Merge branch 'main' into andrei/force_oom

# Conflicts:
#	uv.lock

.envrc

@@ -1 +1,13 @@
+# installs nix-direnv if doesn't exist (speeds up evaluation)
+if ! has nix_direnv_version || ! nix_direnv_version 3.1.1; then
+  source_url "https://raw.githubusercontent.com/nix-community/nix-direnv/3.1.1/direnvrc" "sha256-p+fzQdrms/hDa7g+soShAybJNo4bN4SIAeSfqNKgD5I="
+fi
+
 use flake
+
+# creates .venv if doesn't exist and loads its environment
+export VIRTUAL_ENV=".venv"
+if ! [ -d "./$VIRTUAL_ENV" ]; then
+  uv venv
+fi
+layout python

.gitignore

@@ -38,8 +38,6 @@ bench/**/*.json
 # tmp
 tmp/models
 /build/exo
-/.agents
 /.claude/skills
 /.claude
 /.codex
-skills-lock.json

Cargo.toml

@@ -1,6 +1,6 @@
 [workspace]
 resolver = "3"
-members = ["rust/networking", "rust/exo_rs", "rust/util"]
+members = ["rust/networking", "rust/exo_pyo3_bindings", "rust/util"]
 
 [workspace.package]
 version = "0.0.1"
@@ -28,7 +28,6 @@ delegate = "0.13"
 
 # Utility dependencies
 keccak-const = "0.2"
-nix = "0.31"
 
 # Async dependencies
 async-stream = "0.3"

README.md

@@ -201,12 +201,6 @@ This starts the exo dashboard and API at http://localhost:52415/
   uv run exo --no-worker
   ```
 
-- `--legacy-daemon`: Run exo as a legacy SysV-style background daemon using double-fork daemonization. This is intended for legacy init scripts; systemd and launchd should run exo in the foreground without this flag.
-
-  ```bash
-  uv run exo --legacy-daemon
-  ```
-
 **File Locations (Linux):**
 
 exo follows the [XDG Base Directory Specification](https://specifications.freedesktop.org/basedir-spec/basedir-spec-latest.html) on Linux:
@@ -401,18 +395,6 @@ Sample response:
 }
 ```
 
-This command is asynchronous. Before sending inference requests, wait until the
-API sees the new instance for this model:
-
-```bash
-curl -N "http://localhost:52415/instance/await?model_id=mlx-community/Llama-3.2-1B-Instruct-4bit"
-```
-
-The endpoint returns an SSE stream. A successful wait emits a message with
-`"type": "ready"` and the matching instance; a timeout emits `"type": "timeout"`.
-By default it waits indefinitely. Set `timeout_seconds` to a positive value to
-bound the wait.
-
 ---
 
 **3. Send a chat completion**

app/EXO/EXO/Services/ClusterStateService.swift

@@ -37,7 +37,7 @@ final class ClusterStateService: ObservableObject {
     /// gain nothing from being cached on disk. Use an ephemeral session
     /// with `urlCache = nil` so neither response bodies nor metadata
     /// touch disk.
-    nonisolated private static func makeNonCachingSession() -> URLSession {
+    private static func makeNonCachingSession() -> URLSession {
         let config = URLSessionConfiguration.ephemeral
         config.urlCache = nil
         config.requestCachePolicy = .reloadIgnoringLocalCacheData

bench/METHODOLOGY.md

@@ -125,7 +125,7 @@ A background thread polls each node at 1 Hz, collecting:
 - System power draw (W)
 - CPU cluster usage (performance and efficiency cores)
 
-**Energy** is computed via trapezoidal integration of the power samples over each inference window (the wall-clock span of each benchmark request or concurrent batch). Average power is `total_joules / total_inference_seconds`. The server additionally returns a `power_usage` block in each non-stream `/bench/chat/completions` response that splits energy into prefill and generation phases, with the boundary anchored to the first non-`PrefillProgressChunk` from the runner.
+**Energy** is computed via trapezoidal integration of the power samples over each inference window (the wall-clock span of each benchmark request or concurrent batch). Average power is `total_joules / total_inference_seconds`.
 
 ---
 
@@ -136,7 +136,6 @@ Results are written as JSON with three top-level keys:
 - **`runs`**: Array of per-request result objects, each containing:
   - `elapsed_s`, `output_text_preview` (first 200 chars)
   - `stats`: `{ prompt_tps, generation_tps, prompt_tokens, generation_tokens, peak_memory_usage }`
-  - `power_usage`: server-side total + prefill/generation split, per-node breakdown (non-stream requests only)
   - Placement metadata: `model_id`, `placement_sharding`, `placement_instance_meta`, `placement_nodes`
   - Run metadata: `pp_tokens`, `tg`, `repeat_index`, `concurrency`, `concurrent_index`
   - `download_duration_s` (if model was freshly downloaded)

bench/exo_bench.py

@@ -295,7 +295,6 @@ def run_one_completion(
         elapsed = time.perf_counter() - t0
 
         stats = out.get("generation_stats")
-        power_usage = out.get("power_usage")
         choices = out.get("choices") or [{}]
         message = choices[0].get("message", {}) if choices else {}
         content = message.get("content") or ""
@@ -331,7 +330,6 @@ def run_one_completion(
 
         elapsed = time.perf_counter() - t0
         preview = "".join(text_parts)[:200]
-        power_usage = None
 
         if not stats:
             ttft = (first_token_time - t0) if first_token_time else elapsed
@@ -350,7 +348,6 @@ def run_one_completion(
         "elapsed_s": elapsed,
         "output_text_preview": preview,
         "stats": stats,
-        "power_usage": power_usage,
     }, pp_tokens
 
 
@@ -767,7 +764,6 @@ def main() -> int:
                                 out = c.post_bench_chat_completions(_payload)
                                 elapsed = time.perf_counter() - t0
                                 stats = out.get("generation_stats")
-                                power_usage = out.get("power_usage")
                                 choices = out.get("choices") or [{}]
                                 message = (
                                     choices[0].get("message", {}) if choices else {}
@@ -777,7 +773,6 @@ def main() -> int:
                                     "elapsed_s": elapsed,
                                     "output_text_preview": text[:200],
                                     "stats": stats,
-                                    "power_usage": power_usage,
                                 }, _actual_pp
 
                             inf_t0 = time.monotonic()
@@ -873,27 +868,6 @@ def main() -> int:
                             inf_seconds = sum(t1 - t0 for t0, t1 in inference_windows)
                             avg_watts = joules / inf_seconds if inf_seconds > 0 else 0
                             summary += f"    energy={joules:.1f}J ({avg_watts:.1f}W avg over {inf_seconds:.1f}s inference)"
-
-                        # mean() not sum() across concurrent runs: each
-                        # request's PowerSampler observes the same shared
-                        # cluster state, so they all report the same figure.
-                        prefill_energies = [
-                            (x.get("power_usage") or {}).get("prefill_energy_joules")
-                            for x in runs
-                        ]
-                        gen_energies = [
-                            (x.get("power_usage") or {}).get("generation_energy_joules")
-                            for x in runs
-                        ]
-                        prefill_vals = [e for e in prefill_energies if e is not None]
-                        gen_vals = [e for e in gen_energies if e is not None]
-                        if prefill_vals and gen_vals:
-                            avg_pref = mean(prefill_vals)
-                            avg_gen = mean(gen_vals)
-                            summary += (
-                                f"    prefill_energy={avg_pref:.1f}J  "
-                                f"gen_energy={avg_gen:.1f}J"
-                            )
                         logger.info(f"{summary}\n")
                     time.sleep(2)
         finally:

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

@@ -2253,9 +2253,10 @@ class AppStore {
    * @returns The model ID to use, or null if none available
    */
   private getModelForRequest(modelId?: string): string | null {
-    const requestedModelId = modelId || this.selectedChatModel;
+    if (modelId) return modelId;
+    if (this.selectedChatModel) return this.selectedChatModel;
 
-    // Only models with a placed instance can receive requests; disk downloads alone are not enough.
+    // Try to get model from first running instance
     for (const [, instanceWrapper] of Object.entries(this.instances)) {
       if (instanceWrapper && typeof instanceWrapper === "object") {
         const keys = Object.keys(instanceWrapper as Record<string, unknown>);
@@ -2263,15 +2264,8 @@ class AppStore {
           const instance = (instanceWrapper as Record<string, unknown>)[
             keys[0]
           ] as { shardAssignments?: { modelId?: string } };
-          const instanceModelId = instance?.shardAssignments?.modelId;
-
-          // ensure to only return requestedModelId that matches an instance
-          // or fall back to first instance
-          if (
-            instanceModelId &&
-            (!requestedModelId || requestedModelId === instanceModelId)
-          ) {
-            return instanceModelId;
+          if (instance?.shardAssignments?.modelId) {
+            return instance.shardAssignments.modelId;
           }
         }
       }

dashboard/src/routes/+page.svelte

@@ -1461,9 +1461,6 @@
         addToast({ type: "info", message: `Launching model...` });
         // Always auto-select the newly launched model so the user chats to what they just launched
         setSelectedChatModel(modelId);
-        userForcedIdle = false;
-        pendingChatModelId = modelId;
-        chatLaunchState = "launching";
 
         // Record the launch in recent models history
         recordRecentLaunch(modelId);
@@ -2550,10 +2547,12 @@
   ];
 
   // ── Seamless chat: launch models from chat view ──
-  type InFlightChatLaunchState = "launching" | "downloading" | "loading";
-  type ReadyLikeChatLaunchState = "idle" | "ready";
-  type ChatLaunchState = InFlightChatLaunchState | ReadyLikeChatLaunchState;
-
+  type ChatLaunchState =
+    | "idle"
+    | "launching"
+    | "downloading"
+    | "loading"
+    | "ready";
   let chatLaunchState = $state<ChatLaunchState>("idle");
   let pendingChatModelId = $state<string | null>(null);
   let selectedChatCategory = $state<string | null>(null);
@@ -3130,15 +3129,6 @@
     if (model) {
       pendingAutoMessage = { content, files };
       userForcedIdle = false;
-      // The selected model is already being placed or loaded; keep the queued
-      // message and let the existing launch state effects send it once ready.
-      if (
-        pendingChatModelId === model &&
-        chatLaunchState !== "idle" &&
-        chatLaunchState !== "ready"
-      ) {
-        return;
-      }
       launchModelForChat(model, "picker", messages().length > 0);
       return;
     }
@@ -4613,7 +4603,7 @@
                 type="button"
                 onclick={() => {
                   completeOnboarding();
-                  handleChatSend(chip);
+                  sendMessage(chip, undefined, thinkingEnabled());
                 }}
                 class="px-4 py-2 rounded-full border border-white/10 bg-white/5 text-sm text-white/60 hover:bg-white/10 hover:text-white/80 hover:border-white/20 transition-all duration-200 cursor-pointer"
               >
@@ -6110,7 +6100,7 @@
                           onclick={() => {
                             chatLaunchState = "idle";
                             selectedChatCategory = null;
-                            handleChatSend(prompt);
+                            sendMessage(prompt, undefined, thinkingEnabled());
                           }}
                           class="text-left px-3 py-2.5 text-xs text-exo-light-gray hover:text-white font-mono rounded-lg border border-exo-medium-gray/30 hover:border-exo-yellow/30 bg-exo-dark-gray/30 hover:bg-exo-dark-gray/60 transition-all duration-200 cursor-pointer"
                         >

docs/api.md

@@ -66,9 +66,7 @@ Creates a new model instance in the cluster.
 ```
 
 **Response:**
-Command acknowledgement. Instance creation is asynchronous; clients should wait
-for the model to appear through `/instance/await` before sending inference
-requests for that model.
+JSON description of the created instance.
 
 ### Delete Instance
 
@@ -96,31 +94,6 @@ Returns details of a specific instance.
 **Response:**
 JSON description of the instance.
 
-### Await Instance
-
-**GET** `/instance/await?model_id=...&timeout_seconds=0`
-
-Waits until API state contains an instance for the requested model. The response
-is an SSE stream so clients receive keep-alive comments while waiting.
-
-**Query parameters:**
-
-* `model_id`: string, required
-* `timeout_seconds`: float, optional, default `0`. `0` waits indefinitely;
-  positive values time out after that many seconds. Maximum positive value:
-  `300`.
-
-**Stream messages:**
-
-```text
-data: {"type": "ready", "instance": {...}}
-
-data: {"type": "timeout", "message": "No instance found for model ..."}
-```
-
-The HTTP status is `200` for both messages because the stream starts before the
-final result is known. The `type` field disambiguates the terminal message.
-
 ### Preview Placements
 
 **GET** `/instance/previews?model_id=...`
@@ -150,18 +123,17 @@ Computes a placement for a potential instance without creating it.
 **Response:**
 JSON object describing the proposed placement / instance configuration.
 
-### Place Instance
+### Place Instance (Dry Operation)
 
 **POST** `/place_instance`
 
-Places an instance for a model using the server's placement logic.
+Performs a placement operation for an instance (planning step), without necessarily creating it.
 
 **Request body:**
 JSON describing the instance to be placed.
 
 **Response:**
-Command acknowledgement. The instance may not be ready immediately; wait for it
-to appear through `/instance/await` before sending inference requests.
+Placement result.
 
 ## 3. Models
 
@@ -667,11 +639,10 @@ GET     /events
 
 # Instance Management
 POST    /instance
-GET     /instance/await
-GET     /instance/previews
-GET     /instance/placement
 GET     /instance/{instance_id}
 DELETE  /instance/{instance_id}
+GET     /instance/previews
+GET     /instance/placement
 POST    /place_instance
 
 # Models

flake.nix

@@ -110,7 +110,7 @@
               nixpkgs-fmt.enable = true;
               ruff-format = {
                 enable = true;
-                excludes = [ "rust/exo_rs/exo_rs.pyi" ];
+                excludes = [ "rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi" ];
               };
               rustfmt = {
                 enable = true;

justfile

@@ -23,7 +23,7 @@ sync-clean:
 
 rust-rebuild:
     PYO3_PYTHON="$(uv run python -c 'import sys; print(sys.executable)')" cargo run --bin stub_gen
-    uv sync --reinstall-package exo_rs
+    uv sync --reinstall-package exo_pyo3_bindings
 
 build-dashboard:
     #!/usr/bin/env bash
@@ -37,7 +37,7 @@ package: build-dashboard
     rm -rf build
 
 build-app: rust-rebuild sync-clean package
-    env -u LD xcodebuild build -project app/EXO/EXO.xcodeproj -scheme EXO -configuration Debug -derivedDataPath app/EXO/build
+    xcodebuild build -project app/EXO/EXO.xcodeproj -scheme EXO -configuration Debug -derivedDataPath app/EXO/build
     @echo "\nBuild complete. Run with:\n  open {{justfile_directory()}}/app/EXO/build/Build/Products/Debug/EXO.app"
 
 clean:

pyproject.toml

@@ -15,7 +15,7 @@ dependencies = [
   "huggingface-hub>=1.8.0",
   "psutil>=7.0.0",
   "loguru>=0.7.3",
-  "exo-rs",                          # rust bindings
+  "exo-pyo3-bindings",               # rust bindings
   "anyio==4.11.0",
   "tiktoken>=0.12.0",                # required for kimi k2 tokenizer
   "hypercorn>=0.18.0",
@@ -26,7 +26,6 @@ dependencies = [
   "msgspec>=0.19.0",
   "zstandard>=0.23.0",
   "transformers>=5.6.2",
-  "python-daemon>=3.1.2",
 ]
 
 [project.scripts]
@@ -76,14 +75,15 @@ mlx-cuda13 = [
 ###
 
 [tool.uv.workspace]
-members = ["rust/exo_rs", "bench", "tools"]
+members = ["rust/exo_pyo3_bindings", "bench", "tools"]
 
 [tool.uv.sources]
-exo-rs = { workspace = true }
+exo-pyo3-bindings = { workspace = true }
 mlx = [
   { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git", branch = "address-rdma-gpu-locks", marker = "sys_platform == 'darwin'" },
   { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl", marker = "sys_platform == 'linux' and platform_machine == 'aarch64'" },
   { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl", marker = "sys_platform == 'linux' and platform_machine != 'aarch64'" },
+
 ]
 mlx-lm = { git = "https://github.com/rltakashige/mlx-lm", branch = "leo/deepseek-v4" }
 mflux = { git = "https://github.com/evanev7/mflux", branch = "exo2" }
@@ -240,7 +240,7 @@ torchaudio = ["torch"]
 ###
 
 [tool.ruff]
-extend-exclude = [".typings/**", "rust/exo_rs/**", "bench/vendor/**"]
+extend-exclude = [".typings/**", "rust/exo_pyo3_bindings/**", "bench/vendor/**"]
 
 [tool.ruff.lint]
 extend-select = ["I", "N", "B", "A", "PIE", "SIM"]

python/parts.nix

@@ -44,21 +44,20 @@ let
         paths = builtins.concatMap (p: [ (lib.getBin p) (lib.getLib p) (lib.getDev p) ]) (cudaLibs ++ [ cudaPackages.cuda_nvcc cuda_cccl_compat ]);
       };
       exoOverlay = final: prev: {
-        # Replace workspace exo_rs with Nix-built wheel.
+        # Replace workspace exo_pyo3_bindings with Nix-built wheel.
         # Preserve passthru so mkVirtualEnv can resolve dependency groups.
         # Copy .pyi stub + py.typed marker so basedpyright can find the types.
-        exo-rs = pkgs.stdenv.mkDerivation {
-          pname = "exo-rs";
+        exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
+          pname = "exo-pyo3-bindings";
           version = "0.1.0";
-          src = self'.packages.exo-rs;
+          src = self'.packages.exo_pyo3_bindings;
           # Install from pre-built wheel
           nativeBuildInputs = [ final.pyprojectWheelHook ];
           dontStrip = true;
-          passthru = prev.exo-rs.passthru or { };
-
+          passthru = prev.exo-pyo3-bindings.passthru or { };
           postInstall = ''
-            local siteDir=$out/${final.python.sitePackages}/exo_rs
-            cp ${inputs.self}/rust/exo_rs/exo_rs.pyi $siteDir/
+            local siteDir=$out/${final.python.sitePackages}/exo_pyo3_bindings
+            cp ${inputs.self}/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi $siteDir/
             touch $siteDir/py.typed
           '';
         };
@@ -224,7 +223,7 @@ let
           ++ lib.optionals isDarwin [ pkgs.macmon ];
           passthru = {
             venv = venv name;
-            evenv = ((pythonSet.overrideScope editableOverlay).mkVirtualEnv "${name}-evenv" (members // { exo = (members.exo or [ ]) ++ [ "dev" ]; exo-rs = [ ]; })).overrideAttrs (_: {
+            evenv = ((pythonSet.overrideScope editableOverlay).mkVirtualEnv "${name}-evenv" (members // { exo = (members.exo or [ ]) ++ [ "dev" ]; })).overrideAttrs (_: {
               venvSkip = [ "lib/python${python.pythonVersion}/site-packages/mlx/share/cmake/*" "lib/python${python.pythonVersion}/site-packages/build_backend.py" ];
             });
           } // lib.optionalAttrs cudaSupport {

rust/exo_pyo3_bindings/Cargo.toml

@@ -1,5 +1,5 @@
 [package]
-name = "exo_rs"
+name = "exo_pyo3_bindings"
 version = { workspace = true }
 edition = { workspace = true }
 publish = false
@@ -7,7 +7,7 @@ publish = false
 [lib]
 doctest = false
 path = "src/lib.rs"
-name = "exo_rs"
+name = "exo_pyo3_bindings"
 
 # "cdylib" needed to produce shared library for Python to import
 # "rlib" needed for stub-gen to run
@@ -25,7 +25,7 @@ workspace = true
 networking = { workspace = true }
 
 # interop
-pyo3 = { version = "0.28.3", features = [
+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]
@@ -38,15 +38,15 @@ pyo3 = { version = "0.28.3", features = [
   # "ordered-float", "rust_decimal", "smallvec",
   # "anyhow", "chrono", "chrono-local", "chrono-tz", "eyre", "jiff-02", "lock_api", "parking-lot", "time",  "serde",
 ] }
-pyo3-stub-gen = { version = "0.22.3" }
-pyo3-async-runtimes = { version = "0.28.0", features = [
+pyo3-stub-gen = { version = "0.17.2" }
+pyo3-async-runtimes = { version = "0.27.0", features = [
   "attributes",
   "tokio-runtime",
   "testing",
 ] }
-pyo3-log = "0.13.3"
+pyo3-log = "0.13.2"
 
-pidfile-rs = { git = "https://github.com/AndreiCravtov/pidfile-rs" }
+pidfile-rs = "0.3"
 
 # macro dependencies
 extend = { workspace = true }

rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi

@@ -1,20 +1,10 @@
 # This file is automatically generated by pyo3_stub_gen
-# ruff: noqa: E501, F401, F403, F405
+# ruff: noqa: E501, F401
 
 import builtins
 import os
 import pathlib
 import typing
-__all__ = [
-    "AllQueuesFullError",
-    "FromSwarm",
-    "Keypair",
-    "MessageTooLargeError",
-    "NetworkingHandle",
-    "NoPeersSubscribedToTopicError",
-    "Pidfile",
-    "PidfileError",
-]
 
 @typing.final
 class AllQueuesFullError(builtins.Exception):
@@ -22,29 +12,6 @@ class AllQueuesFullError(builtins.Exception):
     def __repr__(self) -> builtins.str: ...
     def __str__(self) -> builtins.str: ...
 
-class FromSwarm:
-    @typing.final
-    class Connection(FromSwarm):
-        __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) -> FromSwarm.Connection: ...
-    
-    @typing.final
-    class Message(FromSwarm):
-        __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) -> FromSwarm.Message: ...
-    
-    ...
-
 @typing.final
 class Keypair:
     r"""
@@ -78,7 +45,6 @@ class MessageTooLargeError(builtins.Exception):
 @typing.final
 class NetworkingHandle:
     def __new__(cls, identity: Keypair, bootstrap_peers: typing.Sequence[builtins.str], listen_port: builtins.int) -> NetworkingHandle: ...
-    def recv(self) -> typing.Awaitable[FromSwarm]: ...
     async def gossipsub_subscribe(self, topic: builtins.str) -> builtins.bool:
         r"""
         Subscribe to a `GossipSub` topic.
@@ -97,6 +63,7 @@ class NetworkingHandle:
         
         If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
         """
+    async def recv(self) -> PyFromSwarm: ...
 
 @typing.final
 class NoPeersSubscribedToTopicError(builtins.Exception):
@@ -110,7 +77,7 @@ class Pidfile:
     A PID file protected with a lock.
     
     An instance of `Pidfile` can be used to manage a PID file: create it,
-    lock it, detect already running daemons. It is backed by [`pidfile`]
+    lock it, detect already running daemons. It is backed by [`pidfile`][]
     functions of `libbsd`/`libutil` which use `flopen` to lock the PID
     file.
     
@@ -140,23 +107,32 @@ class Pidfile:
         
         The file is truncated before writing.
         """
-    def as_raw_fd(self) -> builtins.int:
-        r"""
-        Extracts the raw file descriptor.
-        
-        This function is typically used to **borrow** an owned file descriptor.
-        When used in this way, this method does **not** pass ownership of the
-        raw file descriptor to the caller, and the file descriptor is only
-        guaranteed to be valid while the original object has not yet been
-        destroyed.
-        """
-    def close(self) -> None:
-        r"""
-        Closes the PID file and releases associated resources.
-        """
 
 @typing.final
 class PidfileError(builtins.Exception):
     def __repr__(self) -> builtins.str: ...
     def __str__(self) -> builtins.str: ...
 
+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

@@ -3,8 +3,8 @@ requires = ["maturin>=1.0,<2.0"]
 build-backend = "maturin"
 
 [project]
-name = "exo_rs"
-version = "0.2.16"
+name = "exo_pyo3_bindings"
+version = "0.2.2"
 description = "Add your description here"
 readme = "README.md"
 authors = [
@@ -15,17 +15,14 @@ requires-python = ">=3.13"
 dependencies = []
 
 [dependency-groups]
-dev = ["exo_rs", "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
 #python-source = "python"
-module-name = "exo_rs"
+module-name = "exo_pyo3_bindings"
 features = ["pyo3/extension-module", "pyo3/experimental-async"]
 
-[tool.pyo3-stub-gen]
-generate-init-py = true
-
 [tool.pytest.ini_options]
 log_cli = true
 log_cli_level = "INFO"

rust/exo_pyo3_bindings/src/bin/stub_gen.rs

@@ -2,7 +2,7 @@ use pyo3_stub_gen::Result;
 
 fn main() -> Result<()> {
     env_logger::Builder::from_env(env_logger::Env::default().filter_or("RUST_LOG", "info")).init();
-    let stub = exo_rs::stub_info()?;
+    let stub = exo_pyo3_bindings::stub_info()?;
     stub.generate()?;
     Ok(())
 }

rust/exo_pyo3_bindings/src/lib.rs

@@ -153,7 +153,7 @@ pub(crate) mod ext {
 /// A Python module implemented in Rust. The name of this function must match
 /// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
 /// import the module.
-#[pymodule(name = "exo_rs", gil_used = true)]
+#[pymodule(name = "exo_pyo3_bindings")]
 fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
     // install logger
     pyo3_log::init();

rust/exo_pyo3_bindings/src/networking.rs

@@ -16,7 +16,9 @@ use pyo3::exceptions::PyRuntimeError;
 use pyo3::prelude::{PyModule, PyModuleMethods as _};
 use pyo3::types::PyBytes;
 use pyo3::{Bound, Py, PyAny, PyErr, PyResult, Python, pymethods};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pyclass_complex_enum, gen_stub_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 {
@@ -136,7 +138,7 @@ struct PyNetworkingHandle {
 }
 
 #[gen_stub_pyclass_complex_enum]
-#[pyclass(name = "FromSwarm")]
+#[pyclass]
 enum PyFromSwarm {
     Connection {
         peer_id: String,
@@ -202,11 +204,9 @@ impl PyNetworkingHandle {
         })
     }
 
-    #[gen_stub(override_return_type(
-        type_repr="typing.Awaitable[FromSwarm]", imports=("typing")
-    ))]
+    #[gen_stub(skip)]
     fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        let swarm = self.swarm.clone();
+        let swarm = Arc::clone(&self.swarm);
         pyo3_async_runtimes::tokio::future_into_py(py, async move {
             swarm
                 .try_lock()
@@ -297,6 +297,15 @@ impl PyNetworkingHandle {
     }
 }
 
+pyo3_stub_gen::inventory::submit! {
+    gen_methods_from_python! {
+        r#"
+            class PyNetworkingHandle:
+                async def recv() -> PyFromSwarm: ...
+        "#
+    }
+}
+
 pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_class::<exception::PyNoPeersSubscribedToTopicError>()?;
     m.add_class::<exception::PyAllQueuesFullError>()?;

rust/exo_pyo3_bindings/src/pidfile.rs

@@ -3,9 +3,7 @@ use pyo3::exceptions::PyException;
 use pyo3::prelude::{PyModule, PyModuleMethods};
 use pyo3::{Bound, PyErr, PyResult, Python, pyclass, pymethods};
 use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
-use std::fs;
 use std::fs::Permissions;
-use std::os::fd::{AsRawFd, RawFd};
 use std::os::unix::prelude::PermissionsExt;
 use std::path::PathBuf;
 
@@ -38,7 +36,7 @@ impl PyPidfileError {
 /// A PID file protected with a lock.
 ///
 /// An instance of `Pidfile` can be used to manage a PID file: create it,
-/// lock it, detect already running daemons. It is backed by [`pidfile`]
+/// lock it, detect already running daemons. It is backed by [`pidfile`][]
 /// functions of `libbsd`/`libutil` which use `flopen` to lock the PID
 /// file.
 ///
@@ -55,23 +53,7 @@ impl PyPidfileError {
 /// [`daemon`(3)]: https://linux.die.net/man/3/daemon
 #[gen_stub_pyclass]
 #[pyclass(name = "Pidfile")]
-pub struct PyPidfile(Option<Pidfile>);
-
-impl PyPidfile {
-    #[inline(always)]
-    fn get(&self) -> &Pidfile {
-        self.0
-            .as_ref()
-            .expect("cannot use resource after exiting context")
-    }
-
-    #[inline(always)]
-    fn get_mut(&mut self) -> &mut Pidfile {
-        self.0
-            .as_mut()
-            .expect("cannot use resource after exiting context")
-    }
-}
+pub struct PyPidfile(Pidfile);
 
 #[gen_stub_pymethods]
 #[pymethods]
@@ -83,40 +65,17 @@ impl PyPidfile {
     /// the PID file yet.
     #[new]
     fn py_new(py: Python, path: PathBuf, mode: u32) -> PyResult<Self> {
-        // create all parent directories if don't exist
-        if let Some(parent) = path.parent() {
-            fs::create_dir_all(parent)
-                .map_err(|e| PyPidfileError(PidfileError::Io(e)).into_pyerr(py))?;
-        }
-
-        let pidfile = Pidfile::new(&path, Permissions::from_mode(mode))
-            .map_err(|e| PyPidfileError(e).into_pyerr(py))?;
-        Ok(Self(Some(pidfile)))
+        Ok(Self(
+            Pidfile::new(&path, Permissions::from_mode(mode))
+                .map_err(|e| PyPidfileError(e).into_pyerr(py))?,
+        ))
     }
 
     /// Writes the current process ID to the PID file.
     ///
     /// The file is truncated before writing.
     fn write<'py>(&mut self, py: Python<'py>) -> PyResult<()> {
-        self.get_mut()
-            .write()
-            .map_err(|e| PyPidfileError(e).into_pyerr(py))
-    }
-
-    /// Extracts the raw file descriptor.
-    ///
-    /// This function is typically used to **borrow** an owned file descriptor.
-    /// When used in this way, this method does **not** pass ownership of the
-    /// raw file descriptor to the caller, and the file descriptor is only
-    /// guaranteed to be valid while the original object has not yet been
-    /// destroyed.
-    fn as_raw_fd(&self) -> RawFd {
-        self.get().as_raw_fd()
-    }
-
-    /// Closes the PID file and releases associated resources.
-    fn close(&mut self) {
-        self.0 = None;
+        self.0.write().map_err(|e| PyPidfileError(e).into_pyerr(py))
     }
 }
 

rust/exo_pyo3_bindings/tests/test_python.py

@@ -2,12 +2,12 @@ import asyncio
 
 import pytest
 from _pytest.capture import CaptureFixture
-from exo_rs import (
+from exo_pyo3_bindings import (
     Keypair,
     NetworkingHandle,
     NoPeersSubscribedToTopicError,
     Pidfile,
-    FromSwarm,
+    PyFromSwarm,
 )
 
 
@@ -39,9 +39,9 @@ async def _await_recv(h: NetworkingHandle):
     while True:
         event = await h.recv()
         match event:
-            case FromSwarm.Connection() as c:
+            case PyFromSwarm.Connection() as c:
                 print(f"PYTHON: connection update: {c}")
-            case FromSwarm.Message() as m:
+            case PyFromSwarm.Message() as m:
                 print(f"PYTHON: message: {m}")
 
 

rust/parts.nix

@@ -55,7 +55,6 @@
         ];
 
         OPENSSL_NO_VENDOR = "1";
-        MATURIN_NO_INSTALL_RUST = "1";
 
         # Required for pyo3 tests to find libpython
         LD_LIBRARY_PATH = lib.makeLibraryPath [ pkgs.python313 ];
@@ -82,11 +81,11 @@
       config = {
         packages = {
           # Python bindings wheel via maturin
-          exo-rs = craneLib.buildPackage (
+          exo_pyo3_bindings = craneLib.buildPackage (
             commonArgs
             // {
               inherit cargoArtifacts;
-              pname = "exo-rs";
+              pname = "exo_pyo3_bindings";
 
               nativeBuildInputs = commonArgs.nativeBuildInputs ++ [
                 pkgs.maturin
@@ -96,7 +95,7 @@
                 maturin build \
                   --release \
                   --manylinux off \
-                  --manifest-path rust/exo_rs/Cargo.toml \
+                  --manifest-path rust/exo_pyo3_bindings/Cargo.toml \
                   --features "pyo3/extension-module,pyo3/experimental-async" \
                   --interpreter ${pkgs.python313}/bin/python \
                   --out dist

src/exo/api/main.py

@@ -20,7 +20,7 @@ from fastapi.staticfiles import StaticFiles
 from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType]
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
-from hypercorn.utils import LifespanTimeoutError, ShutdownError
+from hypercorn.utils import LifespanTimeoutError
 from loguru import logger
 
 from exo.api.adapters.chat_completions import (
@@ -50,8 +50,6 @@ from exo.api.keepalive import with_sse_keepalive
 from exo.api.types import (
     AddCustomModelParams,
     AdvancedImageParams,
-    AwaitInstanceReadyMessage,
-    AwaitInstanceTimeoutMessage,
     BenchChatCompletionRequest,
     BenchChatCompletionResponse,
     BenchImageGenerationResponse,
@@ -346,7 +344,6 @@ class API:
         self.app.post("/place_instance")(self.place_instance)
         self.app.get("/instance/placement")(self.get_placement)
         self.app.get("/instance/previews")(self.get_placement_previews)
-        self.app.get("/instance/await", response_model=None)(self.await_instance)
         self.app.get("/instance/{instance_id}")(self.get_instance)
         self.app.delete("/instance/{instance_id}")(self.delete_instance)
         self.app.get("/v1/instance-links")(self.list_instance_links)
@@ -636,48 +633,6 @@ class API:
             raise HTTPException(status_code=404, detail="Instance not found")
         return self.state.instances[instance_id]
 
-    async def await_instance(
-        self,
-        model_id: ModelId,
-        timeout_seconds: float = Query(default=0.0, ge=0.0, le=300.0),
-    ) -> StreamingResponse:
-        _sleep = 0.1
-
-        async def _stream() -> AsyncGenerator[str, None]:
-            deadline = (
-                None if timeout_seconds == 0 else anyio.current_time() + timeout_seconds
-            )
-
-            while True:
-                for instance in self.state.instances.values():
-                    if instance.shard_assignments.model_id == model_id:
-                        payload = AwaitInstanceReadyMessage(instance=instance)
-                        yield f"data: {payload.model_dump_json()}\n\n"
-                        return
-
-                if deadline is None:
-                    await anyio.sleep(_sleep)
-                else:
-                    remaining = deadline - anyio.current_time()
-                    if remaining <= 0:
-                        payload = AwaitInstanceTimeoutMessage(
-                            message=f"No instance found for model {model_id}"
-                        )
-                        yield f"data: {payload.model_dump_json()}\n\n"
-                        return
-
-                    await anyio.sleep(min(_sleep, remaining))
-
-        return StreamingResponse(
-            with_sse_keepalive(_stream()),
-            media_type="text/event-stream",
-            headers={
-                "Cache-Control": "no-cache",
-                "Connection": "close",
-                "X-Accel-Buffering": "no",
-            },
-        )
-
     async def delete_instance(self, instance_id: InstanceId) -> DeleteInstanceResponse:
         if instance_id not in self.state.instances:
             raise HTTPException(status_code=404, detail="Instance not found")
@@ -806,8 +761,6 @@ class API:
                 if isinstance(chunk, PrefillProgressChunk):
                     continue
 
-                sampler.mark_prefill_done()
-
                 if chunk.finish_reason == "error":
                     raise HTTPException(
                         status_code=500,
@@ -918,8 +871,10 @@ class API:
     ) -> ChatCompletionResponse | StreamingResponse:
         """OpenAI Chat Completions API - adapter."""
         task_params = await chat_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(task_params.model)
-        task_params = task_params.model_copy(update={"model": validated_model})
+        resolved_model = await self._resolve_and_validate_text_model(
+            ModelId(task_params.model)
+        )
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         command = await self._send_text_generation_with_images(task_params)
 
@@ -951,10 +906,10 @@ class API:
         self, payload: BenchChatCompletionRequest
     ) -> BenchChatCompletionResponse | StreamingResponse:
         task_params = await chat_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(
+        resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
-        task_params = task_params.model_copy(update={"model": validated_model})
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         task_params = task_params.model_copy(
             update={
@@ -984,10 +939,8 @@ class API:
 
         return await self._collect_text_generation_with_stats(command.command_id)
 
-    async def _validate_model_has_instance(self, model_id: ModelId) -> ModelId:
-        """Validate a model has an active instance.
-        If the model isn't even downloaded, triggers notification to user to download model.
-
+    async def _resolve_and_validate_text_model(self, model_id: ModelId) -> ModelId:
+        """Validate a text model exists and return the resolved model ID.
 
         Raises HTTPException 404 if no instance is found for the model.
         """
@@ -995,21 +948,30 @@ class API:
             instance.shard_assignments.model_id == model_id
             for instance in self.state.instances.values()
         ):
-            # Check if model is actually downloaded
-            model_is_downloaded = any(
-                isinstance(download, DownloadCompleted)
-                and download.shard_metadata.model_card.model_id == model_id
-                for node_downloads in self.state.downloads.values()
-                for download in node_downloads
-            )
-            if not model_is_downloaded:
-                await self._trigger_notify_user_to_download_model(model_id)
-
+            await self._trigger_notify_user_to_download_model(model_id)
             raise HTTPException(
-                status_code=404, detail=f"No instance found for model {model_id}"
+                status_code=404,
+                detail=f"No instance found for model {model_id}",
             )
         return model_id
 
+    async def _validate_image_model(self, model: ModelId) -> ModelId:
+        """Validate model exists and return resolved model ID.
+
+        Raises HTTPException 404 if no instance is found for the model.
+        """
+        model_card = await ModelCard.load(model)
+        resolved_model = model_card.model_id
+        if not any(
+            instance.shard_assignments.model_id == resolved_model
+            for instance in self.state.instances.values()
+        ):
+            await self._trigger_notify_user_to_download_model(resolved_model)
+            raise HTTPException(
+                status_code=404, detail=f"No instance found for model {resolved_model}"
+            )
+        return resolved_model
+
     def stream_events(self) -> StreamingResponse:
         def _generate_json_array(events: Iterable[Event]) -> Iterable[str]:
             yield "["
@@ -1062,9 +1024,7 @@ class API:
         """
         payload = payload.model_copy(
             update={
-                "model": await self._validate_model_has_instance(
-                    ModelId(payload.model)
-                ),
+                "model": await self._validate_image_model(ModelId(payload.model)),
                 "advanced_params": _ensure_seed(payload.advanced_params),
             }
         )
@@ -1332,9 +1292,7 @@ class API:
     ) -> BenchImageGenerationResponse:
         payload = payload.model_copy(
             update={
-                "model": await self._validate_model_has_instance(
-                    ModelId(payload.model)
-                ),
+                "model": await self._validate_image_model(ModelId(payload.model)),
                 "stream": False,
                 "partial_images": 0,
                 "advanced_params": _ensure_seed(payload.advanced_params),
@@ -1370,7 +1328,7 @@ class API:
         advanced_params: AdvancedImageParams | None,
     ) -> ImageEdits:
         """Prepare and send an image edits command with chunked image upload."""
-        validated_model = await self._validate_model_has_instance(model)
+        resolved_model = await self._validate_image_model(model)
         advanced_params = _ensure_seed(advanced_params)
 
         image_content = await image.read()
@@ -1389,7 +1347,7 @@ class API:
                 image_data="",
                 total_input_chunks=total_chunks,
                 prompt=prompt,
-                model=validated_model,
+                model=resolved_model,
                 n=n,
                 size=size,
                 response_format=response_format,
@@ -1410,7 +1368,7 @@ class API:
             await self._send(
                 SendInputChunk(
                     chunk=InputImageChunk(
-                        model=validated_model,
+                        model=resolved_model,
                         command_id=command.command_id,
                         data=chunk_data,
                         chunk_index=chunk_index,
@@ -1534,10 +1492,10 @@ class API:
     ) -> ClaudeMessagesResponse | StreamingResponse:
         """Claude Messages API - adapter."""
         task_params = await claude_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(
+        resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
-        task_params = task_params.model_copy(update={"model": validated_model})
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         command = await self._send_text_generation_with_images(task_params)
 
@@ -1572,8 +1530,8 @@ class API:
     ) -> ResponsesResponse | StreamingResponse:
         """OpenAI Responses API."""
         task_params = await responses_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(task_params.model)
-        task_params = task_params.model_copy(update={"model": validated_model})
+        resolved_model = await self._resolve_and_validate_text_model(task_params.model)
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         command = await self._send_text_generation_with_images(task_params)
 
@@ -1615,10 +1573,10 @@ class API:
         body = await request.body()
         payload = OllamaChatRequest.model_validate_json(body)
         task_params = ollama_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(
+        resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
-        task_params = task_params.model_copy(update={"model": validated_model})
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         command = await self._send_text_generation_with_images(task_params)
 
@@ -1651,10 +1609,10 @@ class API:
         body = await request.body()
         payload = OllamaGenerateRequest.model_validate_json(body)
         task_params = ollama_generate_request_to_text_generation(payload)
-        validated_model = await self._validate_model_has_instance(
+        resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
-        task_params = task_params.model_copy(update={"model": validated_model})
+        task_params = task_params.model_copy(update={"model": resolved_model})
 
         command = await self._send_text_generation_with_images(task_params)
 
@@ -1956,10 +1914,6 @@ class API:
                     cfg,
                     shutdown_trigger=ev.wait,
                 )
-                if not ev.is_set():
-                    raise ShutdownError(
-                        "Server exited without shutdown trigger - exiting abnormally"
-                    )
             except LifespanTimeoutError as e:
                 logger.warning(
                     "Graceful server shutdown timed out, some connections forcebly closed"

src/exo/api/types/__init__.py

@@ -1,7 +1,5 @@
 from .api import AddCustomModelParams as AddCustomModelParams
 from .api import AdvancedImageParams as AdvancedImageParams
-from .api import AwaitInstanceReadyMessage as AwaitInstanceReadyMessage
-from .api import AwaitInstanceTimeoutMessage as AwaitInstanceTimeoutMessage
 from .api import BenchChatCompletionRequest as BenchChatCompletionRequest
 from .api import BenchChatCompletionResponse as BenchChatCompletionResponse
 from .api import BenchImageGenerationResponse as BenchImageGenerationResponse

src/exo/api/types/api.py

@@ -186,12 +186,6 @@ class NodePowerStats(BaseModel, frozen=True):
     node_id: NodeId
     samples: int
     avg_sys_power: float
-    # Per-phase breakdown. Populated only when the caller marks a phase
-    # boundary (e.g. prefill -> generation); None otherwise.
-    prefill_avg_sys_power: float | None = None
-    generation_avg_sys_power: float | None = None
-    prefill_energy_joules: float | None = None
-    generation_energy_joules: float | None = None
 
 
 class PowerUsage(BaseModel, frozen=True):
@@ -199,16 +193,6 @@ class PowerUsage(BaseModel, frozen=True):
     nodes: list[NodePowerStats]
     total_avg_sys_power_watts: float
     total_energy_joules: float
-    # Split between the prefill (prompt-processing) phase and the
-    # generation/decode phase. Populated only when the caller marks a phase
-    # boundary; None otherwise. The two phase energies should sum to
-    # approximately `total_energy_joules` (modulo interpolation rounding).
-    prefill_seconds: float | None = None
-    generation_seconds: float | None = None
-    prefill_energy_joules: float | None = None
-    generation_energy_joules: float | None = None
-    prefill_avg_sys_power_watts: float | None = None
-    generation_avg_sys_power_watts: float | None = None
 
 
 class BenchChatCompletionResponse(ChatCompletionResponse):
@@ -307,16 +291,6 @@ class DeleteInstanceResponse(BaseModel):
     instance_id: InstanceId
 
 
-class AwaitInstanceReadyMessage(BaseModel):
-    type: Literal["ready"] = "ready"
-    instance: Instance
-
-
-class AwaitInstanceTimeoutMessage(BaseModel):
-    type: Literal["timeout"] = "timeout"
-    message: str
-
-
 class CancelCommandResponse(BaseModel):
     message: str
     command_id: CommandId

src/exo/download/coordinator.py

@@ -15,10 +15,6 @@ from exo.download.download_utils import (
     resolve_existing_model,
 )
 from exo.download.shard_downloader import ShardDownloader
-from exo.routing.event_router import (
-    EventRouterBrokenResourceError,
-    EventRouterClosedResourceError,
-)
 from exo.shared.constants import EXO_DEFAULT_MODELS_DIR, EXO_MODELS_READ_ONLY_DIRS
 from exo.shared.models import model_cards
 from exo.shared.models.model_cards import ModelId
@@ -143,14 +139,7 @@ class DownloadCoordinator:
             async with self._tg as tg:
                 tg.start_soon(self._command_processor)
                 tg.start_soon(self._emit_existing_download_progress)
-        except* (EventRouterBrokenResourceError, EventRouterClosedResourceError):
-            # Event router has been closed (try-star syntax handles error groups)
-            pass
         finally:
-            # don't forget to clean up resources
-            self.download_command_receiver.close()
-            self.event_sender.close()
-
             self._stopped.set()
 
     async def shutdown(self) -> None:

src/exo/main.py

@@ -8,9 +8,6 @@ from dataclasses import dataclass, field
 from typing import Self
 
 import anyio
-from anyio.lowlevel import checkpoint as anyio_checkpoint
-from daemon import DaemonContext  # pyright: ignore[reportMissingTypeStubs]
-from exo_rs import Pidfile, PidfileError
 from loguru import logger
 from pydantic import PositiveInt
 
@@ -21,12 +18,13 @@ from exo.download.impl_shard_downloader import exo_shard_downloader
 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_DEFAULT_MODELS_DIR, EXO_LOG, EXO_PID_FILE
+from exo.shared.constants import EXO_DEFAULT_MODELS_DIR, EXO_LOG
 from exo.shared.election import Election, ElectionResult
 from exo.shared.logging import logger_cleanup, logger_setup
 from exo.shared.types.common import NodeId, SessionId
-from exo.utils import STDIO_FDS
 from exo.utils.channels import Receiver, channel
+from exo.utils.daemon import detach_stdio_to_devnull
+from exo.utils.pidfile import PidfileLockError, acquire_exo_pidfile
 from exo.utils.pydantic_ext import FrozenModel
 from exo.utils.task_group import TaskGroup
 from exo.worker.main import Worker
@@ -192,7 +190,7 @@ class Node:
                 # - Shut down and re-create the API
 
                 if result.is_new_master:
-                    await anyio_checkpoint()
+                    await anyio.sleep(0)
                     self.event_router.shutdown()
                     self.event_router = EventRouter(
                         result.session_id,
@@ -205,10 +203,7 @@ class Node:
                     result.session_id.master_node_id == self.node_id
                     and self.master is not None
                 ):
-                    assert not result.is_new_master, (
-                        "cannot be new master if we remain master"
-                    )
-                    logger.info("Node elected Master - maintaining self")
+                    logger.info("Node elected Master")
                 elif (
                     result.session_id.master_node_id == self.node_id
                     and self.master is None
@@ -275,60 +270,14 @@ class Node:
 
 
 def main():
-    # Parse args first => --help or bad args don't require PID-locking
+    # Exit early if no PID file (not compatible with double-for daemonization yet)
+    try:
+        pidfile = acquire_exo_pidfile()
+    except PidfileLockError as exception:
+        print(exception, file=sys.stderr)
+        raise SystemExit(1) from exception
+
     args = Args.parse()
-
-    # Exit early if cannot acquire PID file
-    try:
-        pidfile = Pidfile(EXO_PID_FILE, 0o0600)
-    except PidfileError as e:
-        print(e, file=sys.stderr)
-        raise SystemExit(1) from e
-
-    try:
-        if args.legacy_daemon:
-            # keep stdio backed by explicit /dev/null streams. multiprocessing spawn expects
-            # valid stdio FDs; letting DaemonContext close/reopen them can break runner startup.
-            for stream in (sys.stdout, sys.stderr, sys.__stdout__, sys.__stderr__):
-                if stream is not None:
-                    stream.flush()
-            stdin = open(os.devnull, "r")  # noqa: SIM115
-            stdout = open(os.devnull, "w")  # noqa: SIM115
-            stderr = open(os.devnull, "w")  # noqa: SIM115
-
-            with DaemonContext(
-                detach_process=True,
-                files_preserve=[pidfile.as_raw_fd()],
-                stdin=stdin,
-                stdout=stdout,
-                stderr=stderr,
-            ):
-                # cleanup loose file descriptors (as long as they aren't stdio)
-                for f in (
-                    f for f in (stdin, stdout, stderr) if f.fileno() not in STDIO_FDS
-                ):
-                    f.close()
-
-                # 1) if daemonizing => fork then write PID
-                try:
-                    pidfile.write()
-                except PidfileError as e:
-                    print(e, file=sys.stderr)
-                    raise SystemExit(1) from e
-                main_inner(args)
-        else:
-            # 2) otherwise      => just write PID
-            try:
-                pidfile.write()
-            except PidfileError as e:
-                print(e, file=sys.stderr)
-                raise SystemExit(1) from e
-            main_inner(args)
-    finally:
-        pidfile.close()
-
-
-def main_inner(args: "Args"):
     soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
     target = min(max(soft, 65535), hard)
     resource.setrlimit(resource.RLIMIT_NOFILE, (target, hard))
@@ -337,6 +286,9 @@ def main_inner(args: "Args"):
 
     # TODO: Refactor the current verbosity system
     logger_setup(EXO_LOG, args.verbosity)
+    if args.no_stdio:
+        detach_stdio_to_devnull()
+        logger.info("Detached stdio to /dev/null")
 
     logger.info(f"{'=' * 40}")
     logger.info(f"Starting EXO | pid={os.getpid()}")
@@ -372,6 +324,7 @@ def main_inner(args: "Args"):
     finally:
         logger.info("EXO Shutdown complete")
         logger_cleanup()
+        del pidfile
 
 
 class Args(FrozenModel):
@@ -385,7 +338,7 @@ class Args(FrozenModel):
     offline: bool = os.getenv("EXO_OFFLINE", "false").lower() == "true"
     no_batch: bool = False
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
-    legacy_daemon: bool = False
+    no_stdio: bool = False
     bootstrap_peers: list[str] = []
     libp2p_port: int
 
@@ -446,9 +399,9 @@ class Args(FrozenModel):
             help="Disable continuous batching, use sequential generation",
         )
         parser.add_argument(
-            "--legacy-daemon",
+            "--no-stdio",
             action="store_true",
-            help="Run as a legacy SysV-style background daemon using double-fork daemonization",
+            help="Detach stdin/stdout/stderr to /dev/null after logging is configured",
         )
         parser.add_argument(
             "--bootstrap-peers",

src/exo/master/main.py

@@ -11,10 +11,6 @@ from exo.master.placement import (
     place_instance,
 )
 from exo.master.placement_utils import find_ip_prioritised
-from exo.routing.event_router import (
-    EventRouterBrokenResourceError,
-    EventRouterClosedResourceError,
-)
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
@@ -155,9 +151,6 @@ class Master:
                 tg.start_soon(self._event_processor)
                 tg.start_soon(self._command_processor)
                 tg.start_soon(self._plan)
-        except* (EventRouterBrokenResourceError, EventRouterClosedResourceError):
-            # Event router has been closed (try-star syntax handles error groups)
-            pass
         finally:
             self._event_log.close()
             self.global_event_sender.close()
@@ -181,7 +174,6 @@ class Master:
                         case TestCommand():
                             pass
                         case TextGeneration():
-                            # set-difference => prefill-only nodes
                             prefill_only: set[InstanceId] = set()
                             for link in self.state.instance_links.values():
                                 prefill_only.update(link.prefill_instances)
@@ -189,13 +181,11 @@ class Master:
                                 prefill_only.difference_update(link.decode_instances)
 
                             for instance in self.state.instances.values():
-                                # NON-prefill-only instances matching the model ID
                                 if (
                                     instance.shard_assignments.model_id
                                     == command.task_params.model
                                     and instance.instance_id not in prefill_only
                                 ):
-                                    # count in-flight tasks of that instance
                                     in_flight = {TaskStatus.Pending, TaskStatus.Running}
                                     task_count = sum(
                                         1
@@ -207,7 +197,6 @@ class Master:
                                         task_count
                                     )
 
-                            # there are no NON-prefill-only instances matching this model ID
                             if not instance_task_counts:
                                 raise ValueError(
                                     f"No instance found for model {command.task_params.model}"
@@ -459,9 +448,7 @@ class Master:
                                 self._event_log.read_range(command.since_idx, end),
                                 start=command.since_idx,
                             ):
-                                await self._send_indexed_event(
-                                    IndexedEvent(idx=i, event=event)
-                                )
+                                await self._send_event(IndexedEvent(idx=i, event=event))
                     for event in generated_events:
                         await self.event_sender.send(event)
                 except ValueError as e:
@@ -519,10 +506,10 @@ class Master:
                     self.state = apply(self.state, indexed)
 
                     self._event_log.append(event)
-                    await self._send_indexed_event(indexed)
+                    await self._send_event(indexed)
 
     # This function is re-entrant, take care!
-    async def _send_indexed_event(self, event: IndexedEvent):
+    async def _send_event(self, event: IndexedEvent):
         # Convenience method since this line is ugly
         await self.global_event_sender.send(
             GlobalForwarderEvent(

src/exo/routing/connection_message.py

@@ -1,4 +1,4 @@
-from exo_rs import FromSwarm
+from exo_pyo3_bindings import PyFromSwarm
 
 from exo.shared.types.common import NodeId
 from exo.utils.pydantic_ext import FrozenModel
@@ -11,5 +11,5 @@ class ConnectionMessage(FrozenModel):
     connected: bool
 
     @classmethod
-    def from_update(cls, update: FromSwarm.Connection) -> "ConnectionMessage":
+    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

@@ -15,30 +15,11 @@ from exo.shared.types.events import (
     IndexedEvent,
     LocalForwarderEvent,
 )
-from exo.utils import channels
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import OrderedBuffer
 from exo.utils.task_group import TaskGroup
 
 
-class EventRouterClosedResourceError(ClosedResourceError):
-    pass
-
-
-class EventRouterBrokenResourceError(BrokenResourceError):
-    pass
-
-
-# Event Router is created and destroyed before consumers of its channels are,
-# hence its nice to have tagged errors for event-router channels being closed
-#
-# so consumers can catch specifically these errors, rather than the generic ones
-_ERROR_CFG = channels.ErrorOverride(
-    closed_resource_error=EventRouterClosedResourceError,
-    broken_resource_error=EventRouterBrokenResourceError,
-)
-
-
 @dataclass
 class EventRouter:
     session_id: SessionId
@@ -83,7 +64,7 @@ class EventRouter:
                     await self.external_outbound.send(event)
 
     def sender(self) -> Sender[Event]:
-        send, recv = channel[Event](error_override_config=_ERROR_CFG)
+        send, recv = channel[Event]()
         if self._tg.is_running():
             self._tg.start_soon(self._ingest, SystemId(), recv)
         else:
@@ -92,7 +73,7 @@ class EventRouter:
 
     def receiver(self) -> Receiver[IndexedEvent]:
         assert not self._tg.is_running()
-        send, recv = channel[IndexedEvent](error_override_config=_ERROR_CFG)
+        send, recv = channel[IndexedEvent]()
         self.internal_outbound.append(send)
         return recv
 

src/exo/routing/router.py

@@ -12,13 +12,13 @@ from anyio import (
     move_on_after,
     sleep_forever,
 )
-from exo_rs import (
+from exo_pyo3_bindings import (
     AllQueuesFullError,
-    FromSwarm,
     Keypair,
     MessageTooLargeError,
     NetworkingHandle,
     NoPeersSubscribedToTopicError,
+    PyFromSwarm,
 )
 from filelock import FileLock
 from loguru import logger
@@ -191,7 +191,7 @@ class Router:
                 from_swarm = await self._net.recv()
                 logger.debug(from_swarm)
                 match from_swarm:
-                    case FromSwarm.Message(origin, topic, data):
+                    case PyFromSwarm.Message(origin, topic, data):
                         logger.trace(
                             f"Received message on {topic} from {origin} with payload {data}"
                         )
@@ -202,7 +202,7 @@ class Router:
                             continue
                         router = self.topic_routers[topic]
                         await router.publish_bytes(data)
-                    case FromSwarm.Connection():
+                    case PyFromSwarm.Connection():
                         message = ConnectionMessage.from_update(from_swarm)
                         logger.trace(
                             f"Received message on connection_messages with payload {message}"

src/exo/shared/logging.py

@@ -46,7 +46,7 @@ class _InterceptHandler(logging.Handler):
 def logger_setup(log_file: Path | None, verbosity: int = 0):
     """Set up logging for this process - formatting, file handles, verbosity and output"""
 
-    logging.getLogger("exo_rs").setLevel(logging.WARNING)
+    logging.getLogger("exo_pyo3_bindings").setLevel(logging.WARNING)
     logging.getLogger("httpx").setLevel(logging.WARNING)
     logging.getLogger("httpcore").setLevel(logging.WARNING)
 

src/exo/shared/models/model_cards.py

@@ -81,10 +81,8 @@ class _CardCache:
                     card = card.model_copy(update={"is_custom": True})
                 if self.get(card.model_id) is None:
                     self.cc[card.model_id] = card
-            except (ValidationError, TOMLKitError) as e:
-                logger.opt(exception=e).warning(
-                    f"failed to validate model card at {toml_file}"
-                )
+            except (ValidationError, TOMLKitError):
+                pass
 
     async def refresh(self) -> None:
         for path in _BUILTIN_CARD_DIRS:
@@ -199,11 +197,6 @@ class ModelCard(FrozenModel):
     def _validate_tasks(cls, v: list[str | ModelTask]) -> list[ModelTask]:
         return [item if isinstance(item, ModelTask) else ModelTask(item) for item in v]
 
-    @field_validator("backends", mode="before")
-    @classmethod
-    def _validate_backends(cls, v: list[str | Backend]) -> list[Backend]:
-        return [item if isinstance(item, Backend) else Backend(item) for item in v]
-
     async def save(self, path: Path) -> None:
         async with await open_file(path, "w") as f:
             py = self.model_dump(exclude_none=True, exclude={"is_custom"})

src/exo/shared/types/chunks.py

@@ -13,6 +13,7 @@ from exo.shared.models.model_cards import ModelId
 from exo.utils.pydantic_ext import TaggedModel
 from exo.worker.runner.diagnostics import KnownRunnerDiagnostic
 
+from ...worker.runner.diagnostics import KnownRunnerDiagnostic
 from .common import CommandId
 
 

src/exo/utils/__init__.py

@@ -2,11 +2,6 @@ from typing import Any, Type
 
 from .phantom import PhantomData
 
-STDIN_FD = 0
-STDOUT_FD = 1
-STDERR_FD = 2
-STDIO_FDS = (STDIN_FD, STDOUT_FD, STDERR_FD)
-
 
 def ensure_type[T](obj: Any, expected_type: Type[T]) -> T:  # type: ignore
     if not isinstance(obj, expected_type):

src/exo/utils/async_process.py

@@ -25,9 +25,10 @@ from anyio import (
 from anyio.abc import TaskStatus
 from loguru import logger
 
-from exo.utils import STDERR_FD, STDIO_FDS, STDOUT_FD
 from exo.utils.channels import Receiver, Sender, channel
 
+_STDOUT_FD = 1
+_STDERR_FD = 2
 _READ_CHUNK_SIZE = 64 * 1024
 _JOIN_GRACE_SECONDS = 3.0
 _TERMINATE_GRACE_SECONDS = 5.0
@@ -255,11 +256,11 @@ def _run_with_captured_stdio(
     stderr_fd = stderr.detach()
 
     try:
-        os.dup2(stdout_fd, STDOUT_FD)
-        os.dup2(stderr_fd, STDERR_FD)
+        os.dup2(stdout_fd, _STDOUT_FD)
+        os.dup2(stderr_fd, _STDERR_FD)
     finally:
         for fd in (stdout_fd, stderr_fd):
-            if fd not in STDIO_FDS:
+            if fd not in (_STDOUT_FD, _STDERR_FD):
                 _close_fd(fd)
 
     faulthandler.enable(file=sys.stderr, all_threads=True)

src/exo/utils/channels.py

@@ -1,16 +1,13 @@
 import contextlib
 import multiprocessing as mp
 from dataclasses import dataclass, field
-from functools import wraps
-from inspect import iscoroutinefunction
 from math import inf
 from multiprocessing.synchronize import Event
 from queue import Empty, Full
-from types import CoroutineType, TracebackType
-from typing import Any, Callable, NoReturn, Self, cast, overload, override
+from types import TracebackType
+from typing import Any, Self
 
 from anyio import (
-    BrokenResourceError,
     CapacityLimiter,
     ClosedResourceError,
     EndOfStream,
@@ -23,172 +20,35 @@ from anyio.streams.memory import (
 from anyio.streams.memory import (
     MemoryObjectSendStream as AnyioSender,
 )
-from anyio.streams.memory import (
-    MemoryObjectStreamState,
-)
 from anyio.streams.memory import (
     MemoryObjectStreamState as AnyioState,
 )
 
 
-@dataclass(eq=False)
-class ErrorOverride:
-    closed_resource_error: type[ClosedResourceError] = field(
-        default=ClosedResourceError,
-    )
-    broken_resource_error: type[BrokenResourceError] = field(
-        default=BrokenResourceError,
-    )
-    end_of_stream: type[EndOfStream] = field(
-        default=EndOfStream,
-    )
-    would_block: type[WouldBlock] = field(
-        default=WouldBlock,
-    )
-
-    @overload
-    def patch[**P, R](
-        self,
-        fn: Callable[P, CoroutineType[Any, Any, R]],
-        /,
-    ) -> Callable[P, CoroutineType[Any, Any, R]]: ...
-
-    @overload
-    def patch[**P, R](
-        self,
-        fn: Callable[P, R],
-        /,
-    ) -> Callable[P, R]: ...
-
-    def patch[**P, R](self, fn: Callable[P, Any], /) -> Callable[P, Any]:
-        """
-        Returns a function with all these exceptions replaced by their overrides
-        """
-
-        if iscoroutinefunction(fn):
-            async_fn = cast(Callable[P, CoroutineType[Any, Any, R]], fn)
-
-            @wraps(async_fn)
-            async def async_wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
-                try:
-                    return await async_fn(*args, **kwargs)
-                except ClosedResourceError as e:
-                    self._raise_replace(self.closed_resource_error, e)
-                except BrokenResourceError as e:
-                    self._raise_replace(self.broken_resource_error, e)
-                except EndOfStream as e:
-                    self._raise_replace(self.end_of_stream, e)
-                except WouldBlock as e:
-                    self._raise_replace(self.would_block, e)
-
-            return async_wrapper
-        else:
-            sync_fn = cast(Callable[P, R], fn)
-
-            @wraps(sync_fn)
-            def sync_wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
-                try:
-                    return sync_fn(*args, **kwargs)
-                except ClosedResourceError as e:
-                    self._raise_replace(self.closed_resource_error, e)
-                except BrokenResourceError as e:
-                    self._raise_replace(self.broken_resource_error, e)
-                except EndOfStream as e:
-                    self._raise_replace(self.end_of_stream, e)
-                except WouldBlock as e:
-                    self._raise_replace(self.would_block, e)
-
-            return sync_wrapper
-
-    @staticmethod
-    def _raise_replace(replacement: type[BaseException], e: BaseException) -> NoReturn:
-        if isinstance(e, replacement):
-            raise
-        raise replacement() from e
-
-
 class Sender[T](AnyioSender[T]):
-    def __init__(
-        self,
-        state: MemoryObjectStreamState[T],
-        error_override_config: ErrorOverride | None,
-    ):
-        super().__init__(_state=state)
-
-        # patch the methods we want to override errors for
-        #
-        # NOTE: it is very important that new methods which are added,
-        #       and which can throw, are patched in this block
-        if (e := error_override_config) is not None:
-            # new methods of this class
-            self.clone_receiver = e.patch(self.clone_receiver)
-
-            # overridden methods
-            self.clone = e.patch(self.clone)
-
-            # parent methods
-            self.send_nowait = e.patch(self.send_nowait)
-            self.send = e.patch(self.send)
-            self.close = e.patch(self.close)
-            self.aclose = e.patch(self.aclose)
-            self.statistics = e.patch(self.statistics)
-
-        self.err_config = error_override_config
-
-    @override
     def clone(self) -> "Sender[T]":
         if self._closed:
             raise ClosedResourceError
-        return Sender(self._state, self.err_config)
+        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(self._state, self.err_config)
+        return Receiver(_state=self._state)
 
 
 class Receiver[T](AnyioReceiver[T]):
-    def __init__(
-        self,
-        state: MemoryObjectStreamState[T],
-        error_override_config: ErrorOverride | None,
-    ):
-        super().__init__(_state=state)
-
-        # patch the methods we want to override errors for
-        #
-        # NOTE: it is very important that new methods which are added,
-        #       and which can throw, are patched in this block
-        if (e := error_override_config) is not None:
-            # new methods of this class
-            self.clone_sender = e.patch(self.clone_sender)
-            self.collect = e.patch(self.collect)
-            self.receive_at_least = e.patch(self.receive_at_least)
-
-            # overridden methods
-            self.clone = e.patch(self.clone)
-
-            # parent methods
-            self.receive_nowait = e.patch(self.receive_nowait)
-            self.receive = e.patch(self.receive)
-            self.close = e.patch(self.close)
-            self.aclose = e.patch(self.aclose)
-            self.statistics = e.patch(self.statistics)
-
-        self.err_config = error_override_config
-
-    @override
     def clone(self) -> "Receiver[T]":
         if self._closed:
             raise ClosedResourceError
-        return Receiver(self._state, self.err_config)
+        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 sender"""
         if self._closed:
             raise ClosedResourceError
-        return Sender(self._state, self.err_config)
+        return Sender(_state=self._state)
 
     def collect(self) -> list[T]:
         """Collect all currently available items from this receiver"""
@@ -210,7 +70,6 @@ class Receiver[T](AnyioReceiver[T]):
             out.extend(self.collect())
         return out
 
-    @override
     def __enter__(self) -> Self:
         return self
 
@@ -426,17 +285,11 @@ class MpReceiver[T]:
 class channel[T]:  # noqa: N801
     """Create a pair of asynchronous channels for communicating within the same process"""
 
-    def __new__(
-        cls,
-        max_buffer_size: float = inf,
-        error_override_config: ErrorOverride | None = None,
-    ) -> tuple[Sender[T], Receiver[T]]:
+    def __new__(cls, max_buffer_size: float = inf) -> tuple[Sender[T], Receiver[T]]:
         if max_buffer_size != inf and not isinstance(max_buffer_size, int):
             raise ValueError("max_buffer_size must be either an integer or math.inf")
         state = AnyioState[T](max_buffer_size)
-        return Sender(state, error_override_config), Receiver(
-            state, error_override_config
-        )
+        return Sender(_state=state), Receiver(_state=state)
 
 
 class mp_channel[T]:  # noqa: N801

src/exo/utils/daemon.py

@@ -0,0 +1,28 @@
+import os
+import sys
+
+_STDIN_FD = 0
+_STDOUT_FD = 1
+_STDERR_FD = 2
+
+
+def detach_stdio_to_devnull() -> None:
+    """Redirect process stdio file descriptors to /dev/null."""
+
+    for stream in (sys.stdout, sys.stderr, sys.__stdout__, sys.__stderr__):
+        if stream is not None:
+            stream.flush()
+
+    stdin_fd = os.open(os.devnull, os.O_RDONLY)
+    stdout_fd = os.open(os.devnull, os.O_WRONLY)
+    stderr_fd = os.open(os.devnull, os.O_WRONLY)
+
+    try:
+        # dup2 closes the target fd first, but leaves the source fd open.
+        os.dup2(stdin_fd, _STDIN_FD)
+        os.dup2(stdout_fd, _STDOUT_FD)
+        os.dup2(stderr_fd, _STDERR_FD)
+    finally:
+        for fd in (stdin_fd, stdout_fd, stderr_fd):
+            if fd not in (_STDIN_FD, _STDOUT_FD, _STDERR_FD):
+                os.close(fd)

src/exo/utils/info_gatherer/info_gatherer.py

@@ -630,14 +630,6 @@ class InfoGatherer:
                     f"MacMon failed with return code {e.returncode}: {stderr_msg}"
                 )
                 self._tg.start_soon(self._monitor_memory_usage, 1)
-            except ProcessLookupError:
-                # usually throws by the process' context manager on exit
-                # when we ctrl+c, hence usually should be ignored;
-                # if anything else throws it, we explicitly don't care:
-                # process is dead anyways ;)
-                logger.warning(
-                    "Macmon process not found - shutting down macmon monitor"
-                )
             except Exception as e:
                 logger.opt(exception=e).warning("Error in macmon monitor")
                 self._tg.start_soon(self._monitor_memory_usage, 1)

src/exo/utils/info_gatherer/macmon.py

@@ -1,3 +1,6 @@
+import os
+import shutil
+import subprocess
 from typing import Self
 
 from pydantic import BaseModel
@@ -68,3 +71,41 @@ class MacmonMetrics(TaggedModel):
     @classmethod
     def from_raw_json(cls, json: str) -> Self:
         return cls.from_raw(RawMacmonMetrics.model_validate_json(json))
+
+
+def read_macmon_metrics_once(
+    macmon_path: str | None = None,
+    *,
+    timeout: float = 5,
+) -> MacmonMetrics | None:
+    """
+    Read a single macmon sample, returning None when macmon is unavailable.
+    """
+    resolved_macmon_path = (
+        macmon_path or os.getenv("EXO_MACMON_PATH") or shutil.which("macmon")
+    )
+    if resolved_macmon_path is None:
+        return None
+
+    try:
+        result = subprocess.run(
+            [resolved_macmon_path, "pipe", "--samples", "1", "--interval", "100"],
+            capture_output=True,
+            check=False,
+            text=True,
+            timeout=timeout,
+        )
+    except (OSError, subprocess.SubprocessError):
+        return None
+
+    if result.returncode != 0:
+        return None
+
+    lines = result.stdout.strip().splitlines()
+    if not lines:
+        return None
+
+    try:
+        return MacmonMetrics.from_raw_json(lines[0])
+    except ValueError:
+        return None

src/exo/utils/pidfile.py

@@ -0,0 +1,28 @@
+from __future__ import annotations
+
+import os
+from typing import Final
+
+from exo_pyo3_bindings import Pidfile, PidfileError
+
+from exo.shared.constants import EXO_PID_FILE
+
+_PIDFILE_MODE: Final = 0o600
+
+
+class PidfileLockError(RuntimeError):
+    pass
+
+
+def acquire_exo_pidfile() -> Pidfile:
+    path = EXO_PID_FILE
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    try:
+        pidfile = Pidfile(path, _PIDFILE_MODE)
+        pidfile.write()
+    except (OSError, PidfileError) as exception:
+        raise PidfileLockError(
+            f"Failed to acquire EXO pidfile at {path}: {exception}"
+        ) from exception
+
+    return pidfile

src/exo/utils/power_sampler.py

@@ -24,7 +24,6 @@ class PowerSampler:
         ] = defaultdict(list)
         self._start_time: float | None = None
         self._stopped = False
-        self._prefill_done_at: float | None = None
 
     def _take_sample(self, t_rel: float | None = None) -> None:
         assert self._start_time is not None
@@ -39,35 +38,14 @@ class PowerSampler:
             await anyio.sleep(self._interval)
             self._take_sample()
 
-    def mark_prefill_done(self) -> None:
-        """Anchor the prefill→generation boundary on a fresh sample.
-        Idempotent. Safe to call before `run()`; boundary then lands at t=0.
-        """
-        if self._prefill_done_at is not None:
-            return
-        if self._start_time is None:
-            self._prefill_done_at = 0.0
-            return
-        t_rel = time.perf_counter() - self._start_time
-        self._take_sample(t_rel=t_rel)
-        self._prefill_done_at = t_rel
-
     def result(self) -> PowerUsage:
         self._stopped = True
         assert self._start_time is not None, "result() called before run()"
         elapsed = time.perf_counter() - self._start_time
         self._take_sample(t_rel=elapsed)
 
-        # Clamp the split point to [0, elapsed] in case timing is weird (e.g.
-        # mark called after result, or sampler ran for < the prefill window).
-        split = self._prefill_done_at
-        if split is not None:
-            split = max(0.0, min(elapsed, split))
-
         node_stats: list[NodePowerStats] = []
         total_energy_j = 0.0
-        total_prefill_energy_j = 0.0
-        total_generation_energy_j = 0.0
         for node_id, ts_profiles in self._samples.items():
             n = len(ts_profiles)
             if n == 0:
@@ -75,68 +53,20 @@ class PowerSampler:
             node_energy_j = trapezoidal_energy(ts_profiles, elapsed)
             avg_power_w = node_energy_j / elapsed if elapsed > 0 else 0.0
             total_energy_j += node_energy_j
-
-            prefill_e: float | None = None
-            generation_e: float | None = None
-            prefill_avg: float | None = None
-            generation_avg: float | None = None
-            if split is not None:
-                prefill_e = trapezoidal_energy_range(ts_profiles, 0.0, split)
-                generation_e = trapezoidal_energy_range(ts_profiles, split, elapsed)
-                total_prefill_energy_j += prefill_e
-                total_generation_energy_j += generation_e
-                prefill_dt = split
-                generation_dt = elapsed - split
-                prefill_avg = prefill_e / prefill_dt if prefill_dt > 0 else 0.0
-                generation_avg = (
-                    generation_e / generation_dt if generation_dt > 0 else 0.0
-                )
-
             node_stats.append(
                 NodePowerStats(
                     node_id=node_id,
                     samples=n,
                     avg_sys_power=avg_power_w,
-                    prefill_avg_sys_power=prefill_avg,
-                    generation_avg_sys_power=generation_avg,
-                    prefill_energy_joules=prefill_e,
-                    generation_energy_joules=generation_e,
                 )
             )
 
         total_avg_sys_w = total_energy_j / elapsed if elapsed > 0 else 0.0
-
-        prefill_seconds: float | None = None
-        generation_seconds: float | None = None
-        prefill_energy_joules: float | None = None
-        generation_energy_joules: float | None = None
-        prefill_avg_w: float | None = None
-        generation_avg_w: float | None = None
-        if split is not None:
-            prefill_seconds = split
-            generation_seconds = elapsed - split
-            prefill_energy_joules = total_prefill_energy_j
-            generation_energy_joules = total_generation_energy_j
-            prefill_avg_w = (
-                total_prefill_energy_j / prefill_seconds if prefill_seconds > 0 else 0.0
-            )
-            generation_avg_w = (
-                total_generation_energy_j / generation_seconds
-                if generation_seconds > 0
-                else 0.0
-            )
-
         return PowerUsage(
             elapsed_seconds=elapsed,
             nodes=node_stats,
             total_avg_sys_power_watts=total_avg_sys_w,
             total_energy_joules=total_energy_j,
-            prefill_seconds=prefill_seconds,
-            generation_seconds=generation_seconds,
-            prefill_energy_joules=prefill_energy_joules,
-            generation_energy_joules=generation_energy_joules,
-            prefill_avg_sys_power_watts=prefill_avg_w,
-            generation_avg_sys_power_watts=generation_avg_w,
         )
 
 
@@ -159,55 +89,3 @@ def trapezoidal_energy(
             continue
         energy_j += (p_prev.sys_power + p_cur.sys_power) / 2.0 * dt
     return energy_j
-
-
-def trapezoidal_energy_range(
-    ts_profiles: list[tuple[float, SystemPerformanceProfile]],
-    t_start: float,
-    t_end: float,
-) -> float:
-    """Integrate sys_power(t) over [t_start, t_end] using the trapezoidal rule.
-
-    Linearly interpolates power at the endpoints when they fall between
-    existing samples, so callers can integrate over arbitrary sub-windows
-    (e.g. the prefill segment) without losing accuracy. Returns 0 for an
-    empty or zero-length window. Falls back to constant-power assumption
-    when only one sample exists.
-    """
-    if t_end <= t_start:
-        return 0.0
-    if len(ts_profiles) == 0:
-        return 0.0
-    if len(ts_profiles) == 1:
-        return ts_profiles[0][1].sys_power * (t_end - t_start)
-
-    def power_at(t: float) -> float:
-        if t <= ts_profiles[0][0]:
-            return ts_profiles[0][1].sys_power
-        if t >= ts_profiles[-1][0]:
-            return ts_profiles[-1][1].sys_power
-        for i in range(1, len(ts_profiles)):
-            t_cur, p_cur = ts_profiles[i]
-            if t_cur >= t:
-                t_prev, p_prev = ts_profiles[i - 1]
-                span = t_cur - t_prev
-                if span <= 0:
-                    return p_cur.sys_power
-                frac = (t - t_prev) / span
-                return p_prev.sys_power + frac * (p_cur.sys_power - p_prev.sys_power)
-        return ts_profiles[-1][1].sys_power
-
-    p_start = power_at(t_start)
-    p_end = power_at(t_end)
-    in_range: list[tuple[float, float]] = [
-        (t, profile.sys_power) for t, profile in ts_profiles if t_start < t < t_end
-    ]
-    seq: list[tuple[float, float]] = [(t_start, p_start)] + in_range + [(t_end, p_end)]
-
-    energy_j = 0.0
-    for i in range(1, len(seq)):
-        dt = seq[i][0] - seq[i - 1][0]
-        if dt <= 0:
-            continue
-        energy_j += (seq[i - 1][1] + seq[i][1]) / 2.0 * dt
-    return energy_j

src/exo/utils/tests/test_channels.py

@@ -1,121 +0,0 @@
-import multiprocessing as mp
-import time
-
-import pytest
-from anyio import (
-    BrokenResourceError,
-    ClosedResourceError,
-    EndOfStream,
-    WouldBlock,
-    fail_after,
-)
-from loguru import logger
-
-from exo.utils.channels import ErrorOverride, MpReceiver, MpSender, channel, mp_channel
-
-
-class CustomClosedResourceError(ClosedResourceError):
-    pass
-
-
-class CustomBrokenResourceError(BrokenResourceError):
-    pass
-
-
-class CustomEndOfStream(EndOfStream):
-    pass
-
-
-class CustomWouldBlock(WouldBlock):
-    pass
-
-
-ERROR_OVERRIDE = ErrorOverride(
-    closed_resource_error=CustomClosedResourceError,
-    broken_resource_error=CustomBrokenResourceError,
-    end_of_stream=CustomEndOfStream,
-    would_block=CustomWouldBlock,
-)
-
-
-def foo(recv: MpReceiver[str]):
-    expected = ["hi", "hi 2", "bye"]
-    with recv as r:
-        for item in r:
-            assert item == expected.pop(0)
-
-
-def bar(send: MpSender[str]):
-    logger.warning("hi")
-    send.send("hi")
-    time.sleep(0.1)
-    logger.warning("hi 2")
-    send.send("hi 2")
-    time.sleep(0.1)
-    logger.warning("bye")
-    send.send("bye")
-    time.sleep(0.1)
-    send.close()
-
-
-@pytest.mark.anyio
-async def test_channel_ipc():
-    with fail_after(0.5):
-        s, r = mp_channel[str]()
-        p1 = mp.Process(target=foo, args=(r,))
-        p2 = mp.Process(target=bar, args=(s,))
-        p1.start()
-        p2.start()
-        p1.join()
-        p2.join()
-
-
-def test_channel_error_override_replaces_sync_errors_with_subclasses():
-    send, recv = channel[int](0, error_override_config=ERROR_OVERRIDE)
-
-    with pytest.raises(CustomWouldBlock) as would_block_info:
-        send.send_nowait(1)
-    assert type(would_block_info.value.__cause__) is WouldBlock
-
-    recv.close()
-    with pytest.raises(CustomBrokenResourceError) as broken_resource_info:
-        send.send_nowait(1)
-    assert type(broken_resource_info.value.__cause__) is BrokenResourceError
-
-    send.close()
-    with pytest.raises(CustomClosedResourceError) as closed_resource_info:
-        send.send_nowait(1)
-    assert type(closed_resource_info.value.__cause__) is ClosedResourceError
-
-
-@pytest.mark.anyio
-async def test_channel_error_override_replaces_async_errors_with_subclasses():
-    send, recv = channel[int](0, error_override_config=ERROR_OVERRIDE)
-    recv.close()
-
-    with pytest.raises(CustomBrokenResourceError) as broken_resource_info:
-        await send.send(1)
-    assert type(broken_resource_info.value.__cause__) is BrokenResourceError
-
-    send, recv = channel[int](error_override_config=ERROR_OVERRIDE)
-    send.close()
-    with pytest.raises(CustomEndOfStream) as end_of_stream_info:
-        await recv.receive()
-    assert type(end_of_stream_info.value.__cause__) is EndOfStream
-
-
-@pytest.mark.anyio
-async def test_channel_error_override_is_preserved_by_clones():
-    send, recv = channel[int](0, error_override_config=ERROR_OVERRIDE)
-    send_clone = send.clone()
-    recv.close()
-
-    with pytest.raises(CustomBrokenResourceError):
-        await send_clone.send(1)
-
-    send, recv = channel[int](0, error_override_config=ERROR_OVERRIDE)
-    cloned_send = recv.clone_sender()
-    recv.close()
-
-    with pytest.raises(CustomBrokenResourceError):
-        await cloned_send.send(1)

src/exo/utils/tests/test_daemon.py

@@ -0,0 +1,168 @@
+import contextlib
+import os
+from collections.abc import AsyncIterator
+
+import anyio
+import pytest
+from anyio import EndOfStream, create_task_group, fail_after
+
+from exo.utils.async_process import AsyncProcess
+from exo.utils.channels import MpReceiver, MpSender, Receiver, mp_channel
+from exo.utils.daemon import detach_stdio_to_devnull
+
+
+def _write_before_and_after_detach() -> None:
+    os.write(1, b"before stdout\n")
+    os.write(2, b"before stderr\n")
+    detach_stdio_to_devnull()
+    os.write(1, b"after stdout\n")
+    os.write(2, b"after stderr\n")
+
+
+def _write_grandchild_stdio(label: str) -> None:
+    os.write(1, f"{label} stdout\n".encode())
+    os.write(2, f"{label} stderr\n".encode())
+
+
+async def _spawn_grandchild_and_report(
+    result_sender: MpSender[tuple[int, bytes, bytes]],
+    label: str,
+) -> None:
+    result_sender.send(await _collect_spawned_child(label))
+    result_sender.close()
+
+
+async def _collect_spawned_child(label: str) -> tuple[int, bytes, bytes]:
+    process = AsyncProcess(_write_grandchild_stdio, args=(label,))
+    async with _started_process(process):
+        return await _collect_process_output(process)
+
+
+def _detach_stdio_then_spawn_captured_child(
+    result_sender: MpSender[tuple[int, bytes, bytes]],
+) -> None:
+    detach_stdio_to_devnull()
+    anyio.run(_spawn_grandchild_and_report, result_sender, "grandchild")
+
+
+def _detach_stdio_then_spawn_captured_children_sequentially(
+    result_sender: MpSender[list[tuple[int, bytes, bytes]]],
+) -> None:
+    async def run_children() -> list[tuple[int, bytes, bytes]]:
+        results: list[tuple[int, bytes, bytes]] = []
+        for index in range(5):
+            results.append(await _collect_spawned_child(f"grandchild-{index}"))
+        return results
+
+    detach_stdio_to_devnull()
+    result_sender.send(anyio.run(run_children))
+    result_sender.close()
+
+
+async def _collect_stream(stream: Receiver[bytes], output: bytearray) -> None:
+    while True:
+        try:
+            output.extend(await stream.receive())
+        except EndOfStream:
+            return
+
+
+async def _collect_process_output(
+    process: AsyncProcess,
+) -> tuple[int, bytes, bytes]:
+    stdout = bytearray()
+    stderr = bytearray()
+    exitcodes: list[int] = []
+
+    async with create_task_group() as collect_group:
+        collect_group.start_soon(_collect_stream, process.stdout, stdout)
+        collect_group.start_soon(_collect_stream, process.stderr, stderr)
+        exitcodes.append(await process.wait())
+
+    if not exitcodes:
+        raise RuntimeError("process exited without a return code")
+    return exitcodes[0], bytes(stdout), bytes(stderr)
+
+
+@contextlib.asynccontextmanager
+async def _started_process(process: AsyncProcess) -> AsyncIterator[None]:
+    async with create_task_group() as task_group:
+        await task_group.start(process.run)
+        try:
+            yield
+        finally:
+            await process.stop()
+
+
+async def _run_process_and_receive[T](
+    process: AsyncProcess,
+    recv: MpReceiver[T],
+    *,
+    timeout: float,
+) -> tuple[int, T]:
+    async with _started_process(process):
+        with fail_after(timeout):
+            result = await recv.receive_async()
+            exitcode = await process.wait()
+
+    return exitcode, result
+
+
+@pytest.mark.anyio
+async def test_detach_stdio_to_devnull_redirects_stdio_away_from_capture() -> None:
+    process = AsyncProcess(_write_before_and_after_detach)
+
+    async with _started_process(process):
+        exitcode, stdout, stderr = await _collect_process_output(process)
+
+    assert exitcode == 0
+    assert stdout == b"before stdout\n"
+    assert stderr == b"before stderr\n"
+
+
+@pytest.mark.anyio
+async def test_detached_stdio_process_can_spawn_and_capture_child_stdio() -> None:
+    send, recv = mp_channel[tuple[int, bytes, bytes]]()
+    process = AsyncProcess(_detach_stdio_then_spawn_captured_child, args=(send,))
+
+    try:
+        daemonized_parent_exitcode, result = await _run_process_and_receive(
+            process, recv, timeout=5
+        )
+    finally:
+        recv.close()
+
+    child_exitcode, child_stdout, child_stderr = result
+
+    assert daemonized_parent_exitcode == 0
+    assert child_exitcode == 0
+    assert child_stdout == b"grandchild stdout\n"
+    assert child_stderr == b"grandchild stderr\n"
+
+
+@pytest.mark.anyio
+async def test_detached_stdio_process_can_spawn_captured_children_sequentially() -> (
+    None
+):
+    send, recv = mp_channel[list[tuple[int, bytes, bytes]]]()
+    process = AsyncProcess(
+        _detach_stdio_then_spawn_captured_children_sequentially,
+        args=(send,),
+    )
+
+    try:
+        daemonized_parent_exitcode, results = await _run_process_and_receive(
+            process, recv, timeout=10
+        )
+    finally:
+        recv.close()
+
+    assert daemonized_parent_exitcode == 0
+    assert results == [
+        (
+            0,
+            f"grandchild-{index} stdout\n".encode(),
+            f"grandchild-{index} stderr\n".encode(),
+        )
+        for index in range(5)
+    ]

src/exo/utils/tests/test_mp_channel.py

@@ -0,0 +1,40 @@
+import multiprocessing as mp
+import time
+
+import pytest
+from anyio import fail_after
+from loguru import logger
+
+from exo.utils.channels import MpReceiver, MpSender, mp_channel
+
+
+def foo(recv: MpReceiver[str]):
+    expected = ["hi", "hi 2", "bye"]
+    with recv as r:
+        for item in r:
+            assert item == expected.pop(0)
+
+
+def bar(send: MpSender[str]):
+    logger.warning("hi")
+    send.send("hi")
+    time.sleep(0.1)
+    logger.warning("hi 2")
+    send.send("hi 2")
+    time.sleep(0.1)
+    logger.warning("bye")
+    send.send("bye")
+    time.sleep(0.1)
+    send.close()
+
+
+@pytest.mark.anyio
+async def test_channel_ipc():
+    with fail_after(0.5):
+        s, r = mp_channel[str]()
+        p1 = mp.Process(target=foo, args=(r,))
+        p2 = mp.Process(target=bar, args=(s,))
+        p1.start()
+        p2.start()
+        p1.join()
+        p2.join()

src/exo/utils/tests/test_pidfile.py

@@ -8,28 +8,36 @@ import textwrap
 from pathlib import Path
 from typing import Final
 
-from exo_rs import Pidfile
+import pytest
+
+import exo.utils.pidfile as pidfile
+from exo.utils.pidfile import acquire_exo_pidfile
 
 _CHILD_ACQUIRE_PIDFILE_SCRIPT: Final = textwrap.dedent(
     """
     import sys
     from pathlib import Path
+    from unittest.mock import patch
 
-    from exo_rs import Pidfile, PidfileError
+    import exo.utils.pidfile as pidfile
+    from exo.utils.pidfile import PidfileLockError, acquire_exo_pidfile
 
-    path = Path(sys.argv[1])
-    try:
-        handle = Pidfile(path, 0o0600)
-        handle.write()
-    except (OSError, PidfileError) as exception:
-        print(f"Failed to acquire EXO pidfile at {path}: {exception}")
-        raise SystemExit(73) from exception
+    with patch.object(pidfile, "EXO_PID_FILE", Path(sys.argv[1])):
+        try:
+            handle = acquire_exo_pidfile()
+        except PidfileLockError as exception:
+            print(str(exception))
+            raise SystemExit(73) from exception
 
-    del handle
+        del handle
     """
 )
 
 
+def _use_pidfile_path(monkeypatch: pytest.MonkeyPatch, path: Path) -> None:
+    monkeypatch.setattr(pidfile, "EXO_PID_FILE", path)
+
+
 def _run_child_acquire_pidfile(path: Path) -> subprocess.CompletedProcess[str]:
     return subprocess.run(
         [sys.executable, "-c", _CHILD_ACQUIRE_PIDFILE_SCRIPT, str(path)],
@@ -41,11 +49,12 @@ def _run_child_acquire_pidfile(path: Path) -> subprocess.CompletedProcess[str]:
 
 def test_acquire_exo_pidfile_writes_current_pid_and_removes_on_drop(
     tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
     path = tmp_path / "exo.pid"
+    _use_pidfile_path(monkeypatch, path)
 
-    handle = Pidfile(path, 0o0600)
-    handle.write()
+    handle = acquire_exo_pidfile()
     assert path.read_text() == str(os.getpid())
 
     del handle
@@ -56,11 +65,12 @@ def test_acquire_exo_pidfile_writes_current_pid_and_removes_on_drop(
 
 def test_acquire_exo_pidfile_rejects_second_process(
     tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
     path = tmp_path / "exo.pid"
+    _use_pidfile_path(monkeypatch, path)
 
-    handle = Pidfile(path, 0o0600)
-    handle.write()
+    handle = acquire_exo_pidfile()
     try:
         blocked_child = _run_child_acquire_pidfile(path)
         assert blocked_child.returncode == 73

src/exo/utils/tests/test_power_sampler.py

@@ -141,142 +141,6 @@ def test_trapezoidal_unit_single_sample() -> None:
     assert trapezoidal_energy(samples, elapsed=3.0) == 42.0 * 3.0
 
 
-def test_trapezoidal_range_interpolation() -> None:
-    """Sub-window integration should linearly interpolate at the boundaries."""
-    from exo.utils.power_sampler import trapezoidal_energy_range
-
-    # Two samples: t=0 W=10, t=10 W=20 -> power(t) = 10 + t
-    samples = [
-        (0.0, _make_profile(10.0)),
-        (10.0, _make_profile(20.0)),
-    ]
-    # Integral from t=4 to t=6: power goes 14 -> 16, mean 15, dt=2 -> 30 J
-    assert abs(trapezoidal_energy_range(samples, 4.0, 6.0) - 30.0) < 1e-9
-    # Integral over the full window matches the full trapezoidal integral.
-    full = trapezoidal_energy_range(samples, 0.0, 10.0)
-    assert abs(full - 150.0) < 1e-9
-
-
-def test_trapezoidal_range_zero_window() -> None:
-    """Zero-length or reversed windows integrate to zero."""
-    from exo.utils.power_sampler import trapezoidal_energy_range
-
-    samples = [(0.0, _make_profile(10.0)), (5.0, _make_profile(20.0))]
-    assert trapezoidal_energy_range(samples, 3.0, 3.0) == 0.0
-    assert trapezoidal_energy_range(samples, 5.0, 3.0) == 0.0
-
-
-def test_trapezoidal_range_splits_sum_to_full() -> None:
-    """Energy split at an arbitrary boundary should sum back to the full integral."""
-    from exo.utils.power_sampler import (
-        trapezoidal_energy,
-        trapezoidal_energy_range,
-    )
-
-    samples = [
-        (0.0, _make_profile(10.0)),
-        (1.0, _make_profile(20.0)),
-        (3.0, _make_profile(15.0)),
-        (5.0, _make_profile(25.0)),
-    ]
-    full = trapezoidal_energy(samples, elapsed=5.0)
-    # Split at t=2.5 (between samples) — interpolation should be exact.
-    left = trapezoidal_energy_range(samples, 0.0, 2.5)
-    right = trapezoidal_energy_range(samples, 2.5, 5.0)
-    assert abs((left + right) - full) < 1e-9
-
-
-async def test_prefill_generation_split() -> None:
-    """When mark_prefill_done() is called, the result should split energy."""
-    state: dict[NodeId, SystemPerformanceProfile] = {
-        NODE_A: _make_profile(10.0),
-    }
-    sampler = PowerSampler(get_node_system=lambda: state, interval=0.02)
-
-    async with anyio.create_task_group() as tg:
-        tg.start_soon(sampler.run)
-        # "Prefill" phase: power = 10 W
-        await anyio.sleep(0.1)
-        # Mark the boundary BEFORE changing state — this matches what
-        # _collect_text_generation_with_stats does in production: the mark
-        # fires on the first non-prefill chunk, so the boundary sample is
-        # the genuine end-of-prefill reading rather than the new phase's.
-        sampler.mark_prefill_done()
-        state[NODE_A] = _make_profile(30.0)
-        # "Generation" phase: power = 30 W
-        await anyio.sleep(0.1)
-        tg.cancel_scope.cancel()
-
-    result = sampler.result()
-    assert result.prefill_seconds is not None
-    assert result.generation_seconds is not None
-    assert result.prefill_energy_joules is not None
-    assert result.generation_energy_joules is not None
-    assert result.prefill_avg_sys_power_watts is not None
-    assert result.generation_avg_sys_power_watts is not None
-
-    # Phase durations should sum to the elapsed seconds.
-    assert (
-        abs(
-            (result.prefill_seconds + result.generation_seconds)
-            - result.elapsed_seconds
-        )
-        < 1e-6
-    )
-    # Phase energies should sum to (approximately) the total.
-    assert (
-        abs(
-            (result.prefill_energy_joules + result.generation_energy_joules)
-            - result.total_energy_joules
-        )
-        < 1e-6
-    )
-    # With the boundary sample anchored at the genuine end-of-prefill (10 W),
-    # prefill avg should converge tightly on 10 W and generation on 30 W.
-    # 15 W cleanly separates the two and would catch any cross-contamination.
-    assert result.prefill_avg_sys_power_watts < 15.0
-    assert result.generation_avg_sys_power_watts > 15.0
-    assert result.nodes[0].prefill_avg_sys_power is not None
-    assert result.nodes[0].generation_avg_sys_power is not None
-
-
-async def test_no_split_when_unmarked() -> None:
-    """If mark_prefill_done() is never called, phase fields stay None."""
-    state: dict[NodeId, SystemPerformanceProfile] = {
-        NODE_A: _make_profile(10.0),
-    }
-    sampler = PowerSampler(get_node_system=lambda: state, interval=0.02)
-    async with anyio.create_task_group() as tg:
-        tg.start_soon(sampler.run)
-        await anyio.sleep(0.05)
-        tg.cancel_scope.cancel()
-
-    result = sampler.result()
-    assert result.prefill_seconds is None
-    assert result.generation_seconds is None
-    assert result.prefill_energy_joules is None
-    assert result.generation_energy_joules is None
-    assert result.nodes[0].prefill_energy_joules is None
-    assert result.nodes[0].generation_energy_joules is None
-
-
-async def test_mark_prefill_done_is_idempotent() -> None:
-    """Only the first call to mark_prefill_done() should take effect."""
-    state: dict[NodeId, SystemPerformanceProfile] = {
-        NODE_A: _make_profile(10.0),
-    }
-    sampler = PowerSampler(get_node_system=lambda: state, interval=0.02)
-    async with anyio.create_task_group() as tg:
-        tg.start_soon(sampler.run)
-        await anyio.sleep(0.05)
-        sampler.mark_prefill_done()
-        first_prefill_at = sampler._prefill_done_at  # pyright: ignore[reportPrivateUsage]
-        await anyio.sleep(0.05)
-        sampler.mark_prefill_done()
-        assert sampler._prefill_done_at == first_prefill_at  # pyright: ignore[reportPrivateUsage]
-        tg.cancel_scope.cancel()
-
-
 async def test_result_stops_sampling() -> None:
     """Calling result() should stop the sampler's run loop."""
     state: dict[NodeId, SystemPerformanceProfile] = {

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

@@ -229,47 +229,6 @@ def has_non_kv_caches(cache: KVCacheType) -> bool:
     return any(is_non_trimmable_cache_entry(c) for c in cache)
 
 
-# Max snapshots retained per cache entry. Each CacheSnapshot pins detached GPU
-# copies of every non-trimmable (SSM/ArraysCache, RotatingKVCache) layer, so
-# retaining one per ~4096-token prefill chunk makes snapshot memory grow linearly
-# with context — the dominant residual cost when a single entry is grown to long
-# contexts on hybrid models (~56 MB/snapshot on Qwen3.5-122B, so a full 256K
-# context = 64 snapshots ≈ 3.6 GB). A sliding window of the most-recent N caps
-# this at N×per-snapshot (~0.9 GB here) while preserving the restore points
-# in-place grows actually use (they always extend from the tip).
-_MAX_RETAINED_SNAPSHOTS = 16
-
-
-def _bounded_snapshots(snapshots: list[CacheSnapshot]) -> list[CacheSnapshot]:
-    """Deduplicate snapshots by token position and bound the retained count.
-
-    Returned list is sorted ascending by ``token_count``.
-    """
-    # Deduplicate by position, keeping the most-recently-appended snapshot per
-    # position. Repeated in-place grows re-snapshot positions the kept old
-    # snapshots already cover, which would otherwise grow `_snapshots`
-    # unbounded even at constant context.
-    # TODO: keying on token_count alone is safe only while a position uniquely
-    # identifies the prefix within an entry (grows are strict prefix-extensions).
-    # If edit-and-regenerate, sliding-window/prefix trimming, cross-entry
-    # snapshot sharing, per-request adapter/LoRA swap, or branchy decoding
-    # (beam/parallel/speculative) is added, enrich the key to
-    # (token_count, prefix_hash[, media/adapter id]) — else a stale snapshot
-    # could be restored for a different prefix (silent wrong output).
-    by_position: dict[int, CacheSnapshot] = {}
-    for snapshot in snapshots:
-        by_position[snapshot.token_count] = snapshot
-    deduped = [by_position[pos] for pos in sorted(by_position)]
-
-    # Sliding window: keep only the most-recent N positions. In-place grows
-    # always extend from the tip, so the newest snapshots are the ones future
-    # grows restore from — dropping the oldest is never incorrect: a later hit on
-    # a prefix older than the window finds no snapshot <= target, so get_kv_cache
-    # returns a fresh cache (matched_index=None) and the request takes a full cold
-    # prefill — correct, just slower than a partial-hit reuse for that one request.
-    return deduped[-_MAX_RETAINED_SNAPSHOTS:]
-
-
 class KVPrefixCache:
     def __init__(self, group: mx.distributed.Group | None):
         self.prompts: list[mx.array] = []  # mx array of tokens (ints)
@@ -302,9 +261,7 @@ class KVPrefixCache:
         self._evict_if_needed()
         self.prompts.append(prompt_tokens)
         self.caches.append(deepcopy(cache))
-        self._snapshots.append(
-            _bounded_snapshots(ssm_snapshots) if ssm_snapshots else None
-        )
+        self._snapshots.append(ssm_snapshots)
         self._media_regions.append(media_regions or [])
         self.prefill_tps.append(prefill_tps)
         self._access_counter += 1
@@ -331,7 +288,7 @@ class KVPrefixCache:
 
         self.prompts[index] = prompt_tokens
         self.caches[index] = deepcopy(cache)
-        self._snapshots[index] = _bounded_snapshots(merged) or None
+        self._snapshots[index] = merged or None
         self._media_regions[index] = media_regions or []
         self.prefill_tps[index] = prefill_tps
         self._access_counter += 1

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

@@ -1,3 +1,4 @@
+from itertools import pairwise
 import json
 import os
 import re
@@ -5,6 +6,7 @@ import sys
 import tempfile
 import time
 from collections.abc import Generator
+from math import isqrt
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, cast
 
@@ -786,13 +788,75 @@ def mlx_force_oom(size: int = 200000) -> None:
     a = mx.random.uniform(shape=(size, size), dtype=mx.float32)
     b = mx.random.uniform(shape=(size, size), dtype=mx.float32)
     mx.eval(a, b)
-    c = mx.matmul(a, b)
-    d = mx.matmul(a, c)
-    e = mx.matmul(b, c)
-    f = mx.sigmoid(d + e)
+    c = mx.matmul(a, b)  # (size,size)
+    d = mx.matmul(a, c)  # (size,size)
+    e = mx.matmul(b, c)  # (size,size)
+    f = mx.sigmoid(d + e)  # (size,size)
     mx.eval(f)
 
 
+def mlx_force_oom2(bytes_alloc: int = 1024**5):  # the default is 1 petabyte lol
+    """
+    Force an Out-Of-Memory (OOM) error in MLX by performing large tensor operations.
+
+    NOTE: probably only works correctly on Apple unified memory
+    """
+
+    # TODO: I give up. this either uses swap (inconsistently) and only sometimes OOMs
+    #       or if I tune it to be more aggressive then it kerenel panics entirely;
+    #       there MIGHT be a way to make it not use swap memory but I'm not determined enough
+    #       to figure out how :)
+
+    def get_size(memory: int):
+        mat_elem = -(-memory // 4)  # per-matrix elements (4 bytes per elem)
+
+        # square root to get size (round up if not integer)
+        root = isqrt(mat_elem)
+        return root if root**2 == mat_elem else root + 1
+
+    def oom(size: int):
+        mx.set_default_device(mx.gpu)
+        mx.clear_cache()
+
+        # allocate a lot
+        z = mx.zeros(shape=(size, size), dtype=mx.float32)
+        t1 = [mx.random.uniform(shape=(size, size), dtype=mx.float32) for _ in range(2)]
+
+        # mat mul cycle
+        t2: list[mx.array] = []
+        for m1, m2 in pairwise(t1):
+            t2.append(mx.matmul(m1, m2))
+        print("t2-run")
+        mx.eval(*t2)
+        print("t2-eval")
+
+        # sigmoid sum
+        f = mx.sigmoid(sum(t1, start=z) + sum(t2, start=z))
+        print("f-run")
+        mx.eval(f)
+        print("f-eval")
+
+    # use supplied size, or computer appropriate size otherwise
+    fail_num = 0
+    while True:
+        try:
+            print(f"size {bytes_alloc / 1024**3} GB")
+            oom(get_size(bytes_alloc))
+            break
+        except RuntimeError as e:
+            max_bytes = re.compile(
+                r"\[metal::malloc\] Attempting to allocate (?:\d+) bytes which is greater than the maximum allowed buffer size of (?P<max_bytes>\d+) bytes."
+            ).match(str(e))
+            if max_bytes is None:
+                raise RuntimeError(
+                    "Tried to get max buffer, but wrong error format"
+                ) from e
+            bytes_alloc = round(int(max_bytes.group("max_bytes")) * 0.95**fail_num)
+            fail_num += 1
+
+    mlx_force_oom2()
+
+
 def set_wired_limit_for_model(model_size: Memory):
     """
     A context manager to temporarily change the wired limit.

src/exo/worker/main.py

@@ -8,10 +8,6 @@ from loguru import logger
 
 from exo.api.types import ImageEditsTaskParams
 from exo.download.download_utils import is_read_only_model_dir, resolve_existing_model
-from exo.routing.event_router import (
-    EventRouterBrokenResourceError,
-    EventRouterClosedResourceError,
-)
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_MAX_INSTANCE_RETRIES
 from exo.shared.models.model_cards import ModelId, card_cache
@@ -113,9 +109,7 @@ class Worker:
                 tg.start_soon(self._event_applier)
                 tg.start_soon(self._poll_connection_updates)
                 tg.start_soon(self._reconcile_custom_cards)
-        except* (EventRouterBrokenResourceError, EventRouterClosedResourceError):
-            # Event router has been closed (try-star syntax handles error groups)
-            pass
+
         finally:
             # Actual shutdown code - waits for all tasks to complete before executing.
             logger.info("Stopping Worker")

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

@@ -11,7 +11,6 @@ from mlx_lm.sample_utils import make_sampler
 from exo.shared.types.common import ModelId
 from exo.shared.types.text_generation import InputMessage, TextGenerationTaskParams
 from exo.worker.engines.mlx.cache import (
-    CacheSnapshot,
     KVPrefixCache,
     cache_length,
     encode_prompt,
@@ -78,74 +77,6 @@ class TestGetPrefixLength:
         assert get_prefix_length(a, b) == 0
 
 
-class TestSnapshotAccumulation:
-    """Locks in the fix for the actual per-grow Metal leak on hybrid (SSM)
-    models: `update_kv_cache` must not let `_snapshots` grow without bound when
-    the same entry is grown in place many times."""
-
-    def test_repeated_update_does_not_accumulate_snapshots(self):
-        with patch(
-            "exo.worker.engines.mlx.cache.get_memory_used_percentage",
-            return_value=0.0,
-        ):
-            kv_prefix_cache = KVPrefixCache(None)
-            initial = [
-                CacheSnapshot(states=[None], token_count=4096),
-                CacheSnapshot(states=[None], token_count=8192),
-            ]
-            kv_prefix_cache.add_kv_cache(
-                mx.arange(10000), [KVCache()], ssm_snapshots=initial
-            )
-
-            # Each in-place grow re-prefills from restore_pos and produces a
-            # fresh snapshot at a position the retained old snapshots already
-            # cover. Pre-fix this appended one snapshot per grow forever.
-            for _ in range(50):
-                fresh = [CacheSnapshot(states=[None], token_count=8192)]
-                kv_prefix_cache.update_kv_cache(
-                    0, mx.arange(10000), [KVCache()], fresh, restore_pos=8192
-                )
-
-            stored = kv_prefix_cache._snapshots[0]
-            assert stored is not None
-            # Bounded by the number of distinct snapshot positions (here 2),
-            # not by the 50 grows.
-            assert len(stored) == 2
-            assert sorted(s.token_count for s in stored) == [4096, 8192]
-            # The kept 8192 snapshot must be the most recently supplied one.
-            assert stored[1] is fresh[0]
-
-    def test_extension_caps_snapshots_to_sliding_window(self):
-        """Extending a single entry to a long context (one snapshot per ~4096
-        tokens) must cap retained snapshots to a sliding window of the most-recent
-        N, not keep all of them — that linear-in-context retention was the
-        residual OOM cause."""
-        from exo.worker.engines.mlx.cache import _MAX_RETAINED_SNAPSHOTS
-
-        with patch(
-            "exo.worker.engines.mlx.cache.get_memory_used_percentage",
-            return_value=0.0,
-        ):
-            kv_prefix_cache = KVPrefixCache(None)
-            # 64 distinct positions = a 262144-token context at 4096/chunk.
-            num_positions = 64
-            snaps = [
-                CacheSnapshot(states=[None], token_count=4096 * (i + 1))
-                for i in range(num_positions)
-            ]
-            kv_prefix_cache.add_kv_cache(
-                mx.arange(10), [KVCache()], ssm_snapshots=snaps
-            )
-
-            stored = kv_prefix_cache._snapshots[0]
-            assert stored is not None
-            # Capped at the window; the most-recent N positions are retained
-            # (in-place grows extend from the tip, so these are what get used).
-            assert len(stored) == _MAX_RETAINED_SNAPSHOTS
-            assert stored == snaps[-_MAX_RETAINED_SNAPSHOTS:]
-            assert stored[-1] is snaps[-1]  # tip always kept
-
-
 class TestKVPrefix:
     @pytest.fixture
     def mock_tokenizer(self):

tmp/test_trust_remote_code_attack.sh

@@ -7,9 +7,6 @@ set -uo pipefail
 
 HOST="${1:-localhost:52415}"
 MODEL_ID="KevTheHermit/security-testing"
-ENCODED_MODEL_ID=$(
-  python3 -c 'import sys, urllib.parse; print(urllib.parse.quote(sys.argv[1], safe=""))' "$MODEL_ID"
-)
 CUSTOM_CARDS_DIR="$HOME/.exo/custom_model_cards"
 CARD_FILE="$CUSTOM_CARDS_DIR/KevTheHermit--security-testing.toml"
 
@@ -74,30 +71,9 @@ PLACE_BODY=$(echo "$PLACE_RESPONSE" | sed '$d')
 echo "    HTTP $PLACE_CODE"
 echo "    Response: $PLACE_BODY"
 
-if [ "$PLACE_CODE" -ge 400 ]; then
-  echo "    Placement failed; cannot trigger tokenizer loading."
-  exit 1
-fi
-
-# Step 3b: Wait for placement to materialize before inference.
+# Step 3b: Send a chat completion to actually trigger tokenizer loading
 echo ""
-echo "[3b] Waiting for placed instance ..."
-if ! AWAIT_RESPONSE=$(curl -fsS --max-time 65 \
-  "http://$HOST/instance/await?model_id=$ENCODED_MODEL_ID&timeout_seconds=60" |
-  awk '/^data: / { sub(/^data: /, ""); print; exit }'); then
-  echo "    Timed out waiting for an instance for $MODEL_ID"
-  exit 1
-fi
-
-if ! printf '%s' "$AWAIT_RESPONSE" | grep -q '"type":"ready"'; then
-  echo "    Timed out waiting for an instance for $MODEL_ID"
-  exit 1
-fi
-echo "    Instance ready"
-
-# Step 3c: Send a chat completion to actually trigger tokenizer loading
-echo ""
-echo "[3c] Sending chat completion to trigger tokenizer load ..."
+echo "[3b] Sending chat completion to trigger tokenizer load ..."
 CHAT_RESPONSE=$(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}")
@@ -106,7 +82,7 @@ CHAT_BODY=$(echo "$CHAT_RESPONSE" | sed '$d')
 echo "    HTTP $CHAT_CODE"
 echo "    Response: $CHAT_BODY"
 echo ""
-echo "[3d] Checking for RCE proof ..."
+echo "[3c] Checking for RCE proof ..."
 sleep 5
 if [ -f /tmp/exo-rce-proof.txt ]; then
   echo "    VULNERABLE: Remote code executed!"

uv.lock

@@ -24,7 +24,7 @@ prerelease-mode = "allow"
 members = [
     "exo",
     "exo-bench",
-    "exo-rs",
+    "exo-pyo3-bindings",
     "exo-tools",
 ]
 overrides = [{ name = "opencv-python", marker = "python_full_version < '0'" }]
@@ -423,7 +423,7 @@ dependencies = [
     { name = "aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "aiohttp", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "anyio", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "exo-rs", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "exo-pyo3-bindings", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "fastapi", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "httpx", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -434,7 +434,6 @@ dependencies = [
     { name = "openai-harmony", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "psutil", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "pydantic", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "python-daemon", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "python-multipart", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "rustworkx", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "tiktoken", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -450,7 +449,7 @@ build = [
 ]
 mlx = [
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -461,7 +460,7 @@ mlx = [
 ]
 mlx-cpu = [
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cpu') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cpu') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cpu') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cpu') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx-cpu", marker = "sys_platform == 'linux'" },
@@ -473,7 +472,7 @@ mlx-cpu = [
 ]
 mlx-cuda12 = [
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx-cuda-12", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx_cuda_12-0.32.0-py3-none-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra != 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -487,7 +486,7 @@ mlx-cuda12 = [
 ]
 mlx-cuda13 = [
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "(sys_platform == 'darwin' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx-cuda-13", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx_cuda_13-0.32.0-py3-none-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none')" },
@@ -523,7 +522,7 @@ requires-dist = [
     { name = "exo", extras = ["mlx"], marker = "extra == 'mlx-cpu'" },
     { name = "exo", extras = ["mlx"], marker = "extra == 'mlx-cuda12'" },
     { name = "exo", extras = ["mlx"], marker = "extra == 'mlx-cuda13'" },
-    { name = "exo-rs", editable = "rust/exo_rs" },
+    { name = "exo-pyo3-bindings", editable = "rust/exo_pyo3_bindings" },
     { name = "fastapi", specifier = ">=0.116.1" },
     { name = "filelock", specifier = ">=3.18.0" },
     { name = "httpx", specifier = ">=0.28.1" },
@@ -549,7 +548,6 @@ requires-dist = [
     { name = "openai-harmony", specifier = ">=0.0.8" },
     { name = "psutil", specifier = ">=7.0.0" },
     { name = "pydantic", specifier = ">=2.11.7" },
-    { name = "python-daemon", specifier = ">=3.1.2" },
     { name = "python-multipart", specifier = ">=0.0.21" },
     { name = "rustworkx", specifier = ">=0.17.1" },
     { name = "tiktoken", specifier = ">=0.12.0" },
@@ -622,13 +620,13 @@ requires-dist = [
 ]
 
 [[package]]
-name = "exo-rs"
-version = "0.2.16"
-source = { editable = "rust/exo_rs" }
+name = "exo-pyo3-bindings"
+version = "0.2.2"
+source = { editable = "rust/exo_pyo3_bindings" }
 
 [package.dev-dependencies]
 dev = [
-    { name = "exo-rs", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "exo-pyo3-bindings", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "pytest", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "pytest-asyncio", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
 ]
@@ -637,7 +635,7 @@ dev = [
 
 [package.metadata.requires-dev]
 dev = [
-    { name = "exo-rs", editable = "rust/exo_rs" },
+    { name = "exo-pyo3-bindings", editable = "rust/exo_pyo3_bindings" },
     { name = "pytest", specifier = ">=8.4.0" },
     { name = "pytest-asyncio", specifier = ">=1.0.0" },
 ]
@@ -1164,15 +1162,6 @@ math = [
     { name = "sympy", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
 ]
 
-[[package]]
-name = "lockfile"
-version = "0.12.2"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/17/47/72cb04a58a35ec495f96984dddb48232b551aafb95bde614605b754fe6f7/lockfile-0.12.2.tar.gz", hash = "sha256:6aed02de03cba24efabcd600b30540140634fc06cfa603822d508d5361e9f799", size = 20874, upload-time = "2015-11-25T18:29:58.279Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/c8/22/9460e311f340cb62d26a38c419b1381b8593b0bb6b5d1f056938b086d362/lockfile-0.12.2-py2.py3-none-any.whl", hash = "sha256:6c3cb24f344923d30b2785d5ad75182c8ea7ac1b6171b08657258ec7429d50fa", size = 13564, upload-time = "2015-11-25T18:29:51.462Z" },
-]
-
 [[package]]
 name = "loguru"
 version = "0.7.3"
@@ -1343,7 +1332,7 @@ dependencies = [
     { name = "hf-transfer", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "matplotlib", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -1374,7 +1363,7 @@ sdist = { url = "https://files.pythonhosted.org/packages/55/fa/96d4cc7ada2833571
 
 [[package]]
 name = "mlx"
-version = "0.32.0.dev20260506+cc3f3e60"
+version = "0.32.0.dev20260512+cc3f3e60"
 source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }
 resolution-markers = [
     "sys_platform == 'darwin'",
@@ -1549,7 +1538,7 @@ version = "0.31.3"
 source = { git = "https://github.com/rltakashige/mlx-lm?branch=leo%2Fdeepseek-v4#6a3df6cd6b00a347ee40f12d97a182aaf86ea599" }
 dependencies = [
     { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "protobuf", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "pyyaml", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -1565,7 +1554,7 @@ dependencies = [
     { name = "datasets", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "fastapi", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "miniaudio", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-    { name = "mlx", version = "0.32.0.dev20260506+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
+    { name = "mlx", version = "0.32.0.dev20260512+cc3f3e60", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#cc3f3e60be1289506125f2fa19b73b05aa770df8" }, marker = "sys_platform == 'darwin' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_aarch64.whl" }, marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine != 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx", version = "0.32.0", source = { url = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv/releases/download/mlx_cuda/mlx-0.32.0-cp313-cp313-manylinux_2_35_x86_64.whl" }, marker = "(platform_machine != 'aarch64' and sys_platform == 'linux') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (platform_machine == 'aarch64' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (sys_platform != 'linux' and extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
@@ -2311,18 +2300,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/27/98/822b924a4a3eb58aacba84444c7439fce32680592f394de26af9c76e2569/pytest_env-1.2.0-py3-none-any.whl", hash = "sha256:d7e5b7198f9b83c795377c09feefa45d56083834e60d04767efd64819fc9da00", size = 6251, upload-time = "2025-10-09T19:15:46.077Z" },
 ]
 
-[[package]]
-name = "python-daemon"
-version = "3.1.2"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "lockfile", marker = "sys_platform == 'darwin' or sys_platform == 'linux' or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda12') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cpu' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-cuda13') or (extra == 'extra-3-exo-mlx-cuda12' and extra == 'extra-3-exo-mlx-none') or (extra == 'extra-3-exo-mlx-cuda13' and extra == 'extra-3-exo-mlx-none')" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/3d/37/4f10e37bdabc058a32989da2daf29e57dc59dbc5395497f3d36d5f5e2694/python_daemon-3.1.2.tar.gz", hash = "sha256:f7b04335adc473de877f5117e26d5f1142f4c9f7cd765408f0877757be5afbf4", size = 71576, upload-time = "2024-12-03T08:41:07.843Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/45/3c/b88167e2d6785c0e781ee5d498b07472aeb9b6765da3b19e7cc9e0813841/python_daemon-3.1.2-py3-none-any.whl", hash = "sha256:b906833cef63502994ad48e2eab213259ed9bb18d54fa8774dcba2ff7864cec6", size = 30872, upload-time = "2024-12-03T08:41:03.322Z" },
-]
-
 [[package]]
 name = "python-dateutil"
 version = "2.9.0.post0"