yay

2026-01-30 16:51:02 -05:00 · 2026-01-30 11:53:18 +00:00
15 changed files with 367 additions and 303 deletions
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -19,7 +19,7 @@ dependencies = [
    "anyio==4.11.0",
    "mlx==0.30.4; sys_platform == 'darwin'",
    "mlx[cpu]==0.30.4; sys_platform == 'linux'",
-    "mlx-lm==0.30.5",
+    "mlx-lm",
    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
    "hypercorn>=0.18.0",
    "openai-harmony>=0.0.8",
@@ -31,6 +31,8 @@ dependencies = [
 ]

 [project.scripts]
+exo-master = "exo.master.main:main"
+exo-worker = "exo.worker.main:main"
 exo = "exo.main:main"

 # dependencies only required for development
@@ -44,6 +46,12 @@ dev = [
    "ruff>=0.11.13",
 ]

+# mlx[cuda] requires a newer version of mlx. the ideal on linux is: default to mlx[cpu] unless[cuda] specified.
+[project.optional-dependencies]
+# cuda = [
+#     "mlx[cuda]==0.26.3",
+# ]
+
 ###
 # workspace configuration
 ###
@@ -55,8 +63,8 @@ members = [

 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
+mlx-lm = { git = "https://github.com/ml-explore/mlx-lm", branch = "main" }
 # Uncomment to use local mlx/mlx-lm development versions:
-# mlx-lm = { git = "https://github.com/ml-explore/mlx-lm", branch = "main" }
 # mlx = { path = "/Users/Shared/mlx", editable=true }
 # mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }

@@ -108,7 +116,7 @@ environments = [
 ###

 [tool.ruff]
-extend-exclude = ["*mlx_typings/**", "rust/exo_pyo3_bindings/**"]
+extend-exclude = ["shared/protobufs/**", "*mlx_typings/**", "rust/exo_pyo3_bindings/**"]

 [tool.ruff.lint]
 extend-select = ["I", "N", "B", "A", "PIE", "SIM"]
--- a/src/exo/download/impl_shard_downloader.py
+++ b/src/exo/download/impl_shard_downloader.py
@@ -166,8 +166,9 @@ class ResumableShardDownloader(ShardDownloader):
        for task in asyncio.as_completed(tasks):
            try:
                yield await task
+            # TODO: except Exception
            except Exception as e:
-                logger.warning(f"Error downloading shard: {type(e).__name__}")
+                logger.error("Error downloading shard:", e)

    async def get_shard_download_status_for_shard(
        self, shard: ShardMetadata
--- a/src/exo/master/api.py
+++ b/src/exo/master/api.py
@@ -65,9 +65,7 @@ from exo.shared.types.api import (
    StartDownloadParams,
    StartDownloadResponse,
    StreamingChoiceResponse,
-    StreamOptions,
    ToolCall,
-    Usage,
 )
 from exo.shared.types.chunks import (
    ErrorChunk,
@@ -115,9 +113,7 @@ def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None)


 def chunk_to_response(
-    chunk: TokenChunk | ToolCallChunk,
-    command_id: CommandId,
-    usage: Usage | None,
+    chunk: TokenChunk | ToolCallChunk, command_id: CommandId
 ) -> ChatCompletionResponse:
    return ChatCompletionResponse(
        id=command_id,
@@ -142,7 +138,6 @@ def chunk_to_response(
                finish_reason=chunk.finish_reason,
            )
        ],
-        usage=usage,
    )


@@ -527,10 +522,9 @@ class API:
                del self._chat_completion_queues[command_id]

    async def _generate_chat_stream(
-        self, command_id: CommandId, stream_options: StreamOptions | None = None
+        self, command_id: CommandId
    ) -> AsyncGenerator[str, None]:
        """Generate chat completion stream as JSON strings."""
-        include_usage = stream_options.include_usage if stream_options else False

        async for chunk in self._chat_chunk_stream(command_id):
            assert not isinstance(chunk, ImageChunk)
@@ -546,10 +540,8 @@ class API:
                yield "data: [DONE]\n\n"
                return

-            usage = chunk.usage if include_usage else None
-
            chunk_response: ChatCompletionResponse = chunk_to_response(
-                chunk, command_id, usage=usage
+                chunk, command_id
            )
            logger.debug(f"chunk_response: {chunk_response}")

@@ -567,7 +559,6 @@ class API:
        tool_calls: list[ToolCall] = []
        model: str | None = None
        finish_reason: FinishReason | None = None
-        usage: Usage | None = None

        async for chunk in self._chat_chunk_stream(command_id):
            if isinstance(chunk, ErrorChunk):
@@ -592,9 +583,6 @@ class API:
                    for i, tool in enumerate(chunk.tool_calls)
                )

-            if chunk.usage is not None:
-                usage = chunk.usage
-
            if chunk.finish_reason is not None:
                finish_reason = chunk.finish_reason

@@ -616,7 +604,6 @@ class API:
                    finish_reason=finish_reason,
                )
            ],
-            usage=usage,
        )

    async def _collect_chat_completion_with_stats(
@@ -704,7 +691,7 @@ class API:
        await self._send(command)
        if payload.stream:
            return StreamingResponse(
-                self._generate_chat_stream(command.command_id, payload.stream_options),
+                self._generate_chat_stream(command.command_id),
                media_type="text/event-stream",
            )

--- a/src/exo/shared/types/api.py
+++ b/src/exo/shared/types/api.py
@@ -3,7 +3,7 @@ from collections.abc import Generator
 from typing import Annotated, Any, Literal

 from fastapi import UploadFile
-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 from pydantic_core import PydanticUseDefault

 from exo.shared.models.model_cards import ModelCard, ModelId
@@ -11,7 +11,7 @@ from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
-from exo.utils.pydantic_ext import CamelCaseModel, ConfigDict, TaggedModel
+from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel

 FinishReason = Literal[
    "stop", "length", "tool_calls", "content_filter", "function_call", "error"
@@ -116,8 +116,8 @@ class Usage(BaseModel):
    prompt_tokens: int
    completion_tokens: int
    total_tokens: int
-    prompt_tokens_details: PromptTokensDetails
-    completion_tokens_details: CompletionTokensDetails
+    prompt_tokens_details: PromptTokensDetails | None = None
+    completion_tokens_details: CompletionTokensDetails | None = None


 class StreamingChoiceResponse(BaseModel):
@@ -170,10 +170,6 @@ class BenchChatCompletionResponse(ChatCompletionResponse):
    generation_stats: GenerationStats | None = None


-class StreamOptions(BaseModel):
-    include_usage: bool = False
-
-
 class ChatCompletionTaskParams(TaggedModel):
    model_config = ConfigDict(extra="ignore")

@@ -190,7 +186,6 @@ class ChatCompletionTaskParams(TaggedModel):
    seed: int | None = None
    stop: str | list[str] | None = None
    stream: bool = False
-    stream_options: StreamOptions | None = None
    temperature: float | None = None
    top_p: float | None = None
    tools: list[dict[str, Any]] | None = None
--- a/src/exo/shared/types/chunks.py
+++ b/src/exo/shared/types/chunks.py
@@ -2,7 +2,7 @@ from collections.abc import Generator
 from typing import Any, Literal

 from exo.shared.models.model_cards import ModelId
-from exo.shared.types.api import GenerationStats, ImageGenerationStats, Usage
+from exo.shared.types.api import GenerationStats, ImageGenerationStats
 from exo.utils.pydantic_ext import TaggedModel

 from .api import FinishReason
@@ -17,7 +17,6 @@ class BaseChunk(TaggedModel):
 class TokenChunk(BaseChunk):
    text: str
    token_id: int
-    usage: Usage | None
    finish_reason: Literal["stop", "length", "content_filter"] | None = None
    stats: GenerationStats | None = None

@@ -29,7 +28,6 @@ class ErrorChunk(BaseChunk):

 class ToolCallChunk(BaseChunk):
    tool_calls: list[ToolCallItem]
-    usage: Usage | None
    finish_reason: Literal["tool_calls"] = "tool_calls"
    stats: GenerationStats | None = None

--- a/src/exo/shared/types/worker/runner_response.py
+++ b/src/exo/shared/types/worker/runner_response.py
@@ -6,7 +6,6 @@ from exo.shared.types.api import (
    GenerationStats,
    ImageGenerationStats,
    ToolCallItem,
-    Usage,
 )
 from exo.utils.pydantic_ext import TaggedModel

@@ -25,7 +24,6 @@ class GenerationResponse(BaseRunnerResponse):
    # logprobs: list[float] | None = None # too big. we can change to be top-k
    finish_reason: FinishReason | None = None
    stats: GenerationStats | None = None
-    usage: Usage | None


 class ImageGenerationResponse(BaseRunnerResponse):
@@ -59,7 +57,6 @@ class PartialImageResponse(BaseRunnerResponse):

 class ToolCallResponse(BaseRunnerResponse):
    tool_calls: list[ToolCallItem]
-    usage: Usage | None


 class FinishedResponse(BaseRunnerResponse):
--- a/src/exo/worker/engines/mlx/generator/generate.py
+++ b/src/exo/worker/engines/mlx/generator/generate.py
@@ -10,11 +10,8 @@ from mlx_lm.tokenizer_utils import TokenizerWrapper
 from exo.shared.types.api import (
    BenchChatCompletionTaskParams,
    ChatCompletionMessage,
-    CompletionTokensDetails,
    FinishReason,
    GenerationStats,
-    PromptTokensDetails,
-    Usage,
 )
 from exo.shared.types.memory import Memory
 from exo.shared.types.mlx import KVCacheType
@@ -219,37 +216,22 @@ def mlx_generate(
    max_tokens = task.max_tokens or MAX_TOKENS
    generated_text_parts: list[str] = []
    generation_start_time = time.perf_counter()
-    usage: Usage | None = None
-    in_thinking = False
-    reasoning_tokens = 0
-    think_start = tokenizer.think_start
-    think_end = tokenizer.think_end
-    for completion_tokens, out in enumerate(
-        stream_generate(
-            model=model,
-            tokenizer=tokenizer,
-            prompt=last_token,
-            max_tokens=max_tokens,
-            sampler=sampler,
-            logits_processors=logits_processors,
-            prompt_cache=caches,
-            # TODO: Dynamically change prefill step size to be the maximum possible without timing out.
-            prefill_step_size=2048,
-            kv_group_size=KV_GROUP_SIZE,
-            kv_bits=KV_BITS,
-        ),
-        start=1,
+    for out in stream_generate(
+        model=model,
+        tokenizer=tokenizer,
+        prompt=last_token,
+        max_tokens=max_tokens,
+        sampler=sampler,
+        logits_processors=logits_processors,
+        prompt_cache=caches,
+        # TODO: Dynamically change prefill step size to be the maximum possible without timing out.
+        prefill_step_size=2048,
+        kv_group_size=KV_GROUP_SIZE,
+        kv_bits=KV_BITS,
    ):
        generated_text_parts.append(out.text)
        logger.info(out.text)

-        if think_start is not None and out.text == think_start:
-            in_thinking = True
-        elif think_end is not None and out.text == think_end:
-            in_thinking = False
-        if in_thinking:
-            reasoning_tokens += 1
-
        stats: GenerationStats | None = None
        if out.finish_reason is not None:
            stats = GenerationStats(
@@ -267,24 +249,11 @@ def mlx_generate(
                    f"Model generated unexpected finish_reason: {out.finish_reason}"
                )

-            usage = Usage(
-                prompt_tokens=int(out.prompt_tokens),
-                completion_tokens=completion_tokens,
-                total_tokens=int(out.prompt_tokens) + completion_tokens,
-                prompt_tokens_details=PromptTokensDetails(
-                    cached_tokens=prefix_hit_length
-                ),
-                completion_tokens_details=CompletionTokensDetails(
-                    reasoning_tokens=reasoning_tokens
-                ),
-            )
-
        yield GenerationResponse(
            text=out.text,
            token=out.token,
            finish_reason=cast(FinishReason | None, out.finish_reason),
            stats=stats,
-            usage=usage,
        )

        if out.finish_reason is not None:
--- a/src/exo/worker/runner/runner.py
+++ b/src/exo/worker/runner/runner.py
@@ -277,11 +277,9 @@ def main(
                                tokenizer.tool_parser,  # pyright: ignore[reportAny]
                            )

-                        completion_tokens = 0
                        for response in mlx_generator:
                            match response:
                                case GenerationResponse():
-                                    completion_tokens += 1
                                    if (
                                        device_rank == 0
                                        and response.finish_reason == "error"
@@ -309,7 +307,6 @@ def main(
                                                    model=shard_metadata.model_card.model_id,
                                                    text=response.text,
                                                    token_id=response.token,
-                                                    usage=response.usage,
                                                    finish_reason=response.finish_reason,
                                                    stats=response.stats,
                                                ),
@@ -323,7 +320,6 @@ def main(
                                                chunk=ToolCallChunk(
                                                    tool_calls=response.tool_calls,
                                                    model=shard_metadata.model_card.model_id,
-                                                    usage=response.usage,
                                                ),
                                            )
                                        )
@@ -539,8 +535,7 @@ def parse_gpt_oss(
                            name=current_tool_name,
                            arguments="".join(tool_arg_parts).strip(),
                        )
-                    ],
-                    usage=response.usage,
+                    ]
                )
                tool_arg_parts = []
            current_tool_name = recipient
@@ -688,7 +683,7 @@ def parse_tool_calls(
                    tools = [_validate_single_tool(tool) for tool in parsed]
                else:
                    tools = [_validate_single_tool(parsed)]
-                yield ToolCallResponse(tool_calls=tools, usage=response.usage)
+                yield ToolCallResponse(tool_calls=tools)

            except (
                json.JSONDecodeError,
--- a/src/exo/worker/tests/unittests/test_runner/test_event_ordering.py
+++ b/src/exo/worker/tests/unittests/test_runner/test_event_ordering.py
@@ -120,7 +120,7 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
    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)
+        yield GenerationResponse(token=0, text="hi", finish_reason="stop")

    monkeypatch.setattr(mlx_runner, "mlx_generate", fake_generate)

@@ -182,8 +182,6 @@ def test_events_processed_in_correct_order(patch_out_mlx: pytest.MonkeyPatch):
            text="hi",
            token_id=0,
            finish_reason="stop",
-            usage=None,
-            stats=None,
        ),
    )

--- a/tests/headless_runner.py
+++ b/tests/headless_runner.py
@@ -1,26 +1,21 @@
-import multiprocessing as mp
 import socket
-import time
-import typing
+from typing import Literal

 import anyio
+from loguru import logger
 from fastapi import FastAPI
-from fastapi.responses import StreamingResponse
+from fastapi.responses import StreamingResponse, Response
 from hypercorn import Config
 from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType]
-from loguru import logger
 from pydantic import BaseModel

-from exo.download.impl_shard_downloader import (
-    build_full_shard,
-    exo_shard_downloader,
-)
-from exo.shared.logging import InterceptLogger, logger_setup
+from exo.shared.constants import EXO_MODELS_DIR
 from exo.shared.models.model_cards import MODEL_CARDS, ModelId
 from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
+from exo.shared.types.chunks import TokenChunk
 from exo.shared.types.commands import CommandId
 from exo.shared.types.common import Host, NodeId
-from exo.shared.types.events import Event
+from exo.shared.types.events import ChunkGenerated, Event, RunnerStatusUpdated
 from exo.shared.types.tasks import (
    ChatCompletion,
    ConnectToGroup,
@@ -36,9 +31,14 @@ from exo.shared.types.worker.instances import (
    MlxJacclInstance,
    MlxRingInstance,
 )
-from exo.shared.types.worker.runners import RunnerId, ShardAssignments
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerShutdown,
+    ShardAssignments,
+)
 from exo.shared.types.worker.shards import PipelineShardMetadata, TensorShardMetadata
-from exo.utils.channels import MpReceiver, MpSender, channel, mp_channel
+from exo.utils.channels import channel, mp_channel
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo, InfoGatherer
 from exo.worker.runner.bootstrap import entrypoint

@@ -46,37 +46,36 @@ from exo.worker.runner.bootstrap import entrypoint
 class Tests(BaseModel):
    # list[hostname, ip addr]
    devs: list[list[str]]
-    model_id: str
-    kind: typing.Literal["init", "warmup", "inference"]
+    rdma_devs: list[list[str | None]] | None
+    model_id: ModelId
+    kind: Literal["ring", "rdma", "both"]


-mp.set_start_method("spawn", force=True)
-logger_setup(None)
+iid = InstanceId("im testing here")


 async def main():
    logger.info("starting cool server majig")
-    await assert_downloads()
    cfg = Config()
-    cfg.bind = "0.0.0.0:52415"
+    cfg.bind = "0.0.0.0:52414"
    # nb: shared.logging needs updating if any of this changes
    cfg.accesslog = "-"
    cfg.errorlog = "-"
-    cfg.logger_class = InterceptLogger
+    ev = anyio.Event()
    app = FastAPI()
-    app.post("/ring")(ring_backend)
-    app.post("/jaccl")(jaccl_backend)
-    app.post("/tb_detection")(tb_detection)
-    shutdown = anyio.Event()
+    app.post("/run_test")(run_test)
+    app.post("/kill")(lambda: kill(ev))
+    app.get("/tb_detection")(tb_detection)
+    app.get("/models")(list_models)
    await serve(
        app,  # type: ignore
        cfg,
-        shutdown_trigger=lambda: shutdown.wait(),
+        shutdown_trigger = lambda: ev.wait()
    )
-    await anyio.sleep_forever()
-    # gracefully shutdown the api
-    shutdown.set()

+def kill(ev: anyio.Event):
+    ev.set()
+    return Response(status_code=204)

 async def tb_detection():
    send, recv = channel[GatheredInfo]()
@@ -87,29 +86,19 @@ async def tb_detection():
        return recv.collect()


-async def assert_downloads():
-    sd = exo_shard_downloader()
-    # await sd.ensure_shard(await build_full_shard(MODEL_CARDS["qwen3-0.6b"].model_id))
-    await sd.ensure_shard(
-        await build_full_shard(MODEL_CARDS["llama-3.1-8b-bf16"].model_id)
-    )
-    await sd.ensure_shard(await build_full_shard(MODEL_CARDS["qwen3-30b"].model_id))
-    await sd.ensure_shard(
-        await build_full_shard(MODEL_CARDS["gpt-oss-120b-MXFP4-Q8"].model_id)
-    )
-    await sd.ensure_shard(
-        await build_full_shard(MODEL_CARDS["gpt-oss-20b-4bit"].model_id)
-    )
-    await sd.ensure_shard(
-        await build_full_shard(MODEL_CARDS["glm-4.7-8bit-gs32"].model_id)
-    )
-    await sd.ensure_shard(
-        await build_full_shard(MODEL_CARDS["minimax-m2.1-8bit"].model_id)
-    )
+def list_models():
+    sent = set[str]()
+    for path in EXO_MODELS_DIR.rglob("model-*.safetensors"):
+        if "--" not in path.parent.name:
+            continue
+        name = path.parent.name.replace("--", "/")
+        if name in sent:
+            continue
+        sent.add(name)
+        yield ModelId(path.parent.name.replace("--", "/"))


-async def ring_backend(test: Tests):
-    iid = InstanceId(str(hash(str(test.devs))))
+async def run_test(test: Tests):
    weird_hn = socket.gethostname()
    for dev in test.devs:
        if weird_hn.startswith(dev[0]) or dev[0].startswith(weird_hn):
@@ -117,31 +106,67 @@ async def ring_backend(test: Tests):
            break
    else:
        raise ValueError(f"{weird_hn} not in {test.devs}")
-    return await execute_test(test, ring_instance(test, iid, hn), hn)
+
+    async def run():
+        logger.info(f"testing {test.model_id}")
+
+        instances: list[Instance] = []
+        if test.kind in ["ring", "both"]:
+            i = ring_instance(test, hn)
+            if i is None:
+                yield "no model found"
+                return
+            instances.append(i)
+        if test.kind in ["rdma", "both"]:
+            i = jaccl_instance(test)
+            if i is None:
+                yield "no model found"
+                return
+            instances.append(i)
+
+        for instance in instances:
+            recv = await execute_test(test, instance, hn)
+
+            str_out = ""
+
+            for item in recv:
+                if isinstance(item, ChunkGenerated):
+                    assert isinstance(item.chunk, TokenChunk)
+                    str_out += item.chunk.text
+
+                if isinstance(item, RunnerStatusUpdated) and isinstance(
+                    item.runner_status, (RunnerFailed, RunnerShutdown)
+                ):
+                    yield str_out + "\n"
+                    yield item.model_dump_json() + "\n"
+
+    return StreamingResponse(run())


-def ring_instance(test: Tests, iid: InstanceId, hn: str) -> Instance:
-    hbn = [Host(ip="i dont care", port=52416) for _ in test.devs]
+def ring_instance(test: Tests, hn: str) -> Instance | None:
+    hbn = [Host(ip="198.51.100.0", port=52417) for _ in test.devs]
    world_size = len(test.devs)
    for i in range(world_size):
        if test.devs[i][0] == hn:
            hn = test.devs[i][0]
-            if i - 1 >= 0:
-                hbn[i - 1] = Host(ip=test.devs[i - 1][1], port=52416)
-            if i + 1 < len(test.devs):
-                hbn[i + 1] = Host(ip=test.devs[i + 1][1], port=52416)
-            hbn[i] = Host(ip="0.0.0.0", port=52416)
-            break
+        hbn[(i - 1) % world_size] = Host(ip=test.devs[i - 1][1], port=52417)
+        hbn[(i + 1) % world_size] = Host(ip=test.devs[i + 1][1], port=52417)
+        hbn[i] = Host(ip="0.0.0.0", port=52417)
+        break
    else:
        raise ValueError(f"{hn} not in {test.devs}")

-    card = MODEL_CARDS[test.model_id]
+    card = next(
+        (card for card in MODEL_CARDS.values() if card.model_id == test.model_id), None
+    )
+    if card is None:
+        return None
    instance = MlxRingInstance(
        instance_id=iid,
-        ephemeral_port=52416,
+        ephemeral_port=52417,
        hosts_by_node={NodeId(hn): hbn},
        shard_assignments=ShardAssignments(
-            model_id=ModelId(test.model_id),
+            model_id=test.model_id,
            node_to_runner={NodeId(host[0]): RunnerId(host[0]) for host in test.devs},
            runner_to_shard={
                RunnerId(test.devs[i][0]): PipelineShardMetadata(
@@ -163,119 +188,86 @@ def ring_instance(test: Tests, iid: InstanceId, hn: str) -> Instance:
    return instance


-async def execute_test(test: Tests, instance: Instance, hn: str):
+async def execute_test(test: Tests, instance: Instance, hn: str) -> list[Event]:
    world_size = len(test.devs)
-    iid = InstanceId(str(hash(str(test.devs))))
-    _handle, recv, send = new_runner(instance, hn)
-    if world_size > 1:
-        send.send(ConnectToGroup(instance_id=iid))
-    send.send(LoadModel(instance_id=iid))
-
-    match test.kind:
-        case "init":
-            pass
-        case "warmup":
-            send.send(StartWarmup(instance_id=iid))
-        case "inference":
-            send.send(StartWarmup(instance_id=iid))
-            send.send(
-                ChatCompletion(
-                    task_params=ChatCompletionTaskParams(
-                        model=test.model_id,
-                        messages=[
-                            ChatCompletionMessage(
-                                role="system", content="You are a helpful assistant"
-                            ),
-                            ChatCompletionMessage(
-                                role="user", content="What is the capital of France?"
-                            ),
-                        ],
-                    ),
-                    command_id=CommandId("yo"),
-                    instance_id=iid,
-                )
+    commands: list[Task] = [
+        (LoadModel(instance_id=iid)),
+        (StartWarmup(instance_id=iid)),
+        (
+            ChatCompletion(
+                task_params=ChatCompletionTaskParams(
+                    model=test.model_id,
+                    messages=[
+                        ChatCompletionMessage(
+                            role="system", content="You are a helpful assistant"
+                        ),
+                        ChatCompletionMessage(
+                            role="user", content="What is the capital of France?"
+                        ),
+                    ],
+                    max_tokens=50,
+                ),
+                command_id=CommandId("yo"),
+                instance_id=iid,
            )
+        ),
+        (Shutdown(runner_id=RunnerId(hn), instance_id=iid)),
+    ]
+    if world_size > 1:
+        commands.insert(0, ConnectToGroup(instance_id=iid))
+    bound_instance = BoundInstance(
+        instance=instance, bound_runner_id=RunnerId(hn), bound_node_id=NodeId(hn)
+    )
+    ev_send, _ev_recv = mp_channel[Event]()
+    task_send, task_recv = mp_channel[Task]()

-    send.send(Shutdown(runner_id=RunnerId(hn), instance_id=iid))
+    for command in commands:
+        task_send.send(command)

-    async def map_recv():
-        with recv:
-            try:
-                async for item in recv:
-                    yield item.model_dump_json() + "\n"
-            except anyio.ClosedResourceError:
-                pass
+    entrypoint(
+        bound_instance,
+        ev_send,
+        task_recv,
+        logger,  # type: ignore
+    )

-    ret = StreamingResponse(map_recv())
-    ret._pls_dont_gc = _handle  # type: ignore
-    return ret
+    # TODO(evan): return ev_recv.collect()
+    return []


-async def jaccl_backend(test: Tests):
-    iid = InstanceId(str(hash(str(test.devs))))
-    weird_hn = socket.gethostname()
-    for dev in test.devs:
-        if weird_hn.startswith(dev[0]) or dev[0].startswith(weird_hn):
-            hn = dev[0]
-            break
-    else:
-        raise ValueError(f"{weird_hn} not in {test.devs}")
-    return await execute_test(test, jaccl_instance(test, iid), hn)
-
-
-def jaccl_instance(test: Tests, iid: InstanceId):
-    card = MODEL_CARDS[test.model_id]
+def jaccl_instance(test: Tests) -> MlxJacclInstance | None:
+    card = next(
+        (card for card in MODEL_CARDS.values() if card.model_id == test.model_id), None
+    )
+    if card is None:
+        return None
    world_size = len(test.devs)
+    assert test.rdma_devs

    return MlxJacclInstance(
        instance_id=iid,
-        jaccl_devices=[[None, "rdma_en3"], ["rdma_en3", None]],
+        jaccl_devices=test.rdma_devs,
        # rank 0 is always coordinator
        jaccl_coordinators={
-            NodeId(host[0]): test.devs[0][1] + ":52416" for host in test.devs
+            NodeId(host[0]): test.devs[0][1] + ":52417" for host in test.devs
        },
        shard_assignments=ShardAssignments(
-            model_id=ModelId(test.model_id),
+            model_id=test.model_id,
            node_to_runner={NodeId(host[0]): RunnerId(host[0]) for host in test.devs},
            runner_to_shard={
-                RunnerId(test.devs[i][0]): TensorShardMetadata(
+                RunnerId(host[0]): TensorShardMetadata(
                    model_card=card,
                    device_rank=i,
                    world_size=world_size,
-                    start_layer=card.n_layers,
+                    start_layer=0,
                    end_layer=card.n_layers,
                    n_layers=card.n_layers,
                )
-                for i in range(world_size)
+                for i, host in enumerate(test.devs)
            },
        ),
    )


-def new_runner(
-    instance: Instance,
-    hn: str,
-) -> tuple[mp.Process, MpReceiver[Event], MpSender[Task]]:
-    bound_instance = BoundInstance(
-        instance=instance, bound_runner_id=RunnerId(hn), bound_node_id=NodeId(hn)
-    )
-    ev_send, ev_recv = mp_channel[Event]()
-    task_send, task_recv = mp_channel[Task]()
-
-    runner_process = mp.Process(
-        target=entrypoint,
-        args=(
-            bound_instance,
-            ev_send,
-            task_recv,
-            logger,
-        ),
-    )
-    runner_process._pls_dont_gc = (ev_send, task_recv)  # type: ignore
-    runner_process.start()
-    time.sleep(0.1)
-    return (runner_process, ev_recv, task_send)
-
-
 if __name__ == "__main__":
    anyio.run(main)
--- a/tests/run_distributed_test.sh
+++ b/tests/run_distributed_test.sh
@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+[ $# -eq 0 ] && {
+  echo "Usage: $0 host1 [host2 ...]"
+  exit 1
+}
+
+[ -z "$(git status --porcelain)" ] || {
+  echo "Uncommitted changes"
+  exit 1
+}
+commit=$(git rev-parse HEAD)
+git fetch -q origin
+git branch -r --contains "$commit" | grep -qE '^\s*origin/' || {
+  echo "Not pushed to origin"
+  exit 1
+}
+for host; do
+  curl -m 1 -X POST "http://$host:52414/kill" >/dev/null 2>&1 || true &
+done
+wait
+
+echo "Deploying $commit to $# hosts..."
+
+pids=""
+trap 'xargs -r kill 2>/dev/null <<<"$pids" || true' EXIT INT TERM
+colours=($'\e[31m' $'\e[32m' $'\e[33m' $'\e[34m')
+reset=$'\e[0m'
+i=0
+
+for host; do
+  colour=${colours[i++ % 4]}
+  ssh -tt -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" "/usr/bin/env bash -lc '
+    set -euo pipefail
+    cd exo 
+    git fetch -q origin
+    git checkout -q $commit
+    nix develop -c uv sync
+    .venv/bin/python tests/headless_runner.py
+    '" 2>&1 | sed -u "s/^/${colour}[${host}]${reset}/" &
+  pids+=" $!"
+done
+
+for host; do
+  echo "Waiting for $host..."
+  until curl -sf "http://$host:52414/models"; do sleep 1; done
+done
+
+uv run tests/start_distributed_test.py "$@"
--- a/tests/start_distributed_test.py
+++ b/tests/start_distributed_test.py
@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+import itertools
+import json
+import subprocess
+import sys
+from concurrent.futures import ThreadPoolExecutor
+from typing import Any, cast
+from urllib.request import Request, urlopen
+
+if not (args := sys.argv[1:]):
+    sys.exit(
+        f"USAGE: {sys.argv[0]} <kind> [host1] [host2] ...\nkind is optional, and should be rdma or ring"
+    )
+
+kind = args[0] if args[0] in ("rdma", "ring") else "both"
+hosts = args[1:] if kind != "both" else args
+ts = subprocess.run(
+    ["tailscale", "status"], check=True, text=True, capture_output=True
+).stdout.splitlines()
+ip = {sl[1]: sl[0] for line in ts if len(sl := line.split()) >= 2}
+ips = [ip[h] for h in hosts]
+devs = [[h, ip[h]] for h in hosts]
+n = len(hosts)
+
+
+def get_tb(a: str) -> list[dict[str, Any]]:
+    with urlopen(f"http://{a}:52414/tb_detection", timeout=5) as r:  # pyright: ignore[reportAny]
+        return json.loads(r.read())  # pyright: ignore[reportAny]
+
+
+def get_models(a: str) -> set[str]:
+    with urlopen(f"http://{a}:52414/models", timeout=5) as r:  # pyright: ignore[reportAny]
+        return set(json.loads(r.read()))  # pyright: ignore[reportAny]
+
+
+def run(h: str, a: str, body: bytes) -> None:
+    with urlopen(
+        Request(
+            f"http://{a}:52414/run_test",
+            data=body,
+            method="POST",
+            headers={"Content-Type": "application/json"},
+        ),
+        timeout=300,
+    ) as r:  # pyright: ignore[reportAny]
+        for line in r.read().decode(errors="replace").splitlines():  # pyright: ignore[reportAny]
+            print(f"\n{h}@{a}: {line}", flush=True)
+
+
+with ThreadPoolExecutor(n) as exctr:
+    if kind in ("rdma", "both"):
+        payloads = list(exctr.map(get_tb, ips))
+
+        u2e = {
+            ident["domainUuid"]: (i, ident["rdmaInterface"])
+            for i, p in enumerate(payloads)
+            for d in p
+            for ident in cast(
+                list[dict[str, str]],
+                d.get("MacThunderboltIdentifiers", {}).get("idents", []),  # pyright: ignore[reportAny]
+            )
+        }
+        edges = {
+            (u2e[s][0], u2e[t][0]): u2e[t][1]
+            for p in payloads
+            for d in p
+            for c in d.get("MacThunderboltConnections", {}).get("conns", [])  # pyright: ignore[reportAny]
+            if (s := c["sourceUuid"]) in u2e and (t := c["sinkUuid"]) in u2e  # pyright: ignore[reportAny]
+        }
+        rdma_devs = [[edges.get((i, j)) for j in range(n)] for i in range(n)]
+    else:
+        rdma_devs = None
+
+    models = set[str].intersection(*exctr.map(get_models, ips))
+
+    print("\n")
+    print("=" * 70)
+    print(f"Starting test with {models}")
+    print("=" * 70)
+    print("\n")
+    for model in models:
+        body = json.dumps(
+            {"devs": devs, "model_id": model, "rdma_devs": rdma_devs, "kind": kind}
+        ).encode()
+        list(exctr.map(run, hosts, ips, itertools.repeat(body)))
--- a/tests/start_distributed_test.sh
+++ b/tests/start_distributed_test.sh
@@ -1,54 +0,0 @@
-#!/usr/bin/env bash
-
-set -euo pipefail
-
-query() {
-  tailscale status | awk -v find="$1" '$2 == find { print $1 }'
-}
-
-if [[ $# -lt 2 ]]; then
-  echo "USAGE: $0 <test kind> [host1] [host2] ..."
-  exit 1
-fi
-
-kind=$1
-shift
-
-test_kinds="ring jaccl"
-
-if ! echo "$test_kinds" | grep -q "$kind"; then
-  printf "%s is not a known test kind.\nCurrent test kinds are %s" "$kind" "$test_kinds"
-  exit 1
-fi
-
-hostnames=("$@")
-weaved=()
-ips=()
-for name in "${hostnames[@]}"; do
-  ip=$(query "$name")
-  ips+=("$ip")
-  weaved+=("$name" "$ip")
-done
-
-devs_raw=$(printf '["%s", "%s"], ' "${weaved[@]}")
-devs="[${devs_raw%, }]"
-
-model_ids=("qwen3-30b" "gpt-oss-120b-MXFP4-Q8" "kimi-k2-thinking")
-
-for model_id in "${model_ids[@]}"; do
-  for i in "${!ips[@]}"; do
-    {
-      req="{
-        \"model_id\": \"${model_id}\",
-        \"devs\": ${devs},
-        \"kind\": \"inference\"
-       }"
-      echo "req $req"
-      curl -sN \
-        -X POST "http://${ips[$i]}:52415/${kind}" \
-        -H "Content-Type: application/json" -d "$req" \
-        2>&1 | sed "s/^/\n${hostnames[$i]}@${ips[$i]}: /" || echo "curl to ${hostnames[$i]} failed" && exit 1
-    } &
-  done
-  wait
-done
--- a/tmp/run_exo_on.sh
+++ b/tmp/run_exo_on.sh
@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+[ $# -eq 0 ] && {
+  echo "Usage: $0 host1 [host2 ...]"
+  exit 1
+}
+
+[ -z "$(git status --porcelain)" ] || {
+  echo "Uncommitted changes"
+  exit 1
+}
+commit=$(git rev-parse HEAD)
+git fetch -q origin
+git branch -r --contains "$commit" | grep -qE '^\s*origin/' || {
+  echo "Not pushed to origin"
+  exit 1
+}
+echo "Deploying $commit to $# hosts..."
+
+pids=""
+trap 'xargs -r kill 2>/dev/null <<<"$pids" || true' EXIT INT TERM
+colours=($'\e[31m' $'\e[32m' $'\e[33m' $'\e[34m')
+reset=$'\e[0m'
+i=0
+
+for host; do
+  colour=${colours[i++ % 4]}
+  ssh -t -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" "/usr/bin/env bash -lc '
+    set -euo pipefail
+    cd exo 
+    git fetch -q origin
+    git checkout -q $commit
+    nix develop -c uv sync
+    uv run exo
+    '" 2>&1 | sed -u "s/^/${colour}[${host}]${reset}/" &
+  pids+=" $!"
+done
+wait
--- a/uv.lock
+++ b/uv.lock
@@ -192,14 +192,20 @@ sdist = { url = "https://files.pythonhosted.org/packages/eb/56/b1ba7935a17738ae8
 wheels = [
    { url = "https://files.pythonhosted.org/packages/b0/1e/d22cc63332bd59b06481ceaac49d6c507598642e2230f201649058a7e704/cffi-2.0.0-cp313-cp313-manylinux1_i686.manylinux2014_i686.manylinux_2_17_i686.manylinux_2_5_i686.whl", hash = "sha256:07b271772c100085dd28b74fa0cd81c8fb1a3ba18b21e03d7c27f3436a10606b", size = 212446 },
    { url = "https://files.pythonhosted.org/packages/a9/f5/a2c23eb03b61a0b8747f211eb716446c826ad66818ddc7810cc2cc19b3f2/cffi-2.0.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d48a880098c96020b02d5a1f7d9251308510ce8858940e6fa99ece33f610838b", size = 220101 },
+    { url = "https://files.pythonhosted.org/packages/f2/7f/e6647792fc5850d634695bc0e6ab4111ae88e89981d35ac269956605feba/cffi-2.0.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:f93fd8e5c8c0a4aa1f424d6173f14a892044054871c771f8566e4008eaa359d2", size = 207948 },
+    { url = "https://files.pythonhosted.org/packages/cb/1e/a5a1bd6f1fb30f22573f76533de12a00bf274abcdc55c8edab639078abb6/cffi-2.0.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:dd4f05f54a52fb558f1ba9f528228066954fee3ebe629fc1660d874d040ae5a3", size = 206422 },
    { url = "https://files.pythonhosted.org/packages/98/df/0a1755e750013a2081e863e7cd37e0cdd02664372c754e5560099eb7aa44/cffi-2.0.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:c8d3b5532fc71b7a77c09192b4a5a200ea992702734a2e9279a37f2478236f26", size = 219499 },
    { url = "https://files.pythonhosted.org/packages/50/e1/a969e687fcf9ea58e6e2a928ad5e2dd88cc12f6f0ab477e9971f2309b57c/cffi-2.0.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:d9b29c1f0ae438d5ee9acb31cadee00a58c46cc9c0b2f9038c6b0b3470877a8c", size = 222928 },
    { url = "https://files.pythonhosted.org/packages/36/54/0362578dd2c9e557a28ac77698ed67323ed5b9775ca9d3fe73fe191bb5d8/cffi-2.0.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:6d50360be4546678fc1b79ffe7a66265e28667840010348dd69a314145807a1b", size = 221302 },
    { url = "https://files.pythonhosted.org/packages/d6/43/0e822876f87ea8a4ef95442c3d766a06a51fc5298823f884ef87aaad168c/cffi-2.0.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:24b6f81f1983e6df8db3adc38562c83f7d4a0c36162885ec7f7b77c7dcbec97b", size = 220049 },
+    { url = "https://files.pythonhosted.org/packages/b4/89/76799151d9c2d2d1ead63c2429da9ea9d7aac304603de0c6e8764e6e8e70/cffi-2.0.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:12873ca6cb9b0f0d3a0da705d6086fe911591737a59f28b7936bdfed27c0d47c", size = 207793 },
+    { url = "https://files.pythonhosted.org/packages/bb/dd/3465b14bb9e24ee24cb88c9e3730f6de63111fffe513492bf8c808a3547e/cffi-2.0.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:d9b97165e8aed9272a6bb17c01e3cc5871a594a446ebedc996e2397a1c1ea8ef", size = 206300 },
    { url = "https://files.pythonhosted.org/packages/47/d9/d83e293854571c877a92da46fdec39158f8d7e68da75bf73581225d28e90/cffi-2.0.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:afb8db5439b81cf9c9d0c80404b60c3cc9c3add93e114dcae767f1477cb53775", size = 219244 },
    { url = "https://files.pythonhosted.org/packages/2b/0f/1f177e3683aead2bb00f7679a16451d302c436b5cbf2505f0ea8146ef59e/cffi-2.0.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:737fe7d37e1a1bffe70bd5754ea763a62a066dc5913ca57e957824b72a85e205", size = 222828 },
    { url = "https://files.pythonhosted.org/packages/c6/0f/cafacebd4b040e3119dcb32fed8bdef8dfe94da653155f9d0b9dc660166e/cffi-2.0.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:38100abb9d1b1435bc4cc340bb4489635dc2f0da7456590877030c9b3d40b0c1", size = 220926 },
    { url = "https://files.pythonhosted.org/packages/be/b4/c56878d0d1755cf9caa54ba71e5d049479c52f9e4afc230f06822162ab2f/cffi-2.0.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7cc09976e8b56f8cebd752f7113ad07752461f48a58cbba644139015ac24954c", size = 221593 },
+    { url = "https://files.pythonhosted.org/packages/e0/0d/eb704606dfe8033e7128df5e90fee946bbcb64a04fcdaa97321309004000/cffi-2.0.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:92b68146a71df78564e4ef48af17551a5ddd142e5190cdf2c5624d0c3ff5b2e8", size = 209354 },
+    { url = "https://files.pythonhosted.org/packages/d8/19/3c435d727b368ca475fb8742ab97c9cb13a0de600ce86f62eab7fa3eea60/cffi-2.0.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:b1e74d11748e7e98e2f426ab176d4ed720a64412b6a15054378afdb71e0f37dc", size = 208480 },
    { url = "https://files.pythonhosted.org/packages/d0/44/681604464ed9541673e486521497406fadcc15b5217c3e326b061696899a/cffi-2.0.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:28a3a209b96630bca57cce802da70c266eb08c6e97e5afd61a75611ee6c64592", size = 221584 },
    { url = "https://files.pythonhosted.org/packages/25/8e/342a504ff018a2825d395d44d63a767dd8ebc927ebda557fecdaca3ac33a/cffi-2.0.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:7553fb2090d71822f02c629afe6042c299edf91ba1bf94951165613553984512", size = 224443 },
    { url = "https://files.pythonhosted.org/packages/e1/5e/b666bacbbc60fbf415ba9988324a132c9a7a0448a9a8f125074671c0f2c3/cffi-2.0.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:6c6c373cfc5c83a975506110d17457138c8c63016b563cc9ed6e056a82f13ce4", size = 223437 },
@@ -305,8 +311,10 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/5c/49/498c86566a1d80e978b42f0d702795f69887005548c041636df6ae1ca64c/cryptography-46.0.3-cp311-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:01ca9ff2885f3acc98c29f1860552e37f6d7c7d013d7334ff2a9de43a449315d", size = 4450807 },
    { url = "https://files.pythonhosted.org/packages/4b/0a/863a3604112174c8624a2ac3c038662d9e59970c7f926acdcfaed8d61142/cryptography-46.0.3-cp311-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:6eae65d4c3d33da080cff9c4ab1f711b15c1d9760809dad6ea763f3812d254cb", size = 4299615 },
    { url = "https://files.pythonhosted.org/packages/64/02/b73a533f6b64a69f3cd3872acb6ebc12aef924d8d103133bb3ea750dc703/cryptography-46.0.3-cp311-abi3-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e5bf0ed4490068a2e72ac03d786693adeb909981cc596425d09032d372bcc849", size = 4016800 },
+    { url = "https://files.pythonhosted.org/packages/25/d5/16e41afbfa450cde85a3b7ec599bebefaef16b5c6ba4ec49a3532336ed72/cryptography-46.0.3-cp311-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:5ecfccd2329e37e9b7112a888e76d9feca2347f12f37918facbb893d7bb88ee8", size = 4984707 },
    { url = "https://files.pythonhosted.org/packages/c9/56/e7e69b427c3878352c2fb9b450bd0e19ed552753491d39d7d0a2f5226d41/cryptography-46.0.3-cp311-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:a2c0cd47381a3229c403062f764160d57d4d175e022c1df84e168c6251a22eec", size = 4482541 },
    { url = "https://files.pythonhosted.org/packages/78/f6/50736d40d97e8483172f1bb6e698895b92a223dba513b0ca6f06b2365339/cryptography-46.0.3-cp311-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:549e234ff32571b1f4076ac269fcce7a808d3bf98b76c8dd560e42dbc66d7d91", size = 4299464 },
+    { url = "https://files.pythonhosted.org/packages/00/de/d8e26b1a855f19d9994a19c702fa2e93b0456beccbcfe437eda00e0701f2/cryptography-46.0.3-cp311-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:c0a7bb1a68a5d3471880e264621346c48665b3bf1c3759d682fc0864c540bd9e", size = 4950838 },
    { url = "https://files.pythonhosted.org/packages/8f/29/798fc4ec461a1c9e9f735f2fc58741b0daae30688f41b2497dcbc9ed1355/cryptography-46.0.3-cp311-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:10b01676fc208c3e6feeb25a8b83d81767e8059e1fe86e1dc62d10a3018fa926", size = 4481596 },
    { url = "https://files.pythonhosted.org/packages/15/8d/03cd48b20a573adfff7652b76271078e3045b9f49387920e7f1f631d125e/cryptography-46.0.3-cp311-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:0abf1ffd6e57c67e92af68330d05760b7b7efb243aab8377e583284dbab72c71", size = 4426782 },
    { url = "https://files.pythonhosted.org/packages/fa/b1/ebacbfe53317d55cf33165bda24c86523497a6881f339f9aae5c2e13e57b/cryptography-46.0.3-cp311-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:a04bee9ab6a4da801eb9b51f1b708a1b5b5c9eb48c03f74198464c66f0d344ac", size = 4698381 },
@@ -314,8 +322,10 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/c5/fd/bc1daf8230eaa075184cbbf5f8cd00ba9db4fd32d63fb83da4671b72ed8a/cryptography-46.0.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:39b6755623145ad5eff1dab323f4eae2a32a77a7abef2c5089a04a3d04366715", size = 4435078 },
    { url = "https://files.pythonhosted.org/packages/82/98/d3bd5407ce4c60017f8ff9e63ffee4200ab3e23fe05b765cab805a7db008/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:db391fa7c66df6762ee3f00c95a89e6d428f4d60e7abc8328f4fe155b5ac6e54", size = 4293460 },
    { url = "https://files.pythonhosted.org/packages/26/e9/e23e7900983c2b8af7a08098db406cf989d7f09caea7897e347598d4cd5b/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:78a97cf6a8839a48c49271cdcbd5cf37ca2c1d6b7fdd86cc864f302b5e9bf459", size = 3995237 },
+    { url = "https://files.pythonhosted.org/packages/91/15/af68c509d4a138cfe299d0d7ddb14afba15233223ebd933b4bbdbc7155d3/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_ppc64le.whl", hash = "sha256:dfb781ff7eaa91a6f7fd41776ec37c5853c795d3b358d4896fdbb5df168af422", size = 4967344 },
    { url = "https://files.pythonhosted.org/packages/ca/e3/8643d077c53868b681af077edf6b3cb58288b5423610f21c62aadcbe99f4/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:6f61efb26e76c45c4a227835ddeae96d83624fb0d29eb5df5b96e14ed1a0afb7", size = 4466564 },
    { url = "https://files.pythonhosted.org/packages/0e/43/c1e8726fa59c236ff477ff2b5dc071e54b21e5a1e51aa2cee1676f1c986f/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_aarch64.whl", hash = "sha256:23b1a8f26e43f47ceb6d6a43115f33a5a37d57df4ea0ca295b780ae8546e8044", size = 4292415 },
+    { url = "https://files.pythonhosted.org/packages/42/f9/2f8fefdb1aee8a8e3256a0568cffc4e6d517b256a2fe97a029b3f1b9fe7e/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_ppc64le.whl", hash = "sha256:b419ae593c86b87014b9be7396b385491ad7f320bde96826d0dd174459e54665", size = 4931457 },
    { url = "https://files.pythonhosted.org/packages/79/30/9b54127a9a778ccd6d27c3da7563e9f2d341826075ceab89ae3b41bf5be2/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_x86_64.whl", hash = "sha256:50fc3343ac490c6b08c0cf0d704e881d0d660be923fd3076db3e932007e726e3", size = 4466074 },
    { url = "https://files.pythonhosted.org/packages/ac/68/b4f4a10928e26c941b1b6a179143af9f4d27d88fe84a6a3c53592d2e76bf/cryptography-46.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:22d7e97932f511d6b0b04f2bfd818d73dcd5928db509460aaf48384778eb6d20", size = 4420569 },
    { url = "https://files.pythonhosted.org/packages/a3/49/3746dab4c0d1979888f125226357d3262a6dd40e114ac29e3d2abdf1ec55/cryptography-46.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:d55f3dffadd674514ad19451161118fd010988540cee43d8bc20675e775925de", size = 4681941 },
@@ -323,8 +333,10 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/26/42/fa8389d4478368743e24e61eea78846a0006caffaf72ea24a15159215a14/cryptography-46.0.3-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:15ab9b093e8f09daab0f2159bb7e47532596075139dd74365da52ecc9cb46c5d", size = 4440029 },
    { url = "https://files.pythonhosted.org/packages/5f/eb/f483db0ec5ac040824f269e93dd2bd8a21ecd1027e77ad7bdf6914f2fd80/cryptography-46.0.3-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:46acf53b40ea38f9c6c229599a4a13f0d46a6c3fa9ef19fc1a124d62e338dfa0", size = 4297222 },
    { url = "https://files.pythonhosted.org/packages/fd/cf/da9502c4e1912cb1da3807ea3618a6829bee8207456fbbeebc361ec38ba3/cryptography-46.0.3-cp38-abi3-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:10ca84c4668d066a9878890047f03546f3ae0a6b8b39b697457b7757aaf18dbc", size = 4012280 },
+    { url = "https://files.pythonhosted.org/packages/6b/8f/9adb86b93330e0df8b3dcf03eae67c33ba89958fc2e03862ef1ac2b42465/cryptography-46.0.3-cp38-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:36e627112085bb3b81b19fed209c05ce2a52ee8b15d161b7c643a7d5a88491f3", size = 4978958 },
    { url = "https://files.pythonhosted.org/packages/d1/a0/5fa77988289c34bdb9f913f5606ecc9ada1adb5ae870bd0d1054a7021cc4/cryptography-46.0.3-cp38-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:1000713389b75c449a6e979ffc7dcc8ac90b437048766cef052d4d30b8220971", size = 4473714 },
    { url = "https://files.pythonhosted.org/packages/14/e5/fc82d72a58d41c393697aa18c9abe5ae1214ff6f2a5c18ac470f92777895/cryptography-46.0.3-cp38-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:b02cf04496f6576afffef5ddd04a0cb7d49cf6be16a9059d793a30b035f6b6ac", size = 4296970 },
+    { url = "https://files.pythonhosted.org/packages/78/06/5663ed35438d0b09056973994f1aec467492b33bd31da36e468b01ec1097/cryptography-46.0.3-cp38-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:71e842ec9bc7abf543b47cf86b9a743baa95f4677d22baa4c7d5c69e49e9bc04", size = 4940236 },
    { url = "https://files.pythonhosted.org/packages/fc/59/873633f3f2dcd8a053b8dd1d38f783043b5fce589c0f6988bf55ef57e43e/cryptography-46.0.3-cp38-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:402b58fc32614f00980b66d6e56a5b4118e6cb362ae8f3fda141ba4689bd4506", size = 4472642 },
    { url = "https://files.pythonhosted.org/packages/3d/39/8e71f3930e40f6877737d6f69248cf74d4e34b886a3967d32f919cc50d3b/cryptography-46.0.3-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:ef639cb3372f69ec44915fafcd6698b6cc78fbe0c2ea41be867f6ed612811963", size = 4423126 },
    { url = "https://files.pythonhosted.org/packages/cd/c7/f65027c2810e14c3e7268353b1681932b87e5a48e65505d8cc17c99e36ae/cryptography-46.0.3-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:3b51b8ca4f1c6453d8829e1eb7299499ca7f313900dd4d89a24b8b87c0a780d4", size = 4686573 },
@@ -403,7 +415,7 @@ requires-dist = [
    { name = "mflux", specifier = "==0.15.4" },
    { name = "mlx", marker = "sys_platform == 'darwin'", specifier = "==0.30.4" },
    { name = "mlx", extras = ["cpu"], marker = "sys_platform == 'linux'", specifier = "==0.30.4" },
-    { name = "mlx-lm", specifier = "==0.30.5" },
+    { name = "mlx-lm", git = "https://github.com/ml-explore/mlx-lm?branch=main" },
    { name = "openai-harmony", specifier = ">=0.0.8" },
    { name = "pillow", specifier = ">=11.0,<12.0" },
    { name = "psutil", specifier = ">=7.0.0" },
@@ -1061,7 +1073,7 @@ wheels = [
 [[package]]
 name = "mlx-lm"
 version = "0.30.5"
-source = { registry = "https://pypi.org/simple" }
+source = { git = "https://github.com/ml-explore/mlx-lm?branch=main#96699e6dadb13b82b28285bb131a0741997d19ae" }
 dependencies = [
    { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "mlx", marker = "sys_platform == 'darwin'" },
@@ -1071,10 +1083,6 @@ dependencies = [
    { name = "sentencepiece", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "transformers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/0b/90/4469d9f75f196e6255f59a89441abe0079925d30a001462e1c1c4bc4e6a1/mlx_lm-0.30.5.tar.gz", hash = "sha256:9e6cb258c65b766c6af25cb90958aef40acab67139f05839eef19864cb3154f6", size = 262367 }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/89/ba/66db6e1e5f1ef506655b562932f6bd8f72600116d5f31f92d71c1f200b3f/mlx_lm-0.30.5-py3-none-any.whl", hash = "sha256:a80bc8e3efdebe81813b0f6eb403fb66a7a15071e256f4e7102ada986acb75bb", size = 366716 },
-]

 [[package]]
 name = "mlx-metal"
@@ -1495,9 +1503,6 @@ version = "11.3.0"
 source = { registry = "https://pypi.org/simple" }
 sdist = { url = "https://files.pythonhosted.org/packages/f3/0d/d0d6dea55cd152ce3d6767bb38a8fc10e33796ba4ba210cbab9354b6d238/pillow-11.3.0.tar.gz", hash = "sha256:3828ee7586cd0b2091b6209e5ad53e20d0649bbe87164a459d0676e035e8f523", size = 47113069 }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/1e/93/0952f2ed8db3a5a4c7a11f91965d6184ebc8cd7cbb7941a260d5f018cd2d/pillow-11.3.0-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:1c627742b539bba4309df89171356fcb3cc5a9178355b2727d1b74a6cf155fbd", size = 2128328 },
-    { url = "https://files.pythonhosted.org/packages/4b/e8/100c3d114b1a0bf4042f27e0f87d2f25e857e838034e98ca98fe7b8c0a9c/pillow-11.3.0-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:30b7c02f3899d10f13d7a48163c8969e4e653f8b43416d23d13d1bbfdc93b9f8", size = 2170652 },
-    { url = "https://files.pythonhosted.org/packages/aa/86/3f758a28a6e381758545f7cdb4942e1cb79abd271bea932998fc0db93cb6/pillow-11.3.0-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:7859a4cc7c9295f5838015d8cc0a9c215b77e43d07a25e460f35cf516df8626f", size = 2227443 },
    { url = "https://files.pythonhosted.org/packages/01/f4/91d5b3ffa718df2f53b0dc109877993e511f4fd055d7e9508682e8aba092/pillow-11.3.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:ec1ee50470b0d050984394423d96325b744d55c701a439d2bd66089bff963d3c", size = 5278474 },
    { url = "https://files.pythonhosted.org/packages/f9/0e/37d7d3eca6c879fbd9dba21268427dffda1ab00d4eb05b32923d4fbe3b12/pillow-11.3.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:7db51d222548ccfd274e4572fdbf3e810a5e66b00608862f947b163e613b67dd", size = 4686038 },
    { url = "https://files.pythonhosted.org/packages/ff/b0/3426e5c7f6565e752d81221af9d3676fdbb4f352317ceafd42899aaf5d8a/pillow-11.3.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2d6fcc902a24ac74495df63faad1884282239265c6839a0a6416d33faedfae7e", size = 5864407 },
@@ -2273,7 +2278,7 @@ wheels = [

 [[package]]
 name = "transformers"
-version = "5.0.0rc3"
+version = "5.0.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -2282,15 +2287,14 @@ dependencies = [
    { name = "packaging", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "pyyaml", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "regex", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "requests", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "safetensors", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "tokenizers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "tqdm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
    { name = "typer-slim", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/3f/a3/7c116a8d85f69ea7749cf4c2df79e64c35d028e5fc7ea0168f299d03b8c7/transformers-5.0.0rc3.tar.gz", hash = "sha256:a0315b92b7e087617ade42ec9e6e92ee7620541cc5d6a3331886c52cbe306f5c", size = 8388520 }
+sdist = { url = "https://files.pythonhosted.org/packages/bc/79/845941711811789c85fb7e2599cea425a14a07eda40f50896b9d3fda7492/transformers-5.0.0.tar.gz", hash = "sha256:5f5634efed6cf76ad068cc5834c7adbc32db78bbd6211fb70df2325a9c37dec8", size = 8424830 }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/1e/f2/ae2b8968764253bdf38a48dee3c299b8d0bedf7c8ffbe3449fca9bd95338/transformers-5.0.0rc3-py3-none-any.whl", hash = "sha256:383fad27f4f73092d330e45fae384681e5c8521e1dc1cf6cb1a297780e68bf2d", size = 10107087 },
+    { url = "https://files.pythonhosted.org/packages/52/f3/ac976fa8e305c9e49772527e09fbdc27cc6831b8a2f6b6063406626be5dd/transformers-5.0.0-py3-none-any.whl", hash = "sha256:587086f249ce64c817213cf36afdb318d087f790723e9b3d4500b97832afd52d", size = 10142091 },
 ]

 [[package]]