Use 2GB buffer for more accurate bandwidth measurement

- Increase buffer size from 512MB to 2GB to better saturate memory bus
- Use 2D array shape to avoid issues with very large 1D arrays
- Improves accuracy from ~75% to ~82% of theoretical peak

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Cargo.lock

@@ -4340,25 +4340,6 @@ dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "system_custodian"
-version = "0.0.1"
-dependencies = [
- "delegate",
- "derive_more",
- "either",
- "extend",
- "futures",
- "futures-timer",
- "impl-trait-for-tuples",
- "keccak-const",
- "log",
- "thiserror 2.0.17",
- "tokio",
- "tracing-subscriber",
- "util",
-]
-
 [[package]]
 name = "tagptr"
 version = "0.2.0"

Cargo.toml

@@ -3,7 +3,6 @@ resolver = "3"
 members = [
     "rust/networking",
     "rust/exo_pyo3_bindings",
-    "rust/system_custodian",
     "rust/util",
 ]
 
@@ -25,7 +24,6 @@ opt-level = 3
 [workspace.dependencies]
 ## Crate members as common dependencies
 networking = { path = "rust/networking" }
-system_custodian = { path = "rust/system_custodian" }
 util = { path = "rust/util" }
 
 # Proc-macro authoring tools

app/EXO/EXO/ContentView.swift

@@ -56,6 +56,11 @@ struct ContentView: View {
     }
 
     private var shouldShowLocalNetworkWarning: Bool {
+        // Show warning if local network is not working and EXO is running.
+        // The checker uses a longer timeout on first launch to allow time for
+        // the permission prompt, so this correctly handles both:
+        // 1. User denied permission on first launch
+        // 2. Permission broke after restart (macOS TCC bug)
         if case .notWorking = localNetworkChecker.status {
             return controller.status != .stopped
         }

app/EXO/EXO/Services/LocalNetworkChecker.swift

@@ -5,8 +5,8 @@ import os.log
 /// Checks if the app's local network permission is actually functional.
 ///
 /// macOS local network permission can appear enabled in System Preferences but not
-/// actually work after a restart. This service detects this by creating a UDP
-/// connection to the mDNS multicast address (224.0.0.251:5353).
+/// actually work after a restart. This service uses NWConnection to mDNS multicast
+/// to verify actual connectivity.
 @MainActor
 final class LocalNetworkChecker: ObservableObject {
     enum Status: Equatable {
@@ -35,30 +35,43 @@ final class LocalNetworkChecker: ObservableObject {
     }
 
     private static let logger = Logger(subsystem: "io.exo.EXO", category: "LocalNetworkChecker")
+    private static let hasCompletedInitialCheckKey = "LocalNetworkChecker.hasCompletedInitialCheck"
 
     @Published private(set) var status: Status = .unknown
-    @Published private(set) var lastConnectionState: String = "none"
 
     private var connection: NWConnection?
     private var checkTask: Task<Void, Never>?
 
+    /// Whether we've completed at least one check (stored in UserDefaults)
+    private var hasCompletedInitialCheck: Bool {
+        get { UserDefaults.standard.bool(forKey: Self.hasCompletedInitialCheckKey) }
+        set { UserDefaults.standard.set(newValue, forKey: Self.hasCompletedInitialCheckKey) }
+    }
+
     /// Checks if local network access is working.
     func check() {
         checkTask?.cancel()
         status = .checking
-        lastConnectionState = "connecting"
+
+        // Use longer timeout on first launch to allow time for permission prompt
+        let isFirstCheck = !hasCompletedInitialCheck
+        let timeout: UInt64 = isFirstCheck ? 30_000_000_000 : 3_000_000_000
 
         checkTask = Task { [weak self] in
             guard let self else { return }
-            let result = await self.performCheck()
+
+            Self.logger.info("Checking local network connectivity (first check: \(isFirstCheck))")
+            let result = await self.checkConnectivity(timeout: timeout)
             self.status = result
+            self.hasCompletedInitialCheck = true
+
             Self.logger.info("Local network check complete: \(result.displayText)")
         }
     }
 
-    private func performCheck() async -> Status {
-        Self.logger.info("Checking local network access via UDP multicast")
-
+    /// Checks connectivity using NWConnection to mDNS multicast.
+    /// The connection attempt triggers the permission prompt if not yet shown.
+    private func checkConnectivity(timeout: UInt64) async -> Status {
         connection?.cancel()
         connection = nil
 
@@ -84,22 +97,7 @@ final class LocalNetworkChecker: ObservableObject {
                 continuation.resume(returning: status)
             }
 
-            conn.stateUpdateHandler = { [weak self] state in
-                let stateStr: String
-                switch state {
-                case .setup: stateStr = "setup"
-                case .preparing: stateStr = "preparing"
-                case .ready: stateStr = "ready"
-                case .waiting(let e): stateStr = "waiting(\(e))"
-                case .failed(let e): stateStr = "failed(\(e))"
-                case .cancelled: stateStr = "cancelled"
-                @unknown default: stateStr = "unknown"
-                }
-
-                Task { @MainActor in
-                    self?.lastConnectionState = stateStr
-                }
-
+            conn.stateUpdateHandler = { state in
                 switch state {
                 case .ready:
                     resumeOnce(.working)
@@ -108,6 +106,7 @@ final class LocalNetworkChecker: ObservableObject {
                     if errorStr.contains("54") || errorStr.contains("ECONNRESET") {
                         resumeOnce(.notWorking(reason: "Connection blocked"))
                     }
+                // Otherwise keep waiting - might be showing permission prompt
                 case .failed(let error):
                     let errorStr = "\(error)"
                     if errorStr.contains("65") || errorStr.contains("EHOSTUNREACH")
@@ -127,7 +126,7 @@ final class LocalNetworkChecker: ObservableObject {
             conn.start(queue: .main)
 
             Task {
-                try? await Task.sleep(nanoseconds: 3_000_000_000)
+                try? await Task.sleep(nanoseconds: timeout)
                 let state = conn.state
                 switch state {
                 case .ready:

bench/exo_bench.py

@@ -241,6 +241,9 @@ class PromptSizer:
             ids = tokenizer.apply_chat_template(
                 messages, tokenize=True, add_generation_prompt=True
             )
+            # Fix for transformers 5.x
+            if hasattr(ids, "input_ids"):
+                ids = ids.input_ids
             return int(len(ids))
 
         return count_fn

dashboard/package-lock.json

@@ -863,6 +863,7 @@
 			"integrity": "sha512-oH8tXw7EZnie8FdOWYrF7Yn4IKrqTFHhXvl8YxXxbKwTMcD/5NNCryUSEXRk2ZR4ojnub0P8rNrsVGHXWqIDtA==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@standard-schema/spec": "^1.0.0",
 				"@sveltejs/acorn-typescript": "^1.0.5",
@@ -902,6 +903,7 @@
 			"integrity": "sha512-Y1Cs7hhTc+a5E9Va/xwKlAJoariQyHY+5zBgCZg4PFWNYQ1nMN9sjK1zhw1gK69DuqVP++sht/1GZg1aRwmAXQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@sveltejs/vite-plugin-svelte-inspector": "^4.0.1",
 				"debug": "^4.4.1",
@@ -1518,6 +1520,7 @@
 			"integrity": "sha512-LCCV0HdSZZZb34qifBsyWlUmok6W7ouER+oQIGBScS8EsZsQbrtFTUrDX4hOl+CS6p7cnNC4td+qrSVGSCTUfQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"undici-types": "~6.21.0"
 			}
@@ -1527,6 +1530,7 @@
 			"resolved": "https://registry.npmjs.org/acorn/-/acorn-8.15.0.tgz",
 			"integrity": "sha512-NZyJarBfL7nWwIq+FDL6Zp/yHEhePMNnnJ0y3qfieCrmNvYct8uvtiV41UvlSe6apAfk0fY1FbWx+NwfmpvtTg==",
 			"license": "MIT",
+			"peer": true,
 			"bin": {
 				"acorn": "bin/acorn"
 			},
@@ -1939,6 +1943,7 @@
 			"integrity": "sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==",
 			"dev": true,
 			"license": "ISC",
+			"peer": true,
 			"engines": {
 				"node": ">=12"
 			}
@@ -2646,6 +2651,7 @@
 			"integrity": "sha512-5gTmgEY/sqK6gFXLIsQNH19lWb4ebPDLA4SdLP7dsWkIXHWlG66oPuVvXSGFPppYZz8ZDZq0dYYrbHfBCVUb1Q==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"engines": {
 				"node": ">=12"
 			},
@@ -2833,6 +2839,7 @@
 			"resolved": "https://registry.npmjs.org/svelte/-/svelte-5.45.3.tgz",
 			"integrity": "sha512-ngKXNhNvwPzF43QqEhDOue7TQTrG09em1sd4HBxVF0Wr2gopAmdEWan+rgbdgK4fhBtSOTJO8bYU4chUG7VXZQ==",
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@jridgewell/remapping": "^2.3.4",
 				"@jridgewell/sourcemap-codec": "^1.5.0",
@@ -2977,6 +2984,7 @@
 			"integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==",
 			"dev": true,
 			"license": "Apache-2.0",
+			"peer": true,
 			"bin": {
 				"tsc": "bin/tsc",
 				"tsserver": "bin/tsserver"
@@ -2998,6 +3006,7 @@
 			"integrity": "sha512-+Oxm7q9hDoLMyJOYfUYBuHQo+dkAloi33apOPP56pzj+vsdJDzr+j1NISE5pyaAuKL4A3UD34qd0lx5+kfKp2g==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"esbuild": "^0.25.0",
 				"fdir": "^6.4.4",

dashboard/src/lib/components/ChatForm.svelte

@@ -60,12 +60,39 @@
 		return models;
 	});
 
-	// Auto-select the first available model if none is selected
+	// Track previous model IDs to detect newly added models (plain variable to avoid reactive loop)
+	let previousModelIds: Set<string> = new Set();
+
+	// Auto-select the first available model if none is selected, if current selection is stale, or if a new model is added
 	$effect(() => {
 		const models = availableModels();
-		if (models.length > 0 && !currentModel) {
-			setSelectedChatModel(models[0].id);
+		const currentModelIds = new Set(models.map(m => m.id));
+
+		if (models.length > 0) {
+			// Find newly added models (in current but not in previous)
+			const newModels = models.filter(m => !previousModelIds.has(m.id));
+
+			// If no model selected, select the first available
+			if (!currentModel) {
+				setSelectedChatModel(models[0].id);
+			}
+			// If current model is stale (no longer has a running instance), reset to first available
+			else if (!models.some(m => m.id === currentModel)) {
+				setSelectedChatModel(models[0].id);
+			}
+			// If a new model was just added, select it
+			else if (newModels.length > 0 && previousModelIds.size > 0) {
+				setSelectedChatModel(newModels[0].id);
+			}
+		} else {
+			// No instances running - clear the selected model
+			if (currentModel) {
+				setSelectedChatModel('');
+			}
 		}
+
+		// Update previous model IDs for next comparison
+		previousModelIds = currentModelIds;
 	});
 
 	function getInstanceModelId(instanceWrapped: unknown): string {

dashboard/src/routes/+page.svelte

@@ -400,10 +400,8 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 				const errorText = await response.text();
 				console.error('Failed to launch instance:', errorText);
 			} else {
-				// Auto-select the launched model only if no model is currently selected
-				if (!selectedChatModel()) {
-					setSelectedChatModel(modelId);
-				}
+				// Always auto-select the newly launched model so the user chats to what they just launched
+				setSelectedChatModel(modelId);
 				
 				// Scroll to the bottom of instances container to show the new instance
 				// Use multiple attempts to ensure DOM has updated with the new instance
@@ -763,6 +761,10 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 	async function deleteInstance(instanceId: string) {
 		if (!confirm(`Delete instance ${instanceId.slice(0, 8)}...?`)) return;
 		
+		// Get the model ID of the instance being deleted before we delete it
+		const deletedInstanceModelId = getInstanceModelId(instanceData[instanceId]);
+		const wasSelected = selectedChatModel() === deletedInstanceModelId;
+		
 		try {
 			const response = await fetch(`/instance/${instanceId}`, {
 				method: 'DELETE',
@@ -771,6 +773,24 @@ function toggleInstanceDownloadDetails(nodeId: string): void {
 			
 			if (!response.ok) {
 				console.error('Failed to delete instance:', response.status);
+			} else if (wasSelected) {
+				// If we deleted the currently selected model, switch to another available model
+				// Find another instance that isn't the one we just deleted
+				const remainingInstances = Object.entries(instanceData).filter(([id]) => id !== instanceId);
+				if (remainingInstances.length > 0) {
+					// Select the last instance (most recently added, since objects preserve insertion order)
+					const [, lastInstance] = remainingInstances[remainingInstances.length - 1];
+					const newModelId = getInstanceModelId(lastInstance);
+					if (newModelId && newModelId !== 'Unknown' && newModelId !== 'Unknown Model') {
+						setSelectedChatModel(newModelId);
+					} else {
+						// Clear selection if no valid model found
+						setSelectedChatModel('');
+					}
+				} else {
+					// No more instances, clear the selection
+					setSelectedChatModel('');
+				}
 			}
 		} catch (error) {
 			console.error('Error deleting instance:', error);

justfile

@@ -1,3 +1,5 @@
+export NIX_CONFIG := "extra-experimental-features = nix-command flakes"
+
 fmt:
     nix fmt
 

pyproject.toml

@@ -23,6 +23,7 @@ dependencies = [
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
+    "httpx>=0.28.1",
 ]
 
 [project.scripts]

rust/parts.nix

@@ -81,20 +81,6 @@
 
       config = {
         packages = {
-          # The system_custodian binary
-          system_custodian = craneLib.buildPackage (
-            commonArgs
-            // {
-              inherit cargoArtifacts;
-              cargoExtraArgs = "-p system_custodian";
-
-              meta = {
-                description = "System custodian daemon for exo";
-                mainProgram = "system_custodian";
-              };
-            }
-          );
-
           # Python bindings wheel via maturin
           exo_pyo3_bindings = craneLib.buildPackage (
             commonArgs

rust/system_custodian/Cargo.toml

@@ -1,47 +0,0 @@
-[package]
-name = "system_custodian"
-version = { workspace = true }
-edition = { workspace = true }
-publish = false
-
-[lib]
-doctest = false
-name = "system_custodian"
-path = "src/lib.rs"
-
-[[bin]]
-path = "src/bin/main.rs"
-name = "system_custodian"
-doc = false
-
-[lints]
-workspace = true
-
-[dependencies]
-# datastructures
-either = { workspace = true }
-
-# macro dependencies
-extend = { workspace = true }
-delegate = { workspace = true }
-impl-trait-for-tuples = { workspace = true }
-derive_more = { workspace = true }
-
-# async
-tokio = { workspace = true, features = ["full"] }
-futures = { workspace = true }
-futures-timer = { workspace = true }
-
-# utility dependencies
-util = { workspace = true }
-thiserror = { workspace = true }
-#internment = { workspace = true }
-#recursion = { workspace = true }
-#generativity = { workspace = true }
-#itertools = { workspace = true }
-tracing-subscriber = { version = "0.3.19", features = ["default", "env-filter"] }
-keccak-const = { workspace = true }
-
-# tracing/logging
-log = { workspace = true }
-

rust/system_custodian/src/bin/main.rs

@@ -1,4 +0,0 @@
-//! TODO: documentation
-//!
-
-fn main() {}

rust/system_custodian/src/lib.rs

@@ -1,69 +0,0 @@
-//! This crate defines the logic of, and ways to interact with, Exo's **_System Custodian_** daemon.
-//!
-//! The **_System Custodian_** daemon is supposed to be a long-living process that precedes the
-//! launch of the Exo application, and responsible for ensuring the system (configuration, settings,
-//! etc.) is in an appropriate state to facilitate the running of Exo application.
-//! The **_System Custodian_** daemon shall expose a [D-Bus](https://www.freedesktop.org/wiki/Software/dbus/)
-//! service which Exo application use to _control & query_ it.
-//!
-//! # Lifecycle
-//! When the Exo application starts, it will _wake_ the **_System Custodian_** daemon for the
-//! duration of its lifetime, and after it has terminated the daemon will go back to sleep. When
-//! the daemon wakes up, it will configure the system into a state suitable for the Exo Application;
-//! When the daemon goes to sleep, it will revert those changes as much as it can in case they were
-//! destructive to the user's pre-existing configurations.
-//!
-//! # Responsibilities
-//! TODO: these are purely on MacOS, but change to be more broad
-//! The **_System Custodian_** daemon is responsible for using System Configuration framework to
-//!  1. duplicate the current network set
-//!  2. modify existing services to turn on IPv6 if not there
-//!  3. remove any bridge services & add any missing services that AREN'T bridge
-//! TODO: In the future:
-//!  1. run a dummy AWDL service to [allow for macOS peer-to-peer wireless networking](https://yggdrasil-network.github.io/2019/08/19/awdl.html)
-//!  2. toggle some GPU/memory configurations to speed up GPU (ask Alex what those configurations are)
-//!  3. if we ever decide to provide our **own network interfaces** that abstract over some userland
-//!     logic, this would be the place to spin that up.
-//!
-//! Then it will watch the SCDynamicStore for:
-//!  1. all __actual__ network interfaces -> collect information on them e.g. their BSD name, MAC
-//!     address, MTU, IPv6 addresses, etc. -> and set up watchers/notifiers to inform the DBus
-//!     interface of any changes
-//!  2. watch for any __undesirable__ changes to configuration and revert it
-//!
-//! It should somehow (probably through system sockets and/or BSD interface) trigger IPv6 NDP on
-//! each of the interfaces & also listen to/query for any changes on the OS routing cache??
-//! Basically emulate the `ping6 ff02::1%enX` and `ndp -an` commands BUT BETTER!!!
-//!  1. all that info should coalesce back to the overall state colleted -> should be queryable
-//!     over D-Bus
-//! TODO:
-//!  1. we might potentially add to this step a handshake of some kind...? To ensure that we can
-//!     ACTUALLY communicate with that machine over that link over e.g. TCP, UDP, etc. Will the
-//!     handshake require to know Node ID? Will the handshake require heartbeats? Who knows...
-//!  2. if we ever decide to write proprietary L2/L3 protocols for quicker communication,
-//!     e.g. [AF_NDRV](https://www.zerotier.com/blog/how-zerotier-eliminated-kernel-extensions-on-macos/)
-//!     for raw ethernet frame communication, or even a [custom thunderbolt PCIe driver](https://developer.apple.com/documentation/pcidriverkit/creating-custom-pcie-drivers-for-thunderbolt-devices),
-//!     then this would be the place to carry out discovery and propper handshakes with devices
-//!     on the other end of the link.
-//!
-
-// enable Rust-unstable features for convenience
-#![feature(trait_alias)]
-#![feature(stmt_expr_attributes)]
-#![feature(type_alias_impl_trait)]
-#![feature(specialization)]
-#![feature(unboxed_closures)]
-#![feature(const_trait_impl)]
-#![feature(fn_traits)]
-
-pub(crate) mod private {
-    // sealed traits support
-    pub trait Sealed {}
-    impl<T: ?Sized> Sealed for T {}
-}
-
-/// Namespace for all the type/trait aliases used by this crate.
-pub(crate) mod alias {}
-
-/// Namespace for crate-wide extension traits/methods
-pub(crate) mod ext {}

src/exo/master/api.py

@@ -13,12 +13,6 @@ from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
-from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
-    HarmonyEncodingName,
-    Role,
-    StreamableParser,
-    load_harmony_encoding,
-)
 
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
@@ -67,8 +61,6 @@ from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
-encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
-
 
 def chunk_to_response(
     chunk: TokenChunk, command_id: CommandId
@@ -381,35 +373,8 @@ class API:
             instance_id=instance_id,
         )
 
-    async def _process_gpt_oss(self, token_chunks: Receiver[TokenChunk]):
-        stream = StreamableParser(encoding, role=Role.ASSISTANT)
-        thinking = False
-
-        async for chunk in token_chunks:
-            stream.process(chunk.token_id)
-
-            delta = stream.last_content_delta
-            ch = stream.current_channel
-
-            if ch == "analysis" and not thinking:
-                thinking = True
-                yield chunk.model_copy(update={"text": "<think>"})
-
-            if ch != "analysis" and thinking:
-                thinking = False
-                yield chunk.model_copy(update={"text": "</think>"})
-
-            if delta:
-                yield chunk.model_copy(update={"text": delta})
-
-            if chunk.finish_reason is not None:
-                if thinking:
-                    yield chunk.model_copy(update={"text": "</think>"})
-                yield chunk
-                break
-
     async def _chat_chunk_stream(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> AsyncGenerator[TokenChunk, None]:
         """Yield `TokenChunk`s for a given command until completion."""
 
@@ -417,16 +382,10 @@ class API:
             self._chat_completion_queues[command_id], recv = channel[TokenChunk]()
 
             with recv as token_chunks:
-                if parse_gpt_oss:
-                    async for chunk in self._process_gpt_oss(token_chunks):
-                        yield chunk
-                        if chunk.finish_reason is not None:
-                            break
-                else:
-                    async for chunk in token_chunks:
-                        yield chunk
-                        if chunk.finish_reason is not None:
-                            break
+                async for chunk in token_chunks:
+                    yield chunk
+                    if chunk.finish_reason is not None:
+                        break
 
         except anyio.get_cancelled_exc_class():
             # TODO: TaskCancelled
@@ -442,11 +401,11 @@ class API:
             del self._chat_completion_queues[command_id]
 
     async def _generate_chat_stream(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> AsyncGenerator[str, None]:
         """Generate chat completion stream as JSON strings."""
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             chunk_response: ChatCompletionResponse = chunk_to_response(
                 chunk, command_id
             )
@@ -458,7 +417,7 @@ class API:
                 yield "data: [DONE]\n\n"
 
     async def _collect_chat_completion(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> ChatCompletionResponse:
         """Collect all token chunks for a chat completion and return a single response."""
 
@@ -466,7 +425,7 @@ class API:
         model: str | None = None
         finish_reason: FinishReason | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             if model is None:
                 model = chunk.model
 
@@ -495,7 +454,7 @@ class API:
         )
 
     async def _collect_chat_completion_with_stats(
-        self, command_id: CommandId, parse_gpt_oss: bool
+        self, command_id: CommandId
     ) -> BenchChatCompletionResponse:
         text_parts: list[str] = []
         model: str | None = None
@@ -503,7 +462,7 @@ class API:
 
         stats: GenerationStats | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id, parse_gpt_oss):
+        async for chunk in self._chat_chunk_stream(command_id):
             if model is None:
                 model = chunk.model
 
@@ -544,8 +503,6 @@ class API:
         """Handle chat completions, supporting both streaming and non-streaming responses."""
         model_meta = await resolve_model_meta(payload.model)
         payload.model = model_meta.model_id
-        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
-        logger.info(f"{parse_gpt_oss=}")
 
         if not any(
             instance.shard_assignments.model_id == payload.model
@@ -562,17 +519,16 @@ class API:
         await self._send(command)
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id, parse_gpt_oss),
+                self._generate_chat_stream(command.command_id),
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id, parse_gpt_oss)
+        return await self._collect_chat_completion(command.command_id)
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionTaskParams
     ) -> BenchChatCompletionResponse:
         model_meta = await resolve_model_meta(payload.model)
-        parse_gpt_oss = "gpt-oss" in model_meta.model_id.lower()
         payload.model = model_meta.model_id
 
         if not any(
@@ -589,10 +545,7 @@ class API:
         command = ChatCompletion(request_params=payload)
         await self._send(command)
 
-        response = await self._collect_chat_completion_with_stats(
-            command.command_id,
-            parse_gpt_oss,
-        )
+        response = await self._collect_chat_completion_with_stats(command.command_id)
         return response
 
     def _calculate_total_available_memory(self) -> Memory:

src/exo/master/placement_utils.py

@@ -49,20 +49,22 @@ def get_smallest_cycles(cycles: list[list[NodeInfo]]) -> list[list[NodeInfo]]:
     return [cycle for cycle in cycles if len(cycle) == min_nodes]
 
 
-def get_shard_assignments_for_pipeline_parallel(
+def _assign_layers_by_ram(
     model_meta: ModelMetadata,
     selected_cycle: list[NodeWithProfile],
-):
-    cycle_memory = sum(
-        (node.node_profile.memory.ram_available for node in selected_cycle),
-        start=Memory(),
-    )
+) -> ShardAssignments:
+    """Assign layers proportionally based on available RAM."""
     total_layers = model_meta.n_layers
     world_size = len(selected_cycle)
     runner_to_shard: dict[RunnerId, ShardMetadata] = {}
     node_to_runner: dict[NodeId, RunnerId] = {}
 
+    cycle_memory = sum(
+        (node.node_profile.memory.ram_available for node in selected_cycle),
+        start=Memory(),
+    )
     layers_assigned = 0
+
     for i, node in enumerate(selected_cycle):
         if i == len(selected_cycle) - 1:
             node_layers = total_layers - layers_assigned
@@ -77,7 +79,6 @@ def get_shard_assignments_for_pipeline_parallel(
             node_layers = max(1, node_layers)
 
         runner_id = RunnerId()
-
         shard = PipelineShardMetadata(
             model_meta=model_meta,
             device_rank=i,
@@ -86,18 +87,143 @@ def get_shard_assignments_for_pipeline_parallel(
             end_layer=layers_assigned + node_layers,
             n_layers=total_layers,
         )
-
         runner_to_shard[runner_id] = shard
         node_to_runner[node.node_id] = runner_id
         layers_assigned += node_layers
 
-    shard_assignments = ShardAssignments(
+    return ShardAssignments(
         model_id=model_meta.model_id,
         runner_to_shard=runner_to_shard,
         node_to_runner=node_to_runner,
     )
 
-    return shard_assignments
+
+def _reserve_base_layers(world_size: int, total_layers: int) -> dict[int, int]:
+    """Reserve 1 layer per node to ensure connectivity."""
+    assignments = {i: 0 for i in range(world_size)}
+    remaining_layers = total_layers
+
+    for i in range(world_size):
+        assignments[i] = 1
+        remaining_layers -= 1
+
+    if remaining_layers < 0:
+        logger.warning(
+            "Fewer layers than nodes! Reducing to 1 layer per node where possible."
+        )
+        assignments = {i: 1 if i < total_layers else 0 for i in range(world_size)}
+        remaining_layers = 0
+
+    return assignments
+
+
+def _distribute_layers_by_bandwidth(
+    selected_cycle: list[NodeWithProfile],
+    assignments: dict[int, int],
+    remaining_layers: int,
+    model_meta: ModelMetadata,
+) -> None:
+    """Distribute remaining layers based on bandwidth and RAM capacity."""
+    indexed_nodes = list(enumerate(selected_cycle))
+    sorted_nodes = sorted(
+        indexed_nodes,
+        key=lambda x: x[1].node_profile.memory_bandwidth or 0,
+        reverse=True,
+    )
+
+    for original_idx, node in sorted_nodes:
+        if remaining_layers <= 0:
+            break
+
+        layer_size_bytes = model_meta.storage_size.in_bytes / model_meta.n_layers
+        max_layers_by_ram = int(
+            node.node_profile.memory.ram_available.in_bytes // layer_size_bytes
+        )
+        can_take = max(0, max_layers_by_ram - assignments[original_idx])
+        take = min(can_take, remaining_layers)
+        assignments[original_idx] += take
+        remaining_layers -= take
+
+    if remaining_layers > 0:
+        logger.warning(
+            "All nodes maxed out on RAM estimation, dumping remaining layers on fastest nodes."
+        )
+        for original_idx, _ in sorted_nodes:
+            assignments[original_idx] += 1
+            remaining_layers -= 1
+            if remaining_layers == 0:
+                break
+
+
+def _create_shard_assignments(
+    model_meta: ModelMetadata,
+    selected_cycle: list[NodeWithProfile],
+    assignments: dict[int, int],
+) -> ShardAssignments:
+    """Create shard assignments from layer assignments."""
+    world_size = len(selected_cycle)
+    runner_to_shard: dict[RunnerId, ShardMetadata] = {}
+    node_to_runner: dict[NodeId, RunnerId] = {}
+
+    current_start = 0
+    for i, node in enumerate(selected_cycle):
+        count = assignments[i]
+        runner_id = RunnerId()
+        shard = PipelineShardMetadata(
+            model_meta=model_meta,
+            device_rank=i,
+            world_size=world_size,
+            start_layer=current_start,
+            end_layer=current_start + count,
+            n_layers=model_meta.n_layers,
+        )
+        runner_to_shard[runner_id] = shard
+        node_to_runner[node.node_id] = runner_id
+        current_start += count
+
+    return ShardAssignments(
+        model_id=model_meta.model_id,
+        runner_to_shard=runner_to_shard,
+        node_to_runner=node_to_runner,
+    )
+
+
+def _assign_layers_by_bandwidth(
+    model_meta: ModelMetadata,
+    selected_cycle: list[NodeWithProfile],
+) -> ShardAssignments:
+    """Assign layers based on memory bandwidth."""
+    logger.info("Using bandwidth-aware shard assignment")
+
+    total_layers = model_meta.n_layers
+    world_size = len(selected_cycle)
+
+    assignments = _reserve_base_layers(world_size, total_layers)
+    remaining_layers = total_layers - sum(assignments.values())
+
+    if remaining_layers > 0:
+        _distribute_layers_by_bandwidth(
+            selected_cycle, assignments, remaining_layers, model_meta
+        )
+
+    return _create_shard_assignments(model_meta, selected_cycle, assignments)
+
+
+def get_shard_assignments_for_pipeline_parallel(
+    model_meta: ModelMetadata,
+    selected_cycle: list[NodeWithProfile],
+):
+    has_bandwidth = all(
+        node.node_profile.memory_bandwidth is not None for node in selected_cycle
+    )
+
+    if not has_bandwidth:
+        logger.info(
+            "Bandwidth data missing for some nodes, falling back to RAM-proportional assignment"
+        )
+        return _assign_layers_by_ram(model_meta, selected_cycle)
+
+    return _assign_layers_by_bandwidth(model_meta, selected_cycle)
 
 
 def get_shard_assignments_for_tensor_parallel(

src/exo/master/tests/test_placement_utils.py

@@ -397,3 +397,106 @@ def test_get_mlx_jaccl_coordinators(
     assert coordinators[node_c_id] == (
         f"{conn_c_a.send_back_multiaddr.ip_address}:5000"
     ), "node_c should use the IP from conn_c_a"
+
+
+def test_get_shard_assignments_bandwidth_aware(
+    topology: Topology,
+    create_node: Callable[[int, NodeId | None], NodeInfo],
+    create_connection: Callable[[NodeId, NodeId], Connection],
+):
+    # arrange
+    node_a_id = NodeId()
+    node_b_id = NodeId()
+    node_c_id = NodeId()
+
+    # Create nodes with identical RAM (plenty of it)
+    # Using 1GB to ensure no RAM constraints (model is small)
+    node_a = create_node(1024 * 1024 * 1024, node_a_id)
+    node_b = create_node(1024 * 1024 * 1024, node_b_id)
+    node_c = create_node(1024 * 1024 * 1024, node_c_id)
+
+    # Set Bandwidths: A=400 (Fastest), B=200, C=100 (Slowest)
+    assert node_a.node_profile is not None
+    assert node_b.node_profile is not None
+    assert node_c.node_profile is not None
+
+    node_a.node_profile.memory_bandwidth = 400_000_000_000
+    node_b.node_profile.memory_bandwidth = 200_000_000_000
+    node_c.node_profile.memory_bandwidth = 100_000_000_000
+
+    topology.add_node(node_a)
+    topology.add_node(node_b)
+    topology.add_node(node_c)
+
+    topology.add_connection(create_connection(node_a_id, node_b_id))
+    topology.add_connection(create_connection(node_b_id, node_c_id))
+    topology.add_connection(create_connection(node_c_id, node_a_id))
+
+    # Needs full cycle edges for get_cycles/get_shard_assignments if strict?
+    # Actually get_cycles just looks for cycles.
+    # But let's follow the pattern of other tests if they add bidirectional.
+    # checking test_filter_cycles_by_memory, it adds both directions.
+    topology.add_connection(create_connection(node_b_id, node_a_id))
+    topology.add_connection(create_connection(node_c_id, node_b_id))
+    topology.add_connection(create_connection(node_a_id, node_c_id))
+
+    model_meta = ModelMetadata(
+        model_id=ModelId("test-model"),
+        pretty_name="Test Model",
+        n_layers=30,  # 30 layers
+        storage_size=Memory.from_kb(
+            300
+        ),  # 10KB per layer. Nodes have 100MB RAM (100*1024 in create_node usually means KB? other tests use 1000*1024).
+        # create_node arg is likely KB or Bytes.
+        # test_filter_cycles_by_memory: create_node(1000 * 1024, ...) -> Memory.from_bytes(1) passes.
+        # Let's assume create_node takes Bytes or KB consistently.
+        # If I give 100*1024*1024 bytes = 100MB.
+        # Model storage = 300KB.
+        # So capacity is definitely not an issue.
+        hidden_size=1000,
+        supports_tensor=True,
+    )
+
+    cycles = topology.get_cycles()
+    # Depending on how get_cycles works and order of addition, we might get multiple cycles.
+    # filtering by memory usually done in master.
+    # Here we just pick one.
+    selected_cycle = cycles[0]
+
+    # act
+    shard_assignments = get_shard_assignments(
+        model_meta, selected_cycle, Sharding.Pipeline
+    )
+
+    # assert
+    runner_id_a = shard_assignments.node_to_runner[node_a_id]
+    runner_id_b = shard_assignments.node_to_runner[node_b_id]
+    runner_id_c = shard_assignments.node_to_runner[node_c_id]
+
+    # Get layer counts
+    layers_a = (
+        shard_assignments.runner_to_shard[runner_id_a].end_layer
+        - shard_assignments.runner_to_shard[runner_id_a].start_layer
+    )
+    layers_b = (
+        shard_assignments.runner_to_shard[runner_id_b].end_layer
+        - shard_assignments.runner_to_shard[runner_id_b].start_layer
+    )
+    layers_c = (
+        shard_assignments.runner_to_shard[runner_id_c].end_layer
+        - shard_assignments.runner_to_shard[runner_id_c].start_layer
+    )
+
+    # Check total
+    assert layers_a + layers_b + layers_c == 30
+
+    # Check that the fastest node (A with 400GB/s) gets saturated first.
+    # With strict greedy assignment and plenty of RAM:
+    # 1. Reserve: A=1, B=1, C=1. Remaining=27.
+    # 2. Sort: [A, B, C]
+    # 3. A takes min(remaining=27, capacity=huge) = 27.
+    # 4. A=28, B=1, C=1.
+
+    assert layers_a == 28
+    assert layers_b == 1
+    assert layers_c == 1

src/exo/shared/logging.py

@@ -29,6 +29,11 @@ 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_pyo3_bindings").setLevel(logging.WARNING)
+    logging.getLogger("httpx").setLevel(logging.WARNING)
+    logging.getLogger("httpcore").setLevel(logging.WARNING)
+
     logger.remove()
 
     # replace all stdlib loggers with _InterceptHandlers that log to loguru

src/exo/shared/models/model_cards.py

@@ -14,32 +14,6 @@ class ModelCard(CamelCaseModel):
 
 MODEL_CARDS: dict[str, ModelCard] = {
     # deepseek v3
-    # "deepseek-v3-0324:4bit": ModelCard(
-    #     short_id="deepseek-v3-0324:4bit",
-    #     model_id="mlx-community/DeepSeek-V3-0324-4bit",
-    #     name="DeepSeek V3 0324 (4-bit)",
-    #     description="""DeepSeek V3 is a large language model trained on the DeepSeek V3 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-V3-0324-4bit"),
-    #         pretty_name="DeepSeek V3 0324 (4-bit)",
-    #         storage_size=Memory.from_kb(409706307),
-    #         n_layers=61,
-    #     ),
-    # ),
-    # "deepseek-v3-0324": ModelCard(
-    #     short_id="deepseek-v3-0324",
-    #     model_id="mlx-community/DeepSeek-v3-0324-8bit",
-    #     name="DeepSeek V3 0324 (8-bit)",
-    #     description="""DeepSeek V3 is a large language model trained on the DeepSeek V3 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-v3-0324-8bit"),
-    #         pretty_name="DeepSeek V3 0324 (8-bit)",
-    #         storage_size=Memory.from_kb(754706307),
-    #         n_layers=61,
-    #     ),
-    # ),
     "deepseek-v3.1-4bit": ModelCard(
         short_id="deepseek-v3.1-4bit",
         model_id=ModelId("mlx-community/DeepSeek-V3.1-4bit"),
@@ -70,65 +44,6 @@ MODEL_CARDS: dict[str, ModelCard] = {
             supports_tensor=True,
         ),
     ),
-    # "deepseek-v3.2": ModelCard(
-    #     short_id="deepseek-v3.2",
-    #     model_id=ModelId("mlx-community/DeepSeek-V3.2-8bit"),
-    #     name="DeepSeek V3.2 (8-bit)",
-    #     description="""DeepSeek V3.2 is a large language model trained on the DeepSeek V3.2 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-V3.2-8bit"),
-    #         pretty_name="DeepSeek V3.2 (8-bit)",
-    #         storage_size=Memory.from_kb(754706307),
-    #         n_layers=61,
-    #         hidden_size=7168,
-    #         supports_tensor=True,
-    #     ),
-    # ),
-    # "deepseek-v3.2-4bit": ModelCard(
-    #     short_id="deepseek-v3.2-4bit",
-    #     model_id=ModelId("mlx-community/DeepSeek-V3.2-4bit"),
-    #     name="DeepSeek V3.2 (4-bit)",
-    #     description="""DeepSeek V3.2 is a large language model trained on the DeepSeek V3.2 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-V3.2-4bit"),
-    #         pretty_name="DeepSeek V3.2 (4-bit)",
-    #         storage_size=Memory.from_kb(754706307 // 2),  # TODO !!!!!
-    #         n_layers=61,
-    #         hidden_size=7168,
-    #         supports_tensor=True,
-    #     ),
-    # ),
-    # deepseek r1
-    # "deepseek-r1-0528-4bit": ModelCard(
-    #     short_id="deepseek-r1-0528-4bit",
-    #     model_id="mlx-community/DeepSeek-R1-0528-4bit",
-    #     name="DeepSeek-R1-0528 (4-bit)",
-    #     description="""DeepSeek R1 is a large language model trained on the DeepSeek R1 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-R1-0528-4bit"),
-    #         pretty_name="DeepSeek R1 671B (4-bit)",
-    #         storage_size=Memory.from_kb(409706307),
-    #         n_layers=61,
-    #         hidden_size=7168,
-    #     ),
-    # ),
-    # "deepseek-r1-0528": ModelCard(
-    #     short_id="deepseek-r1-0528",
-    #     model_id="mlx-community/DeepSeek-R1-0528-8bit",
-    #     name="DeepSeek-R1-0528 (8-bit)",
-    #     description="""DeepSeek R1 is a large language model trained on the DeepSeek R1 dataset.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/DeepSeek-R1-0528-8bit"),
-    #         pretty_name="DeepSeek R1 671B (8-bit)",
-    #         storage_size=Memory.from_bytes(754998771712),
-    #         n_layers=61,
-    # .       hidden_size=7168,
-    #     ),
-    # ),
     # kimi k2
     "kimi-k2-instruct-4bit": ModelCard(
         short_id="kimi-k2-instruct-4bit",
@@ -510,23 +425,24 @@ MODEL_CARDS: dict[str, ModelCard] = {
             supports_tensor=True,
         ),
     ),
-    "gpt-oss-20b-4bit": ModelCard(
-        short_id="gpt-oss-20b-4bit",
-        model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q4"),
-        name="GPT-OSS 20B (MXFP4-Q4, MLX)",
-        description="""OpenAI's GPT-OSS 20B is a medium-sized MoE model for lower-latency and local or specialized use cases; this MLX variant uses MXFP4 4-bit quantization.""",
+    "gpt-oss-20b-MXFP4-Q8": ModelCard(
+        short_id="gpt-oss-20b-MXFP4-Q8",
+        model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
+        name="GPT-OSS 20B (MXFP4-Q8, MLX)",
+        description="""OpenAI's GPT-OSS 20B is a medium-sized MoE model for lower-latency and local or specialized use cases; this variant is a 4-bit MLX conversion for Apple Silicon.""",
         tags=[],
         metadata=ModelMetadata(
-            model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q4"),
-            pretty_name="GPT-OSS 20B (MXFP4-Q4, MLX)",
+            model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
+            pretty_name="GPT-OSS 20B (MXFP4-Q8, MLX)",
             storage_size=Memory.from_kb(11_744_051),
             n_layers=24,
             hidden_size=2880,
             supports_tensor=True,
         ),
     ),
-    # Needs to be quantized g32 or g16.
+    # glm 4.5
     "glm-4.5-air-8bit": ModelCard(
+        # Needs to be quantized g32 or g16 to work with tensor parallel
         short_id="glm-4.5-air-8bit",
         model_id=ModelId("mlx-community/GLM-4.5-Air-8bit"),
         name="GLM 4.5 Air 8bit",
@@ -556,6 +472,7 @@ MODEL_CARDS: dict[str, ModelCard] = {
             supports_tensor=True,
         ),
     ),
+    # glm 4.7
     "glm-4.7-4bit": ModelCard(
         short_id="glm-4.7-4bit",
         model_id=ModelId("mlx-community/GLM-4.7-4bit"),
@@ -601,6 +518,7 @@ MODEL_CARDS: dict[str, ModelCard] = {
             supports_tensor=True,
         ),
     ),
+    # minimax-m2
     "minimax-m2.1-8bit": ModelCard(
         short_id="minimax-m2.1-8bit",
         model_id=ModelId("mlx-community/MiniMax-M2.1-8bit"),
@@ -631,19 +549,4 @@ MODEL_CARDS: dict[str, ModelCard] = {
             supports_tensor=True,
         ),
     ),
-    # "devstral-2-123b-instruct-2512-8bit": ModelCard(
-    #     short_id="devstral-2-123b-instruct-2512-8bit",
-    #     model_id=ModelId("mlx-community/Devstral-2-123B-Instruct-2512-8bit"),
-    #     name="Devstral 2 123B Instruct 2512 (8-bit, MLX)",
-    #     description="""Mistral AI's Devstral 2 123B Instruct (2512) is an agentic coding model.""",
-    #     tags=[],
-    #     metadata=ModelMetadata(
-    #         model_id=ModelId("mlx-community/Devstral-2-123B-Instruct-2512-8bit"),
-    #         pretty_name="Devstral 2 123B Instruct 2512 (8-bit, MLX)",
-    #         storage_size=Memory.from_kb(133_000_000),
-    #         n_layers=88,
-    #         hidden_size=12288,
-    #         supports_tensor=True,
-    #     ),
-    # ),
 }

src/exo/shared/types/profiling.py

@@ -57,6 +57,7 @@ class NodePerformanceProfile(CamelCaseModel):
     chip_id: str
     friendly_name: str
     memory: MemoryPerformanceProfile
+    memory_bandwidth: int | None = None
     network_interfaces: list[NetworkInterfaceInfo] = []
     system: SystemPerformanceProfile
 

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

@@ -20,6 +20,7 @@ except ImportError:
 
 from mlx_lm.models.cache import KVCache, QuantizedKVCache, RotatingKVCache
 from mlx_lm.models.deepseek_v3 import DeepseekV3Model
+from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.worker.engines.mlx.constants import (
@@ -365,6 +366,8 @@ def apply_chat_template(
         tools=chat_task_data.tools,
     )
 
+    logger.info(prompt)
+
     return prompt
 
 
@@ -396,6 +399,11 @@ def make_kv_cache(
 ) -> list[KVCache | RotatingKVCache | QuantizedKVCache]:
     assert hasattr(model, "layers")
 
+    # TODO: Do this for all models
+    if hasattr(model, "make_cache") and isinstance(model, GptOssModel):
+        logger.info("Using MLX LM's make cache")
+        return model.make_cache()  # type: ignore
+
     if max_kv_size is None:
         if KV_CACHE_BITS is None:
             logger.info("Using default KV cache")

src/exo/worker/runner/runner.py

@@ -1,6 +1,15 @@
 import time
+from collections.abc import Generator
+from functools import cache
 
 import mlx.core as mx
+from mlx_lm.models.gpt_oss import Model as GptOssModel
+from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
+    HarmonyEncodingName,
+    Role,
+    StreamableParser,
+    load_harmony_encoding,
+)
 
 from exo.shared.types.api import ChatCompletionMessageText
 from exo.shared.types.chunks import TokenChunk
@@ -153,11 +162,19 @@ def main(
                     _check_for_debug_prompts(task_params.messages[0].content)
 
                     # Generate responses using the actual MLX generation
-                    for response in mlx_generate(
+                    mlx_generator = mlx_generate(
                         model=model,
                         tokenizer=tokenizer,
                         task=task_params,
-                    ):
+                    )
+
+                    # GPT-OSS specific parsing to match other model formats.
+                    if isinstance(model, GptOssModel):
+                        mlx_generator = parse_gpt_oss(mlx_generator)
+
+                    # TODO: Add tool call parser here
+
+                    for response in mlx_generator:
                         match response:
                             case GenerationResponse():
                                 if shard_metadata.device_rank == 0:
@@ -207,6 +224,43 @@ def main(
                 break
 
 
+@cache
+def get_gpt_oss_encoding():
+    encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
+    return encoding
+
+
+def parse_gpt_oss(
+    responses: Generator[GenerationResponse],
+) -> Generator[GenerationResponse]:
+    encoding = get_gpt_oss_encoding()
+    stream = StreamableParser(encoding, role=Role.ASSISTANT)
+    thinking = False
+
+    for response in responses:
+        stream.process(response.token)
+
+        delta = stream.last_content_delta
+        ch = stream.current_channel
+
+        if ch == "analysis" and not thinking:
+            thinking = True
+            yield response.model_copy(update={"text": "<think>"})
+
+        if ch != "analysis" and thinking:
+            thinking = False
+            yield response.model_copy(update={"text": "</think>"})
+
+        if delta:
+            yield response.model_copy(update={"text": delta})
+
+        if response.finish_reason is not None:
+            if thinking:
+                yield response.model_copy(update={"text": "</think>"})
+            yield response
+            break
+
+
 EXO_RUNNER_MUST_FAIL = "EXO RUNNER MUST FAIL"
 EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
 EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"

src/exo/worker/utils/net_profile.py

@@ -1,49 +1,64 @@
-import http.client
-
-from anyio import create_task_group, to_thread
+import anyio
+import httpx
+from anyio import create_task_group
 from loguru import logger
 
 from exo.shared.topology import Topology
 from exo.shared.types.common import NodeId
 
+REACHABILITY_ATTEMPTS = 3
+
 
 async def check_reachability(
     target_ip: str,
     expected_node_id: NodeId,
-    self_node_id: NodeId,
     out: dict[NodeId, set[str]],
+    client: httpx.AsyncClient,
 ) -> None:
     """Check if a node is reachable at the given IP and verify its identity."""
+    if ":" in target_ip:
+        # TODO: use real IpAddress types
+        target_ip = f"[{target_ip}]"
+    url = f"http://{target_ip}:52415/node_id"
 
-    def _fetch_remote_node_id() -> NodeId | None:
-        connection = http.client.HTTPConnection(target_ip, 52415, timeout=1)
+    remote_node_id = None
+    last_error = None
+
+    for _ in range(REACHABILITY_ATTEMPTS):
         try:
-            connection.request("GET", "/node_id")
-            response = connection.getresponse()
-            if response.status != 200:
-                return None
+            r = await client.get(url)
+            if r.status_code != 200:
+                await anyio.sleep(1)
+                continue
 
-            body = response.read().decode("utf-8").strip()
+            body = r.text.strip().strip('"')
+            if not body:
+                await anyio.sleep(1)
+                continue
 
-            # Strip quotes if present (JSON string response)
-            if body.startswith('"') and body.endswith('"') and len(body) >= 2:
-                body = body[1:-1]
+            remote_node_id = NodeId(body)
+            break
 
-            return NodeId(body) or None
-        except OSError:
-            return None
-        except http.client.HTTPException:
-            return None
-        finally:
-            connection.close()
+        # expected failure cases
+        except (
+            httpx.TimeoutException,
+            httpx.NetworkError,
+        ):
+            await anyio.sleep(1)
+
+        # other failures should be logged on last attempt
+        except httpx.HTTPError as e:
+            last_error = e
+            await anyio.sleep(1)
+
+    if last_error is not None:
+        logger.warning(
+            f"connect error {type(last_error).__name__} from {target_ip} after {REACHABILITY_ATTEMPTS} attempts; treating as down"
+        )
 
-    remote_node_id = await to_thread.run_sync(_fetch_remote_node_id)
     if remote_node_id is None:
         return
 
-    if remote_node_id == self_node_id:
-        return
-
     if remote_node_id != expected_node_id:
         logger.warning(
             f"Discovered node with unexpected node_id; "
@@ -61,18 +76,33 @@ async def check_reachable(
     topology: Topology, self_node_id: NodeId
 ) -> dict[NodeId, set[str]]:
     """Check which nodes are reachable and return their IPs."""
+
     reachable: dict[NodeId, set[str]] = {}
-    async with create_task_group() as tg:
+
+    # these are intentionally httpx's defaults so we can tune them later
+    timeout = httpx.Timeout(timeout=5.0)
+    limits = httpx.Limits(
+        max_connections=100,
+        max_keepalive_connections=20,
+        keepalive_expiry=5,
+    )
+
+    async with (
+        httpx.AsyncClient(timeout=timeout, limits=limits) as client,
+        create_task_group() as tg,
+    ):
         for node in topology.list_nodes():
             if not node.node_profile:
                 continue
+            if node.node_id == self_node_id:
+                continue
             for iface in node.node_profile.network_interfaces:
                 tg.start_soon(
                     check_reachability,
                     iface.ip_address,
                     node.node_id,
-                    self_node_id,
                     reachable,
+                    client,
                 )
 
     return reachable

src/exo/worker/utils/profile.py

@@ -4,6 +4,7 @@ import platform
 from typing import Any, Callable, Coroutine
 
 import anyio
+from anyio import to_thread
 from loguru import logger
 
 from exo.shared.types.memory import Memory
@@ -24,8 +25,61 @@ from .system_info import (
     get_friendly_name,
     get_model_and_chip,
     get_network_interfaces,
+    profile_memory_bandwidth,
 )
 
+# Module-level cache for memory bandwidth (doesn't change at runtime)
+_cached_bandwidth: int | None = None
+_bandwidth_profiled: bool = False
+_bandwidth_profiling_task: asyncio.Task[int | None] | None = None
+
+
+async def profile_bandwidth_once() -> int | None:
+    """Profile bandwidth once in a background thread and cache the result.
+
+    This function is non-blocking - it runs the profiling in a thread pool.
+    Subsequent calls return the cached result immediately.
+    """
+    global _cached_bandwidth, _bandwidth_profiled, _bandwidth_profiling_task
+
+    # Already profiled, return cached value
+    if _bandwidth_profiled:
+        return _cached_bandwidth
+
+    # Profiling already in progress, wait for it
+    if _bandwidth_profiling_task is not None:
+        return await _bandwidth_profiling_task
+
+    # Start profiling in background thread
+    async def _do_profile() -> int | None:
+        global _cached_bandwidth, _bandwidth_profiled
+        try:
+            logger.info("Starting memory bandwidth profiling in background thread...")
+            bandwidth = await to_thread.run_sync(profile_memory_bandwidth, cancellable=True)
+            _cached_bandwidth = bandwidth
+            _bandwidth_profiled = True
+            if bandwidth:
+                logger.info(f"Memory bandwidth profiled: {bandwidth / 1e9:.1f} GB/s")
+            else:
+                logger.warning("Memory bandwidth profiling returned None")
+            return bandwidth
+        except Exception as e:
+            logger.opt(exception=e).error("Memory bandwidth profiling failed")
+            _bandwidth_profiled = True  # Mark as done to avoid retrying
+            return None
+
+    _bandwidth_profiling_task = asyncio.create_task(_do_profile())
+    return await _bandwidth_profiling_task
+
+
+def get_memory_bandwidth_cached() -> int | None:
+    """Return cached bandwidth or None if not yet profiled.
+
+    This is a non-blocking synchronous function that returns immediately.
+    Call profile_bandwidth_once() first to trigger profiling.
+    """
+    return _cached_bandwidth if _bandwidth_profiled else None
+
 
 async def get_metrics_async() -> Metrics | None:
     """Return detailed Metrics on macOS or a minimal fallback elsewhere."""
@@ -71,6 +125,8 @@ async def start_polling_node_metrics(
     callback: Callable[[NodePerformanceProfile], Coroutine[Any, Any, None]],
 ):
     poll_interval_s = 1.0
+    bandwidth_profile_started = False
+
     while True:
         try:
             metrics = await get_metrics_async()
@@ -85,6 +141,15 @@ async def start_polling_node_metrics(
             # do the memory profile last to get a fresh reading to not conflict with the other memory profiling loop
             memory_profile = get_memory_profile()
 
+            # Start bandwidth profiling in background on first poll (non-blocking)
+            if not bandwidth_profile_started:
+                bandwidth_profile_started = True
+                # Fire and forget - don't await, let it run in background
+                asyncio.create_task(profile_bandwidth_once())
+
+            # Use cached bandwidth (None until profiling completes)
+            memory_bandwidth = get_memory_bandwidth_cached()
+
             await callback(
                 NodePerformanceProfile(
                     model_id=model_id,
@@ -92,6 +157,7 @@ async def start_polling_node_metrics(
                     friendly_name=friendly_name,
                     network_interfaces=network_interfaces,
                     memory=memory_profile,
+                    memory_bandwidth=memory_bandwidth,
                     system=SystemPerformanceProfile(
                         gpu_usage=metrics.gpu_usage[1],
                         temp=metrics.temp.gpu_temp_avg,

src/exo/worker/utils/system_info.py

@@ -1,5 +1,6 @@
 import socket
 import sys
+import time
 from subprocess import CalledProcessError
 
 import psutil
@@ -81,3 +82,68 @@ async def get_model_and_chip() -> tuple[str, str]:
     chip = chip_line.split(": ")[1] if chip_line else "Unknown Chip"
 
     return (model, chip)
+
+
+def profile_memory_bandwidth() -> int | None:
+    """
+    Profile device memory bandwidth using MLX GPU operations.
+
+    Uses a large array copy on the GPU to measure unified memory bandwidth.
+    Returns measured bandwidth in bytes/second, or None if MLX is unavailable.
+    """
+    try:
+        import mlx.core as mx
+
+        if not mx.metal.is_available():
+            return None
+
+        # Use 2GB buffer to better saturate memory bandwidth
+        # Use 2D shape to avoid potential issues with very large 1D arrays
+        size_bytes = 2 * 1024 * 1024 * 1024
+        side = int((size_bytes // 4) ** 0.5)  # Square 2D array of float32
+        shape = (side, side)
+        actual_bytes = side * side * 4
+        bytes_transferred = actual_bytes * 2  # read + write
+
+        # Warm-up: run the full benchmark operation multiple times to stabilize GPU
+        for _ in range(3):
+            src = mx.random.uniform(shape=shape, dtype=mx.float32)
+            mx.eval(src)
+            dst = src + 0.0
+            mx.eval(dst)
+            mx.synchronize()
+            del src, dst
+
+        # Benchmark: measure time to copy array
+        best_bandwidth = 0.0
+        num_runs = 4
+
+        for _ in range(num_runs):
+            src = mx.random.uniform(shape=shape, dtype=mx.float32)
+            mx.eval(src)
+            mx.synchronize()
+
+            # Time the copy operation (src + 0.0 forces read of src, write of dst)
+            start = time.perf_counter()
+            dst = src + 0.0
+            mx.eval(dst)
+            mx.synchronize()
+            end = time.perf_counter()
+
+            bandwidth = bytes_transferred / (end - start)
+            best_bandwidth = max(best_bandwidth, bandwidth)
+
+            del src, dst
+
+        return int(best_bandwidth)
+    except Exception:
+        return None
+
+
+def get_memory_bandwidth(_chip_id: str) -> int | None:
+    """
+    Returns measured memory bandwidth in bytes/second.
+
+    Uses MLX GPU operations for accurate unified memory bandwidth measurement.
+    """
+    return profile_memory_bandwidth()