add telechat 2.5

Author: foldl

convert.py

@@ -1824,7 +1824,7 @@ def state_dict_pp(cls, config, state_dict):
     def dump_config(f, config, ggml_type):
         assert config.hidden_act is None, 'hidden_act must be None'
         assert not config.tie_word_embeddings, 'tie_word_embeddings must be False'
-        assert config.embed_layernorm is None, 'embed_layernorm must be None'
+        assert (config.embed_layernorm is None) or not config.embed_layernorm, 'embed_layernorm must be None'
         assert not config.apply_residual_connection_post_layernorm, 'apply_residual_connection_post_layernorm must be False'
         assert config.training_seqlen == config.base_seqlen, 'training_seqlen must equal to base_seqlen'
 

docs/models.md

@@ -245,6 +245,7 @@
 
 * TeleChat (`TeleChat2ForCausalLM`)
     * [x] v2: [3B](https://huggingface.co/Tele-AI/TeleChat2-3B), [7B](https://huggingface.co/Tele-AI/TeleChat2-7B), [115B](https://huggingface.co/Tele-AI/TeleChat2-115B)
+    * [x] v2.5 [35B](https://huggingface.co/Tele-AI/TeleChat2.5-35B/commit/e53676611f3c5072f7696a359132eaf456272151), [115B](https://huggingface.co/Tele-AI/TeleChat2-115B/tree/8be654fe28bfe60fca4cd483297167a6e570f93b)
 
 * XVERSE (`XverseForCausalLM`)
     * [x] [Chat-7B](https://huggingface.co/xverse/XVERSE-7B-Chat), [Chat-13B](https://huggingface.co/xverse/XVERSE-13B-Chat), [Chat-65B](https://huggingface.co/xverse/XVERSE-65B-Chat)

scripts/models.json

@@ -1779,6 +1779,22 @@
             }
         }
     },
+    "telechat2.5": {
+        "brief": "TeleChat2.5 is a large language model trained by the Artificial Intelligence Research Institute of China Telecom.",
+        "default": "35b",
+        "license": "Apache License 2.0",
+        "variants": {
+            "35b": {
+                "default": "q4_0",
+                "quantized": {
+                    "q8": {
+                        "size": 19939670112,
+                        "url": "chatllm_quantized_telechat/telechat2.5-35b-q4_0.bin"
+                    }
+                }
+            }
+        }
+    },
     "alphageometry-lm": {
         "brief": "TeleChat2 is a large language model trained by the Artificial Intelligence Research Institute of China Telecom.",
         "default": "0.2b",