remove built-in cpp chat template

Author: ngxson

common/CMakeLists.txt

@@ -47,52 +47,6 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-#
-# inline chat templates
-#
-# Built-in Jinja chat templates for model architectures whose GGUF files ship
-# without an embedded chat_template. The header is generated from the manifest
-# common/chat-inline-templates.h.in into the build directory and included by
-# common/chat.cpp (a quoted include resolves to the generated copy because no
-# chat-inline-templates.h exists next to chat.cpp).
-
-set(CHAT_INLINE_TEMPLATES_MANIFEST ${CMAKE_CURRENT_SOURCE_DIR}/chat-inline-templates.h.in)
-set(CHAT_INLINE_TEMPLATES_GEN_DIR  ${CMAKE_CURRENT_BINARY_DIR}/generated)
-set(CHAT_INLINE_TEMPLATES_HEADER   ${CHAT_INLINE_TEMPLATES_GEN_DIR}/chat-inline-templates.h)
-set(CHAT_INLINE_TEMPLATES_SCRIPT   ${PROJECT_SOURCE_DIR}/scripts/gen-chat-inline-templates.py)
-
-# every template the manifest may reference, so the header regenerates on edits
-file(GLOB CHAT_INLINE_TEMPLATE_FILES ${PROJECT_SOURCE_DIR}/models/templates/*.jinja)
-
-find_package(Python3 COMPONENTS Interpreter QUIET)
-
-if (Python3_FOUND)
-    add_custom_command(
-        OUTPUT ${CHAT_INLINE_TEMPLATES_HEADER}
-        COMMAND ${CMAKE_COMMAND} -E make_directory ${CHAT_INLINE_TEMPLATES_GEN_DIR}
-        COMMAND ${Python3_EXECUTABLE} ${CHAT_INLINE_TEMPLATES_SCRIPT}
-                --manifest  ${CHAT_INLINE_TEMPLATES_MANIFEST}
-                --repo-root ${PROJECT_SOURCE_DIR}
-                --output    ${CHAT_INLINE_TEMPLATES_HEADER}
-        DEPENDS ${CHAT_INLINE_TEMPLATES_MANIFEST}
-                ${CHAT_INLINE_TEMPLATES_SCRIPT}
-                ${CHAT_INLINE_TEMPLATE_FILES}
-        COMMENT "Generating chat-inline-templates.h"
-        VERBATIM
-    )
-else()
-    message(WARNING "Python3 not found: inline chat templates disabled (generating empty chat-inline-templates.h)")
-    file(MAKE_DIRECTORY ${CHAT_INLINE_TEMPLATES_GEN_DIR})
-    file(WRITE ${CHAT_INLINE_TEMPLATES_HEADER}
-"// AUTO-GENERATED fallback (Python3 not found) - no inline templates available.
-#pragma once
-#include <cstddef>
-struct common_chat_inline_template_entry { const char * arch; const char * tmpl; };
-static const common_chat_inline_template_entry COMMON_CHAT_INLINE_TEMPLATES[] = { { nullptr, nullptr } };
-static const size_t COMMON_CHAT_INLINE_TEMPLATES_COUNT = 0;
-")
-endif()
-
 #
 # llama-common
 #
@@ -174,7 +128,6 @@ set_target_properties(${TARGET} PROPERTIES
 )
 
 target_include_directories(${TARGET} PUBLIC . ../vendor)
-target_include_directories(${TARGET} PRIVATE ${CHAT_INLINE_TEMPLATES_GEN_DIR})
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
 if (BUILD_SHARED_LIBS)

common/chat-inline-templates.h.in

@@ -653,15 +653,6 @@ std::string common_chat_templates_source(const struct common_chat_templates * tm
     return tmpls->template_default->source();
 }
 
-std::string common_chat_template_inline(const std::string & arch) {
-    for (size_t i = 0; i < COMMON_CHAT_INLINE_TEMPLATES_COUNT; ++i) {
-        if (arch == COMMON_CHAT_INLINE_TEMPLATES[i].arch) {
-            return COMMON_CHAT_INLINE_TEMPLATES[i].tmpl;
-        }
-    }
-    return "";
-}
-
 common_chat_templates_ptr common_chat_templates_init(const struct llama_model * model,
                                                      const std::string &        chat_template_override,
                                                      const std::string &        bos_token_override,
@@ -686,21 +677,6 @@ common_chat_templates_ptr common_chat_templates_init(const struct llama_model *
         default_template_src = chat_template_override;
     }
 
-    // No template override and the model has no embedded chat template: fall back
-    // to a built-in inline template selected by the model architecture, if any.
-    if (chat_template_override.empty() && default_template_src.empty() && template_tool_use_src.empty() && model != nullptr) {
-        char arch_buf[128] = { 0 };
-        if (llama_model_meta_val_str(model, "general.architecture", arch_buf, sizeof(arch_buf)) > 0) {
-            std::string inline_template = common_chat_template_inline(arch_buf);
-            if (!inline_template.empty()) {
-                LOG_INF("%s: no chat template found in model, using built-in inline template for arch '%s'\n",
-                        __func__, arch_buf);
-                default_template_src  = inline_template;
-                has_explicit_template = true;
-            }
-        }
-    }
-
     if (default_template_src.empty() || default_template_src == "chatml") {
         if (!template_tool_use_src.empty()) {
             default_template_src = template_tool_use_src;

common/chat.cpp

@@ -261,12 +261,6 @@ common_chat_templates_ptr common_chat_templates_init(const struct llama_model *
 bool        common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
 std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant = "");
 
-// Returns the built-in ("inline") chat template registered for the given model
-// architecture name (see common/chat-inline-templates.h.in), or an empty string
-// if none is registered. Used as a fallback for models that ship without an
-// embedded chat_template and when no template override is supplied.
-std::string common_chat_template_inline(const std::string & arch);
-
 struct common_chat_params common_chat_templates_apply(const struct common_chat_templates *        tmpls,
                                                       const struct common_chat_templates_inputs & inputs);
 

common/chat.h

@@ -2,6 +2,7 @@
 
 import json
 import re
+from pathlib import Path
 
 from typing import Any, Callable, Iterable, TYPE_CHECKING
 
@@ -489,6 +490,12 @@ def __init__(self, *args, **kwargs):
         if type(self)._skipped_mtp_tensors:
             logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors)
 
+        # add a default chat template; if the model has a built-in template, it will be overridden later
+        template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja"
+        if template_path.is_file():
+            with open(template_path, "r", encoding="utf-8") as f:
+                self.gguf_writer.add_chat_template(f.read())
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, _ = item