diff --git a/include/minja/chat-template.hpp b/include/minja/chat-template.hpp index b53e08f..e7bf82b 100644 --- a/include/minja/chat-template.hpp +++ b/include/minja/chat-template.hpp @@ -194,10 +194,9 @@ class chat_template { const json dummy_args_obj {{"argument_needle", "print('Hello, World!')"}}; const auto contains_arg_needle = [&](const std::string & out_str) { return contains(out_str, "") - || contains(out_str, "\"argument_needle\":") + || contains(out_str, "\"argument_needle\"") || contains(out_str, "'argument_needle':") - || contains(out_str, ">argument_needle<") - || contains(out_str, ""); + || contains(out_str, ">argument_needle<"); }; // Note: the arguments are rendered in both cases, but may be double-escaped, which we don't want. diff --git a/scripts/fetch_templates_and_goldens.py b/scripts/fetch_templates_and_goldens.py index 8361764..9501cf5 100644 --- a/scripts/fetch_templates_and_goldens.py +++ b/scripts/fetch_templates_and_goldens.py @@ -194,10 +194,9 @@ def make_tool_call(tool_name, arguments): dummy_args_obj = {"argument_needle": "print('Hello, World!')"} contains_arg_needle = lambda out_str: ( "" in out_str - or '"argument_needle":' in out_str + or '"argument_needle"' in out_str or "'argument_needle':" in out_str or ">argument_needle<" in out_str - or "" in out_str ) out = self.try_raw_render([ @@ -432,6 +431,16 @@ async def async_hf_download(repo_id: str, filename: str) -> str: async def process_model(output_folder: str, model_id: str, contexts: list[Context]): try: print(f"Processing model {model_id}...", file=sys.stderr) + + # Handle local .jinja files directly (for synthetic test templates) + if model_id.endswith('.jinja') and os.path.isfile(model_id): + async with aiofiles.open(model_id, 'r', encoding='utf-8') as f: + chat_template = await f.read() + # Use filename without extension as model_id for output naming + synthetic_id = os.path.basename(model_id).replace('.jinja', '') + await handle_chat_template(output_folder, synthetic_id, None, chat_template, contexts) + return + config_str = await async_hf_download(model_id, "tokenizer_config.json") try: diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 84ff609..c2e9ed2 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -143,6 +143,7 @@ set(MODEL_IDS deepseek-ai/DeepSeek-V2-Lite deepseek-ai/DeepSeek-V2.5 deepseek-ai/DeepSeek-V3 + # deepseek-ai/DeepSeek-V3.2 # No Jinja template; see synthetic below deepseek-ai/deepseek-coder-7b-instruct-v1.5 dicta-il/dictalm2.0-instruct ehristoforu/Falcon3-8B-Franken-Basestruct @@ -195,6 +196,9 @@ set(MODEL_IDS xwen-team/Xwen-7B-Chat zai-org/GLM-4.6 + # Synthetic templates for models without Jinja templates + ${CMAKE_CURRENT_SOURCE_DIR}/synthetic-deepseek-v3.2-dsml.jinja + # Broken, TODO: # ai21labs/AI21-Jamba-1.5-Large # https://github.com/google/minja/issues/8 # Almawave/Velvet-14B diff --git a/tests/synthetic-deepseek-v3.2-dsml.jinja b/tests/synthetic-deepseek-v3.2-dsml.jinja new file mode 100644 index 0000000..72044f5 --- /dev/null +++ b/tests/synthetic-deepseek-v3.2-dsml.jinja @@ -0,0 +1,42 @@ +{# Synthetic template based on DeepSeek V3.2 DSML format (encoding_dsv32.py) #} +{# V3.2 doesn't provide a Jinja template, so this replicates its Python encoding logic #} +{%- set bos_token = "<｜begin▁of▁sentence｜>" -%} +{%- set eos_token = "<｜end▁of▁sentence｜>" -%} +{%- set dsml_token = "｜DSML｜" -%} +{{ bos_token }} +{%- for message in messages -%} +{%- if message.role == 'system' -%} +{{ message.content }} +{%- elif message.role == 'user' -%} +<｜User｜>{{ message.content }}<｜Assistant｜> +{%- elif message.role == 'assistant' -%} +{%- if message.tool_calls is defined and message.tool_calls -%} +<{{ dsml_token }}function_calls> +{%- for tool_call in message.tool_calls -%} +{%- if tool_call.type == 'function' -%} +<{{ dsml_token }}invoke name="{{ tool_call.function.name }}"> +{%- if tool_call.function.arguments is mapping -%} +{%- for key, value in tool_call.function.arguments.items() -%} +{%- if value is string -%} +<{{ dsml_token }}parameter name="{{ key }}" string="true">{{ value }} +{%- else -%} +<{{ dsml_token }}parameter name="{{ key }}" string="false">{{ value | tojson }} +{%- endif -%} +{%- endfor -%} +{%- endif -%} + +{%- endif -%} +{%- endfor -%} + +{%- endif -%} +{%- if message.content -%} +{{ message.content }} +{%- endif -%} +{{ eos_token }} +{%- elif message.role == 'tool' -%} +<{{ dsml_token }}tool_result>{{ message.content }} +{%- endif -%} +{%- endfor -%} +{%- if add_generation_prompt -%} +<｜Assistant｜> +{%- endif -%} diff --git a/tests/test-capabilities.cpp b/tests/test-capabilities.cpp index aa17993..8c10eaa 100644 --- a/tests/test-capabilities.cpp +++ b/tests/test-capabilities.cpp @@ -287,3 +287,20 @@ TEST(CapabilitiesTest, GLM46) { EXPECT_FALSE(caps.requires_non_null_content); EXPECT_FALSE(caps.requires_typed_content); } + +// Synthetic template based on DeepSeek V3.2's DSML format (encoding_dsv32.py) +// V3.2 doesn't provide a Jinja template, so we replicate its Python encoding logic +// DSML format: <｜DSML｜parameter name="argument_needle" string="true"> +TEST(CapabilitiesTest, SyntheticDeepSeekV3_2_DSML) { + auto caps = get_caps("tests/synthetic-deepseek-v3.2-dsml.jinja"); + EXPECT_TRUE(caps.supports_system_role); + EXPECT_FALSE(caps.supports_tools); // No native tools block in template + EXPECT_TRUE(caps.supports_tool_calls); // Has tool_calls rendering with DSML format + EXPECT_FALSE(caps.supports_tool_call_id); + EXPECT_TRUE(caps.supports_tool_responses); + EXPECT_TRUE(caps.supports_parallel_tool_calls); // Iterates over tool_calls array + EXPECT_TRUE(caps.requires_object_arguments); // DSML iterates over argument keys + EXPECT_FALSE(caps.requires_non_null_content); + EXPECT_FALSE(caps.requires_typed_content); +} +