sloppin it up

## Motivation

Clicking on the mini topology view in the right sidebar during chat was
selecting/filtering nodes instead of navigating back to the home view.
The entire mini topology section is wrapped in a button that calls
`handleGoHome()`, but individual node clicks were intercepting the event
via `stopPropagation()`.

## Changes

- Removed `onNodeClick={togglePreviewNodeFilter}` from the mini
sidebar's `TopologyGraph` component
- Added `pointer-events-none` to the topology graph container so all
clicks pass through to the parent button

## Why It Works

`TopologyGraph` only registers click/hover handlers when `onNodeClick`
is provided. Removing it disables node interaction entirely in the mini
view. The `pointer-events-none` class ensures no SVG element can
intercept clicks — they all bubble up to the parent `<button
onclick={handleGoHome}>`, which resets the chat state and returns to the
main topology view.

## Test Plan

### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
- Start exo and open the dashboard
- Start a chat to trigger the mini topology sidebar
- Click on the mini topology view → should navigate back to home
- Confirm nodes in the mini topology are not selectable or hoverable

### Automated Testing
- Dashboard builds successfully (`npm run build`)

.mlx_typings/mlx_lm/generate.pyi

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

.mlx_typings/mlx_lm/models/cache.pyi

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

Cargo.lock

@@ -216,6 +216,28 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
+[[package]]
+name = "async-stream"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b5a71a6f37880a80d1d7f19efd781e4b5de42c88f0722cc13bcb6cc2cfe8476"
+dependencies = [
+ "async-stream-impl",
+ "futures-core",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "async-stream-impl"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c7c24de15d275a1ecfd47a380fb4d5ec9bfe0933f309ed5e705b775596a3574d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "async-trait"
 version = "0.1.89"
@@ -2759,6 +2781,7 @@ dependencies = [
 name = "networking"
 version = "0.0.1"
 dependencies = [
+ "async-stream",
  "delegate",
  "either",
  "extend",
@@ -2767,6 +2790,7 @@ dependencies = [
  "keccak-const",
  "libp2p",
  "log",
+ "pin-project",
  "tokio",
  "tracing-subscriber",
  "util",

Cargo.toml

@@ -1,10 +1,6 @@
 [workspace]
 resolver = "3"
-members = [
-    "rust/networking",
-    "rust/exo_pyo3_bindings",
-    "rust/util",
-]
+members = ["rust/networking", "rust/exo_pyo3_bindings", "rust/util"]
 
 [workspace.package]
 version = "0.0.1"
@@ -34,6 +30,7 @@ delegate = "0.13"
 keccak-const = "0.2"
 
 # Async dependencies
+async-stream = "0.3"
 tokio = "1.46"
 futures-lite = "2.6.1"
 futures-timer = "3.0"

app/EXO/EXO/ExoProcessController.swift

@@ -5,6 +5,7 @@ import Foundation
 private let customNamespaceKey = "EXOCustomNamespace"
 private let hfTokenKey = "EXOHFToken"
 private let enableImageModelsKey = "EXOEnableImageModels"
+private let offlineModeKey = "EXOOfflineMode"
 private let onboardingCompletedKey = "EXOOnboardingCompleted"
 
 @MainActor
@@ -60,6 +61,14 @@ final class ExoProcessController: ObservableObject {
             UserDefaults.standard.set(enableImageModels, forKey: enableImageModelsKey)
         }
     }
+    @Published var offlineMode: Bool = {
+        return UserDefaults.standard.bool(forKey: offlineModeKey)
+    }()
+    {
+        didSet {
+            UserDefaults.standard.set(offlineMode, forKey: offlineModeKey)
+        }
+    }
 
     /// Fires once when EXO transitions to `.running` for the very first time (fresh install).
     @Published private(set) var isFirstLaunchReady = false
@@ -267,6 +276,9 @@ final class ExoProcessController: ObservableObject {
         if enableImageModels {
             environment["EXO_ENABLE_IMAGE_MODELS"] = "true"
         }
+        if offlineMode {
+            environment["EXO_OFFLINE"] = "true"
+        }
 
         var paths: [String] = []
         if let existing = environment["PATH"], !existing.isEmpty {

app/EXO/EXO/Views/SettingsView.swift

@@ -13,6 +13,7 @@ struct SettingsView: View {
     @State private var pendingNamespace: String = ""
     @State private var pendingHFToken: String = ""
     @State private var pendingEnableImageModels = false
+    @State private var pendingOfflineMode = false
     @State private var needsRestart = false
     @State private var bugReportInFlight = false
     @State private var bugReportMessage: String?
@@ -42,6 +43,7 @@ struct SettingsView: View {
             pendingNamespace = controller.customNamespace
             pendingHFToken = controller.hfToken
             pendingEnableImageModels = controller.enableImageModels
+            pendingOfflineMode = controller.offlineMode
             needsRestart = false
         }
     }
@@ -72,6 +74,13 @@ struct SettingsView: View {
                     .foregroundColor(.secondary)
             }
 
+            Section {
+                Toggle("Offline Mode", isOn: $pendingOfflineMode)
+                Text("Skip internet checks and use only locally available models.")
+                    .font(.caption)
+                    .foregroundColor(.secondary)
+            }
+
             Section {
                 HStack {
                     Spacer()
@@ -445,6 +454,7 @@ struct SettingsView: View {
 
     private var hasGeneralChanges: Bool {
         pendingNamespace != controller.customNamespace || pendingHFToken != controller.hfToken
+            || pendingOfflineMode != controller.offlineMode
     }
 
     private var hasModelChanges: Bool {
@@ -454,6 +464,7 @@ struct SettingsView: View {
     private func applyGeneralSettings() {
         controller.customNamespace = pendingNamespace
         controller.hfToken = pendingHFToken
+        controller.offlineMode = pendingOfflineMode
         restartIfRunning()
     }
 

bench/bench.toml

@@ -2,6 +2,4 @@
 #
 # Lists the suite files to include. Each file defines benchmarks
 # with shared constraints, topology, and default args.
-include = [
-    "single-m3-ultra.toml",
-]
+include = ["single-m3-ultra.toml"]

bench/exo_bench.py

@@ -75,7 +75,7 @@ def load_tokenizer_for_bench(model_id: str) -> Any:
         model_path = Path(
             snapshot_download(
                 model_id,
-                allow_patterns=["*.json", "*.py", "*.tiktoken"],
+                allow_patterns=["*.json", "*.py", "*.tiktoken", "*.model"],
             )
         )
 

bench/pyproject.toml

@@ -4,13 +4,13 @@ version = "0.1.0"
 description = "Benchmarking tool for exo distributed inference"
 requires-python = ">=3.13"
 dependencies = [
-    "httpx>=0.27.0",
-    "loguru>=0.7.3",
-    "transformers>=5.0.0",
-    "huggingface-hub>=0.33.4",
-    "tiktoken>=0.12.0",
-    "jinja2>=3.1.0",
-    "protobuf>=5.29.0",
+  "httpx>=0.27.0",
+  "loguru>=0.7.3",
+  "transformers>=5.0.0",
+  "huggingface-hub>=0.33.4",
+  "tiktoken>=0.12.0",
+  "jinja2>=3.1.0",
+  "protobuf>=5.29.0",
 ]
 
 [build-system]

bench/single-m3-ultra.toml

@@ -2,10 +2,10 @@
 #
 # Shared constraints applied to ALL benchmarks in this file.
 constraints = [
-    "All(MacOsBuild(=25D125))",
-    "Hosts(=1)",
-    "All(Chip(m3_ultra))",
-    "All(GpuCores(=80))",
+  "All(MacOsBuild(=25D125))",
+  "Hosts(=1)",
+  "All(Chip(m3_ultra))",
+  "All(GpuCores(=80))",
 ]
 
 [topology]

dashboard/src/lib/components/HeaderNav.svelte

@@ -95,7 +95,7 @@
     {#if showHome}
       <button
         onclick={handleHome}
-        class="text-sm text-exo-light-gray hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+        class="text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
         title="Back to topology view"
       >
         <svg
@@ -116,7 +116,7 @@
     {/if}
     <a
       href="/#/downloads"
-      class="text-sm text-exo-light-gray hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+      class="text-sm text-white/70 hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
       title="View downloads overview"
     >
       {#if downloadProgress}

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

@@ -3158,6 +3158,23 @@ class AppStore {
     return (await response.json()) as TraceStatsResponse;
   }
 
+  /**
+   * Delete traces by task IDs
+   */
+  async deleteTraces(
+    taskIds: string[],
+  ): Promise<{ deleted: string[]; notFound: string[] }> {
+    const response = await fetch("/v1/traces/delete", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ taskIds }),
+    });
+    if (!response.ok) {
+      throw new Error(`Failed to delete traces: ${response.status}`);
+    }
+    return await response.json();
+  }
+
   /**
    * Get the URL for the raw trace file (for Perfetto)
    */
@@ -3301,3 +3318,5 @@ export const fetchTraceStats = (taskId: string) =>
   appStore.fetchTraceStats(taskId);
 export const getTraceRawUrl = (taskId: string) =>
   appStore.getTraceRawUrl(taskId);
+export const deleteTraces = (taskIds: string[]) =>
+  appStore.deleteTraces(taskIds);

dashboard/src/routes/+page.svelte

@@ -42,6 +42,7 @@
     setSelectedChatModel,
     selectedChatModel,
     sendMessage,
+    messages,
     debugMode,
     toggleDebugMode,
     topologyOnlyMode,
@@ -889,6 +890,7 @@
       availableModels.some((m) => m.id === defaults.modelId)
     ) {
       selectPreviewModel(defaults.modelId);
+      setSelectedChatModel(defaults.modelId);
     }
   }
 
@@ -1324,6 +1326,7 @@
 
   function handleModelPickerSelect(modelId: string) {
     selectPreviewModel(modelId);
+    setSelectedChatModel(modelId);
     saveLaunchDefaults();
     isModelPickerOpen = false;
   }
@@ -2284,7 +2287,8 @@
     selectedChatCategory = null;
     pendingAutoMessage = null;
     userForcedIdle = true;
-    setSelectedChatModel("");
+    // Restore chat model from the sidebar preview selection so both selectors stay in sync
+    setSelectedChatModel(selectedModelId ?? "");
     clearChat();
   }
 
@@ -2514,6 +2518,11 @@
         return;
       }
     }
+
+    // Fallthrough: model exists but has no active instance/download/loading state
+    chatLaunchState = "idle";
+    pendingChatModelId = null;
+    selectedChatCategory = null;
   });
 
   // Suggested prompts per category
@@ -2637,7 +2646,11 @@
   }
 
   // Launch a model for seamless chat
-  async function launchModelForChat(modelId: string, category: string) {
+  async function launchModelForChat(
+    modelId: string,
+    category: string,
+    skipCreate = false,
+  ) {
     userForcedIdle = false;
     pendingChatModelId = modelId;
     selectedChatCategory = category;
@@ -2645,7 +2658,7 @@
     // Check if already running — skip straight to chat
     if (hasRunningInstance(modelId)) {
       setSelectedChatModel(modelId);
-      createConversation();
+      if (!skipCreate) createConversation();
       chatLaunchState = "ready";
       return;
     }
@@ -2654,7 +2667,7 @@
     if (hasExistingInstance(modelId)) {
       setSelectedChatModel(modelId);
       pendingChatModelId = modelId;
-      createConversation();
+      if (!skipCreate) createConversation();
       const dlStatus = getModelDownloadStatus(modelId);
       if (dlStatus.isDownloading) {
         chatLaunchState = "downloading";
@@ -2707,7 +2720,7 @@
 
       setSelectedChatModel(modelId);
       recordRecentLaunch(modelId);
-      createConversation();
+      if (!skipCreate) createConversation();
       chatLaunchState = "downloading";
     } catch (error) {
       addToast({ type: "error", message: `Network error: ${error}` });
@@ -2993,6 +3006,7 @@
   // Handle model selection from the picker when opened from chat context
   function handleChatPickerSelect(modelId: string) {
     setSelectedChatModel(modelId);
+    selectPreviewModel(modelId);
     userForcedIdle = false;
     isModelPickerOpen = false;
   }
@@ -3012,6 +3026,7 @@
 
     // Model is selected and running — send directly
     if (model && hasRunningInstance(model)) {
+      chatLaunchState = "ready";
       sendMessage(content, files, null);
       return;
     }
@@ -3020,7 +3035,7 @@
     if (model) {
       pendingAutoMessage = { content, files };
       userForcedIdle = false;
-      launchModelForChat(model, "picker");
+      launchModelForChat(model, "picker", messages().length > 0);
       return;
     }
 
@@ -5804,43 +5819,7 @@
           class="flex-1 flex flex-col min-w-0 overflow-hidden"
           in:fade={{ duration: 300, delay: 100 }}
         >
-          {#if chatLaunchState === "idle"}
-            <!-- No running instance: show model selector -->
-            <div
-              class="flex-1 overflow-y-auto flex items-center justify-center px-8 py-6"
-            >
-              <ChatModelSelector
-                models={models.map((m) => ({
-                  id: m.id,
-                  name: m.name ?? "",
-                  base_model: m.base_model ?? "",
-                  storage_size_megabytes: m.storage_size_megabytes ?? 0,
-                  capabilities: m.capabilities ?? [],
-                  family: m.family ?? "",
-                  quantization: m.quantization ?? "",
-                }))}
-                clusterLabel={chatClusterLabel}
-                totalMemoryGB={availableMemoryGB()}
-                onSelect={handleChatModelSelect}
-                onAddModel={handleChatAddModel}
-              />
-            </div>
-            <div
-              class="flex-shrink-0 px-8 pb-6 pt-4 bg-gradient-to-t from-exo-black via-exo-black to-transparent"
-            >
-              <div class="max-w-7xl mx-auto">
-                <ChatForm
-                  placeholder="Ask anything — we'll pick the best model automatically"
-                  showModelSelector={!!bestRunningModelId}
-                  modelDisplayOverride={bestRunningModelId ?? undefined}
-                  modelTasks={modelTasks()}
-                  modelCapabilities={modelCapabilities()}
-                  onAutoSend={handleAutoSend}
-                  onOpenModelPicker={openChatModelPicker}
-                />
-              </div>
-            </div>
-          {:else if chatLaunchState !== "idle" && chatLaunchState !== "ready"}
+          {#if chatLaunchState !== "idle" && chatLaunchState !== "ready"}
             <!-- Model launching/downloading/loading: show progress -->
             <div class="flex-1 flex items-center justify-center px-8 py-6">
               <div class="flex flex-col items-center gap-6 max-w-md w-full">
@@ -5947,8 +5926,8 @@
                 />
               </div>
             </div>
-          {:else}
-            <!-- Normal chat: model is running -->
+          {:else if messages().length > 0 || chatLaunchState === "ready"}
+            <!-- Normal chat: show messages -->
             <div
               class="flex-1 overflow-y-auto px-8 py-6"
               bind:this={chatScrollRef}
@@ -6004,6 +5983,42 @@
                 />
               </div>
             </div>
+          {:else}
+            <!-- No running instance, no messages: show model selector -->
+            <div
+              class="flex-1 overflow-y-auto flex items-center justify-center px-8 py-6"
+            >
+              <ChatModelSelector
+                models={models.map((m) => ({
+                  id: m.id,
+                  name: m.name ?? "",
+                  base_model: m.base_model ?? "",
+                  storage_size_megabytes: m.storage_size_megabytes ?? 0,
+                  capabilities: m.capabilities ?? [],
+                  family: m.family ?? "",
+                  quantization: m.quantization ?? "",
+                }))}
+                clusterLabel={chatClusterLabel}
+                totalMemoryGB={availableMemoryGB()}
+                onSelect={handleChatModelSelect}
+                onAddModel={handleChatAddModel}
+              />
+            </div>
+            <div
+              class="flex-shrink-0 px-8 pb-6 pt-4 bg-gradient-to-t from-exo-black via-exo-black to-transparent"
+            >
+              <div class="max-w-7xl mx-auto">
+                <ChatForm
+                  placeholder="Ask anything — we'll pick the best model automatically"
+                  showModelSelector={!!bestRunningModelId}
+                  modelDisplayOverride={bestRunningModelId ?? undefined}
+                  modelTasks={modelTasks()}
+                  modelCapabilities={modelCapabilities()}
+                  onAutoSend={handleAutoSend}
+                  onOpenModelPicker={openChatModelPicker}
+                />
+              </div>
+            </div>
           {/if}
         </div>
 
@@ -6035,12 +6050,11 @@
               </div>
 
               <div
-                class="relative aspect-square bg-exo-dark-gray rounded-lg overflow-hidden"
+                class="relative aspect-square bg-exo-dark-gray rounded-lg overflow-hidden pointer-events-none"
               >
                 <TopologyGraph
                   highlightedNodes={highlightedNodes()}
                   filteredNodes={nodeFilter}
-                  onNodeClick={togglePreviewNodeFilter}
                 />
 
                 {@render clusterWarningsCompact()}

dashboard/src/routes/downloads/+page.svelte

@@ -412,7 +412,7 @@
                   <div>{col.label}</div>
                   {#if col.diskAvailable != null}
                     <div
-                      class="text-[9px] text-exo-light-gray/60 normal-case tracking-normal mt-0.5"
+                      class="text-[9px] text-white/70 normal-case tracking-normal mt-0.5"
                     >
                       {formatBytes(col.diskAvailable)} free
                     </div>
@@ -436,7 +436,7 @@
                       </div>
                       {#if row.prettyName}
                         <div
-                          class="text-[10px] text-exo-light-gray/60"
+                          class="text-[10px] text-white/60"
                           title={row.modelId}
                         >
                           {row.modelId}
@@ -450,7 +450,7 @@
                       title="View model details"
                     >
                       <svg
-                        class="w-4 h-4 text-white/30 hover:text-white/60"
+                        class="w-4 h-4 text-white/60 hover:text-white/80"
                         viewBox="0 0 24 24"
                         fill="currentColor"
                       >
@@ -469,11 +469,11 @@
                   <td class="px-4 py-3 text-center align-middle">
                     {#if cell.kind === "completed"}
                       <div
-                        class="flex flex-col items-center gap-0.5"
+                        class="flex flex-col items-center gap-1"
                         title="Completed ({formatBytes(cell.totalBytes)})"
                       >
                         <svg
-                          class="w-5 h-5 text-green-400"
+                          class="w-7 h-7 text-green-400"
                           viewBox="0 0 20 20"
                           fill="currentColor"
                         >
@@ -483,18 +483,18 @@
                             clip-rule="evenodd"
                           ></path>
                         </svg>
-                        <span class="text-[10px] text-exo-light-gray/70"
+                        <span class="text-xs text-white/70"
                           >{formatBytes(cell.totalBytes)}</span
                         >
                         <button
                           type="button"
-                          class="text-exo-light-gray/40 hover:text-red-400 transition-colors mt-0.5"
+                          class="text-white/50 hover:text-red-400 transition-colors mt-0.5 cursor-pointer"
                           onclick={() =>
                             deleteDownload(col.nodeId, row.modelId)}
                           title="Delete from this node"
                         >
                           <svg
-                            class="w-3.5 h-3.5"
+                            class="w-5 h-5"
                             viewBox="0 0 20 20"
                             fill="none"
                             stroke="currentColor"
@@ -517,11 +517,11 @@
                           cell.speed,
                         )} - ETA {formatEta(cell.etaMs)}"
                       >
-                        <span class="text-exo-yellow text-xs font-medium"
+                        <span class="text-exo-yellow text-sm font-medium"
                           >{clampPercent(cell.percentage).toFixed(1)}%</span
                         >
                         <div
-                          class="w-14 h-1.5 bg-exo-black/60 rounded-sm overflow-hidden"
+                          class="w-16 h-2 bg-exo-black/60 rounded-sm overflow-hidden"
                         >
                           <div
                             class="h-full bg-gradient-to-r from-exo-yellow to-exo-yellow/70 transition-all duration-300"
@@ -530,25 +530,25 @@
                             ).toFixed(1)}%"
                           ></div>
                         </div>
-                        <span class="text-[9px] text-exo-light-gray/60"
+                        <span class="text-[10px] text-white/70"
                           >{formatSpeed(cell.speed)}</span
                         >
                       </div>
                     {:else if cell.kind === "pending"}
                       <div
-                        class="flex flex-col items-center gap-0.5"
+                        class="flex flex-col items-center gap-1"
                         title={cell.downloaded > 0
-                          ? `${formatBytes(cell.downloaded)} / ${formatBytes(cell.total)} downloaded`
+                          ? `${formatBytes(cell.downloaded)} / ${formatBytes(cell.total)} downloaded (paused)`
                           : "Download pending"}
                       >
                         {#if cell.downloaded > 0 && cell.total > 0}
-                          <span class="text-exo-light-gray/70 text-[10px]"
+                          <span class="text-white/70 text-xs"
                             >{formatBytes(cell.downloaded)} / {formatBytes(
                               cell.total,
                             )}</span
                           >
                           <div
-                            class="w-full h-1 bg-white/10 rounded-full overflow-hidden"
+                            class="w-full h-1.5 bg-white/10 rounded-full overflow-hidden"
                           >
                             <div
                               class="h-full bg-exo-light-gray/40 rounded-full"
@@ -558,21 +558,65 @@
                               ).toFixed(1)}%"
                             ></div>
                           </div>
-                          <span class="text-exo-light-gray/40 text-[9px]"
-                            >paused</span
+                          {#if row.shardMetadata}
+                            <button
+                              type="button"
+                              class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
+                              onclick={() =>
+                                startDownload(col.nodeId, row.shardMetadata!)}
+                              title="Resume download on this node"
+                            >
+                              <svg
+                                class="w-5 h-5"
+                                viewBox="0 0 20 20"
+                                fill="none"
+                                stroke="currentColor"
+                                stroke-width="2"
+                              >
+                                <path
+                                  d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
+                                  stroke-linecap="round"
+                                  stroke-linejoin="round"
+                                ></path>
+                              </svg>
+                            </button>
+                          {:else}
+                            <span class="text-white/50 text-[10px]">paused</span
+                            >
+                          {/if}
+                        {:else if row.shardMetadata}
+                          <button
+                            type="button"
+                            class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
+                            onclick={() =>
+                              startDownload(col.nodeId, row.shardMetadata!)}
+                            title="Start download on this node"
                           >
+                            <svg
+                              class="w-6 h-6"
+                              viewBox="0 0 20 20"
+                              fill="none"
+                              stroke="currentColor"
+                              stroke-width="2"
+                            >
+                              <path
+                                d="M10 3v10m0 0l-3-3m3 3l3-3M3 17h14"
+                                stroke-linecap="round"
+                                stroke-linejoin="round"
+                              ></path>
+                            </svg>
+                          </button>
                         {:else}
-                          <span class="text-exo-light-gray/50 text-sm">...</span
-                          >
+                          <span class="text-white/40 text-sm">...</span>
                         {/if}
                       </div>
                     {:else if cell.kind === "failed"}
                       <div
-                        class="flex flex-col items-center gap-0.5"
+                        class="flex flex-col items-center gap-1"
                         title="Download failed"
                       >
                         <svg
-                          class="w-5 h-5 text-red-400"
+                          class="w-7 h-7 text-red-400"
                           viewBox="0 0 20 20"
                           fill="currentColor"
                         >
@@ -585,13 +629,13 @@
                         {#if row.shardMetadata}
                           <button
                             type="button"
-                            class="text-exo-light-gray/40 hover:text-exo-yellow transition-colors"
+                            class="text-white/50 hover:text-exo-yellow transition-colors cursor-pointer"
                             onclick={() =>
                               startDownload(col.nodeId, row.shardMetadata!)}
                             title="Retry download on this node"
                           >
                             <svg
-                              class="w-3.5 h-3.5"
+                              class="w-5 h-5"
                               viewBox="0 0 20 20"
                               fill="none"
                               stroke="currentColor"
@@ -617,13 +661,13 @@
                         {#if row.shardMetadata}
                           <button
                             type="button"
-                            class="text-exo-light-gray/30 hover:text-exo-yellow transition-colors mt-0.5 opacity-0 group-hover:opacity-100"
+                            class="text-white/50 hover:text-exo-yellow transition-colors mt-0.5 opacity-0 group-hover:opacity-100 cursor-pointer"
                             onclick={() =>
                               startDownload(col.nodeId, row.shardMetadata!)}
                             title="Download to this node"
                           >
                             <svg
-                              class="w-3.5 h-3.5"
+                              class="w-5 h-5"
                               viewBox="0 0 20 20"
                               fill="none"
                               stroke="currentColor"

dashboard/src/routes/traces/+page.svelte

@@ -3,6 +3,7 @@
   import {
     listTraces,
     getTraceRawUrl,
+    deleteTraces,
     type TraceListItem,
   } from "$lib/stores/app.svelte";
   import HeaderNav from "$lib/components/HeaderNav.svelte";
@@ -10,6 +11,51 @@
   let traces = $state<TraceListItem[]>([]);
   let loading = $state(true);
   let error = $state<string | null>(null);
+  let selectedIds = $state<Set<string>>(new Set());
+  let deleting = $state(false);
+
+  let allSelected = $derived(
+    traces.length > 0 && selectedIds.size === traces.length,
+  );
+
+  function toggleSelect(taskId: string) {
+    const next = new Set(selectedIds);
+    if (next.has(taskId)) {
+      next.delete(taskId);
+    } else {
+      next.add(taskId);
+    }
+    selectedIds = next;
+  }
+
+  function toggleSelectAll() {
+    if (allSelected) {
+      selectedIds = new Set();
+    } else {
+      selectedIds = new Set(traces.map((t) => t.taskId));
+    }
+  }
+
+  async function handleDelete() {
+    if (selectedIds.size === 0) return;
+    const count = selectedIds.size;
+    if (
+      !confirm(
+        `Delete ${count} trace${count === 1 ? "" : "s"}? This cannot be undone.`,
+      )
+    )
+      return;
+    deleting = true;
+    try {
+      await deleteTraces([...selectedIds]);
+      selectedIds = new Set();
+      await refresh();
+    } catch (e) {
+      error = e instanceof Error ? e.message : "Failed to delete traces";
+    } finally {
+      deleting = false;
+    }
+  }
 
   function formatBytes(bytes: number): string {
     if (!bytes || bytes <= 0) return "0B";
@@ -109,6 +155,16 @@
         </h1>
       </div>
       <div class="flex items-center gap-3">
+        {#if selectedIds.size > 0}
+          <button
+            type="button"
+            class="text-xs font-mono text-red-400 hover:text-red-300 transition-colors uppercase border border-red-500/40 px-2 py-1 rounded"
+            onclick={handleDelete}
+            disabled={deleting}
+          >
+            {deleting ? "Deleting..." : `Delete (${selectedIds.size})`}
+          </button>
+        {/if}
         <button
           type="button"
           class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"
@@ -143,14 +199,41 @@
       </div>
     {:else}
       <div class="space-y-3">
+        <div class="flex items-center gap-2 px-1">
+          <button
+            type="button"
+            class="text-xs font-mono uppercase transition-colors {allSelected
+              ? 'text-exo-yellow'
+              : 'text-exo-light-gray hover:text-exo-yellow'}"
+            onclick={toggleSelectAll}
+          >
+            {allSelected ? "Deselect all" : "Select all"}
+          </button>
+        </div>
         {#each traces as trace}
+          {@const isSelected = selectedIds.has(trace.taskId)}
+          <!-- svelte-ignore a11y_no_static_element_interactions -->
           <div
-            class="rounded border border-exo-medium-gray/30 bg-exo-black/30 p-4 flex items-center justify-between gap-4"
+            role="button"
+            tabindex="0"
+            class="w-full text-left rounded border-l-2 border-r border-t border-b transition-all p-4 flex items-center justify-between gap-4 cursor-pointer {isSelected
+              ? 'bg-exo-yellow/10 border-l-exo-yellow border-r-exo-medium-gray/30 border-t-exo-medium-gray/30 border-b-exo-medium-gray/30'
+              : 'bg-exo-black/30 border-l-transparent border-r-exo-medium-gray/30 border-t-exo-medium-gray/30 border-b-exo-medium-gray/30 hover:bg-white/[0.03]'}"
+            onclick={() => toggleSelect(trace.taskId)}
+            onkeydown={(e) => {
+              if (e.key === "Enter" || e.key === " ") {
+                e.preventDefault();
+                toggleSelect(trace.taskId);
+              }
+            }}
           >
             <div class="min-w-0 flex-1">
               <a
                 href="#/traces/{trace.taskId}"
-                class="text-sm font-mono text-white hover:text-exo-yellow transition-colors truncate block"
+                class="text-sm font-mono transition-colors truncate block {isSelected
+                  ? 'text-exo-yellow'
+                  : 'text-white hover:text-exo-yellow'}"
+                onclick={(e) => e.stopPropagation()}
               >
                 {trace.taskId}
               </a>
@@ -160,7 +243,11 @@
                 )}
               </div>
             </div>
-            <div class="flex items-center gap-2 shrink-0">
+            <!-- svelte-ignore a11y_click_events_have_key_events -->
+            <div
+              class="flex items-center gap-2 shrink-0"
+              onclick={(e) => e.stopPropagation()}
+            >
               <a
                 href="#/traces/{trace.taskId}"
                 class="text-xs font-mono text-exo-light-gray hover:text-exo-yellow transition-colors uppercase border border-exo-medium-gray/40 px-2 py-1 rounded"

docs/api.md

@@ -183,6 +183,27 @@ Same schema as `/v1/chat/completions`.
 **Response:**
 Chat completion plus benchmarking metrics.
 
+### Cancel Command
+
+**POST** `/v1/cancel/{command_id}`
+
+Cancels an active generation command (text or image). Notifies workers and closes the stream.
+
+**Path parameters:**
+
+* `command_id`: string, ID of the command to cancel
+
+**Response (example):**
+
+```json
+{
+  "message": "Command cancelled.",
+  "command_id": "cmd-abc-123"
+}
+```
+
+Returns 404 if the command is not found or already completed.
+
 ## 5. Image Generation & Editing
 
 ### Image Generation
@@ -266,6 +287,7 @@ GET     /v1/models
 
 POST    /v1/chat/completions
 POST    /bench/chat/completions
+POST    /v1/cancel/{command_id}
 
 POST    /v1/images/generations
 POST    /bench/images/generations

flake.nix

@@ -108,6 +108,7 @@
                 package = pkgsSwift.swiftPackages.swift-format;
               };
               shfmt.enable = true;
+              taplo.enable = true;
             };
           };
 

nix/mlx.nix

@@ -41,7 +41,7 @@ let
 
   mlx = stdenv.mkDerivation rec {
     pname = "mlx";
-    version = let v = "0.30.7.dev20260220+13998a05"; in
+    version = let v = "0.30.7.dev20260225+257d5692"; 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;
@@ -49,8 +49,8 @@ let
     src = fetchFromGitHub {
       owner = "rltakashige";
       repo = "mlx-jaccl-fix-small-recv";
-      rev = "13998a054715edcdc93618fb1496c79c7c25ff7c";
-      hash = "sha256-fAqA3hFwNBx7FcoGnhQsIFpAIRbC2EerACm4Fvne0Cc=";
+      rev = "257d5692fc7af6bba3b8afaeb63c549b7d1e43d5";
+      hash = "sha256-GosFIWxIB48Egb1MqJrR3xhsUsQeWdRk5rV93USY6wQ=";
     };
 
     patches = [

pyproject.toml

@@ -1,35 +1,35 @@
 [project]
 name = "exo"
-version = "0.3.0"
+version = "0.3.68"
 description = "Exo"
 readme = "README.md"
 requires-python = ">=3.13"
 dependencies = [
-    "aiofiles>=24.1.0",
-    "aiohttp>=3.12.14",
-    "types-aiofiles>=24.1.0.20250708",
-    "pydantic>=2.11.7",
-    "fastapi>=0.116.1",
-    "filelock>=3.18.0",
-    "rustworkx>=0.17.1",
-    "huggingface-hub>=0.33.4",
-    "psutil>=7.0.0",
-    "loguru>=0.7.3",
-    "exo_pyo3_bindings", # rust bindings
-    "anyio==4.11.0",
-    "mlx; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.6; sys_platform == 'linux'",
-    "mlx-lm==0.30.7",
-    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
-    "hypercorn>=0.18.0",
-    "openai-harmony>=0.0.8",
-    "httpx>=0.28.1",
-    "tomlkit>=0.14.0",
-    "pillow>=11.0,<12.0", # compatibility with mflux
-    "mflux==0.15.5",
-    "python-multipart>=0.0.21",
-    "msgspec>=0.19.0",
-    "zstandard>=0.23.0",
+  "aiofiles>=24.1.0",
+  "aiohttp>=3.12.14",
+  "types-aiofiles>=24.1.0.20250708",
+  "pydantic>=2.11.7",
+  "fastapi>=0.116.1",
+  "filelock>=3.18.0",
+  "rustworkx>=0.17.1",
+  "huggingface-hub>=0.33.4",
+  "psutil>=7.0.0",
+  "loguru>=0.7.3",
+  "exo_pyo3_bindings",                         # rust bindings
+  "anyio==4.11.0",
+  "mlx; sys_platform == 'darwin'",
+  "mlx[cpu]==0.30.6; sys_platform == 'linux'",
+  "mlx-lm==0.30.7",
+  "tiktoken>=0.12.0",                          # required for kimi k2 tokenizer
+  "hypercorn>=0.18.0",
+  "openai-harmony>=0.0.8",
+  "httpx>=0.28.1",
+  "tomlkit>=0.14.0",
+  "pillow>=11.0,<12.0",                        # compatibility with mflux
+  "mflux==0.15.5",
+  "python-multipart>=0.0.21",
+  "msgspec>=0.19.0",
+  "zstandard>=0.23.0",
 ]
 
 [project.scripts]
@@ -38,12 +38,12 @@ exo = "exo.main:main"
 # dependencies only required for development
 [dependency-groups]
 dev = [
-    "basedpyright>=1.29.0",
-    "pyinstaller>=6.17.0",
-    "pytest>=8.4.0",
-    "pytest-asyncio>=1.0.0",
-    "pytest-env",
-    "ruff>=0.11.13",
+  "basedpyright>=1.29.0",
+  "pyinstaller>=6.17.0",
+  "pytest>=8.4.0",
+  "pytest-asyncio>=1.0.0",
+  "pytest-env",
+  "ruff>=0.11.13",
 ]
 
 # mlx[cuda] requires a newer version of mlx. the ideal on linux is: default to mlx[cpu] unless[cuda] specified.
@@ -57,10 +57,7 @@ dev = [
 ###
 
 [tool.uv.workspace]
-members = [
-    "rust/exo_pyo3_bindings",
-    "bench",
-]
+members = ["rust/exo_pyo3_bindings", "bench"]
 
 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
@@ -95,7 +92,15 @@ reportUnnecessaryTypeIgnoreComment = "error"
 pythonVersion = "3.13"
 pythonPlatform = "Darwin"
 
-exclude = ["**/.venv", "**/venv", "**/__pycache__", "**/exo_scripts", "**/.direnv", "**/rust", "**/.github"]
+exclude = [
+  "**/.venv",
+  "**/venv",
+  "**/__pycache__",
+  "**/exo_scripts",
+  "**/.direnv",
+  "**/rust",
+  "**/.github",
+]
 stubPath = ".mlx_typings"
 
 [[tool.basedpyright.executionEnvironments]]
@@ -109,17 +114,18 @@ root = "src"
 [tool.uv]
 required-version = ">=0.8.6"
 prerelease = "allow"
-environments = [
-    "sys_platform == 'darwin'",
-    "sys_platform == 'linux'",
-]
+environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]
 
 ###
 # ruff configuration
 ###
 
 [tool.ruff]
-extend-exclude = ["shared/protobufs/**", "*mlx_typings/**", "rust/exo_pyo3_bindings/**"]
+extend-exclude = [
+  "shared/protobufs/**",
+  "*mlx_typings/**",
+  "rust/exo_pyo3_bindings/**",
+]
 
 [tool.ruff.lint]
 extend-select = ["I", "N", "B", "A", "PIE", "SIM"]
@@ -127,13 +133,7 @@ extend-select = ["I", "N", "B", "A", "PIE", "SIM"]
 [tool.pytest.ini_options]
 pythonpath = "."
 asyncio_mode = "auto"
-markers = [
-    "slow: marks tests as slow (deselected by default)"
-]
-env = [
-  "EXO_TESTS=1"
-]
+markers = ["slow: marks tests as slow (deselected by default)"]
+env = ["EXO_TESTS=1"]
 addopts = "-m 'not slow' --ignore=tests/start_distributed_test.py"
-filterwarnings = [
-    "ignore:builtin type Swig:DeprecationWarning",
-]
+filterwarnings = ["ignore:builtin type Swig:DeprecationWarning"]

rust/exo_pyo3_bindings/Cargo.toml

@@ -26,20 +26,24 @@ networking = { workspace = true }
 
 # interop
 pyo3 = { version = "0.27.2", features = [
-    # "abi3-py313", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.13
-    # "nightly", # enables better-supported GIL integration
-    "experimental-async", # async support in #[pyfunction] & #[pymethods]
-    #"experimental-inspect", # inspection of generated binary => easier to automate type-hint generation
-    #"py-clone", # adding Clone-ing of `Py<T>` without GIL (may cause panics - remove if panics happen)
-    # "multiple-pymethods", # allows multiple #[pymethods] sections per class
+  # "abi3-py313", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.13
+  # "nightly", # enables better-supported GIL integration
+  "experimental-async", # async support in #[pyfunction] & #[pymethods]
+  #"experimental-inspect", # inspection of generated binary => easier to automate type-hint generation
+  #"py-clone", # adding Clone-ing of `Py<T>` without GIL (may cause panics - remove if panics happen)
+  # "multiple-pymethods", # allows multiple #[pymethods] sections per class
 
-    # integrations with other libraries
-    # "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
-    # "ordered-float", "rust_decimal", "smallvec",
-    # "anyhow", "chrono", "chrono-local", "chrono-tz", "eyre", "jiff-02", "lock_api", "parking-lot", "time",  "serde",
+  # integrations with other libraries
+  # "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
+  # "ordered-float", "rust_decimal", "smallvec",
+  # "anyhow", "chrono", "chrono-local", "chrono-tz", "eyre", "jiff-02", "lock_api", "parking-lot", "time",  "serde",
 ] }
 pyo3-stub-gen = { version = "0.17.2" }
-pyo3-async-runtimes = { version = "0.27.0", features = ["attributes", "tokio-runtime", "testing"] }
+pyo3-async-runtimes = { version = "0.27.0", features = [
+  "attributes",
+  "tokio-runtime",
+  "testing",
+] }
 pyo3-log = "0.13.2"
 
 # macro dependencies

rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi

@@ -2,7 +2,6 @@
 # ruff: noqa: E501, F401
 
 import builtins
-import enum
 import typing
 
 @typing.final
@@ -11,29 +10,6 @@ class AllQueuesFullError(builtins.Exception):
     def __repr__(self) -> builtins.str: ...
     def __str__(self) -> builtins.str: ...
 
-@typing.final
-class ConnectionUpdate:
-    @property
-    def update_type(self) -> ConnectionUpdateType:
-        r"""
-        Whether this is a connection or disconnection event
-        """
-    @property
-    def peer_id(self) -> builtins.str:
-        r"""
-        Identity of the peer that we have connected to or disconnected from.
-        """
-    @property
-    def remote_ipv4(self) -> builtins.str:
-        r"""
-        Remote connection's IPv4 address.
-        """
-    @property
-    def remote_tcp_port(self) -> builtins.int:
-        r"""
-        Remote connection's TCP port.
-        """
-
 @typing.final
 class Keypair:
     r"""
@@ -58,21 +34,15 @@ class Keypair:
         Convert the `Keypair` into the corresponding `PeerId` string, which we use as our `NodeId`.
         """
 
+@typing.final
+class MessageTooLargeError(builtins.Exception):
+    def __new__(cls, *args: typing.Any) -> MessageTooLargeError: ...
+    def __repr__(self) -> builtins.str: ...
+    def __str__(self) -> builtins.str: ...
+
 @typing.final
 class NetworkingHandle:
     def __new__(cls, identity: Keypair) -> NetworkingHandle: ...
-    async def connection_update_recv(self) -> ConnectionUpdate:
-        r"""
-        Receives the next `ConnectionUpdate` from networking.
-        """
-    async def connection_update_recv_many(self, limit: builtins.int) -> builtins.list[ConnectionUpdate]:
-        r"""
-        Receives at most `limit` `ConnectionUpdate`s from networking and returns them.
-        
-        For `limit = 0`, an empty collection of `ConnectionUpdate`s will be returned immediately.
-        For `limit > 0`, if there are no `ConnectionUpdate`s in the channel's queue this method
-        will sleep until a `ConnectionUpdate`s is sent.
-        """
     async def gossipsub_subscribe(self, topic: builtins.str) -> builtins.bool:
         r"""
         Subscribe to a `GossipSub` topic.
@@ -91,24 +61,7 @@ class NetworkingHandle:
         
         If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
         """
-    async def gossipsub_recv(self) -> tuple[builtins.str, bytes]:
-        r"""
-        Receives the next message from the `GossipSub` network.
-        """
-    async def gossipsub_recv_many(self, limit: builtins.int) -> builtins.list[tuple[builtins.str, bytes]]:
-        r"""
-        Receives at most `limit` messages from the `GossipSub` network and returns them.
-        
-        For `limit = 0`, an empty collection of messages will be returned immediately.
-        For `limit > 0`, if there are no messages in the channel's queue this method
-        will sleep until a message is sent.
-        """
-
-@typing.final
-class MessageTooLargeError(builtins.Exception):
-    def __new__(cls, *args: typing.Any) -> MessageTooLargeError: ...
-    def __repr__(self) -> builtins.str: ...
-    def __str__(self) -> builtins.str: ...
+    async def recv(self) -> PyFromSwarm: ...
 
 @typing.final
 class NoPeersSubscribedToTopicError(builtins.Exception):
@@ -116,11 +69,26 @@ class NoPeersSubscribedToTopicError(builtins.Exception):
     def __repr__(self) -> builtins.str: ...
     def __str__(self) -> builtins.str: ...
 
-@typing.final
-class ConnectionUpdateType(enum.Enum):
-    r"""
-    Connection or disconnection event discriminant type.
-    """
-    Connected = ...
-    Disconnected = ...
+class PyFromSwarm:
+    @typing.final
+    class Connection(PyFromSwarm):
+        __match_args__ = ("peer_id", "connected",)
+        @property
+        def peer_id(self) -> builtins.str: ...
+        @property
+        def connected(self) -> builtins.bool: ...
+        def __new__(cls, peer_id: builtins.str, connected: builtins.bool) -> PyFromSwarm.Connection: ...
+    
+    @typing.final
+    class Message(PyFromSwarm):
+        __match_args__ = ("origin", "topic", "data",)
+        @property
+        def origin(self) -> builtins.str: ...
+        @property
+        def topic(self) -> builtins.str: ...
+        @property
+        def data(self) -> bytes: ...
+        def __new__(cls, origin: builtins.str, topic: builtins.str, data: bytes) -> PyFromSwarm.Message: ...
+    
+    ...
 

rust/exo_pyo3_bindings/pyproject.toml

@@ -4,21 +4,18 @@ build-backend = "maturin"
 
 [project]
 name = "exo_pyo3_bindings"
-version = "0.1.0"
+version = "0.2.0"
 description = "Add your description here"
 readme = "README.md"
 authors = [
-    { name = "Andrei Cravtov", email = "the.andrei.cravtov@gmail.com" }
+  { name = "Andrei Cravtov", email = "the.andrei.cravtov@gmail.com" },
+  { name = "Evan Quiney", email = "evanev7@gmail.com" },
 ]
 requires-python = ">=3.13"
 dependencies = []
 
 [dependency-groups]
-dev = [
-    "exo_pyo3_bindings",
-    "pytest>=8.4.0",
-    "pytest-asyncio>=1.0.0",
-]
+dev = ["exo_pyo3_bindings", "pytest>=8.4.0", "pytest-asyncio>=1.0.0"]
 
 [tool.maturin]
 #purelib = true

rust/exo_pyo3_bindings/src/lib.rs

@@ -155,6 +155,9 @@ pub(crate) mod ext {
 fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
     // install logger
     pyo3_log::init();
+    let mut builder = tokio::runtime::Builder::new_multi_thread();
+    builder.enable_all();
+    pyo3_async_runtimes::tokio::init(builder);
 
     // TODO: for now this is all NOT a submodule, but figure out how to make the submodule system
     //       work with maturin, where the types generate correctly, in the right folder, without

rust/exo_pyo3_bindings/src/networking.rs

@@ -1,26 +1,24 @@
-#![allow(
-    clippy::multiple_inherent_impl,
-    clippy::unnecessary_wraps,
-    clippy::unused_self,
-    clippy::needless_pass_by_value
-)]
+use std::pin::Pin;
+use std::sync::Arc;
 
 use crate::r#const::MPSC_CHANNEL_SIZE;
 use crate::ext::{ByteArrayExt as _, FutureExt, PyErrExt as _};
-use crate::ext::{ResultExt as _, TokioMpscReceiverExt as _, TokioMpscSenderExt as _};
+use crate::ext::{ResultExt as _, TokioMpscSenderExt as _};
 use crate::ident::PyKeypair;
+use crate::networking::exception::{
+    PyAllQueuesFullError, PyMessageTooLargeError, PyNoPeersSubscribedToTopicError,
+};
 use crate::pyclass;
-use libp2p::futures::StreamExt as _;
-use libp2p::gossipsub;
-use libp2p::gossipsub::{IdentTopic, Message, MessageId, PublishError};
-use libp2p::swarm::SwarmEvent;
-use networking::discovery;
-use networking::swarm::create_swarm;
+use futures_lite::{Stream, StreamExt as _};
+use libp2p::gossipsub::PublishError;
+use networking::swarm::{FromSwarm, ToSwarm, create_swarm};
+use pyo3::exceptions::PyRuntimeError;
 use pyo3::prelude::{PyModule, PyModuleMethods as _};
 use pyo3::types::PyBytes;
-use pyo3::{Bound, Py, PyErr, PyResult, PyTraverseError, PyVisit, Python, pymethods};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pyclass_enum, gen_stub_pymethods};
-use std::net::IpAddr;
+use pyo3::{Bound, Py, PyAny, PyErr, PyResult, Python, pymethods};
+use pyo3_stub_gen::derive::{
+    gen_methods_from_python, gen_stub_pyclass, gen_stub_pyclass_complex_enum, gen_stub_pymethods,
+};
 use tokio::sync::{Mutex, mpsc, oneshot};
 
 mod exception {
@@ -131,237 +129,45 @@ mod exception {
     }
 }
 
-/// Connection or disconnection event discriminant type.
-#[gen_stub_pyclass_enum]
-#[pyclass(eq, eq_int, name = "ConnectionUpdateType")]
-#[derive(Debug, Clone, PartialEq)]
-enum PyConnectionUpdateType {
-    Connected = 0,
-    Disconnected,
-}
-
-#[gen_stub_pyclass]
-#[pyclass(frozen, name = "ConnectionUpdate")]
-#[derive(Debug, Clone)]
-struct PyConnectionUpdate {
-    /// Whether this is a connection or disconnection event
-    #[pyo3(get)]
-    update_type: PyConnectionUpdateType,
-
-    /// Identity of the peer that we have connected to or disconnected from.
-    #[pyo3(get)]
-    peer_id: String,
-
-    /// Remote connection's IPv4 address.
-    #[pyo3(get)]
-    remote_ipv4: String,
-
-    /// Remote connection's TCP port.
-    #[pyo3(get)]
-    remote_tcp_port: u16,
-}
-
-enum ToTask {
-    GossipsubSubscribe {
-        topic: String,
-        result_tx: oneshot::Sender<PyResult<bool>>,
-    },
-    GossipsubUnsubscribe {
-        topic: String,
-        result_tx: oneshot::Sender<bool>,
-    },
-    GossipsubPublish {
-        topic: String,
-        data: Vec<u8>,
-        result_tx: oneshot::Sender<PyResult<MessageId>>,
-    },
-}
-
-#[allow(clippy::enum_glob_use)]
-async fn networking_task(
-    mut swarm: networking::swarm::Swarm,
-    mut to_task_rx: mpsc::Receiver<ToTask>,
-    connection_update_tx: mpsc::Sender<PyConnectionUpdate>,
-    gossipsub_message_tx: mpsc::Sender<(String, Vec<u8>)>,
-) {
-    use SwarmEvent::*;
-    use ToTask::*;
-    use networking::swarm::BehaviourEvent::*;
-
-    log::info!("RUST: networking task started");
-
-    loop {
-        tokio::select! {
-            message = to_task_rx.recv() => {
-                // handle closed channel
-                let Some(message) = message else {
-                    log::info!("RUST: channel closed");
-                    break;
-                };
-
-                // dispatch incoming messages
-                match message {
-                    GossipsubSubscribe { topic, result_tx } => {
-                        // try to subscribe
-                        let result = swarm.behaviour_mut()
-                            .gossipsub.subscribe(&IdentTopic::new(topic));
-
-                        // send response oneshot
-                        if let Err(e) = result_tx.send(result.pyerr()) {
-                            log::error!("RUST: could not subscribe to gossipsub topic since channel already closed: {e:?}");
-                            continue;
-                        }
-                    }
-                    GossipsubUnsubscribe { topic, result_tx } => {
-                        // try to unsubscribe from the topic
-                        let result = swarm.behaviour_mut()
-                            .gossipsub.unsubscribe(&IdentTopic::new(topic));
-
-                        // send response oneshot (or exit if connection closed)
-                        if let Err(e) = result_tx.send(result) {
-                            log::error!("RUST: could not unsubscribe from gossipsub topic since channel already closed: {e:?}");
-                            continue;
-                        }
-                    }
-                    GossipsubPublish { topic, data, result_tx } => {
-                        // try to publish the data -> catch NoPeersSubscribedToTopic error & convert to correct exception
-                        let result = swarm.behaviour_mut().gossipsub.publish(
-                            IdentTopic::new(topic), data);
-                        let pyresult: PyResult<MessageId> = if let Err(PublishError::NoPeersSubscribedToTopic) = result {
-                            Err(exception::PyNoPeersSubscribedToTopicError::new_err())
-                        } else if let Err(PublishError::AllQueuesFull(_)) = result {
-                            Err(exception::PyAllQueuesFullError::new_err())
-                        } else if let Err(PublishError::MessageTooLarge) = result {
-                            Err(exception::PyMessageTooLargeError::new_err())
-                        } else {
-                            result.pyerr()
-                        };
-
-                        // send response oneshot (or exit if connection closed)
-                        if let Err(e) = result_tx.send(pyresult) {
-                            log::error!("RUST: could not publish gossipsub message since channel already closed: {e:?}");
-                            continue;
-                        }
-                    }
-                }
-            }
-
-            // architectural solution to this problem:
-            // create keep_alive behavior who's job it is to dial peers discovered by mDNS (and drop when expired)
-            //   -> it will emmit TRUE connected/disconnected events consumable elsewhere
-            //
-            // gossipsub will feed off-of dial attempts created by networking, and that will bootstrap its' peers list
-            // then for actual communication it will dial those peers if need-be
-            swarm_event = swarm.select_next_some() => {
-                match swarm_event {
-                    Behaviour(Gossipsub(gossipsub::Event::Message {
-                        message: Message {
-                            topic,
-                            data,
-                            ..
-                        },
-                        ..
-                    })) => {
-                        // topic-ID is just the topic hash!!! (since we used identity hasher)
-                        let message = (topic.into_string(), data);
-
-                        // send incoming message to channel (or exit if connection closed)
-                        if let Err(e) = gossipsub_message_tx.send(message).await {
-                            log::error!("RUST: could not send incoming gossipsub message since channel already closed: {e}");
-                            continue;
-                        }
-                    },
-                    Behaviour(Discovery(discovery::Event::ConnectionEstablished { peer_id, remote_ip, remote_tcp_port, .. })) => {
-                        // grab IPv4 string
-                        let remote_ipv4 = match remote_ip {
-                            IpAddr::V4(ip) => ip.to_string(),
-                            IpAddr::V6(ip) => {
-                                log::warn!("RUST: ignoring connection to IPv6 address: {ip}");
-                                continue;
-                            }
-                        };
-
-                        // send connection event to channel (or exit if connection closed)
-                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
-                            update_type: PyConnectionUpdateType::Connected,
-                            peer_id: peer_id.to_base58(),
-                            remote_ipv4,
-                            remote_tcp_port,
-                        }).await {
-                            log::error!("RUST: could not send connection update since channel already closed: {e}");
-                            continue;
-                        }
-                    },
-                    Behaviour(Discovery(discovery::Event::ConnectionClosed { peer_id, remote_ip, remote_tcp_port, .. })) => {
-                        // grab IPv4 string
-                        let remote_ipv4 = match remote_ip {
-                            IpAddr::V4(ip) => ip.to_string(),
-                            IpAddr::V6(ip) => {
-                                log::warn!("RUST: ignoring disconnection from IPv6 address: {ip}");
-                                continue;
-                            }
-                        };
-
-                        // send disconnection event to channel (or exit if connection closed)
-                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
-                            update_type: PyConnectionUpdateType::Disconnected,
-                            peer_id: peer_id.to_base58(),
-                            remote_ipv4,
-                            remote_tcp_port,
-                        }).await {
-                            log::error!("RUST: could not send connection update since channel already closed: {e}");
-                            continue;
-                        }
-                    },
-                    e => {
-                        log::info!("RUST: other event {e:?}");
-                    }
-                }
-            }
-        }
-    }
-
-    log::info!("RUST: networking task stopped");
-}
-
 #[gen_stub_pyclass]
 #[pyclass(name = "NetworkingHandle")]
-#[derive(Debug)]
 struct PyNetworkingHandle {
     // channels
-    to_task_tx: Option<mpsc::Sender<ToTask>>,
-    connection_update_rx: Mutex<mpsc::Receiver<PyConnectionUpdate>>,
-    gossipsub_message_rx: Mutex<mpsc::Receiver<(String, Vec<u8>)>>,
+    pub to_swarm: mpsc::Sender<ToSwarm>,
+    pub swarm: Arc<Mutex<Pin<Box<dyn Stream<Item = FromSwarm> + Send>>>>,
 }
 
-impl Drop for PyNetworkingHandle {
-    fn drop(&mut self) {
-        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
-        //       to ensure that the networking task is done BEFORE exiting the clear function...
-        //       but this may require GIL?? and it may not be safe to call GIL here??
-        self.to_task_tx = None; // Using Option<T> as a trick to force channel to be dropped
-    }
+#[gen_stub_pyclass_complex_enum]
+#[pyclass]
+enum PyFromSwarm {
+    Connection {
+        peer_id: String,
+        connected: bool,
+    },
+    Message {
+        origin: String,
+        topic: String,
+        data: Py<PyBytes>,
+    },
 }
-
-#[allow(clippy::expect_used)]
-impl PyNetworkingHandle {
-    fn new(
-        to_task_tx: mpsc::Sender<ToTask>,
-        connection_update_rx: mpsc::Receiver<PyConnectionUpdate>,
-        gossipsub_message_rx: mpsc::Receiver<(String, Vec<u8>)>,
-    ) -> Self {
-        Self {
-            to_task_tx: Some(to_task_tx),
-            connection_update_rx: Mutex::new(connection_update_rx),
-            gossipsub_message_rx: Mutex::new(gossipsub_message_rx),
+impl From<FromSwarm> for PyFromSwarm {
+    fn from(value: FromSwarm) -> Self {
+        match value {
+            FromSwarm::Discovered { peer_id } => Self::Connection {
+                peer_id: peer_id.to_base58(),
+                connected: true,
+            },
+            FromSwarm::Expired { peer_id } => Self::Connection {
+                peer_id: peer_id.to_base58(),
+                connected: false,
+            },
+            FromSwarm::Message { from, topic, data } => Self::Message {
+                origin: from.to_base58(),
+                topic: topic,
+                data: data.pybytes(),
+            },
         }
     }
-
-    const fn to_task_tx(&self) -> &mpsc::Sender<ToTask> {
-        self.to_task_tx
-            .as_ref()
-            .expect("The sender should only be None after de-initialization.")
-    }
 }
 
 #[gen_stub_pymethods]
@@ -375,97 +181,36 @@ impl PyNetworkingHandle {
 
     #[new]
     fn py_new(identity: Bound<'_, PyKeypair>) -> PyResult<Self> {
-        use pyo3_async_runtimes::tokio::get_runtime;
-
         // create communication channels
-        let (to_task_tx, to_task_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
-        let (connection_update_tx, connection_update_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
-        let (gossipsub_message_tx, gossipsub_message_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
+        let (to_swarm, from_client) = mpsc::channel(MPSC_CHANNEL_SIZE);
 
         // get identity
         let identity = identity.borrow().0.clone();
 
         // create networking swarm (within tokio context!! or it crashes)
-        let swarm = get_runtime()
-            .block_on(async { create_swarm(identity) })
-            .pyerr()?;
+        let _guard = pyo3_async_runtimes::tokio::get_runtime().enter();
+        let swarm = create_swarm(identity, from_client).pyerr()?.into_stream();
 
-        // spawn tokio task running the networking logic
-        get_runtime().spawn(async move {
-            networking_task(
-                swarm,
-                to_task_rx,
-                connection_update_tx,
-                gossipsub_message_tx,
-            )
-            .await;
-        });
-        Ok(Self::new(
-            to_task_tx,
-            connection_update_rx,
-            gossipsub_message_rx,
-        ))
+        Ok(Self {
+            swarm: Arc::new(Mutex::new(swarm)),
+            to_swarm,
+        })
     }
 
     #[gen_stub(skip)]
-    const fn __traverse__(&self, _visit: PyVisit<'_>) -> Result<(), PyTraverseError> {
-        Ok(()) // This is needed purely so `__clear__` can work
+    fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
+        let swarm = Arc::clone(&self.swarm);
+        pyo3_async_runtimes::tokio::future_into_py(py, async move {
+            swarm
+                .try_lock()
+                .map_err(|_| PyRuntimeError::new_err("called recv twice concurrently"))?
+                .next()
+                .await
+                .ok_or(PyErr::receiver_channel_closed())
+                .map(PyFromSwarm::from)
+        })
     }
 
-    #[gen_stub(skip)]
-    fn __clear__(&mut self) {
-        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
-        //       to ensure that the networking task is done BEFORE exiting the clear function...
-        //       but this may require GIL?? and it may not be safe to call GIL here??
-        self.to_task_tx = None; // Using Option<T> as a trick to force channel to be dropped
-    }
-
-    // ---- Connection update receiver methods ----
-
-    /// Receives the next `ConnectionUpdate` from networking.
-    async fn connection_update_recv(&self) -> PyResult<PyConnectionUpdate> {
-        self.connection_update_rx
-            .lock()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .recv_py()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-    }
-
-    /// Receives at most `limit` `ConnectionUpdate`s from networking and returns them.
-    ///
-    /// For `limit = 0`, an empty collection of `ConnectionUpdate`s will be returned immediately.
-    /// For `limit > 0`, if there are no `ConnectionUpdate`s in the channel's queue this method
-    /// will sleep until a `ConnectionUpdate`s is sent.
-    async fn connection_update_recv_many(&self, limit: usize) -> PyResult<Vec<PyConnectionUpdate>> {
-        self.connection_update_rx
-            .lock()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .recv_many_py(limit)
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-    }
-
-    // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)
-    //       so its too dangerous to expose just yet. figure out a better semantics for handling this,
-    //       so things don't randomly block
-    // /// Tries to receive the next `ConnectionUpdate` from networking.
-    // fn connection_update_try_recv(&self) -> PyResult<Option<PyConnectionUpdate>> {
-    //     self.connection_update_rx.blocking_lock().try_recv_py()
-    // }
-    //
-    // /// Checks if the `ConnectionUpdate` channel is empty.
-    // fn connection_update_is_empty(&self) -> bool {
-    //     self.connection_update_rx.blocking_lock().is_empty()
-    // }
-    //
-    // /// Returns the number of `ConnectionUpdate`s in the channel.
-    // fn connection_update_len(&self) -> usize {
-    //     self.connection_update_rx.blocking_lock().len()
-    // }
-
     // ---- Gossipsub management methods ----
 
     /// Subscribe to a `GossipSub` topic.
@@ -475,10 +220,10 @@ impl PyNetworkingHandle {
         let (tx, rx) = oneshot::channel();
 
         // send off request to subscribe
-        self.to_task_tx()
-            .send_py(ToTask::GossipsubSubscribe {
+        self.to_swarm
+            .send_py(ToSwarm::Subscribe {
                 topic,
-                result_tx: tx,
+                result_sender: tx,
             })
             .allow_threads_py() // allow-threads-aware async call
             .await?;
@@ -487,6 +232,7 @@ impl PyNetworkingHandle {
         rx.allow_threads_py() // allow-threads-aware async call
             .await
             .map_err(|_| PyErr::receiver_channel_closed())?
+            .pyerr()
     }
 
     /// Unsubscribes from a `GossipSub` topic.
@@ -496,10 +242,10 @@ impl PyNetworkingHandle {
         let (tx, rx) = oneshot::channel();
 
         // send off request to unsubscribe
-        self.to_task_tx()
-            .send_py(ToTask::GossipsubUnsubscribe {
+        self.to_swarm
+            .send_py(ToSwarm::Unsubscribe {
                 topic,
-                result_tx: tx,
+                result_sender: tx,
             })
             .allow_threads_py() // allow-threads-aware async call
             .await?;
@@ -518,11 +264,11 @@ impl PyNetworkingHandle {
 
         // send off request to subscribe
         let data = Python::attach(|py| Vec::from(data.as_bytes(py)));
-        self.to_task_tx()
-            .send_py(ToTask::GossipsubPublish {
+        self.to_swarm
+            .send_py(ToSwarm::Publish {
                 topic,
                 data,
-                result_tx: tx,
+                result_sender: tx,
             })
             .allow_threads_py() // allow-threads-aware async call
             .await?;
@@ -531,64 +277,26 @@ impl PyNetworkingHandle {
         let _ = rx
             .allow_threads_py() // allow-threads-aware async call
             .await
-            .map_err(|_| PyErr::receiver_channel_closed())??;
+            .map_err(|_| PyErr::receiver_channel_closed())?
+            .map_err(|e| match e {
+                PublishError::AllQueuesFull(_) => PyAllQueuesFullError::new_err(),
+                PublishError::MessageTooLarge => PyMessageTooLargeError::new_err(),
+                PublishError::NoPeersSubscribedToTopic => {
+                    PyNoPeersSubscribedToTopicError::new_err()
+                }
+                e => PyRuntimeError::new_err(e.to_string()),
+            })?;
         Ok(())
     }
+}
 
-    // ---- Gossipsub message receiver methods ----
-
-    /// Receives the next message from the `GossipSub` network.
-    async fn gossipsub_recv(&self) -> PyResult<(String, Py<PyBytes>)> {
-        self.gossipsub_message_rx
-            .lock()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .recv_py()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .map(|(t, d)| (t, d.pybytes()))
+pyo3_stub_gen::inventory::submit! {
+    gen_methods_from_python! {
+        r#"
+            class PyNetworkingHandle:
+                async def recv() -> PyFromSwarm: ...
+        "#
     }
-
-    /// Receives at most `limit` messages from the `GossipSub` network and returns them.
-    ///
-    /// For `limit = 0`, an empty collection of messages will be returned immediately.
-    /// For `limit > 0`, if there are no messages in the channel's queue this method
-    /// will sleep until a message is sent.
-    async fn gossipsub_recv_many(&self, limit: usize) -> PyResult<Vec<(String, Py<PyBytes>)>> {
-        Ok(self
-            .gossipsub_message_rx
-            .lock()
-            .allow_threads_py() // allow-threads-aware async call
-            .await
-            .recv_many_py(limit)
-            .allow_threads_py() // allow-threads-aware async call
-            .await?
-            .into_iter()
-            .map(|(t, d)| (t, d.pybytes()))
-            .collect())
-    }
-
-    // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)
-    //       so its too dangerous to expose just yet. figure out a better semantics for handling this,
-    //       so things don't randomly block
-    // /// Tries to receive the next message from the `GossipSub` network.
-    // fn gossipsub_try_recv(&self) -> PyResult<Option<(String, Py<PyBytes>)>> {
-    //     Ok(self
-    //         .gossipsub_message_rx
-    //         .blocking_lock()
-    //         .try_recv_py()?
-    //         .map(|(t, d)| (t, d.pybytes())))
-    // }
-    //
-    // /// Checks if the `GossipSub` message channel is empty.
-    // fn gossipsub_is_empty(&self) -> bool {
-    //     self.gossipsub_message_rx.blocking_lock().is_empty()
-    // }
-    //
-    // /// Returns the number of `GossipSub` messages in the channel.
-    // fn gossipsub_len(&self) -> usize {
-    //     self.gossipsub_message_rx.blocking_lock().len()
-    // }
 }
 
 pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
@@ -596,10 +304,8 @@ pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_class::<exception::PyAllQueuesFullError>()?;
     m.add_class::<exception::PyMessageTooLargeError>()?;
 
-    m.add_class::<PyConnectionUpdateType>()?;
-    m.add_class::<PyConnectionUpdate>()?;
-    m.add_class::<PyConnectionUpdateType>()?;
     m.add_class::<PyNetworkingHandle>()?;
+    m.add_class::<PyFromSwarm>()?;
 
     Ok(())
 }

rust/networking/Cargo.toml

@@ -21,13 +21,17 @@ extend = { workspace = true }
 delegate = { workspace = true }
 
 # async
-tokio = { workspace = true, features = ["full"] }
+async-stream = { workspace = true }
 futures-lite = { workspace = true }
 futures-timer = { workspace = true }
+tokio = { workspace = true, features = ["full"] }
 
 # utility dependencies
 util = { workspace = true }
-tracing-subscriber = { version = "0.3.19", features = ["default", "env-filter"] }
+tracing-subscriber = { version = "0.3.19", features = [
+  "default",
+  "env-filter",
+] }
 keccak-const = { workspace = true }
 
 # tracing/logging
@@ -35,3 +39,4 @@ log = { workspace = true }
 
 # networking
 libp2p = { workspace = true, features = ["full"] }
+pin-project = "1.1.10"

rust/networking/examples/chatroom.rs

@@ -1,7 +1,9 @@
 use futures_lite::StreamExt;
-use libp2p::{gossipsub, identity, swarm::SwarmEvent};
-use networking::{discovery, swarm};
-use tokio::{io, io::AsyncBufReadExt as _, select};
+use libp2p::identity;
+use networking::swarm;
+use networking::swarm::{FromSwarm, ToSwarm};
+use tokio::sync::{mpsc, oneshot};
+use tokio::{io, io::AsyncBufReadExt as _};
 use tracing_subscriber::EnvFilter;
 use tracing_subscriber::filter::LevelFilter;
 
@@ -11,64 +13,69 @@ async fn main() {
         .with_env_filter(EnvFilter::from_default_env().add_directive(LevelFilter::INFO.into()))
         .try_init();
 
+    let (to_swarm, from_client) = mpsc::channel(20);
+
     // Configure swarm
-    let mut swarm =
-        swarm::create_swarm(identity::Keypair::generate_ed25519()).expect("Swarm creation failed");
+    let mut swarm = swarm::create_swarm(identity::Keypair::generate_ed25519(), from_client)
+        .expect("Swarm creation failed")
+        .into_stream();
 
     // Create a Gossipsub topic & subscribe
-    let topic = gossipsub::IdentTopic::new("test-net");
-    swarm
-        .behaviour_mut()
-        .gossipsub
-        .subscribe(&topic)
-        .expect("Subscribing to topic failed");
+    let (tx, rx) = oneshot::channel();
+    _ = to_swarm
+        .send(ToSwarm::Subscribe {
+            topic: "test-net".to_string(),
+            result_sender: tx,
+        })
+        .await
+        .expect("should send");
 
     // Read full lines from stdin
     let mut stdin = io::BufReader::new(io::stdin()).lines();
     println!("Enter messages via STDIN and they will be sent to connected peers using Gossipsub");
 
+    tokio::task::spawn(async move {
+        rx.await
+            .expect("tx not dropped")
+            .expect("subscribe shouldn't fail");
+        loop {
+            if let Ok(Some(line)) = stdin.next_line().await {
+                let (tx, rx) = oneshot::channel();
+                if let Err(e) = to_swarm
+                    .send(swarm::ToSwarm::Publish {
+                        topic: "test-net".to_string(),
+                        data: line.as_bytes().to_vec(),
+                        result_sender: tx,
+                    })
+                    .await
+                {
+                    println!("Send error: {e:?}");
+                    return;
+                };
+                match rx.await {
+                    Ok(Err(e)) => println!("Publish error: {e:?}"),
+                    Err(e) => println!("Publish error: {e:?}"),
+                    Ok(_) => {}
+                }
+            }
+        }
+    });
+
     // Kick it off
     loop {
-        select! {
-            // on gossipsub outgoing
-            Ok(Some(line)) = stdin.next_line() => {
-                if let Err(e) = swarm
-                    .behaviour_mut().gossipsub
-                    .publish(topic.clone(), line.as_bytes()) {
-                    println!("Publish error: {e:?}");
-                }
+        // on gossipsub outgoing
+        match swarm.next().await {
+            // on gossipsub incoming
+            Some(FromSwarm::Discovered { peer_id }) => {
+                println!("\n\nconnected to {peer_id}\n\n")
             }
-            event = swarm.next() => match event {
-                // on gossipsub incoming
-                Some(SwarmEvent::Behaviour(swarm::BehaviourEvent::Gossipsub(gossipsub::Event::Message {
-                    propagation_source: peer_id,
-                    message_id: id,
-                    message,
-                }))) => println!(
-                        "\n\nGot message: '{}' with id: {id} from peer: {peer_id}\n\n",
-                        String::from_utf8_lossy(&message.data),
-                    ),
-
-                // on discovery
-                Some(SwarmEvent::Behaviour(swarm::BehaviourEvent::Discovery(e)) )=> match e {
-                    discovery::Event::ConnectionEstablished {
-                        peer_id, connection_id, remote_ip, remote_tcp_port
-                    } => {
-                        println!("\n\nConnected to: {peer_id}; connection ID: {connection_id}; remote IP: {remote_ip}; remote TCP port: {remote_tcp_port}\n\n");
-                    }
-                    discovery::Event::ConnectionClosed {
-                        peer_id, connection_id, remote_ip, remote_tcp_port
-                    } => {
-                        eprintln!("\n\nDisconnected from: {peer_id}; connection ID: {connection_id}; remote IP: {remote_ip}; remote TCP port: {remote_tcp_port}\n\n");
-                    }
-                }
-
-                // ignore outgoing errors: those are normal
-                e@Some(SwarmEvent::OutgoingConnectionError { .. }) => { log::debug!("Outgoing connection error: {e:?}"); }
-
-                // otherwise log any other event
-                e => { log::info!("Other event {e:?}"); }
+            Some(FromSwarm::Expired { peer_id }) => {
+                println!("\n\ndisconnected from {peer_id}\n\n")
             }
+            Some(FromSwarm::Message { from, topic, data }) => {
+                println!("{topic}/{from}:\n{}", String::from_utf8_lossy(&data))
+            }
+            None => {}
         }
     }
 }

rust/networking/src/swarm.rs

@@ -1,9 +1,11 @@
-use crate::alias;
-use crate::swarm::transport::tcp_transport;
-pub use behaviour::{Behaviour, BehaviourEvent};
-use libp2p::{SwarmBuilder, identity};
+use std::pin::Pin;
 
-pub type Swarm = libp2p::Swarm<Behaviour>;
+use crate::swarm::transport::tcp_transport;
+use crate::{alias, discovery};
+pub use behaviour::{Behaviour, BehaviourEvent};
+use futures_lite::{Stream, StreamExt};
+use libp2p::{PeerId, SwarmBuilder, gossipsub, identity, swarm::SwarmEvent};
+use tokio::sync::{mpsc, oneshot};
 
 /// The current version of the network: this prevents devices running different versions of the
 /// software from interacting with each other.
@@ -15,8 +17,136 @@ pub type Swarm = libp2p::Swarm<Behaviour>;
 pub const NETWORK_VERSION: &[u8] = b"v0.0.1";
 pub const OVERRIDE_VERSION_ENV_VAR: &str = "EXO_LIBP2P_NAMESPACE";
 
+// Uses oneshot senders to emulate function calling apis while avoiding requiring unique ownership
+// of the Swarm.
+pub enum ToSwarm {
+    Unsubscribe {
+        topic: String,
+        result_sender: oneshot::Sender<bool>,
+    },
+    Subscribe {
+        topic: String,
+        result_sender: oneshot::Sender<Result<bool, gossipsub::SubscriptionError>>,
+    },
+    Publish {
+        topic: String,
+        data: Vec<u8>,
+        result_sender: oneshot::Sender<Result<gossipsub::MessageId, gossipsub::PublishError>>,
+    },
+}
+pub enum FromSwarm {
+    Message {
+        from: PeerId,
+        topic: String,
+        data: Vec<u8>,
+    },
+    Discovered {
+        peer_id: PeerId,
+    },
+    Expired {
+        peer_id: PeerId,
+    },
+}
+
+pub struct Swarm {
+    swarm: libp2p::Swarm<Behaviour>,
+    from_client: mpsc::Receiver<ToSwarm>,
+}
+
+impl Swarm {
+    pub fn into_stream(self) -> Pin<Box<dyn Stream<Item = FromSwarm> + Send>> {
+        let Swarm {
+            mut swarm,
+            mut from_client,
+        } = self;
+        let stream = async_stream::stream! {
+            loop {
+                tokio::select! {
+                    msg = from_client.recv() => {
+                        let Some(msg) = msg else { break };
+                        on_message(&mut swarm, msg);
+                    }
+                    event = swarm.next() => {
+                        let Some(event) = event else { break };
+                        if let Some(item) = filter_swarm_event(event) {
+                            yield item;
+                        }
+                    }
+                }
+            }
+        };
+        Box::pin(stream)
+    }
+}
+
+fn on_message(swarm: &mut libp2p::Swarm<Behaviour>, message: ToSwarm) {
+    match message {
+        ToSwarm::Subscribe {
+            topic,
+            result_sender,
+        } => {
+            let result = swarm
+                .behaviour_mut()
+                .gossipsub
+                .subscribe(&gossipsub::IdentTopic::new(topic));
+            _ = result_sender.send(result);
+        }
+        ToSwarm::Unsubscribe {
+            topic,
+            result_sender,
+        } => {
+            let result = swarm
+                .behaviour_mut()
+                .gossipsub
+                .unsubscribe(&gossipsub::IdentTopic::new(topic));
+            _ = result_sender.send(result);
+        }
+        ToSwarm::Publish {
+            topic,
+            data,
+            result_sender,
+        } => {
+            let result = swarm
+                .behaviour_mut()
+                .gossipsub
+                .publish(gossipsub::IdentTopic::new(topic), data);
+            _ = result_sender.send(result);
+        }
+    }
+}
+
+fn filter_swarm_event(event: SwarmEvent<BehaviourEvent>) -> Option<FromSwarm> {
+    match event {
+        SwarmEvent::Behaviour(BehaviourEvent::Gossipsub(gossipsub::Event::Message {
+            message:
+                gossipsub::Message {
+                    source: Some(peer_id),
+                    topic,
+                    data,
+                    ..
+                },
+            ..
+        })) => Some(FromSwarm::Message {
+            from: peer_id,
+            topic: topic.into_string(),
+            data,
+        }),
+        SwarmEvent::Behaviour(BehaviourEvent::Discovery(
+            discovery::Event::ConnectionEstablished { peer_id, .. },
+        )) => Some(FromSwarm::Discovered { peer_id }),
+        SwarmEvent::Behaviour(BehaviourEvent::Discovery(discovery::Event::ConnectionClosed {
+            peer_id,
+            ..
+        })) => Some(FromSwarm::Expired { peer_id }),
+        _ => None,
+    }
+}
+
 /// Create and configure a swarm which listens to all ports on OS
-pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
+pub fn create_swarm(
+    keypair: identity::Keypair,
+    from_client: mpsc::Receiver<ToSwarm>,
+) -> alias::AnyResult<Swarm> {
     let mut swarm = SwarmBuilder::with_existing_identity(keypair)
         .with_tokio()
         .with_other_transport(tcp_transport)?
@@ -25,7 +155,7 @@ pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
 
     // Listen on all interfaces and whatever port the OS assigns
     swarm.listen_on("/ip4/0.0.0.0/tcp/0".parse()?)?;
-    Ok(swarm)
+    Ok(Swarm { swarm, from_client })
 }
 
 mod transport {

src/exo/__main__.py

@@ -32,8 +32,55 @@ def _maybe_run_inline_code(argv: Sequence[str]) -> bool:
     return True
 
 
+def _is_frozen() -> bool:
+    """Check if running in a frozen environment (e.g., PyInstaller)."""
+    return getattr(sys, "frozen", False) or hasattr(sys, "_MEIPASS")
+
+
+def _maybe_run_runner(argv: Sequence[str]) -> bool:
+    """
+    Handle the --exo-runner flag for spawning runner subprocesses in frozen builds.
+    """
+    if "--exo-runner" not in argv:
+        return False
+
+    try:
+        flag_index = argv.index("--exo-runner")
+        args_index = flag_index + 1
+        if args_index >= len(argv):
+            return False
+
+        # Set up the runner environment
+        import os
+
+        os.environ["EXO_FROZEN"] = "1"
+
+        # Import and run the bootstrap entrypoint
+        from exo.worker.runner.bootstrap import entrypoint
+
+        # Inject the JSON args into sys.argv as expected by entrypoint
+        sys.argv = [argv[0], argv[args_index]]
+        entrypoint()
+        return True
+    except Exception as e:
+        print(f"Failed to start runner: {e}", file=sys.stderr)
+        return False
+
+
 if __name__ == "__main__":
     if _maybe_run_inline_code(sys.argv):
         sys.exit(0)
+
+    # Check if we should run as a runner subprocess (frozen builds)
+    if _maybe_run_runner(sys.argv):
+        sys.exit(0)
+
     freeze_support()
+
+    # Store frozen status for subprocess spawning
+    if _is_frozen():
+        import os
+
+        os.environ["EXO_FROZEN"] = "1"
+
     main()

src/exo/download/coordinator.py

@@ -1,7 +1,5 @@
 import asyncio
-import socket
 from dataclasses import dataclass, field
-from random import random
 
 import anyio
 from anyio import current_time
@@ -22,13 +20,9 @@ from exo.shared.types.commands import (
     ForwarderDownloadCommand,
     StartDownload,
 )
-from exo.shared.types.common import NodeId, SessionId, SystemId
+from exo.shared.types.common import NodeId
 from exo.shared.types.events import (
     Event,
-    EventId,
-    # TODO(evan): just for acks, should delete this ASAP
-    GlobalForwarderEvent,
-    LocalForwarderEvent,
     NodeDownloadProgress,
 )
 from exo.shared.types.worker.downloads import (
@@ -39,42 +33,28 @@ from exo.shared.types.worker.downloads import (
     DownloadProgress,
 )
 from exo.shared.types.worker.shards import PipelineShardMetadata, ShardMetadata
-from exo.utils.channels import Receiver, Sender, channel
+from exo.utils.channels import Receiver, Sender
 from exo.utils.task_group import TaskGroup
 
 
 @dataclass
 class DownloadCoordinator:
     node_id: NodeId
-    session_id: SessionId
     shard_downloader: ShardDownloader
     download_command_receiver: Receiver[ForwarderDownloadCommand]
-    local_event_sender: Sender[LocalForwarderEvent]
-
-    # ack stuff
-    _global_event_receiver: Receiver[GlobalForwarderEvent]
-    _out_for_delivery: dict[EventId, LocalForwarderEvent] = field(default_factory=dict)
-
+    event_sender: Sender[Event]
     offline: bool = False
 
-    _system_id: SystemId = field(default_factory=SystemId)
-
     # Local state
     download_status: dict[ModelId, DownloadProgress] = field(default_factory=dict)
     active_downloads: dict[ModelId, asyncio.Task[None]] = field(default_factory=dict)
 
-    # Internal event channel for forwarding (initialized in __post_init__)
-    event_sender: Sender[Event] = field(init=False)
-    event_receiver: Receiver[Event] = field(init=False)
     _tg: TaskGroup = field(init=False, default_factory=TaskGroup)
 
     # Per-model throttle for download progress events
     _last_progress_time: dict[ModelId, float] = field(default_factory=dict)
 
     def __post_init__(self) -> None:
-        self.event_sender, self.event_receiver = channel[Event]()
-        if self.offline:
-            self.shard_downloader.set_internet_connection(False)
         self.shard_downloader.on_progress(self._download_progress_callback)
 
     def _model_dir(self, model_id: ModelId) -> str:
@@ -123,66 +103,17 @@ class DownloadCoordinator:
         logger.info(
             f"Starting DownloadCoordinator{' (offline mode)' if self.offline else ''}"
         )
-        if not self.offline:
-            self._test_internet_connection()
         try:
             async with self._tg as tg:
                 tg.start_soon(self._command_processor)
-                tg.start_soon(self._forward_events)
                 tg.start_soon(self._emit_existing_download_progress)
-                tg.start_soon(self._resend_out_for_delivery)
-                tg.start_soon(self._clear_ofd)
-                if not self.offline:
-                    tg.start_soon(self._check_internet_connection)
         finally:
             for task in self.active_downloads.values():
                 task.cancel()
 
-    def _test_internet_connection(self) -> None:
-        # Try multiple endpoints since some ISPs/networks block specific IPs
-        for host in ("1.1.1.1", "8.8.8.8", "1.0.0.1"):
-            try:
-                socket.create_connection((host, 443), timeout=3).close()
-                self.shard_downloader.set_internet_connection(True)
-                logger.debug(f"Internet connectivity: True (via {host})")
-                return
-            except OSError:
-                continue
-        self.shard_downloader.set_internet_connection(False)
-        logger.debug("Internet connectivity: False")
-
-    async def _check_internet_connection(self) -> None:
-        first_connection = True
-        while True:
-            await asyncio.sleep(10)
-
-            # Assume that internet connection is set to False on 443 errors.
-            if self.shard_downloader.internet_connection:
-                continue
-
-            self._test_internet_connection()
-
-            if first_connection and self.shard_downloader.internet_connection:
-                first_connection = False
-                self._tg.start_soon(self._emit_existing_download_progress)
-
     def shutdown(self) -> None:
         self._tg.cancel_tasks()
 
-    # directly copied from worker
-    async def _resend_out_for_delivery(self) -> None:
-        # This can also be massively tightened, we should check events are at least a certain age before resending.
-        # Exponential backoff would also certainly help here.
-        while True:
-            await anyio.sleep(1 + random())
-            for event in self._out_for_delivery.copy().values():
-                await self.local_event_sender.send(event)
-
-    async def _clear_ofd(self) -> None:
-        with self._global_event_receiver as events:
-            async for event in events:
-                self._out_for_delivery.pop(event.event.event_id, None)
-
     async def _command_processor(self) -> None:
         with self.download_command_receiver as commands:
             async for cmd in commands:
@@ -355,23 +286,6 @@ class DownloadCoordinator:
             )
             del self.download_status[model_id]
 
-    async def _forward_events(self) -> None:
-        idx = 0
-        with self.event_receiver as events:
-            async for event in events:
-                fe = LocalForwarderEvent(
-                    origin_idx=idx,
-                    origin=self._system_id,
-                    session=self.session_id,
-                    event=event,
-                )
-                idx += 1
-                logger.debug(
-                    f"DownloadCoordinator published event {idx}: {str(event)[:100]}"
-                )
-                await self.local_event_sender.send(fe)
-                self._out_for_delivery[event.event_id] = fe
-
     async def _emit_existing_download_progress(self) -> None:
         try:
             while True:

src/exo/download/download_utils.py

@@ -314,9 +314,13 @@ async def fetch_file_list_with_cache(
         _fetched_file_lists_this_session.add(cache_key)
         return file_list
     except Exception as e:
+        logger.opt(exception=e).warning(
+            "Ran into exception when fetching file list from HF."
+        )
+
         if await aios.path.exists(cache_file):
             logger.warning(
-                f"No internet and no cached file list for {model_id} - using local file list"
+                f"No cached file list for {model_id} - using local file list"
             )
             async with aiofiles.open(cache_file, "r") as f:
                 return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
@@ -823,6 +827,7 @@ async def download_shard(
 
     for file in filtered_file_list:
         downloaded_bytes = await get_downloaded_size(target_dir / file.path)
+        final_file_exists = await aios.path.exists(target_dir / file.path)
         file_progress[file.path] = RepoFileDownloadProgress(
             repo_id=shard.model_card.model_id,
             repo_revision=revision,
@@ -832,7 +837,9 @@ async def download_shard(
             total=Memory.from_bytes(file.size or 0),
             speed=0,
             eta=timedelta(0),
-            status="complete" if downloaded_bytes == file.size else "not_started",
+            status="complete"
+            if final_file_exists and downloaded_bytes == file.size
+            else "not_started",
             start_time=time.time(),
         )
 

src/exo/download/impl_shard_downloader.py

@@ -15,9 +15,13 @@ from exo.shared.types.worker.shards import (
 )
 
 
-def exo_shard_downloader(max_parallel_downloads: int = 8) -> ShardDownloader:
+def exo_shard_downloader(
+    max_parallel_downloads: int = 8, offline: bool = False
+) -> ShardDownloader:
     return SingletonShardDownloader(
-        CachedShardDownloader(ResumableShardDownloader(max_parallel_downloads))
+        CachedShardDownloader(
+            ResumableShardDownloader(max_parallel_downloads, offline=offline)
+        )
     )
 
 
@@ -50,10 +54,6 @@ class SingletonShardDownloader(ShardDownloader):
         self.shard_downloader = shard_downloader
         self.active_downloads: dict[ShardMetadata, asyncio.Task[Path]] = {}
 
-    def set_internet_connection(self, value: bool) -> None:
-        self.internet_connection = value
-        self.shard_downloader.set_internet_connection(value)
-
     def on_progress(
         self,
         callback: Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]],
@@ -90,10 +90,6 @@ class CachedShardDownloader(ShardDownloader):
         self.shard_downloader = shard_downloader
         self.cache: dict[tuple[str, ShardMetadata], Path] = {}
 
-    def set_internet_connection(self, value: bool) -> None:
-        self.internet_connection = value
-        self.shard_downloader.set_internet_connection(value)
-
     def on_progress(
         self,
         callback: Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]],
@@ -123,8 +119,9 @@ class CachedShardDownloader(ShardDownloader):
 
 
 class ResumableShardDownloader(ShardDownloader):
-    def __init__(self, max_parallel_downloads: int = 8):
+    def __init__(self, max_parallel_downloads: int = 8, offline: bool = False):
         self.max_parallel_downloads = max_parallel_downloads
+        self.offline = offline
         self.on_progress_callbacks: list[
             Callable[[ShardMetadata, RepoDownloadProgress], Awaitable[None]]
         ] = []
@@ -151,8 +148,7 @@ class ResumableShardDownloader(ShardDownloader):
             self.on_progress_wrapper,
             max_parallel_downloads=self.max_parallel_downloads,
             allow_patterns=allow_patterns,
-            skip_internet=not self.internet_connection,
-            on_connection_lost=lambda: self.set_internet_connection(False),
+            skip_internet=self.offline,
         )
         return target_dir
 
@@ -168,8 +164,7 @@ class ResumableShardDownloader(ShardDownloader):
                 shard,
                 self.on_progress_wrapper,
                 skip_download=True,
-                skip_internet=not self.internet_connection,
-                on_connection_lost=lambda: self.set_internet_connection(False),
+                skip_internet=self.offline,
             )
 
         semaphore = asyncio.Semaphore(self.max_parallel_downloads)
@@ -198,7 +193,6 @@ class ResumableShardDownloader(ShardDownloader):
             shard,
             self.on_progress_wrapper,
             skip_download=True,
-            skip_internet=not self.internet_connection,
-            on_connection_lost=lambda: self.set_internet_connection(False),
+            skip_internet=self.offline,
         )
         return progress

src/exo/download/shard_downloader.py

@@ -16,11 +16,6 @@ from exo.shared.types.worker.shards import (
 
 # TODO: the PipelineShardMetadata getting reinstantiated is a bit messy. Should this be a classmethod?
 class ShardDownloader(ABC):
-    internet_connection: bool = False
-
-    def set_internet_connection(self, value: bool) -> None:
-        self.internet_connection = value
-
     @abstractmethod
     async def ensure_shard(
         self, shard: ShardMetadata, config_only: bool = False

src/exo/download/tests/test_coordinator_ack.py

@@ -1,98 +0,0 @@
-from typing import Any
-
-import anyio
-import pytest
-
-from exo.download.coordinator import DownloadCoordinator
-from exo.download.shard_downloader import NoopShardDownloader
-from exo.shared.models.model_cards import ModelCard, ModelTask
-from exo.shared.types.common import ModelId, NodeId, SessionId
-from exo.shared.types.events import (
-    GlobalForwarderEvent,
-    LocalForwarderEvent,
-    NodeDownloadProgress,
-)
-from exo.shared.types.memory import Memory
-from exo.shared.types.worker.downloads import (
-    DownloadPending,
-)
-from exo.shared.types.worker.shards import PipelineShardMetadata
-from exo.utils.channels import channel
-
-# Use the built‑in NoopShardDownloader directly – it already implements the required abstract interface.
-# No additional subclass is needed for this test.
-
-
-@pytest.mark.anyio
-async def test_ack_behaviour():
-    # Create channels (type Any for simplicity)
-    _, command_receiver = channel[Any]()
-    local_sender, _ = channel[Any]()
-    global_sender, global_receiver = channel[Any]()
-
-    # Minimal identifiers
-    node_id = NodeId()
-    session_id = SessionId(master_node_id=node_id, election_clock=0)
-
-    # Create a dummy model card and shard metadata
-    model_id = ModelId("test/model")
-    model_card = ModelCard(
-        model_id=model_id,
-        storage_size=Memory.from_bytes(0),
-        n_layers=1,
-        hidden_size=1,
-        supports_tensor=True,
-        tasks=[ModelTask.TextGeneration],
-    )
-    shard = PipelineShardMetadata(
-        model_card=model_card,
-        device_rank=0,
-        world_size=1,
-        start_layer=0,
-        end_layer=1,
-        n_layers=1,
-    )
-
-    # Instantiate the coordinator with the dummy downloader
-    coord = DownloadCoordinator(
-        node_id=node_id,
-        session_id=session_id,
-        shard_downloader=NoopShardDownloader(),
-        download_command_receiver=command_receiver,
-        local_event_sender=local_sender,
-        _global_event_receiver=global_receiver,
-    )
-
-    async with anyio.create_task_group() as tg:
-        # Start the forwarding and ack‑clearing loops
-        tg.start_soon(coord._forward_events)  # pyright: ignore[reportPrivateUsage]
-        tg.start_soon(coord._clear_ofd)  # pyright: ignore[reportPrivateUsage]
-
-        # Send a pending download progress event via the internal event sender
-        pending = DownloadPending(
-            node_id=node_id,
-            shard_metadata=shard,
-            model_directory="/tmp/model",
-        )
-        await coord.event_sender.send(NodeDownloadProgress(download_progress=pending))
-        # Allow the forwarder to process the event
-        await anyio.sleep(0.1)
-
-        # There should be exactly one entry awaiting ACK
-        assert len(coord._out_for_delivery) == 1  # pyright: ignore[reportPrivateUsage]
-        # Retrieve the stored LocalForwarderEvent
-        stored_fe: LocalForwarderEvent = next(iter(coord._out_for_delivery.values()))  # pyright: ignore[reportPrivateUsage]
-        # Simulate receiving a global ack for this event
-        ack = GlobalForwarderEvent(
-            origin_idx=0,
-            origin=node_id,
-            session=session_id,
-            event=stored_fe.event,
-        )
-        await global_sender.send(ack)
-        # Give the clear‑ofd task a moment to process the ack
-        await anyio.sleep(0.1)
-        # The out‑for‑delivery map should now be empty
-        assert len(coord._out_for_delivery) == 0  # pyright: ignore[reportPrivateUsage]
-        # Cancel background tasks
-        tg.cancel_scope.cancel()

src/exo/main.py

@@ -15,6 +15,7 @@ from exo.download.coordinator import DownloadCoordinator
 from exo.download.impl_shard_downloader import exo_shard_downloader
 from exo.master.api import API  # TODO: should API be in master?
 from exo.master.main import Master
+from exo.routing.event_router import EventRouter
 from exo.routing.router import Router, get_node_id_keypair
 from exo.shared.constants import EXO_LOG
 from exo.shared.election import Election, ElectionResult
@@ -29,6 +30,7 @@ from exo.worker.main import Worker
 @dataclass
 class Node:
     router: Router
+    event_router: EventRouter
     download_coordinator: DownloadCoordinator | None
     worker: Worker | None
     election: Election  # Every node participates in election, as we do want a node to become master even if it isn't a master candidate if no master candidates are present.
@@ -52,6 +54,12 @@ class Node:
         await router.register_topic(topics.ELECTION_MESSAGES)
         await router.register_topic(topics.CONNECTION_MESSAGES)
         await router.register_topic(topics.DOWNLOAD_COMMANDS)
+        event_router = EventRouter(
+            session_id,
+            command_sender=router.sender(topics.COMMANDS),
+            external_outbound=router.sender(topics.LOCAL_EVENTS),
+            external_inbound=router.receiver(topics.GLOBAL_EVENTS),
+        )
 
         logger.info(f"Starting node {node_id}")
 
@@ -59,13 +67,10 @@ class Node:
         if not args.no_downloads:
             download_coordinator = DownloadCoordinator(
                 node_id,
-                session_id,
-                exo_shard_downloader(),
+                exo_shard_downloader(offline=args.offline),
+                event_sender=event_router.sender(),
                 download_command_receiver=router.receiver(topics.DOWNLOAD_COMMANDS),
-                local_event_sender=router.sender(topics.LOCAL_EVENTS),
                 offline=args.offline,
-                # TODO(evan): remove
-                _global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
             )
         else:
             download_coordinator = None
@@ -73,9 +78,8 @@ class Node:
         if args.spawn_api:
             api = API(
                 node_id,
-                session_id,
                 port=args.api_port,
-                global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
+                event_receiver=event_router.receiver(),
                 command_sender=router.sender(topics.COMMANDS),
                 download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                 election_receiver=router.receiver(topics.ELECTION_MESSAGES),
@@ -86,9 +90,8 @@ class Node:
         if not args.no_worker:
             worker = Worker(
                 node_id,
-                session_id,
-                global_event_receiver=router.receiver(topics.GLOBAL_EVENTS),
-                local_event_sender=router.sender(topics.LOCAL_EVENTS),
+                event_receiver=event_router.receiver(),
+                event_sender=event_router.sender(),
                 command_sender=router.sender(topics.COMMANDS),
                 download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
             )
@@ -99,6 +102,7 @@ class Node:
         master = Master(
             node_id,
             session_id,
+            event_sender=event_router.sender(),
             global_event_sender=router.sender(topics.GLOBAL_EVENTS),
             local_event_receiver=router.receiver(topics.LOCAL_EVENTS),
             command_receiver=router.receiver(topics.COMMANDS),
@@ -121,6 +125,7 @@ class Node:
 
         return cls(
             router,
+            event_router,
             download_coordinator,
             worker,
             election,
@@ -136,6 +141,7 @@ class Node:
             signal.signal(signal.SIGINT, lambda _, __: self.shutdown())
             signal.signal(signal.SIGTERM, lambda _, __: self.shutdown())
             tg.start_soon(self.router.run)
+            tg.start_soon(self.event_router.run)
             tg.start_soon(self.election.run)
             if self.download_coordinator:
                 tg.start_soon(self.download_coordinator.run)
@@ -170,6 +176,17 @@ class Node:
                 # - Shutdown and re-create the worker
                 # - Shut down and re-create the API
 
+                if result.is_new_master:
+                    await anyio.sleep(0)
+                    self.event_router.shutdown()
+                    self.event_router = EventRouter(
+                        result.session_id,
+                        self.router.sender(topics.COMMANDS),
+                        self.router.receiver(topics.GLOBAL_EVENTS),
+                        self.router.sender(topics.LOCAL_EVENTS),
+                    )
+                    self._tg.start_soon(self.event_router.run)
+
                 if (
                     result.session_id.master_node_id == self.node_id
                     and self.master is not None
@@ -183,6 +200,7 @@ class Node:
                     self.master = Master(
                         self.node_id,
                         result.session_id,
+                        event_sender=self.event_router.sender(),
                         global_event_sender=self.router.sender(topics.GLOBAL_EVENTS),
                         local_event_receiver=self.router.receiver(topics.LOCAL_EVENTS),
                         command_receiver=self.router.receiver(topics.COMMANDS),
@@ -205,22 +223,16 @@ class Node:
                         f"Node {result.session_id.master_node_id} elected master"
                     )
                 if result.is_new_master:
-                    await anyio.sleep(0)
                     if self.download_coordinator:
                         self.download_coordinator.shutdown()
                         self.download_coordinator = DownloadCoordinator(
                             self.node_id,
-                            result.session_id,
-                            exo_shard_downloader(),
+                            exo_shard_downloader(offline=self.offline),
+                            event_sender=self.event_router.sender(),
                             download_command_receiver=self.router.receiver(
                                 topics.DOWNLOAD_COMMANDS
                             ),
-                            local_event_sender=self.router.sender(topics.LOCAL_EVENTS),
                             offline=self.offline,
-                            # TODO(evan): remove
-                            _global_event_receiver=self.router.receiver(
-                                topics.GLOBAL_EVENTS
-                            ),
                         )
                         self._tg.start_soon(self.download_coordinator.run)
                     if self.worker:
@@ -228,11 +240,8 @@ class Node:
                         # TODO: add profiling etc to resource monitor
                         self.worker = Worker(
                             self.node_id,
-                            result.session_id,
-                            global_event_receiver=self.router.receiver(
-                                topics.GLOBAL_EVENTS
-                            ),
-                            local_event_sender=self.router.sender(topics.LOCAL_EVENTS),
+                            event_receiver=self.event_router.receiver(),
+                            event_sender=self.event_router.sender(),
                             command_sender=self.router.sender(topics.COMMANDS),
                             download_command_sender=self.router.sender(
                                 topics.DOWNLOAD_COMMANDS
@@ -240,7 +249,7 @@ class Node:
                         )
                         self._tg.start_soon(self.worker.run)
                     if self.api:
-                        self.api.reset(result.session_id, result.won_clock)
+                        self.api.reset(result.won_clock, self.event_router.receiver())
                 else:
                     if self.api:
                         self.api.unpause(result.won_clock)
@@ -252,7 +261,7 @@ 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")
@@ -270,9 +279,16 @@ def main():
         logger.info("FAST_SYNCH forced OFF")
 
     node = anyio.run(Node.create, args)
-    anyio.run(node.run)
-    logger.info("EXO Shutdown complete")
-    logger_cleanup()
+    try:
+        anyio.run(node.run)
+    except BaseException as exception:
+        logger.opt(exception=exception).critical(
+            "EXO terminated due to unhandled exception"
+        )
+        raise
+    finally:
+        logger.info("EXO Shutdown complete")
+        logger_cleanup()
 
 
 class Args(CamelCaseModel):
@@ -283,7 +299,7 @@ class Args(CamelCaseModel):
     tb_only: bool = False
     no_worker: bool = False
     no_downloads: bool = False
-    offline: bool = False
+    offline: bool = os.getenv("EXO_OFFLINE", "false").lower() == "true"
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
 
     @classmethod
@@ -334,6 +350,7 @@ class Args(CamelCaseModel):
         parser.add_argument(
             "--offline",
             action="store_true",
+            default=os.getenv("EXO_OFFLINE", "false").lower() == "true",
             help="Run in offline/air-gapped mode: skip internet checks, use only pre-staged local models",
         )
         fast_synch_group = parser.add_mutually_exclusive_group()

src/exo/master/api.py

@@ -73,6 +73,7 @@ from exo.shared.types.api import (
     BenchChatCompletionResponse,
     BenchImageGenerationResponse,
     BenchImageGenerationTaskParams,
+    CancelCommandResponse,
     ChatCompletionChoice,
     ChatCompletionMessage,
     ChatCompletionRequest,
@@ -81,6 +82,8 @@ from exo.shared.types.api import (
     CreateInstanceResponse,
     DeleteDownloadResponse,
     DeleteInstanceResponse,
+    DeleteTracesRequest,
+    DeleteTracesResponse,
     ErrorInfo,
     ErrorResponse,
     FinishReason,
@@ -140,11 +143,10 @@ from exo.shared.types.commands import (
     TaskFinished,
     TextGeneration,
 )
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId, SystemId
+from exo.shared.types.common import CommandId, Id, NodeId, SystemId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
-    GlobalForwarderEvent,
     IndexedEvent,
     TracesMerged,
 )
@@ -172,7 +174,6 @@ 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
 from exo.utils.channels import Receiver, Sender, channel
-from exo.utils.event_buffer import OrderedBuffer
 from exo.utils.task_group import TaskGroup
 
 _API_EVENT_LOG_DIR = EXO_EVENT_LOG_DIR / "api"
@@ -196,10 +197,9 @@ class API:
     def __init__(
         self,
         node_id: NodeId,
-        session_id: SessionId,
         *,
         port: int,
-        global_event_receiver: Receiver[GlobalForwarderEvent],
+        event_receiver: Receiver[IndexedEvent],
         command_sender: Sender[ForwarderCommand],
         download_command_sender: Sender[ForwarderDownloadCommand],
         # This lets us pause the API if an election is running
@@ -210,11 +210,9 @@ class API:
         self._system_id = SystemId()
         self.command_sender = command_sender
         self.download_command_sender = download_command_sender
-        self.global_event_receiver = global_event_receiver
+        self.event_receiver = event_receiver
         self.election_receiver = election_receiver
-        self.event_buffer: OrderedBuffer[Event] = OrderedBuffer[Event]()
         self.node_id: NodeId = node_id
-        self.session_id: SessionId = session_id
         self.last_completed_election: int = 0
         self.port = port
 
@@ -254,17 +252,18 @@ class API:
         self._image_store = ImageStore(EXO_IMAGE_CACHE_DIR)
         self._tg: TaskGroup = TaskGroup()
 
-    def reset(self, new_session_id: SessionId, result_clock: int):
+    def reset(self, result_clock: int, event_receiver: Receiver[IndexedEvent]):
         logger.info("Resetting API State")
         self._event_log.close()
         self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
         self.state = State()
         self._system_id = SystemId()
-        self.session_id = new_session_id
-        self.event_buffer = OrderedBuffer[Event]()
         self._text_generation_queues = {}
         self._image_generation_queues = {}
         self.unpause(result_clock)
+        self.event_receiver.close()
+        self.event_receiver = event_receiver
+        self._tg.start_soon(self._apply_state)
 
     def unpause(self, result_clock: int):
         logger.info("Unpausing API")
@@ -324,6 +323,7 @@ class API:
         self.app.get("/images/{image_id}")(self.get_image)
         self.app.post("/v1/messages", response_model=None)(self.claude_messages)
         self.app.post("/v1/responses", response_model=None)(self.openai_responses)
+        self.app.post("/v1/cancel/{command_id}")(self.cancel_command)
 
         # Ollama API
         self.app.head("/ollama/")(self.ollama_version)
@@ -344,6 +344,7 @@ class API:
         self.app.post("/download/start")(self.start_download)
         self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
         self.app.get("/v1/traces")(self.list_traces)
+        self.app.post("/v1/traces/delete")(self.delete_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)
@@ -568,6 +569,25 @@ class API:
             instance_id=instance_id,
         )
 
+    async def cancel_command(self, command_id: CommandId) -> CancelCommandResponse:
+        """Cancel an active command by closing its stream and notifying workers."""
+        sender = self._text_generation_queues.get(
+            command_id
+        ) or self._image_generation_queues.get(command_id)
+        if sender is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Command not found or already completed",
+            )
+
+        await self._send(TaskCancelled(cancelled_command_id=command_id))
+        sender.close()
+
+        return CancelCommandResponse(
+            message="Command cancelled.",
+            command_id=command_id,
+        )
+
     async def _token_chunk_stream(
         self, command_id: CommandId
     ) -> AsyncGenerator[
@@ -1606,7 +1626,7 @@ class API:
         finally:
             self._event_log.close()
             self.command_sender.close()
-            self.global_event_receiver.close()
+            self.event_receiver.close()
 
     async def run_api(self, ev: anyio.Event):
         cfg = Config()
@@ -1623,38 +1643,31 @@ class API:
             )
 
     async def _apply_state(self):
-        with self.global_event_receiver as events:
-            async for f_event in events:
-                if f_event.session != self.session_id:
-                    continue
-                if f_event.origin != self.session_id.master_node_id:
-                    continue
-                self.event_buffer.ingest(f_event.origin_idx, f_event.event)
-                for idx, event in self.event_buffer.drain_indexed():
-                    self._event_log.append(event)
-                    self.state = apply(self.state, IndexedEvent(event=event, idx=idx))
+        with self.event_receiver as events:
+            async for i_event in events:
+                self._event_log.append(i_event.event)
+                self.state = apply(self.state, i_event)
+                event = i_event.event
 
-                    if isinstance(event, ChunkGenerated):
-                        if queue := self._image_generation_queues.get(
-                            event.command_id, None
-                        ):
-                            assert isinstance(event.chunk, ImageChunk)
-                            try:
-                                await queue.send(event.chunk)
-                            except BrokenResourceError:
-                                self._image_generation_queues.pop(
-                                    event.command_id, None
-                                )
-                        if queue := self._text_generation_queues.get(
-                            event.command_id, None
-                        ):
-                            assert not isinstance(event.chunk, ImageChunk)
-                            try:
-                                await queue.send(event.chunk)
-                            except BrokenResourceError:
-                                self._text_generation_queues.pop(event.command_id, None)
-                    if isinstance(event, TracesMerged):
-                        self._save_merged_trace(event)
+                if isinstance(event, ChunkGenerated):
+                    if queue := self._image_generation_queues.get(
+                        event.command_id, None
+                    ):
+                        assert isinstance(event.chunk, ImageChunk)
+                        try:
+                            await queue.send(event.chunk)
+                        except BrokenResourceError:
+                            self._image_generation_queues.pop(event.command_id, None)
+                    if queue := self._text_generation_queues.get(
+                        event.command_id, None
+                    ):
+                        assert not isinstance(event.chunk, ImageChunk)
+                        try:
+                            await queue.send(event.chunk)
+                        except BrokenResourceError:
+                            self._text_generation_queues.pop(event.command_id, None)
+                if isinstance(event, TracesMerged):
+                    self._save_merged_trace(event)
 
     def _save_merged_trace(self, event: TracesMerged) -> None:
         traces = [
@@ -1718,8 +1731,12 @@ class API:
         await self._send_download(command)
         return DeleteDownloadResponse(command_id=command.command_id)
 
-    def _get_trace_path(self, task_id: str) -> Path:
-        return EXO_TRACING_CACHE_DIR / f"trace_{task_id}.json"
+    @staticmethod
+    def _get_trace_path(task_id: str) -> Path:
+        trace_path = EXO_TRACING_CACHE_DIR / f"trace_{task_id}.json"
+        if not trace_path.resolve().is_relative_to(EXO_TRACING_CACHE_DIR.resolve()):
+            raise HTTPException(status_code=400, detail=f"Invalid task ID: {task_id}")
+        return trace_path
 
     async def list_traces(self) -> TraceListResponse:
         traces: list[TraceListItem] = []
@@ -1818,6 +1835,18 @@ class API:
             filename=f"trace_{task_id}.json",
         )
 
+    async def delete_traces(self, request: DeleteTracesRequest) -> DeleteTracesResponse:
+        deleted: list[str] = []
+        not_found: list[str] = []
+        for task_id in request.task_ids:
+            trace_path = self._get_trace_path(task_id)
+            if trace_path.exists():
+                trace_path.unlink()
+                deleted.append(task_id)
+            else:
+                not_found.append(task_id)
+        return DeleteTracesResponse(deleted=deleted, not_found=not_found)
+
     async def get_onboarding(self) -> JSONResponse:
         return JSONResponse({"completed": ONBOARDING_COMPLETE_FILE.exists()})
 

src/exo/master/main.py

@@ -60,7 +60,7 @@ 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.utils.channels import Receiver, Sender
 from exo.utils.event_buffer import MultiSourceBuffer
 from exo.utils.task_group import TaskGroup
 
@@ -72,25 +72,21 @@ class Master:
         session_id: SessionId,
         *,
         command_receiver: Receiver[ForwarderCommand],
+        event_sender: Sender[Event],
         local_event_receiver: Receiver[LocalForwarderEvent],
         global_event_sender: Sender[GlobalForwarderEvent],
         download_command_sender: Sender[ForwarderDownloadCommand],
     ):
-        self.state = State()
-        self._tg: TaskGroup = TaskGroup()
         self.node_id = node_id
         self.session_id = session_id
+        self.state = State()
+        self._tg: TaskGroup = TaskGroup()
         self.command_task_mapping: dict[CommandId, TaskId] = {}
         self.command_receiver = command_receiver
         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[LocalForwarderEvent] = (
-            local_event_receiver.clone_sender()
-        )
+        self.event_sender = event_sender
         self._system_id = SystemId()
         self._multi_buffer = MultiSourceBuffer[SystemId, Event]()
         self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
@@ -104,15 +100,12 @@ 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)
         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")
@@ -409,22 +402,6 @@ class Master:
                     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(
-                    LocalForwarderEvent(
-                        origin=self._system_id,
-                        origin_idx=local_index,
-                        session=self.session_id,
-                        event=event,
-                    )
-                )
-                local_index += 1
-
     # This function is re-entrant, take care!
     async def _send_event(self, event: IndexedEvent):
         # Convenience method since this line is ugly

src/exo/master/tests/test_cancel_command.py

@@ -0,0 +1,77 @@
+# pyright: reportUnusedFunction=false, reportAny=false
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from exo.shared.types.common import CommandId
+
+
+def _make_api() -> Any:
+    """Create a minimal API instance with cancel route and error handler."""
+    from exo.master.api import API
+
+    app = FastAPI()
+    api = object.__new__(API)
+    api.app = app
+    api._text_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
+    api._image_generation_queues = {}  # pyright: ignore[reportPrivateUsage]
+    api._send = AsyncMock()  # pyright: ignore[reportPrivateUsage]
+    api._setup_exception_handlers()  # pyright: ignore[reportPrivateUsage]
+    app.post("/v1/cancel/{command_id}")(api.cancel_command)
+    return api
+
+
+def test_cancel_nonexistent_command_returns_404() -> None:
+    """Cancel for an unknown command_id returns 404 in OpenAI error format."""
+    api = _make_api()
+    client = TestClient(api.app)
+
+    response = client.post("/v1/cancel/nonexistent-id")
+    assert response.status_code == 404
+    data: dict[str, Any] = response.json()
+    assert "error" in data
+    assert data["error"]["message"] == "Command not found or already completed"
+    assert data["error"]["type"] == "Not Found"
+    assert data["error"]["code"] == 404
+
+
+def test_cancel_active_text_generation() -> None:
+    """Cancel an active text generation command: returns 200, sender.close() called."""
+    api = _make_api()
+    client = TestClient(api.app)
+
+    cid = CommandId("text-cmd-123")
+    sender = MagicMock()
+    api._text_generation_queues[cid] = sender
+
+    response = client.post(f"/v1/cancel/{cid}")
+    assert response.status_code == 200
+    data: dict[str, Any] = response.json()
+    assert data["message"] == "Command cancelled."
+    assert data["command_id"] == str(cid)
+    sender.close.assert_called_once()
+    api._send.assert_called_once()
+    task_cancelled = api._send.call_args[0][0]
+    assert task_cancelled.cancelled_command_id == cid
+
+
+def test_cancel_active_image_generation() -> None:
+    """Cancel an active image generation command: returns 200, sender.close() called."""
+    api = _make_api()
+    client = TestClient(api.app)
+
+    cid = CommandId("img-cmd-456")
+    sender = MagicMock()
+    api._image_generation_queues[cid] = sender
+
+    response = client.post(f"/v1/cancel/{cid}")
+    assert response.status_code == 200
+    data: dict[str, Any] = response.json()
+    assert data["message"] == "Command cancelled."
+    assert data["command_id"] == str(cid)
+    sender.close.assert_called_once()
+    api._send.assert_called_once()
+    task_cancelled = api._send.call_args[0][0]
+    assert task_cancelled.cancelled_command_id == cid

src/exo/master/tests/test_master.py

@@ -17,6 +17,7 @@ from exo.shared.types.commands import (
 )
 from exo.shared.types.common import ModelId, NodeId, SessionId, SystemId
 from exo.shared.types.events import (
+    Event,
     GlobalForwarderEvent,
     IndexedEvent,
     InstanceCreated,
@@ -50,6 +51,22 @@ async def test_master():
     command_sender, co_receiver = channel[ForwarderCommand]()
     local_event_sender, le_receiver = channel[LocalForwarderEvent]()
     fcds, _fcdr = channel[ForwarderDownloadCommand]()
+    ev_send, ev_recv = channel[Event]()
+
+    async def mock_event_router():
+        idx = 0
+        sid = SystemId()
+        with ev_recv as master_events:
+            async for event in master_events:
+                await local_event_sender.send(
+                    LocalForwarderEvent(
+                        origin=sid,
+                        origin_idx=idx,
+                        session=session_id,
+                        event=event,
+                    )
+                )
+                idx += 1
 
     all_events: list[IndexedEvent] = []
 
@@ -67,6 +84,7 @@ async def test_master():
     master = Master(
         node_id,
         session_id,
+        event_sender=ev_send,
         global_event_sender=ge_sender,
         local_event_receiver=le_receiver,
         command_receiver=co_receiver,
@@ -75,6 +93,7 @@ async def test_master():
     logger.info("run the master")
     async with anyio.create_task_group() as tg:
         tg.start_soon(master.run)
+        tg.start_soon(mock_event_router)
 
         # inject a NodeGatheredInfo event
         logger.info("inject a NodeGatheredInfo event")
@@ -197,4 +216,5 @@ async def test_master():
             input=[InputMessage(role="user", content="Hello, how are you?")],
         )
 
+        ev_send.close()
         await master.shutdown()

src/exo/master/tests/test_placement.py

@@ -14,10 +14,12 @@ from exo.shared.models.model_cards import ModelCard, ModelId, ModelTask
 from exo.shared.topology import Topology
 from exo.shared.types.commands import PlaceInstance
 from exo.shared.types.common import CommandId, NodeId
-from exo.shared.types.events import InstanceCreated, InstanceDeleted
+from exo.shared.types.events import InstanceCreated, InstanceDeleted, TaskStatusUpdated
 from exo.shared.types.memory import Memory
 from exo.shared.types.multiaddr import Multiaddr
 from exo.shared.types.profiling import NetworkInterfaceInfo, NodeNetworkInfo
+from exo.shared.types.tasks import TaskId, TaskStatus, TextGeneration
+from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.shared.types.topology import Connection, SocketConnection
 from exo.shared.types.worker.instances import (
     Instance,
@@ -456,3 +458,117 @@ def test_tensor_rdma_backend_connectivity_matrix(
         else:
             ip_part = coordinator.split(":")[0]
             assert len(ip_part.split(".")) == 4
+
+
+def _make_task(
+    instance_id: InstanceId,
+    status: TaskStatus = TaskStatus.Running,
+) -> TextGeneration:
+    return TextGeneration(
+        task_id=TaskId(),
+        task_status=status,
+        instance_id=instance_id,
+        command_id=CommandId(),
+        task_params=TextGenerationTaskParams(
+            model=ModelId("test-model"),
+            input=[InputMessage(role="user", content="hello")],
+        ),
+    )
+
+
+def test_get_transition_events_delete_instance_cancels_running_tasks(
+    instance: Instance,
+):
+    # arrange
+    instance_id = InstanceId()
+    current_instances: dict[InstanceId, Instance] = {instance_id: instance}
+    target_instances: dict[InstanceId, Instance] = {}
+    task = _make_task(instance_id, TaskStatus.Running)
+    tasks = {task.task_id: task}
+
+    # act
+    events = get_transition_events(current_instances, target_instances, tasks)
+
+    # assert – cancellation event should come before the deletion event
+    assert len(events) == 2
+    assert isinstance(events[0], TaskStatusUpdated)
+    assert events[0].task_id == task.task_id
+    assert events[0].task_status == TaskStatus.Cancelled
+    assert isinstance(events[1], InstanceDeleted)
+    assert events[1].instance_id == instance_id
+
+
+def test_get_transition_events_delete_instance_cancels_pending_tasks(
+    instance: Instance,
+):
+    # arrange
+    instance_id = InstanceId()
+    current_instances: dict[InstanceId, Instance] = {instance_id: instance}
+    target_instances: dict[InstanceId, Instance] = {}
+    task = _make_task(instance_id, TaskStatus.Pending)
+    tasks = {task.task_id: task}
+
+    # act
+    events = get_transition_events(current_instances, target_instances, tasks)
+
+    # assert
+    assert len(events) == 2
+    assert isinstance(events[0], TaskStatusUpdated)
+    assert events[0].task_id == task.task_id
+    assert events[0].task_status == TaskStatus.Cancelled
+    assert isinstance(events[1], InstanceDeleted)
+
+
+def test_get_transition_events_delete_instance_ignores_completed_tasks(
+    instance: Instance,
+):
+    # arrange
+    instance_id = InstanceId()
+    current_instances: dict[InstanceId, Instance] = {instance_id: instance}
+    target_instances: dict[InstanceId, Instance] = {}
+    tasks = {
+        t.task_id: t
+        for t in [
+            _make_task(instance_id, TaskStatus.Complete),
+            _make_task(instance_id, TaskStatus.Failed),
+            _make_task(instance_id, TaskStatus.TimedOut),
+            _make_task(instance_id, TaskStatus.Cancelled),
+        ]
+    }
+
+    # act
+    events = get_transition_events(current_instances, target_instances, tasks)
+
+    # assert – only the InstanceDeleted event, no cancellations
+    assert len(events) == 1
+    assert isinstance(events[0], InstanceDeleted)
+
+
+def test_get_transition_events_delete_instance_cancels_only_matching_tasks(
+    instance: Instance,
+):
+    # arrange
+    instance_id_a = InstanceId()
+    instance_id_b = InstanceId()
+    current_instances: dict[InstanceId, Instance] = {
+        instance_id_a: instance,
+        instance_id_b: instance,
+    }
+    # only delete instance A, keep instance B
+    target_instances: dict[InstanceId, Instance] = {instance_id_b: instance}
+
+    task_a = _make_task(instance_id_a, TaskStatus.Running)
+    task_b = _make_task(instance_id_b, TaskStatus.Running)
+    tasks = {task_a.task_id: task_a, task_b.task_id: task_b}
+
+    # act
+    events = get_transition_events(current_instances, target_instances, tasks)
+
+    # assert – only task_a should be cancelled
+    cancel_events = [e for e in events if isinstance(e, TaskStatusUpdated)]
+    delete_events = [e for e in events if isinstance(e, InstanceDeleted)]
+    assert len(cancel_events) == 1
+    assert cancel_events[0].task_id == task_a.task_id
+    assert cancel_events[0].task_status == TaskStatus.Cancelled
+    assert len(delete_events) == 1
+    assert delete_events[0].instance_id == instance_id_a

src/exo/routing/connection_message.py

@@ -1,6 +1,4 @@
-from enum import Enum
-
-from exo_pyo3_bindings import ConnectionUpdate, ConnectionUpdateType
+from exo_pyo3_bindings import PyFromSwarm
 
 from exo.shared.types.common import NodeId
 from exo.utils.pydantic_ext import CamelCaseModel
@@ -8,30 +6,10 @@ from exo.utils.pydantic_ext import CamelCaseModel
 """Serialisable types for Connection Updates/Messages"""
 
 
-class ConnectionMessageType(Enum):
-    Connected = 0
-    Disconnected = 1
-
-    @staticmethod
-    def from_update_type(update_type: ConnectionUpdateType):
-        match update_type:
-            case ConnectionUpdateType.Connected:
-                return ConnectionMessageType.Connected
-            case ConnectionUpdateType.Disconnected:
-                return ConnectionMessageType.Disconnected
-
-
 class ConnectionMessage(CamelCaseModel):
     node_id: NodeId
-    connection_type: ConnectionMessageType
-    remote_ipv4: str
-    remote_tcp_port: int
+    connected: bool
 
     @classmethod
-    def from_update(cls, update: ConnectionUpdate) -> "ConnectionMessage":
-        return cls(
-            node_id=NodeId(update.peer_id),
-            connection_type=ConnectionMessageType.from_update_type(update.update_type),
-            remote_ipv4=update.remote_ipv4,
-            remote_tcp_port=update.remote_tcp_port,
-        )
+    def from_update(cls, update: PyFromSwarm.Connection) -> "ConnectionMessage":
+        return cls(node_id=NodeId(update.peer_id), connected=update.connected)

src/exo/routing/event_router.py

@@ -0,0 +1,161 @@
+from dataclasses import dataclass, field
+from random import random
+
+import anyio
+from anyio import BrokenResourceError, ClosedResourceError
+from anyio.abc import CancelScope
+from loguru import logger
+
+from exo.shared.types.commands import ForwarderCommand, RequestEventLog
+from exo.shared.types.common import SessionId, SystemId
+from exo.shared.types.events import (
+    Event,
+    EventId,
+    GlobalForwarderEvent,
+    IndexedEvent,
+    LocalForwarderEvent,
+)
+from exo.utils.channels import Receiver, Sender, channel
+from exo.utils.event_buffer import OrderedBuffer
+from exo.utils.task_group import TaskGroup
+
+
+@dataclass
+class EventRouter:
+    session_id: SessionId
+    command_sender: Sender[ForwarderCommand]
+    external_inbound: Receiver[GlobalForwarderEvent]
+    external_outbound: Sender[LocalForwarderEvent]
+    _system_id: SystemId = field(init=False, default_factory=SystemId)
+    internal_outbound: list[Sender[IndexedEvent]] = field(
+        init=False, default_factory=list
+    )
+    event_buffer: OrderedBuffer[Event] = field(
+        init=False, default_factory=OrderedBuffer
+    )
+    out_for_delivery: dict[EventId, tuple[float, LocalForwarderEvent]] = field(
+        init=False, default_factory=dict
+    )
+    _tg: TaskGroup = field(init=False, default_factory=TaskGroup)
+
+    _nack_cancel_scope: CancelScope | None = field(init=False, default=None)
+    _nack_attempts: int = field(init=False, default=0)
+    _nack_base_seconds: float = field(init=False, default=0.5)
+    _nack_cap_seconds: float = field(init=False, default=10.0)
+
+    async def run(self):
+        try:
+            async with self._tg as tg:
+                tg.start_soon(self._run_ext_in)
+                tg.start_soon(self._simple_retry)
+        finally:
+            self.external_outbound.close()
+            for send in self.internal_outbound:
+                send.close()
+
+    # can make this better in future
+    async def _simple_retry(self):
+        while True:
+            await anyio.sleep(1 + random())
+            # list here is a shallow clone for shared mutation
+            for e_id, (time, event) in list(self.out_for_delivery.items()):
+                if anyio.current_time() > time + 5:
+                    self.out_for_delivery[e_id] = (anyio.current_time(), event)
+                    await self.external_outbound.send(event)
+
+    def sender(self) -> Sender[Event]:
+        send, recv = channel[Event]()
+        if self._tg.is_running():
+            self._tg.start_soon(self._ingest, SystemId(), recv)
+        else:
+            self._tg.queue(self._ingest, SystemId(), recv)
+        return send
+
+    def receiver(self) -> Receiver[IndexedEvent]:
+        send, recv = channel[IndexedEvent]()
+        self.internal_outbound.append(send)
+        return recv
+
+    def shutdown(self) -> None:
+        self._tg.cancel_tasks()
+
+    async def _ingest(self, system_id: SystemId, recv: Receiver[Event]):
+        idx = 0
+        with recv as events:
+            async for event in events:
+                f_ev = LocalForwarderEvent(
+                    origin_idx=idx,
+                    origin=system_id,
+                    session=self.session_id,
+                    event=event,
+                )
+                idx += 1
+                await self.external_outbound.send(f_ev)
+                self.out_for_delivery[event.event_id] = (anyio.current_time(), f_ev)
+
+    async def _run_ext_in(self):
+        buf = OrderedBuffer[Event]()
+        with self.external_inbound as events:
+            async for event in events:
+                if event.session != self.session_id:
+                    continue
+                if event.origin != self.session_id.master_node_id:
+                    continue
+
+                buf.ingest(event.origin_idx, event.event)
+                event_id = event.event.event_id
+                if event_id in self.out_for_delivery:
+                    self.out_for_delivery.pop(event_id)
+
+                drained = buf.drain_indexed()
+                if drained:
+                    self._nack_attempts = 0
+                    if self._nack_cancel_scope:
+                        self._nack_cancel_scope.cancel()
+
+                if not drained and (
+                    self._nack_cancel_scope is None
+                    or self._nack_cancel_scope.cancel_called
+                ):
+                    # Request the next index.
+                    self._tg.start_soon(self._nack_request, buf.next_idx_to_release)
+                    continue
+
+                for idx, event in drained:
+                    to_clear = set[int]()
+                    for i, sender in enumerate(self.internal_outbound):
+                        try:
+                            await sender.send(IndexedEvent(idx=idx, event=event))
+                        except (ClosedResourceError, BrokenResourceError):
+                            to_clear.add(i)
+                    for i in sorted(to_clear, reverse=True):
+                        self.internal_outbound.pop(i)
+
+    async def _nack_request(self, since_idx: int) -> None:
+        # We request all events after (and including) the missing index.
+        # This function is started whenever we receive an event that is out of sequence.
+        # It is cancelled as soon as we receiver an event that is in sequence.
+
+        if since_idx < 0:
+            logger.warning(f"Negative value encountered for nack request {since_idx=}")
+            since_idx = 0
+
+        with CancelScope() as scope:
+            self._nack_cancel_scope = scope
+            delay: float = self._nack_base_seconds * (2.0**self._nack_attempts)
+            delay = min(self._nack_cap_seconds, delay)
+            self._nack_attempts += 1
+            try:
+                await anyio.sleep(delay)
+                logger.info(
+                    f"Nack attempt {self._nack_attempts}: Requesting Event Log from {since_idx}"
+                )
+                await self.command_sender.send(
+                    ForwarderCommand(
+                        origin=self._system_id,
+                        command=RequestEventLog(since_idx=since_idx),
+                    )
+                )
+            finally:
+                if self._nack_cancel_scope is scope:
+                    self._nack_cancel_scope = None

src/exo/routing/router.py

@@ -17,6 +17,7 @@ from exo_pyo3_bindings import (
     MessageTooLargeError,
     NetworkingHandle,
     NoPeersSubscribedToTopicError,
+    PyFromSwarm,
 )
 from filelock import FileLock
 from loguru import logger
@@ -121,7 +122,8 @@ class Router:
             send = self.networking_receiver.clone_sender()
         router = TopicRouter[T](topic, send)
         self.topic_routers[topic.topic] = cast(TopicRouter[CamelCaseModel], router)
-        await self._networking_subscribe(str(topic.topic))
+        if self._tg.is_running():
+            await self._networking_subscribe(topic.topic)
 
     def sender[T: CamelCaseModel](self, topic: TypedTopic[T]) -> Sender[T]:
         router = self.topic_routers.get(topic.topic, None)
@@ -152,8 +154,10 @@ class Router:
                     router = self.topic_routers[topic]
                     tg.start_soon(router.run)
                 tg.start_soon(self._networking_recv)
-                tg.start_soon(self._networking_recv_connection_messages)
                 tg.start_soon(self._networking_publish)
+                # subscribe to pending topics
+                for topic in self.topic_routers:
+                    await self._networking_subscribe(topic)
                 # Router only shuts down if you cancel it.
                 await sleep_forever()
         finally:
@@ -174,28 +178,41 @@ class Router:
         logger.info(f"Unsubscribed from {topic}")
 
     async def _networking_recv(self):
-        while True:
-            topic, data = await self._net.gossipsub_recv()
-            logger.trace(f"Received message on {topic} with payload {data}")
-            if topic not in self.topic_routers:
-                logger.warning(f"Received message on unknown or inactive topic {topic}")
-                continue
-
-            router = self.topic_routers[topic]
-            await router.publish_bytes(data)
-
-    async def _networking_recv_connection_messages(self):
-        while True:
-            update = await self._net.connection_update_recv()
-            message = ConnectionMessage.from_update(update)
-            logger.trace(
-                f"Received message on connection_messages with payload {message}"
+        try:
+            while True:
+                from_swarm = await self._net.recv()
+                logger.debug(from_swarm)
+                match from_swarm:
+                    case PyFromSwarm.Message(origin, topic, data):
+                        logger.trace(
+                            f"Received message on {topic} from {origin} with payload {data}"
+                        )
+                        if topic not in self.topic_routers:
+                            logger.warning(
+                                f"Received message on unknown or inactive topic {topic}"
+                            )
+                            continue
+                        router = self.topic_routers[topic]
+                        await router.publish_bytes(data)
+                    case PyFromSwarm.Connection():
+                        message = ConnectionMessage.from_update(from_swarm)
+                        logger.trace(
+                            f"Received message on connection_messages with payload {message}"
+                        )
+                        if CONNECTION_MESSAGES.topic in self.topic_routers:
+                            router = self.topic_routers[CONNECTION_MESSAGES.topic]
+                            assert router.topic.model_type == ConnectionMessage
+                            router = cast(TopicRouter[ConnectionMessage], router)
+                            await router.publish(message)
+                    case _:
+                        logger.critical(
+                            "failed to exhaustively check FromSwarm messages - logic error"
+                        )
+        except Exception as exception:
+            logger.opt(exception=exception).error(
+                "Gossipsub receive loop terminated unexpectedly"
             )
-            if CONNECTION_MESSAGES.topic in self.topic_routers:
-                router = self.topic_routers[CONNECTION_MESSAGES.topic]
-                assert router.topic.model_type == ConnectionMessage
-                router = cast(TopicRouter[ConnectionMessage], router)
-                await router.publish(message)
+            raise
 
     async def _networking_publish(self):
         with self.networking_receiver as networked_items:

src/exo/shared/constants.py

@@ -78,4 +78,6 @@ EXO_ENABLE_IMAGE_MODELS = (
     os.getenv("EXO_ENABLE_IMAGE_MODELS", "false").lower() == "true"
 )
 
+EXO_OFFLINE = os.getenv("EXO_OFFLINE", "false").lower() == "true"
+
 EXO_TRACING_ENABLED = os.getenv("EXO_TRACING_ENABLED", "false").lower() == "true"

src/exo/shared/models/model_cards.py

@@ -90,6 +90,7 @@ class ModelCard(CamelCaseModel):
     base_model: str = ""
     capabilities: list[str] = []
     uses_cfg: bool = False
+    trust_remote_code: bool = True
 
     @field_validator("tasks", mode="before")
     @classmethod
@@ -137,6 +138,7 @@ class ModelCard(CamelCaseModel):
             hidden_size=config_data.hidden_size or 0,
             supports_tensor=config_data.supports_tensor,
             tasks=[ModelTask.TextGeneration],
+            trust_remote_code=False,
         )
         await mc.save_to_custom_dir()
         _card_cache[model_id] = mc

src/exo/shared/tests/test_election.py

@@ -1,7 +1,7 @@
 import pytest
 from anyio import create_task_group, fail_after, move_on_after
 
-from exo.routing.connection_message import ConnectionMessage, ConnectionMessageType
+from exo.routing.connection_message import ConnectionMessage
 from exo.shared.election import Election, ElectionMessage, ElectionResult
 from exo.shared.types.commands import ForwarderCommand, TestCommand
 from exo.shared.types.common import NodeId, SessionId, SystemId
@@ -327,14 +327,7 @@ async def test_connection_message_triggers_new_round_broadcast() -> None:
             tg.start_soon(election.run)
 
             # Send any connection message object; we close quickly to cancel before result creation
-            await cm_tx.send(
-                ConnectionMessage(
-                    node_id=NodeId(),
-                    connection_type=ConnectionMessageType.Connected,
-                    remote_ipv4="",
-                    remote_tcp_port=0,
-                )
-            )
+            await cm_tx.send(ConnectionMessage(node_id=NodeId(), connected=True))
 
             # Expect a broadcast for the new round at clock=1
             while True:

src/exo/shared/types/api.py

@@ -262,6 +262,11 @@ class DeleteInstanceResponse(BaseModel):
     instance_id: InstanceId
 
 
+class CancelCommandResponse(BaseModel):
+    message: str
+    command_id: CommandId
+
+
 ImageSize = Literal[
     "auto",
     "512x512",
@@ -437,3 +442,12 @@ class TraceListItem(CamelCaseModel):
 
 class TraceListResponse(CamelCaseModel):
     traces: list[TraceListItem]
+
+
+class DeleteTracesRequest(CamelCaseModel):
+    task_ids: list[str]
+
+
+class DeleteTracesResponse(CamelCaseModel):
+    deleted: list[str]
+    not_found: list[str]

src/exo/shared/types/mlx.py

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

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

@@ -0,0 +1,22 @@
+"""CLI arguments for runner processes passed as JSON."""
+
+from exo.shared.types.worker.instances import BoundInstance
+from exo.utils.pydantic_ext import FrozenModel
+
+
+class RunnerCliArgs(FrozenModel):
+    """
+    Arguments passed to the runner process via JSON on the command line.
+
+    File descriptors for communication are passed separately via pass_fds.
+    Each logical channel uses a single pipe:
+    - Events: child -> parent (child writes, parent reads)
+    - Tasks: parent -> child (parent writes, child reads)
+    - Cancels: parent -> child (parent writes, child reads)
+    """
+
+    bound_instance: BoundInstance
+    event_fd: int  # Child writes events here
+    task_fd: int  # Child reads tasks from here
+    cancel_fd: int  # Child reads cancels from here
+    log_level: str = "INFO"

src/exo/utils/fd_channels.py

@@ -0,0 +1,320 @@
+import contextlib
+import os
+import struct
+from dataclasses import dataclass
+from typing import Any, Self
+
+from anyio import (
+    BrokenResourceError,
+    ClosedResourceError,
+    EndOfStream,
+    WouldBlock,
+    to_thread,
+)
+from anyio.streams.memory import (
+    MemoryObjectReceiveStream as AnyioReceiver,
+)
+from anyio.streams.memory import (
+    MemoryObjectSendStream as AnyioSender,
+)
+from pydantic import TypeAdapter
+
+
+def _encode_message(data: bytes) -> bytes:
+    """Encode a message with its length prefix."""
+    return struct.pack("!I", len(data)) + data
+
+
+def _decode_message(data: bytes) -> tuple[bytes, bytes]:
+    """Decode a message from a buffer, returning (message, remaining_buffer)."""
+    if len(data) < 4:
+        return b"", data
+    length = struct.unpack("!I", data[:4])[0]
+    if len(data) < 4 + length:
+        return b"", data
+    return data[4 : 4 + length], data[4 + length :]
+
+
+class FdSender[T]:
+    """A sender that writes JSON-encoded Pydantic models to a file descriptor."""
+
+    def __init__(
+        self,
+        fd: int,
+        type_adapter: TypeAdapter[T],
+        max_buffer_size: int = 1024 * 1024,
+    ) -> None:
+        self._fd = fd
+        self._type_adapter = type_adapter
+        self._max_buffer_size = max_buffer_size
+        self._closed = False
+        self._write_buffer = bytearray()
+
+    def send_nowait(self, item: T) -> None:
+        """Send an item without blocking."""
+        if self._closed:
+            raise ClosedResourceError
+
+        data = self._type_adapter.dump_json(item)
+        message = _encode_message(data)
+
+        if len(self._write_buffer) + len(message) > self._max_buffer_size:
+            raise WouldBlock
+
+        self._write_buffer.extend(message)
+        self._flush()
+
+    def send(self, item: T) -> None:
+        """Send an item, blocking if necessary."""
+        if self._closed:
+            raise ClosedResourceError
+
+        data = self._type_adapter.dump_json(item)
+        message = _encode_message(data)
+        self._write_buffer.extend(message)
+        self._flush_blocking()
+
+    async def send_async(self, item: T) -> None:
+        """Send an item asynchronously."""
+        await to_thread.run_sync(self.send, item)
+
+    def _flush(self) -> None:
+        """Flush the write buffer non-blocking."""
+        if not self._write_buffer:
+            return
+
+        try:
+            written = os.write(self._fd, self._write_buffer)
+            self._write_buffer = self._write_buffer[written:]
+        except (OSError, BrokenPipeError) as e:
+            raise BrokenResourceError from e
+
+    def _flush_blocking(self) -> None:
+        """Flush the write buffer, blocking until complete."""
+        while self._write_buffer:
+            self._flush()
+
+    def close(self) -> None:
+        """Close the sender."""
+        if not self._closed:
+            self._closed = True
+            try:
+                self._flush_blocking()
+                os.close(self._fd)
+            except (OSError, BrokenPipeError):
+                pass
+
+    def __enter__(self) -> Self:
+        return self
+
+    def __exit__(self, *args: Any) -> None:
+        self.close()
+
+
+class FdReceiver[T]:
+    """A receiver that reads JSON-encoded Pydantic models from a file descriptor."""
+
+    def __init__(
+        self,
+        fd: int,
+        type_adapter: TypeAdapter[T],
+        max_buffer_size: int = 1024 * 1024,
+    ) -> None:
+        self._fd = fd
+        self._type_adapter = type_adapter
+        self._max_buffer_size = max_buffer_size
+        self._closed = False
+        self._read_buffer = bytearray()
+
+    def receive_nowait(self) -> T:
+        """Receive an item without blocking."""
+        if self._closed:
+            raise ClosedResourceError
+
+        # Try to read any available data
+        self._read_available()
+
+        # Try to decode a message
+        message, remaining = _decode_message(bytes(self._read_buffer))
+        if message:
+            self._read_buffer = bytearray(remaining)
+            return self._type_adapter.validate_json(message)
+
+        raise WouldBlock
+
+    def receive(self) -> T:
+        """Receive an item, blocking if necessary."""
+        if self._closed:
+            raise ClosedResourceError
+
+        while True:
+            # Try to decode a message
+            message, remaining = _decode_message(bytes(self._read_buffer))
+            if message:
+                self._read_buffer = bytearray(remaining)
+                return self._type_adapter.validate_json(message)
+
+            # Read more data
+            try:
+                chunk = os.read(self._fd, 8192)
+                if not chunk:
+                    # EOF
+                    self.close()
+                    raise EndOfStream
+                self._read_buffer.extend(chunk)
+            except (OSError, BrokenPipeError) as e:
+                self.close()
+                raise BrokenResourceError from e
+
+    async def receive_async(self) -> T:
+        """Receive an item asynchronously."""
+        return await to_thread.run_sync(self.receive)
+
+    def _read_available(self) -> None:
+        """Read any available data without blocking."""
+        try:
+            import fcntl
+
+            # Set non-blocking mode temporarily
+            flags = fcntl.fcntl(self._fd, fcntl.F_GETFL)
+            fcntl.fcntl(self._fd, fcntl.F_SETFL, flags | os.O_NONBLOCK)
+            try:
+                chunk = os.read(self._fd, 8192)
+                if chunk:
+                    self._read_buffer.extend(chunk)
+            except BlockingIOError:
+                pass
+            finally:
+                fcntl.fcntl(self._fd, fcntl.F_SETFL, flags)
+        except ImportError:
+            # Windows doesn't have fcntl, just try reading
+            pass
+
+    def close(self) -> None:
+        """Close the receiver."""
+        if not self._closed:
+            self._closed = True
+            with contextlib.suppress(OSError, BrokenPipeError):
+                os.close(self._fd)
+
+    def __iter__(self) -> Self:
+        return self
+
+    def __next__(self) -> T:
+        try:
+            return self.receive()
+        except EndOfStream:
+            raise StopIteration from None
+
+    def __aiter__(self) -> Self:
+        return self
+
+    async def __anext__(self) -> T:
+        try:
+            return await self.receive_async()
+        except EndOfStream:
+            raise StopAsyncIteration from None
+
+    def __enter__(self) -> Self:
+        return self
+
+    def __exit__(self, *args: Any) -> None:
+        self.close()
+
+    def collect(self) -> list[T]:
+        """Collect all currently available items from this receiver."""
+        out: list[T] = []
+        while True:
+            try:
+                item = self.receive_nowait()
+                out.append(item)
+            except WouldBlock:
+                break
+        return out
+
+
+@dataclass(eq=False)
+class FdChannel:
+    """A pair of file descriptors for bidirectional communication."""
+
+    sender_fd: int
+    receiver_fd: int
+
+    def close(self) -> None:
+        """Close both file descriptors."""
+        for fd in (self.sender_fd, self.receiver_fd):
+            with contextlib.suppress(OSError, BrokenPipeError):
+                os.close(fd)
+
+
+def fd_channel_pair() -> tuple[FdChannel, FdChannel]:
+    """
+    Create a pair of connected file descriptor channels.
+    Returns (parent_channel, child_channel).
+    """
+    # Create two pipes: one for parent->child, one for child->parent
+    parent_to_child_read, parent_to_child_write = os.pipe()
+    child_to_parent_read, child_to_parent_write = os.pipe()
+
+    # Parent writes to parent_to_child_write, reads from child_to_parent_read
+    parent_channel = FdChannel(
+        sender_fd=parent_to_child_write,
+        receiver_fd=child_to_parent_read,
+    )
+
+    # Child reads from parent_to_child_read, writes to child_to_parent_write
+    child_channel = FdChannel(
+        sender_fd=child_to_parent_write,
+        receiver_fd=parent_to_child_read,
+    )
+
+    return parent_channel, child_channel
+
+
+class Sender[T](AnyioSender[T]):
+    def clone(self) -> "Sender[T]":
+        if self._closed:
+            raise ClosedResourceError
+        return Sender(_state=self._state)
+
+    def clone_receiver(self) -> "Receiver[T]":
+        """Constructs a Receiver using a Senders shared state - similar to calling Receiver.clone() without needing the receiver"""
+        if self._closed:
+            raise ClosedResourceError
+        return Receiver(_state=self._state)
+
+
+class Receiver[T](AnyioReceiver[T]):
+    def clone(self) -> "Receiver[T]":
+        if self._closed:
+            raise ClosedResourceError
+        return Receiver(_state=self._state)
+
+    def clone_sender(self) -> Sender[T]:
+        """Constructs a Sender using a Receivers shared state - similar to calling Sender.clone() without needing the receiver"""
+        if self._closed:
+            raise ClosedResourceError
+        return Sender(_state=self._state)
+
+    def collect(self) -> list[T]:
+        """Collect all currently available items from this receiver"""
+        out: list[T] = []
+        while True:
+            try:
+                item = self.receive_nowait()
+                out.append(item)
+            except WouldBlock:
+                break
+        return out
+
+    async def receive_at_least(self, n: int) -> list[T]:
+        out: list[T] = []
+        out.append(await self.receive())
+        out.extend(self.collect())
+        while len(out) < n:
+            out.append(await self.receive())
+            out.extend(self.collect())
+        return out
+
+    def __enter__(self) -> Self:
+        return self

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

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

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

@@ -49,6 +49,21 @@ TimeoutCallback = Callable[[], None]
 LayerLoadedCallback = Callable[[int, int], None]  # (layers_loaded, total_layers)
 
 
+_pending_prefill_sends: list[tuple[mx.array, int, mx.distributed.Group]] = []
+
+
+def flush_prefill_sends() -> None:
+    for output, dst, group in _pending_prefill_sends:
+        sent = mx.distributed.send(output, dst, group=group)
+        mx.async_eval(sent)
+    _pending_prefill_sends.clear()
+
+
+def clear_prefill_sends() -> None:
+    # Discard pending sends (e.g. on cancellation).
+    _pending_prefill_sends.clear()
+
+
 def eval_with_timeout(
     mlx_item: Any,  # pyright: ignore[reportAny]
     timeout_seconds: float = 60.0,
@@ -128,11 +143,11 @@ class PipelineFirstLayer(CustomMlxLayer):
 
     def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
         if self.r != 0:
+            # We want to avoid GPU timeout errors by evalling the distributed operation
+            # so that it stays on CPU, which does not have a timeout.
+            mx.eval(x)
             x = mx.distributed.recv_like(x, (self.r - 1), group=self.group)
-            if self.is_prefill:
-                # We want to avoid GPU timeout errors by evalling the distributed operation
-                # so that it stays on CPU, which does not have a timeout.
-                mx.eval(x)
+            mx.eval(x)
         return self.original_layer(x, *args, **kwargs)
 
 
@@ -150,6 +165,7 @@ class PipelineLastLayer(CustomMlxLayer):
         self.group = group
         self.original_layer_signature = signature(self.original_layer.__call__)
         self.is_prefill: bool = False
+        self.queue_sends: bool = False
 
     def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
         cache = self.original_layer_signature.bind_partial(
@@ -158,24 +174,33 @@ class PipelineLastLayer(CustomMlxLayer):
 
         output: mx.array = self.original_layer(x, *args, **kwargs)
 
+        # Eval layer output to materialize it before send — this splits the graph
+        # so the send is isolated and the receiving rank's recv can complete.
+        mx.eval(output)
+
         if self.r != self.s - 1:
-            output = mx.distributed.send(
-                output, (self.r + 1) % self.s, group=self.group
-            )
+            if self.queue_sends:
+                _pending_prefill_sends.append(
+                    (output, (self.r + 1) % self.s, self.group)
+                )
+            else:
+                output = mx.distributed.send(
+                    output, (self.r + 1) % self.s, group=self.group
+                )
             if cache is not None:
                 # CacheList (used by MLA models like DeepSeekV32, GLM MoE DSA)
                 # doesn't have .keys directly; access via first sub-cache.
                 _cache = cache[0] if hasattr(cache, "caches") else cache  # type: ignore
                 _cache.keys = mx.depends(_cache.keys, output)  # type: ignore
-            if self.is_prefill:
-                mx.eval(output)
-                if cache is not None:
-                    mx.eval(_cache.keys)  # type: ignore
+            mx.eval(output)
+            if cache is not None:
+                mx.eval(_cache.keys)  # type: ignore
 
         if not self.is_prefill:
             output = mx.distributed.all_gather(output, group=self.group)[
                 -output.shape[0] :
             ]
+            mx.eval(output)
 
         return output
 
@@ -186,6 +211,12 @@ def set_pipeline_prefill(model: nn.Module, is_prefill: bool) -> None:
             layer.is_prefill = is_prefill
 
 
+def set_pipeline_queue_sends(model: nn.Module, queue_sends: bool) -> None:
+    for layer in model.layers:  # type: ignore
+        if isinstance(layer, PipelineLastLayer):
+            layer.queue_sends = queue_sends
+
+
 def get_inner_model(model: nn.Module) -> nn.Module:
     inner = getattr(model, "model", None)
     if isinstance(inner, nn.Module):
@@ -852,6 +883,8 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
                 layer.self_attn.o_proj = self.sharded_to_all_linear(
                     layer.self_attn.o_proj
                 )
+                layer.self_attn.n_heads //= self.N
+                layer.self_attn.n_kv_heads //= self.N
             else:
                 assert isinstance(layer, Qwen3NextDecoderLayer)
                 if hasattr(layer, "linear_attn"):

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

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

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

@@ -1,9 +1,14 @@
+import functools
+import math
 import time
 from copy import deepcopy
 from typing import Callable, Generator, cast, get_args
 
 import mlx.core as mx
-from mlx_lm.generate import stream_generate
+from mlx_lm.generate import (
+    maybe_quantize_kv_cache,
+    stream_generate,
+)
 from mlx_lm.models.cache import ArraysCache, RotatingKVCache
 from mlx_lm.sample_utils import make_sampler
 from mlx_lm.tokenizer_utils import TokenizerWrapper
@@ -18,13 +23,19 @@ from exo.shared.types.api import (
 )
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
-from exo.shared.types.mlx import KVCacheType
+from exo.shared.types.mlx import KVCacheType, Model
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.shared.types.worker.runner_response import (
     GenerationResponse,
 )
-from exo.worker.engines.mlx import Model
-from exo.worker.engines.mlx.auto_parallel import set_pipeline_prefill
+from exo.worker.engines.mlx.auto_parallel import (
+    PipelineFirstLayer,
+    PipelineLastLayer,
+    clear_prefill_sends,
+    flush_prefill_sends,
+    set_pipeline_prefill,
+    set_pipeline_queue_sends,
+)
 from exo.worker.engines.mlx.cache import (
     CacheSnapshot,
     KVPrefixCache,
@@ -55,6 +66,130 @@ class PrefillCancelled(BaseException):
     """Raised when prefill is cancelled via the progress callback."""
 
 
+def _has_pipeline_communication_layer(model: Model):
+    for layer in model.layers:
+        if isinstance(layer, (PipelineFirstLayer, PipelineLastLayer)):
+            return True
+    return False
+
+
+def pipeline_parallel_prefill(
+    model: Model,
+    prompt: mx.array,
+    prompt_cache: KVCacheType,
+    prefill_step_size: int,
+    kv_group_size: int | None,
+    kv_bits: int | None,
+    prompt_progress_callback: Callable[[int, int], None],
+    distributed_prompt_progress_callback: Callable[[], None] | None,
+    group: mx.distributed.Group,
+) -> None:
+    """Prefill the KV cache for pipeline parallel with overlapping stages.
+
+    Each rank processes the full prompt through its real cache, offset by leading
+    and trailing dummy iterations.
+
+    Total iterations per rank = N_real_chunks + world_size - 1:
+      - rank r leading dummies  (skip_pipeline_io, throwaway cache)
+      - N_real_chunks real      (pipeline IO active, real cache)
+      - (world_size-1-r) trailing dummies (skip_pipeline_io, throwaway cache)
+
+    e.g.
+    Timeline (2 ranks, 3 chunks of 10240 tokens @ step=4096):
+        iter 0: R0 real[0:4096]     R1 dummy
+        iter 1: R0 real[4096:8192]  R1 real[0:4096]
+        iter 2: R0 real[8192:10240] R1 real[4096:8192]
+        iter 3: R0 dummy            R1 real[8192:10240]
+
+    This function is designed to match mlx_lm's stream_generate exactly in terms of
+    side effects (given the same prefill step size)
+    """
+    prefill_step_size = prefill_step_size // min(4, group.size())
+
+    quantize_cache_fn: Callable[..., None] = functools.partial(
+        maybe_quantize_kv_cache,
+        quantized_kv_start=0,
+        kv_group_size=kv_group_size,
+        kv_bits=kv_bits,
+    )
+
+    _prompt_cache: KVCacheType = prompt_cache
+    rank = group.rank()
+    world_size = group.size()
+
+    # Build list of real prompt chunk sizes
+    total = len(prompt)
+    real_chunk_sizes: list[int] = []
+    remaining = total - 1
+    while remaining:
+        n = min(prefill_step_size, remaining)
+        real_chunk_sizes.append(n)
+        remaining -= n
+    n_real = len(real_chunk_sizes)
+
+    # Each rank does: [rank leading dummies] [N real chunks] [world_size-1-rank trailing dummies]
+    n_leading = rank
+    n_trailing = world_size - 1 - rank
+    n_total = n_leading + n_real + n_trailing
+
+    t_start = time.perf_counter()
+    processed = 0
+    logger.info(
+        f"[R{rank}] Pipeline prefill: {n_real} real + {n_leading} leading + {n_trailing} trailing = {n_total} iterations"
+    )
+    clear_prefill_sends()
+
+    # Initial callback matching generate_step
+    prompt_progress_callback(0, total)
+
+    try:
+        with mx.stream(generation_stream):
+            for _ in range(n_leading):
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+            for i in range(n_real):
+                chunk_size = real_chunk_sizes[i]
+                model(
+                    prompt[processed : processed + chunk_size][None],
+                    cache=_prompt_cache,
+                )
+                quantize_cache_fn(_prompt_cache)
+                processed += chunk_size
+
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+                flush_prefill_sends()
+
+                prompt_progress_callback(processed, total)
+
+            for _ in range(n_trailing):
+                if distributed_prompt_progress_callback is not None:
+                    distributed_prompt_progress_callback()
+
+    finally:
+        clear_prefill_sends()
+
+    # Post-loop: process remaining 1 token + add +1 entry to match stream_generate.
+    for _ in range(2):
+        with mx.stream(generation_stream):
+            model(prompt[-1:][None], cache=_prompt_cache)
+            quantize_cache_fn(_prompt_cache)
+        flush_prefill_sends()
+
+    assert _prompt_cache is not None
+    mx.eval([c.state for c in _prompt_cache])  # type: ignore
+
+    # Final callback matching generate_step
+    prompt_progress_callback(total, total)
+
+    logger.info(
+        f"[R{rank}] Prefill: {n_real} real + {n_leading}+{n_trailing} dummy iterations, "
+        f"Processed {processed} tokens in {(time.perf_counter() - t_start) * 1000:.1f}ms"
+    )
+
+
 def prefill(
     model: Model,
     tokenizer: TokenizerWrapper,
@@ -63,6 +198,7 @@ def prefill(
     cache: KVCacheType,
     group: mx.distributed.Group | None,
     on_prefill_progress: Callable[[int, int], None] | None,
+    distributed_prompt_progress_callback: Callable[[], None] | None,
 ) -> tuple[float, int, list[CacheSnapshot]]:
     """Prefill the KV cache with prompt tokens.
 
@@ -94,31 +230,57 @@ def prefill(
         if on_prefill_progress is not None:
             on_prefill_progress(processed, total)
 
+    def combined_progress_callback(processed: int, total: int) -> None:
+        if distributed_prompt_progress_callback is not None:
+            distributed_prompt_progress_callback()
+        progress_callback(processed, total)
+
     set_pipeline_prefill(model, is_prefill=True)
 
     mx_barrier(group)
     logger.info("Starting prefill")
 
-    # Use max_tokens=1 because max_tokens=0 does not work.
-    # We just throw away the generated token - we only care about filling the cache
+    is_pipeline = _has_pipeline_communication_layer(model)
+
+    prefill_step_size = 4096
+
     try:
-        for _ in stream_generate(
-            model=model,
-            tokenizer=tokenizer,
-            prompt=prompt_tokens,
-            max_tokens=1,
-            sampler=sampler,
-            prompt_cache=cache,
-            prefill_step_size=4096,
-            kv_group_size=KV_GROUP_SIZE,
-            kv_bits=KV_BITS,
-            prompt_progress_callback=progress_callback,
-        ):
-            break  # Stop after first iteration - cache is now filled
+        if is_pipeline and num_tokens >= prefill_step_size:
+            set_pipeline_queue_sends(model, queue_sends=True)
+            assert group is not None, "Pipeline prefill requires a distributed group"
+            pipeline_parallel_prefill(
+                model=model,
+                prompt=prompt_tokens,
+                prompt_cache=cache,
+                prefill_step_size=prefill_step_size,
+                kv_group_size=KV_GROUP_SIZE,
+                kv_bits=KV_BITS,
+                prompt_progress_callback=progress_callback,
+                distributed_prompt_progress_callback=distributed_prompt_progress_callback,
+                group=group,
+            )
+        else:
+            # Use max_tokens=1 because max_tokens=0 does not work.
+            # We just throw away the generated token - we only care about filling the cache
+            for _ in stream_generate(
+                model=model,
+                tokenizer=tokenizer,
+                prompt=prompt_tokens,
+                max_tokens=1,
+                sampler=sampler,
+                prompt_cache=cache,
+                prefill_step_size=prefill_step_size,
+                kv_group_size=KV_GROUP_SIZE,
+                kv_bits=KV_BITS,
+                prompt_progress_callback=combined_progress_callback,
+            ):
+                break  # Stop after first iteration - cache is now filled
     except PrefillCancelled:
+        set_pipeline_queue_sends(model, queue_sends=False)
         set_pipeline_prefill(model, is_prefill=False)
         raise
 
+    set_pipeline_queue_sends(model, queue_sends=False)
     set_pipeline_prefill(model, is_prefill=False)
 
     # stream_generate added 1 extra generated token to the cache, so we should trim it.
@@ -131,7 +293,7 @@ def prefill(
                 cache[i] = deepcopy(pre_gen.states[i])  # type: ignore
         else:
             assert not isinstance(c, (ArraysCache, RotatingKVCache))
-            c.trim(2)  # pyright: ignore[reportUnknownMemberType]
+            c.trim(2)
 
     elapsed = time.perf_counter() - start_time
     tokens_per_sec = num_tokens / elapsed if elapsed > 0 else 0.0
@@ -248,6 +410,9 @@ def extract_top_logprobs(
     for i in range(top_logprobs):
         token_id = int(top_indices[i].item())
         token_logprob = float(top_values[i].item())
+        if math.isnan(token_logprob):
+            continue
+
         # Decode token ID to string
         token_str = tokenizer.decode([token_id])
         # Get byte representation
@@ -271,6 +436,7 @@ def mlx_generate(
     kv_prefix_cache: KVPrefixCache | None,
     group: mx.distributed.Group | None,
     on_prefill_progress: Callable[[int, int], None] | None = None,
+    distributed_prompt_progress_callback: Callable[[], None] | None = None,
 ) -> Generator[GenerationResponse]:
     # Ensure that generation stats only contains peak memory for this generation
     mx.reset_peak_memory()
@@ -332,6 +498,7 @@ def mlx_generate(
         caches,
         group,
         on_prefill_progress,
+        distributed_prompt_progress_callback,
     )
     cache_snapshots: list[CacheSnapshot] | None = ssm_snapshots_list or None
 

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

@@ -2,6 +2,7 @@ import json
 import os
 import re
 import sys
+import tempfile
 import time
 from pathlib import Path
 from typing import Any, cast
@@ -23,9 +24,7 @@ from mlx_lm.models.deepseek_v3 import DeepseekV3Model
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.shared.models.model_cards import ModelId
-from exo.worker.engines.mlx.constants import (
-    TRUST_REMOTE_CODE,
-)
+from exo.worker.engines.mlx.constants import TRUST_REMOTE_CODE
 
 try:
     from mlx_lm.tokenizer_utils import load_tokenizer
@@ -41,6 +40,7 @@ from pydantic import RootModel
 from exo.download.download_utils import build_model_path
 from exo.shared.types.common import Host
 from exo.shared.types.memory import Memory
+from exo.shared.types.mlx import Model
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import (
     BoundInstance,
@@ -53,7 +53,6 @@ from exo.shared.types.worker.shards import (
     ShardMetadata,
     TensorShardMetadata,
 )
-from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.auto_parallel import (
     LayerLoadedCallback,
     TimeoutCallback,
@@ -100,14 +99,13 @@ def mlx_distributed_init(
     rank = bound_instance.bound_shard.device_rank
     logger.info(f"Starting initialization for rank {rank}")
 
-    coordination_file = None
-    try:
+    with tempfile.TemporaryDirectory() as tmpdir:
+        coordination_file = str(
+            Path(tmpdir) / f"hosts_{bound_instance.instance.instance_id}_{rank}.json"
+        )
         # TODO: singleton instances
         match bound_instance.instance:
             case MlxRingInstance(hosts_by_node=hosts_by_node, ephemeral_port=_):
-                coordination_file = (
-                    f"./hosts_{bound_instance.instance.instance_id}_{rank}.json"
-                )
                 hosts_for_node = hosts_by_node[bound_instance.bound_node_id]
                 hosts_json = HostList.from_hosts(hosts_for_node).model_dump_json()
 
@@ -130,9 +128,6 @@ def mlx_distributed_init(
                     jaccl_devices[i][i] is None for i in range(len(jaccl_devices))
                 )
                 # Use RDMA connectivity matrix
-                coordination_file = (
-                    f"./hosts_{bound_instance.instance.instance_id}_{rank}.json"
-                )
                 jaccl_devices_json = json.dumps(jaccl_devices)
 
                 with open(coordination_file, "w") as f:
@@ -152,10 +147,6 @@ def mlx_distributed_init(
         logger.info(f"Rank {rank} mlx distributed initialization complete")
 
         return group
-    finally:
-        with contextlib.suppress(FileNotFoundError):
-            if coordination_file:
-                os.remove(coordination_file)
 
 
 def initialize_mlx(
@@ -216,6 +207,8 @@ def load_mlx_items(
 
     set_wired_limit_for_model(get_weights_size(bound_instance.bound_shard))
 
+    mx.clear_cache()
+
     return cast(Model, model), tokenizer
 
 
@@ -293,7 +286,11 @@ def shard_and_load(
 
 def get_tokenizer(model_path: Path, shard_metadata: ShardMetadata) -> TokenizerWrapper:
     """Load tokenizer for a model shard. Delegates to load_tokenizer_for_model_id."""
-    return load_tokenizer_for_model_id(shard_metadata.model_card.model_id, model_path)
+    return load_tokenizer_for_model_id(
+        shard_metadata.model_card.model_id,
+        model_path,
+        trust_remote_code=shard_metadata.model_card.trust_remote_code,
+    )
 
 
 def get_eos_token_ids_for_model(model_id: ModelId) -> list[int] | None:
@@ -325,7 +322,7 @@ def get_eos_token_ids_for_model(model_id: ModelId) -> list[int] | None:
 
 
 def load_tokenizer_for_model_id(
-    model_id: ModelId, model_path: Path
+    model_id: ModelId, model_path: Path, *, trust_remote_code: bool = TRUST_REMOTE_CODE
 ) -> TokenizerWrapper:
     """
     Load tokenizer for a model given its ID and local path.
@@ -394,7 +391,7 @@ def load_tokenizer_for_model_id(
 
     tokenizer = load_tokenizer(
         model_path,
-        tokenizer_config_extra={"trust_remote_code": TRUST_REMOTE_CODE},
+        tokenizer_config_extra={"trust_remote_code": trust_remote_code},
         eos_token_ids=eos_token_ids,
     )
 

src/exo/worker/main.py

@@ -1,9 +1,8 @@
 from collections import defaultdict
 from datetime import datetime, timezone
-from random import random
 
 import anyio
-from anyio import CancelScope, fail_after
+from anyio import fail_after
 from loguru import logger
 
 from exo.download.download_utils import resolve_model_in_path
@@ -13,17 +12,13 @@ from exo.shared.types.api import ImageEditsTaskParams
 from exo.shared.types.commands import (
     ForwarderCommand,
     ForwarderDownloadCommand,
-    RequestEventLog,
     StartDownload,
 )
-from exo.shared.types.common import CommandId, NodeId, SessionId, SystemId
+from exo.shared.types.common import CommandId, NodeId, SystemId
 from exo.shared.types.events import (
     Event,
-    EventId,
-    GlobalForwarderEvent,
     IndexedEvent,
     InputChunkReceived,
-    LocalForwarderEvent,
     NodeDownloadProgress,
     NodeGatheredInfo,
     TaskCreated,
@@ -46,7 +41,6 @@ from exo.shared.types.topology import Connection, SocketConnection
 from exo.shared.types.worker.downloads import DownloadCompleted
 from exo.shared.types.worker.runners import RunnerId
 from exo.utils.channels import Receiver, Sender, channel
-from exo.utils.event_buffer import OrderedBuffer
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo, InfoGatherer
 from exo.utils.info_gatherer.net_profile import check_reachable
 from exo.utils.keyed_backoff import KeyedBackoff
@@ -59,38 +53,26 @@ class Worker:
     def __init__(
         self,
         node_id: NodeId,
-        session_id: SessionId,
         *,
-        global_event_receiver: Receiver[GlobalForwarderEvent],
-        local_event_sender: Sender[LocalForwarderEvent],
+        event_receiver: Receiver[IndexedEvent],
+        event_sender: Sender[Event],
         # This is for requesting updates. It doesn't need to be a general command sender right now,
         # but I think it's the correct way to be thinking about commands
         command_sender: Sender[ForwarderCommand],
         download_command_sender: Sender[ForwarderDownloadCommand],
     ):
         self.node_id: NodeId = node_id
-        self.session_id: SessionId = session_id
-
-        self.global_event_receiver = global_event_receiver
-        self.local_event_sender = local_event_sender
+        self.event_receiver = event_receiver
+        self.event_sender = event_sender
         self.command_sender = command_sender
         self.download_command_sender = download_command_sender
-        self.event_buffer = OrderedBuffer[Event]()
-        self.out_for_delivery: dict[EventId, LocalForwarderEvent] = {}
 
         self.state: State = State()
         self.runners: dict[RunnerId, RunnerSupervisor] = {}
         self._tg: TaskGroup = TaskGroup()
 
-        self._nack_cancel_scope: CancelScope | None = None
-        self._nack_attempts: int = 0
-        self._nack_base_seconds: float = 0.5
-        self._nack_cap_seconds: float = 10.0
-
         self._system_id = SystemId()
 
-        self.event_sender, self.event_receiver = channel[Event]()
-
         # Buffer for input image chunks (for image editing)
         self.input_chunk_buffer: dict[CommandId, dict[int, str]] = {}
         self.input_chunk_counts: dict[CommandId, int] = {}
@@ -108,18 +90,16 @@ class Worker:
                 tg.start_soon(info_gatherer.run)
                 tg.start_soon(self._forward_info, info_recv)
                 tg.start_soon(self.plan_step)
-                tg.start_soon(self._resend_out_for_delivery)
                 tg.start_soon(self._event_applier)
-                tg.start_soon(self._forward_events)
                 tg.start_soon(self._poll_connection_updates)
         finally:
             # Actual shutdown code - waits for all tasks to complete before executing.
             logger.info("Stopping Worker")
-            self.local_event_sender.close()
+            self.event_sender.close()
             self.command_sender.close()
             self.download_command_sender.close()
             for runner in self.runners.values():
-                runner.shutdown()
+                await runner.shutdown()
 
     async def _forward_info(self, recv: Receiver[GatheredInfo]):
         with recv as info_stream:
@@ -133,47 +113,22 @@ class Worker:
                 )
 
     async def _event_applier(self):
-        with self.global_event_receiver as events:
-            async for f_event in events:
-                if f_event.session != self.session_id:
-                    continue
-                if f_event.origin != self.session_id.master_node_id:
-                    continue
-                self.event_buffer.ingest(f_event.origin_idx, f_event.event)
-                event_id = f_event.event.event_id
-                if event_id in self.out_for_delivery:
-                    del self.out_for_delivery[event_id]
-
+        with self.event_receiver as events:
+            async for event in events:
                 # 2. for each event, apply it to the state
-                indexed_events = self.event_buffer.drain_indexed()
-                if indexed_events:
-                    self._nack_attempts = 0
+                self.state = apply(self.state, event=event)
+                event = event.event
 
-                if not indexed_events and (
-                    self._nack_cancel_scope is None
-                    or self._nack_cancel_scope.cancel_called
-                ):
-                    # Request the next index.
-                    self._tg.start_soon(
-                        self._nack_request, self.state.last_event_applied_idx + 1
+                # Buffer input image chunks for image editing
+                if isinstance(event, InputChunkReceived):
+                    cmd_id = event.command_id
+                    if cmd_id not in self.input_chunk_buffer:
+                        self.input_chunk_buffer[cmd_id] = {}
+                        self.input_chunk_counts[cmd_id] = event.chunk.total_chunks
+
+                    self.input_chunk_buffer[cmd_id][event.chunk.chunk_index] = (
+                        event.chunk.data
                     )
-                    continue
-                elif indexed_events and self._nack_cancel_scope:
-                    self._nack_cancel_scope.cancel()
-
-                for idx, event in indexed_events:
-                    self.state = apply(self.state, IndexedEvent(idx=idx, event=event))
-
-                    # Buffer input image chunks for image editing
-                    if isinstance(event, InputChunkReceived):
-                        cmd_id = event.command_id
-                        if cmd_id not in self.input_chunk_buffer:
-                            self.input_chunk_buffer[cmd_id] = {}
-                            self.input_chunk_counts[cmd_id] = event.chunk.total_chunks
-
-                        self.input_chunk_buffer[cmd_id][event.chunk.chunk_index] = (
-                            event.chunk.data
-                        )
 
     async def plan_step(self):
         while True:
@@ -205,7 +160,7 @@ class Worker:
             # lets not kill the worker if a runner is unresponsive
             match task:
                 case CreateRunner():
-                    self._create_supervisor(task)
+                    await self._create_supervisor(task)
                     await self.event_sender.send(
                         TaskStatusUpdated(
                             task_id=task.task_id, task_status=TaskStatus.Complete
@@ -265,7 +220,7 @@ class Worker:
                             )
                         )
                     finally:
-                        runner.shutdown()
+                        await runner.shutdown()
                 case CancelTask(
                     cancelled_task_id=cancelled_task_id, runner_id=runner_id
                 ):
@@ -325,46 +280,9 @@ class Worker:
                 instance.shard_assignments.node_to_runner[self.node_id]
             ].start_task(task)
 
-    async def _nack_request(self, since_idx: int) -> None:
-        # We request all events after (and including) the missing index.
-        # This function is started whenever we receive an event that is out of sequence.
-        # It is cancelled as soon as we receiver an event that is in sequence.
-
-        if since_idx < 0:
-            logger.warning(f"Negative value encountered for nack request {since_idx=}")
-            since_idx = 0
-
-        with CancelScope() as scope:
-            self._nack_cancel_scope = scope
-            delay: float = self._nack_base_seconds * (2.0**self._nack_attempts)
-            delay = min(self._nack_cap_seconds, delay)
-            self._nack_attempts += 1
-            try:
-                await anyio.sleep(delay)
-                logger.info(
-                    f"Nack attempt {self._nack_attempts}: Requesting Event Log from {since_idx}"
-                )
-                await self.command_sender.send(
-                    ForwarderCommand(
-                        origin=self._system_id,
-                        command=RequestEventLog(since_idx=since_idx),
-                    )
-                )
-            finally:
-                if self._nack_cancel_scope is scope:
-                    self._nack_cancel_scope = None
-
-    async def _resend_out_for_delivery(self) -> None:
-        # This can also be massively tightened, we should check events are at least a certain age before resending.
-        # Exponential backoff would also certainly help here.
-        while True:
-            await anyio.sleep(1 + random())
-            for event in self.out_for_delivery.copy().values():
-                await self.local_event_sender.send(event)
-
-    def _create_supervisor(self, task: CreateRunner) -> RunnerSupervisor:
+    async def _create_supervisor(self, task: CreateRunner) -> RunnerSupervisor:
         """Creates and stores a new AssignedRunner with initial downloading status."""
-        runner = RunnerSupervisor.create(
+        runner = await RunnerSupervisor.create(
             bound_instance=task.bound_instance,
             event_sender=self.event_sender.clone(),
         )
@@ -372,21 +290,6 @@ class Worker:
         self._tg.start_soon(runner.run)
         return runner
 
-    async def _forward_events(self) -> None:
-        idx = 0
-        with self.event_receiver as events:
-            async for event in events:
-                fe = LocalForwarderEvent(
-                    origin_idx=idx,
-                    origin=self._system_id,
-                    session=self.session_id,
-                    event=event,
-                )
-                idx += 1
-                logger.debug(f"Worker published event {idx}: {str(event)[:100]}")
-                await self.local_event_sender.send(fe)
-                self.out_for_delivery[event.event_id] = fe
-
     async def _poll_connection_updates(self):
         while True:
             edges = set(

src/exo/worker/runner/bootstrap.py

@@ -1,26 +1,44 @@
+import contextlib
 import os
+import sys
 
 import loguru
+from pydantic import TypeAdapter
 
 from exo.shared.types.events import Event, RunnerStatusUpdated
 from exo.shared.types.tasks import Task, TaskId
-from exo.shared.types.worker.instances import BoundInstance
+from exo.shared.types.worker.runner_args import RunnerCliArgs
 from exo.shared.types.worker.runners import RunnerFailed
-from exo.utils.channels import ClosedResourceError, MpReceiver, MpSender
+from exo.utils.fd_channels import FdReceiver, FdSender
 
 logger: "loguru.Logger" = loguru.logger
 
+# Type adapters for union types
+event_adapter: TypeAdapter[Event] = TypeAdapter(Event)
+task_adapter: TypeAdapter[Task] = TypeAdapter(Task)
+task_id_adapter: TypeAdapter[TaskId] = TypeAdapter(TaskId)
 
-def entrypoint(
-    bound_instance: BoundInstance,
-    event_sender: MpSender[Event],
-    task_receiver: MpReceiver[Task],
-    cancel_receiver: MpReceiver[TaskId],
-    _logger: "loguru.Logger",
-) -> None:
-    global logger
-    logger = _logger
 
+def entrypoint() -> None:
+    """
+    Entrypoint for the runner process.
+
+    Expects a single command-line argument containing JSON-serialized RunnerCliArgs.
+    File descriptors for communication are inherited via pass_fds.
+    """
+    if len(sys.argv) < 2:
+        print("Usage: runner <json_args>", file=sys.stderr)
+        sys.exit(1)
+
+    # Parse CLI arguments
+    args_json = sys.argv[1]
+    args = RunnerCliArgs.model_validate_json(args_json)
+
+    # Configure logging
+    logger.remove()
+    logger.add(sys.stderr, level=args.log_level)
+
+    # Set up environment
     fast_synch_override = os.environ.get("EXO_FAST_SYNCH")
     if fast_synch_override != "off":
         os.environ["MLX_METAL_FAST_SYNCH"] = "1"
@@ -29,8 +47,16 @@ def entrypoint(
 
     logger.info(f"Fast synch flag: {os.environ['MLX_METAL_FAST_SYNCH']}")
 
-    # Import main after setting global logger - this lets us just import logger from this module
+    # Create FD-based channels with TypeAdapters
+    # Child uses: event_fd for writing events, task_fd for reading tasks, cancel_fd for reading cancels
+    event_sender = FdSender(args.event_fd, event_adapter)
+    task_receiver = FdReceiver(args.task_fd, task_adapter)
+    cancel_receiver = FdReceiver(args.cancel_fd, task_id_adapter)
+
+    bound_instance = args.bound_instance
+
     try:
+        # Import main after setting global logger - this lets us just import logger from this module
         if bound_instance.is_image_model:
             from exo.worker.runner.image_models.runner import main
         else:
@@ -38,23 +64,28 @@ def entrypoint(
 
         main(bound_instance, event_sender, task_receiver, cancel_receiver)
 
-    except ClosedResourceError:
-        logger.warning("Runner communication closed unexpectedly")
     except Exception as e:
         logger.opt(exception=e).warning(
             f"Runner {bound_instance.bound_runner_id} crashed with critical exception {e}"
         )
-        event_sender.send(
-            RunnerStatusUpdated(
-                runner_id=bound_instance.bound_runner_id,
-                runner_status=RunnerFailed(error_message=str(e)),
+        with contextlib.suppress(Exception):
+            event_sender.send(
+                RunnerStatusUpdated(
+                    runner_id=bound_instance.bound_runner_id,
+                    runner_status=RunnerFailed(error_message=str(e)),
+                )
             )
-        )
     finally:
-        try:
+        # Clean up channels
+        with contextlib.suppress(Exception):
             event_sender.close()
+        with contextlib.suppress(Exception):
             task_receiver.close()
-        finally:
-            event_sender.join()
-            task_receiver.join()
-            logger.info("bye from the runner")
+        with contextlib.suppress(Exception):
+            cancel_receiver.close()
+
+        logger.info("bye from the runner")
+
+
+if __name__ == "__main__":
+    entrypoint()

src/exo/worker/runner/image_models/runner.py

@@ -55,7 +55,7 @@ from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
 )
-from exo.utils.channels import MpReceiver, MpSender
+from exo.utils.fd_channels import FdReceiver, FdSender
 from exo.worker.engines.image import (
     DistributedImageModel,
     generate_image,
@@ -88,7 +88,7 @@ def _process_image_response(
     response: ImageGenerationResponse | PartialImageResponse,
     command_id: CommandId,
     shard_metadata: ShardMetadata,
-    event_sender: MpSender[Event],
+    event_sender: FdSender[Event],
     image_index: int,
 ) -> None:
     """Process a single image response and send chunks."""
@@ -111,7 +111,7 @@ def _process_image_response(
 
 
 def _send_traces_if_enabled(
-    event_sender: MpSender[Event],
+    event_sender: FdSender[Event],
     task_id: TaskId,
     rank: int,
 ) -> None:
@@ -144,7 +144,7 @@ def _send_image_chunk(
     encoded_data: str,
     command_id: CommandId,
     model_id: ModelId,
-    event_sender: MpSender[Event],
+    event_sender: FdSender[Event],
     image_index: int,
     is_partial: bool,
     partial_index: int | None = None,
@@ -184,9 +184,9 @@ def _send_image_chunk(
 
 def main(
     bound_instance: BoundInstance,
-    event_sender: MpSender[Event],
-    task_receiver: MpReceiver[Task],
-    cancel_receiver: MpReceiver[TaskId],
+    event_sender: FdSender[Event],
+    task_receiver: FdReceiver[Task],
+    cancel_receiver: FdReceiver[TaskId],
 ):
     soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
     resource.setrlimit(resource.RLIMIT_NOFILE, (min(max(soft, 2048), hard), hard))

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

@@ -31,6 +31,7 @@ from exo.shared.types.events import (
     TaskAcknowledged,
     TaskStatusUpdated,
 )
+from exo.shared.types.mlx import Model
 from exo.shared.types.tasks import (
     ConnectToGroup,
     LoadModel,
@@ -62,8 +63,7 @@ from exo.shared.types.worker.runners import (
     RunnerStatus,
     RunnerWarmingUp,
 )
-from exo.utils.channels import MpReceiver, MpSender
-from exo.worker.engines.mlx import Model
+from exo.utils.fd_channels import FdReceiver, FdSender
 from exo.worker.engines.mlx.cache import KVPrefixCache
 from exo.worker.engines.mlx.generator.generate import (
     PrefillCancelled,
@@ -85,9 +85,9 @@ from .tool_parsers import ToolParser, make_mlx_parser
 
 def main(
     bound_instance: BoundInstance,
-    event_sender: MpSender[Event],
-    task_receiver: MpReceiver[Task],
-    cancel_receiver: MpReceiver[TaskId],
+    event_sender: FdSender[Event],
+    task_receiver: FdReceiver[Task],
+    cancel_receiver: FdReceiver[TaskId],
 ):
     soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
     resource.setrlimit(resource.RLIMIT_NOFILE, (min(max(soft, 2048), hard), hard))
@@ -97,6 +97,7 @@ def main(
         bound_instance.bound_runner_id,
         bound_instance.bound_shard,
     )
+    model_id = shard_metadata.model_card.model_id
     device_rank = shard_metadata.device_rank
     logger.info("hello from the runner")
     if getattr(shard_metadata, "immediate_exception", False):
@@ -273,20 +274,23 @@ def main(
                     def on_prefill_progress(
                         processed: int,
                         total: int,
-                        _task_id: TaskId = task.task_id,
-                        _group: mx.distributed.Group | None = group,
                     ) -> None:
                         if device_rank == 0:
                             event_sender.send(
                                 ChunkGenerated(
                                     command_id=command_id,
                                     chunk=PrefillProgressChunk(
-                                        model=shard_metadata.model_card.model_id,
+                                        model=model_id,
                                         processed_tokens=processed,
                                         total_tokens=total,
                                     ),
                                 )
                             )
+
+                    def distributed_prompt_progress_callback(
+                        _task_id: TaskId = task.task_id,
+                        _group: mx.distributed.Group | None = group,
+                    ) -> None:
                         cancelled_tasks.update(cancel_receiver.collect())
                         want_to_cancel = (_task_id in cancelled_tasks) or (
                             TaskId("CANCEL_CURRENT_TASK") in cancelled_tasks
@@ -308,6 +312,7 @@ def main(
                             prompt=prompt,
                             kv_prefix_cache=kv_prefix_cache,
                             on_prefill_progress=on_prefill_progress,
+                            distributed_prompt_progress_callback=distributed_prompt_progress_callback,
                             group=group,
                         )
 
@@ -325,7 +330,10 @@ def main(
                         # Model-specific output parsing for tool calls.
                         if isinstance(inference_model, GptOssModel):
                             mlx_generator = parse_gpt_oss(mlx_generator)
-                        elif isinstance(inference_model, DeepseekV32Model):
+                        elif (
+                            isinstance(inference_model, DeepseekV32Model)
+                            and "deepseek" in model_id.normalize().lower()
+                        ):
                             mlx_generator = parse_deepseek_v32(mlx_generator)
                         elif tool_parser:
                             mlx_generator = parse_tool_calls(mlx_generator, tool_parser)
@@ -355,7 +363,7 @@ def main(
                                                 command_id=command_id,
                                                 chunk=ErrorChunk(
                                                     error_message=response.text,
-                                                    model=shard_metadata.model_card.model_id,
+                                                    model=model_id,
                                                 ),
                                             )
                                         )
@@ -370,7 +378,7 @@ def main(
                                             ChunkGenerated(
                                                 command_id=command_id,
                                                 chunk=TokenChunk(
-                                                    model=shard_metadata.model_card.model_id,
+                                                    model=model_id,
                                                     text=response.text,
                                                     token_id=response.token,
                                                     usage=response.usage,
@@ -389,7 +397,7 @@ def main(
                                                 command_id=command_id,
                                                 chunk=ToolCallChunk(
                                                     tool_calls=response.tool_calls,
-                                                    model=shard_metadata.model_card.model_id,
+                                                    model=model_id,
                                                     usage=response.usage,
                                                     stats=response.stats,
                                                 ),
@@ -405,7 +413,7 @@ def main(
                                 ChunkGenerated(
                                     command_id=command_id,
                                     chunk=ErrorChunk(
-                                        model=shard_metadata.model_card.model_id,
+                                        model=model_id,
                                         finish_reason="error",
                                         error_message=str(e),
                                     ),
@@ -727,7 +735,7 @@ def parse_tool_calls(
     in_tool_call = False
     tool_call_text_parts: list[str] = []
     for response in responses:
-        if response.text.startswith(tool_parser.start_parsing):
+        if not in_tool_call and response.text.startswith(tool_parser.start_parsing):
             in_tool_call = True
 
         if in_tool_call:
@@ -765,10 +773,9 @@ def parse_tool_calls(
                 )
                 yield response
 
-            continue
-
-        # fallthrough
-        yield response
+        else:
+            # fallthrough
+            yield response
 
 
 EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"

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

@@ -58,15 +58,16 @@ def _flatten(p: dict[str, Any]) -> dict[str, str]:
     }
 
 
-json_tool_parser = ToolParser(
-    start_parsing="<tool_call>",
-    end_parsing="</tool_call>",
-    parse_tool_calls=_parse_json_calls,
-)
+def make_json_parser() -> ToolParser:
+    return ToolParser(
+        start_parsing="<tool_call>",
+        end_parsing="</tool_call>",
+        parse_tool_calls=_parse_json_calls,
+    )
 
 
 def infer_tool_parser(chat_template: str) -> ToolParser | None:
     """Attempt to auto-infer a tool parser from the chat template."""
     if "<tool_call>" in chat_template and "tool_call.name" in chat_template:
-        return json_tool_parser
+        return make_json_parser()
     return None

src/exo/worker/runner/runner_supervisor.py

@@ -1,6 +1,8 @@
 import contextlib
-import multiprocessing as mp
+import json
+import os
 import signal
+import sys
 from dataclasses import dataclass, field
 from typing import Self
 
@@ -8,9 +10,11 @@ import anyio
 from anyio import (
     BrokenResourceError,
     ClosedResourceError,
-    to_thread,
 )
+from anyio.abc import Process
+from anyio.streams.text import TextReceiveStream
 from loguru import logger
+from pydantic import TypeAdapter
 
 from exo.shared.types.events import (
     Event,
@@ -20,6 +24,7 @@ from exo.shared.types.events import (
 )
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.instances import BoundInstance
+from exo.shared.types.worker.runner_args import RunnerCliArgs
 from exo.shared.types.worker.runners import (
     RunnerConnecting,
     RunnerFailed,
@@ -31,24 +36,29 @@ from exo.shared.types.worker.runners import (
     RunnerWarmingUp,
 )
 from exo.shared.types.worker.shards import ShardMetadata
-from exo.utils.channels import MpReceiver, MpSender, Sender, mp_channel
+from exo.utils.channels import Sender
+from exo.utils.fd_channels import FdReceiver, FdSender
 from exo.utils.task_group import TaskGroup
-from exo.worker.runner.bootstrap import entrypoint
 
 PREFILL_TIMEOUT_SECONDS = 60
 DECODE_TIMEOUT_SECONDS = 5
 
+# Type adapters for union types
+event_adapter: TypeAdapter[Event] = TypeAdapter(Event)
+task_adapter: TypeAdapter[Task] = TypeAdapter(Task)
+task_id_adapter: TypeAdapter[TaskId] = TypeAdapter(TaskId)
+
 
 @dataclass(eq=False)
 class RunnerSupervisor:
     shard_metadata: ShardMetadata
     bound_instance: BoundInstance
-    runner_process: mp.Process
+    process: Process
     initialize_timeout: float
-    _ev_recv: MpReceiver[Event]
-    _task_sender: MpSender[Task]
+    _ev_recv: FdReceiver[Event]
+    _task_sender: FdSender[Task]
     _event_sender: Sender[Event]
-    _cancel_sender: MpSender[TaskId]
+    _cancel_sender: FdSender[TaskId]
     _tg: TaskGroup = field(default_factory=TaskGroup, init=False)
     status: RunnerStatus = field(default_factory=RunnerIdle, init=False)
     pending: dict[TaskId, anyio.Event] = field(default_factory=dict, init=False)
@@ -59,35 +69,75 @@ class RunnerSupervisor:
     )
 
     @classmethod
-    def create(
+    async def create(
         cls,
         *,
         bound_instance: BoundInstance,
         event_sender: Sender[Event],
         initialize_timeout: float = 400,
     ) -> Self:
-        ev_send, ev_recv = mp_channel[Event]()
-        task_sender, task_recv = mp_channel[Task]()
-        cancel_sender, cancel_recv = mp_channel[TaskId]()
+        # Create pipes for each communication direction:
+        # - Events: child -> parent (child writes, parent reads)
+        # - Tasks: parent -> child (parent writes, child reads)
+        # - Cancels: parent -> child (parent writes, child reads)
+        event_read_fd, event_write_fd = os.pipe()  # Child writes, parent reads
+        task_read_fd, task_write_fd = os.pipe()  # Parent writes, child reads
+        cancel_read_fd, cancel_write_fd = os.pipe()  # Parent writes, child reads
 
-        runner_process = mp.Process(
-            target=entrypoint,
-            args=(
-                bound_instance,
-                ev_send,
-                task_recv,
-                cancel_recv,
-                logger,
-            ),
-            daemon=True,
+        # Build CLI args with only the FDs the child needs
+        args = RunnerCliArgs(
+            bound_instance=bound_instance,
+            event_fd=event_write_fd,  # Child writes events here
+            task_fd=task_read_fd,  # Child reads tasks from here
+            cancel_fd=cancel_read_fd,  # Child reads cancels from here
+            log_level="INFO",
         )
 
+        # File descriptors to pass to the child process
+        pass_fds = [event_write_fd, task_read_fd, cancel_read_fd]
+
         shard_metadata = bound_instance.bound_shard
 
-        self = cls(
+        # Create FD-based channels for the parent side using TypeAdapters
+        ev_recv = FdReceiver[Event](event_read_fd, event_adapter)
+        task_sender = FdSender[Task](task_write_fd, task_adapter)
+        cancel_sender = FdSender[TaskId](cancel_write_fd, task_id_adapter)
+
+        # Close the child-side FDs in the parent process
+        with contextlib.suppress(OSError):
+            os.close(event_write_fd)
+        with contextlib.suppress(OSError):
+            os.close(task_read_fd)
+        with contextlib.suppress(OSError):
+            os.close(cancel_read_fd)
+
+        # Start the process using anyio.open_process
+        # Check if running in a frozen environment (e.g., PyInstaller)
+        is_frozen = os.environ.get("EXO_FROZEN") == "1" or getattr(sys, "frozen", False)
+
+        if is_frozen:
+            # In frozen builds, use the same executable with a special flag
+            # The __main__.py handles this by checking sys.argv
+            cmd = [
+                sys.executable,
+                "--exo-runner",
+                json.dumps(args.model_dump(mode="json")),
+            ]
+        else:
+            # In normal Python, use -m to run the module
+            cmd = [
+                sys.executable,
+                "-m",
+                "exo.worker.runner.bootstrap",
+                json.dumps(args.model_dump(mode="json")),
+            ]
+
+        process = await anyio.open_process(cmd, pass_fds=pass_fds)
+
+        return cls(
             bound_instance=bound_instance,
             shard_metadata=shard_metadata,
-            runner_process=runner_process,
+            process=process,
             initialize_timeout=initialize_timeout,
             _ev_recv=ev_recv,
             _task_sender=task_sender,
@@ -95,38 +145,49 @@ class RunnerSupervisor:
             _event_sender=event_sender,
         )
 
-        return self
-
     async def run(self):
-        self.runner_process.start()
         async with self._tg as tg:
             tg.start_soon(self._watch_runner)
             tg.start_soon(self._forward_events)
+            tg.start_soon(self._read_stderr)
 
-    def shutdown(self):
+    async def _read_stderr(self):
+        """Read and log stderr from the runner process."""
+        if self.process.stderr:
+            async with TextReceiveStream(self.process.stderr) as stderr_stream:
+                async for line in stderr_stream:
+                    logger.debug(f"Runner stderr: {line.rstrip()}")
+
+    async def shutdown(self):
         logger.info("Runner supervisor shutting down")
         self._tg.cancel_tasks()
-        self._ev_recv.close()
-        self._task_sender.close()
         if not self._cancel_watch_runner.cancel_called:
             self._cancel_watch_runner.cancel()
+        with contextlib.suppress(ClosedResourceError):
+            self._ev_recv.close()
+        with contextlib.suppress(ClosedResourceError):
+            self._task_sender.close()
+        with contextlib.suppress(ClosedResourceError):
+            self._event_sender.close()
         with contextlib.suppress(ClosedResourceError):
             self._cancel_sender.send(TaskId("CANCEL_CURRENT_TASK"))
-        self._cancel_sender.close()
-        self.runner_process.join(5)
-        if not self.runner_process.is_alive():
-            logger.info("Runner process succesfully terminated")
-            return
+        with contextlib.suppress(ClosedResourceError):
+            self._cancel_sender.close()
 
-        # This is overkill but it's not technically bad, just unnecessary.
-        logger.warning("Runner process didn't shutdown succesfully, terminating")
-        self.runner_process.terminate()
-        self.runner_process.join(1)
-        if not self.runner_process.is_alive():
-            return
-
-        logger.critical("Runner process didn't respond to SIGTERM, killing")
-        self.runner_process.kill()
+        # Try graceful termination
+        if self.process.returncode is None:
+            self.process.terminate()
+            try:
+                with anyio.fail_after(5):
+                    await self.process.wait()
+            except TimeoutError:
+                logger.warning("Runner process didn't shutdown gracefully, killing")
+                self.process.kill()
+                try:
+                    with anyio.fail_after(1):
+                        await self.process.wait()
+                except TimeoutError:
+                    logger.critical("Runner process didn't respond to SIGKILL")
 
     async def start_task(self, task: Task):
         if task.task_id in self.pending:
@@ -193,38 +254,48 @@ class RunnerSupervisor:
                 self.pending[tid].set()
 
     def __del__(self) -> None:
-        if self.runner_process.is_alive():
+        if self.process.returncode is None:
             logger.critical("RunnerSupervisor was not stopped cleanly.")
-            with contextlib.suppress(ValueError):
-                self.runner_process.kill()
+            with contextlib.suppress(Exception):
+                self.process.kill()
 
     async def _watch_runner(self) -> None:
         with self._cancel_watch_runner:
             while True:
                 await anyio.sleep(5)
-                if not self.runner_process.is_alive():
+                if self.process.returncode is not None:
                     await self._check_runner(RuntimeError("Runner found to be dead"))
 
     async def _check_runner(self, e: Exception) -> None:
         if not self._cancel_watch_runner.cancel_called:
             self._cancel_watch_runner.cancel()
         logger.info("Checking runner's status")
-        if self.runner_process.is_alive():
-            logger.info("Runner was found to be alive, attempting to join process")
-            await to_thread.run_sync(self.runner_process.join, 5)
-        rc = self.runner_process.exitcode
-        logger.info(f"Runner exited with exit code {rc}")
-        if rc == 0:
+
+        returncode = self.process.returncode
+
+        if returncode is None:
+            # Process is still running, try to wait for it
+            self.process.terminate()
+            try:
+                with anyio.fail_after(5):
+                    await self.process.wait()
+                    returncode = self.process.returncode
+            except TimeoutError:
+                pass
+
+        logger.info(f"Runner exited with return code {returncode}")
+
+        if returncode == 0:
             return
 
-        if isinstance(rc, int) and rc < 0:
-            sig = -rc
+        if isinstance(returncode, int) and returncode < 0:
+            sig = -returncode
             try:
                 cause = f"signal={sig} ({signal.strsignal(sig)})"
             except Exception:
                 cause = f"signal={sig}"
         else:
-            cause = f"exitcode={rc}"
+            cause = f"returncode={returncode}"
 
         logger.opt(exception=e).error(f"Runner terminated with {cause}")
 
@@ -243,4 +314,4 @@ class RunnerSupervisor:
             logger.warning(
                 "Event sender already closed, unable to report runner failure"
             )
-        self.shutdown()
+        await self.shutdown()

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

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

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

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

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

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

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

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

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

@@ -39,7 +39,6 @@ from exo.shared.types.worker.runners import (
     RunnerShuttingDown,
     RunnerWarmingUp,
 )
-from exo.utils.channels import mp_channel
 
 from ...constants import (
     CHAT_COMPLETION_TASK_ID,
@@ -139,8 +138,46 @@ class EventCollector:
     def close(self) -> None:
         pass
 
-    def join(self) -> None:
-        pass
+
+class MockFdReceiver[T]:
+    """Mock FdReceiver that iterates over a list of items."""
+
+    def __init__(self, items: list[T]) -> None:
+        self._items = items
+        self._index = 0
+        self._closed = False
+
+    def receive(self) -> T:
+        if self._closed:
+            raise StopIteration
+        if self._index >= len(self._items):
+            raise StopIteration
+        item = self._items[self._index]
+        self._index += 1
+        return item
+
+    def __iter__(self):
+        return self
+
+    def __next__(self) -> T:
+        return self.receive()
+
+    def close(self) -> None:
+        self._closed = True
+
+    def collect(self) -> list[T]:
+        """Collect all remaining items."""
+        items = []
+        while self._index < len(self._items):
+            items.append(self._items[self._index])
+            self._index += 1
+        return items
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        self.close()
 
 
 class MockTokenizer:
@@ -167,28 +204,21 @@ def _run(tasks: Iterable[Task]):
         node_id=NODE_A,
     )
 
-    task_sender, task_receiver = mp_channel[Task]()
-    _cancel_sender, cancel_receiver = mp_channel[TaskId]()
+    task_list = list(tasks)
+    task_receiver = MockFdReceiver[Task](task_list)
+    cancel_receiver = MockFdReceiver[TaskId]([])
     event_sender = EventCollector()
 
-    with task_sender:
-        for t in tasks:
-            task_sender.send(t)
-
-        # worst monkeypatch known to man
-        # this is some c++ nonsense
-        task_receiver.close = nothin
-        task_receiver.join = nothin
-        with unittest.mock.patch(
-            "exo.worker.runner.llm_inference.runner.mx.distributed.all_gather",
-            make_nothin(mx.array([1])),
-        ):
-            mlx_runner.main(
-                bound_instance,
-                event_sender,  # pyright: ignore[reportArgumentType]
-                task_receiver,
-                cancel_receiver,
-            )
+    with unittest.mock.patch(
+        "exo.worker.runner.llm_inference.runner.mx.distributed.all_gather",
+        make_nothin(mx.array([1])),
+    ):
+        mlx_runner.main(
+            bound_instance,
+            event_sender,  # pyright: ignore[reportArgumentType]
+            task_receiver,  # pyright: ignore[reportArgumentType]
+            cancel_receiver,  # pyright: ignore[reportArgumentType]
+        )
 
         return event_sender.events
 

tests/auto_bench.sh

@@ -17,6 +17,13 @@ git branch -r --contains "$commit" | grep -qE '^\s*origin/' || {
   exit 1
 }
 hosts=("$@")
+
+for host; do
+  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
+    "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix build github:exo-explore/exo/$commit" &
+done
+wait
+
 cleanup() {
   for host in "${hosts[@]}"; do
     ssh -T -o BatchMode=yes "$host@$host" "pkill -f bin/exo" &
@@ -26,11 +33,6 @@ cleanup() {
 }
 trap 'cleanup' EXIT INT TERM
 
-for host; do
-  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
-    "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix build github:exo-explore/exo/$commit" &
-done
-wait
 for host; do
   ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
     "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix run github:exo-explore/exo/$commit" &>/dev/null &

tmp/test_trust_remote_code_attack.sh

@@ -0,0 +1,125 @@
+#!/usr/bin/env bash
+# Test that models added via API get trust_remote_code=false
+# Run this against a running exo instance.
+# Usage: ./test_trust_remote_code_attack.sh [host:port]
+
+set -uo pipefail
+
+HOST="${1:-localhost:52415}"
+MODEL_ID="KevTheHermit/security-testing"
+CUSTOM_CARDS_DIR="$HOME/.exo/custom_model_cards"
+CARD_FILE="$CUSTOM_CARDS_DIR/KevTheHermit--security-testing.toml"
+
+echo "=== Test: trust_remote_code attack via API ==="
+echo "Target: $HOST"
+echo ""
+
+# Clean up RCE proof from previous runs
+rm -f /tmp/exo-rce-proof.txt
+
+# Step 0: Clean up any stale card from previous runs
+if [ -f "$CARD_FILE" ]; then
+  echo "[0] Removing stale card from previous run ..."
+  curl -s -X DELETE \
+    "http://$HOST/models/custom/$(python3 -c 'import urllib.parse; print(urllib.parse.quote("'"$MODEL_ID"'", safe=""))')" >/dev/null
+  rm -f "$CARD_FILE"
+  echo "    Done"
+  echo ""
+fi
+
+# Step 1: Add the malicious model via API
+echo "[1] Adding model via POST /models/add ..."
+ADD_RESPONSE=$(curl -s -w "\n%{http_code}" -X POST "http://$HOST/models/add" \
+  -H "Content-Type: application/json" \
+  -d "{\"model_id\":\"$MODEL_ID\"}")
+HTTP_CODE=$(echo "$ADD_RESPONSE" | tail -1)
+BODY=$(echo "$ADD_RESPONSE" | sed '$d')
+echo "    HTTP $HTTP_CODE"
+
+if [ "$HTTP_CODE" -ge 400 ]; then
+  echo "    Model add failed (HTTP $HTTP_CODE) — that's fine if model doesn't exist on HF."
+  echo "    Response: $BODY"
+  echo ""
+  echo "RESULT: Model was rejected at add time. Attack blocked."
+  exit 0
+fi
+
+# Step 2: Verify the saved TOML has trust_remote_code = false
+echo ""
+echo "[2] Checking saved model card TOML ..."
+if [ ! -f "$CARD_FILE" ]; then
+  echo "    FAIL: Card file not found at $CARD_FILE"
+  exit 1
+fi
+
+if grep -q 'trust_remote_code = false' "$CARD_FILE"; then
+  echo "    SAFE: trust_remote_code = false (fix is active)"
+else
+  echo "    VULNERABLE: trust_remote_code is not false — remote code WILL be trusted"
+fi
+echo "    Contents:"
+cat "$CARD_FILE"
+
+# Step 3: Place the instance
+echo ""
+echo "[3] Attempting POST /place_instance ..."
+PLACE_RESPONSE=$(curl -s -w "\n%{http_code}" -X POST "http://$HOST/place_instance" \
+  -H "Content-Type: application/json" \
+  -d "{\"model_id\":\"$MODEL_ID\"}")
+PLACE_CODE=$(echo "$PLACE_RESPONSE" | tail -1)
+PLACE_BODY=$(echo "$PLACE_RESPONSE" | sed '$d')
+echo "    HTTP $PLACE_CODE"
+echo "    Response: $PLACE_BODY"
+
+# Step 3b: Send a chat completion to actually trigger tokenizer loading
+echo ""
+echo "[3b] Sending chat completion to trigger tokenizer load ..."
+CHAT_RESPONSE=$(curl -s -w "\n%{http_code}" --max-time 30 -X POST "http://$HOST/v1/chat/completions" \
+  -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL_ID\",\"messages\":[{\"role\":\"user\",\"content\":\"hello\"}],\"max_tokens\":1}")
+CHAT_CODE=$(echo "$CHAT_RESPONSE" | tail -1)
+CHAT_BODY=$(echo "$CHAT_RESPONSE" | sed '$d')
+echo "    HTTP $CHAT_CODE"
+echo "    Response: $CHAT_BODY"
+echo ""
+echo "[3c] Checking for RCE proof ..."
+sleep 5
+if [ -f /tmp/exo-rce-proof.txt ]; then
+  echo "    VULNERABLE: Remote code executed!"
+  echo "    Contents:"
+  cat /tmp/exo-rce-proof.txt
+else
+  echo "    SAFE: /tmp/exo-rce-proof.txt does not exist — remote code was NOT executed"
+fi
+
+# Step 4: Clean up — delete instance and custom model
+echo ""
+echo "[4] Cleaning up ..."
+
+# Find and delete any instance for this model
+INSTANCE_ID=$(curl -s "http://$HOST/state" | python3 -c "
+import sys, json
+state = json.load(sys.stdin)
+for iid, wrapper in state.get('instances', {}).items():
+    for tag, inst in wrapper.items():
+        sa = inst.get('shardAssignments', {})
+        if sa.get('modelId', '') == '$MODEL_ID':
+            print(iid)
+            sys.exit(0)
+" 2>/dev/null || true)
+
+if [ -n "$INSTANCE_ID" ]; then
+  echo "    Deleting instance $INSTANCE_ID ..."
+  curl -s -X DELETE "http://$HOST/instance/$INSTANCE_ID" >/dev/null
+  echo "    Done"
+else
+  echo "    No instance found to delete"
+fi
+
+echo "    Deleting custom model card ..."
+curl -s -X DELETE \
+  "http://$HOST/models/custom/$(python3 -c 'import urllib.parse; print(urllib.parse.quote("'"$MODEL_ID"'", safe=""))')" >/dev/null
+echo "    Done"
+
+echo ""
+echo "=== DONE ==="

uv.lock

@@ -363,7 +363,7 @@ wheels = [
 
 [[package]]
 name = "exo"
-version = "0.3.0"
+version = "0.3.68"
 source = { editable = "." }
 dependencies = [
     { name = "aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -378,7 +378,7 @@ dependencies = [
     { name = "loguru", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "mflux", marker = "sys_platform == 'darwin' or 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.dev20260220+13998a05", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#13998a054715edcdc93618fb1496c79c7c25ff7c" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.30.7.dev20260225+257d5692", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#257d5692fc7af6bba3b8afaeb63c549b7d1e43d5" }, 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'" },
@@ -469,7 +469,7 @@ requires-dist = [
 
 [[package]]
 name = "exo-pyo3-bindings"
-version = "0.1.0"
+version = "0.2.0"
 source = { editable = "rust/exo_pyo3_bindings" }
 
 [package.dev-dependencies]
@@ -1025,7 +1025,7 @@ dependencies = [
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "matplotlib", marker = "sys_platform == 'darwin' or 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.dev20260220+13998a05", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#13998a054715edcdc93618fb1496c79c7c25ff7c" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.30.7.dev20260225+257d5692", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#257d5692fc7af6bba3b8afaeb63c549b7d1e43d5" }, 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'" },
@@ -1072,8 +1072,8 @@ cuda13 = [
 
 [[package]]
 name = "mlx"
-version = "0.30.7.dev20260220+13998a05"
-source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#13998a054715edcdc93618fb1496c79c7c25ff7c" }
+version = "0.30.7.dev20260225+257d5692"
+source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#257d5692fc7af6bba3b8afaeb63c549b7d1e43d5" }
 resolution-markers = [
     "sys_platform == 'darwin'",
 ]
@@ -1108,7 +1108,7 @@ version = "0.30.7"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", version = "0.30.7.dev20260220+13998a05", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#13998a054715edcdc93618fb1496c79c7c25ff7c" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.30.7.dev20260225+257d5692", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#257d5692fc7af6bba3b8afaeb63c549b7d1e43d5" }, 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'" },