Parse GPT OSS in runner (#1160)

## Motivation

Simplification of API + moving model specific code to the runner

<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->

## Test Plan

### Manual Testing
Tested that GPT OSS outputs are parsed correctly on the dashboard.

### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->

justfile

@@ -1,3 +1,5 @@
+export NIX_CONFIG := "extra-experimental-features = nix-command flakes"
+
 fmt:
     nix fmt
 

src/exo/master/api.py

@@ -13,12 +13,6 @@ from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
-from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
-    HarmonyEncodingName,
-    Role,
-    StreamableParser,
-    load_harmony_encoding,
-)
 
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
@@ -67,8 +61,6 @@ from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
-encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
-
 
 def chunk_to_response(
     chunk: TokenChunk, command_id: CommandId
@@ -381,35 +373,8 @@ class API:
             instance_id=instance_id,
         )
 
-    async def _process_gpt_oss(self, token_chunks: Receiver[TokenChunk]):
-        stream = StreamableParser(encoding, role=Role.ASSISTANT)
-        thinking = False
-
-        async for chunk in token_chunks:
-            stream.process(chunk.token_id)
-
-            delta = stream.last_content_delta
-            ch = stream.current_channel
-
-            if ch == "analysis" and not thinking:
-                thinking = True
-                yield chunk.model_copy(update={"text": "<think>"})
-
-            if ch != "analysis" and thinking:
-                thinking = False
-                yield chunk.model_copy(update={"text": "</think>"})
-
-            if delta:
-                yield chunk.model_copy(update={"text": delta})
-
-            if chunk.finish_reason is not None:
-                if thinking:
-                    yield chunk.model_copy(update={"text": "</think>"})
-                yield chunk
-                break
-
     async def _chat_chunk_stream(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> AsyncGenerator[TokenChunk, None]:
         """Yield `TokenChunk`s for a given command until completion."""
 
@@ -417,16 +382,10 @@ class API:
             self._chat_completion_queues[command_id], recv = channel[TokenChunk]()
 
             with recv as token_chunks:
-                if parse_gpt_oss:
-                    async for chunk in self._process_gpt_oss(token_chunks):
-                        yield chunk
-                        if chunk.finish_reason is not None:
-                            break
-                else:
-                    async for chunk in token_chunks:
-                        yield chunk
-                        if chunk.finish_reason is not None:
-                            break
+                async for chunk in token_chunks:
+                    yield chunk
+                    if chunk.finish_reason is not None:
+                        break
 
         except anyio.get_cancelled_exc_class():
             # TODO: TaskCancelled
@@ -442,11 +401,11 @@ class API:
             del self._chat_completion_queues[command_id]
 
     async def _generate_chat_stream(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> AsyncGenerator[str, None]:
         """Generate chat completion stream as JSON strings."""
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             chunk_response: ChatCompletionResponse = chunk_to_response(
                 chunk, command_id
             )
@@ -458,7 +417,7 @@ class API:
                 yield "data: [DONE]\n\n"
 
     async def _collect_chat_completion(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> ChatCompletionResponse:
         """Collect all token chunks for a chat completion and return a single response."""
 
@@ -466,7 +425,7 @@ class API:
         model: str | None = None
         finish_reason: FinishReason | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             if model is None:
                 model = chunk.model
 
@@ -495,7 +454,7 @@ class API:
         )
 
     async def _collect_chat_completion_with_stats(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> BenchChatCompletionResponse:
         text_parts: list[str] = []
         model: str | None = None
@@ -503,7 +462,7 @@ class API:
 
         stats: GenerationStats | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             if model is None:
                 model = chunk.model
 
@@ -544,8 +503,6 @@ class API:
         """Handle chat completions, supporting both streaming and non-streaming responses."""
         model_meta = await resolve_model_meta(payload.model)
         payload.model = model_meta.model_id
-        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
-        logger.info(f"{parse_gpt_oss=}")
 
         if not any(
             instance.shard_assignments.model_id == payload.model
@@ -562,17 +519,16 @@ class API:
         await self._send(command)
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id, parse_gpt_oss),
+                self._generate_chat_stream(command.command_id),
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id, parse_gpt_oss)
+        return await self._collect_chat_completion(command.command_id)
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionTaskParams
     ) -> BenchChatCompletionResponse:
         model_meta = await resolve_model_meta(payload.model)
-        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
         payload.model = model_meta.model_id
 
         if not any(
@@ -589,10 +545,7 @@ class API:
         command = ChatCompletion(request_params=payload)
         await self._send(command)
 
-        response = await self._collect_chat_completion_with_stats(
-            command.command_id,
-            parse_gpt_oss,
-        )
+        response = await self._collect_chat_completion_with_stats(command.command_id)
         return response
 
     def _calculate_total_available_memory(self) -> Memory:

src/exo/worker/engines/mlx/utils_mlx.py

@@ -366,6 +366,8 @@ def apply_chat_template(
         tools=chat_task_data.tools,
     )
 
+    logger.info(prompt)
+
     return prompt
 
 

src/exo/worker/runner/runner.py

@@ -1,6 +1,15 @@
 import time
+from collections.abc import Generator
+from functools import cache
 
 import mlx.core as mx
+from mlx_lm.models.gpt_oss import Model as GptOssModel
+from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
+    HarmonyEncodingName,
+    Role,
+    StreamableParser,
+    load_harmony_encoding,
+)
 
 from exo.shared.types.api import ChatCompletionMessageText
 from exo.shared.types.chunks import TokenChunk
@@ -153,11 +162,19 @@ def main(
                     _check_for_debug_prompts(task_params.messages[0].content)
 
                     # Generate responses using the actual MLX generation
-                    for response in mlx_generate(
+                    mlx_generator = mlx_generate(
                         model=model,
                         tokenizer=tokenizer,
                         task=task_params,
-                    ):
+                    )
+
+                    # GPT-OSS specific parsing to match other model formats.
+                    if isinstance(model, GptOssModel):
+                        mlx_generator = parse_gpt_oss(mlx_generator)
+
+                    # TODO: Add tool call parser here
+
+                    for response in mlx_generator:
                         match response:
                             case GenerationResponse():
                                 if shard_metadata.device_rank == 0:
@@ -207,6 +224,43 @@ def main(
                 break
 
 
+@cache
+def get_gpt_oss_encoding():
+    encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
+    return encoding
+
+
+def parse_gpt_oss(
+    responses: Generator[GenerationResponse],
+) -> Generator[GenerationResponse]:
+    encoding = get_gpt_oss_encoding()
+    stream = StreamableParser(encoding, role=Role.ASSISTANT)
+    thinking = False
+
+    for response in responses:
+        stream.process(response.token)
+
+        delta = stream.last_content_delta
+        ch = stream.current_channel
+
+        if ch == "analysis" and not thinking:
+            thinking = True
+            yield response.model_copy(update={"text": "<think>"})
+
+        if ch != "analysis" and thinking:
+            thinking = False
+            yield response.model_copy(update={"text": "</think>"})
+
+        if delta:
+            yield response.model_copy(update={"text": delta})
+
+        if response.finish_reason is not None:
+            if thinking:
+                yield response.model_copy(update={"text": "</think>"})
+            yield response
+            break
+
+
 EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"
 EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
 EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"