fix: add missing tasks arg to get_transition_events()

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

AGENTS.md

@@ -194,3 +194,40 @@ GitHub's API doesn't support direct image upload for PR comments. Workaround:
    git push origin <branch>
    ```
    The images still render in the PR comment because they reference the permanent commit SHA.
+
+## Running exo Remotely via SSH (macOS mDNS)
+
+**CRITICAL: On macOS, mDNS multicast (used for peer discovery) only works when the process runs in a proper macOS user session.** Background processes started via `nohup ... &`, `screen`, or plain SSH commands will NOT send mDNS packets and nodes will never discover each other.
+
+### The Problem
+When you SSH into a Mac and run `nohup uv run exo &`, the process runs in a detached session without access to macOS multicast networking. The exo node will start but will never discover peers, even if they're on the same network.
+
+### The Solution: Use `open` with a `.command` wrapper
+
+Create a `.command` script that `open` will execute in the proper macOS GUI session context:
+
+```bash
+# 1. Create wrapper script on the remote machine
+ssh user@remote-mac "cat > /tmp/run_exo.command << 'SCRIPT'
+#!/bin/bash
+export PATH=/opt/homebrew/bin:\$HOME/.local/bin:\$PATH
+export EXO_LIBP2P_NAMESPACE=your-namespace  # must match across all nodes
+cd ~/path/to/exo
+exec uv run exo -vv 2>&1 | tee /tmp/exo.log
+SCRIPT
+chmod +x /tmp/run_exo.command"
+
+# 2. Launch it via `open` (runs in macOS GUI session with proper mDNS)
+ssh user@remote-mac "open /tmp/run_exo.command"
+
+# 3. Check logs
+ssh user@remote-mac "tail -f /tmp/exo.log"
+```
+
+### Key Details
+- **`EXO_LIBP2P_NAMESPACE`**: All nodes in a cluster MUST use the same namespace value. The EXO.app uses a build-specific namespace (check with `ps eww <pid> | grep NAMESPACE`). If mixing dev builds with EXO.app, set the dev build's namespace to match.
+- **`open *.command`**: This is the macOS equivalent of double-clicking the script in Finder. It runs in the user's GUI session with full network access.
+- **Do NOT use**: `nohup ... &`, `screen -dm`, `tmux new-session -d`, or `sshpass`. These all create detached sessions where mDNS won't work.
+- **Killing**: `ssh user@remote-mac "pkill -f 'python.*exo'"` works fine for stopping.
+- **Dashboard**: Must be built before running: `cd dashboard && npm install && npm run build && cd ..`. Node.js is at `/opt/homebrew/bin/node` on Apple Silicon Macs.
+- **Verifying cluster**: `curl -s http://localhost:52415/state | python3 -c "import json,sys; s=json.load(sys.stdin); print(len(s['topology']['nodes']), 'nodes')"` — should show 2+ nodes.

README.md

@@ -72,16 +72,23 @@ There are two ways to run exo:
 
 ### Run from Source (macOS)
 
+If you have [Nix](https://nixos.org/) installed, you can skip most of the steps below and run exo directly (after accepting the Cachix cache):
+
+```bash
+nix run .#exo
+```
+
 **Prerequisites:**
+- [Xcode](https://developer.apple.com/xcode/) (provides the Metal ToolChain required for MLX compilation)
 - [brew](https://github.com/Homebrew/brew) (for simple package management on macOS)
-  
+
   ```bash
   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
   ```
 - [uv](https://github.com/astral-sh/uv) (for Python dependency management)
 - [macmon](https://github.com/vladkens/macmon) (for hardware monitoring on Apple Silicon)
 - [node](https://github.com/nodejs/node) (for building the dashboard)
-  
+
   ```bash
   brew install uv macmon node
   ```

dashboard/src/lib/components/ChatSidebar.svelte

@@ -185,11 +185,7 @@
 
     let instanceType: string | null = null;
     if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";
 
     let sharding: string | null = null;
     const inst = instance as {

dashboard/src/lib/components/ModelCard.svelte

@@ -21,7 +21,7 @@
     } | null;
     nodes?: Record<string, NodeInfo>;
     sharding?: "Pipeline" | "Tensor";
-    runtime?: "MlxRing" | "MlxIbv" | "MlxJaccl";
+    runtime?: "MlxRing" | "MlxJaccl";
     onLaunch?: () => void;
     tags?: string[];
     apiPreview?: PlacementPreview | null;
@@ -348,7 +348,7 @@
   // Debug mode state
   const isDebugMode = $derived(debugMode());
   const topology = $derived(topologyData());
-  const isRdma = $derived(runtime === "MlxIbv" || runtime === "MlxJaccl");
+  const isRdma = $derived(runtime === "MlxJaccl");
 
   // Get interface name for an IP from node data
   function getInterfaceForIp(nodeId: string, ip?: string): string | null {
@@ -575,7 +575,7 @@
       >
         {runtime === "MlxRing"
           ? "MLX Ring"
-          : runtime === "MlxIbv" || runtime === "MlxJaccl"
+          : runtime === "MlxJaccl"
             ? "MLX RDMA"
             : runtime}
       </span>

dashboard/src/lib/stores/app.svelte.ts

@@ -168,7 +168,7 @@ export interface ModelDownloadStatus {
 export interface PlacementPreview {
   model_id: string;
   sharding: "Pipeline" | "Tensor";
-  instance_meta: "MlxRing" | "MlxIbv" | "MlxJaccl";
+  instance_meta: "MlxRing" | "MlxJaccl";
   instance: unknown | null;
   memory_delta_by_node: Record<string, number> | null;
   error: string | null;
@@ -219,7 +219,6 @@ interface RawStateResponse {
     string,
     {
       MlxRingInstance?: Instance;
-      MlxIbvInstance?: Instance;
       MlxJacclInstance?: Instance;
     }
   >;
@@ -250,6 +249,20 @@ interface RawStateResponse {
   >;
   // Thunderbolt bridge cycles (nodes with bridge enabled forming loops)
   thunderboltBridgeCycles?: string[][];
+  // MetaInstances (declarative instance constraints)
+  metaInstances?: Record<string, MetaInstanceData>;
+}
+
+export interface MetaInstanceData {
+  metaInstanceId: string;
+  modelId: string;
+  sharding: string;
+  instanceMeta: string;
+  minNodes: number;
+  nodeIds: string[] | null;
+  placementError: string | null;
+  consecutiveFailures: number;
+  lastFailureError: string | null;
 }
 
 export interface MessageAttachment {
@@ -537,6 +550,7 @@ class AppStore {
   previewNodeFilter = $state<Set<string>>(new Set());
   lastUpdate = $state<number | null>(null);
   nodeIdentities = $state<Record<string, RawNodeIdentity>>({});
+  metaInstances = $state<Record<string, MetaInstanceData>>({});
   thunderboltBridgeCycles = $state<string[][]>([]);
   nodeThunderbolt = $state<
     Record<
@@ -895,11 +909,7 @@ class AppStore {
 
     let instanceType: string | null = null;
     if (instanceTag === "MlxRingInstance") instanceType = "MLX Ring";
-    else if (
-      instanceTag === "MlxIbvInstance" ||
-      instanceTag === "MlxJacclInstance"
-    )
-      instanceType = "MLX RDMA";
+    else if (instanceTag === "MlxJacclInstance") instanceType = "MLX RDMA";
 
     let sharding: string | null = null;
     const inst = instance as {
@@ -1273,6 +1283,8 @@ class AppStore {
       this.nodeThunderbolt = data.nodeThunderbolt ?? {};
       // RDMA ctl status per node
       this.nodeRdmaCtl = data.nodeRdmaCtl ?? {};
+      // MetaInstances
+      this.metaInstances = data.metaInstances ?? {};
       // Thunderbolt bridge cycles
       this.thunderboltBridgeCycles = data.thunderboltBridgeCycles ?? [];
       // Thunderbolt bridge status per node
@@ -3044,6 +3056,7 @@ export const tps = () => appStore.tps;
 export const totalTokens = () => appStore.totalTokens;
 export const topologyData = () => appStore.topologyData;
 export const instances = () => appStore.instances;
+export const metaInstances = () => appStore.metaInstances;
 export const runners = () => appStore.runners;
 export const downloads = () => appStore.downloads;
 export const nodeDisk = () => appStore.nodeDisk;

flake.nix

@@ -115,7 +115,7 @@
           packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
             let
               uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);
-              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx") uvLock.package);
+              mlxPackage = builtins.head (builtins.filter (p: p.name == "mlx" && p.source ? git) uvLock.package);
               uvLockMlxVersion = mlxPackage.version;
             in
             {

nix/mlx.nix

@@ -41,16 +41,16 @@ let
 
   mlx = stdenv.mkDerivation rec {
     pname = "mlx";
-    version = let v = "0.30.6"; in
+    version = let v = "0.30.7.dev20260217+50487b41"; in
       assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
       v;
     pyproject = true;
 
     src = fetchFromGitHub {
-      owner = "ml-explore";
-      repo = "mlx";
-      tag = "v${version}";
-      hash = "sha256-avD5EGhwgmPdXLAyQSqTO6AXk/W3ziH+f6AetjK3Sdo=";
+      owner = "rltakashige";
+      repo = "mlx-jaccl-fix-small-recv";
+      rev = "50487b4141f3c951122655db3b83df5146c1fbeb";
+      hash = "sha256-IL4a9vMX5nocgJU1WG4zE8hArHkHJtnh4sdYh3od5zU=";
     };
 
     patches = [

pyproject.toml

@@ -17,7 +17,7 @@ dependencies = [
     "loguru>=0.7.3",
     "exo_pyo3_bindings", # rust bindings
     "anyio==4.11.0",
-    "mlx==0.30.6; sys_platform == 'darwin'",
+    "mlx; sys_platform == 'darwin'",
     "mlx[cpu]==0.30.6; sys_platform == 'linux'",
     "mlx-lm==0.30.6",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
@@ -64,6 +64,7 @@ members = [
 
 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
+mlx = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" }
 #mlx-lm = { git = "https://github.com/davidmcc73/mlx-lm", branch = "stable" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }

python/parts.nix

@@ -58,6 +58,21 @@
         lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux (
           (lib.mapAttrs (_: ignoreMissing) nvidiaPackages) // {
             mlx = ignoreMissing prev.mlx;
+            mlx-cuda-13 = prev.mlx-cuda-13.overrideAttrs (old: {
+              buildInputs = (old.buildInputs or [ ]) ++ [
+                final.nvidia-cublas
+                final.nvidia-cuda-nvrtc
+                final.nvidia-cudnn-cu13
+                final.nvidia-nccl-cu13
+              ];
+              preFixup = ''
+                addAutoPatchelfSearchPath ${final.nvidia-cublas}
+                addAutoPatchelfSearchPath ${final.nvidia-cuda-nvrtc}
+                addAutoPatchelfSearchPath ${final.nvidia-cudnn-cu13}
+                addAutoPatchelfSearchPath ${final.nvidia-nccl-cu13}
+              '';
+              autoPatchelfIgnoreMissingDeps = [ "libcuda.so.1" ];
+            });
             torch = ignoreMissing prev.torch;
             triton = ignoreMissing prev.triton;
           }
@@ -74,14 +89,25 @@
           linuxOverlay
         ]
       );
-      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
+      # mlx-cpu and mlx-cuda-13 both ship mlx/ site-packages files; keep first.
+      # mlx-cpu/mlx-cuda-13 and nvidia-cudnn-cu12/cu13 ship overlapping files.
+      venvCollisionPaths = lib.optionals pkgs.stdenv.hostPlatform.isLinux [
+        "lib/python3.13/site-packages/mlx*"
+        "lib/python3.13/site-packages/nvidia*"
+      ];
+
+      exoVenv = (pythonSet.mkVirtualEnv "exo-env" workspace.deps.default).overrideAttrs {
+        venvIgnoreCollisions = venvCollisionPaths;
+      };
 
       # Virtual environment with dev dependencies for testing
-      testVenv = pythonSet.mkVirtualEnv "exo-test-env" (
+      testVenv = (pythonSet.mkVirtualEnv "exo-test-env" (
         workspace.deps.default // {
           exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
         }
-      );
+      )).overrideAttrs {
+        venvIgnoreCollisions = venvCollisionPaths;
+      };
 
       mkPythonScript = name: path: pkgs.writeShellApplication {
         inherit name;

src/exo/download/coordinator.py

@@ -314,7 +314,17 @@ class DownloadCoordinator:
                             ),
                         )
                     elif progress.status in ["in_progress", "not_started"]:
-                        if progress.downloaded_bytes_this_session.in_bytes == 0:
+                        if (
+                            progress.downloaded_bytes.in_bytes
+                            >= progress.total_bytes.in_bytes
+                            > 0
+                        ):
+                            status = DownloadCompleted(
+                                node_id=self.node_id,
+                                shard_metadata=progress.shard,
+                                total_bytes=progress.total_bytes,
+                            )
+                        elif progress.downloaded_bytes_this_session.in_bytes == 0:
                             status = DownloadPending(
                                 node_id=self.node_id,
                                 shard_metadata=progress.shard,

src/exo/main.py

@@ -254,7 +254,7 @@ def main():
     target = min(max(soft, 65535), hard)
     resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))
 
-    mp.set_start_method("spawn")
+    mp.set_start_method("spawn", force=True)
     # TODO: Refactor the current verbosity system
     logger_setup(EXO_LOG, args.verbosity)
     logger.info("Starting EXO")

src/exo/master/api.py

@@ -71,8 +71,13 @@ from exo.shared.types.api import (
     ChatCompletionResponse,
     CreateInstanceParams,
     CreateInstanceResponse,
+    CreateMetaInstanceParams,
+    CreateMetaInstanceResponse,
     DeleteDownloadResponse,
     DeleteInstanceResponse,
+    DeleteMetaInstanceResponse,
+    DistributeModelParams,
+    DistributeModelResponse,
     ErrorInfo,
     ErrorResponse,
     FinishReason,
@@ -115,8 +120,11 @@ from exo.shared.types.claude_api import (
 from exo.shared.types.commands import (
     Command,
     CreateInstance,
+    CreateMetaInstance,
     DeleteDownload,
     DeleteInstance,
+    DeleteMetaInstance,
+    DistributeModel,
     DownloadCommand,
     ForwarderCommand,
     ForwarderDownloadCommand,
@@ -129,7 +137,7 @@ from exo.shared.types.commands import (
     TaskFinished,
     TextGeneration,
 )
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId
+from exo.shared.types.common import CommandId, Id, MetaInstanceId, NodeId, SessionId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -138,11 +146,13 @@ from exo.shared.types.events import (
     TracesMerged,
 )
 from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.openai_responses import (
     ResponsesRequest,
     ResponsesResponse,
 )
 from exo.shared.types.state import State
+from exo.shared.types.worker.downloads import DownloadCompleted
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
@@ -276,6 +286,9 @@ class API:
         self.app.get("/instance/previews")(self.get_placement_previews)
         self.app.get("/instance/{instance_id}")(self.get_instance)
         self.app.delete("/instance/{instance_id}")(self.delete_instance)
+        self.app.get("/meta_instances")(self.list_meta_instances)
+        self.app.post("/meta_instance")(self.create_meta_instance)
+        self.app.delete("/meta_instance/{meta_instance_id}")(self.delete_meta_instance)
         self.app.get("/models")(self.get_models)
         self.app.get("/v1/models")(self.get_models)
         self.app.post("/models/add")(self.add_custom_model)
@@ -299,18 +312,34 @@ class API:
         self.app.get("/events")(self.stream_events)
         self.app.post("/download/start")(self.start_download)
         self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
+        self.app.post("/v1/models/{model_id:path}/distribute")(self.distribute_model)
         self.app.get("/v1/traces")(self.list_traces)
         self.app.get("/v1/traces/{task_id}")(self.get_trace)
         self.app.get("/v1/traces/{task_id}/stats")(self.get_trace_stats)
         self.app.get("/v1/traces/{task_id}/raw")(self.get_trace_raw)
 
     async def place_instance(self, payload: PlaceInstanceParams):
+        model_card = await ModelCard.load(payload.model_id)
         command = PlaceInstance(
-            model_card=await ModelCard.load(payload.model_id),
+            model_card=model_card,
             sharding=payload.sharding,
             instance_meta=payload.instance_meta,
             min_nodes=payload.min_nodes,
         )
+
+        # Validate placement before sending — fail fast with a clear error
+        # instead of silently dropping the command in the master.
+        try:
+            get_instance_placements(
+                command,
+                topology=self.state.topology,
+                current_instances=self.state.instances,
+                node_memory=self.state.node_memory,
+                node_network=self.state.node_network,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc)) from exc
+
         await self._send(command)
 
         return CreateInstanceResponse(
@@ -522,6 +551,44 @@ class API:
             instance_id=instance_id,
         )
 
+    def list_meta_instances(self) -> dict[MetaInstanceId, MetaInstance]:
+        return dict(self.state.meta_instances)
+
+    async def create_meta_instance(
+        self, payload: CreateMetaInstanceParams
+    ) -> CreateMetaInstanceResponse:
+        meta_instance = MetaInstance(
+            model_id=payload.model_id,
+            sharding=payload.sharding,
+            instance_meta=payload.instance_meta,
+            min_nodes=payload.min_nodes,
+            node_ids=payload.node_ids,
+        )
+        command = CreateMetaInstance(meta_instance=meta_instance)
+        await self._send(command)
+        return CreateMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance.meta_instance_id,
+        )
+
+    async def delete_meta_instance(
+        self, meta_instance_id: MetaInstanceId
+    ) -> DeleteMetaInstanceResponse:
+        meta = self.state.meta_instances.get(meta_instance_id)
+        if not meta:
+            raise HTTPException(status_code=404, detail="MetaInstance not found")
+
+        # Command processor handles cascade-deleting backing instances
+        command = DeleteMetaInstance(meta_instance_id=meta_instance_id)
+        await self._send(command)
+
+        return DeleteMetaInstanceResponse(
+            message="Command received.",
+            command_id=command.command_id,
+            meta_instance_id=meta_instance_id,
+        )
+
     async def _token_chunk_stream(
         self, command_id: CommandId
     ) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
@@ -541,10 +608,10 @@ class API:
                         break
 
         except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
+            cancel_command = TaskCancelled(cancelled_command_id=command_id)
             with anyio.CancelScope(shield=True):
                 await self.command_sender.send(
-                    ForwarderCommand(origin=self.node_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=cancel_command)
                 )
             raise
         finally:
@@ -884,10 +951,10 @@ class API:
                         del image_metadata[key]
 
         except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
+            cancel_command = TaskCancelled(cancelled_command_id=command_id)
             with anyio.CancelScope(shield=True):
                 await self.command_sender.send(
-                    ForwarderCommand(origin=self.node_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=cancel_command)
                 )
             raise
         finally:
@@ -970,10 +1037,10 @@ class API:
 
             return (images, stats if capture_stats else None)
         except anyio.get_cancelled_exc_class():
-            command = TaskCancelled(cancelled_command_id=command_id)
+            cancel_command = TaskCancelled(cancelled_command_id=command_id)
             with anyio.CancelScope(shield=True):
                 await self.command_sender.send(
-                    ForwarderCommand(origin=self.node_id, command=command)
+                    ForwarderCommand(origin=self.node_id, command=cancel_command)
                 )
             raise
         finally:
@@ -1495,6 +1562,57 @@ class API:
         await self._send_download(command)
         return DeleteDownloadResponse(command_id=command.command_id)
 
+    async def distribute_model(
+        self, model_id: ModelId, payload: DistributeModelParams
+    ) -> DistributeModelResponse:
+        """Distribute model files from one node to others via MLX distributed."""
+        # Find a source node that has the model downloaded
+        source_node_id: NodeId | None = None
+        for nid, downloads in self.state.downloads.items():
+            for dp in downloads:
+                if (
+                    isinstance(dp, DownloadCompleted)
+                    and dp.shard_metadata.model_card.model_id == model_id
+                ):
+                    source_node_id = nid
+                    break
+            if source_node_id is not None:
+                break
+
+        if source_node_id is None:
+            raise HTTPException(
+                status_code=404,
+                detail=f"No node has model {model_id} downloaded",
+            )
+
+        # Determine target nodes
+        if payload.target_node_ids is not None:
+            target_node_ids = [
+                nid for nid in payload.target_node_ids if nid != source_node_id
+            ]
+        else:
+            target_node_ids = [
+                nid for nid in self.state.topology.list_nodes() if nid != source_node_id
+            ]
+
+        if not target_node_ids:
+            raise HTTPException(
+                status_code=400,
+                detail="No target nodes to distribute to",
+            )
+
+        command = DistributeModel(
+            model_id=model_id,
+            source_node_id=source_node_id,
+            target_node_ids=target_node_ids,
+        )
+        await self._send(command)
+
+        return DistributeModelResponse(
+            command_id=command.command_id,
+            message=f"Distributing {model_id} from {source_node_id} to {len(target_node_ids)} node(s)",
+        )
+
     def _get_trace_path(self, task_id: str) -> Path:
         return EXO_TRACING_CACHE_DIR / f"trace_{task_id}.json"
 

src/exo/master/main.py

@@ -1,4 +1,5 @@
-from datetime import datetime, timedelta, timezone
+from collections.abc import Sequence
+from datetime import datetime, timezone
 
 import anyio
 from anyio.abc import TaskGroup
@@ -12,11 +13,23 @@ from exo.master.placement import (
     get_transition_events,
     place_instance,
 )
+from exo.master.process_managers import ProcessManager
+from exo.master.process_managers.instance_health import InstanceHealthReconciler
+from exo.master.process_managers.meta_instance import MetaInstanceReconciler
+from exo.master.process_managers.node_timeout import NodeTimeoutReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    try_place_for_meta_instance,
+)
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
+from exo.shared.models.model_cards import ModelCard
 from exo.shared.types.commands import (
     CreateInstance,
+    CreateMetaInstance,
     DeleteInstance,
+    DeleteMetaInstance,
+    DistributeModel,
     ForwarderCommand,
     ForwarderDownloadCommand,
     ImageEdits,
@@ -36,8 +49,12 @@ from exo.shared.types.events import (
     IndexedEvent,
     InputChunkReceived,
     InstanceDeleted,
+    JacclSideChannelData,
+    JacclSideChannelGathered,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
     NodeGatheredInfo,
-    NodeTimedOut,
     TaskCreated,
     TaskDeleted,
     TaskStatusUpdated,
@@ -60,7 +77,8 @@ from exo.shared.types.tasks import (
     TextGeneration as TextGenerationTask,
 )
 from exo.shared.types.worker.instances import InstanceId
-from exo.utils.channels import Receiver, Sender, channel
+from exo.shared.types.worker.runners import RunnerId
+from exo.utils.channels import Receiver, Sender
 from exo.utils.event_buffer import MultiSourceBuffer
 
 
@@ -84,16 +102,16 @@ class Master:
         self.local_event_receiver = local_event_receiver
         self.global_event_sender = global_event_sender
         self.download_command_sender = download_command_sender
-        send, recv = channel[Event]()
-        self.event_sender: Sender[Event] = send
-        self._loopback_event_receiver: Receiver[Event] = recv
-        self._loopback_event_sender: Sender[ForwarderEvent] = (
-            local_event_receiver.clone_sender()
-        )
         self._multi_buffer = MultiSourceBuffer[NodeId, Event]()
         self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
         self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
         self._expected_ranks: dict[TaskId, set[int]] = {}
+        self._jaccl_pending: dict[InstanceId, dict[int, dict[RunnerId, bytes]]] = {}
+        self._process_managers: Sequence[ProcessManager] = [
+            InstanceHealthReconciler(),
+            NodeTimeoutReconciler(),
+            MetaInstanceReconciler(),
+        ]
 
     async def run(self):
         logger.info("Starting Master")
@@ -102,15 +120,12 @@ class Master:
             async with self._tg as tg:
                 tg.start_soon(self._event_processor)
                 tg.start_soon(self._command_processor)
-                tg.start_soon(self._loopback_processor)
-                tg.start_soon(self._plan)
+                tg.start_soon(self._reconcile)
         finally:
             self._event_log.close()
             self.global_event_sender.close()
             self.local_event_receiver.close()
             self.command_receiver.close()
-            self._loopback_event_sender.close()
-            self._loopback_event_receiver.close()
 
     async def shutdown(self):
         logger.info("Stopping Master")
@@ -292,6 +307,86 @@ class Master:
                                     )
                                 )
                             generated_events.extend(transition_events)
+                        case CreateMetaInstance():
+                            logger.info(
+                                f"Creating MetaInstance for {command.meta_instance.model_id}"
+                                f" (min_nodes={command.meta_instance.min_nodes},"
+                                f" sharding={command.meta_instance.sharding})"
+                            )
+                            # Apply immediately so self.state is fresh across
+                            # the await below and the reconciler won't race.
+                            await self._apply_and_broadcast(
+                                MetaInstanceCreated(meta_instance=command.meta_instance)
+                            )
+                            # Immediate placement attempt for responsiveness
+                            model_card = await ModelCard.load(
+                                command.meta_instance.model_id
+                            )
+                            # Re-check: reconciler may have satisfied it during the await
+                            meta_id = command.meta_instance.meta_instance_id
+                            still_unsatisfied = any(
+                                m.meta_instance_id == meta_id
+                                for m in find_unsatisfied_meta_instances(
+                                    self.state.meta_instances,
+                                    self.state.instances,
+                                    self.state.topology,
+                                )
+                            )
+                            if still_unsatisfied:
+                                result = try_place_for_meta_instance(
+                                    command.meta_instance,
+                                    model_card,
+                                    self.state.topology,
+                                    self.state.instances,
+                                    self.state.node_memory,
+                                    self.state.node_network,
+                                    self.state.tasks,
+                                )
+                                generated_events.extend(result.events)
+                                if result.error is not None:
+                                    generated_events.append(
+                                        MetaInstancePlacementFailed(
+                                            meta_instance_id=meta_id,
+                                            reason=result.error,
+                                        )
+                                    )
+                        case DeleteMetaInstance():
+                            backing_count = sum(
+                                1
+                                for inst in self.state.instances.values()
+                                if inst.meta_instance_id == command.meta_instance_id
+                            )
+                            logger.info(
+                                f"Deleting MetaInstance {command.meta_instance_id}"
+                                f" (cascade-deleting {backing_count} backing instance(s))"
+                            )
+                            generated_events.append(
+                                MetaInstanceDeleted(
+                                    meta_instance_id=command.meta_instance_id
+                                )
+                            )
+                            # Cascade-delete backing instances atomically,
+                            # cancelling any active tasks first.
+                            for iid, inst in self.state.instances.items():
+                                if inst.meta_instance_id == command.meta_instance_id:
+                                    for task in self.state.tasks.values():
+                                        if (
+                                            task.instance_id == iid
+                                            and task.task_status
+                                            in (
+                                                TaskStatus.Pending,
+                                                TaskStatus.Running,
+                                            )
+                                        ):
+                                            generated_events.append(
+                                                TaskStatusUpdated(
+                                                    task_status=TaskStatus.Cancelled,
+                                                    task_id=task.task_id,
+                                                )
+                                            )
+                                    generated_events.append(
+                                        InstanceDeleted(instance_id=iid)
+                                    )
                         case PlaceInstance():
                             placement = place_instance(
                                 command,
@@ -314,6 +409,36 @@ class Master:
                                 self.state.instances, placement, self.state.tasks
                             )
                             generated_events.extend(transition_events)
+                        case DistributeModel():
+                            from exo.shared.types.worker.instances import InstanceMeta
+                            from exo.shared.types.worker.shards import Sharding
+
+                            model_card = await ModelCard.load(command.model_id)
+                            all_node_ids = set(
+                                [command.source_node_id] + list(command.target_node_ids)
+                            )
+                            place_command = PlaceInstance(
+                                model_card=model_card,
+                                sharding=Sharding.Pipeline,
+                                instance_meta=InstanceMeta.MlxRing,
+                                min_nodes=len(all_node_ids),
+                            )
+                            placement = place_instance(
+                                place_command,
+                                self.state.topology,
+                                self.state.instances,
+                                self.state.node_memory,
+                                self.state.node_network,
+                                required_nodes=all_node_ids,
+                            )
+                            # Mark new instances as transfer-only
+                            for instance_id, instance in placement.items():
+                                if instance_id not in self.state.instances:
+                                    instance.shard_assignments.transfer_only = True
+                            transition_events = get_transition_events(
+                                self.state.instances, placement, self.state.tasks
+                            )
+                            generated_events.extend(transition_events)
                         case SendInputChunk(chunk=chunk):
                             generated_events.append(
                                 InputChunkReceived(
@@ -323,16 +448,19 @@ class Master:
                             )
                         case TaskCancelled():
                             if (
-                                task_id := self.command_task_mapping.get(
-                                    command.cancelled_command_id
-                                )
-                            ) is not None:
+                                command.cancelled_command_id
+                                in self.command_task_mapping
+                            ):
                                 generated_events.append(
-                                    TaskStatusUpdated(
-                                        task_status=TaskStatus.Cancelled,
-                                        task_id=task_id,
+                                    TaskDeleted(
+                                        task_id=self.command_task_mapping[
+                                            command.cancelled_command_id
+                                        ]
                                     )
                                 )
+                                del self.command_task_mapping[
+                                    command.cancelled_command_id
+                                ]
                         case TaskFinished():
                             generated_events.append(
                                 TaskDeleted(
@@ -341,9 +469,10 @@ class Master:
                                     ]
                                 )
                             )
-                            self.command_task_mapping.pop(
-                                command.finished_command_id, None
-                            )
+                            if command.finished_command_id in self.command_task_mapping:
+                                del self.command_task_mapping[
+                                    command.finished_command_id
+                                ]
                         case RequestEventLog():
                             # We should just be able to send everything, since other buffers will ignore old messages
                             # rate limit to 1000 at a time
@@ -354,31 +483,32 @@ class Master:
                             ):
                                 await self._send_event(IndexedEvent(idx=i, event=event))
                     for event in generated_events:
-                        await self.event_sender.send(event)
+                        await self._apply_and_broadcast(event)
                 except ValueError as e:
                     logger.opt(exception=e).warning("Error in command processor")
 
-    # These plan loops are the cracks showing in our event sourcing architecture - more things could be commands
-    async def _plan(self) -> None:
+    async def _apply_and_broadcast(self, event: Event) -> None:
+        """Apply event to state, persist to disk, and broadcast to workers.
+
+        State is updated synchronously (before any await), so callers can
+        rely on ``self.state`` reflecting this event immediately after the
+        call.  Python's cooperative scheduling guarantees no interleaving
+        between the state read and write.
+        """
+        logger.debug(f"Master indexing event: {str(event)[:100]}")
+        indexed = IndexedEvent(event=event, idx=len(self._event_log))
+        self.state = apply(self.state, indexed)
+        event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
+        self._event_log.append(event)
+        await self._send_event(indexed)
+
+    async def _reconcile(self) -> None:
         while True:
-            # kill broken instances
-            connected_node_ids = set(self.state.topology.list_nodes())
-            for instance_id, instance in self.state.instances.items():
-                for node_id in instance.shard_assignments.node_to_runner:
-                    if node_id not in connected_node_ids:
-                        await self.event_sender.send(
-                            InstanceDeleted(instance_id=instance_id)
-                        )
-                        break
-
-            # time out dead nodes
-            for node_id, time in self.state.last_seen.items():
-                now = datetime.now(tz=timezone.utc)
-                if now - time > timedelta(seconds=30):
-                    logger.info(f"Manually removing node {node_id} due to inactivity")
-                    await self.event_sender.send(NodeTimedOut(node_id=node_id))
-
-            await anyio.sleep(10)
+            for pm in self._process_managers:
+                events = await pm.reconcile(self.state)
+                for event in events:
+                    await self._apply_and_broadcast(event)
+            await anyio.sleep(1)
 
     async def _event_processor(self) -> None:
         with self.local_event_receiver as local_events:
@@ -396,32 +526,15 @@ class Master:
                         await self._handle_traces_collected(event)
                         continue
 
-                    logger.debug(f"Master indexing event: {str(event)[:100]}")
-                    indexed = IndexedEvent(event=event, idx=len(self._event_log))
-                    self.state = apply(self.state, indexed)
+                    if isinstance(event, JacclSideChannelData):
+                        await self._apply_and_broadcast(event)
+                        await self._handle_jaccl_side_channel(event)
+                        continue
 
-                    event._master_time_stamp = datetime.now(tz=timezone.utc)  # pyright: ignore[reportPrivateUsage]
                     if isinstance(event, NodeGatheredInfo):
                         event.when = str(datetime.now(tz=timezone.utc))
 
-                    self._event_log.append(event)
-                    await self._send_event(indexed)
-
-    async def _loopback_processor(self) -> None:
-        # this would ideally not be necessary.
-        # this is WAY less hacky than how I was working around this before
-        local_index = 0
-        with self._loopback_event_receiver as events:
-            async for event in events:
-                await self._loopback_event_sender.send(
-                    ForwarderEvent(
-                        origin=NodeId(f"master_{self.node_id}"),
-                        origin_idx=local_index,
-                        session=self.session_id,
-                        event=event,
-                    )
-                )
-                local_index += 1
+                    await self._apply_and_broadcast(event)
 
     # This function is re-entrant, take care!
     async def _send_event(self, event: IndexedEvent):
@@ -453,10 +566,49 @@ class Master:
         for trace_data in self._pending_traces[task_id].values():
             all_trace_data.extend(trace_data)
 
-        await self.event_sender.send(
+        await self._apply_and_broadcast(
             TracesMerged(task_id=task_id, traces=all_trace_data)
         )
 
         del self._pending_traces[task_id]
         if task_id in self._expected_ranks:
             del self._expected_ranks[task_id]
+
+    async def _handle_jaccl_side_channel(self, event: JacclSideChannelData) -> None:
+        """Accumulate SideChannel contributions; when all runners for an instance
+        have submitted for the same sequence, emit JacclSideChannelGathered."""
+        iid = event.instance_id
+        seq = event.sequence
+
+        if iid not in self._jaccl_pending:
+            self._jaccl_pending[iid] = {}
+        if seq not in self._jaccl_pending[iid]:
+            self._jaccl_pending[iid][seq] = {}
+        self._jaccl_pending[iid][seq][event.runner_id] = event.data
+
+        instance = self.state.instances.get(iid)
+        if instance is None:
+            logger.warning(f"JacclSideChannelData for unknown instance {iid}")
+            return
+
+        expected_runners = set(instance.shard_assignments.runner_to_shard.keys())
+        submitted = set(self._jaccl_pending[iid][seq].keys())
+
+        logger.info(
+            f"JACCL side channel: instance={iid} seq={seq} "
+            f"submitted={len(submitted)}/{len(expected_runners)}"
+        )
+
+        if submitted >= expected_runners:
+            gathered = dict(self._jaccl_pending[iid][seq])
+            del self._jaccl_pending[iid][seq]
+            if not self._jaccl_pending[iid]:
+                del self._jaccl_pending[iid]
+
+            await self._apply_and_broadcast(
+                JacclSideChannelGathered(
+                    instance_id=iid,
+                    sequence=seq,
+                    gathered_data=gathered,
+                )
+            )

src/exo/master/placement.py

@@ -6,11 +6,11 @@ from typing import Sequence
 from exo.master.placement_utils import (
     Cycle,
     filter_cycles_by_memory,
+    get_largest_cycles,
     get_mlx_jaccl_coordinators,
     get_mlx_jaccl_devices_matrix,
     get_mlx_ring_hosts_by_node,
     get_shard_assignments,
-    get_smallest_cycles,
 )
 from exo.shared.models.model_cards import ModelId
 from exo.shared.topology import Topology
@@ -106,23 +106,27 @@ def place_instance(
             "Pipeline parallelism is not supported for DeepSeek V3.1 (8-bit)"
         )
 
-    smallest_cycles = get_smallest_cycles(cycles_with_sufficient_memory)
+    largest_cycles = get_largest_cycles(cycles_with_sufficient_memory)
 
-    smallest_rdma_cycles = [
-        cycle for cycle in smallest_cycles if topology.is_rdma_cycle(cycle)
+    largest_rdma_cycles = [
+        cycle for cycle in largest_cycles if topology.is_rdma_cycle(cycle)
     ]
 
-    if command.instance_meta == InstanceMeta.MlxJaccl and smallest_rdma_cycles != []:
-        smallest_cycles = smallest_rdma_cycles
+    if command.instance_meta == InstanceMeta.MlxJaccl:
+        if not largest_rdma_cycles:
+            raise ValueError(
+                "Requested RDMA (MlxJaccl) but no RDMA-connected cycles available"
+            )
+        largest_cycles = largest_rdma_cycles
 
     cycles_with_leaf_nodes: list[Cycle] = [
         cycle
-        for cycle in smallest_cycles
+        for cycle in largest_cycles
         if any(topology.node_is_leaf(node_id) for node_id in cycle)
     ]
 
     selected_cycle = max(
-        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else smallest_cycles,
+        cycles_with_leaf_nodes if cycles_with_leaf_nodes != [] else largest_cycles,
         key=lambda cycle: sum(
             (node_memory[node_id].ram_available for node_id in cycle),
             start=Memory(),

src/exo/master/placement_utils.py

@@ -37,11 +37,11 @@ def filter_cycles_by_memory(
     return filtered_cycles
 
 
-def get_smallest_cycles(
+def get_largest_cycles(
     cycles: list[Cycle],
 ) -> list[Cycle]:
-    min_nodes = min(len(cycle) for cycle in cycles)
-    return [cycle for cycle in cycles if len(cycle) == min_nodes]
+    max_nodes = max(len(cycle) for cycle in cycles)
+    return [cycle for cycle in cycles if len(cycle) == max_nodes]
 
 
 def allocate_layers_proportionally(

src/exo/master/process_managers/__init__.py

@@ -0,0 +1,12 @@
+from collections.abc import Sequence
+from typing import Protocol, runtime_checkable
+
+from exo.shared.types.events import Event
+from exo.shared.types.state import State
+
+
+@runtime_checkable
+class ProcessManager(Protocol):
+    """A reconciliation step that examines state and returns corrective events."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]: ...

src/exo/master/process_managers/instance_health.py

@@ -0,0 +1,62 @@
+from collections.abc import Sequence
+from typing import final
+
+from loguru import logger
+
+from exo.master.reconcile import instance_connections_healthy, instance_runners_failed
+from exo.shared.types.events import Event, InstanceDeleted, InstanceRetrying
+from exo.shared.types.state import State
+
+MAX_INSTANCE_RETRIES = 3
+
+
+@final
+class InstanceHealthReconciler:
+    """Delete instances whose network connections are broken or whose runners have all failed."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        events: list[Event] = []
+        for instance_id, instance in state.instances.items():
+            if not instance_connections_healthy(instance, state.topology):
+                events.append(
+                    InstanceDeleted(
+                        instance_id=instance_id,
+                        failure_error="Network connection lost",
+                    )
+                )
+                continue
+
+            is_failed, error_message = instance_runners_failed(
+                instance, state.runners, state.node_identities
+            )
+            if is_failed:
+                # Retry within the same instance if backed by a MetaInstance
+                mid = instance.meta_instance_id
+                mi = state.meta_instances.get(mid) if mid else None
+                if mid and mi and mi.consecutive_failures < MAX_INSTANCE_RETRIES:
+                    logger.info(
+                        f"Instance {instance_id} failed (attempt"
+                        f" {mi.consecutive_failures + 1}/{MAX_INSTANCE_RETRIES}),"
+                        f" retrying: {error_message}"
+                    )
+                    events.append(
+                        InstanceRetrying(
+                            instance_id=instance_id,
+                            meta_instance_id=mid,
+                            failure_error=error_message or "Runner failed",
+                        )
+                    )
+                else:
+                    if mid and mi:
+                        logger.warning(
+                            f"Instance {instance_id} exceeded retry limit"
+                            f" ({MAX_INSTANCE_RETRIES}), deleting:"
+                            f" {error_message}"
+                        )
+                    events.append(
+                        InstanceDeleted(
+                            instance_id=instance_id,
+                            failure_error=error_message,
+                        )
+                    )
+        return events

src/exo/master/process_managers/meta_instance.py

@@ -0,0 +1,92 @@
+from collections.abc import Sequence
+from typing import final
+
+import anyio
+from loguru import logger
+
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    try_place_for_meta_instance,
+)
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.types.events import Event, InstanceCreated, MetaInstancePlacementFailed
+from exo.shared.types.state import State
+from exo.shared.types.worker.instances import Instance, InstanceId
+
+MODEL_CARD_LOAD_TIMEOUT_SECONDS = 10
+
+
+@final
+class MetaInstanceReconciler:
+    """Place instances for unsatisfied MetaInstances."""
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        all_events: list[Event] = []
+        # Local copy for intermediate tracking — so placement of B
+        # sees A's instance and doesn't double-place on same resources.
+        current_instances: dict[InstanceId, Instance] = dict(state.instances)
+
+        unsatisfied = find_unsatisfied_meta_instances(
+            state.meta_instances,
+            current_instances,
+            state.topology,
+        )
+        for meta_instance in unsatisfied:
+            try:
+                with anyio.fail_after(MODEL_CARD_LOAD_TIMEOUT_SECONDS):
+                    model_card = await ModelCard.load(meta_instance.model_id)
+            except TimeoutError:
+                logger.warning(
+                    f"ModelCard.load timed out for {meta_instance.model_id}, skipping this cycle"
+                )
+                continue
+            except Exception as exc:
+                logger.warning(
+                    f"ModelCard.load failed for {meta_instance.model_id}: {exc}"
+                )
+                error = f"Failed to load model card: {exc}"
+                if meta_instance.placement_error != error:
+                    all_events.append(
+                        MetaInstancePlacementFailed(
+                            meta_instance_id=meta_instance.meta_instance_id,
+                            reason=error,
+                        )
+                    )
+                continue
+
+            result = try_place_for_meta_instance(
+                meta_instance,
+                model_card,
+                state.topology,
+                current_instances,
+                state.node_memory,
+                state.node_network,
+                state.tasks,
+            )
+            # Update local instance map so next placement sees this one
+            for event in result.events:
+                if isinstance(event, InstanceCreated):
+                    logger.info(
+                        f"MetaInstance reconciler placed instance"
+                        f" {event.instance.instance_id} for"
+                        f" {meta_instance.model_id}"
+                    )
+                    current_instances[event.instance.instance_id] = event.instance
+            all_events.extend(result.events)
+
+            # Emit placement failure if error differs from what's already in state
+            if (
+                result.error is not None
+                and meta_instance.placement_error != result.error
+            ):
+                logger.warning(
+                    f"MetaInstance placement failed for"
+                    f" {meta_instance.model_id}: {result.error}"
+                )
+                all_events.append(
+                    MetaInstancePlacementFailed(
+                        meta_instance_id=meta_instance.meta_instance_id,
+                        reason=result.error,
+                    )
+                )
+        return all_events

src/exo/master/process_managers/node_timeout.py

@@ -0,0 +1,27 @@
+from collections.abc import Sequence
+from datetime import datetime, timedelta, timezone
+from typing import final
+
+from loguru import logger
+
+from exo.shared.types.events import Event, NodeTimedOut
+from exo.shared.types.state import State
+
+_DEFAULT_TIMEOUT = timedelta(seconds=30)
+
+
+@final
+class NodeTimeoutReconciler:
+    """Time out nodes that haven't been seen recently."""
+
+    def __init__(self, timeout: timedelta = _DEFAULT_TIMEOUT) -> None:
+        self.timeout = timeout
+
+    async def reconcile(self, state: State) -> Sequence[Event]:
+        now = datetime.now(tz=timezone.utc)
+        events: list[Event] = []
+        for node_id, last_seen in state.last_seen.items():
+            if now - last_seen > self.timeout:
+                logger.info(f"Removing node {node_id} due to inactivity")
+                events.append(NodeTimedOut(node_id=node_id))
+        return events

src/exo/master/reconcile.py

@@ -0,0 +1,244 @@
+from collections.abc import Mapping, Sequence
+from typing import NamedTuple
+
+from loguru import logger
+
+from exo.master.placement import get_transition_events, place_instance
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.topology import Topology
+from exo.shared.types.commands import PlaceInstance
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.events import Event
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.profiling import MemoryUsage, NodeIdentity, NodeNetworkInfo
+from exo.shared.types.tasks import Task, TaskId
+from exo.shared.types.topology import RDMAConnection, SocketConnection
+from exo.shared.types.worker.instances import (
+    BaseInstance,
+    Instance,
+    InstanceId,
+    MlxJacclInstance,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerShutdown,
+    RunnerStatus,
+)
+
+
+class PlacementResult(NamedTuple):
+    """Result of a placement attempt: events to apply and optional error reason."""
+
+    events: Sequence[Event]
+    error: str | None
+
+
+def _get_ring_order(instance: BaseInstance) -> list[NodeId]:
+    """Reconstruct ring order from shard device_rank."""
+    node_ranks: list[tuple[NodeId, int]] = []
+    for node_id, runner_id in instance.shard_assignments.node_to_runner.items():
+        shard = instance.shard_assignments.runner_to_shard[runner_id]
+        node_ranks.append((node_id, shard.device_rank))
+    node_ranks.sort(key=lambda x: x[1])
+    return [node_id for node_id, _ in node_ranks]
+
+
+def _ring_connections_healthy(instance: MlxRingInstance, topology: Topology) -> bool:
+    """Check that the specific IPs used by a ring instance still exist in the topology."""
+    ring = _get_ring_order(instance)
+    n = len(ring)
+    for node in ring:
+        hosts = instance.hosts_by_node[node]
+        for idx in range(n):
+            host = hosts[idx]
+            if host.ip in ("0.0.0.0", "198.51.100.1"):
+                continue  # self or placeholder
+            # Real connection: node → ring[idx]. Check specific IP.
+            connections = topology.get_all_connections_between(node, ring[idx])
+            if not any(
+                isinstance(c, SocketConnection)
+                and c.sink_multiaddr.ip_address == host.ip
+                for c in connections
+            ):
+                return False
+    return True
+
+
+def _jaccl_connections_healthy(instance: MlxJacclInstance, topology: Topology) -> bool:
+    """Check that the specific RDMA interfaces used by a JACCL instance still exist."""
+    ring = _get_ring_order(instance)
+    n = len(ring)
+    for i in range(n):
+        for j in range(n):
+            iface = instance.jaccl_devices[i][j]
+            if iface is None:
+                continue
+            connections = topology.get_all_connections_between(ring[i], ring[j])
+            if not any(
+                isinstance(c, RDMAConnection) and c.source_rdma_iface == iface
+                for c in connections
+            ):
+                return False
+    return True
+
+
+def instance_connections_healthy(instance: Instance, topology: Topology) -> bool:
+    """Check that an instance's nodes and specific connections are still in the topology."""
+    instance_nodes = set(instance.shard_assignments.node_to_runner.keys())
+    if not all(topology.contains_node(n) for n in instance_nodes):
+        return False
+    if len(instance_nodes) <= 1:
+        return True
+    match instance:
+        case MlxRingInstance():
+            return _ring_connections_healthy(instance, topology)
+        case MlxJacclInstance():
+            return _jaccl_connections_healthy(instance, topology)
+
+
+def instance_runners_failed(
+    instance: Instance,
+    runners: Mapping[RunnerId, RunnerStatus],
+    node_identities: Mapping[NodeId, NodeIdentity],
+) -> tuple[bool, str | None]:
+    """Check if an instance's runners have all reached terminal failure states.
+
+    Returns ``(True, error_message)`` when ALL runners are terminal
+    (``RunnerFailed`` or ``RunnerShutdown``) and at least one is ``RunnerFailed``.
+
+    Returns ``(False, None)`` when runners are still active, haven't reported
+    yet, or all gracefully shut down (no ``RunnerFailed``).
+    """
+    instance_runner_ids = set(instance.shard_assignments.node_to_runner.values())
+
+    if not instance_runner_ids:
+        return False, None
+
+    # Build reverse mapping: runner_id -> node_id
+    runner_to_node: dict[RunnerId, NodeId] = {
+        runner_id: node_id
+        for node_id, runner_id in instance.shard_assignments.node_to_runner.items()
+    }
+
+    has_any_failed = False
+    error_messages: list[str] = []
+
+    for runner_id in instance_runner_ids:
+        status = runners.get(runner_id)
+        if status is None:
+            # Runner hasn't reported yet — instance is still starting
+            return False, None
+        if isinstance(status, RunnerFailed):
+            has_any_failed = True
+            if status.error_message:
+                node_id = runner_to_node.get(runner_id)
+                name = (
+                    node_identities[node_id].friendly_name
+                    if node_id and node_id in node_identities
+                    else node_id or "unknown"
+                )
+                error_messages.append(f"{name}: {status.error_message}")
+        elif isinstance(status, RunnerShutdown):
+            pass  # Terminal but not a failure indicator on its own
+        else:
+            # Runner is still active (connecting, loading, running, etc.)
+            return False, None
+
+    if has_any_failed:
+        return True, "; ".join(error_messages) if error_messages else "Runner failed"
+
+    # All runners are Shutdown but none Failed — graceful shutdown, not a failure
+    return False, None
+
+
+def instance_satisfies_meta_instance(
+    meta_instance: MetaInstance,
+    instance: Instance,
+) -> bool:
+    """Check if a single instance satisfies a meta-instance's constraints.
+
+    This is a pure constraint check (model, min_nodes, node_ids).
+    Use ``instance_connections_healthy`` separately for topology health.
+    """
+    if instance.shard_assignments.model_id != meta_instance.model_id:
+        return False
+
+    instance_nodes = set(instance.shard_assignments.node_to_runner.keys())
+
+    if len(instance_nodes) < meta_instance.min_nodes:
+        return False
+
+    return meta_instance.node_ids is None or set(meta_instance.node_ids).issubset(
+        instance_nodes
+    )
+
+
+def find_unsatisfied_meta_instances(
+    meta_instances: Mapping[MetaInstanceId, MetaInstance],
+    instances: Mapping[InstanceId, Instance],
+    topology: Topology,
+) -> Sequence[MetaInstance]:
+    """Return meta-instances that have no healthy backing instance."""
+    unsatisfied: list[MetaInstance] = []
+    for meta_id, meta_instance in meta_instances.items():
+        has_healthy_backing = any(
+            instance.meta_instance_id == meta_id
+            and instance_connections_healthy(instance, topology)
+            for instance in instances.values()
+        )
+        if not has_healthy_backing:
+            unsatisfied.append(meta_instance)
+    return unsatisfied
+
+
+def try_place_for_meta_instance(
+    meta_instance: MetaInstance,
+    model_card: ModelCard,
+    topology: Topology,
+    current_instances: Mapping[InstanceId, Instance],
+    node_memory: Mapping[NodeId, MemoryUsage],
+    node_network: Mapping[NodeId, NodeNetworkInfo],
+    tasks: Mapping[TaskId, Task],
+) -> PlacementResult:
+    """Try to place an instance satisfying the meta-instance constraints.
+
+    Returns a :class:`PlacementResult` with events on success, or an error
+    reason on failure.
+    """
+    command = PlaceInstance(
+        model_card=model_card,
+        sharding=meta_instance.sharding,
+        instance_meta=meta_instance.instance_meta,
+        min_nodes=meta_instance.min_nodes,
+    )
+    try:
+        target_instances = place_instance(
+            command,
+            topology,
+            current_instances,
+            node_memory,
+            node_network,
+            required_nodes=(
+                set(meta_instance.node_ids) if meta_instance.node_ids else None
+            ),
+        )
+        # Tag the new instance with meta_instance_id
+        new_instance_ids = set(target_instances.keys()) - set(current_instances.keys())
+        if new_instance_ids:
+            new_id = next(iter(new_instance_ids))
+            target_instances[new_id] = target_instances[new_id].model_copy(
+                update={"meta_instance_id": meta_instance.meta_instance_id}
+            )
+        return PlacementResult(
+            events=list(
+                get_transition_events(current_instances, target_instances, tasks)
+            ),
+            error=None,
+        )
+    except ValueError as e:
+        logger.debug(
+            f"MetaInstance placement not possible for {meta_instance.model_id}: {e}"
+        )
+        return PlacementResult(events=[], error=str(e))

src/exo/master/tests/test_meta_instance_edge_cases.py

@@ -0,0 +1,778 @@
+"""Edge-case and regression tests for MetaInstance lifecycle, concurrent operations, and error handling."""
+
+import pytest
+
+from exo.master.process_managers.instance_health import (
+    MAX_INSTANCE_RETRIES,
+    InstanceHealthReconciler,
+)
+from exo.master.process_managers.meta_instance import MetaInstanceReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    instance_connections_healthy,
+    instance_runners_failed,
+    instance_satisfies_meta_instance,
+)
+from exo.shared.apply import apply
+from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
+from exo.shared.topology import Topology
+from exo.shared.types.common import Host, MetaInstanceId, NodeId
+from exo.shared.types.events import (
+    IndexedEvent,
+    InstanceCreated,
+    InstanceDeleted,
+    InstanceRetrying,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
+    TaskStatusUpdated,
+)
+from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.multiaddr import Multiaddr
+from exo.shared.types.profiling import NodeIdentity
+from exo.shared.types.state import State
+from exo.shared.types.tasks import LoadModel, TaskId, TaskStatus
+from exo.shared.types.topology import Connection, SocketConnection
+from exo.shared.types.worker.instances import (
+    InstanceId,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerReady,
+    ShardAssignments,
+)
+from exo.shared.types.worker.shards import PipelineShardMetadata
+
+# --- Helpers (copied from test_reconcile.py for independence) ---
+
+
+def _model_card(model_id: str = "test-org/test-model") -> ModelCard:
+    return ModelCard(
+        model_id=ModelId(model_id),
+        storage_size=Memory.from_kb(1000),
+        n_layers=10,
+        hidden_size=30,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    )
+
+
+def _topology(*node_ids: str, connect: bool = True) -> Topology:
+    t = Topology()
+    nodes = [NodeId(n) for n in node_ids]
+    for n in nodes:
+        t.add_node(n)
+    if connect and len(nodes) > 1:
+        for i in range(len(nodes)):
+            j = (i + 1) % len(nodes)
+            t.add_connection(
+                Connection(
+                    source=nodes[i],
+                    sink=nodes[j],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{j + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+            t.add_connection(
+                Connection(
+                    source=nodes[j],
+                    sink=nodes[i],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{i + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+    return t
+
+
+def _meta_instance(
+    model_id: str = "test-org/test-model",
+    *,
+    min_nodes: int = 1,
+    node_ids: list[NodeId] | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+    consecutive_failures: int = 0,
+    last_failure_error: str | None = None,
+    placement_error: str | None = None,
+) -> MetaInstance:
+    return MetaInstance(
+        meta_instance_id=meta_instance_id or MetaInstanceId(),
+        model_id=ModelId(model_id),
+        min_nodes=min_nodes,
+        node_ids=node_ids,
+        consecutive_failures=consecutive_failures,
+        last_failure_error=last_failure_error,
+        placement_error=placement_error,
+    )
+
+
+def _instance(
+    model_id: str = "test-org/test-model",
+    node_ids: list[str] | None = None,
+    instance_id: InstanceId | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> tuple[InstanceId, MlxRingInstance]:
+    iid = instance_id or InstanceId()
+    nodes = node_ids or ["node-a"]
+    n = len(nodes)
+    mc = _model_card(model_id)
+    ephemeral_port = 50000
+    node_to_runner = {NodeId(nd): RunnerId() for nd in nodes}
+    runner_to_shard = {
+        runner_id: PipelineShardMetadata(
+            model_card=mc,
+            device_rank=i,
+            world_size=n,
+            start_layer=0,
+            end_layer=mc.n_layers,
+            n_layers=mc.n_layers,
+        )
+        for i, runner_id in enumerate(node_to_runner.values())
+    }
+    hosts_by_node: dict[NodeId, list[Host]] = {}
+    for r, node_str in enumerate(nodes):
+        hosts: list[Host] = []
+        for idx in range(n):
+            if idx == r:
+                hosts.append(Host(ip="0.0.0.0", port=ephemeral_port))
+            elif n > 1 and idx in ((r - 1) % n, (r + 1) % n):
+                hosts.append(Host(ip=f"10.0.0.{idx + 1}", port=ephemeral_port))
+            else:
+                hosts.append(Host(ip="198.51.100.1", port=0))
+        hosts_by_node[NodeId(node_str)] = hosts
+    return iid, MlxRingInstance(
+        instance_id=iid,
+        shard_assignments=ShardAssignments(
+            model_id=ModelId(model_id),
+            runner_to_shard=runner_to_shard,
+            node_to_runner=node_to_runner,
+        ),
+        hosts_by_node=hosts_by_node,
+        ephemeral_port=ephemeral_port,
+        meta_instance_id=meta_instance_id,
+    )
+
+
+# =============================================================================
+# 1. MetaInstance lifecycle edge cases
+# =============================================================================
+
+
+def test_meta_instance_model_is_frozen():
+    """MetaInstance should be immutable (frozen model)."""
+    meta = _meta_instance()
+    try:
+        meta.model_id = ModelId("something-else")
+        raise AssertionError("Should have raised")
+    except Exception:
+        pass  # Expected — frozen model
+
+
+def test_meta_instance_created_then_deleted_roundtrip():
+    """Create and delete a MetaInstance through apply — state should be clean."""
+    state = State()
+    meta = _meta_instance()
+    state = apply(
+        state, IndexedEvent(idx=0, event=MetaInstanceCreated(meta_instance=meta))
+    )
+    assert meta.meta_instance_id in state.meta_instances
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=1, event=MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+        ),
+    )
+    assert meta.meta_instance_id not in state.meta_instances
+    assert len(state.meta_instances) == 0
+
+
+def test_delete_nonexistent_meta_instance_is_safe():
+    """Deleting a MetaInstance that doesn't exist should not crash."""
+    state = State()
+    event = MetaInstanceDeleted(meta_instance_id=MetaInstanceId("nonexistent"))
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+def test_placement_failed_for_nonexistent_meta_instance_is_safe():
+    """MetaInstancePlacementFailed for unknown ID should not crash."""
+    state = State()
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=MetaInstanceId("nonexistent"),
+        reason="test",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+def test_multiple_meta_instances_for_same_model():
+    """Multiple MetaInstances for the same model are tracked independently."""
+    state = State()
+    meta_a = _meta_instance("test-org/model-x")
+    meta_b = _meta_instance("test-org/model-x")
+    state = apply(
+        state, IndexedEvent(idx=0, event=MetaInstanceCreated(meta_instance=meta_a))
+    )
+    state = apply(
+        state, IndexedEvent(idx=1, event=MetaInstanceCreated(meta_instance=meta_b))
+    )
+    assert len(state.meta_instances) == 2
+    assert meta_a.meta_instance_id in state.meta_instances
+    assert meta_b.meta_instance_id in state.meta_instances
+
+
+# =============================================================================
+# 2. Retry logic edge cases
+# =============================================================================
+
+
+def test_retry_counter_resets_on_successful_instance_creation():
+    """When a new instance is created for a meta-instance, failures should reset."""
+    meta = _meta_instance(consecutive_failures=2, last_failure_error="old")
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    state = apply(state, IndexedEvent(idx=0, event=InstanceCreated(instance=inst)))
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    # last_failure_error is preserved (for UI display)
+    assert mi.last_failure_error == "old"
+
+
+async def test_retry_count_increments_through_full_cycle():
+    """Walk through MAX_INSTANCE_RETRIES worth of retries, then verify delete."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    topology = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        topology=topology,
+    )
+
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    for idx, i in enumerate(range(MAX_INSTANCE_RETRIES)):
+        # Simulate runners failing
+        state_with_runners = state.model_copy(
+            update={"runners": {runner_ids[0]: RunnerFailed(error_message=f"fail-{i}")}}
+        )
+        reconciler = InstanceHealthReconciler()
+        events = await reconciler.reconcile(state_with_runners)
+        assert len(events) == 1
+        assert isinstance(events[0], InstanceRetrying), f"iteration {i}"
+        state = apply(state, IndexedEvent(idx=idx, event=events[0]))
+
+    # After MAX_INSTANCE_RETRIES retries, failure counter should be at max
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == MAX_INSTANCE_RETRIES
+
+    # Next failure should result in deletion
+    state_with_runners = state.model_copy(
+        update={"runners": {runner_ids[0]: RunnerFailed(error_message="final")}}
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state_with_runners)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_respects_exact_limit():
+    """At exactly MAX_INSTANCE_RETRIES, reconciler should delete, not retry."""
+    meta = _meta_instance(consecutive_failures=MAX_INSTANCE_RETRIES)
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_at_limit_minus_one_retries():
+    """At MAX_INSTANCE_RETRIES - 1, reconciler should still retry."""
+    meta = _meta_instance(consecutive_failures=MAX_INSTANCE_RETRIES - 1)
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+
+
+# =============================================================================
+# 3. Error handling edge cases
+# =============================================================================
+
+
+def test_runners_failed_with_empty_error_message():
+    """RunnerFailed with empty error_message should still report as failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerFailed(error_message="")
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    # Empty error message means we get the fallback
+    assert error == "Runner failed"
+
+
+def test_runners_failed_with_none_error_message():
+    """RunnerFailed with None error_message should still report as failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerFailed(error_message=None)
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error == "Runner failed"
+
+
+def test_runners_failed_collects_all_error_messages():
+    """With multiple failed runners, all error messages should be collected."""
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM on GPU 0"),
+        runner_ids[1]: RunnerFailed(error_message="OOM on GPU 1"),
+        runner_ids[2]: RunnerFailed(error_message="OOM on GPU 2"),
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "OOM on GPU 0" in error
+    assert "OOM on GPU 1" in error
+    assert "OOM on GPU 2" in error
+
+
+def test_runners_failed_includes_friendly_name():
+    """Error messages should include node friendly names when available."""
+    _, inst = _instance(node_ids=["node-a"])
+    node_id = NodeId("node-a")
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {runner_ids[0]: RunnerFailed(error_message="OOM")}
+    identities = {node_id: NodeIdentity(friendly_name="My Mac Studio")}
+    is_failed, error = instance_runners_failed(inst, runners, identities)
+    assert is_failed is True
+    assert error is not None
+    assert "My Mac Studio" in error
+
+
+def test_instance_retrying_for_missing_instance_is_safe():
+    """InstanceRetrying for an instance not in state should not crash.
+
+    NOTE: When the instance is missing, the handler returns early WITHOUT
+    incrementing the MetaInstance failure counter. This means stale retry
+    events for already-deleted instances are silently dropped. This is
+    acceptable since the InstanceDeleted handler already increments failures.
+    """
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = InstanceRetrying(
+        instance_id=InstanceId("nonexistent"),
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    # Does not crash, but failure count is NOT incremented (early return)
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+
+
+# =============================================================================
+# 4. Backward compatibility
+# =============================================================================
+
+
+def test_instance_without_meta_instance_id_works():
+    """Instances created without meta_instance_id should still function normally."""
+    _, inst = _instance(node_ids=["node-a"])
+    assert inst.meta_instance_id is None
+    topology = _topology("node-a")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_instance_deleted_without_meta_does_not_affect_meta_instances():
+    """Deleting an instance without meta_instance_id should not affect meta_instances."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"])  # no meta_instance_id
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="crash")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0  # unchanged
+
+
+def test_satisfies_ignores_meta_instance_id_binding():
+    """instance_satisfies_meta_instance checks constraints only, not binding."""
+    meta = _meta_instance()
+    _, inst = _instance(node_ids=["node-a"])  # no meta_instance_id set
+    # Should match on constraints (model, min_nodes) regardless of binding
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+def test_find_unsatisfied_uses_binding_not_constraints():
+    """find_unsatisfied checks meta_instance_id binding, not just constraint matching."""
+    meta = _meta_instance()
+    # Instance matches constraints but is NOT bound to this meta_instance
+    iid, inst = _instance(node_ids=["node-a"])
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta}, {iid: inst}, topology
+    )
+    # Should be unsatisfied because instance.meta_instance_id != meta.meta_instance_id
+    assert list(result) == [meta]
+
+
+# =============================================================================
+# 5. Concurrent / multi-instance scenarios
+# =============================================================================
+
+
+async def test_health_reconciler_handles_multiple_failing_instances():
+    """Multiple instances failing simultaneously should each get their own event."""
+    meta_a = _meta_instance()
+    meta_b = _meta_instance()
+    iid_a, inst_a = _instance(
+        node_ids=["node-a"], meta_instance_id=meta_a.meta_instance_id
+    )
+    iid_b, inst_b = _instance(
+        node_ids=["node-b"], meta_instance_id=meta_b.meta_instance_id
+    )
+    runner_ids_a = list(inst_a.shard_assignments.node_to_runner.values())
+    runner_ids_b = list(inst_b.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        instances={iid_a: inst_a, iid_b: inst_b},
+        runners={
+            runner_ids_a[0]: RunnerFailed(error_message="OOM"),
+            runner_ids_b[0]: RunnerFailed(error_message="OOM"),
+        },
+        topology=_topology("node-a", "node-b"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 2
+    # Both should be InstanceRetrying since failures < MAX
+    assert all(isinstance(e, InstanceRetrying) for e in events)
+    instance_ids = {e.instance_id for e in events}  # type: ignore[union-attr]
+    assert instance_ids == {iid_a, iid_b}
+
+
+async def test_health_reconciler_mixed_healthy_and_failing():
+    """Only failing instances should produce events; healthy ones should not."""
+    meta_healthy = _meta_instance()
+    meta_failing = _meta_instance()
+    iid_h, inst_h = _instance(
+        node_ids=["node-a"], meta_instance_id=meta_healthy.meta_instance_id
+    )
+    iid_f, inst_f = _instance(
+        node_ids=["node-b"], meta_instance_id=meta_failing.meta_instance_id
+    )
+    runner_ids_h = list(inst_h.shard_assignments.node_to_runner.values())
+    runner_ids_f = list(inst_f.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={
+            meta_healthy.meta_instance_id: meta_healthy,
+            meta_failing.meta_instance_id: meta_failing,
+        },
+        instances={iid_h: inst_h, iid_f: inst_f},
+        runners={
+            runner_ids_h[0]: RunnerReady(),
+            runner_ids_f[0]: RunnerFailed(error_message="crash"),
+        },
+        topology=_topology("node-a", "node-b"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+    assert events[0].instance_id == iid_f
+
+
+async def test_meta_instance_reconciler_empty_state():
+    """MetaInstanceReconciler with no meta_instances should produce no events."""
+    state = State()
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 0
+
+
+# =============================================================================
+# 6. Placement error tracking
+# =============================================================================
+
+
+def test_placement_failed_sets_error():
+    """MetaInstancePlacementFailed should set placement_error on the MetaInstance."""
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=meta.meta_instance_id,
+        reason="Not enough memory",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.placement_error == "Not enough memory"
+
+
+def test_instance_created_clears_placement_error():
+    """InstanceCreated should clear placement_error on the MetaInstance."""
+    meta = _meta_instance(placement_error="Not enough memory")
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    state = apply(state, IndexedEvent(idx=0, event=InstanceCreated(instance=inst)))
+    mi = state.meta_instances[meta.meta_instance_id]
+    assert mi.placement_error is None
+
+
+def test_placement_error_does_not_increment_failures():
+    """Placement failures should only set placement_error, not increment consecutive_failures."""
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstancePlacementFailed(
+        meta_instance_id=meta.meta_instance_id,
+        reason="No resources",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    assert mi.placement_error == "No resources"
+
+
+# =============================================================================
+# 7. State serialization roundtrip
+# =============================================================================
+
+
+def test_state_with_meta_instances_serializes():
+    """State with meta_instances should serialize and deserialize correctly."""
+    meta = _meta_instance(consecutive_failures=2, last_failure_error="test")
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    json_str = state.model_dump_json()
+    restored = State.model_validate_json(json_str)
+    assert meta.meta_instance_id in restored.meta_instances
+    mi = restored.meta_instances[meta.meta_instance_id]
+    assert mi.model_id == meta.model_id
+    assert mi.consecutive_failures == 2
+    assert mi.last_failure_error == "test"
+    assert iid in restored.instances
+    assert restored.instances[iid].meta_instance_id == meta.meta_instance_id
+
+
+# =============================================================================
+# 8. MetaInstanceReconciler error handling
+# =============================================================================
+
+
+async def test_meta_instance_reconciler_model_load_error_emits_placement_failed(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """When ModelCard.load raises, reconciler emits MetaInstancePlacementFailed."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta = _meta_instance()
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        topology=topo,
+    )
+
+    async def _failing_load(_model_id: ModelId) -> ModelCard:
+        raise RuntimeError("Network error")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_failing_load)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    placement_failed = [e for e in events if isinstance(e, MetaInstancePlacementFailed)]
+    assert len(placement_failed) == 1
+    assert "Failed to load model card" in placement_failed[0].reason
+    assert meta.meta_instance_id == placement_failed[0].meta_instance_id
+
+
+async def test_meta_instance_reconciler_model_load_error_skips_dedup(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """When ModelCard.load error matches existing placement_error, no duplicate event."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta = _meta_instance(placement_error="Failed to load model card: Network error")
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        topology=topo,
+    )
+
+    async def _failing_load(_model_id: ModelId) -> ModelCard:
+        raise RuntimeError("Network error")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_failing_load)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    # Error matches existing placement_error, so no duplicate event emitted
+    assert len(events) == 0
+
+
+async def test_meta_instance_reconciler_continues_after_error(
+    monkeypatch: "pytest.MonkeyPatch",
+):
+    """Reconciler should continue to next meta-instance after one fails to load."""
+    import exo.master.process_managers.meta_instance as mi_mod
+
+    meta_a = _meta_instance(model_id="org/model-a")
+    meta_b = _meta_instance(model_id="org/model-b")
+    topo = _topology("node-a")
+    state = State(
+        meta_instances={
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        topology=topo,
+    )
+
+    call_count = 0
+
+    async def _load_second_fails(model_id: ModelId) -> ModelCard:
+        nonlocal call_count
+        call_count += 1
+        raise RuntimeError(f"Cannot load {model_id}")
+
+    monkeypatch.setattr(
+        mi_mod, "ModelCard", type("MC", (), {"load": staticmethod(_load_second_fails)})
+    )
+
+    reconciler = MetaInstanceReconciler()
+    events = await reconciler.reconcile(state)
+
+    # Both meta-instances should have been attempted (not short-circuited)
+    assert call_count == 2
+    # Both should have placement failed events
+    placement_failed = [e for e in events if isinstance(e, MetaInstancePlacementFailed)]
+    assert len(placement_failed) == 2
+
+
+# =============================================================================
+# 8. Cascade delete with task cancellation
+# =============================================================================
+
+
+def test_cascade_delete_cancels_active_tasks():
+    """Deleting a MetaInstance should cancel tasks on backing instances.
+
+    Regression test: previously, cascade-deleting backing instances via
+    DeleteMetaInstance did not emit TaskStatusUpdated(Cancelled) for active
+    tasks, leaving orphaned task references in state.
+    """
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    task_id = TaskId()
+    task = LoadModel(task_id=task_id, instance_id=iid, task_status=TaskStatus.Running)
+
+    # Build state with meta-instance, backing instance, and active task
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        tasks={task_id: task},
+        topology=_topology("node-a"),
+    )
+
+    # Simulate the cascade-delete event sequence produced by main.py:
+    # 1. MetaInstanceDeleted
+    # 2. TaskStatusUpdated(Cancelled) for active tasks
+    # 3. InstanceDeleted
+    idx = 0
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=idx,
+            event=MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id),
+        ),
+    )
+    idx += 1
+    state = apply(
+        state,
+        IndexedEvent(
+            idx=idx,
+            event=TaskStatusUpdated(task_id=task_id, task_status=TaskStatus.Cancelled),
+        ),
+    )
+    idx += 1
+    state = apply(
+        state,
+        IndexedEvent(idx=idx, event=InstanceDeleted(instance_id=iid)),
+    )
+
+    # Verify everything is cleaned up
+    assert len(state.meta_instances) == 0
+    assert len(state.instances) == 0
+    assert state.tasks[task_id].task_status == TaskStatus.Cancelled
+
+
+def test_cascade_delete_skips_completed_tasks():
+    """Cascade delete should only cancel Pending/Running tasks, not completed ones."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+
+    running_task_id = TaskId()
+    completed_task_id = TaskId()
+    running_task = LoadModel(
+        task_id=running_task_id, instance_id=iid, task_status=TaskStatus.Running
+    )
+    completed_task = LoadModel(
+        task_id=completed_task_id, instance_id=iid, task_status=TaskStatus.Complete
+    )
+
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        tasks={running_task_id: running_task, completed_task_id: completed_task},
+        topology=_topology("node-a"),
+    )
+
+    # Only the running task should be cancelled — we verify the logic pattern
+    # by checking which tasks are Pending or Running
+    active_tasks = [
+        t
+        for t in state.tasks.values()
+        if t.instance_id == iid
+        and t.task_status in (TaskStatus.Pending, TaskStatus.Running)
+    ]
+    assert len(active_tasks) == 1
+    assert active_tasks[0].task_id == running_task_id

src/exo/master/tests/test_placement_utils.py

@@ -3,10 +3,10 @@ import pytest
 from exo.master.placement_utils import (
     allocate_layers_proportionally,
     filter_cycles_by_memory,
+    get_largest_cycles,
     get_mlx_jaccl_coordinators,
     get_shard_assignments,
     get_shard_assignments_for_pipeline_parallel,
-    get_smallest_cycles,
 )
 from exo.master.tests.conftest import (
     create_node_memory,
@@ -143,7 +143,7 @@ def test_filter_multiple_cycles_by_memory():
     }
 
 
-def test_get_smallest_cycles():
+def test_get_largest_cycles():
     # arrange
     node_a_id = NodeId()
     node_b_id = NodeId()
@@ -175,12 +175,12 @@ def test_get_smallest_cycles():
     cycles = [c for c in topology.get_cycles() if len(c) != 1]  # ignore singletons
 
     # act
-    smallest_cycles = get_smallest_cycles(cycles)
+    largest_cycles = get_largest_cycles(cycles)
 
     # assert
-    assert len(smallest_cycles) == 1
-    assert len(smallest_cycles[0]) == 2
-    assert set(n for n in smallest_cycles[0]) == {node_a_id, node_b_id}
+    assert len(largest_cycles) == 1
+    assert len(largest_cycles[0]) == 3
+    assert set(n for n in largest_cycles[0]) == {node_a_id, node_b_id, node_c_id}
 
 
 @pytest.mark.parametrize(

src/exo/master/tests/test_reconcile.py

@@ -0,0 +1,742 @@
+from exo.master.process_managers.instance_health import InstanceHealthReconciler
+from exo.master.reconcile import (
+    find_unsatisfied_meta_instances,
+    instance_connections_healthy,
+    instance_runners_failed,
+    instance_satisfies_meta_instance,
+)
+from exo.shared.apply import apply
+from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
+from exo.shared.topology import Topology
+from exo.shared.types.common import Host, MetaInstanceId, NodeId
+from exo.shared.types.events import (
+    IndexedEvent,
+    InstanceCreated,
+    InstanceDeleted,
+    InstanceRetrying,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+)
+from exo.shared.types.memory import Memory
+from exo.shared.types.meta_instance import MetaInstance
+from exo.shared.types.multiaddr import Multiaddr
+from exo.shared.types.state import State
+from exo.shared.types.topology import Connection, SocketConnection
+from exo.shared.types.worker.instances import (
+    InstanceId,
+    MlxRingInstance,
+)
+from exo.shared.types.worker.runners import (
+    RunnerFailed,
+    RunnerId,
+    RunnerLoading,
+    RunnerReady,
+    RunnerShutdown,
+    ShardAssignments,
+)
+from exo.shared.types.worker.shards import PipelineShardMetadata
+
+
+def _model_card(model_id: str = "test-org/test-model") -> ModelCard:
+    return ModelCard(
+        model_id=ModelId(model_id),
+        storage_size=Memory.from_kb(1000),
+        n_layers=10,
+        hidden_size=30,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    )
+
+
+def _topology(*node_ids: str, connect: bool = True) -> Topology:
+    """Build a topology with nodes connected in a bidirectional ring with unique IPs.
+
+    Node at index ``i`` gets IP ``10.0.0.{i+1}``. Edges go in both directions
+    between consecutive nodes (including wrap-around).
+    """
+    t = Topology()
+    nodes = [NodeId(n) for n in node_ids]
+    for n in nodes:
+        t.add_node(n)
+    if connect and len(nodes) > 1:
+        for i in range(len(nodes)):
+            j = (i + 1) % len(nodes)
+            t.add_connection(
+                Connection(
+                    source=nodes[i],
+                    sink=nodes[j],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{j + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+            t.add_connection(
+                Connection(
+                    source=nodes[j],
+                    sink=nodes[i],
+                    edge=SocketConnection(
+                        sink_multiaddr=Multiaddr(
+                            address=f"/ip4/10.0.0.{i + 1}/tcp/50000"
+                        )
+                    ),
+                )
+            )
+    return t
+
+
+def _meta_instance(
+    model_id: str = "test-org/test-model",
+    *,
+    min_nodes: int = 1,
+    node_ids: list[NodeId] | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> MetaInstance:
+    return MetaInstance(
+        meta_instance_id=meta_instance_id or MetaInstanceId(),
+        model_id=ModelId(model_id),
+        min_nodes=min_nodes,
+        node_ids=node_ids,
+    )
+
+
+def _instance(
+    model_id: str = "test-org/test-model",
+    node_ids: list[str] | None = None,
+    instance_id: InstanceId | None = None,
+    meta_instance_id: MetaInstanceId | None = None,
+) -> tuple[InstanceId, MlxRingInstance]:
+    """Create a test instance with hosts_by_node matching ``_topology()`` IPs."""
+    iid = instance_id or InstanceId()
+    nodes = node_ids or ["node-a"]
+    n = len(nodes)
+    mc = _model_card(model_id)
+    ephemeral_port = 50000
+    node_to_runner = {NodeId(nd): RunnerId() for nd in nodes}
+    runner_to_shard = {
+        runner_id: PipelineShardMetadata(
+            model_card=mc,
+            device_rank=i,
+            world_size=n,
+            start_layer=0,
+            end_layer=mc.n_layers,
+            n_layers=mc.n_layers,
+        )
+        for i, runner_id in enumerate(node_to_runner.values())
+    }
+    # Build hosts_by_node with IPs matching _topology() convention:
+    # node at index idx has IP 10.0.0.{idx+1}
+    hosts_by_node: dict[NodeId, list[Host]] = {}
+    for r, node_str in enumerate(nodes):
+        hosts: list[Host] = []
+        for idx in range(n):
+            if idx == r:
+                hosts.append(Host(ip="0.0.0.0", port=ephemeral_port))
+            elif n > 1 and idx in ((r - 1) % n, (r + 1) % n):
+                hosts.append(Host(ip=f"10.0.0.{idx + 1}", port=ephemeral_port))
+            else:
+                hosts.append(Host(ip="198.51.100.1", port=0))
+        hosts_by_node[NodeId(node_str)] = hosts
+    return iid, MlxRingInstance(
+        instance_id=iid,
+        shard_assignments=ShardAssignments(
+            model_id=ModelId(model_id),
+            runner_to_shard=runner_to_shard,
+            node_to_runner=node_to_runner,
+        ),
+        hosts_by_node=hosts_by_node,
+        ephemeral_port=ephemeral_port,
+        meta_instance_id=meta_instance_id,
+    )
+
+
+# --- instance_satisfies_meta_instance (pure constraint matching) ---
+
+
+def test_satisfies_matching_model():
+    meta = _meta_instance()
+    _, inst = _instance(node_ids=["node-a"])
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+def test_not_satisfies_wrong_model():
+    meta = _meta_instance("test-org/model-a")
+    _, inst = _instance("test-org/model-b")
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_not_satisfies_missing_required_node():
+    meta = _meta_instance(node_ids=[NodeId("node-c")])
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_not_satisfies_fewer_than_min_nodes():
+    meta = _meta_instance(min_nodes=3)
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    assert instance_satisfies_meta_instance(meta, inst) is False
+
+
+def test_satisfies_with_node_ids_specified():
+    meta = _meta_instance(node_ids=[NodeId("node-a"), NodeId("node-b")], min_nodes=2)
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    assert instance_satisfies_meta_instance(meta, inst) is True
+
+
+# --- instance_connections_healthy ---
+
+
+def test_healthy_single_node_present():
+    _, inst = _instance(node_ids=["node-a"])
+    topology = _topology("node-a")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_single_node_missing():
+    _, inst = _instance(node_ids=["node-a"])
+    topology = Topology()  # empty
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_two_node_ring():
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_two_node_edge_removed():
+    """Nodes present but edge removed — ring broken."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b", connect=False)
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_unhealthy_two_node_ip_changed():
+    """Edge exists but with a different IP than instance was configured with."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    # Build topology with different IPs than _instance() expects
+    topology = Topology()
+    topology.add_node(NodeId("node-a"))
+    topology.add_node(NodeId("node-b"))
+    topology.add_connection(
+        Connection(
+            source=NodeId("node-a"),
+            sink=NodeId("node-b"),
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/192.168.99.99/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=NodeId("node-b"),
+            sink=NodeId("node-a"),
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/192.168.99.98/tcp/50000")
+            ),
+        )
+    )
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_three_node_ring():
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    topology = _topology("node-a", "node-b", "node-c")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+def test_unhealthy_three_node_one_edge_removed():
+    """Remove one edge from a three-node ring — instance unhealthy."""
+    _, inst = _instance(node_ids=["node-a", "node-b", "node-c"])
+    # Build topology with one direction of one edge missing
+    topology = Topology()
+    nodes = [NodeId("node-a"), NodeId("node-b"), NodeId("node-c")]
+    for n in nodes:
+        topology.add_node(n)
+    # Add all edges except node-a → node-b
+    topology.add_connection(
+        Connection(
+            source=nodes[1],
+            sink=nodes[0],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[1],
+            sink=nodes[2],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.3/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[2],
+            sink=nodes[1],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.2/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[2],
+            sink=nodes[0],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/50000")
+            ),
+        )
+    )
+    topology.add_connection(
+        Connection(
+            source=nodes[0],
+            sink=nodes[2],
+            edge=SocketConnection(
+                sink_multiaddr=Multiaddr(address="/ip4/10.0.0.3/tcp/50000")
+            ),
+        )
+    )
+    # Missing: node-a → node-b (ip 10.0.0.2)
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_unhealthy_node_missing_from_topology():
+    """Instance has a node that's not in the topology at all."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a")  # node-b not present
+    assert instance_connections_healthy(inst, topology) is False
+
+
+def test_healthy_extra_nodes_in_topology():
+    """Extra nodes in topology don't affect instance health."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    topology = _topology("node-a", "node-b", "node-c")
+    assert instance_connections_healthy(inst, topology) is True
+
+
+# --- find_unsatisfied_meta_instances ---
+
+
+def test_unsatisfied_no_meta_instances():
+    result = find_unsatisfied_meta_instances({}, {}, Topology())
+    assert list(result) == []
+
+
+def test_unsatisfied_one_satisfied():
+    meta = _meta_instance()
+    id_a, inst_a = _instance(meta_instance_id=meta.meta_instance_id)
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == []
+
+
+def test_unsatisfied_one_not_satisfied():
+    meta = _meta_instance("test-org/model-x")
+    id_a, inst_a = _instance("test-org/model-y")
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta}, {id_a: inst_a}, topology
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_mix():
+    meta_satisfied = _meta_instance("test-org/model-a")
+    meta_unsatisfied = _meta_instance("test-org/model-b")
+    id_a, inst_a = _instance(
+        "test-org/model-a", meta_instance_id=meta_satisfied.meta_instance_id
+    )
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {
+            meta_satisfied.meta_instance_id: meta_satisfied,
+            meta_unsatisfied.meta_instance_id: meta_unsatisfied,
+        },
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta_unsatisfied]
+
+
+def test_unsatisfied_node_disconnect():
+    meta = _meta_instance()
+    id_a, inst_a = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    topology = _topology("node-a")  # node-b disconnected
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_edge_break():
+    """Instance exists but its connections broke — meta-instance becomes unsatisfied."""
+    meta = _meta_instance()
+    id_a, inst_a = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    topology = _topology("node-a", "node-b", connect=False)  # nodes present, no edges
+    result = find_unsatisfied_meta_instances(
+        {meta.meta_instance_id: meta},
+        {id_a: inst_a},
+        topology,
+    )
+    assert list(result) == [meta]
+
+
+def test_unsatisfied_idempotent():
+    meta = _meta_instance("test-org/model-x")
+    topology = _topology("node-a")
+    meta_instances = {meta.meta_instance_id: meta}
+    instances: dict[InstanceId, MlxRingInstance] = {}
+    result_1 = list(
+        find_unsatisfied_meta_instances(meta_instances, instances, topology)
+    )
+    result_2 = list(
+        find_unsatisfied_meta_instances(meta_instances, instances, topology)
+    )
+    assert result_1 == result_2
+
+
+def test_unsatisfied_exclusive_binding():
+    """Two MetaInstances for the same model: one is bound via meta_instance_id, the other is unsatisfied."""
+    meta_a = _meta_instance("test-org/model-x")
+    meta_b = _meta_instance("test-org/model-x")
+    id_inst, inst = _instance(
+        "test-org/model-x", meta_instance_id=meta_a.meta_instance_id
+    )
+    topology = _topology("node-a")
+    result = find_unsatisfied_meta_instances(
+        {
+            meta_a.meta_instance_id: meta_a,
+            meta_b.meta_instance_id: meta_b,
+        },
+        {id_inst: inst},
+        topology,
+    )
+    assert list(result) == [meta_b]
+
+
+# --- apply handlers ---
+
+
+def test_apply_meta_instance_created():
+    state = State()
+    meta = _meta_instance()
+    event = MetaInstanceCreated(meta_instance=meta)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id in new_state.meta_instances
+    assert new_state.meta_instances[meta.meta_instance_id] == meta
+
+
+def test_apply_meta_instance_deleted():
+    meta = _meta_instance()
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id not in new_state.meta_instances
+
+
+def test_apply_meta_instance_deleted_clears_failure_info():
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 2, "last_failure_error": "OOM"}
+    )
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = MetaInstanceDeleted(meta_instance_id=meta.meta_instance_id)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert meta.meta_instance_id not in new_state.meta_instances
+
+
+# --- instance_runners_failed ---
+
+
+def test_runners_failed_all_failed():
+    """All runners in RunnerFailed -> instance is failed."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runners = {
+        rid: RunnerFailed(error_message="OOM")
+        for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "OOM" in error
+
+
+def test_runners_failed_mixed_failed_shutdown():
+    """One Failed + one Shutdown = failed."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="crash"),
+        runner_ids[1]: RunnerShutdown(),
+    }
+    is_failed, error = instance_runners_failed(inst, runners, {})
+    assert is_failed is True
+    assert error is not None
+    assert "crash" in error
+
+
+def test_runners_not_failed_all_shutdown():
+    """All Shutdown (graceful) = not a failure."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerShutdown() for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_still_active():
+    """Some runners still active = not failed yet."""
+    _, inst = _instance(node_ids=["node-a", "node-b"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM"),
+        runner_ids[1]: RunnerLoading(),
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_no_status():
+    """Runner not yet reported = not failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    is_failed, _ = instance_runners_failed(inst, {}, {})
+    assert is_failed is False
+
+
+def test_runners_not_failed_healthy():
+    """Runners in Ready state = not failed."""
+    _, inst = _instance(node_ids=["node-a"])
+    runners = {
+        rid: RunnerReady() for rid in inst.shard_assignments.node_to_runner.values()
+    }
+    is_failed, _ = instance_runners_failed(inst, runners, {})
+    assert is_failed is False
+
+
+# --- failure tracking in apply_instance_deleted ---
+
+
+def test_apply_instance_deleted_tracks_failure():
+    """InstanceDeleted with failure_error increments meta instance failure count."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="Runner OOM")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 1
+    assert mi.last_failure_error == "Runner OOM"
+
+
+def test_apply_instance_deleted_increments_failure():
+    """Subsequent failures increment the counter."""
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 2, "last_failure_error": "previous error"}
+    )
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid, failure_error="new error")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 3
+    assert mi.last_failure_error == "new error"
+
+
+def test_apply_instance_deleted_no_failure_no_tracking():
+    """InstanceDeleted without failure_error does not track."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceDeleted(instance_id=iid)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+
+
+def test_apply_instance_deleted_orphan_no_tracking():
+    """InstanceDeleted for orphan instance (no meta_instance_id) does not track."""
+    iid, inst = _instance(node_ids=["node-a"])
+    state = State(instances={iid: inst})
+    event = InstanceDeleted(instance_id=iid, failure_error="crash")
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert len(new_state.meta_instances) == 0
+
+
+# --- InstanceRetrying ---
+
+
+def test_apply_instance_retrying_removes_runners():
+    """InstanceRetrying removes the instance's runners from state but keeps the instance."""
+    meta = _meta_instance()
+    iid, inst = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    runners = {
+        runner_ids[0]: RunnerFailed(error_message="OOM"),
+        runner_ids[1]: RunnerShutdown(),
+    }
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners=runners,
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="OOM",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    # Instance still exists
+    assert iid in new_state.instances
+    # Runners removed
+    assert runner_ids[0] not in new_state.runners
+    assert runner_ids[1] not in new_state.runners
+
+
+def test_apply_instance_retrying_increments_failure():
+    """InstanceRetrying increments consecutive_failures on the MetaInstance."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 1
+    assert mi.last_failure_error == "crash"
+
+
+def test_apply_instance_retrying_skips_missing_runners():
+    """InstanceRetrying doesn't assert if runners haven't reported yet."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    # No runners in state at all
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+    )
+    event = InstanceRetrying(
+        instance_id=iid,
+        meta_instance_id=meta.meta_instance_id,
+        failure_error="crash",
+    )
+    # Should not raise
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    assert iid in new_state.instances
+
+
+def test_apply_instance_created_resets_failure_counter():
+    """InstanceCreated resets consecutive_failures but preserves last_failure_error."""
+    meta = _meta_instance().model_copy(
+        update={"consecutive_failures": 3, "last_failure_error": "old error"}
+    )
+    _, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    state = State(meta_instances={meta.meta_instance_id: meta})
+    event = InstanceCreated(instance=inst)
+    new_state = apply(state, IndexedEvent(idx=0, event=event))
+    mi = new_state.meta_instances[meta.meta_instance_id]
+    assert mi.consecutive_failures == 0
+    assert mi.last_failure_error == "old error"
+    assert mi.placement_error is None
+
+
+# --- InstanceHealthReconciler retry-vs-delete ---
+
+
+async def test_health_reconciler_retries_when_under_limit():
+    """InstanceHealthReconciler emits InstanceRetrying when consecutive_failures < 3."""
+    meta = _meta_instance()
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceRetrying)
+    assert events[0].instance_id == iid
+    assert events[0].meta_instance_id == meta.meta_instance_id
+
+
+async def test_health_reconciler_deletes_when_limit_reached():
+    """InstanceHealthReconciler emits InstanceDeleted when consecutive_failures >= 3."""
+    meta = _meta_instance().model_copy(update={"consecutive_failures": 3})
+    iid, inst = _instance(node_ids=["node-a"], meta_instance_id=meta.meta_instance_id)
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="OOM")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_deletes_without_meta_instance():
+    """Instances without a MetaInstance are deleted immediately on runner failure."""
+    iid, inst = _instance(node_ids=["node-a"])
+    runner_ids = list(inst.shard_assignments.node_to_runner.values())
+    state = State(
+        instances={iid: inst},
+        runners={runner_ids[0]: RunnerFailed(error_message="crash")},
+        topology=_topology("node-a"),
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+async def test_health_reconciler_network_failure_always_deletes():
+    """Network failure always triggers InstanceDeleted regardless of retry count."""
+    meta = _meta_instance()
+    iid, inst = _instance(
+        node_ids=["node-a", "node-b"], meta_instance_id=meta.meta_instance_id
+    )
+    state = State(
+        meta_instances={meta.meta_instance_id: meta},
+        instances={iid: inst},
+        topology=_topology("node-a"),  # node-b missing
+    )
+    reconciler = InstanceHealthReconciler()
+    events = await reconciler.reconcile(state)
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+    assert events[0].failure_error == "Network connection lost"

src/exo/shared/apply.py

@@ -4,7 +4,7 @@ from datetime import datetime
 
 from loguru import logger
 
-from exo.shared.types.common import NodeId
+from exo.shared.types.common import MetaInstanceId, NodeId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -12,6 +12,12 @@ from exo.shared.types.events import (
     InputChunkReceived,
     InstanceCreated,
     InstanceDeleted,
+    InstanceRetrying,
+    JacclSideChannelData,
+    JacclSideChannelGathered,
+    MetaInstanceCreated,
+    MetaInstanceDeleted,
+    MetaInstancePlacementFailed,
     NodeDownloadProgress,
     NodeGatheredInfo,
     NodeTimedOut,
@@ -28,6 +34,7 @@ from exo.shared.types.events import (
     TracesCollected,
     TracesMerged,
 )
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.profiling import (
     NodeIdentity,
     NodeNetworkInfo,
@@ -66,12 +73,22 @@ def event_apply(event: Event, state: State) -> State:
             | InputChunkReceived()
             | TracesCollected()
             | TracesMerged()
+            | JacclSideChannelData()
+            | JacclSideChannelGathered()
         ):  # Pass-through events that don't modify state
             return state
         case InstanceCreated():
             return apply_instance_created(event, state)
         case InstanceDeleted():
             return apply_instance_deleted(event, state)
+        case InstanceRetrying():
+            return apply_instance_retrying(event, state)
+        case MetaInstanceCreated():
+            return apply_meta_instance_created(event, state)
+        case MetaInstanceDeleted():
+            return apply_meta_instance_deleted(event, state)
+        case MetaInstancePlacementFailed():
+            return apply_meta_instance_placement_failed(event, state)
         case NodeTimedOut():
             return apply_node_timed_out(event, state)
         case NodeDownloadProgress():
@@ -174,20 +191,123 @@ def apply_task_failed(event: TaskFailed, state: State) -> State:
     return state.model_copy(update={"tasks": new_tasks})
 
 
+def _update_meta_instance(
+    state: State, mid: MetaInstanceId, **fields: object
+) -> Mapping[MetaInstanceId, MetaInstance]:
+    mi = state.meta_instances[mid]
+    return {**state.meta_instances, mid: mi.model_copy(update=fields)}
+
+
 def apply_instance_created(event: InstanceCreated, state: State) -> State:
     instance = event.instance
     new_instances: Mapping[InstanceId, Instance] = {
         **state.instances,
         instance.instance_id: instance,
     }
-    return state.model_copy(update={"instances": new_instances})
+    update: dict[str, object] = {"instances": new_instances}
+    # Reset failure tracking when a new instance is created for a meta-instance
+    if instance.meta_instance_id and instance.meta_instance_id in state.meta_instances:
+        mi = state.meta_instances[instance.meta_instance_id]
+        if mi.placement_error is not None or mi.consecutive_failures > 0:
+            update["meta_instances"] = _update_meta_instance(
+                state,
+                instance.meta_instance_id,
+                placement_error=None,
+                consecutive_failures=0,
+            )
+    return state.model_copy(update=update)
 
 
 def apply_instance_deleted(event: InstanceDeleted, state: State) -> State:
+    deleted_instance = state.instances.get(event.instance_id)
     new_instances: Mapping[InstanceId, Instance] = {
         iid: inst for iid, inst in state.instances.items() if iid != event.instance_id
     }
-    return state.model_copy(update={"instances": new_instances})
+    update: dict[str, object] = {"instances": new_instances}
+
+    # Track failure on the MetaInstance itself
+    if (
+        event.failure_error
+        and deleted_instance
+        and deleted_instance.meta_instance_id
+        and deleted_instance.meta_instance_id in state.meta_instances
+    ):
+        mid = deleted_instance.meta_instance_id
+        mi = state.meta_instances[mid]
+        update["meta_instances"] = {
+            **state.meta_instances,
+            mid: mi.model_copy(
+                update={
+                    "consecutive_failures": mi.consecutive_failures + 1,
+                    "last_failure_error": event.failure_error,
+                }
+            ),
+        }
+
+    return state.model_copy(update=update)
+
+
+def apply_instance_retrying(event: InstanceRetrying, state: State) -> State:
+    """Runners failed but retry limit not reached — remove runners, keep instance."""
+    instance = state.instances.get(event.instance_id)
+    if instance is None:
+        # Instance was already deleted (e.g. cascade from DeleteMetaInstance).
+        # The InstanceDeleted handler already incremented consecutive_failures
+        # on the MetaInstance, so skipping here avoids double-counting.
+        return state
+
+    # Remove all runners belonging to this instance from state
+    runner_ids_to_remove = set(instance.shard_assignments.node_to_runner.values())
+    new_runners: Mapping[RunnerId, RunnerStatus] = {
+        rid: rs for rid, rs in state.runners.items() if rid not in runner_ids_to_remove
+    }
+
+    update: dict[str, object] = {"runners": new_runners}
+
+    # Increment failure count on the MetaInstance
+    if event.meta_instance_id in state.meta_instances:
+        update["meta_instances"] = _update_meta_instance(
+            state,
+            event.meta_instance_id,
+            consecutive_failures=state.meta_instances[
+                event.meta_instance_id
+            ].consecutive_failures
+            + 1,
+            last_failure_error=event.failure_error,
+        )
+
+    return state.model_copy(update=update)
+
+
+def apply_meta_instance_created(event: MetaInstanceCreated, state: State) -> State:
+    new_meta: Mapping[MetaInstanceId, MetaInstance] = {
+        **state.meta_instances,
+        event.meta_instance.meta_instance_id: event.meta_instance,
+    }
+    return state.model_copy(update={"meta_instances": new_meta})
+
+
+def apply_meta_instance_deleted(event: MetaInstanceDeleted, state: State) -> State:
+    new_meta: Mapping[MetaInstanceId, MetaInstance] = {
+        mid: mi
+        for mid, mi in state.meta_instances.items()
+        if mid != event.meta_instance_id
+    }
+    return state.model_copy(update={"meta_instances": new_meta})
+
+
+def apply_meta_instance_placement_failed(
+    event: MetaInstancePlacementFailed, state: State
+) -> State:
+    if event.meta_instance_id not in state.meta_instances:
+        return state
+    return state.model_copy(
+        update={
+            "meta_instances": _update_meta_instance(
+                state, event.meta_instance_id, placement_error=event.reason
+            )
+        }
+    )
 
 
 def apply_runner_status_updated(event: RunnerStatusUpdated, state: State) -> State:

src/exo/shared/models/model_cards.py

@@ -165,7 +165,6 @@ def is_custom_card(model_id: ModelId) -> bool:
 class ConfigData(BaseModel):
     model_config = {"extra": "ignore"}  # Allow unknown fields
 
-    model_type: str | None = None
     architectures: list[str] | None = None
     hidden_size: Annotated[int, Field(ge=0)] | None = None
     layer_count: int = Field(
@@ -201,7 +200,6 @@ class ConfigData(BaseModel):
             return data
 
         for field in [
-            "model_type",
             "architectures",
             "hidden_size",
             "num_hidden_layers",

src/exo/shared/types/api.py

@@ -6,7 +6,7 @@ from uuid import uuid4
 from pydantic import BaseModel, Field
 
 from exo.shared.models.model_cards import ModelCard, ModelId
-from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.common import CommandId, MetaInstanceId, 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
@@ -262,6 +262,26 @@ class DeleteInstanceResponse(BaseModel):
     instance_id: InstanceId
 
 
+class CreateMetaInstanceParams(BaseModel):
+    model_id: ModelId
+    sharding: Sharding = Sharding.Pipeline
+    instance_meta: InstanceMeta = InstanceMeta.MlxRing
+    min_nodes: int = 1
+    node_ids: list[NodeId] | None = None
+
+
+class CreateMetaInstanceResponse(BaseModel):
+    message: str
+    command_id: CommandId
+    meta_instance_id: MetaInstanceId
+
+
+class DeleteMetaInstanceResponse(BaseModel):
+    message: str
+    command_id: CommandId
+    meta_instance_id: MetaInstanceId
+
+
 class AdvancedImageParams(BaseModel):
     seed: Annotated[int, Field(ge=0)] | None = None
     num_inference_steps: Annotated[int, Field(ge=1, le=100)] | None = None
@@ -366,6 +386,15 @@ class DeleteDownloadResponse(CamelCaseModel):
     command_id: CommandId
 
 
+class DistributeModelParams(CamelCaseModel):
+    target_node_ids: list[NodeId] | None = None  # None = all connected nodes
+
+
+class DistributeModelResponse(CamelCaseModel):
+    command_id: CommandId
+    message: str
+
+
 class TraceEventResponse(CamelCaseModel):
     name: str
     start_us: int

src/exo/shared/types/commands.py

@@ -6,7 +6,8 @@ from exo.shared.types.api import (
     ImageGenerationTaskParams,
 )
 from exo.shared.types.chunks import InputImageChunk
-from exo.shared.types.common import CommandId, NodeId
+from exo.shared.types.common import CommandId, MetaInstanceId, NodeId
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
@@ -52,6 +53,14 @@ class TaskCancelled(BaseCommand):
     cancelled_command_id: CommandId
 
 
+class CreateMetaInstance(BaseCommand):
+    meta_instance: MetaInstance
+
+
+class DeleteMetaInstance(BaseCommand):
+    meta_instance_id: MetaInstanceId
+
+
 class TaskFinished(BaseCommand):
     finished_command_id: CommandId
 
@@ -81,6 +90,14 @@ class CancelDownload(BaseCommand):
     model_id: ModelId
 
 
+class DistributeModel(BaseCommand):
+    """Distribute model files from one node to others via MLX distributed."""
+
+    model_id: ModelId
+    source_node_id: NodeId
+    target_node_ids: list[NodeId]
+
+
 DownloadCommand = StartDownload | DeleteDownload | CancelDownload
 
 
@@ -94,8 +111,11 @@ Command = (
     | CreateInstance
     | DeleteInstance
     | TaskCancelled
+    | CreateMetaInstance
+    | DeleteMetaInstance
     | TaskFinished
     | SendInputChunk
+    | DistributeModel
 )
 
 

src/exo/shared/types/common.py

@@ -42,6 +42,10 @@ class CommandId(Id):
     pass
 
 
+class MetaInstanceId(Id):
+    """Identifier for a MetaInstance."""
+
+
 class Host(CamelCaseModel):
     ip: str
     port: int

src/exo/shared/types/events.py

@@ -1,11 +1,14 @@
+import base64
+from collections.abc import Mapping
 from datetime import datetime
-from typing import final
+from typing import Annotated, final
 
-from pydantic import Field
+from pydantic import BeforeValidator, Field, PlainSerializer
 
 from exo.shared.topology import Connection
 from exo.shared.types.chunks import GenerationChunk, InputImageChunk
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId
+from exo.shared.types.common import CommandId, Id, MetaInstanceId, NodeId, SessionId
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
@@ -14,6 +17,28 @@ from exo.utils.info_gatherer.info_gatherer import GatheredInfo
 from exo.utils.pydantic_ext import CamelCaseModel, FrozenModel, TaggedModel
 
 
+def _decode_base64_bytes(v: bytes | str) -> bytes:
+    if isinstance(v, bytes):
+        return v
+    return base64.b64decode(v)
+
+
+def _encode_base64_bytes(v: bytes) -> str:
+    return base64.b64encode(v).decode("ascii")
+
+
+Base64Bytes = Annotated[
+    bytes,
+    BeforeValidator(_decode_base64_bytes),
+    PlainSerializer(_encode_base64_bytes, return_type=str),
+]
+"""bytes that serialize to/from base64 strings in JSON.
+
+Needed because TaggedModel's wrap validator converts JSON→Python validation
+context, which breaks strict-mode bytes deserialization from JSON strings.
+"""
+
+
 class EventId(Id):
     """
     Newtype around `ID`
@@ -66,6 +91,30 @@ class InstanceCreated(BaseEvent):
 
 class InstanceDeleted(BaseEvent):
     instance_id: InstanceId
+    failure_error: str | None = None
+
+
+class MetaInstanceCreated(BaseEvent):
+    meta_instance: MetaInstance
+
+
+class MetaInstanceDeleted(BaseEvent):
+    meta_instance_id: MetaInstanceId
+
+
+@final
+class MetaInstancePlacementFailed(BaseEvent):
+    meta_instance_id: MetaInstanceId
+    reason: str
+
+
+@final
+class InstanceRetrying(BaseEvent):
+    """Runners failed but retry count is below the limit — restart runners, keep instance."""
+
+    instance_id: InstanceId
+    meta_instance_id: MetaInstanceId
+    failure_error: str
 
 
 class RunnerStatusUpdated(BaseEvent):
@@ -132,6 +181,25 @@ class TracesMerged(BaseEvent):
     traces: list[TraceEventData]
 
 
+@final
+class JacclSideChannelData(BaseEvent):
+    """A runner's local contribution to a JACCL SideChannel all_gather round."""
+
+    instance_id: InstanceId
+    runner_id: RunnerId
+    sequence: int
+    data: Base64Bytes
+
+
+@final
+class JacclSideChannelGathered(BaseEvent):
+    """Gathered result of a JACCL SideChannel all_gather round."""
+
+    instance_id: InstanceId
+    sequence: int
+    gathered_data: Mapping[RunnerId, Base64Bytes]
+
+
 Event = (
     TestEvent
     | TaskCreated
@@ -141,6 +209,10 @@ Event = (
     | TaskAcknowledged
     | InstanceCreated
     | InstanceDeleted
+    | InstanceRetrying
+    | MetaInstanceCreated
+    | MetaInstanceDeleted
+    | MetaInstancePlacementFailed
     | RunnerStatusUpdated
     | RunnerDeleted
     | NodeTimedOut
@@ -152,6 +224,8 @@ Event = (
     | TopologyEdgeDeleted
     | TracesCollected
     | TracesMerged
+    | JacclSideChannelData
+    | JacclSideChannelGathered
 )
 
 

src/exo/shared/types/meta_instance.py

@@ -0,0 +1,25 @@
+from typing import final
+
+from pydantic import Field
+
+from exo.shared.models.model_cards import ModelId
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.worker.instances import InstanceMeta
+from exo.shared.types.worker.shards import Sharding
+from exo.utils.pydantic_ext import FrozenModel
+
+
+@final
+class MetaInstance(FrozenModel):
+    """Declarative constraint: ensure an instance matching these parameters always exists."""
+
+    meta_instance_id: MetaInstanceId = Field(default_factory=MetaInstanceId)
+    model_id: ModelId
+    sharding: Sharding = Sharding.Pipeline
+    instance_meta: InstanceMeta = InstanceMeta.MlxRing
+    min_nodes: int = 1
+    node_ids: list[NodeId] | None = None
+    # Failure tracking
+    placement_error: str | None = None
+    consecutive_failures: int = 0
+    last_failure_error: str | None = None

src/exo/shared/types/state.py

@@ -6,7 +6,8 @@ from pydantic import ConfigDict, Field, field_serializer, field_validator
 from pydantic.alias_generators import to_camel
 
 from exo.shared.topology import Topology, TopologySnapshot
-from exo.shared.types.common import NodeId
+from exo.shared.types.common import MetaInstanceId, NodeId
+from exo.shared.types.meta_instance import MetaInstance
 from exo.shared.types.profiling import (
     DiskUsage,
     MemoryUsage,
@@ -41,6 +42,7 @@ class State(CamelCaseModel):
         arbitrary_types_allowed=True,
     )
     instances: Mapping[InstanceId, Instance] = {}
+    meta_instances: Mapping[MetaInstanceId, MetaInstance] = {}
     runners: Mapping[RunnerId, RunnerStatus] = {}
     downloads: Mapping[NodeId, Sequence[DownloadProgress]] = {}
     tasks: Mapping[TaskId, Task] = {}

src/exo/shared/types/tasks.py

@@ -42,7 +42,7 @@ class DownloadModel(BaseTask):  # emitted by Worker
 
 
 class LoadModel(BaseTask):  # emitted by Worker
-    pass
+    has_local_model: bool = Field(default=True)
 
 
 class ConnectToGroup(BaseTask):  # emitted by Worker
@@ -61,7 +61,7 @@ class TextGeneration(BaseTask):  # emitted by Master
     error_message: str | None = Field(default=None)
 
 
-class CancelTask(BaseTask):
+class CancelTask(BaseTask):  # emitted by Worker when master cancels a task
     cancelled_task_id: TaskId
     runner_id: RunnerId
 
@@ -82,6 +82,13 @@ class ImageEdits(BaseTask):  # emitted by Master
     error_message: str | None = Field(default=None)
 
 
+class TransferModelToDisk(BaseTask):  # emitted by Worker
+    """Transfer all model files from source to receivers' disk via MLX distributed."""
+
+    shard_metadata: ShardMetadata
+    has_local_model: bool = Field(default=True)
+
+
 class Shutdown(BaseTask):  # emitted by Worker
     runner_id: RunnerId
 
@@ -91,6 +98,7 @@ Task = (
     | DownloadModel
     | ConnectToGroup
     | LoadModel
+    | TransferModelToDisk
     | StartWarmup
     | TextGeneration
     | CancelTask

src/exo/shared/types/worker/instances.py

@@ -2,7 +2,7 @@ from enum import Enum
 
 from pydantic import model_validator
 
-from exo.shared.types.common import Host, Id, NodeId
+from exo.shared.types.common import Host, Id, MetaInstanceId, NodeId
 from exo.shared.types.worker.runners import RunnerId, ShardAssignments, ShardMetadata
 from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
 
@@ -19,6 +19,7 @@ class InstanceMeta(str, Enum):
 class BaseInstance(TaggedModel):
     instance_id: InstanceId
     shard_assignments: ShardAssignments
+    meta_instance_id: MetaInstanceId | None = None
 
     def shard(self, runner_id: RunnerId) -> ShardMetadata | None:
         return self.shard_assignments.runner_to_shard.get(runner_id, None)

src/exo/shared/types/worker/runners.py

@@ -84,6 +84,7 @@ class ShardAssignments(CamelCaseModel):
     model_id: ModelId
     runner_to_shard: Mapping[RunnerId, ShardMetadata]
     node_to_runner: Mapping[NodeId, RunnerId]
+    transfer_only: bool = False
 
     @model_validator(mode="after")
     def validate_runners_exist(self) -> "ShardAssignments":

src/exo/utils/channels.py

@@ -125,9 +125,7 @@ class MpSender[T]:
             self._state.buffer.put(item, block=True)
 
     async def send_async(self, item: T) -> None:
-        await to_thread.run_sync(
-            self.send, item, limiter=CapacityLimiter(1), abandon_on_cancel=True
-        )
+        await to_thread.run_sync(self.send, item, limiter=CapacityLimiter(1))
 
     def close(self) -> None:
         if not self._state.closed.is_set():

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

@@ -47,6 +47,7 @@ if TYPE_CHECKING:
     from mlx_lm.models.cache import Cache
 
 TimeoutCallback = Callable[[], None]
+WeightLoader = Callable[[nn.Module, int], None] | None
 
 
 def eval_with_timeout(
@@ -346,6 +347,7 @@ def tensor_auto_parallel(
     group: mx.distributed.Group,
     timeout_seconds: float = 60.0,
     on_timeout: TimeoutCallback | None = None,
+    weight_loader: WeightLoader = None,
 ) -> nn.Module:
     all_to_sharded_linear = partial(
         shard_linear,
@@ -455,7 +457,7 @@ def tensor_auto_parallel(
         raise ValueError(f"Unsupported model type: {type(model)}")
 
     model = tensor_parallel_sharding_strategy.shard_model(
-        model, timeout_seconds, on_timeout
+        model, timeout_seconds, on_timeout, weight_loader
     )
     return patch_tensor_model(model)
 
@@ -482,6 +484,7 @@ class TensorParallelShardingStrategy(ABC):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module: ...
 
 
@@ -491,9 +494,12 @@ class LlamaShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(LlamaModel, model)
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             # Force load weights before sharding to avoid FAST_SYNCH deadlock
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
@@ -545,9 +551,12 @@ class DeepSeekShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(DeepseekV3Model, model)
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
@@ -620,9 +629,12 @@ class GLM4MoeLiteShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(GLM4MoeLiteModel, model)
-        for layer in model.layers:  # type: ignore
+        for i, layer in enumerate(model.layers):  # type: ignore
+            if weight_loader is not None:
+                weight_loader(model, i)
             layer = cast(Glm4MoeLiteDecoderLayer, layer)
             eval_with_timeout(
                 layer.parameters(),
@@ -762,9 +774,12 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(MiniMaxModel, model)
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
@@ -802,9 +817,12 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(Qwen3MoeModel | Qwen3NextModel, model)
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
@@ -926,9 +944,12 @@ class Glm4MoeShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(Glm4MoeModel, model)
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
@@ -972,10 +993,13 @@ class GptOssShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(GptOssMoeModel, model)
 
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
@@ -1013,10 +1037,13 @@ class Step35ShardingStrategy(TensorParallelShardingStrategy):
         model: nn.Module,
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
+        weight_loader: WeightLoader = None,
     ) -> nn.Module:
         model = cast(Step35Model, model)
 
-        for layer in model.layers:
+        for i, layer in enumerate(model.layers):
+            if weight_loader is not None:
+                weight_loader(model, i)
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )

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

@@ -0,0 +1,507 @@
+"""
+Model transfer via MLX distributed all_sum.
+
+Three transfer modes:
+1. Metadata file transfer: broadcast small files (config.json, tokenizer, etc.) to disk
+2. Weight tensor broadcast: stream weight tensors directly into memory via all_sum
+3. Full file transfer: broadcast all files (including safetensors) to disk
+
+All functions are collective operations — every rank in the group must call them.
+
+Protocol relies on all_sum: source has real data, receivers have zeros.
+all_sum(source + zeros) = source data on all ranks.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+import tempfile
+from functools import partial
+from pathlib import Path
+from typing import Any, Final, cast
+
+import mlx.core as mx
+
+from exo.shared.constants import EXO_MODELS_DIR
+from exo.shared.models.model_cards import ModelId
+from exo.worker.runner.bootstrap import logger
+
+Group = mx.distributed.Group
+
+CHUNK_SIZE: Final[int] = 100 * 1024 * 1024  # 100 MB
+_LAYER_RE: Final[re.Pattern[str]] = re.compile(r"(?:^|\.)(layers|h)\.(\d+)\.")
+
+
+def _all_sum_cpu(x: mx.array, group: Group) -> mx.array:
+    """all_sum on CPU stream to avoid GPU memory pressure."""
+    return mx.distributed.all_sum(
+        x, stream=mx.default_stream(mx.Device(mx.cpu)), group=group
+    )
+
+
+def _is_metadata_file(filename: str) -> bool:
+    """A metadata file is anything that isn't a weight file or weight index.
+
+    Weight indices (.safetensors.index.json) reference safetensors shard paths.
+    Transferring them to a receiver that has no safetensors files is harmless
+    today (load_model's glob doesn't match them), but excluding them avoids
+    stale references and keeps the transfer minimal.
+    """
+    if filename.endswith(".safetensors"):
+        return False
+    return not filename.endswith(".safetensors.index.json")
+
+
+def model_path_for_id(model_id: ModelId) -> Path:
+    """Get model path without requiring directory to exist (unlike build_model_path)."""
+    return EXO_MODELS_DIR / model_id.normalize()
+
+
+def coordinate_transfer(group: Group, has_local_model: bool) -> tuple[bool, int]:
+    """
+    Determine if a transfer is needed and which rank is the source.
+
+    All ranks must call this function (uses collective all_sum).
+
+    Returns:
+        (needs_transfer, source_rank) — source_rank is the lowest rank
+        that has the model. needs_transfer is True if any rank is missing it.
+    """
+    all_sum = partial(_all_sum_cpu, group=group)
+    world_size = group.size()
+
+    # Each rank broadcasts a one-hot vector at its position if it has the model
+    bitmask = mx.zeros(world_size, dtype=mx.int32)
+    if has_local_model:
+        bitmask = bitmask.at[group.rank()].add(1)
+    summed = all_sum(bitmask)
+    mx.eval(summed)
+
+    has_model_flags: list[int] = summed.tolist()  # type: ignore[assignment]
+    total_have = sum(has_model_flags)
+
+    if total_have == 0:
+        raise RuntimeError(
+            "No rank has the model files — cannot transfer. "
+            "At least one node must have downloaded the model."
+        )
+
+    if total_have == world_size:
+        logger.info("All ranks have model files, no transfer needed")
+        return False, 0
+
+    source_rank = next(i for i, flag in enumerate(has_model_flags) if flag > 0)
+    logger.info(
+        f"Transfer needed: source_rank={source_rank}, "
+        f"{total_have}/{world_size} ranks have model"
+    )
+    return True, source_rank
+
+
+def _broadcast_json(obj: object, group: Group, is_source: bool) -> object:
+    """Broadcast a JSON-serializable object from source to all ranks."""
+    all_sum = partial(_all_sum_cpu, group=group)
+
+    data = json.dumps(obj, separators=(",", ":")).encode("utf-8") if is_source else b""
+
+    # Broadcast length
+    len_arr = mx.array([len(data) if is_source else 0], dtype=mx.int64)
+    len_result = all_sum(len_arr)
+    mx.eval(len_result)
+    length = int(len_result.item())
+    if length == 0:
+        return None
+
+    # Broadcast payload
+    if is_source:
+        arr = mx.array(list(data), dtype=mx.uint8)
+    else:
+        arr = mx.zeros(length, dtype=mx.uint8)
+    result = all_sum(arr)
+    mx.eval(result)
+    return json.loads(bytes(cast(list[int], result.tolist())))  # pyright: ignore[reportAny]
+
+
+def _build_manifest(
+    model_path: Path, metadata_only: bool = False
+) -> list[dict[str, str | int]]:
+    """Build a list of files in the model directory with their relative paths and sizes."""
+    manifest: list[dict[str, str | int]] = []
+    for root, _dirs, files in os.walk(model_path):
+        for fname in sorted(files):
+            if metadata_only and not _is_metadata_file(fname):
+                continue
+            full_path = Path(root) / fname
+            rel_path = str(full_path.relative_to(model_path))
+            manifest.append(
+                {
+                    "path": rel_path,
+                    "size": full_path.stat().st_size,
+                }
+            )
+    return manifest
+
+
+def _transfer_file_to_disk(
+    source_path: Path,
+    rel_path: str,
+    file_size: int,
+    group: Group,
+    is_source: bool,
+    dest_path: Path,
+) -> None:
+    """Transfer a single file chunk-by-chunk via all_sum. Source reads from disk, receivers write to dest_path."""
+    all_sum = partial(_all_sum_cpu, group=group)
+
+    if is_source:
+        src_file = source_path / rel_path
+        with open(src_file, "rb") as f:
+            offset = 0
+            while offset < file_size:
+                chunk_bytes = min(CHUNK_SIZE, file_size - offset)
+                data = f.read(chunk_bytes)
+                if not data:
+                    break
+                size_arr = mx.array([len(data)], dtype=mx.int64)
+                mx.eval(all_sum(size_arr))
+                chunk_arr = mx.array(list(data), dtype=mx.uint8)
+                result = all_sum(chunk_arr)
+                mx.eval(result)
+                offset += len(data)
+            # Signal end of file
+            mx.eval(all_sum(mx.array([0], dtype=mx.int64)))
+    else:
+        dst_file = dest_path / rel_path
+        os.makedirs(dst_file.parent, exist_ok=True)
+        with open(dst_file, "wb") as f:
+            while True:
+                size_arr = all_sum(mx.zeros(1, dtype=mx.int64))
+                mx.eval(size_arr)
+                chunk_size = int(size_arr.item())
+                if chunk_size == 0:
+                    break
+                chunk_data = all_sum(mx.zeros(chunk_size, dtype=mx.uint8))
+                mx.eval(chunk_data)
+                f.write(bytes(cast(list[int], chunk_data.tolist())))
+
+
+def _transfer_files_to_disk(
+    model_path: Path,
+    group: Group,
+    is_source: bool,
+    metadata_only: bool = False,
+) -> None:
+    """
+    Transfer files from source to all receivers' disk.
+
+    Source broadcasts a manifest then each file. Receivers write to a temp dir
+    then atomically move files to model_path.
+    """
+    if is_source:
+        source_manifest = _build_manifest(model_path, metadata_only=metadata_only)
+    else:
+        source_manifest = []
+    manifest = cast(
+        list[dict[str, str | int]],
+        _broadcast_json(source_manifest if is_source else None, group, is_source),
+    )
+
+    if not manifest:
+        logger.info("No files to transfer")
+        return
+
+    logger.info(
+        f"Transferring {len(manifest)} files ({'metadata only' if metadata_only else 'all'})"
+    )
+
+    temp_dir: Path | None = None
+    if not is_source:
+        os.makedirs(model_path.parent, exist_ok=True)
+        temp_dir = Path(
+            tempfile.mkdtemp(
+                dir=model_path.parent,
+                prefix=f".transfer_{model_path.name}_",
+            )
+        )
+
+    try:
+        for entry in manifest:
+            rel_path = str(entry["path"])
+            file_size = int(entry["size"])
+            logger.info(f"  {rel_path} ({file_size} bytes)")
+            _transfer_file_to_disk(
+                source_path=model_path,
+                rel_path=rel_path,
+                file_size=file_size,
+                group=group,
+                is_source=is_source,
+                dest_path=temp_dir if temp_dir is not None else model_path,
+            )
+
+        if temp_dir is not None:
+            os.makedirs(model_path, exist_ok=True)
+            for entry in manifest:
+                rel_path = str(entry["path"])
+                src = temp_dir / rel_path
+                dst = model_path / rel_path
+                os.makedirs(dst.parent, exist_ok=True)
+                os.replace(src, dst)
+            logger.info(
+                f"Transfer complete: {len(manifest)} files moved to {model_path}"
+            )
+    finally:
+        if temp_dir is not None and temp_dir.exists():
+            shutil.rmtree(temp_dir, ignore_errors=True)
+
+
+def transfer_metadata_files(model_path: Path, group: Group, is_source: bool) -> None:
+    """
+    Transfer metadata files (config.json, tokenizer files, etc.) to receivers' disk.
+
+    All ranks must call this function (collective operation).
+    Only the designated source (is_source=True) should send; all others receive.
+    """
+    _transfer_files_to_disk(model_path, group, is_source=is_source, metadata_only=True)
+
+
+def transfer_all_files(model_path: Path, group: Group, is_source: bool) -> None:
+    """
+    Transfer ALL model files (including safetensors) to receivers' disk.
+
+    All ranks must call this function (collective operation).
+    Only the designated source (is_source=True) should send; all others receive.
+    """
+    _transfer_files_to_disk(model_path, group, is_source=is_source, metadata_only=False)
+
+
+def _parse_mx_dtype(dtype_str: str) -> mx.Dtype:
+    """Convert a dtype string like 'float16' or 'mlx.core.float16' to mx.Dtype."""
+    name = dtype_str.split(".")[-1]
+    dtype = getattr(mx, name, None)
+    if dtype is None:
+        raise ValueError(f"Unknown MLX dtype: {dtype_str}")
+    return dtype  # type: ignore[return-value]
+
+
+def _extract_layer_index(name: str) -> int | None:
+    """Extract layer index from a weight name, or None for non-layer weights.
+
+    Matches patterns like ``model.layers.5.self_attn.q_proj.weight``
+    or ``transformer.h.12.mlp.gate_proj.scales``.
+    """
+    m = _LAYER_RE.search(name)
+    return int(m.group(2)) if m else None
+
+
+class WeightBroadcastState:
+    """Holds state for layer-by-layer weight broadcasting.
+
+    Created by :func:`prepare_weight_broadcast`.  Callers stream weights
+    incrementally via :meth:`broadcast_non_layer_weights` and
+    :meth:`broadcast_layer` so that at most one layer's worth of un-sharded
+    weight data is resident at a time.
+    """
+
+    def __init__(
+        self,
+        meta: dict[str, dict[str, Any]],
+        source_weights: dict[str, mx.array] | None,
+        group: Group,
+        is_source: bool,
+    ) -> None:
+        self.meta = meta
+        self.source_weights = source_weights
+        self.group = group
+        self.is_source = is_source
+
+        # Partition weight names into layer vs. non-layer
+        self.layer_names: dict[int, list[str]] = {}
+        self.non_layer_names: list[str] = []
+        for name in sorted(meta.keys()):
+            layer_idx = _extract_layer_index(name)
+            if layer_idx is not None:
+                self.layer_names.setdefault(layer_idx, []).append(name)
+            else:
+                self.non_layer_names.append(name)
+
+        logger.info(
+            f"WeightBroadcastState: {len(self.non_layer_names)} non-layer weights, "
+            f"{len(self.layer_names)} layers"
+        )
+
+    # ------------------------------------------------------------------
+    # Internal helpers
+    # ------------------------------------------------------------------
+
+    def _broadcast_names(self, names: list[str]) -> dict[str, mx.array]:
+        """Broadcast a specific set of weight tensors by name."""
+        all_sum = partial(_all_sum_cpu, group=self.group)
+        result: dict[str, mx.array] = {}
+        for name in names:
+            info = self.meta[name]
+            shape = cast(list[int], info["s"])
+            dtype = _parse_mx_dtype(cast(str, info["d"]))
+
+            if self.is_source:
+                assert self.source_weights is not None
+                tensor = self.source_weights.pop(name)
+                mx.eval(tensor)  # loads from disk (lazy)
+            else:
+                tensor = mx.zeros(shape, dtype=dtype)
+
+            broadcasted = all_sum(tensor)
+            mx.eval(broadcasted)
+            result[name] = broadcasted
+        return result
+
+    # ------------------------------------------------------------------
+    # Public API
+    # ------------------------------------------------------------------
+
+    def broadcast_non_layer_weights(self) -> dict[str, mx.array]:
+        """Broadcast non-layer weights (embeddings, norms, lm_head)."""
+        if not self.non_layer_names:
+            return {}
+        logger.info(
+            f"Broadcasting {len(self.non_layer_names)} non-layer weight tensors"
+        )
+        return self._broadcast_names(self.non_layer_names)
+
+    def broadcast_layer(self, layer_idx: int) -> dict[str, mx.array]:
+        """Broadcast weights for a single transformer layer."""
+        names = self.layer_names.get(layer_idx, [])
+        if not names:
+            return {}
+        return self._broadcast_names(names)
+
+
+def prepare_weight_broadcast(
+    model_path: Path,
+    group: Group,
+    is_source: bool,
+) -> WeightBroadcastState:
+    """Prepare for layer-by-layer weight broadcasting.
+
+    Source loads safetensors lazily and broadcasts weight metadata (names,
+    shapes, dtypes) as JSON.  Returns a :class:`WeightBroadcastState` that
+    can then stream weights incrementally via ``broadcast_layer()``.
+
+    All ranks must call this function (collective operation).
+    """
+    source_weights: dict[str, mx.array] | None = None
+    if is_source:
+        source_weights = {}
+        weight_files = sorted(model_path.glob("*.safetensors"))
+        if not weight_files:
+            weight_files = sorted(model_path.glob("**/*.safetensors"))
+        for wf in weight_files:
+            try:
+                loaded = cast(
+                    dict[str, mx.array],
+                    mx.load(str(wf), lazy=True),  # pyright: ignore[reportCallIssue]
+                )
+            except TypeError:
+                loaded = cast(dict[str, mx.array], mx.load(str(wf)))
+            source_weights.update(loaded)
+        logger.info(
+            f"Source loaded {len(source_weights)} weight tensors (lazy) "
+            f"from {len(weight_files)} files"
+        )
+
+    # Broadcast metadata
+    if is_source and source_weights is not None:
+        source_meta: dict[str, dict[str, Any]] = {
+            name: {"s": list(tensor.shape), "d": str(tensor.dtype)}
+            for name, tensor in source_weights.items()
+        }
+    else:
+        source_meta = {}
+
+    meta = cast(
+        dict[str, dict[str, Any]],
+        _broadcast_json(source_meta if is_source else None, group, is_source),
+    )
+
+    logger.info(f"Weight broadcast prepared: {len(meta)} tensors")
+    return WeightBroadcastState(meta, source_weights, group, is_source)
+
+
+def broadcast_model_weights(
+    model_path: Path,
+    group: Group,
+    is_source: bool,
+) -> dict[str, mx.array]:
+    """
+    Broadcast model weight tensors from source rank to all receivers' memory.
+
+    Source loads weights from .safetensors files on disk and broadcasts each
+    tensor via all_sum. Receivers receive tensors directly as mx.arrays in
+    memory — no disk write for weight data.
+
+    All ranks must call this function (collective operation).
+    Only the designated source (is_source=True) should send; all others receive.
+
+    Returns:
+        dict mapping weight names to mx.arrays (on all ranks).
+    """
+    all_sum = partial(_all_sum_cpu, group=group)
+
+    # Source loads weights (lazy if supported, so only one tensor in memory at a time)
+    weights: dict[str, mx.array] = {}
+    if is_source:
+        weight_files = sorted(model_path.glob("*.safetensors"))
+        if not weight_files:
+            weight_files = sorted(model_path.glob("**/*.safetensors"))
+        for wf in weight_files:
+            try:
+                loaded = cast(dict[str, mx.array], mx.load(str(wf), lazy=True))  # pyright: ignore[reportCallIssue]
+            except TypeError:
+                loaded = cast(dict[str, mx.array], mx.load(str(wf)))
+            weights.update(loaded)
+        logger.info(
+            f"Source loaded {len(weights)} weight tensors from {len(weight_files)} files"
+        )
+
+    # Broadcast weight metadata: {name: {shape, dtype}}
+    if is_source:
+        source_meta: dict[str, dict[str, Any]] = {
+            name: {"s": list(tensor.shape), "d": str(tensor.dtype)}
+            for name, tensor in weights.items()
+        }
+    else:
+        source_meta = {}
+    meta = cast(
+        dict[str, dict[str, Any]],
+        _broadcast_json(source_meta if is_source else None, group, is_source),
+    )
+
+    logger.info(f"Broadcasting {len(meta)} weight tensors")
+
+    # Broadcast each tensor in sorted order (deterministic across ranks).
+    # Source loads one tensor at a time from disk (lazy), broadcasts it,
+    # then drops the reference so only one tensor is in flight at a time.
+    result: dict[str, mx.array] = {}
+    for i, name in enumerate(sorted(meta.keys())):
+        info = meta[name]
+        shape = cast(list[int], info["s"])
+        dtype_str = cast(str, info["d"])
+        dtype = _parse_mx_dtype(dtype_str)
+
+        if is_source:
+            tensor = weights.pop(name)  # pop to free lazy ref after broadcast
+            mx.eval(tensor)  # loads from disk
+        else:
+            tensor = mx.zeros(shape, dtype=dtype)
+
+        broadcasted = all_sum(tensor)
+        mx.eval(broadcasted)
+        result[name] = broadcasted
+
+        if (i + 1) % 100 == 0:
+            logger.info(f"  Broadcast {i + 1}/{len(meta)} tensors")
+
+    logger.info(f"Weight broadcast complete: {len(result)} tensors")
+    return result

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

@@ -2,6 +2,7 @@ import json
 import os
 import sys
 import time
+from collections.abc import Callable
 from pathlib import Path
 from typing import Any, cast
 
@@ -59,6 +60,13 @@ from exo.worker.engines.mlx.auto_parallel import (
     pipeline_auto_parallel,
     tensor_auto_parallel,
 )
+from exo.worker.engines.mlx.model_transfer import (
+    WeightBroadcastState,
+    coordinate_transfer,
+    model_path_for_id,
+    prepare_weight_broadcast,
+    transfer_metadata_files,
+)
 from exo.worker.runner.bootstrap import logger
 
 Group = mx.distributed.Group
@@ -171,6 +179,7 @@ def load_mlx_items(
     bound_instance: BoundInstance,
     group: Group | None,
     on_timeout: TimeoutCallback | None = None,
+    has_local_model: bool = True,
 ) -> tuple[Model, TokenizerWrapper]:
     if group is None:
         logger.info(f"Single device used for {bound_instance.instance}")
@@ -185,7 +194,10 @@ def load_mlx_items(
         logger.info("Starting distributed init")
         start_time = time.perf_counter()
         model, tokenizer = shard_and_load(
-            bound_instance.bound_shard, group=group, on_timeout=on_timeout
+            bound_instance.bound_shard,
+            group=group,
+            on_timeout=on_timeout,
+            has_local_model=has_local_model,
         )
         end_time = time.perf_counter()
         logger.info(
@@ -201,30 +213,89 @@ def shard_and_load(
     shard_metadata: ShardMetadata,
     group: Group,
     on_timeout: TimeoutCallback | None = None,
+    has_local_model: bool = True,
 ) -> tuple[nn.Module, TokenizerWrapper]:
-    model_path = build_model_path(shard_metadata.model_card.model_id)
+    model_id = shard_metadata.model_card.model_id
+    model_path = model_path_for_id(model_id)
 
-    model, _ = load_model(model_path, lazy=True, strict=False)
+    # Coordinate: does any rank need a transfer?
+    needs_transfer, source_rank = coordinate_transfer(group, has_local_model)
+    is_source = group.rank() == source_rank
+
+    # Step 1: Always ensure all nodes have metadata files (config, tokenizer, etc.).
+    # This is cheap (~20MB, ~1s) and guarantees config.json is present for load_model().
+    transfer_metadata_files(model_path, group, is_source)
+
+    # Step 2: Only broadcast weights if some rank is missing the model
+    broadcast_state: WeightBroadcastState | None = None
+    if needs_transfer:
+        logger.info(
+            f"Model transfer needed (source_rank={source_rank}, "
+            f"is_source={is_source}, local_weights={has_local_model})"
+        )
+        broadcast_state = prepare_weight_broadcast(model_path, group, is_source)
+
+    # Create model architecture (all ranks have config.json on disk now).
+    # Always use lazy=True when we have broadcast state: load_model's internal
+    # nn.quantize skips quantization when weights dict is empty (no safetensors),
+    # leaving the model un-quantized. lazy=False would then mx.eval() the full
+    # fp16 model (~72GB for a 36B-param model), causing OOM on the receiver.
+    # We handle quantization ourselves below before loading broadcast weights.
+    use_lazy = has_local_model or broadcast_state is not None
+    model, _ = load_model(model_path, lazy=use_lazy, strict=False)
     logger.debug(model)
     if hasattr(model, "model") and isinstance(model.model, DeepseekV3Model):  # type: ignore
         pass
         # TODO: See if we should quantize the model.
-        # def is_attention_layer(path: str) -> bool:
-        #     path = path.lower()
-
-        #     return "self_attn" in path and "layernorm" not in path
-
-        # def quant_predicate(path: str, module: nn.Module):
-        #     if not isinstance(module, nn.Linear):
-        #         return False
-
-        #     return is_attention_layer(path)
-        # model, config = quantize_model(
-        #        model, config, group_size=KV_GROUP_SIZE, bits=ATTENTION_KV_BITS, quant_predicate=quant_predicate, mode=QUANTIZE_MODEL_MODE
-        #    )
 
     assert isinstance(model, nn.Module)
 
+    if broadcast_state is not None:
+        # When receiver has no weight files, load_model skips quantization
+        # (its class_predicate checks `f"{p}.scales" in weights`, which is
+        # always False when weights is empty).  Apply quantization explicitly
+        # using the broadcast metadata to determine which layers are quantized,
+        # matching load_model's selective quantization logic exactly.
+        if not has_local_model:
+            config_path = model_path / "config.json"
+            with open(config_path) as f:
+                config = json.load(f)  # pyright: ignore[reportAny]
+            quant_config: dict[str, Any] | None = config.get(  # pyright: ignore[reportAny]
+                "quantization", None
+            )
+            if quant_config is not None:
+                logger.info(f"Applying quantization to receiver model: {quant_config}")
+                broadcast_weight_names = set(broadcast_state.meta.keys())
+
+                def _class_predicate(p: str, m: nn.Module) -> bool | dict[str, Any]:
+                    # Per-layer overrides from config (e.g. "lm_head": false)
+                    assert quant_config is not None
+                    if p in quant_config:
+                        return quant_config[p]  # pyright: ignore[reportAny]
+                    if not hasattr(m, "to_quantized"):
+                        return False
+                    # Only quantize layers whose .scales exist in broadcast weights
+                    return f"{p}.scales" in broadcast_weight_names
+
+                group_size = int(quant_config.get("group_size", 64))  # pyright: ignore[reportAny]
+                bits = int(quant_config.get("bits", 4))  # pyright: ignore[reportAny]
+                mode: str = quant_config.get("mode", "affine")  # pyright: ignore[reportAny]
+                nn.quantize(  # pyright: ignore[reportUnknownMemberType]
+                    model,
+                    group_size=group_size,
+                    bits=bits,
+                    mode=mode,
+                    class_predicate=_class_predicate,
+                )
+
+        # Broadcast and load non-layer weights (embeddings, norms, lm_head) upfront.
+        # These are small (~600MB) and needed before the sharding loop.
+        non_layer_weights = broadcast_state.broadcast_non_layer_weights()
+        if non_layer_weights:
+            model.load_weights(list(non_layer_weights.items()), strict=False)
+            logger.info(f"Loaded {len(non_layer_weights)} non-layer weight tensors")
+        del non_layer_weights
+
     tokenizer = get_tokenizer(model_path, shard_metadata)
 
     logger.info(f"Group size: {group.size()}, group rank: {group.rank()}")
@@ -238,12 +309,43 @@ def shard_and_load(
         f"(model size: {model_size_gb:.1f}GB)"
     )
 
+    # Build per-layer weight loader for streaming broadcast during sharding.
+    # Each layer's weights are broadcast via all_sum just before that layer is
+    # sharded, so at most one un-sharded layer is in memory at a time.
+    weight_loader_fn: Callable[[nn.Module, int], None] | None = None
+    if broadcast_state is not None:
+        _state = broadcast_state  # capture for closure
+
+        def _load_layer_weights(mdl: nn.Module, layer_idx: int) -> None:
+            layer_weights = _state.broadcast_layer(layer_idx)
+            if layer_weights:
+                mdl.load_weights(list(layer_weights.items()), strict=False)
+
+        weight_loader_fn = _load_layer_weights
+
     match shard_metadata:
         case TensorShardMetadata():
             logger.info(f"loading model from {model_path} with tensor parallelism")
-            model = tensor_auto_parallel(model, group, timeout_seconds, on_timeout)
+            model = tensor_auto_parallel(
+                model, group, timeout_seconds, on_timeout, weight_loader_fn
+            )
         case PipelineShardMetadata():
             logger.info(f"loading model from {model_path} with pipeline parallelism")
+            # Broadcast all layers (all_sum is collective — all ranks must
+            # participate) but only load weights for layers this node will
+            # keep after pipeline slicing.  Out-of-range results are discarded,
+            # keeping peak memory proportional to this node's layer count.
+            if broadcast_state is not None:
+                for layer_idx in sorted(broadcast_state.layer_names.keys()):
+                    layer_weights = broadcast_state.broadcast_layer(layer_idx)
+                    if (
+                        shard_metadata.start_layer
+                        <= layer_idx
+                        < shard_metadata.end_layer
+                        and layer_weights
+                    ):
+                        model.load_weights(list(layer_weights.items()), strict=False)
+                    del layer_weights
             model = pipeline_auto_parallel(model, group, shard_metadata)
             eval_with_timeout(model.parameters(), timeout_seconds, on_timeout)
         case CfgShardMetadata():
@@ -252,6 +354,8 @@ def shard_and_load(
                 "this metadata type is only for image generation models"
             )
 
+    del broadcast_state
+
     # TODO: Do we need this?
     mx.eval(model)
 
@@ -269,52 +373,19 @@ def get_tokenizer(model_path: Path, shard_metadata: ShardMetadata) -> TokenizerW
     return load_tokenizer_for_model_id(shard_metadata.model_card.model_id, model_path)
 
 
-def _read_model_type_from_config(model_path: Path) -> str | None:
-    """Read the model_type field from config.json at the given model path.
-
-    Returns None if config.json doesn't exist or doesn't contain model_type.
+def get_eos_token_ids_for_model(model_id: ModelId) -> list[int] | None:
     """
-    config_path = model_path / "config.json"
-    if not config_path.exists():
-        return None
-    try:
-        with open(config_path) as f:
-            config: dict[str, Any] = json.load(f)  # pyright: ignore[reportAny]
-        model_type: Any = config.get("model_type")
-        if model_type is None:
-            text_config: Any = config.get("text_config")
-            if isinstance(text_config, dict):
-                model_type = text_config.get("model_type")  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
-        return model_type if isinstance(model_type, str) else None
-    except (json.JSONDecodeError, OSError):
-        return None
+    Get the EOS token IDs for a model based on its ID.
 
-
-def get_eos_token_ids_for_model(
-    model_id: ModelId, model_type: str | None = None
-) -> list[int] | None:
-    """Get the EOS token IDs for a model based on its architecture type.
-
-    Uses model_type from config.json when available, falls back to model_id
-    string matching for backward compatibility.
+    Some models require explicit EOS token configuration that isn't in their
+    tokenizer config. This function returns the known EOS token IDs for such models.
 
     Args:
         model_id: The HuggingFace model ID
-        model_type: The model_type field from config.json (e.g., "kimi", "glm4")
 
     Returns:
         List of EOS token IDs, or None if the model uses standard tokenizer config
     """
-    if model_type is not None:
-        if model_type == "kimi":
-            return [163586]
-        elif model_type == "glm4_moe_lite":
-            # 154820: <|endoftext|>, 154827: <|user|>, 154829: <|observation|>
-            return [154820, 154827, 154829]
-        elif model_type.startswith("glm"):
-            return [151336, 151329, 151338]
-
-    # Fallback: string matching on model_id
     model_id_lower = model_id.lower()
     if "kimi-k2" in model_id_lower:
         return [163586]
@@ -329,10 +400,11 @@ def get_eos_token_ids_for_model(
 def load_tokenizer_for_model_id(
     model_id: ModelId, model_path: Path
 ) -> TokenizerWrapper:
-    """Load tokenizer for a model given its ID and local path.
+    """
+    Load tokenizer for a model given its ID and local path.
 
-    Uses model_type from config.json for architecture detection when available,
-    falling back to model_id string matching for backward compatibility.
+    This is the core tokenizer loading logic, handling special cases for different
+    model families (Kimi, GLM, etc.) and transformers 5.x compatibility.
 
     Args:
         model_id: The HuggingFace model ID (e.g., "moonshotai/Kimi-K2-Instruct")
@@ -341,21 +413,11 @@ def load_tokenizer_for_model_id(
     Returns:
         TokenizerWrapper instance configured for the model
     """
-    model_type = _read_model_type_from_config(model_path)
     model_id_lower = model_id.lower()
-    eos_token_ids = get_eos_token_ids_for_model(model_id, model_type=model_type)
-
-    is_kimi = (
-        model_type == "kimi" if model_type is not None else "kimi-k2" in model_id_lower
-    )
-    is_gemma3 = (
-        model_type == "gemma3"
-        if model_type is not None
-        else "gemma-3" in model_id_lower
-    )
+    eos_token_ids = get_eos_token_ids_for_model(model_id)
 
     # Kimi uses a custom TikTokenTokenizer that transformers 5.x can't load via AutoTokenizer
-    if is_kimi:
+    if "kimi-k2" in model_id_lower:
         import importlib.util
         import types
 
@@ -409,7 +471,7 @@ def load_tokenizer_for_model_id(
         eos_token_ids=eos_token_ids,
     )
 
-    if is_gemma3:
+    if "gemma-3" in model_id_lower:
         gemma_3_eos_id = 1
         gemma_3_end_of_turn_id = 106
         if tokenizer.eos_token_ids is not None:
@@ -616,6 +678,11 @@ def mlx_cleanup(
 
 
 def mx_any(bool_: bool, group: Group | None) -> bool:
+    """Synchronize a boolean across all distributed nodes.
+
+    Returns True if any node has bool_=True.  Uses all_sum so every
+    node participates in the collective — preventing GPU deadlocks.
+    """
     if group is None:
         return bool_
     num_true = mx.distributed.all_sum(

src/exo/worker/main.py

@@ -24,6 +24,7 @@ from exo.shared.types.events import (
     ForwarderEvent,
     IndexedEvent,
     InputChunkReceived,
+    JacclSideChannelGathered,
     NodeGatheredInfo,
     TaskCreated,
     TaskStatusUpdated,
@@ -33,7 +34,6 @@ from exo.shared.types.events import (
 from exo.shared.types.multiaddr import Multiaddr
 from exo.shared.types.state import State
 from exo.shared.types.tasks import (
-    CancelTask,
     CreateRunner,
     DownloadModel,
     ImageEdits,
@@ -159,6 +159,15 @@ class Worker:
                 for idx, event in indexed_events:
                     self.state = apply(self.state, IndexedEvent(idx=idx, event=event))
 
+                    # Dispatch JACCL gathered events to the relevant RunnerSupervisor
+                    if isinstance(event, JacclSideChannelGathered):
+                        for runner in self.runners.values():
+                            if (
+                                runner.bound_instance.instance.instance_id
+                                == event.instance_id
+                            ):
+                                runner.notify_gathered(event)
+
                     # Buffer input image chunks for image editing
                     if isinstance(event, InputChunkReceived):
                         cmd_id = event.command_id
@@ -225,22 +234,15 @@ class Worker:
                         )
                     )
                 case Shutdown(runner_id=runner_id):
-                    runner = self.runners.pop(runner_id)
                     try:
                         with fail_after(3):
-                            await runner.start_task(task)
+                            await self.runners.pop(runner_id).start_task(task)
                     except TimeoutError:
                         await self.event_sender.send(
                             TaskStatusUpdated(
                                 task_id=task.task_id, task_status=TaskStatus.TimedOut
                             )
                         )
-                    finally:
-                        runner.shutdown()
-                case CancelTask(
-                    cancelled_task_id=cancelled_task_id, runner_id=runner_id
-                ):
-                    await self.runners[runner_id].cancel_task(cancelled_task_id)
                 case ImageEdits() if task.task_params.total_input_chunks > 0:
                     # Assemble image from chunks and inject into task
                     cmd_id = task.command_id
@@ -278,18 +280,18 @@ class Worker:
                         del self.input_chunk_buffer[cmd_id]
                     if cmd_id in self.input_chunk_counts:
                         del self.input_chunk_counts[cmd_id]
-                    await self._start_runner_task(modified_task)
+                    await self.runners[self._task_to_runner_id(task)].start_task(
+                        modified_task
+                    )
                 case task:
-                    await self._start_runner_task(task)
+                    await self.runners[self._task_to_runner_id(task)].start_task(task)
 
     def shutdown(self):
         self._tg.cancel_scope.cancel()
 
-    async def _start_runner_task(self, task: Task):
-        if (instance := self.state.instances.get(task.instance_id)) is not None:
-            await self.runners[
-                instance.shard_assignments.node_to_runner[self.node_id]
-            ].start_task(task)
+    def _task_to_runner_id(self, task: Task):
+        instance = self.state.instances[task.instance_id]
+        return instance.shard_assignments.node_to_runner[self.node_id]
 
     async def _nack_request(self, since_idx: int) -> None:
         # We request all events after (and including) the missing index.

src/exo/worker/plan.py

@@ -2,6 +2,7 @@
 
 from collections.abc import Mapping, Sequence
 
+from exo.shared.models.model_cards import ModelId
 from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.tasks import (
     CancelTask,
@@ -17,6 +18,7 @@ from exo.shared.types.tasks import (
     TaskId,
     TaskStatus,
     TextGeneration,
+    TransferModelToDisk,
 )
 from exo.shared.types.worker.downloads import (
     DownloadCompleted,
@@ -35,8 +37,11 @@ from exo.shared.types.worker.runners import (
     RunnerLoading,
     RunnerReady,
     RunnerRunning,
+    RunnerShutdown,
+    RunnerShuttingDown,
     RunnerStatus,
     RunnerWarmingUp,
+    ShardAssignments,
 )
 from exo.worker.runner.runner_supervisor import RunnerSupervisor
 
@@ -56,9 +61,10 @@ def plan(
     return (
         _cancel_tasks(runners, tasks)
         or _kill_runner(runners, all_runners, instances)
-        or _create_runner(node_id, runners, instances)
+        or _create_runner(node_id, runners, instances, all_runners)
         or _model_needs_download(node_id, runners, global_download_status)
         or _init_distributed_backend(runners, all_runners)
+        or _transfer_model_to_disk(runners, all_runners, global_download_status)
         or _load_model(runners, all_runners, global_download_status)
         or _ready_to_warmup(runners, all_runners)
         or _pending_tasks(runners, tasks, all_runners, input_chunk_buffer or {})
@@ -75,6 +81,12 @@ def _kill_runner(
         if (instance_id := runner.bound_instance.instance.instance_id) not in instances:
             return Shutdown(instance_id=instance_id, runner_id=runner_id)
 
+        # Master removed our runner from state (retry signal) and process is dead
+        if runner_id not in all_runners and isinstance(
+            runner.status, (RunnerFailed, RunnerShutdown)
+        ):
+            return Shutdown(instance_id=instance_id, runner_id=runner_id)
+
         for (
             global_runner_id
         ) in runner.bound_instance.instance.shard_assignments.node_to_runner.values():
@@ -92,6 +104,7 @@ def _create_runner(
     node_id: NodeId,
     runners: Mapping[RunnerId, RunnerSupervisor],
     instances: Mapping[InstanceId, Instance],
+    all_runners: Mapping[RunnerId, RunnerStatus],
 ) -> CreateRunner | None:
     for instance in instances.values():
         runner_id = instance.shard_assignments.node_to_runner.get(node_id, None)
@@ -101,6 +114,16 @@ def _create_runner(
         if runner_id in runners:
             continue
 
+        # Don't create while any peer runner is in a terminal state — wait for
+        # the master to emit InstanceRetrying which removes them from state.
+        has_terminal_peer = any(
+            isinstance(all_runners.get(peer_rid), (RunnerFailed, RunnerShutdown))
+            for peer_rid in instance.shard_assignments.node_to_runner.values()
+            if peer_rid != runner_id
+        )
+        if has_terminal_peer:
+            continue
+
         shard = instance.shard(runner_id)
         assert shard is not None
 
@@ -123,6 +146,10 @@ def _model_needs_download(
     }
 
     for runner in runners.values():
+        # Transfer-only instances don't need downloads
+        if runner.bound_instance.instance.shard_assignments.transfer_only:
+            continue
+
         model_id = runner.bound_instance.bound_shard.model_card.model_id
         if isinstance(runner.status, RunnerIdle) and (
             model_id not in download_status
@@ -131,6 +158,15 @@ def _model_needs_download(
                 (DownloadOngoing, DownloadCompleted, DownloadFailed),
             )
         ):
+            # For multi-node instances, skip download if a peer already has the model.
+            # The model will be transferred via MLX distributed during LoadModel.
+            instance = runner.bound_instance.instance
+            is_multi_node = len(instance.shard_assignments.node_to_runner) > 1
+            if is_multi_node and _any_peer_has_model(
+                node_id, model_id, instance, global_download_status
+            ):
+                continue
+
             # We don't invalidate download_status randomly in case a file gets deleted on disk
             return DownloadModel(
                 instance_id=runner.bound_instance.instance.instance_id,
@@ -188,6 +224,43 @@ def _init_distributed_backend(
     return None
 
 
+def _transfer_model_to_disk(
+    runners: Mapping[RunnerId, RunnerSupervisor],
+    all_runners: Mapping[RunnerId, RunnerStatus],
+    global_download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> TransferModelToDisk | None:
+    """For transfer-only instances: after all ranks are connected, emit TransferModelToDisk."""
+    for runner in runners.values():
+        instance = runner.bound_instance.instance
+        shard_assignments = instance.shard_assignments
+
+        if not shard_assignments.transfer_only:
+            continue
+
+        is_runner_connected = isinstance(runner.status, RunnerConnected)
+        all_connected_or_further = all(
+            isinstance(
+                all_runners.get(global_runner_id, None),
+                (RunnerConnected, RunnerLoading, RunnerShuttingDown, RunnerShutdown),
+            )
+            for global_runner_id in shard_assignments.runner_to_shard
+        )
+
+        if is_runner_connected and all_connected_or_further:
+            has_local = _node_has_download(
+                runner.bound_instance.bound_node_id,
+                shard_assignments.model_id,
+                global_download_status,
+            )
+            return TransferModelToDisk(
+                instance_id=instance.instance_id,
+                shard_metadata=runner.bound_instance.bound_shard,
+                has_local_model=has_local,
+            )
+
+    return None
+
+
 def _load_model(
     runners: Mapping[RunnerId, RunnerSupervisor],
     all_runners: Mapping[RunnerId, RunnerStatus],
@@ -197,38 +270,97 @@ def _load_model(
         instance = runner.bound_instance.instance
         shard_assignments = instance.shard_assignments
 
-        all_local_downloads_complete = all(
-            nid in global_download_status
-            and any(
-                isinstance(dp, DownloadCompleted)
-                and dp.shard_metadata.model_card.model_id == shard_assignments.model_id
-                for dp in global_download_status[nid]
-            )
-            for nid in shard_assignments.node_to_runner
-        )
-        if not all_local_downloads_complete:
+        # Transfer-only instances don't load models for inference
+        if shard_assignments.transfer_only:
             continue
 
-        is_single_node_instance = len(instance.shard_assignments.runner_to_shard) == 1
-        if is_single_node_instance and isinstance(runner.status, RunnerIdle):
-            return LoadModel(instance_id=instance.instance_id)
+        is_single_node_instance = len(shard_assignments.runner_to_shard) == 1
 
-        is_runner_waiting = isinstance(runner.status, RunnerConnected)
+        if is_single_node_instance:
+            # Single-node: require local download complete
+            if not _all_downloads_complete(shard_assignments, global_download_status):
+                continue
+            if isinstance(runner.status, RunnerIdle):
+                return LoadModel(instance_id=instance.instance_id, has_local_model=True)
+        else:
+            # Multi-node: require at least one node to have the model downloaded.
+            # Nodes without the model will receive it via MLX distributed transfer
+            # during model loading.
+            if not _any_download_complete(shard_assignments, global_download_status):
+                continue
 
-        all_ready_for_model = all(
-            isinstance(
-                all_runners.get(global_runner_id, None),
-                (RunnerConnected, RunnerLoading, RunnerLoaded),
+            is_runner_waiting = isinstance(runner.status, RunnerConnected)
+            all_ready_for_model = all(
+                isinstance(
+                    all_runners.get(global_runner_id, None),
+                    (RunnerConnected, RunnerLoading, RunnerLoaded),
+                )
+                for global_runner_id in shard_assignments.runner_to_shard
             )
-            for global_runner_id in shard_assignments.runner_to_shard
-        )
 
-        if is_runner_waiting and all_ready_for_model:
-            return LoadModel(instance_id=instance.instance_id)
+            if is_runner_waiting and all_ready_for_model:
+                has_local = _node_has_download(
+                    runner.bound_instance.bound_node_id,
+                    shard_assignments.model_id,
+                    global_download_status,
+                )
+                return LoadModel(
+                    instance_id=instance.instance_id,
+                    has_local_model=has_local,
+                )
 
     return None
 
 
+def _node_has_download(
+    nid: NodeId,
+    model_id: ModelId,
+    global_download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> bool:
+    """Check if a specific node has completed downloading the given model."""
+    return any(
+        isinstance(dp, DownloadCompleted)
+        and dp.shard_metadata.model_card.model_id == model_id
+        for dp in global_download_status.get(nid, [])
+    )
+
+
+def _any_peer_has_model(
+    node_id: NodeId,
+    model_id: ModelId,
+    instance: Instance,
+    global_download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> bool:
+    """Check if any other node in the instance already has the model downloaded."""
+    return any(
+        _node_has_download(nid, model_id, global_download_status)
+        for nid in instance.shard_assignments.node_to_runner
+        if nid != node_id
+    )
+
+
+def _all_downloads_complete(
+    shard_assignments: ShardAssignments,
+    global_download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> bool:
+    """Check if ALL nodes in the instance have completed downloading the model."""
+    return all(
+        _node_has_download(nid, shard_assignments.model_id, global_download_status)
+        for nid in shard_assignments.node_to_runner
+    )
+
+
+def _any_download_complete(
+    shard_assignments: ShardAssignments,
+    global_download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> bool:
+    """Check if at least one node in the instance has completed downloading the model."""
+    return any(
+        _node_has_download(nid, shard_assignments.model_id, global_download_status)
+        for nid in shard_assignments.node_to_runner
+    )
+
+
 def _ready_to_warmup(
     runners: Mapping[RunnerId, RunnerSupervisor],
     all_runners: Mapping[RunnerId, RunnerStatus],
@@ -236,6 +368,11 @@ def _ready_to_warmup(
     for runner in runners.values():
         instance = runner.bound_instance.instance
         shard_assignments = instance.shard_assignments
+
+        # Transfer-only instances don't go through warmup
+        if shard_assignments.transfer_only:
+            continue
+
         shard = runner.bound_instance.bound_shard
         device_rank = shard.device_rank
         runner_id = runner.bound_instance.bound_runner_id
@@ -310,7 +447,8 @@ def _pending_tasks(
 def _cancel_tasks(
     runners: Mapping[RunnerId, RunnerSupervisor],
     tasks: Mapping[TaskId, Task],
-) -> Task | None:
+) -> CancelTask | None:
+    """Find a cancelled task that hasn't been sent to the runner yet."""
     for task in tasks.values():
         if task.task_status != TaskStatus.Cancelled:
             continue

src/exo/worker/runner/bootstrap.py

@@ -17,6 +17,7 @@ def entrypoint(
     task_receiver: MpReceiver[Task],
     cancel_receiver: MpReceiver[TaskId],
     _logger: "loguru.Logger",
+    pipe_fifo_paths: tuple[str, str] | None = None,
 ) -> None:
     fast_synch_override = os.environ.get("EXO_FAST_SYNCH")
     if fast_synch_override == "on" or (
@@ -30,6 +31,16 @@ def entrypoint(
     else:
         os.environ["MLX_METAL_FAST_SYNCH"] = "0"
 
+    # Open JACCL FIFOs by path and set env vars for C++ SideChannel.
+    # Named pipes (FIFOs) work across multiprocessing spawn (macOS default).
+    if pipe_fifo_paths is not None:
+        fifo_c2p, fifo_p2c = pipe_fifo_paths
+        # C++ reads gathered data from p2c (PIPE_IN), writes local data to c2p (PIPE_OUT)
+        pipe_in_fd = os.open(fifo_p2c, os.O_RDONLY)
+        pipe_out_fd = os.open(fifo_c2p, os.O_WRONLY)
+        os.environ["MLX_JACCL_PIPE_IN"] = str(pipe_in_fd)
+        os.environ["MLX_JACCL_PIPE_OUT"] = str(pipe_out_fd)
+
     global logger
     logger = _logger
 
@@ -56,7 +67,9 @@ def entrypoint(
         try:
             event_sender.close()
             task_receiver.close()
+            cancel_receiver.close()
         finally:
             event_sender.join()
             task_receiver.join()
+            cancel_receiver.join()
             logger.info("bye from the runner")

src/exo/worker/runner/runner.py

@@ -42,6 +42,7 @@ from exo.shared.types.tasks import (
     TaskId,
     TaskStatus,
     TextGeneration,
+    TransferModelToDisk,
 )
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import BoundInstance
@@ -81,6 +82,11 @@ from exo.worker.engines.image import (
 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, warmup_inference
+from exo.worker.engines.mlx.model_transfer import (
+    coordinate_transfer,
+    model_path_for_id,
+    transfer_all_files,
+)
 from exo.worker.engines.mlx.utils_mlx import (
     apply_chat_template,
     detect_thinking_prompt_suffix,
@@ -201,7 +207,10 @@ def main(
 
                     if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
                         inference_model, tokenizer = load_mlx_items(
-                            bound_instance, group, on_timeout=on_model_load_timeout
+                            bound_instance,
+                            group,
+                            on_timeout=on_model_load_timeout,
+                            has_local_model=task.has_local_model,
                         )
                         logger.info(
                             f"model has_tool_calling={tokenizer.has_tool_calling} using tokens {tokenizer.tool_call_start}, {tokenizer.tool_call_end}"
@@ -243,7 +252,7 @@ def main(
                         assert inference_model
                         assert tokenizer
 
-                        t = time.monotonic()
+                        t = time.perf_counter()
                         toks = warmup_inference(
                             model=inference_model,
                             tokenizer=tokenizer,
@@ -251,7 +260,7 @@ def main(
                         )
                         logger.info(f"warmed up by generating {toks} tokens")
                         check_for_cancel_every = min(
-                            math.ceil(toks / min(time.monotonic() - t, 0.001)), 100
+                            math.ceil(toks / max(time.perf_counter() - t, 0.001)), 100
                         )
                         if group is not None:
                             check_for_cancel_every = int(
@@ -535,6 +544,27 @@ def main(
 
                     current_status = RunnerReady()
                     logger.info("runner ready")
+                case TransferModelToDisk() if (
+                    isinstance(current_status, RunnerConnected) and group is not None
+                ):
+                    logger.info("starting disk-to-disk model transfer")
+                    event_sender.send(TaskAcknowledged(task_id=task.task_id))
+
+                    model_path = model_path_for_id(
+                        task.shard_metadata.model_card.model_id
+                    )
+                    _, source_rank = coordinate_transfer(group, task.has_local_model)
+                    is_source = group.rank() == source_rank
+                    transfer_all_files(model_path, group, is_source)
+
+                    logger.info("disk-to-disk model transfer complete")
+                    current_status = RunnerShuttingDown()
+                    event_sender.send(
+                        RunnerStatusUpdated(
+                            runner_id=runner_id, runner_status=current_status
+                        )
+                    )
+                    current_status = RunnerShutdown()
                 case Shutdown():
                     current_status = RunnerShuttingDown()
                     logger.info("runner shutting down")

src/exo/worker/runner/runner_supervisor.py

@@ -1,6 +1,10 @@
 import contextlib
+import os
 import signal
+import struct
+import tempfile
 from dataclasses import dataclass, field
+from functools import partial
 from multiprocessing import Process
 from typing import Self
 
@@ -14,12 +18,14 @@ from loguru import logger
 
 from exo.shared.types.events import (
     Event,
+    JacclSideChannelData,
+    JacclSideChannelGathered,
     RunnerStatusUpdated,
     TaskAcknowledged,
     TaskStatusUpdated,
 )
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
-from exo.shared.types.worker.instances import BoundInstance
+from exo.shared.types.worker.instances import BoundInstance, MlxJacclInstance
 from exo.shared.types.worker.runners import (
     RunnerConnecting,
     RunnerFailed,
@@ -34,6 +40,26 @@ from exo.shared.types.worker.shards import ShardMetadata
 from exo.utils.channels import MpReceiver, MpSender, Sender, mp_channel
 from exo.worker.runner.bootstrap import entrypoint
 
+
+def _pipe_read_exact(fd: int, n: int) -> bytes | None:
+    """Read exactly n bytes from a file descriptor. Returns None on EOF."""
+    data = b""
+    while len(data) < n:
+        chunk = os.read(fd, n - len(data))
+        if not chunk:
+            return None
+        data += chunk
+    return data
+
+
+def _pipe_write_all(fd: int, data: bytes) -> None:
+    """Write all bytes to a file descriptor."""
+    view = memoryview(data)
+    while view:
+        written = os.write(fd, view)
+        view = view[written:]
+
+
 PREFILL_TIMEOUT_SECONDS = 60
 DECODE_TIMEOUT_SECONDS = 5
 
@@ -46,12 +72,21 @@ class RunnerSupervisor:
     initialize_timeout: float
     _ev_recv: MpReceiver[Event]
     _task_sender: MpSender[Task]
-    _event_sender: Sender[Event]
     _cancel_sender: MpSender[TaskId]
+    _event_sender: Sender[Event]
+    _pipe_read_fd: int | None = None  # Python reads runner's pipe output
+    _pipe_write_fd: int | None = None  # Python writes gathered data to runner
+    _child_pipe_fds: tuple[int, int] | None = None  # fds to close after fork
+    _fifo_dir: str | None = None  # Temp dir for FIFO files (for cleanup)
+    _fifo_c2p: str | None = None  # FIFO path: C++ writes → Python reads
+    _fifo_p2c: str | None = None  # FIFO path: Python writes → C++ reads
     status: RunnerStatus = field(default_factory=RunnerIdle, init=False)
     pending: dict[TaskId, anyio.Event] = field(default_factory=dict, init=False)
     completed: set[TaskId] = field(default_factory=set, init=False)
     cancelled: set[TaskId] = field(default_factory=set, init=False)
+    _gathered_waiters: dict[
+        int, tuple[anyio.Event, JacclSideChannelGathered | None]
+    ] = field(default_factory=dict, init=False)
 
     @classmethod
     def create(
@@ -65,6 +100,23 @@ class RunnerSupervisor:
         task_sender, task_recv = mp_channel[Task]()
         cancel_sender, cancel_recv = mp_channel[TaskId]()
 
+        # For MlxJaccl instances, create named pipes (FIFOs) for SideChannel relay.
+        # Named pipes work across multiprocessing.Process spawn (macOS default).
+        # FIFO c2p: C++ writes local data → Python reads it
+        # FIFO p2c: Python writes gathered data → C++ reads it
+        fifo_dir: str | None = None
+        fifo_c2p: str | None = None
+        fifo_p2c: str | None = None
+        pipe_fifo_paths: tuple[str, str] | None = None
+
+        if isinstance(bound_instance.instance, MlxJacclInstance):
+            fifo_dir = tempfile.mkdtemp(prefix="exo_jaccl_")
+            fifo_c2p = os.path.join(fifo_dir, "c2p")  # C++ → Python
+            fifo_p2c = os.path.join(fifo_dir, "p2c")  # Python → C++
+            os.mkfifo(fifo_c2p)
+            os.mkfifo(fifo_p2c)
+            pipe_fifo_paths = (fifo_c2p, fifo_p2c)
+
         runner_process = Process(
             target=entrypoint,
             args=(
@@ -73,6 +125,7 @@ class RunnerSupervisor:
                 task_recv,
                 cancel_recv,
                 logger,
+                pipe_fifo_paths,
             ),
             daemon=True,
         )
@@ -88,21 +141,54 @@ class RunnerSupervisor:
             _task_sender=task_sender,
             _cancel_sender=cancel_sender,
             _event_sender=event_sender,
+            _fifo_dir=fifo_dir,
+            _fifo_c2p=fifo_c2p,
+            _fifo_p2c=fifo_p2c,
         )
 
         return self
 
     async def run(self):
         self.runner_process.start()
-        await self._forward_events()
+
+        if self._fifo_c2p is not None and self._fifo_p2c is not None:
+            # Open FIFOs from parent side. These block until child opens the other end,
+            # so we run them in threads concurrently to avoid deadlock.
+            fifo_c2p = self._fifo_c2p
+            fifo_p2c = self._fifo_p2c
+
+            async def open_read() -> None:
+                self._pipe_read_fd = await to_thread.run_sync(
+                    partial(os.open, fifo_c2p, os.O_RDONLY)
+                )
+
+            async def open_write() -> None:
+                self._pipe_write_fd = await to_thread.run_sync(
+                    partial(os.open, fifo_p2c, os.O_WRONLY)
+                )
+
+            async with anyio.create_task_group() as open_tg:
+                open_tg.start_soon(open_read)
+                open_tg.start_soon(open_write)
+
+            logger.info(
+                f"JACCL pipe relay: FIFOs opened (read_fd={self._pipe_read_fd}, write_fd={self._pipe_write_fd})"
+            )
+
+            async with anyio.create_task_group() as tg:
+                tg.start_soon(self._pipe_relay)
+                tg.start_soon(self._forward_events)
+        else:
+            await self._forward_events()
 
     def shutdown(self):
         logger.info("Runner supervisor shutting down")
         self._ev_recv.close()
         self._task_sender.close()
-        self._event_sender.close()
         self._cancel_sender.send(TaskId("CANCEL_CURRENT_TASK"))
         self._cancel_sender.close()
+        self._event_sender.close()
+        self._close_pipe_fds()
         self.runner_process.join(1)
         if not self.runner_process.is_alive():
             logger.info("Runner process succesfully terminated")
@@ -140,6 +226,7 @@ class RunnerSupervisor:
         await event.wait()
 
     async def cancel_task(self, task_id: TaskId):
+        """Send a cancellation signal to the runner process."""
         if task_id in self.completed:
             logger.info(f"Unable to cancel {task_id} as it has been completed")
             return
@@ -181,6 +268,110 @@ class RunnerSupervisor:
                 for tid in self.pending:
                     self.pending[tid].set()
 
+    def _close_pipe_fds(self) -> None:
+        if self._pipe_read_fd is not None:
+            with contextlib.suppress(OSError):
+                os.close(self._pipe_read_fd)
+            self._pipe_read_fd = None
+        if self._pipe_write_fd is not None:
+            with contextlib.suppress(OSError):
+                os.close(self._pipe_write_fd)
+            self._pipe_write_fd = None
+        if self._child_pipe_fds is not None:
+            for fd in self._child_pipe_fds:
+                with contextlib.suppress(OSError):
+                    os.close(fd)
+            self._child_pipe_fds = None
+        # Clean up FIFO files
+        if self._fifo_c2p is not None:
+            with contextlib.suppress(OSError):
+                os.unlink(self._fifo_c2p)
+            self._fifo_c2p = None
+        if self._fifo_p2c is not None:
+            with contextlib.suppress(OSError):
+                os.unlink(self._fifo_p2c)
+            self._fifo_p2c = None
+        if self._fifo_dir is not None:
+            with contextlib.suppress(OSError):
+                os.rmdir(self._fifo_dir)
+            self._fifo_dir = None
+
+    async def _pipe_relay(self) -> None:
+        """Relay JACCL SideChannel all_gather rounds between runner pipes and exo events."""
+        assert self._pipe_read_fd is not None
+        assert self._pipe_write_fd is not None
+        read_fd = self._pipe_read_fd
+        write_fd = self._pipe_write_fd
+        sequence = 0
+
+        try:
+            while True:
+                # 1. Read local data from runner: [uint32 size][size bytes]
+                header = await to_thread.run_sync(partial(_pipe_read_exact, read_fd, 4))
+                if header is None:
+                    logger.info("JACCL pipe relay: runner closed pipe (EOF)")
+                    break
+                data_size: int = struct.unpack("<I", header)[0]  # pyright: ignore[reportAny]
+                local_data = await to_thread.run_sync(
+                    partial(_pipe_read_exact, read_fd, data_size)
+                )
+                if local_data is None:
+                    logger.warning("JACCL pipe relay: EOF reading data payload")
+                    break
+
+                logger.info(
+                    f"JACCL pipe relay: read {data_size} bytes from runner, seq={sequence}"
+                )
+
+                # 2. Emit JacclSideChannelData event
+                waiter = anyio.Event()
+                self._gathered_waiters[sequence] = (waiter, None)
+                await self._event_sender.send(
+                    JacclSideChannelData(
+                        instance_id=self.bound_instance.instance.instance_id,
+                        runner_id=self.bound_instance.bound_runner_id,
+                        sequence=sequence,
+                        data=local_data,
+                    )
+                )
+
+                # 3. Wait for gathered result
+                await waiter.wait()
+                _, gathered_event = self._gathered_waiters.pop(sequence)
+                assert gathered_event is not None
+
+                # 4. Order gathered data by runner rank and concatenate
+                instance = self.bound_instance.instance
+                assert isinstance(instance, MlxJacclInstance)
+                runner_order = list(instance.shard_assignments.runner_to_shard.keys())
+                ordered_data = b"".join(
+                    gathered_event.gathered_data[rid] for rid in runner_order
+                )
+
+                # 5. Write gathered data to runner: [uint32 total_size][total_size bytes]
+                total_size = len(ordered_data)
+                response = struct.pack("<I", total_size) + ordered_data
+                await to_thread.run_sync(partial(_pipe_write_all, write_fd, response))
+
+                logger.info(
+                    f"JACCL pipe relay: wrote {total_size} bytes to runner, seq={sequence}"
+                )
+                sequence += 1
+        except OSError as e:
+            logger.warning(f"JACCL pipe relay: OS error: {e}")
+        except Exception as e:
+            logger.opt(exception=e).error("JACCL pipe relay: unexpected error")
+
+    def notify_gathered(self, event: JacclSideChannelGathered) -> None:
+        """Called by the worker when a JacclSideChannelGathered event arrives."""
+        seq = event.sequence
+        if seq not in self._gathered_waiters:
+            logger.warning(f"JACCL: received gathered event for unknown sequence {seq}")
+            return
+        waiter, _ = self._gathered_waiters[seq]
+        self._gathered_waiters[seq] = (waiter, event)
+        waiter.set()
+
     def __del__(self) -> None:
         if self.runner_process.is_alive():
             logger.warning("RunnerSupervisor was not stopped cleanly.")

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

@@ -24,7 +24,6 @@ from exo.worker.engines.mlx.utils_mlx import (
 
 # Files needed for tokenizer functionality
 TOKENIZER_FILE_PATTERNS = [
-    "config.json",
     "tokenizer.json",
     "tokenizer_config.json",
     "special_tokens_map.json",
@@ -339,9 +338,6 @@ async def test_kimi_tokenizer_specifically():
     # Verify EOS token is set
     assert eos_token_ids == [163586], "Kimi EOS token should be [163586]"
 
-    # Verify architecture-based detection gives same result
-    assert get_eos_token_ids_for_model(model_id, model_type="kimi") == [163586]
-
 
 # Test GLM tokenizer since it also has special handling
 @pytest.mark.asyncio
@@ -382,10 +378,3 @@ async def test_glm_tokenizer_specifically():
         151329,
         151338,
     ], "GLM EOS tokens should be correct"
-
-    # Verify architecture-based detection gives same result
-    assert get_eos_token_ids_for_model(model_id, model_type="glm4") == [
-        151336,
-        151329,
-        151338,
-    ]

src/exo/worker/tests/unittests/test_plan/test_download_and_loading.py

@@ -112,6 +112,7 @@ def test_plan_loads_model_when_all_shards_downloaded_and_waiting():
 
     assert isinstance(result, LoadModel)
     assert result.instance_id == INSTANCE_1_ID
+    assert result.has_local_model is True
 
 
 def test_plan_does_not_request_download_when_shard_already_downloaded():
@@ -157,10 +158,11 @@ def test_plan_does_not_request_download_when_shard_already_downloaded():
     assert not isinstance(result, plan_mod.DownloadModel)
 
 
-def test_plan_does_not_load_model_until_all_shards_downloaded_globally():
+def test_plan_loads_model_when_any_node_has_download_for_multi_node():
     """
-    LoadModel should not be emitted while some shards are still missing from
-    the global_download_status.
+    For multi-node instances, LoadModel should be emitted when at least one
+    node has the model downloaded. Nodes without the model will receive it
+    via MLX distributed transfer during model loading.
     """
     shard1 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=0, world_size=2)
     shard2 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=1, world_size=2)
@@ -185,6 +187,7 @@ def test_plan_does_not_load_model_until_all_shards_downloaded_globally():
         RUNNER_2_ID: RunnerConnected(),
     }
 
+    # Only NODE_A has the model — LoadModel should still fire
     global_download_status = {
         NODE_A: [
             DownloadCompleted(
@@ -203,19 +206,42 @@ def test_plan_does_not_load_model_until_all_shards_downloaded_globally():
         tasks={},
     )
 
-    assert result is None
+    assert isinstance(result, LoadModel)
+    assert result.instance_id == INSTANCE_1_ID
+    assert result.has_local_model is True
 
-    global_download_status = {
-        NODE_A: [
-            DownloadCompleted(
-                shard_metadata=shard1, node_id=NODE_A, total_bytes=Memory()
-            )
-        ],
-        NODE_B: [
-            DownloadCompleted(
-                shard_metadata=shard2, node_id=NODE_B, total_bytes=Memory()
-            )
-        ],  # NODE_B has no downloads completed yet
+
+def test_plan_does_not_load_model_when_no_node_has_download():
+    """
+    LoadModel should not be emitted when no node has the model downloaded.
+    """
+    shard1 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=0, world_size=2)
+    shard2 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=1, world_size=2)
+    instance = get_mlx_ring_instance(
+        instance_id=INSTANCE_1_ID,
+        model_id=MODEL_A_ID,
+        node_to_runner={NODE_A: RUNNER_1_ID, NODE_B: RUNNER_2_ID},
+        runner_to_shard={RUNNER_1_ID: shard1, RUNNER_2_ID: shard2},
+    )
+
+    bound_instance = BoundInstance(
+        instance=instance, bound_runner_id=RUNNER_1_ID, bound_node_id=NODE_A
+    )
+    local_runner = FakeRunnerSupervisor(
+        bound_instance=bound_instance, status=RunnerConnected()
+    )
+
+    runners = {RUNNER_1_ID: local_runner}
+    instances = {INSTANCE_1_ID: instance}
+    all_runners = {
+        RUNNER_1_ID: RunnerConnected(),
+        RUNNER_2_ID: RunnerConnected(),
+    }
+
+    # No node has the model
+    global_download_status: dict[NodeId, list[DownloadProgress]] = {
+        NODE_A: [],
+        NODE_B: [],
     }
 
     result = plan_mod.plan(
@@ -227,4 +253,57 @@ def test_plan_does_not_load_model_until_all_shards_downloaded_globally():
         tasks={},
     )
 
-    assert result is not None
+    assert result is None
+
+
+def test_plan_load_model_has_local_model_false_when_node_missing_download():
+    """
+    For multi-node instances, when the local node does NOT have the model
+    but a peer does, LoadModel should be emitted with has_local_model=False.
+    """
+    shard1 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=0, world_size=2)
+    shard2 = get_pipeline_shard_metadata(MODEL_A_ID, device_rank=1, world_size=2)
+    instance = get_mlx_ring_instance(
+        instance_id=INSTANCE_1_ID,
+        model_id=MODEL_A_ID,
+        node_to_runner={NODE_A: RUNNER_1_ID, NODE_B: RUNNER_2_ID},
+        runner_to_shard={RUNNER_1_ID: shard1, RUNNER_2_ID: shard2},
+    )
+
+    # NODE_B is the local node (bound_node_id=NODE_B), it does NOT have the model
+    bound_instance = BoundInstance(
+        instance=instance, bound_runner_id=RUNNER_2_ID, bound_node_id=NODE_B
+    )
+    local_runner = FakeRunnerSupervisor(
+        bound_instance=bound_instance, status=RunnerConnected()
+    )
+
+    runners = {RUNNER_2_ID: local_runner}
+    instances = {INSTANCE_1_ID: instance}
+    all_runners = {
+        RUNNER_1_ID: RunnerConnected(),
+        RUNNER_2_ID: RunnerConnected(),
+    }
+
+    # Only NODE_A has the model, NODE_B does not
+    global_download_status: dict[NodeId, list[DownloadProgress]] = {
+        NODE_A: [
+            DownloadCompleted(
+                shard_metadata=shard1, node_id=NODE_A, total_bytes=Memory()
+            )
+        ],
+        NODE_B: [],
+    }
+
+    result = plan_mod.plan(
+        node_id=NODE_B,
+        runners=runners,  # type: ignore
+        global_download_status=global_download_status,
+        instances=instances,
+        all_runners=all_runners,
+        tasks={},
+    )
+
+    assert isinstance(result, LoadModel)
+    assert result.instance_id == INSTANCE_1_ID
+    assert result.has_local_model is False

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

@@ -1,9 +1,7 @@
 # Check tasks are complete before runner is ever ready.
-import unittest.mock
 from collections.abc import Iterable
 from typing import Callable
 
-import mlx.core as mx
 import pytest
 
 import exo.worker.runner.runner as mlx_runner
@@ -117,6 +115,12 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
     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))
+
+    # Mock mx.distributed.all_gather so MockGroup doesn't hit real MLX C++ bindings.
+    def _mock_all_gather(x: object, **_kw: object) -> object:
+        return x
+
+    monkeypatch.setattr(mlx_runner.mx.distributed, "all_gather", _mock_all_gather)
     # 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"))
@@ -178,16 +182,15 @@ def _run(tasks: Iterable[Task]):
         # this is some c++ nonsense
         task_receiver.close = nothin
         task_receiver.join = nothin
-        with unittest.mock.patch(
-            "exo.worker.runner.runner.mx.distributed.all_gather",
-            make_nothin(mx.array([1])),
-        ):
-            mlx_runner.main(
-                bound_instance,
-                event_sender,  # pyright: ignore[reportArgumentType]
-                task_receiver,
-                cancel_receiver,
-            )
+        cancel_receiver.close = nothin
+        cancel_receiver.join = nothin
+
+        mlx_runner.main(
+            bound_instance,
+            event_sender,  # pyright: ignore[reportArgumentType]
+            task_receiver,
+            cancel_receiver,
+        )
 
         return event_sender.events
 

uv.lock

@@ -377,8 +377,8 @@ dependencies = [
     { name = "hypercorn", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "loguru", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", extra = ["cpu"], marker = "sys_platform == 'linux'" },
+    { name = "mlx", version = "0.30.6", source = { registry = "https://pypi.org/simple" }, extra = ["cpu"], marker = "sys_platform == 'linux'" },
+    { name = "mlx", version = "0.30.7.dev20260217+50487b41", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#50487b4141f3c951122655db3b83df5146c1fbeb" }, marker = "sys_platform == 'darwin'" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "msgspec", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "openai-harmony", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -416,7 +416,7 @@ requires-dist = [
     { name = "hypercorn", specifier = ">=0.18.0" },
     { name = "loguru", specifier = ">=0.7.3" },
     { name = "mflux", specifier = "==0.15.5" },
-    { name = "mlx", marker = "sys_platform == 'darwin'", specifier = "==0.30.6" },
+    { name = "mlx", marker = "sys_platform == 'darwin'", git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks" },
     { name = "mlx", extras = ["cpu"], marker = "sys_platform == 'linux'", specifier = "==0.30.6" },
     { name = "mlx-lm", specifier = "==0.30.6" },
     { name = "msgspec", specifier = ">=0.19.0" },
@@ -1020,8 +1020,8 @@ dependencies = [
     { name = "fonttools", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "matplotlib", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", extra = ["cuda13"], marker = "sys_platform == 'linux'" },
+    { name = "mlx", version = "0.30.6", source = { registry = "https://pypi.org/simple" }, extra = ["cuda13"], marker = "sys_platform == 'linux'" },
+    { name = "mlx", version = "0.30.7.dev20260217+50487b41", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#50487b4141f3c951122655db3b83df5146c1fbeb" }, marker = "sys_platform == 'darwin'" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "opencv-python", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "piexif", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -1048,18 +1048,12 @@ wheels = [
 name = "mlx"
 version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "mlx-metal", marker = "sys_platform == 'darwin'" },
+resolution-markers = [
+    "sys_platform == 'linux'",
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/ae/5b/e460e144a34d5529e010056cccf50b538d56ed001473bc6b246018fd58cb/mlx-0.30.6-cp313-cp313-macosx_14_0_arm64.whl", hash = "sha256:ed86f8bffc174c2f259ca589ea25464c96cf69d1bb457074a2bf2ef53737e54f", size = 573515, upload-time = "2026-02-06T03:45:23.405Z" },
-    { url = "https://files.pythonhosted.org/packages/60/25/69833fefb9a3fef30b56792b1bcd022496c4fea83e45411d289b77ef7546/mlx-0.30.6-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:c52294958269e20f300639a17c1900ca8fc737d859ddda737f9811e94bd040e5", size = 573516, upload-time = "2026-02-06T03:45:24.618Z" },
-    { url = "https://files.pythonhosted.org/packages/9c/6a/7e7fbeebc5cb51b6a5eba96b263a6298707bcbdc059f4b0b73e088bc3dea/mlx-0.30.6-cp313-cp313-macosx_26_0_arm64.whl", hash = "sha256:b5b6636f7c49a4d86d8ec82643b972f45a144a7a9f3a967b27b2e6e22cf71e6a", size = 573592, upload-time = "2026-02-06T03:45:25.928Z" },
     { url = "https://files.pythonhosted.org/packages/93/06/280f6f2ba80520a7109730425eda0d966658793aa0d02d8be8d351f75253/mlx-0.30.6-cp313-cp313-manylinux_2_35_aarch64.whl", hash = "sha256:67e6c9e30a9faeacc209917ef5523177cf9b086914b6b5d83ff886e4294b727d", size = 622011, upload-time = "2026-02-06T03:45:28.165Z" },
     { url = "https://files.pythonhosted.org/packages/fe/35/f872afbee9c079cc69924d9e9c46f5663adb7da58cba3511db082dd307c1/mlx-0.30.6-cp313-cp313-manylinux_2_35_x86_64.whl", hash = "sha256:47db8b16fcb6f6c5a47c0bdb24ed377b41237017ac93aa6cb6aa206c9bdf82e4", size = 663650, upload-time = "2026-02-06T03:45:30.315Z" },
-    { url = "https://files.pythonhosted.org/packages/60/23/361dc7a5797634e4d7e9bdd6564c6b28f9b1246672632def2f91bf066b18/mlx-0.30.6-cp314-cp314-macosx_14_0_arm64.whl", hash = "sha256:78804a89dcff4a838f7c2da72392fe87a523e95122a3c840e53df019122aad45", size = 575028, upload-time = "2026-02-06T03:45:31.549Z" },
-    { url = "https://files.pythonhosted.org/packages/a8/69/1854484d414171586814dfbe8def95f75c4ea2c7341ba13ba8ee675f7c62/mlx-0.30.6-cp314-cp314-macosx_15_0_arm64.whl", hash = "sha256:ec13584ab069665cc7ad34a05494d9291cd623aef6ae96be48875fc87cfc25d6", size = 575026, upload-time = "2026-02-06T03:45:33.072Z" },
-    { url = "https://files.pythonhosted.org/packages/6b/b8/3adbc441924209a7e4c568308b2a0b54bd09aee6a68db5bae85304791e54/mlx-0.30.6-cp314-cp314-macosx_26_0_arm64.whl", hash = "sha256:b2c5e8a090a753ef99a1380a4d059c983083f36198864f6df9faaf1223d083df", size = 575041, upload-time = "2026-02-06T03:45:34.814Z" },
     { url = "https://files.pythonhosted.org/packages/3f/54/9d9e06804fb2088202a2cdf60458e00b221f71420bea285720b60f9e82b5/mlx-0.30.6-cp314-cp314-manylinux_2_35_aarch64.whl", hash = "sha256:9ceddede4af0de31d1f6b3099f70e5469d60cd7c546975dedbdbeab3519cab3f", size = 624002, upload-time = "2026-02-06T03:45:36Z" },
     { url = "https://files.pythonhosted.org/packages/42/92/3140a15a50cb1f9267a6552171e1dfa577861de53e093124bc43707f2a0e/mlx-0.30.6-cp314-cp314-manylinux_2_35_x86_64.whl", hash = "sha256:4a6ffd2d16728cf95f63a1b555d7c2eaeea686a0e6b73228bd265411cb5d77a4", size = 663569, upload-time = "2026-02-06T03:45:37.242Z" },
 ]
@@ -1072,6 +1066,14 @@ cuda13 = [
     { name = "mlx-cuda-13", marker = "sys_platform == 'linux'" },
 ]
 
+[[package]]
+name = "mlx"
+version = "0.30.7.dev20260217+50487b41"
+source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#50487b4141f3c951122655db3b83df5146c1fbeb" }
+resolution-markers = [
+    "sys_platform == 'darwin'",
+]
+
 [[package]]
 name = "mlx-cpu"
 version = "0.30.6"
@@ -1102,7 +1104,7 @@ version = "0.30.6"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.30.7.dev20260217+50487b41", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#50487b4141f3c951122655db3b83df5146c1fbeb" }, marker = "sys_platform == 'darwin'" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "protobuf", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "pyyaml", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -1114,16 +1116,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/20/5f/01d281f1fa8a1521d5936659beb4f5ab1f32b463d059263cf9d4cef969d9/mlx_lm-0.30.6-py3-none-any.whl", hash = "sha256:a7405bd581eacc4bf8209d7a6b7f23629585a0d7c6740c2a97e51fee35b3b0e1", size = 379451, upload-time = "2026-02-04T21:27:43.222Z" },
 ]
 
-[[package]]
-name = "mlx-metal"
-version = "0.30.6"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/f3/85/44406b521f920248fad621334d4dc15e77660a494edf890e7cbee33bf38d/mlx_metal-0.30.6-py3-none-macosx_14_0_arm64.whl", hash = "sha256:ea6d0c973def9a5b4f652cc77036237db3f88c9d0af63701d76b5fddde99b820", size = 38437818, upload-time = "2026-02-06T03:44:56.19Z" },
-    { url = "https://files.pythonhosted.org/packages/d0/cb/10a516995f7d0c154b0d7e633c54b51e96977a86a355105b6474cfcbe0d0/mlx_metal-0.30.6-py3-none-macosx_15_0_arm64.whl", hash = "sha256:0f8cb94634d07e06a372d6ad9a090f38a18bab1ff19a140aede60eacf707bb94", size = 38433701, upload-time = "2026-02-06T03:44:59.678Z" },
-    { url = "https://files.pythonhosted.org/packages/4c/7d/70cb272f7373c334709f210ed8420511fc9d64d05a7a646c0b3b94c29c04/mlx_metal-0.30.6-py3-none-macosx_26_0_arm64.whl", hash = "sha256:d761ae26304f2c4b454eeea7f612a56919d9e5e57dbb1dc0788f8e34aa6f41c2", size = 47718448, upload-time = "2026-02-06T03:45:03.133Z" },
-]
-
 [[package]]
 name = "more-itertools"
 version = "10.8.0"