value_heads)
 
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        if "exps" in new_name:
-            return gguf.GGMLQuantizationType.F16
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # Modify correction bias name as in DeepseekV2
         if name.endswith("e_score_correction_bias"):
@@ -2863,7 +2859,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             return []
 
         # process the experts separately
-        if name.find("experts.") != -1 and name.find("shared") == -1:
+        if name.find("mlp.experts") != -1:
             n_experts = self.hparams["moe_num_experts"]
             assert bid is not None
 

gguf-py/gguf/constants.py

@@ -678,7 +678,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
     MODEL_ARCH.ERNIE4_5:         "ernie4_5",
-    MODEL_ARCH.ERNIE4_5_MOE:     "ernie4_5_moe",
+    MODEL_ARCH.ERNIE4_5_MOE:     "ernie4_5-moe",
     MODEL_ARCH.FALCON_H1:        "falcon-h1",
     MODEL_ARCH.HUNYUAN_MOE:      "hunyuan-moe",
     MODEL_ARCH.SMOLLM3:          "smollm3",

src/llama-arch.cpp

@@ -81,7 +81,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
-    { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5_moe"     },
+    { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
     { LLM_ARCH_SMOLLM3,          "smollm3"          },
     { LLM_ARCH_LFM2,             "lfm2"             },

src/llama-model.cpp

@@ -8365,6 +8365,7 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
+        GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
             // norm
@@ -8403,17 +8404,15 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
-                const float freq_base_l  = model.get_rope_freq_base (cparams, il);
-                const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );
 
                 Kcur = ggml_rope_ext(
                         ctx0, Kcur, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );