support ling-lite-1.5 & ring-lite

Author: foldl

convert.py

@@ -6325,7 +6325,8 @@ def state_dict_pp(cls, config, state_dict):
             if name == 'model.word_embeddings.weight':
                 r['model.embed_tokens.weight'] = tensor
             elif name == "lm_head.weight":
-                tensor = tensor / (torch.norm(tensor, p=2, dim=0, keepdim=True) + 1e-7)
+                if config.norm_head:
+                    tensor = tensor / (torch.norm(tensor, p=2, dim=0, keepdim=True) + 1e-7)
                 r[name] = tensor
             elif name.endswith('query_key_value.weight'):
                 head_dim = config.head_dim
@@ -6346,7 +6347,6 @@ def state_dict_pp(cls, config, state_dict):
 
     @staticmethod
     def dump_config(f, config, ggml_type):
-        assert config.norm_head
         if config.moe_layer_freq is None: config.moe_layer_freq = 1
         if config.attention_bias is None: config.attention_bias = False
         if config.scoring_func is None: config.scoring_func = 'softmax'

docs/models.md

@@ -112,8 +112,9 @@
     [Coder-8B-Instruct](https://huggingface.co/JT-LM/JT-Coder-8B-Instruct/tree/9160d51e9acaae266cfef8493ea25d15e7ed6904),
     [DA-8B](https://huggingface.co/JT-LM/JT-DA-8B/commit/8bd5bb1a76305dcc777786b65c239b362cee808e)
 
-* Ling (`BailingMoeForCausalLM`)
+* Ling/Ring (`BailingMoeForCausalLM`)
     * [x] [Lite](https://huggingface.co/inclusionAI/Ling-lite/tree/a80ae6c479251f1ae33dda517ab83cdc6a312f99), [Coder-Lite](https://huggingface.co/inclusionAI/Ling-Coder-lite/tree/4a8647acf9d3855d599adaaaf4bf6ca14239d2ab)
+    * [x] v1.5: [Ling-lite-1.5-2507](https://huggingface.co/inclusionAI/Ling-lite-1.5-2507/tree/6656efdc763a77102207fc66b176e4c5d07a316b), [Ring-lite2507](https://huggingface.co/inclusionAI/Ring-lite-2507/commit/8cf0ec244871c90102b353cef3568e061fd2504f)
 
 * LlaMA-like (`LlamaForCausalLM`, `Llama4ForConditionalGeneration`):
     * [x] All LlaMA-1 models

models/bailing.cpp

@@ -23,7 +23,7 @@ namespace chatllm::bailing::moe
         Tokenizer(const Config &config)
             : BaseTokenizer(config, &_chat_encoder)
         {
-            sys_prompt = "";
+            sys_prompt = "You are Ling, an assistant created by inclusionAI";
         }
 
         size_t load(tokenizer::DataReader *buffer, int n_vocab) override
@@ -47,6 +47,16 @@ namespace chatllm::bailing::moe
             if (role_open_token_id >= 0)
                 terminate_ids.insert(role_open_token_id);
 
+            int t = tp->PieceToId("<think>");
+            if (t >= 0)
+            {
+                tp->OverrideTokenDecoding(t, "<think>");
+                sys_prompt = "You are Ring, an assistant created by inclusionAI";
+            }
+            t = tp->PieceToId("</think>");
+            if (t >= 0)
+                tp->OverrideTokenDecoding(t, "</think>");
+
             return size;
         }
 

scripts/models.json

@@ -2694,6 +2694,31 @@
                         "url": "chatllm_quantized_bailing/bailing-lite-0415.bin"
                     }
                 }
+            },
+            "lite-1.5-2507": {
+                "default": "q8",
+                "quantized": {
+                    "q8": {
+                        "size": 17856685024,
+                        "url": "chatllm_quantized_bailing/ling-lite-1.5-2507.bin"
+                    }
+                }
+            }
+        }
+    },
+    "ring": {
+        "brief": "Ring-lite is a lightweight, fully open-sourced MoE (Mixture of Experts) LLM designed for complex reasoning tasks.",
+        "default": "lite-2507",
+        "license": "MIT",
+        "variants": {
+            "lite-2507": {
+                "default": "q8",
+                "quantized": {
+                    "q8": {
+                        "size": 17856684480,
+                        "url": "chatllm_quantized_bailing/ring-lite-2507.bin"
+                    }
+                }
             }
         }
     },