gpt oss harmony template

Author: LostRuins

CMakeLists.txt

@@ -61,7 +61,7 @@ add_compile_definitions(LOG_DISABLE_LOGS)
 add_compile_definitions(GGML_USE_CPU)
 add_compile_definitions(GGML_USE_CPU_REPACK)
 add_compile_definitions(NOMINMAX)
-add_compile_definitions(_REGEX_MAX_STACK_COUNT=80000)
+add_compile_definitions(_REGEX_MAX_STACK_COUNT=32000)
 
 if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)

gpttype_adapter.cpp

@@ -2293,15 +2293,21 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         //     std::string forced = "per_layer_token_embd.weight=CPU"; //this tensor on gpu is problematic on unsloth q4_0
         //     tensoroverrides = (tensoroverrides=="" ? forced: (forced+","+tensoroverrides));
         // }
-        if(tensoroverrides=="" && ggml_backend_dev_count()>1 && inputs.moecpu>0)
+        if(ggml_backend_dev_count()>1 && inputs.moecpu>0)
         {
+            std::string toadd = "";
             for (int i = 0; i < inputs.moecpu; ++i) {
                 std::string tmp = string_format("blk\\.%d\\.ffn_(up|down|gate)_exps=CPU", i);
                 if(i>0)
                 {
                     tmp = "," + tmp;
                 }
-                tensoroverrides += tmp;
+                toadd += tmp;
+            }
+            if (tensoroverrides == "") {
+                tensoroverrides = toadd;
+            } else {
+                tensoroverrides += "," + toadd;
             }
             printf("Overriding %d MoE layers to CPU...\n",inputs.moecpu);
         }

kcpp_adapters/AutoGuess.json

@@ -186,6 +186,17 @@
         "assistant_start": "<|response|>",
         "assistant_end": "<|endofresponse|>"
     }
+}, {
+    "search": ["<|start|>user<|message|>", "<|channel|>", "<|end|>"],
+    "name": "OpenAI Harmony",
+    "adapter": {
+        "system_start": "<|start|>system<|message|>",
+        "system_end": "<|end|>\n",
+        "user_start": "<|start|>user<|message|>",
+        "user_end": "<|end|>\n",
+        "assistant_start": "<|start|>assistant<|channel|>final<|message|>",
+        "assistant_end": "<|return|>\n"
+    }
 }, {
     "search": ["rwkv_", "'User: '"],
     "name": "RWKV World",

kcpp_adapters/OpenAI-Harmony.json

@@ -0,0 +1,8 @@
+{
+  "system_start": "<|start|>system<|message|>",
+  "system_end": "<|end|>\n",
+  "user_start": "<|start|>user<|message|>",
+  "user_end": "<|end|>\n",
+  "assistant_start": "<|start|>assistant<|channel|>final<|message|>",
+  "assistant_end": "<|return|>\n"
+}

klite.embd

@@ -3629,6 +3629,15 @@ Current version indicated by LITEVER below.
 		"assistant_end":"",
 		"system":"",
 		"system_end":"",
+	},
+	{
+		"name":"OpenAI Harmony",
+		"user":"<|start|>user<|message|>",
+		"user_end":"<|end|>\\n",
+		"assistant":"<|start|>assistant<|channel|>final<|message|>",
+		"assistant_end":"<|return|>\\n",
+		"system":"<|start|>system<|message|>",
+		"system_end":"<|end|>\\n",
 	}
 	];