feat: added support for deepseek-reasoner

gptme/llm/llm_openai.py

@@ -90,7 +90,7 @@ def get_client(provider: Provider) -> "OpenAI":
     return clients[provider]
 
 
-def _prep_o1(msgs: list[Message]) -> Generator[Message, None, None]:
+def _prep_o1(msgs: Iterable[Message]) -> Generator[Message, None, None]:
     # prepare messages for OpenAI O1, which doesn't support the system role
     # and requires the first message to be from the user
     for msg in msgs:
@@ -101,6 +101,44 @@ def _prep_o1(msgs: list[Message]) -> Generator[Message, None, None]:
         yield msg
 
 
+def _merge_consecutive(msgs: Iterable[Message]) -> Generator[Message, None, None]:
+    # if consecutive messages from same role, merge them
+    last_message = None
+    for msg in msgs:
+        if last_message is None:
+            last_message = msg
+            continue
+
+        if last_message.role == msg.role:
+            last_message = last_message.replace(
+                content=f"{last_message.content}\n{msg.content}"
+            )
+            continue
+        else:
+            yield last_message
+            last_message = msg
+
+    if last_message:
+        yield last_message
+
+
+assert (
+    len(
+        list(
+            _merge_consecutive(
+                [Message(role="user", content="a"), Message(role="user", content="b")]
+            )
+        )
+    )
+    == 1
+)
+
+
+def _prep_deepseek_reasoner(msgs: list[Message]) -> Generator[Message, None, None]:
+    yield msgs[0]
+    yield from _merge_consecutive(_prep_o1(msgs[1:]))
+
+
 def chat(messages: list[Message], model: str, tools: list[ToolSpec] | None) -> str:
     # This will generate code and such, so we need appropriate temperature and top_p params
     # top_p controls diversity, temperature controls randomness
@@ -114,14 +152,16 @@ def chat(messages: list[Message], model: str, tools: list[ToolSpec] | None) -> s
     from openai import NOT_GIVEN  # fmt: skip
 
     is_o1 = base_model.startswith("o1")
+    is_deepseek_reasoner = base_model == "deepseek-reasoner"
+    is_reasoner = is_o1 or is_deepseek_reasoner
 
     messages_dicts, tools_dict = _prepare_messages_for_api(messages, model, tools)
 
     response = client.chat.completions.create(
         model=base_model,
         messages=messages_dicts,  # type: ignore
-        temperature=TEMPERATURE if not is_o1 else NOT_GIVEN,
-        top_p=TOP_P if not is_o1 else NOT_GIVEN,
+        temperature=TEMPERATURE if not is_reasoner else NOT_GIVEN,
+        top_p=TOP_P if not is_reasoner else NOT_GIVEN,
         tools=tools_dict if tools_dict else NOT_GIVEN,
         extra_headers=(openrouter_headers if provider == "openrouter" else {}),
     )
@@ -153,14 +193,16 @@ def stream(
     from openai import NOT_GIVEN  # fmt: skip
 
     is_o1 = base_model.startswith("o1")
+    is_deepseek_reasoner = base_model == "deepseek-reasoner"
+    is_reasoner = is_o1 or is_deepseek_reasoner
 
     messages_dicts, tools_dict = _prepare_messages_for_api(messages, model, tools)
 
     for chunk_raw in client.chat.completions.create(
         model=base_model,
         messages=messages_dicts,  # type: ignore
-        temperature=TEMPERATURE if not is_o1 else NOT_GIVEN,
-        top_p=TOP_P if not is_o1 else NOT_GIVEN,
+        temperature=TEMPERATURE if not is_reasoner else NOT_GIVEN,
+        top_p=TOP_P if not is_reasoner else NOT_GIVEN,
         stream=True,
         tools=tools_dict if tools_dict else NOT_GIVEN,
         # the llama-cpp-python server needs this explicitly set, otherwise unreliable results
@@ -449,15 +491,16 @@ def _spec2tool(spec: ToolSpec, model: ModelMeta) -> "ChatCompletionToolParam":
 def _prepare_messages_for_api(
     messages: list[Message], model: str, tools: list[ToolSpec] | None
 ) -> tuple[Iterable[dict], Iterable["ChatCompletionToolParam"] | None]:
-    from . import _get_base_model, get_provider_from_model  # fmt: skip
+    from . import _get_base_model  # fmt: skip
     from .models import get_model  # fmt: skip
 
-    get_provider_from_model(model)
     model_meta = get_model(model)
 
     is_o1 = _get_base_model(model).startswith("o1")
     if is_o1:
         messages = list(_prep_o1(messages))
+    if model_meta.model == "deepseek-reasoner":
+        messages = list(_prep_deepseek_reasoner(messages))
 
     messages_dicts: Iterable[dict] = (
         _process_file(msg, model_meta) for msg in msgs2dicts(messages)

gptme/llm/models.py

@@ -126,11 +126,17 @@ class _ModelDictMeta(TypedDict):
     "deepseek": {
         "deepseek-chat": {
             "context": 64_000,
-            "max_output": 4000,
+            "max_output": 8192,
             # 10x better price for cache hits
             "price_input": 0.14,
-            "price_output": 0.28,
-        }
+            "price_output": 1.1,
+        },
+        "deepseek-reasoner": {
+            "context": 64_000,
+            "max_output": 8192,
+            "price_input": 0.55,
+            "price_output": 2.19,
+        },
     },
     # https://groq.com/pricing/
     "groq": {