Refactor runner for implementing batching

## Motivation

Tests caught a bug. It was a real bug.

.mlx_typings/mlx_lm/generate.pyi

@@ -73,9 +73,11 @@ class GenerationResponse:
     finish_reason: Optional[str] = ...
 
 def maybe_quantize_kv_cache(
-    prompt_cache, quantized_kv_start, kv_group_size, kv_bits
-):  # -> None:
-    ...
+    prompt_cache: Any,
+    quantized_kv_start: int | None,
+    kv_group_size: int | None,
+    kv_bits: int | None,
+) -> None: ...
 def generate_step(
     prompt: mx.array,
     model: nn.Module,

.mlx_typings/mlx_lm/models/cache.pyi

@@ -16,7 +16,7 @@ class Cache(Protocol):
         self, keys: mx.array, values: mx.array
     ) -> tuple[mx.array, mx.array]: ...
     @property
-    def state(self) -> tuple[mx.array, mx.array]: ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
 
@@ -92,13 +92,14 @@ class _BaseCache(Cache):
     values: mx.array
     offset: int
     @property
-    def state(self) -> tuple[mx.array, mx.array]: ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
     @property
     def meta_state(self) -> Literal[""]: ...
     @meta_state.setter
     def meta_state(self, v) -> None: ...
+    def trim(self, n: int) -> int: ...
     def is_trimmable(self) -> Literal[False]: ...
     @classmethod
     def from_state(cls, state, meta_state) -> Self: ...
@@ -114,15 +115,13 @@ class ConcatenateKVCache(_BaseCache):
     def update_and_fetch(self, keys, values):  # -> tuple[Any | array, Any | array]:
         ...
     @property
-    def state(self):  # -> tuple[Any | array | None, Any | array | None]:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
         ...
 
@@ -132,10 +131,7 @@ class QuantizedKVCache(_BaseCache):
     def update_and_fetch(self, keys, values):  # -> Any:
         ...
     @property
-    def state(
-        self,
-    ):  # -> tuple[Any | tuple[array, array, array] | None, Any | tuple[array, array, array] | None] | Any:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -147,8 +143,7 @@ class QuantizedKVCache(_BaseCache):
         ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
         ...
 
@@ -160,13 +155,12 @@ class KVCache(_BaseCache):
     @property
     def state(
         self,
-    ) -> tuple[array, array]: ...
+    ) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v) -> None: ...
     def is_trimmable(self):  # -> Literal[True]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def to_quantized(
         self, group_size: int = ..., bits: int = ...
     ) -> QuantizedKVCache: ...
@@ -183,8 +177,7 @@ class RotatingKVCache(_BaseCache):
     @property
     def state(
         self,
-    ):  # -> tuple[Any | array, Any | array] | tuple[Any | array | None, Any | array | None]:
-        ...
+    ) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -196,8 +189,7 @@ class RotatingKVCache(_BaseCache):
         ...
     def is_trimmable(self):  # -> bool:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     def to_quantized(
         self, group_size: int = ..., bits: int = ...
     ) -> QuantizedKVCache: ...
@@ -212,8 +204,7 @@ class ArraysCache(_BaseCache):
         ...
     def __getitem__(self, idx): ...
     @property
-    def state(self):  # -> list[Any | array] | list[array]:
-        ...
+    def state(self) -> tuple[mx.array | None, mx.array | None]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...
@@ -239,8 +230,7 @@ class ChunkedKVCache(KVCache):
         ...
     def update_and_fetch(self, keys, values):  # -> tuple[array, array]:
         ...
-    def trim(self, n):  # -> int:
-        ...
+    def trim(self, n: int) -> int: ...
     @property
     def meta_state(self):  # -> tuple[str, ...]:
         ...
@@ -253,10 +243,9 @@ class CacheList(_BaseCache):
     def __getitem__(self, idx): ...
     def is_trimmable(self):  # -> bool:
         ...
-    def trim(self, n): ...
+    def trim(self, n: int) -> int: ...
     @property
-    def state(self):  # -> list[Any]:
-        ...
+    def state(self) -> list[tuple[mx.array | None, mx.array | None]]: ...
     @state.setter
     def state(self, v):  # -> None:
         ...

src/exo/download/download_utils.py

@@ -314,9 +314,13 @@ async def fetch_file_list_with_cache(
         _fetched_file_lists_this_session.add(cache_key)
         return file_list
     except Exception as e:
+        logger.opt(exception=e).warning(
+            "Ran into exception when fetching file list from HF."
+        )
+
         if await aios.path.exists(cache_file):
             logger.warning(
-                f"No internet and no cached file list for {model_id} - using local file list"
+                f"No cached file list for {model_id} - using local file list"
             )
             async with aiofiles.open(cache_file, "r") as f:
                 return TypeAdapter(list[FileListEntry]).validate_json(await f.read())

src/exo/routing/router.py

@@ -258,6 +258,6 @@ def get_node_id_keypair(
 
         # if no valid credentials, create new ones and persist
         with open(path, "w+b") as f:
-            keypair = Keypair.generate_ed25519()
+            keypair = Keypair.generate()
             f.write(keypair.to_bytes())
             return keypair

src/exo/shared/types/mlx.py

@@ -2,6 +2,8 @@
 
 from collections.abc import Sequence
 
+from mlx import core as mx
+from mlx import nn as nn
 from mlx_lm.models.cache import (
     ArraysCache,
     CacheList,
@@ -14,3 +16,16 @@ from mlx_lm.models.cache import (
 KVCacheType = Sequence[
     KVCache | RotatingKVCache | QuantizedKVCache | ArraysCache | CacheList
 ]
+
+
+# Model is a wrapper function to fix the fact that mlx is not strongly typed in the same way that EXO is.
+# For example - MLX has no guarantee of the interface that nn.Module will expose. But we need a guarantee that it has a __call__() function
+class Model(nn.Module):
+    layers: list[nn.Module]
+
+    def __call__(
+        self,
+        x: mx.array,
+        cache: KVCacheType | None,
+        input_embeddings: mx.array | None = None,
+    ) -> mx.array: ...

src/exo/utils/channels.py

@@ -1,5 +1,6 @@
 import contextlib
 import multiprocessing as mp
+from collections.abc import Generator
 from dataclasses import dataclass, field
 from math import inf
 from multiprocessing.synchronize import Event
@@ -282,6 +283,54 @@ class MpReceiver[T]:
         return d
 
 
+class NonBlockingGenerator[T](Generator[T | None, None, None]):
+    def __init__(self, source: MpReceiver[T] | Generator[T | None, None, None]) -> None:
+        self._receiver: MpReceiver[T] | None = None
+        self._inner: Generator[T | None, None, None] | None = None
+        if isinstance(source, MpReceiver):
+            self._receiver = source
+        else:
+            self._inner = source
+        self._exhausted = False
+
+    def send(self, value: None, /) -> T | None:
+        if self._exhausted:
+            raise StopIteration
+        if self._inner is not None:
+            try:
+                return next(self._inner)
+            except (StopIteration, ClosedResourceError):
+                self._exhausted = True
+                raise StopIteration from None
+        assert self._receiver is not None
+        try:
+            return self._receiver.receive_nowait()
+        except WouldBlock:
+            return None
+        except (EndOfStream, ClosedResourceError):
+            self._exhausted = True
+            raise StopIteration from None
+
+    def throw(
+        self,
+        typ: type[BaseException] | BaseException,
+        val: BaseException | object = None,
+        tb: TracebackType | None = None,
+        /,
+    ) -> T | None:
+        raise StopIteration
+
+    @property
+    def is_exhausted(self) -> bool:
+        return self._exhausted
+
+    def try_receive(self) -> T | None:
+        try:
+            return next(self)
+        except StopIteration:
+            return None
+
+
 class channel[T]:  # noqa: N801
     """Create a pair of asynchronous channels for communicating within the same process"""
 

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

@@ -1,17 +0,0 @@
-import mlx.core as mx
-import mlx.nn as nn
-from mlx_lm.models.cache import KVCache
-
-# These are wrapper functions to fix the fact that mlx is not strongly typed in the same way that EXO is.
-# For example - MLX has no guarantee of the interface that nn.Module will expose. But we need a guarantee that it has a __call__() function
-
-
-class Model(nn.Module):
-    layers: list[nn.Module]
-
-    def __call__(
-        self,
-        x: mx.array,
-        cache: list[KVCache] | None,
-        input_embeddings: mx.array | None = None,
-    ) -> mx.array: ...

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

@@ -49,6 +49,21 @@ TimeoutCallback = Callable[[], None]
 LayerLoadedCallback = Callable[[int, int], None]  # (layers_loaded, total_layers)
 
 
+_pending_prefill_sends: list[tuple[mx.array, int, mx.distributed.Group]] = []
+
+
+def flush_prefill_sends() -> None:
+    for output, dst, group in _pending_prefill_sends:
+        sent = mx.distributed.send(output, dst, group=group)
+        mx.async_eval(sent)
+    _pending_prefill_sends.clear()
+
+
+def clear_prefill_sends() -> None:
+    # Discard pending sends (e.g. on cancellation).
+    _pending_prefill_sends.clear()
+
+
 def eval_with_timeout(
     mlx_item: Any,  # pyright: ignore[reportAny]
     timeout_seconds: float = 60.0,
@@ -150,6 +165,7 @@ class PipelineLastLayer(CustomMlxLayer):
         self.group = group
         self.original_layer_signature = signature(self.original_layer.__call__)
         self.is_prefill: bool = False
+        self.queue_sends: bool = False
 
     def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
         cache = self.original_layer_signature.bind_partial(
@@ -163,9 +179,14 @@ class PipelineLastLayer(CustomMlxLayer):
         mx.eval(output)
 
         if self.r != self.s - 1:
-            output = mx.distributed.send(
-                output, (self.r + 1) % self.s, group=self.group
-            )
+            if self.queue_sends:
+                _pending_prefill_sends.append(
+                    (output, (self.r + 1) % self.s, self.group)
+                )
+            else:
+                output = mx.distributed.send(
+                    output, (self.r + 1) % self.s, group=self.group
+                )
             if cache is not None:
                 # CacheList (used by MLA models like DeepSeekV32, GLM MoE DSA)
                 # doesn't have .keys directly; access via first sub-cache.
@@ -190,6 +211,12 @@ def set_pipeline_prefill(model: nn.Module, is_prefill: bool) -> None:
             layer.is_prefill = is_prefill
 
 
+def set_pipeline_queue_sends(model: nn.Module, queue_sends: bool) -> None:
+    for layer in model.layers:  # type: ignore
+        if isinstance(layer, PipelineLastLayer):
+            layer.queue_sends = queue_sends
+
+
 def get_inner_model(model: nn.Module) -> nn.Module:
     inner = getattr(model, "model", None)
     if isinstance(inner, nn.Module):

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

@@ -13,8 +13,7 @@ from mlx_lm.models.cache import (
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.shared.types.memory import Memory
-from exo.shared.types.mlx import KVCacheType
-from exo.worker.engines.mlx import Model
+from exo.shared.types.mlx import KVCacheType, Model
 from exo.worker.engines.mlx.constants import CACHE_GROUP_SIZE, KV_CACHE_BITS
 from exo.worker.runner.bootstrap import logger
 
@@ -254,9 +253,9 @@ def trim_cache(
             if snapshot is not None and snapshot.states[i] is not None:
                 cache[i] = deepcopy(snapshot.states[i])  # type: ignore
             else:
-                c.state = [None] * len(c.state)  # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType]
+                c.state = [None] * len(c.state)
         else:
-            c.trim(num_tokens)  # pyright: ignore[reportUnknownMemberType]
+            c.trim(num_tokens)
 
 
 def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:

src/exo/worker/engines/mlx/generator/generate.py

@@ -1,10 +1,14 @@
+import functools
 import math
 import time
 from copy import deepcopy
 from typing import Callable, Generator, cast, get_args
 
 import mlx.core as mx
-from mlx_lm.generate import stream_generate
+from mlx_lm.generate import (
+    maybe_quantize_kv_cache,
+    stream_generate,
+)
 from mlx_lm.models.cache import ArraysCache, RotatingKVCache
 from mlx_lm.sample_utils import make_sampler
 from mlx_lm.tokenizer_utils import TokenizerWrapper
@@ -19,13 +23,19 @@ from exo.shared.types.api import (
 )
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
-from exo.shared.types.mlx import KVCacheType
+from exo.shared.types.mlx import KVCacheType, Model
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.shared.types.worker.runner_response import (
     GenerationResponse,
 )
-from exo.worker.engines.mlx import Model
-from exo.worker.engines.mlx.auto_parallel import set_pipeline_prefill
+from exo.worker.engines.mlx.auto_parallel import (
+    PipelineFirstLayer,
+    PipelineLastLayer,
+    clear_prefill_sends,
+    flush_prefill_sends,
+    set_pipeline_prefill,
+    set_pipeline_queue_sends,
+)
 from exo.worker.engines.mlx.cache import (
     CacheSnapshot,
     KVPrefixCache,
@@ -56,6 +66,130 @@ class PrefillCancelled(BaseException):
     """Raised when prefill is cancelled via the progress callback."""
 
 
+def _has_pipeline_communication_layer(model: Model):
+    for layer in model.layers:
+        if isinstance(layer, (PipelineFirstLayer, PipelineLastLayer)):
+            return True
+    return False
+
+
+def pipeline_parallel_prefill(
+    model: Model,
+    prompt: mx.array,
+    prompt_cache: KVCacheType,
+    prefill_step_size: int,
+    kv_group_size: int | None,
+    kv_bits: int | None,
+    prompt_progress_callback: Callable[[int, int], None],
+    distributed_prompt_progress_callback: Callable[[], None] | None,
+    group: mx.distributed.Group,
+) -> None:
+    """Prefill the KV cache for pipeline parallel with overlapping stages.
+
+    Each rank processes the full prompt through its real cache, offset by leading
+    and trailing dummy iterations.
+
+    Total iterations per rank = N_real_chunks + world_size - 1:
+      - rank r leading dummies  (skip_pipeline_io, throwaway cache)
+      - N_real_chunks real      (pipeline IO active, real cache)
+      - (world_size-1-r) trailing dummies (skip_pipeline_io, throwaway cache)
+
+    e.g.
+    Timeline (2 ranks, 3 chunks of 10240 tokens @ step=4096):
+        iter 0: R0 real[0:4096]     R1 dummy
+        iter 1: R0 real[4096:8192]  R1 real[0:4096]
+        iter 2: R0 real[8192:10240] R1 real[4096:8192]
+        iter 3: R0 dummy            R1 real[8192:10240]
+
+    This function is designed to match mlx_lm's stream_generate exactly in terms of
+    side effects (given the same prefill step size)
+    """
+    prefill_step_size = prefill_step_size // min(4, group.size())
+
+    quantize_cache_fn: Callable[..., None] = functools.partial(
+        maybe_quantize_kv_cache,
+        quantized_kv_start=0,
+        kv_group_size=kv_group_size,
+        kv_bits=kv_bits,
+    )
+
+    _prompt_cache: KVCacheType = prompt_cache
+    rank = group.rank()
+    world_size = group.size()
+
+    # Build list of real prompt chunk sizes
+    total = len(prompt)
+    real_chunk_sizes: list[int] = []
+    remaining = total - 1
+    while remaining:
+        n = min(prefill_step_size, remaining)
+        real_chunk_sizes.append(n)
+        remaining -= n
+    n_real = len(real_chunk_sizes)
+
+    # Each rank does: [rank leading dummies] [N real chunks] [world_size-1-rank trailing dummies]
+    n_leading = rank
+    n_trailing = world_size - 1 - rank
+    n_total = n_leading + n_real + n_trailing
+
+    t_start = time.perf_counter()
+    processed = 0
+    logger.info(
+        f"[R{rank}] Pipeline prefill: {n_real} real + {n_leading} leading + {n_trailing} trailing = {n_total} iterations"
+    )
+    clear_prefill_sends()
+
+    # Initial callback matching generate_step
+    prompt_progress_callback(0, total)
+
+    try:
+        with mx.stream(generation_stream):
+            for _ in range(n_leading):
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+            for i in range(n_real):
+                chunk_size = real_chunk_sizes[i]
+                model(
+                    prompt[processed : processed + chunk_size][None],
+                    cache=_prompt_cache,
+                )
+                quantize_cache_fn(_prompt_cache)
+                processed += chunk_size
+
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+                flush_prefill_sends()
+
+                prompt_progress_callback(processed, total)
+
+            for _ in range(n_trailing):
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+    finally:
+        clear_prefill_sends()
+
+    # Post-loop: process remaining 1 token + add +1 entry to match stream_generate.
+    for _ in range(2):
+        with mx.stream(generation_stream):
+            model(prompt[-1:][None], cache=_prompt_cache)
+            quantize_cache_fn(_prompt_cache)
+        flush_prefill_sends()
+
+    assert _prompt_cache is not None
+    mx.eval([c.state for c in _prompt_cache])  # type: ignore
+
+    # Final callback matching generate_step
+    prompt_progress_callback(total, total)
+
+    logger.info(
+        f"[R{rank}] Prefill: {n_real} real + {n_leading}+{n_trailing} dummy iterations, "
+        f"Processed {processed} tokens in {(time.perf_counter() - t_start) * 1000:.1f}ms"
+    )
+
+
 def prefill(
     model: Model,
     tokenizer: TokenizerWrapper,
@@ -64,6 +198,7 @@ def prefill(
     cache: KVCacheType,
     group: mx.distributed.Group | None,
     on_prefill_progress: Callable[[int, int], None] | None,
+    distributed_prompt_progress_callback: Callable[[], None] | None,
 ) -> tuple[float, int, list[CacheSnapshot]]:
     """Prefill the KV cache with prompt tokens.
 
@@ -95,31 +230,57 @@ def prefill(
         if on_prefill_progress is not None:
             on_prefill_progress(processed, total)
 
+    def combined_progress_callback(processed: int, total: int) -> None:
+        if distributed_prompt_progress_callback is not None:
+            distributed_prompt_progress_callback()
+        progress_callback(processed, total)
+
     set_pipeline_prefill(model, is_prefill=True)
 
     mx_barrier(group)
     logger.info("Starting prefill")
 
-    # Use max_tokens=1 because max_tokens=0 does not work.
-    # We just throw away the generated token - we only care about filling the cache
+    is_pipeline = _has_pipeline_communication_layer(model)
+
+    prefill_step_size = 4096
+
     try:
-        for _ in stream_generate(
-            model=model,
-            tokenizer=tokenizer,
-            prompt=prompt_tokens,
-            max_tokens=1,
-            sampler=sampler,
-            prompt_cache=cache,
-            prefill_step_size=4096,
-            kv_group_size=KV_GROUP_SIZE,
-            kv_bits=KV_BITS,
-            prompt_progress_callback=progress_callback,
-        ):
-            break  # Stop after first iteration - cache is now filled
+        if is_pipeline and num_tokens >= prefill_step_size:
+            set_pipeline_queue_sends(model, queue_sends=True)
+            assert group is not None, "Pipeline prefill requires a distributed group"
+            pipeline_parallel_prefill(
+                model=model,
+                prompt=prompt_tokens,
+                prompt_cache=cache,
+                prefill_step_size=prefill_step_size,
+                kv_group_size=KV_GROUP_SIZE,
+                kv_bits=KV_BITS,
+                prompt_progress_callback=progress_callback,
+                distributed_prompt_progress_callback=distributed_prompt_progress_callback,
+                group=group,
+            )
+        else:
+            # Use max_tokens=1 because max_tokens=0 does not work.
+            # We just throw away the generated token - we only care about filling the cache
+            for _ in stream_generate(
+                model=model,
+                tokenizer=tokenizer,
+                prompt=prompt_tokens,
+                max_tokens=1,
+                sampler=sampler,
+                prompt_cache=cache,
+                prefill_step_size=prefill_step_size,
+                kv_group_size=KV_GROUP_SIZE,
+                kv_bits=KV_BITS,
+                prompt_progress_callback=combined_progress_callback,
+            ):
+                break  # Stop after first iteration - cache is now filled
     except PrefillCancelled:
+        set_pipeline_queue_sends(model, queue_sends=False)
         set_pipeline_prefill(model, is_prefill=False)
         raise
 
+    set_pipeline_queue_sends(model, queue_sends=False)
     set_pipeline_prefill(model, is_prefill=False)
 
     # stream_generate added 1 extra generated token to the cache, so we should trim it.
@@ -132,7 +293,7 @@ def prefill(
                 cache[i] = deepcopy(pre_gen.states[i])  # type: ignore
         else:
             assert not isinstance(c, (ArraysCache, RotatingKVCache))
-            c.trim(2)  # pyright: ignore[reportUnknownMemberType]
+            c.trim(2)
 
     elapsed = time.perf_counter() - start_time
     tokens_per_sec = num_tokens / elapsed if elapsed > 0 else 0.0
@@ -275,6 +436,8 @@ def mlx_generate(
     kv_prefix_cache: KVPrefixCache | None,
     group: mx.distributed.Group | None,
     on_prefill_progress: Callable[[int, int], None] | None = None,
+    distributed_prompt_progress_callback: Callable[[], None] | None = None,
+    on_generation_token: Callable[[], None] | None = None,
 ) -> Generator[GenerationResponse]:
     # Ensure that generation stats only contains peak memory for this generation
     mx.reset_peak_memory()
@@ -336,6 +499,7 @@ def mlx_generate(
         caches,
         group,
         on_prefill_progress,
+        distributed_prompt_progress_callback,
     )
     cache_snapshots: list[CacheSnapshot] | None = ssm_snapshots_list or None
 
@@ -481,6 +645,9 @@ def mlx_generate(
                         full_prompt_tokens, caches, cache_snapshots
                     )
 
+        if on_generation_token is not None:
+            on_generation_token()
+
         yield GenerationResponse(
             text=text,
             token=out.token,

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

@@ -40,6 +40,7 @@ from pydantic import RootModel
 from exo.download.download_utils import build_model_path
 from exo.shared.types.common import Host
 from exo.shared.types.memory import Memory
+from exo.shared.types.mlx import Model
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import (
     BoundInstance,
@@ -52,7 +53,6 @@ from exo.shared.types.worker.shards import (
     ShardMetadata,
     TensorShardMetadata,
 )
-from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.auto_parallel import (
     LayerLoadedCallback,
     TimeoutCallback,

src/exo/worker/plan.py

@@ -297,10 +297,10 @@ def _pending_tasks(
             # the task status _should_ be set to completed by the LAST runner
             # it is currently set by the first
             # this is definitely a hack
-            if task.task_id in runner.completed:
+            if task.task_id in runner.completed or task.task_id in runner.pending:
                 continue
 
-            if isinstance(runner.status, RunnerReady) and all(
+            if isinstance(runner.status, (RunnerReady, RunnerRunning)) and all(
                 isinstance(all_runners[global_runner_id], (RunnerReady, RunnerRunning))
                 for global_runner_id in runner.bound_instance.instance.shard_assignments.runner_to_shard
             ):

src/exo/worker/runner/bootstrap.py

@@ -33,10 +33,15 @@ def entrypoint(
     try:
         if bound_instance.is_image_model:
             from exo.worker.runner.image_models.runner import main
-        else:
-            from exo.worker.runner.llm_inference.runner import main
 
-        main(bound_instance, event_sender, task_receiver, cancel_receiver)
+            main(bound_instance, event_sender, task_receiver, cancel_receiver)
+        else:
+            from exo.worker.runner.llm_inference.runner import Runner
+
+            runner = Runner(
+                bound_instance, event_sender, task_receiver, cancel_receiver
+            )
+            runner.main()
 
     except ClosedResourceError:
         logger.warning("Runner communication closed unexpectedly")

src/exo/worker/runner/llm_inference/batch_generator.py

@@ -0,0 +1,178 @@
+from collections import deque
+from collections.abc import Generator
+from dataclasses import dataclass, field
+
+import mlx.core as mx
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+
+from exo.shared.types.chunks import ErrorChunk, PrefillProgressChunk
+from exo.shared.types.common import ModelId
+from exo.shared.types.events import ChunkGenerated, Event
+from exo.shared.types.mlx import Model
+from exo.shared.types.tasks import TaskId, TextGeneration
+from exo.shared.types.text_generation import TextGenerationTaskParams
+from exo.shared.types.worker.runner_response import GenerationResponse
+from exo.utils.channels import MpReceiver, MpSender
+from exo.worker.engines.mlx.cache import KVPrefixCache
+from exo.worker.engines.mlx.generator.generate import PrefillCancelled, mlx_generate
+from exo.worker.engines.mlx.utils_mlx import (
+    apply_chat_template,
+    mx_any,
+)
+
+EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"
+EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
+EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"
+
+
+def _check_for_debug_prompts(task_params: TextGenerationTaskParams) -> None:
+    """Check for debug prompt triggers in the input."""
+    import time
+
+    from exo.worker.engines.mlx.utils_mlx import mlx_force_oom
+
+    if len(task_params.input) == 0:
+        return
+    prompt = task_params.input[0].content
+    if not prompt:
+        return
+    if EXO_RUNNER_MUST_FAIL in prompt:
+        raise Exception("Artificial runner exception - for testing purposes only.")
+    if EXO_RUNNER_MUST_OOM in prompt:
+        mlx_force_oom()
+    if EXO_RUNNER_MUST_TIMEOUT in prompt:
+        time.sleep(100)
+
+
+@dataclass(eq=False)
+class BatchGenerator:
+    model: Model
+    tokenizer: TokenizerWrapper
+    group: mx.distributed.Group | None
+    kv_prefix_cache: KVPrefixCache | None
+    model_id: ModelId
+    device_rank: int
+    cancel_receiver: MpReceiver[TaskId]
+    cancelled_tasks: set[TaskId]
+    event_sender: MpSender[Event]
+    check_for_cancel_every: int
+
+    _queue: deque[tuple[TextGeneration, MpSender[GenerationResponse]]] = field(
+        default_factory=deque, init=False
+    )
+    _active: (
+        tuple[
+            TextGeneration,
+            MpSender[GenerationResponse],
+            Generator[GenerationResponse],
+        ]
+        | None
+    ) = field(default=None, init=False)
+
+    def submit(
+        self,
+        task: TextGeneration,
+        sender: MpSender[GenerationResponse],
+    ) -> None:
+        self._queue.append((task, sender))
+        if self._active is None:
+            self._start_next()
+
+    def step(self) -> None:
+        if self._active is None:
+            if self._queue:
+                self._start_next()
+            else:
+                return
+
+        if self._active is None:
+            return
+
+        task, sender, gen = self._active
+        try:
+            response = next(gen)
+            sender.send(response)
+        except (StopIteration, PrefillCancelled):
+            sender.close()
+            self._active = None
+            if self._queue:
+                self._start_next()
+        except Exception as e:
+            self._send_error(task, e)
+            sender.close()
+            self._active = None
+            raise
+
+    def _start_next(self) -> None:
+        task, sender = self._queue.popleft()
+        try:
+            gen = self._build_generator(task)
+        except Exception as e:
+            self._send_error(task, e)
+            sender.close()
+            raise
+        self._active = (task, sender, gen)
+
+    def _send_error(self, task: TextGeneration, e: Exception) -> None:
+        if self.device_rank == 0:
+            self.event_sender.send(
+                ChunkGenerated(
+                    command_id=task.command_id,
+                    chunk=ErrorChunk(
+                        model=self.model_id,
+                        finish_reason="error",
+                        error_message=str(e),
+                    ),
+                )
+            )
+
+    def _build_generator(self, task: TextGeneration) -> Generator[GenerationResponse]:
+        _check_for_debug_prompts(task.task_params)
+        prompt = apply_chat_template(self.tokenizer, task.task_params)
+
+        def on_prefill_progress(processed: int, total: int) -> None:
+            if self.device_rank == 0:
+                self.event_sender.send(
+                    ChunkGenerated(
+                        command_id=task.command_id,
+                        chunk=PrefillProgressChunk(
+                            model=self.model_id,
+                            processed_tokens=processed,
+                            total_tokens=total,
+                        ),
+                    )
+                )
+
+        def distributed_prompt_progress_callback() -> None:
+            self.cancelled_tasks.update(self.cancel_receiver.collect())
+            want_to_cancel = (task.task_id in self.cancelled_tasks) or (
+                TaskId("CANCEL_CURRENT_TASK") in self.cancelled_tasks
+            )
+            if mx_any(want_to_cancel, self.group):
+                raise PrefillCancelled()
+
+        tokens_since_cancel_check = self.check_for_cancel_every
+
+        def on_generation_token() -> None:
+            nonlocal tokens_since_cancel_check
+            tokens_since_cancel_check += 1
+            if tokens_since_cancel_check >= self.check_for_cancel_every:
+                tokens_since_cancel_check = 0
+                self.cancelled_tasks.update(self.cancel_receiver.collect())
+                want_to_cancel = (task.task_id in self.cancelled_tasks) or (
+                    TaskId("CANCEL_CURRENT_TASK") in self.cancelled_tasks
+                )
+                if mx_any(want_to_cancel, self.group):
+                    raise PrefillCancelled()
+
+        return mlx_generate(
+            model=self.model,
+            tokenizer=self.tokenizer,
+            task=task.task_params,
+            prompt=prompt,
+            kv_prefix_cache=self.kv_prefix_cache,
+            on_prefill_progress=on_prefill_progress,
+            distributed_prompt_progress_callback=distributed_prompt_progress_callback,
+            on_generation_token=on_generation_token,
+            group=self.group,
+        )

src/exo/worker/runner/llm_inference/model_output_parsers.py

@@ -0,0 +1,341 @@
+from collections.abc import Generator
+from functools import cache
+
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
+    HarmonyEncodingName,
+    HarmonyError,  # pyright: ignore[reportUnknownVariableType]
+    Role,
+    StreamableParser,
+    load_harmony_encoding,
+)
+
+from exo.shared.types.api import ToolCallItem
+from exo.shared.types.worker.runner_response import GenerationResponse, ToolCallResponse
+from exo.worker.runner.bootstrap import logger
+from exo.worker.runner.llm_inference.tool_parsers import ToolParser
+
+
+@cache
+def get_gpt_oss_encoding():
+    encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
+    return encoding
+
+
+def parse_gpt_oss(
+    responses: Generator[GenerationResponse | None],
+) -> Generator[GenerationResponse | ToolCallResponse | None]:
+    encoding = get_gpt_oss_encoding()
+    stream = StreamableParser(encoding, role=Role.ASSISTANT)
+    thinking = False
+    current_tool_name: str | None = None
+    tool_arg_parts: list[str] = []
+
+    for response in responses:
+        if response is None:
+            yield None
+            continue
+        try:
+            stream.process(response.token)
+        except HarmonyError:
+            logger.error("Encountered critical Harmony Error, returning early")
+            return
+
+        delta = stream.last_content_delta
+        ch = stream.current_channel
+        recipient = stream.current_recipient
+
+        # Debug: log every token with state
+        logger.debug(
+            f"parse_gpt_oss token={response.token} text={response.text!r} "
+            f"recipient={recipient!r} ch={ch!r} delta={delta!r} "
+            f"state={stream.state} current_tool={current_tool_name!r}"
+        )
+
+        if recipient != current_tool_name:
+            if current_tool_name is not None:
+                prefix = "functions."
+                if current_tool_name.startswith(prefix):
+                    current_tool_name = current_tool_name[len(prefix) :]
+                logger.info(
+                    f"parse_gpt_oss yielding tool call: name={current_tool_name!r}"
+                )
+                yield ToolCallResponse(
+                    tool_calls=[
+                        ToolCallItem(
+                            name=current_tool_name,
+                            arguments="".join(tool_arg_parts).strip(),
+                        )
+                    ],
+                    usage=response.usage,
+                )
+                tool_arg_parts = []
+            current_tool_name = recipient
+
+        # If inside a tool call, accumulate arguments
+        if current_tool_name is not None:
+            if delta:
+                tool_arg_parts.append(delta)
+            continue
+
+        if ch == "analysis" and not thinking:
+            thinking = True
+
+        if ch != "analysis" and thinking:
+            thinking = False
+
+        if delta:
+            yield response.model_copy(update={"text": delta, "is_thinking": thinking})
+
+        if response.finish_reason is not None:
+            yield response
+
+
+def parse_deepseek_v32(
+    responses: Generator[GenerationResponse | None],
+) -> Generator[GenerationResponse | ToolCallResponse | None]:
+    """Parse DeepSeek V3.2 DSML tool calls from the generation stream.
+
+    Uses accumulated-text matching (not per-token marker checks) because
+    DSML markers like <｜DSML｜function_calls> may span multiple tokens.
+    Also handles <think>...</think> blocks for thinking mode.
+    """
+    from exo.worker.engines.mlx.dsml_encoding import (
+        THINKING_END,
+        THINKING_START,
+        TOOL_CALLS_END,
+        TOOL_CALLS_START,
+        parse_dsml_output,
+    )
+
+    accumulated = ""
+    in_tool_call = False
+    thinking = False
+    # Tokens buffered while we detect the start of a DSML block
+    pending_buffer: list[GenerationResponse] = []
+    # Text accumulated during a tool call block
+    tool_call_text = ""
+
+    for response in responses:
+        if response is None:
+            yield None
+            continue
+
+        # ── Handle thinking tags ──
+        if not thinking and THINKING_START in response.text:
+            thinking = True
+            # Yield any text before the <think> tag
+            before = response.text[: response.text.index(THINKING_START)]
+            if before:
+                yield response.model_copy(update={"text": before})
+            continue
+
+        if thinking and THINKING_END in response.text:
+            thinking = False
+            # Yield any text after the </think> tag
+            after = response.text[
+                response.text.index(THINKING_END) + len(THINKING_END) :
+            ]
+            if after:
+                yield response.model_copy(update={"text": after, "is_thinking": False})
+            continue
+
+        if thinking:
+            yield response.model_copy(update={"is_thinking": True})
+            continue
+
+        # ── Handle tool call accumulation ──
+        if in_tool_call:
+            tool_call_text += response.text
+            if TOOL_CALLS_END in tool_call_text:
+                # Parse the accumulated DSML block
+                parsed = parse_dsml_output(tool_call_text)
+                if parsed is not None:
+                    logger.info(f"parsed DSML tool calls: {parsed}")
+                    yield ToolCallResponse(
+                        tool_calls=parsed,
+                        usage=response.usage,
+                        stats=response.stats,
+                    )
+                else:
+                    logger.warning(
+                        f"DSML tool call parsing failed for: {tool_call_text}"
+                    )
+                    yield response.model_copy(update={"text": tool_call_text})
+                in_tool_call = False
+                tool_call_text = ""
+                continue
+
+            # EOS reached before end marker — yield buffered text as-is
+            if response.finish_reason is not None:
+                logger.info("DSML tool call parsing interrupted by EOS")
+                yield response.model_copy(update={"text": tool_call_text})
+                in_tool_call = False
+                tool_call_text = ""
+            continue
+
+        # ── Detect start of tool call block ──
+        accumulated += response.text
+
+        if TOOL_CALLS_START in accumulated:
+            # The start marker might be split across pending_buffer + current token
+            start_idx = accumulated.index(TOOL_CALLS_START)
+            # Yield any pending tokens that are purely before the marker
+            pre_text = accumulated[:start_idx]
+            if pre_text:
+                # Flush pending buffer tokens that contributed text before the marker
+                for buf_resp in pending_buffer:
+                    if pre_text:
+                        chunk = buf_resp.text
+                        if len(chunk) <= len(pre_text):
+                            yield buf_resp
+                            pre_text = pre_text[len(chunk) :]
+                        else:
+                            yield buf_resp.model_copy(update={"text": pre_text})
+                            pre_text = ""
+            pending_buffer = []
+            tool_call_text = accumulated[start_idx:]
+            accumulated = ""
+
+            # Check if the end marker is already present (entire tool call in one token)
+            if TOOL_CALLS_END in tool_call_text:
+                parsed = parse_dsml_output(tool_call_text)
+                if parsed is not None:
+                    logger.info(f"parsed DSML tool calls: {parsed}")
+                    yield ToolCallResponse(
+                        tool_calls=parsed,
+                        usage=response.usage,
+                        stats=response.stats,
+                    )
+                else:
+                    logger.warning(
+                        f"DSML tool call parsing failed for: {tool_call_text}"
+                    )
+                    yield response.model_copy(update={"text": tool_call_text})
+                tool_call_text = ""
+            else:
+                in_tool_call = True
+            continue
+
+        # Check if accumulated text might be the start of a DSML marker
+        # Buffer tokens if we see a partial match at the end
+        if _could_be_dsml_prefix(accumulated):
+            pending_buffer.append(response)
+            continue
+
+        # No partial match — flush all pending tokens and the current one
+        for buf_resp in pending_buffer:
+            yield buf_resp
+        pending_buffer = []
+        accumulated = ""
+        yield response
+
+    # Flush any remaining pending buffer at generator end
+    for buf_resp in pending_buffer:
+        yield buf_resp
+
+
+def _could_be_dsml_prefix(text: str) -> bool:
+    """Check if the end of text could be the start of a DSML function_calls marker.
+
+    We look for suffixes of text that are prefixes of the TOOL_CALLS_START pattern.
+    This allows us to buffer tokens until we can determine if a tool call is starting.
+    """
+    from exo.worker.engines.mlx.dsml_encoding import TOOL_CALLS_START
+
+    # Only check the last portion of text that could overlap with the marker
+    max_check = len(TOOL_CALLS_START)
+    tail = text[-max_check:] if len(text) > max_check else text
+
+    # Check if any suffix of tail is a prefix of TOOL_CALLS_START
+    for i in range(len(tail)):
+        suffix = tail[i:]
+        if TOOL_CALLS_START.startswith(suffix):
+            return True
+    return False
+
+
+def parse_thinking_models(
+    responses: Generator[GenerationResponse | None],
+    tokenizer: TokenizerWrapper,
+    starts_in_thinking: bool = True,
+) -> Generator[GenerationResponse | None]:
+    """Route thinking tokens via is_thinking flag.
+
+    Swallows think tag tokens, sets is_thinking on all others.
+    Always yields tokens with finish_reason to avoid hanging the chunk stream.
+    """
+    in_thinking = starts_in_thinking
+    for response in responses:
+        if response is None:
+            yield None
+            continue
+        if isinstance(response, ToolCallResponse):
+            yield response
+            continue
+
+        is_think_tag = (
+            tokenizer.think_end is not None and response.text == tokenizer.think_end
+        ) or (
+            tokenizer.think_start is not None and response.text == tokenizer.think_start
+        )
+
+        if is_think_tag:
+            in_thinking = response.text != tokenizer.think_end
+            # Never swallow finish_reason — the chunk stream needs it to terminate.
+            if response.finish_reason is not None:
+                yield response.model_copy(update={"text": "", "is_thinking": False})
+            continue
+        yield response.model_copy(update={"is_thinking": in_thinking})
+
+
+def parse_tool_calls(
+    responses: Generator[GenerationResponse | None], tool_parser: ToolParser
+) -> Generator[GenerationResponse | ToolCallResponse | None]:
+    in_tool_call = False
+    tool_call_text_parts: list[str] = []
+    for response in responses:
+        if response is None:
+            yield None
+            continue
+        if not in_tool_call and response.text.startswith(tool_parser.start_parsing):
+            in_tool_call = True
+
+        if in_tool_call:
+            tool_call_text_parts.append(response.text)
+            if response.text.endswith(tool_parser.end_parsing):
+                # parse the actual tool calls from the tool call text
+                parsed = tool_parser.parse_tool_calls(
+                    "".join(tool_call_text_parts).strip()
+                )
+                logger.info(f"parsed {tool_call_text_parts=} into {parsed=}")
+                if parsed is not None:
+                    yield ToolCallResponse(
+                        tool_calls=parsed, usage=response.usage, stats=response.stats
+                    )
+                else:
+                    logger.warning(
+                        f"tool call parsing failed for text {''.join(tool_call_text_parts)}"
+                    )
+                    response.text = "".join(tool_call_text_parts)
+                    yield response
+
+                in_tool_call = False
+                tool_call_text_parts = []
+                continue
+
+            if response.finish_reason is not None:
+                logger.info(
+                    "tool call parsing interrupted, yield partial tool call as text"
+                )
+                response = response.model_copy(
+                    update={
+                        "text": "".join(tool_call_text_parts),
+                        "token": 0,
+                    }
+                )
+                yield response
+
+        else:
+            # fallthrough
+            yield response

src/exo/worker/runner/runner_supervisor.py

@@ -172,7 +172,7 @@ class RunnerSupervisor:
                     if isinstance(event, RunnerStatusUpdated):
                         self.status = event.runner_status
                     if isinstance(event, TaskAcknowledged):
-                        self.pending.pop(event.task_id).set()
+                        self.pending[event.task_id].set()
                         continue
                     if (
                         isinstance(event, TaskStatusUpdated)
@@ -190,6 +190,7 @@ class RunnerSupervisor:
                             ),
                         )
                         self.completed.add(event.task_id)
+                        self.pending.pop(event.task_id, None)
                     await self._event_sender.send(event)
         except (ClosedResourceError, BrokenResourceError) as e:
             await self._check_runner(e)

src/exo/worker/tests/unittests/conftest.py

@@ -20,6 +20,7 @@ class FakeRunnerSupervisor:
     bound_instance: BoundInstance
     status: RunnerStatus
     completed: set[TaskId] = field(default_factory=set)
+    pending: dict[TaskId, object] = field(default_factory=dict)
 
 
 class OtherTask(BaseTask):

src/exo/worker/tests/unittests/test_mlx/conftest.py

@@ -14,9 +14,9 @@ from exo.shared.constants import EXO_MODELS_DIR
 from exo.shared.models.model_cards import ModelCard, ModelTask
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
+from exo.shared.types.mlx import Model
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.shared.types.worker.shards import PipelineShardMetadata, TensorShardMetadata
-from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.generator.generate import mlx_generate
 from exo.worker.engines.mlx.utils_mlx import apply_chat_template, shard_and_load
 

src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py

@@ -9,8 +9,8 @@ from mlx_lm.models.cache import KVCache
 from mlx_lm.sample_utils import make_sampler
 
 from exo.shared.types.common import ModelId
+from exo.shared.types.mlx import Model
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
-from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.cache import (
     KVPrefixCache,
     cache_length,
@@ -143,7 +143,14 @@ class TestKVPrefixCacheWithModel:
         cache = make_kv_cache(model)
 
         _, _, snapshots = prefill(
-            model, tokenizer, make_sampler(0.0), tokens, cache, group=None
+            model,
+            tokenizer,
+            make_sampler(0.0),
+            tokens,
+            cache,
+            group=None,
+            on_prefill_progress=None,
+            distributed_prompt_progress_callback=None,
         )
 
         # Cache should now hold the prompt tokens minus one
@@ -164,7 +171,14 @@ class TestKVPrefixCacheWithModel:
         cache = make_kv_cache(model)
 
         _, _, snapshots = prefill(
-            model, tokenizer, make_sampler(0.0), tokens, cache, group=None
+            model,
+            tokenizer,
+            make_sampler(0.0),
+            tokens,
+            cache,
+            group=None,
+            on_prefill_progress=None,
+            distributed_prompt_progress_callback=None,
         )
 
         kv_prefix_cache = KVPrefixCache(None)
@@ -200,7 +214,14 @@ class TestKVPrefixCacheWithModel:
         cache = make_kv_cache(model)
 
         _, _, snapshots = prefill(
-            model, tokenizer, make_sampler(0.0), short_tokens, cache, group=None
+            model,
+            tokenizer,
+            make_sampler(0.0),
+            short_tokens,
+            cache,
+            group=None,
+            on_prefill_progress=None,
+            distributed_prompt_progress_callback=None,
         )
 
         kv_prefix_cache = KVPrefixCache(None)
@@ -245,7 +266,14 @@ class TestKVPrefixCacheWithModel:
         cache = make_kv_cache(model)
 
         _, _, snapshots = prefill(
-            model, tokenizer, make_sampler(0.0), tokens, cache, group=None
+            model,
+            tokenizer,
+            make_sampler(0.0),
+            tokens,
+            cache,
+            group=None,
+            on_prefill_progress=None,
+            distributed_prompt_progress_callback=None,
         )
 
         kv_prefix_cache = KVPrefixCache(None)
@@ -285,7 +313,14 @@ class TestKVPrefixCacheWithModel:
         cache = make_kv_cache(model)
 
         _, _, snapshots = prefill(
-            model, tokenizer, make_sampler(0.0), tokens, cache, group=None
+            model,
+            tokenizer,
+            make_sampler(0.0),
+            tokens,
+            cache,
+            group=None,
+            on_prefill_progress=None,
+            distributed_prompt_progress_callback=None,
         )
 
         kv_prefix_cache = KVPrefixCache(None)
@@ -513,7 +548,16 @@ class TestKVPrefixCacheWithModel:
             prompt = apply_chat_template(tokenizer, task)
             tokens = encode_prompt(tokenizer, prompt)
             cache = make_kv_cache(model)
-            prefill(model, tokenizer, make_sampler(0.0), tokens, cache, group=None)
+            prefill(
+                model,
+                tokenizer,
+                make_sampler(0.0),
+                tokens,
+                cache,
+                group=None,
+                on_prefill_progress=None,
+                distributed_prompt_progress_callback=None,
+            )
             kv_prefix_cache.add_kv_cache(tokens, cache)
             # Stagger _last_used so LRU order is deterministic
             kv_prefix_cache._last_used[i] = float(i)
@@ -538,7 +582,16 @@ class TestKVPrefixCacheWithModel:
             prompt = apply_chat_template(tokenizer, task)
             tokens = encode_prompt(tokenizer, prompt)
             cache = make_kv_cache(model)
-            prefill(model, tokenizer, make_sampler(0.0), tokens, cache, group=None)
+            prefill(
+                model,
+                tokenizer,
+                make_sampler(0.0),
+                tokens,
+                cache,
+                group=None,
+                on_prefill_progress=None,
+                distributed_prompt_progress_callback=None,
+            )
             kv_prefix_cache.add_kv_cache(tokens, cache)
 
         # LRU entries should have been evicted (entries 0, 1, 2 in order of _last_used)

src/exo/worker/tests/unittests/test_mlx/test_pipeline_prefill_callbacks.py

@@ -0,0 +1,512 @@
+# type: ignore
+"""Test that pipeline prefill callbacks and output exactly match stream_generate.
+
+Spins up a single-device (non-pipeline) run and a distributed pipeline run,
+then verifies that the prompt_progress_callback sequences are identical
+and that generated text matches.
+"""
+
+import json
+import multiprocessing as mp
+import os
+import tempfile
+import traceback
+from typing import Any, cast
+
+import pytest
+
+from exo.shared.constants import EXO_MODELS_DIR
+from exo.shared.models.model_cards import ModelCard, ModelTask
+from exo.shared.types.common import ModelId
+from exo.shared.types.memory import Memory
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
+
+MODEL_ID = "mlx-community/gpt-oss-20b-MXFP4-Q8"
+MODEL_PATH = EXO_MODELS_DIR / "mlx-community--gpt-oss-20b-MXFP4-Q8"
+TOTAL_LAYERS = 24
+MAX_TOKENS = 10
+SEED = 42
+TEMPERATURE = 0.0
+
+
+def _model_card() -> ModelCard:
+    return ModelCard(
+        model_id=ModelId(MODEL_ID),
+        storage_size=Memory.from_gb(12),
+        n_layers=TOTAL_LAYERS,
+        hidden_size=2880,
+        supports_tensor=False,
+        tasks=[ModelTask.TextGeneration],
+    )
+
+
+def _build_prompt(tokenizer: Any, prompt_tokens: int) -> tuple[str, Any]:
+    """Build a prompt with the given number of user-content tokens, return (chat_prompt, task)."""
+    from exo.worker.engines.mlx.utils_mlx import apply_chat_template
+
+    base_text = "The quick brown fox jumps over the lazy dog. "
+    base_toks = tokenizer.encode(base_text)
+    repeats = (prompt_tokens // len(base_toks)) + 2
+    long_text = base_text * repeats
+    tokens = tokenizer.encode(long_text)[:prompt_tokens]
+    prompt_text = tokenizer.decode(tokens)
+
+    task = TextGenerationTaskParams(
+        model=MODEL_ID,
+        input=[InputMessage(role="user", content=prompt_text)],
+        max_output_tokens=MAX_TOKENS,
+        temperature=TEMPERATURE,
+        seed=SEED,
+    )
+
+    prompt = apply_chat_template(tokenizer, task)
+    return prompt, task
+
+
+# ---------------------------------------------------------------------------
+# Single-device process: uses stream_generate path (no pipeline layers)
+# ---------------------------------------------------------------------------
+def _run_single_device(
+    prompt_tokens: int,
+    result_queue: Any,
+) -> None:
+    """Load full model without pipeline sharding, run mlx_generate, record callbacks."""
+    try:
+        import mlx.core as mx
+        from mlx_lm.utils import load_model
+
+        from exo.shared.types.worker.shards import PipelineShardMetadata
+        from exo.worker.engines.mlx.cache import encode_prompt
+        from exo.worker.engines.mlx.generator.generate import mlx_generate
+        from exo.worker.engines.mlx.utils_mlx import (
+            build_model_path,
+            get_tokenizer,
+        )
+
+        model_path = build_model_path(ModelId(MODEL_ID))
+        model, _ = load_model(model_path, lazy=True, strict=False)
+        mx.eval(model)
+
+        # Use PipelineShardMetadata just for get_tokenizer (needs model_card), but
+        # do NOT apply pipeline sharding — the model keeps all layers unwrapped.
+        dummy_meta = PipelineShardMetadata(
+            model_card=_model_card(),
+            device_rank=0,
+            world_size=1,
+            start_layer=0,
+            end_layer=TOTAL_LAYERS,
+            n_layers=TOTAL_LAYERS,
+        )
+        tokenizer = get_tokenizer(model_path, dummy_meta)
+
+        prompt, task = _build_prompt(tokenizer, prompt_tokens)
+
+        callbacks: list[tuple[int, int]] = []
+
+        def on_progress(processed: int, total: int) -> None:
+            callbacks.append((processed, total))
+
+        generated_text = ""
+        for response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=None,
+            group=None,
+            on_prefill_progress=on_progress,
+        ):
+            generated_text += response.text
+            if response.finish_reason is not None:
+                break
+
+        # Also record the token count that prefill() received (prompt_tokens[:-1])
+        all_tokens = encode_prompt(tokenizer, prompt)
+        prefill_token_count = len(all_tokens) - 1
+
+        result_queue.put(
+            (
+                True,
+                {
+                    "callbacks": callbacks,
+                    "text": generated_text,
+                    "prefill_token_count": prefill_token_count,
+                },
+            )
+        )
+
+    except Exception as e:
+        result_queue.put((False, f"{e}\n{traceback.format_exc()}"))
+
+
+# ---------------------------------------------------------------------------
+# Pipeline device process: uses _pipeline_prefill_cache path
+# ---------------------------------------------------------------------------
+def _run_pipeline_device(
+    rank: int,
+    world_size: int,
+    hostfile_path: str,
+    layer_splits: list[tuple[int, int]],
+    prompt_tokens: int,
+    result_queue: Any,
+) -> None:
+    """Load model with pipeline sharding, run mlx_generate, record callbacks."""
+    os.environ["MLX_HOSTFILE"] = hostfile_path
+    os.environ["MLX_RANK"] = str(rank)
+
+    try:
+        import mlx.core as mx
+
+        from exo.shared.types.worker.shards import PipelineShardMetadata
+        from exo.worker.engines.mlx.cache import encode_prompt
+        from exo.worker.engines.mlx.generator.generate import mlx_generate
+        from exo.worker.engines.mlx.utils_mlx import shard_and_load
+
+        group = mx.distributed.init(backend="ring", strict=True)
+
+        start_layer, end_layer = layer_splits[rank]
+        shard_meta = PipelineShardMetadata(
+            model_card=_model_card(),
+            device_rank=rank,
+            world_size=world_size,
+            start_layer=start_layer,
+            end_layer=end_layer,
+            n_layers=TOTAL_LAYERS,
+        )
+
+        model, tokenizer = shard_and_load(
+            shard_meta, group, on_timeout=None, on_layer_loaded=None
+        )
+        model = cast(Any, model)
+
+        prompt, task = _build_prompt(tokenizer, prompt_tokens)
+
+        callbacks: list[tuple[int, int]] = []
+
+        def on_progress(processed: int, total: int) -> None:
+            callbacks.append((processed, total))
+
+        def distributed_prompt_progress_callback(_group: Any = group) -> None:
+            from exo.worker.engines.mlx.utils_mlx import mx_any
+
+            mx_any(False, _group)
+
+        generated_text = ""
+        for response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=None,
+            group=group,
+            on_prefill_progress=on_progress,
+            distributed_prompt_progress_callback=distributed_prompt_progress_callback,
+        ):
+            generated_text += response.text
+            if response.finish_reason is not None:
+                break
+
+        all_tokens = encode_prompt(tokenizer, prompt)
+        prefill_token_count = len(all_tokens) - 1
+
+        result_queue.put(
+            (
+                rank,
+                True,
+                {
+                    "callbacks": callbacks,
+                    "text": generated_text,
+                    "prefill_token_count": prefill_token_count,
+                },
+            )
+        )
+
+    except Exception as e:
+        result_queue.put((rank, False, f"{e}\n{traceback.format_exc()}"))
+
+
+# ---------------------------------------------------------------------------
+# Test helpers
+# ---------------------------------------------------------------------------
+def _create_hostfile(world_size: int, base_port: int) -> str:
+    hosts = [f"127.0.0.1:{base_port + i}" for i in range(world_size)]
+    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
+        json.dump(hosts, f)
+        return f.name
+
+
+def _run_single_device_test(prompt_tokens: int, timeout: int = 120) -> dict[str, Any]:
+    """Run single-device (stream_generate) prefill and return results."""
+    ctx = mp.get_context("spawn")
+    result_queue: Any = ctx.Queue()
+
+    p = ctx.Process(target=_run_single_device, args=(prompt_tokens, result_queue))
+    p.start()
+    p.join(timeout=timeout)
+
+    if p.is_alive():
+        p.terminate()
+        p.join(timeout=5)
+        pytest.fail("Single-device process timed out")
+
+    assert not result_queue.empty(), "Single-device process produced no result"
+    success, data = result_queue.get()
+    assert success, f"Single-device process failed:\n{data}"
+    return data
+
+
+def _run_pipeline_test(
+    layer_splits: list[tuple[int, int]],
+    prompt_tokens: int,
+    base_port: int,
+    timeout: int = 120,
+) -> dict[int, dict[str, Any]]:
+    """Run pipeline prefill across ranks and return per-rank results."""
+    world_size = len(layer_splits)
+    hostfile_path = _create_hostfile(world_size, base_port)
+    ctx = mp.get_context("spawn")
+    result_queue: Any = ctx.Queue()
+
+    try:
+        processes: list[Any] = []
+        for rank in range(world_size):
+            p = ctx.Process(
+                target=_run_pipeline_device,
+                args=(
+                    rank,
+                    world_size,
+                    hostfile_path,
+                    layer_splits,
+                    prompt_tokens,
+                    result_queue,
+                ),
+            )
+            p.start()
+            processes.append(p)
+
+        for p in processes:
+            p.join(timeout=timeout)
+
+        timed_out = any(p.is_alive() for p in processes)
+        for p in processes:
+            if p.is_alive():
+                p.terminate()
+                p.join(timeout=5)
+
+        assert not timed_out, "Pipeline processes timed out"
+
+        results: dict[int, dict[str, Any]] = {}
+        while not result_queue.empty():
+            rank, success, data = result_queue.get()
+            assert success, f"Pipeline rank {rank} failed:\n{data}"
+            results[rank] = data
+
+        assert len(results) == world_size, (
+            f"Expected {world_size} results, got {len(results)}: missing ranks {set(range(world_size)) - results.keys()}"
+        )
+        return results
+
+    finally:
+        os.unlink(hostfile_path)
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+pytestmark = [
+    pytest.mark.slow,
+    pytest.mark.skipif(
+        not MODEL_PATH.exists(),
+        reason=f"GPT-OSS model not found at {MODEL_PATH}",
+    ),
+]
+
+LAYER_SPLITS_4WAY: list[tuple[int, int]] = [(0, 6), (6, 12), (12, 18), (18, 24)]
+LAYER_SPLITS_2WAY: list[tuple[int, int]] = [(0, 12), (12, 24)]
+
+
+class TestPipelineNoDeadlock:
+    """Pipeline prefill must not deadlock at any rank count or prompt length."""
+
+    @pytest.mark.parametrize(
+        "layer_splits,prompt_tokens",
+        [
+            (LAYER_SPLITS_2WAY, 128),
+            (LAYER_SPLITS_2WAY, 4096),
+            (LAYER_SPLITS_2WAY, 8192),
+            (LAYER_SPLITS_2WAY, 16384),
+            (LAYER_SPLITS_4WAY, 128),
+            (LAYER_SPLITS_4WAY, 4096),
+            (LAYER_SPLITS_4WAY, 8192),
+            (LAYER_SPLITS_4WAY, 16384),
+        ],
+        ids=[
+            "2rank_128tok",
+            "2rank_4096tok",
+            "2rank_8192tok",
+            "2rank_16384tok",
+            "4rank_128tok",
+            "4rank_4096tok",
+            "4rank_8192tok",
+            "4rank_16384tok",
+        ],
+    )
+    def test_no_deadlock(
+        self,
+        layer_splits: list[tuple[int, int]],
+        prompt_tokens: int,
+    ) -> None:
+        """Pipeline must complete without deadlock at various prompt lengths."""
+        pipeline_results = _run_pipeline_test(
+            layer_splits=layer_splits,
+            prompt_tokens=prompt_tokens,
+            base_port=29650,
+            timeout=60,
+        )
+        # If we get here, no deadlock. Verify all ranks produced output.
+        for rank, pipe_data in sorted(pipeline_results.items()):
+            assert pipe_data["text"], f"Rank {rank} produced no output text"
+
+
+class TestPipelinePrefillCallbacks:
+    """Verify that pipeline prefill callbacks exactly match stream_generate callbacks."""
+
+    @pytest.mark.parametrize(
+        "prompt_tokens",
+        [50, 500, 5000],
+        ids=["short_50", "medium_500", "long_5000"],
+    )
+    def test_callbacks_match(self, prompt_tokens: int) -> None:
+        """All pipeline ranks must produce identical callback sequences."""
+        # Run 4-rank pipeline
+        pipeline_results = _run_pipeline_test(
+            layer_splits=LAYER_SPLITS_4WAY,
+            prompt_tokens=prompt_tokens,
+            base_port=29700,
+            timeout=180,
+        )
+
+        # All ranks must agree on prefill token count and callback sequence
+        rank0_data = pipeline_results[0]
+        rank0_callbacks = rank0_data["callbacks"]
+        prefill_count = rank0_data["prefill_token_count"]
+
+        for rank, pipe_data in sorted(pipeline_results.items()):
+            pipe_callbacks = pipe_data["callbacks"]
+
+            assert pipe_data["prefill_token_count"] == prefill_count, (
+                f"Rank {rank} prefill token count mismatch: "
+                f"{pipe_data['prefill_token_count']} vs {prefill_count}"
+            )
+
+            assert pipe_callbacks == rank0_callbacks, (
+                f"Rank {rank} callback mismatch for {prompt_tokens} prompt tokens "
+                f"(prefill M={prefill_count}):\n"
+                f"  pipeline R0 ({len(rank0_callbacks)} callbacks): {rank0_callbacks}\n"
+                f"  pipeline R{rank} ({len(pipe_callbacks)} callbacks): {pipe_callbacks}"
+            )
+
+        # Structural checks: starts with (0, M), ends with (M, M), monotonically increasing
+        assert rank0_callbacks[0] == (0, prefill_count), (
+            f"First callback should be (0, {prefill_count}), got {rank0_callbacks[0]}"
+        )
+        assert rank0_callbacks[-1] == (prefill_count, prefill_count), (
+            f"Last callback should be ({prefill_count}, {prefill_count}), got {rank0_callbacks[-1]}"
+        )
+        for i in range(1, len(rank0_callbacks)):
+            assert rank0_callbacks[i][0] >= rank0_callbacks[i - 1][0], (
+                f"Callbacks not monotonically increasing at index {i}: {rank0_callbacks}"
+            )
+
+    @pytest.mark.parametrize(
+        "prompt_tokens",
+        [50, 500],
+        ids=["short_50", "medium_500"],
+    )
+    def test_output_matches(self, prompt_tokens: int) -> None:
+        """Pipeline-generated text must match single-device output."""
+        single = _run_single_device_test(prompt_tokens, timeout=180)
+
+        pipeline_results = _run_pipeline_test(
+            layer_splits=LAYER_SPLITS_4WAY,
+            prompt_tokens=prompt_tokens,
+            base_port=29800,
+            timeout=180,
+        )
+
+        single_text = single["text"]
+
+        # The last rank produces the final logits, so its output should match.
+        # Due to SDPA tiling non-determinism, allow minor differences in text.
+        last_rank = max(pipeline_results.keys())
+        pipe_text = pipeline_results[last_rank]["text"]
+
+        # For deterministic sampling (temp=0.0), outputs should match exactly
+        # or be very close. Log both for debugging even if they match.
+        if single_text != pipe_text:
+            # Find first divergence point
+            min_len = min(len(single_text), len(pipe_text))
+            diverge_idx = next(
+                (i for i in range(min_len) if single_text[i] != pipe_text[i]),
+                min_len,
+            )
+            pytest.fail(
+                f"Output text diverged at character {diverge_idx} for {prompt_tokens} prompt tokens:\n"
+                f"  single-device: {single_text!r}\n"
+                f"  pipeline R{last_rank}: {pipe_text!r}"
+            )
+
+
+class TestPipelineCallbacksStructure:
+    """Verify structural properties of callbacks independent of model output."""
+
+    def test_callback_structure_matches_generate_step(self) -> None:
+        """Verify callbacks follow generate_step's pattern: (0,M), chunks up to M-1, (M,M)."""
+        prompt_tokens = 200
+        pipeline_results = _run_pipeline_test(
+            layer_splits=LAYER_SPLITS_4WAY,
+            prompt_tokens=prompt_tokens,
+            base_port=29900,
+            timeout=180,
+        )
+
+        for rank, pipe_data in sorted(pipeline_results.items()):
+            callbacks = pipe_data["callbacks"]
+            m = pipe_data["prefill_token_count"]
+            assert m > 0, f"Rank {rank}: prefill token count is 0"
+
+            assert callbacks[0] == (0, m), (
+                f"Rank {rank}: first callback should be (0, {m}), got {callbacks[0]}"
+            )
+
+            assert callbacks[-1] == (m, m), (
+                f"Rank {rank}: last callback should be ({m}, {m}), got {callbacks[-1]}"
+            )
+
+            if len(callbacks) > 2:
+                second_to_last = callbacks[-2]
+                assert second_to_last[0] < m, (
+                    f"Rank {rank}: second-to-last callback should report < {m}, "
+                    f"got {second_to_last}"
+                )
+
+            # All callbacks must have total == M
+            for i, (_, total) in enumerate(callbacks):
+                assert total == m, (
+                    f"Rank {rank}: callback {i} has total={total}, expected {m}"
+                )
+
+            # processed values must be non-decreasing
+            processed_vals = [p for p, _ in callbacks]
+            for i in range(1, len(processed_vals)):
+                assert processed_vals[i] >= processed_vals[i - 1], (
+                    f"Rank {rank}: callbacks not non-decreasing at index {i}: "
+                    f"{processed_vals}"
+                )
+
+            # No duplicate consecutive callbacks (pipeline dummies must not emit callbacks)
+            for i in range(1, len(callbacks)):
+                assert callbacks[i] != callbacks[i - 1], (
+                    f"Rank {rank}: duplicate consecutive callback at index {i}: "
+                    f"{callbacks[i]} (this suggests dummy iterations are emitting callbacks)"
+                )

src/exo/worker/tests/unittests/test_mlx/test_prefix_cache_architectures.py

@@ -15,8 +15,8 @@ from mlx.utils import tree_flatten, tree_unflatten
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.shared.types.common import ModelId
+from exo.shared.types.mlx import Model
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
-from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.cache import KVPrefixCache
 from exo.worker.engines.mlx.generator.generate import mlx_generate
 from exo.worker.engines.mlx.utils_mlx import (

src/exo/worker/tests/unittests/test_runner/test_dsml_e2e.py

@@ -19,7 +19,7 @@ from exo.worker.engines.mlx.dsml_encoding import (
     encode_messages,
     parse_dsml_output,
 )
-from exo.worker.runner.llm_inference.runner import parse_deepseek_v32
+from exo.worker.runner.llm_inference.model_output_parsers import parse_deepseek_v32
 
 # ── Shared fixtures ──────────────────────────────────────────────
 

src/exo/worker/tests/unittests/test_runner/test_event_ordering.py

@@ -6,6 +6,7 @@ from typing import Callable
 import mlx.core as mx
 import pytest
 
+import exo.worker.runner.llm_inference.batch_generator as mlx_batch_generator
 import exo.worker.runner.llm_inference.runner as mlx_runner
 from exo.shared.types.chunks import TokenChunk
 from exo.shared.types.events import (
@@ -115,17 +116,20 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.setattr(mlx_runner, "initialize_mlx", make_nothin(MockGroup()))
     monkeypatch.setattr(mlx_runner, "load_mlx_items", make_nothin((1, MockTokenizer)))
     monkeypatch.setattr(mlx_runner, "warmup_inference", make_nothin(1))
-    monkeypatch.setattr(mlx_runner, "_check_for_debug_prompts", nothin)
-    monkeypatch.setattr(mlx_runner, "mx_any", make_nothin(False))
+    monkeypatch.setattr(mlx_batch_generator, "_check_for_debug_prompts", nothin)
+    monkeypatch.setattr(mlx_batch_generator, "mx_any", make_nothin(False))
     # Mock apply_chat_template since we're using a fake tokenizer (integer 1).
     # Returns a prompt without thinking tag so detect_thinking_prompt_suffix returns None.
     monkeypatch.setattr(mlx_runner, "apply_chat_template", make_nothin("test prompt"))
+    monkeypatch.setattr(
+        mlx_batch_generator, "apply_chat_template", make_nothin("test prompt")
+    )
     monkeypatch.setattr(mlx_runner, "detect_thinking_prompt_suffix", make_nothin(False))
 
     def fake_generate(*_1: object, **_2: object):
         yield GenerationResponse(token=0, text="hi", finish_reason="stop", usage=None)
 
-    monkeypatch.setattr(mlx_runner, "mlx_generate", fake_generate)
+    monkeypatch.setattr(mlx_batch_generator, "mlx_generate", fake_generate)
 
 
 # Use a fake event_sender to remove test flakiness.
@@ -183,12 +187,13 @@ def _run(tasks: Iterable[Task]):
             "exo.worker.runner.llm_inference.runner.mx.distributed.all_gather",
             make_nothin(mx.array([1])),
         ):
-            mlx_runner.main(
+            runner = mlx_runner.Runner(
                 bound_instance,
                 event_sender,  # pyright: ignore[reportArgumentType]
                 task_receiver,
                 cancel_receiver,
             )
+            runner.main()
 
         return event_sender.events
 

src/exo/worker/tests/unittests/test_runner/test_parse_gpt_oss.py

@@ -4,7 +4,7 @@ from exo.shared.types.worker.runner_response import (
     GenerationResponse,
     ToolCallResponse,
 )
-from exo.worker.runner.llm_inference.runner import parse_gpt_oss
+from exo.worker.runner.llm_inference.model_output_parsers import parse_gpt_oss
 
 # Token IDs from mlx-community/gpt-oss-20b-MXFP4-Q8 tokenizer.
 # These are stable since they come from the model's vocabulary.
@@ -107,7 +107,7 @@ def _collect(
     def _gen() -> Generator[GenerationResponse, None, None]:
         yield from _make_gen_responses(tokens)
 
-    return list(parse_gpt_oss(_gen()))
+    return list(x for x in parse_gpt_oss(_gen()) if x is not None)
 
 
 def _get_tool_call(

src/exo/worker/tests/unittests/test_runner/test_parse_tool_calls.py

@@ -4,7 +4,7 @@ from collections.abc import Generator
 from typing import Any
 
 from exo.shared.types.worker.runner_response import GenerationResponse, ToolCallResponse
-from exo.worker.runner.llm_inference.runner import parse_tool_calls
+from exo.worker.runner.llm_inference.model_output_parsers import parse_tool_calls
 from exo.worker.runner.llm_inference.tool_parsers import make_mlx_parser