docs: request the mlx extra in the documented setup commands (#2245 )

Since #2087 moved mlx, mlx-lm, mlx-vlm and mflux out of `dependencies` into the `mlx` extra, the documented setup path never installs them, and `[tool.uv]` sets no default extras. `uv run exo` starts the API, then every runner crashes with `ModuleNotFoundError: No module named 'mlx'` (#2156). Adds the extra to the commands documented as the way to set up: - README macOS: `uv sync --extra mlx` - README Linux: `uv sync --extra mlx-cpu`, with the mlx-cuda12 / mlx-cuda13 alternatives - CONTRIBUTING.md quick start - `just sync` / `just sync-clean`, which `just build-app` runs before pyinstaller — the spec aborts when the mlx Metal libraries are missing Docs plus two justfile recipes; no dependency or lock changes. Backend choice stays explicit, since mlx-cpu / mlx-cuda12 / mlx-cuda13 / mlx-none are declared as conflicting extras. #2234 overlaps on the macOS README block only, via a setup script that runs `uv sync --extra mlx`; the Linux block, CONTRIBUTING.md and the justfile are not covered there.
Add Kimi K2.7-Code model card (official INT4 weights + vision) (#2167 )
2026-09-08 19:41:32 -04:00 · 2026-08-25 18:59:53 +00:00 · 2026-06-22 13:38:54 +00:00 · 2026-06-22 13:29:15 +00:00
41 changed files with 710 additions and 1431 deletions
@@ -29,6 +29,7 @@ To run EXO from source:
 git clone https://github.com/exo-explore/exo.git
 cd exo/dashboard
 npm install && npm run build && cd ..
+uv sync --extra mlx
 uv run exo
 ```

@@ -1007,7 +1007,6 @@ dependencies = [
 "serde_json",
 "tokio",
 "zenoh",
- "zenoh-ext",
 ]

 [[package]]
@@ -1843,12 +1842,6 @@ dependencies = [
 "spin 0.9.8",
 ]

-[[package]]
-name = "leb128"
-version = "0.2.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6cc46bac87ef8093eed6f272babb833b6443374399985ac8ed28471ee0918545"
-
 [[package]]
 name = "leb128fmt"
 version = "0.1.0"
@@ -5114,7 +5107,7 @@ dependencies = [
 [[package]]
 name = "zenoh"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "ahash",
 "arc-swap",
@@ -5164,7 +5157,7 @@ dependencies = [
 [[package]]
 name = "zenoh-buffers"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "zenoh-collections",
 ]
@@ -5172,7 +5165,7 @@ dependencies = [
 [[package]]
 name = "zenoh-codec"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "tracing",
 "uhlc",
@@ -5183,7 +5176,7 @@ dependencies = [
 [[package]]
 name = "zenoh-collections"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "ahash",
 ]
@@ -5191,7 +5184,7 @@ dependencies = [
 [[package]]
 name = "zenoh-config"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "json5",
 "nonempty-collections",
@@ -5216,7 +5209,7 @@ dependencies = [
 [[package]]
 name = "zenoh-core"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "lazy_static",
 "tokio",
@@ -5227,7 +5220,7 @@ dependencies = [
 [[package]]
 name = "zenoh-crypto"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "aes",
 "hmac",
@@ -5237,29 +5230,10 @@ dependencies = [
 "zenoh-result",
 ]

-[[package]]
-name = "zenoh-ext"
-version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
-dependencies = [
- "async-trait",
- "bincode",
- "flume",
- "futures",
- "leb128",
- "serde",
- "tokio",
- "tracing",
- "uhlc",
- "zenoh",
- "zenoh-macros",
- "zenoh-util",
-]
-
 [[package]]
 name = "zenoh-keyexpr"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "getrandom 0.2.17",
 "hashbrown 0.16.1",
@@ -5274,7 +5248,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "zenoh-config",
 "zenoh-link-commons",
@@ -5292,7 +5266,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-commons"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "base64",
@@ -5328,7 +5302,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-quic"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "rustls-webpki",
@@ -5344,7 +5318,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-quic_datagram"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "rustls-webpki",
@@ -5360,7 +5334,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-tcp"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "socket2 0.5.10",
@@ -5377,7 +5351,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-tls"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "base64",
@@ -5406,7 +5380,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-udp"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "libc",
@@ -5428,7 +5402,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-unixsock_stream"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "nix 0.29.0",
@@ -5446,7 +5420,7 @@ dependencies = [
 [[package]]
 name = "zenoh-link-ws"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "futures-util",
@@ -5466,7 +5440,7 @@ dependencies = [
 [[package]]
 name = "zenoh-macros"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5477,7 +5451,8 @@ dependencies = [
 [[package]]
 name = "zenoh-plugin-storage-manager"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b1d9fd55c490af4dc4abb6d51b620ecf6f824a9285a1f9446b59eeec6d66ccf7"
 dependencies = [
 "async-trait",
 "bincode",
@@ -5504,7 +5479,7 @@ dependencies = [
 [[package]]
 name = "zenoh-plugin-trait"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "git-version",
 "libloading",
@@ -5521,7 +5496,7 @@ dependencies = [
 [[package]]
 name = "zenoh-protocol"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "const_format",
 "rand 0.8.6",
@@ -5536,7 +5511,7 @@ dependencies = [
 [[package]]
 name = "zenoh-result"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "anyhow",
 ]
@@ -5544,7 +5519,7 @@ dependencies = [
 [[package]]
 name = "zenoh-runtime"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "lazy_static",
 "ron",
@@ -5558,7 +5533,7 @@ dependencies = [
 [[package]]
 name = "zenoh-sync"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "arc-swap",
 "event-listener",
@@ -5572,7 +5547,7 @@ dependencies = [
 [[package]]
 name = "zenoh-task"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "futures",
 "tokio",
@@ -5585,7 +5560,7 @@ dependencies = [
 [[package]]
 name = "zenoh-transport"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "crossbeam-utils",
@@ -5619,7 +5594,7 @@ dependencies = [
 [[package]]
 name = "zenoh-util"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "const_format",
@@ -5645,7 +5620,7 @@ dependencies = [
 [[package]]
 name = "zenoh_backend_traits"
 version = "1.9.0"
-source = "git+https://github.com/evanev7/zenoh.git?branch=exo#2fed58ef264a8415b29452cc37054b2e04c10396"
+source = "git+https://github.com/evanev7/zenoh.git?branch=exo#aa219b84e55dfbbf55ebd827b492626a470b8ac8"
 dependencies = [
 "async-trait",
 "const_format",
@@ -49,7 +49,6 @@ env_logger = "0.11.10"

 # networking
 zenoh = "=1.9.0"
-zenoh-ext = "=1.9.0"
 zenoh-plugin-storage-manager = { version = "=1.9.0", default-features = false }
 zenoh-plugin-trait = "=1.9.0"
 netwatcher = "0.6.0"
@@ -62,7 +61,6 @@ pidfile-rs = { git = "https://github.com/AndreiCravtov/pidfile-rs" }

 [patch.crates-io]
 zenoh = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
-zenoh-ext = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-buffers = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-codec = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-collections = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
@@ -88,7 +86,6 @@ zenoh-sync = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-task = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-transport = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh-util = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
-zenoh-plugin-storage-manager = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }
 zenoh_backend_traits = { git = "https://github.com/evanev7/zenoh.git", branch = "exo" }

 [workspace.lints.rust]
@@ -118,7 +118,7 @@ Then restart the Nix daemon: `sudo launchctl kickstart -k system/org.nixos.nix-d
    --force
  ```

-Clone the repo, build the dashboard, and run exo:
+Clone the repo, build the dashboard, install the dependencies, and run exo:

 ```bash
 # Clone exo
@@ -127,6 +127,9 @@ git clone https://github.com/exo-explore/exo
 # Build dashboard
 cd exo/dashboard && npm install && npm run build && cd ..

+# Install Python dependencies, including the MLX backend
+uv sync --extra mlx
+
 # Run exo
 uv run exo
 ```
@@ -176,7 +179,7 @@ rustup toolchain install nightly

 **Note:** The `macmon` package is macOS-only and not required for Linux.

-Clone the repo, build the dashboard, and run exo:
+Clone the repo, build the dashboard, install the dependencies, and run exo:

 ```bash
 # Clone exo
@@ -185,6 +188,10 @@ git clone https://github.com/exo-explore/exo
 # Build dashboard
 cd exo/dashboard && npm install && npm run build && cd ..

+# Install Python dependencies with the MLX backend for your hardware
+# (NVIDIA: --extra mlx-cuda13 or --extra mlx-cuda12)
+uv sync --extra mlx-cpu
+
 # Run exo
 uv run exo
 ```
@@ -8,6 +8,7 @@
 			"name": "exo-dashboard",
 			"version": "1.0.0",
 			"dependencies": {
+				"devalue": "^5.6.2",
 				"highlight.js": "^11.11.1",
 				"katex": "^0.16.27",
 				"marked": "^17.0.1",
@@ -2331,9 +2332,9 @@
 			}
 		},
 		"node_modules/devalue": {
-			"version": "5.5.0",
-			"resolved": "https://registry.npmjs.org/devalue/-/devalue-5.5.0.tgz",
-			"integrity": "sha512-69sM5yrHfFLJt0AZ9QqZXGCPfJ7fQjvpln3Rq5+PS03LD32Ost1Q9N+eEnaQwGRIriKkMImXD56ocjQmfjbV3w==",
+			"version": "5.6.2",
+			"resolved": "https://registry.npmjs.org/devalue/-/devalue-5.6.2.tgz",
+			"integrity": "sha512-nPRkjWzzDQlsejL1WVifk5rvcFi/y1onBRxjaFMjZeR9mFpqu2gmAZ9xUB9/IEanEP/vBtGeGganC/GO1fmufg==",
 			"license": "MIT"
 		},
 		"node_modules/enhanced-resolve": {
@@ -11,8 +11,6 @@
 		"check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json"
 	},
 	"devDependencies": {
-		"prettier": "^3.4.2",
-		"prettier-plugin-svelte": "^3.3.3",
 		"@sveltejs/adapter-static": "^3.0.10",
 		"@sveltejs/kit": "^2.48.4",
 		"@sveltejs/vite-plugin-svelte": "^5.0.0",
@@ -20,6 +18,8 @@
 		"@types/d3": "^7.4.3",
 		"@types/node": "^22",
 		"d3": "^7.9.0",
+		"prettier": "^3.4.2",
+		"prettier-plugin-svelte": "^3.3.3",
 		"svelte": "^5.0.0",
 		"svelte-check": "^4.0.0",
 		"tailwindcss": "^4.0.0",
@@ -28,6 +28,7 @@
 		"vite": "^6.0.0"
 	},
 	"dependencies": {
+		"devalue": "^5.6.2",
 		"highlight.js": "^11.11.1",
 		"katex": "^0.16.27",
 		"marked": "^17.0.1",
@@ -16,10 +16,10 @@ check:
    uv run basedpyright --project pyproject.toml

 sync:
-    uv sync --all-packages
+    uv sync --all-packages --extra mlx

 sync-clean:
-    uv sync --all-packages --force-reinstall --no-cache
+    uv sync --all-packages --extra mlx --force-reinstall --no-cache

 rust-rebuild:
    PYO3_PYTHON="$(uv run python -c 'import sys; print(sys.executable)')" cargo run --bin stub_gen
@@ -0,0 +1,36 @@
+model_id = "moonshotai/Kimi-K2.7-Code"
+n_layers = 61
+hidden_size = 7168
+num_key_value_heads = 64
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "kimi"
+quantization = ""
+base_model = "Kimi K2.7 Code"
+capabilities = ["text", "thinking", "thinking_toggle", "vision"]
+
+context_length = 262144
+backends = ["MlxMetal", "MlxCuda", "MlxCpu"]
+[storage_size]
+in_bytes = 595204986173
+
+# Vision tower + mm_projector extracted unmodified (bf16) from the official
+# repo, in the same format as exolabs/Kimi-K2.6-vision; extraction script
+# included in the weights repo. Vision config is identical to Kimi-K2.6's.
+[vision]
+image_token_id = 163605
+model_type = "kimi_vl"
+weights_repo = "aidiffuser/Kimi-K2.7-Code-vision"
+processor_repo = "moonshotai/Kimi-K2.7-Code"
+
+# Source: https://huggingface.co/moonshotai/Kimi-K2.7-Code
+# (recommends temperature 1.0 / top_p 0.95 for thinking mode, same as K2.6)
+[sampling_defaults]
+temperature = 1.0
+top_p = 0.95
+min_p = 0.01
+
+[sampling_defaults.non_thinking]
+temperature = 0.6
+top_p = 0.95
+min_p = 0.01
@@ -48,7 +48,6 @@ env_logger.workspace = true

 # Networking
 zenoh.workspace = true
-zenoh-ext = { workspace = true, features = ["unstable"] }
 rand.workspace = true
 serde_json.workspace = true
 parking_lot.workspace = true
@@ -2,21 +2,14 @@
 # ruff: noqa: E501, F401, F403, F405

 import builtins
-import collections.abc
 import os
 import pathlib
 import typing
 __all__ = [
    "FromSwarm",
-    "LVAggregator",
-    "LVPublisher",
-    "LVSubscriber",
    "NetworkingHandle",
    "Pidfile",
    "PidfileError",
-    "SessionHandle",
-    "Storage",
-    "StorageGetter",
 ]

 class FromSwarm:
@@ -38,18 +31,6 @@ class FromSwarm:
    
    ...

-@typing.final
-class LVAggregator:
-    def dump(self) -> builtins.dict[builtins.str, builtins.str]: ...
-
-@typing.final
-class LVPublisher:
-    def put(self, data: builtins.str) -> collections.abc.Awaitable[None]: ...
-
-@typing.final
-class LVSubscriber:
-    def recv(self) -> collections.abc.Awaitable[tuple[str, str] | None]: ...
-
@typing.final
 class NetworkingHandle:
    @staticmethod
@@ -130,24 +111,3 @@ class PidfileError(builtins.Exception):
    def __repr__(self) -> builtins.str: ...
    def __str__(self) -> builtins.str: ...

-@typing.final
-class SessionHandle:
-    @staticmethod
-    def new(identity: builtins.str, namespace: builtins.str, listen_port: builtins.int, discovery_service_port: builtins.int) -> tuple[SessionHandle, NetworkingHandle]: ...
-    def last_value_aggregator(self, prefix: builtins.str) -> LVAggregator: ...
-    def last_value_subscriber(self, kexpr: builtins.str) -> LVSubscriber: ...
-    def last_value_publisher(self, kexpr: builtins.str) -> LVPublisher: ...
-    def storage_interface(self) -> Storage: ...
-
-@typing.final
-class Storage:
-    def get(self, key: builtins.str) -> collections.abc.Awaitable[str | None]: ...
-    def get_many(self, key: builtins.str) -> StorageGetter: ...
-    def put(self, key: builtins.str, data: builtins.str) -> collections.abc.Awaitable[None]: ...
-    def delete(self, key: builtins.str) -> collections.abc.Awaitable[None]: ...
-    def dump(self, prefix: builtins.str) -> collections.abc.Awaitable[dict[str, str]]: ...
-
-@typing.final
-class StorageGetter:
-    def recv(self) -> collections.abc.Awaitable[tuple[str, str] | None]: ...
-
@@ -1,181 +0,0 @@
-use networking::{Session, liveliness_aggregator::LivelinessAggregator};
-use std::{
-    collections::{HashMap, HashSet},
-    sync::Arc,
-};
-use zenoh::{Result as ZResult, Wait};
-
-use parking_lot::Mutex;
-use pyo3::{
-    exceptions::{PyConnectionError, PyRuntimeError},
-    prelude::*,
-};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
-use zenoh::{
-    handlers::FifoChannelHandler,
-    sample::{Sample, SampleKind},
-};
-use zenoh_ext::{
-    AdvancedPublisher, AdvancedSubscriber, AdvancedSubscriberBuilderExt, HistoryConfig,
-};
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct LVAggregator {
-    pub prefix: Arc<str>,
-    pub store: Arc<Mutex<HashMap<String, String>>>,
-    pub current_live: LivelinessAggregator,
-}
-
-pub fn spawn_lv_aggregator_onto(session: &Session, prefix: Arc<str>) -> ZResult<LVAggregator> {
-    // nota bene: config must be kept in line with SessionHandle::last_value_receiver
-    let store = Arc::new(Mutex::new(HashMap::default()));
-    session
-        .z
-        //assuming all LV aggregators are prefix/node_id/atomic_json
-        .declare_subscriber(format!("{prefix}/*/*"))
-        .advanced()
-        .history(
-            HistoryConfig::default()
-                .max_samples(1)
-                .detect_late_publishers(),
-        )
-        .callback({
-            let store = Arc::clone(&store);
-            let prefix = Arc::clone(&prefix);
-            move |sample| {
-                if let Some(s) = sample
-                    .key_expr()
-                    .to_string()
-                    .strip_prefix(&*prefix)
-                    .and_then(|it| it.strip_prefix('/'))
-                {
-                    let s = s.to_string();
-                    match sample.kind() {
-                        SampleKind::Put => {
-                            store.lock().insert(
-                                s,
-                                sample
-                                    .payload()
-                                    .try_to_string()
-                                    .expect("we only use utf8 encoded strings. someone messed up")
-                                    .to_string(),
-                            );
-                        }
-                        SampleKind::Delete => {
-                            store.lock().remove(&s);
-                        }
-                    }
-                };
-            }
-        })
-        .background()
-        .wait()?;
-    Ok(LVAggregator {
-        prefix,
-        store,
-        current_live: session.liveliness_aggregator.clone(),
-    })
-}
-
-#[gen_stub_pymethods]
-#[pymethods]
-impl LVAggregator {
-    pub fn dump(&self) -> HashMap<String, String> {
-        let mut store = self.store.lock();
-        let currently_alive: HashSet<String> = self.current_live.dump();
-        // remove any keys that are no longer live
-        store.retain(|key, _| {
-            currently_alive.iter().any(|node_id| {
-                key.strip_prefix(node_id)
-                    .is_some_and(|rest| rest.starts_with("/"))
-            })
-        });
-        store.clone()
-    }
-}
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct LVSubscriber {
-    pub subscriber: AdvancedSubscriber<FifoChannelHandler<Sample>>,
-}
-#[gen_stub_pymethods]
-#[pymethods]
-impl LVSubscriber {
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[tuple[str, str] | None]",
-        imports=("collections.abc")
-    ))]
-    pub fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        pyo3_async_runtimes::tokio::future_into_py(py, {
-            if self.subscriber.receiver_count() != 1 {
-                return Err(PyRuntimeError::new_err(
-                    "tried to receive twice on the same receiver",
-                ));
-            }
-            let subscriber = self.subscriber.clone();
-            async move {
-                loop {
-                    match subscriber.recv_async().await {
-                        Ok(sample) if sample.kind() == SampleKind::Delete => continue,
-                        Err(_) => {
-                            return Ok(None);
-                        }
-                        Ok(sample) => {
-                            return Ok(Some((
-                                sample.key_expr().to_string(),
-                                sample
-                                    .payload()
-                                    .try_to_string()
-                                    .expect("we only use utf8 encoded strings. someone messed up")
-                                    .to_string(),
-                            )));
-                        }
-                    }
-                }
-            }
-        })
-    }
-}
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct LVPublisher {
-    pub state: Arc<AdvancedPublisher<'static>>,
-}
-impl LVPublisher {
-    pub fn new(publisher: AdvancedPublisher<'static>) -> Self {
-        Self {
-            state: Arc::new(publisher),
-        }
-    }
-}
-#[gen_stub_pymethods]
-#[pymethods]
-impl LVPublisher {
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[None]",
-        imports=("collections.abc")
-    ))]
-    pub fn put<'py>(&'py self, py: Python<'py>, data: String) -> PyResult<Bound<'py, PyAny>> {
-        let state = Arc::clone(&self.state);
-        pyo3_async_runtimes::tokio::future_into_py(py, {
-            // clone the data so py can have it back
-            async move {
-                state
-                    .put(data)
-                    .await
-                    .map_err(|e| PyConnectionError::new_err(e.to_string()))?;
-                Ok(())
-            }
-        })
-    }
-}
-
-pub fn lv_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
-    m.add_class::<LVPublisher>()?;
-    m.add_class::<LVSubscriber>()?;
-    m.add_class::<LVAggregator>()?;
-    Ok(())
-}
@@ -4,18 +4,13 @@
 //!
 //!

-pub mod allow_threading;
-pub mod last_value;
-pub mod networking;
-pub mod pidfile;
-pub mod session;
-mod storage;
+mod allow_threading;
+// mod ident;
+mod networking;
+mod pidfile;

-use crate::last_value::lv_submodule;
 use crate::networking::networking_submodule;
 use crate::pidfile::pidfile_submodule;
-use crate::session::session_submodule;
-use crate::storage::storage_submodule;
 use pyo3::prelude::PyModule;
 use pyo3::{Bound, PyResult, pymodule};
 use pyo3_stub_gen::define_stub_info_gatherer;
@@ -159,12 +154,15 @@ fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    builder.enable_all();
    pyo3_async_runtimes::tokio::init(builder);

-    // TODO: for now this is all NOT a submodule. KISS
+    // TODO: for now this is all NOT a submodule, but figure out how to make the submodule system
+    //       work with maturin, where the types generate correctly, in the right folder, without
+    //       too many importing issues...
    pidfile_submodule(m)?;
+    // m.add_class::<PyKeypair>()?;
    networking_submodule(m)?;
-    lv_submodule(m)?;
-    session_submodule(m)?;
-    storage_submodule(m)?;
+
+    // top-level constructs
+    // TODO: ...

    Ok(())
 }
@@ -1,111 +0,0 @@
-use networking::Session;
-use pyo3::{
-    exceptions::{PyConnectionError, PyRuntimeError, PyValueError},
-    prelude::*,
-};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
-use zenoh::Wait;
-use zenoh_ext::{
-    AdvancedPublisherBuilderExt, AdvancedSubscriberBuilderExt, CacheConfig, HistoryConfig,
-    MissDetectionConfig,
-};
-
-use crate::{
-    last_value::{LVAggregator, LVPublisher, LVSubscriber, spawn_lv_aggregator_onto},
-    networking::PyNetworkingHandle,
-    storage::Storage,
-};
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct SessionHandle {
-    pub session: Session,
-}
-
-#[gen_stub_pymethods]
-#[pymethods]
-impl SessionHandle {
-    #[staticmethod]
-    pub fn new<'py>(
-        identity: &str,
-        namespace: &str,
-        listen_port: u16,
-        discovery_service_port: u16,
-    ) -> PyResult<(SessionHandle, PyNetworkingHandle)> {
-        // get identity
-        if !identity
-            .chars()
-            .all(|c| ('0'..='9').contains(&c) || ('a'..='f').contains(&c))
-            || identity.len() > 32
-        {
-            return Err(PyValueError::new_err(format!(
-                "{identity} is not a valid zenoh identity"
-            )));
-        }
-
-        let cfg = networking::cfg(identity, listen_port).map_err(|e| {
-            PyValueError::new_err(format!("failed to write config: {}", e.to_string()))
-        })?;
-        let session = pyo3_async_runtimes::tokio::get_runtime()
-            .block_on(networking::open(
-                cfg,
-                namespace,
-                listen_port,
-                discovery_service_port,
-            ))
-            .map_err(|e| {
-                PyRuntimeError::new_err(format!(
-                    "failed to spawn networking on tokio runtime: {}",
-                    e.to_string()
-                ))
-            })?;
-        let legacy = PyNetworkingHandle::from_session(session.clone());
-        Ok((Self { session }, legacy))
-    }
-
-    pub fn last_value_aggregator(&self, prefix: String) -> PyResult<LVAggregator> {
-        spawn_lv_aggregator_onto(&self.session, prefix.into()).map_err(|e| {
-            PyConnectionError::new_err(format!("failed to spawn liveliness aggregator: {e}"))
-        })
-    }
-
-    pub fn last_value_subscriber(&self, kexpr: &str) -> PyResult<LVSubscriber> {
-        // nota bene: config must be kept in track with the LVAggregator
-        self.session
-            .z
-            .declare_subscriber(kexpr)
-            .advanced()
-            .history(
-                HistoryConfig::default()
-                    .max_samples(1)
-                    .detect_late_publishers(),
-            )
-            .wait()
-            .map_err(|e| PyConnectionError::new_err(format!("failed to declare subscriber: {e}")))
-            .map(|subscriber| LVSubscriber { subscriber })
-    }
-
-    pub fn last_value_publisher(&self, kexpr: String) -> PyResult<LVPublisher> {
-        self.session
-            .z
-            .declare_publisher(kexpr)
-            .advanced()
-            .publisher_detection()
-            .sample_miss_detection(MissDetectionConfig::default())
-            .cache(CacheConfig::default().max_samples(1))
-            .wait()
-            .map_err(|e| PyConnectionError::new_err(format!("failed to declare publisher: {e}")))
-            .map(LVPublisher::new)
-    }
-
-    pub fn storage_interface(&self) -> Storage {
-        Storage {
-            session: self.session.z.clone(),
-        }
-    }
-}
-
-pub fn session_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
-    m.add_class::<SessionHandle>()?;
-    Ok(())
-}
@@ -1,199 +0,0 @@
-use std::{collections::HashMap, time::Duration};
-
-use networking::STORAGE_PREFIX;
-use pyo3::{
-    exceptions::{PyConnectionError, PyRuntimeError, PyValueError},
-    prelude::*,
-};
-use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
-use zenoh::{
-    Session as ZSession, Wait, handlers::FifoChannelHandler, query::Reply, sample::SampleKind,
-};
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct Storage {
-    pub session: ZSession,
-}
-
-#[gen_stub_pymethods]
-#[pymethods]
-impl Storage {
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[str | None]",
-        imports=("collections.abc")
-    ))]
-    pub fn get<'py>(&'py self, py: Python<'py>, key: String) -> PyResult<Bound<'py, PyAny>> {
-        if key.contains('*') {
-            return Err(PyValueError::new_err(format!(
-                "{key} is invalid -- Storage.get only supports fixed keys"
-            )));
-        }
-        let session = self.session.clone();
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            let subscriber = session
-                .get(format!("{STORAGE_PREFIX}/{key}"))
-                //.allowed_destination(Locality::SessionLocal)
-                .await
-                .map_err(|e| PyConnectionError::new_err(format!("failed to query storage: {e}")))?;
-            tokio::select! {
-                _ = tokio::time::sleep(Duration::from_secs(1)) => {
-                    Ok(None)
-                }
-                reply = subscriber.recv_async() => {
-                    Ok(reply.ok()
-                        .and_then(|reply| reply.into_result().ok())
-                        .and_then(|sample| {
-                            if sample.kind() == SampleKind::Put {
-                                Some(sample
-                                    .payload()
-                                    .try_to_string()
-                                    .expect("we only use utf8 encoded strings. someone messed up").to_string()
-                                )
-                            } else { None }
-                        })
-                    )
-                }
-            }
-        })
-    }
-    pub fn get_many(&self, key: String) -> PyResult<StorageGetter> {
-        self.session
-            .get(key)
-            .wait()
-            .map_err(|e| PyConnectionError::new_err(format!("failed to query storage: {e}")))
-            .map(StorageGetter)
-    }
-
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[None]",
-        imports=("collections.abc")
-    ))]
-    pub fn put<'py>(
-        &'py self,
-        py: Python<'py>,
-        key: String,
-        data: String,
-    ) -> PyResult<Bound<'py, PyAny>> {
-        if key.contains('*') {
-            return Err(PyValueError::new_err(format!(
-                "{key} is invalid -- Storage.put only supports fixed keys"
-            )));
-        }
-        let session = self.session.clone();
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            session
-                .put(format!("{STORAGE_PREFIX}/{key}"), data)
-                .await
-                .map_err(|e| PyConnectionError::new_err(format!("failed to query storage: {e}")))
-        })
-    }
-
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[None]",
-        imports=("collections.abc")
-    ))]
-    pub fn delete<'py>(&'py self, py: Python<'py>, key: String) -> PyResult<Bound<'py, PyAny>> {
-        if key.contains('*') {
-            return Err(PyValueError::new_err(format!(
-                "{key} is invalid -- Storage.delete only supports fixed keys"
-            )));
-        }
-        let session = self.session.clone();
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            session
-                .delete(format!("{STORAGE_PREFIX}/{key}"))
-                .await
-                .map_err(|e| PyConnectionError::new_err(format!("failed to query storage: {e}")))
-        })
-    }
-
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[dict[str, str]]",
-        imports=("collections.abc")
-    ))]
-    pub fn dump<'py>(&'py self, py: Python<'py>, prefix: String) -> PyResult<Bound<'py, PyAny>> {
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            Ok(networking::read_raw_memory_storage()
-                .await
-                .into_iter()
-                .filter_map(|(key, value)| {
-                    Some((
-                        key?.as_str().strip_prefix(prefix.as_str())?.to_string(),
-                        value
-                            .payload
-                            .try_to_string()
-                            .expect("we only use utf8 encoded strings. someone messed up")
-                            .to_string(),
-                    ))
-                })
-                .collect::<HashMap<String, String>>())
-        })
-    }
-}
-
-#[gen_stub_pyclass]
-#[pyclass]
-pub struct StorageGetter(FifoChannelHandler<Reply>);
-
-#[gen_stub_pymethods]
-#[pymethods]
-impl StorageGetter {
-    #[gen_stub(override_return_type(
-        type_repr="collections.abc.Awaitable[tuple[str, str] | None]",
-        imports=("collections.abc")
-    ))]
-    fn recv<'py>(&'py self, py: Python<'py>) -> PyResult<Bound<'py, PyAny>> {
-        if self.0.receiver_count() != 1 {
-            return Err(PyRuntimeError::new_err(
-                "Tried to call StorageGetter.recv twice concurrently",
-            ));
-        }
-        let dupe = self.0.clone();
-        pyo3_async_runtimes::tokio::future_into_py(py, async move {
-            let sample = loop {
-                match dupe.recv_async().await {
-                    Err(_) => return Ok(None),
-                    Ok(reply) => match reply.into_result() {
-                        Err(e) => {
-                            log::warn!("Ignoring reply error: {e}");
-                            continue;
-                        }
-                        Ok(sample) => match sample.kind() {
-                            SampleKind::Put => break sample,
-                            SampleKind::Delete => {
-                                log::warn!(
-                                    "Received unexpected DELETE from queryable: {}",
-                                    sample.key_expr()
-                                );
-                                continue;
-                            }
-                        },
-                    },
-                };
-            };
-
-            let key = sample
-                .key_expr()
-                .to_string()
-                .strip_prefix(format!("{STORAGE_PREFIX}/").as_str())
-                .expect("invalid storage format encountered")
-                .to_string();
-
-            Ok(Some((
-                key,
-                sample
-                    .payload()
-                    .try_to_string()
-                    .expect("we only use utf8 encoded strings. someone messed up")
-                    .to_string(),
-            )))
-        })
-    }
-}
-
-pub fn storage_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
-    m.add_class::<Storage>()?;
-    m.add_class::<StorageGetter>()?;
-    Ok(())
-}
@@ -1,168 +0,0 @@
-use std::time::{Duration, SystemTime, UNIX_EPOCH};
-
-use zenoh_ext::{
-    AdvancedPublisherBuilderExt, AdvancedSubscriber, AdvancedSubscriberBuilderExt, CacheConfig,
-    HistoryConfig, MissDetectionConfig,
-};
-
-use zenoh::{handlers::FifoChannelHandler, sample::Sample};
-
-// Adjust these imports to your crate/module paths.
-use exo_rs::{
-    last_value::{LVPublisher, LVSubscriber},
-    session::SessionHandle,
-};
-async fn expect_two_values(
-    sub: &AdvancedSubscriber<FifoChannelHandler<Sample>>,
-    key_a: &str,
-    val_a: &str,
-    key_b: &str,
-    val_b: &str,
-) {
-    use std::collections::HashMap;
-    use tokio::time::{Duration, Instant, timeout};
-    use zenoh::sample::SampleKind;
-
-    let deadline = Instant::now() + Duration::from_secs(10);
-    let mut seen: HashMap<String, String> = HashMap::new();
-
-    loop {
-        if seen.get(key_a).map(String::as_str) == Some(val_a)
-            && seen.get(key_b).map(String::as_str) == Some(val_b)
-        {
-            return;
-        }
-
-        let remaining = deadline.saturating_duration_since(Instant::now());
-        assert!(
-            !remaining.is_zero(),
-            "timed out waiting for both historical samples; expected {key_a}={val_a}, {key_b}={val_b}; seen = {seen:?}"
-        );
-
-        match timeout(remaining.min(Duration::from_millis(750)), sub.recv_async()).await {
-            Ok(Ok(sample)) => {
-                if sample.kind() == SampleKind::Delete {
-                    continue;
-                }
-
-                let key = sample.key_expr().to_string();
-                let value = sample
-                    .payload()
-                    .try_to_string()
-                    .expect("payload should be UTF-8")
-                    .to_string();
-
-                if key == key_a || key == key_b {
-                    eprintln!("received relevant {key} = {value}");
-                    seen.insert(key, value);
-                } else {
-                    eprintln!("received unrelated {key} = {value}");
-                }
-            }
-            Ok(Err(e)) => panic!("subscriber receive failed: {e}"),
-            Err(_) => {}
-        }
-    }
-}
-
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn lv_subscriber_receives_last_value_from_multiple_publishers() {
-    let cfg =
-        networking::cfg(&format!("{:x}", rand::random::<u128>()), 52414).expect("create config");
-    let n_session = networking::open(cfg, "exo", 52414, 52413)
-        .await
-        .expect("open session");
-
-    let session = SessionHandle { session: n_session };
-
-    let run_id = SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .unwrap()
-        .as_nanos();
-
-    let base = format!("zenoh_advanced_history_test/{run_id}");
-
-    let key_a = format!("{base}/a");
-    let key_b = format!("{base}/b");
-    let sub_key = format!("{base}/*");
-
-    let pub1: LVPublisher = session
-        .last_value_publisher(key_a.clone())
-        .expect("declare LV publisher a");
-
-    pub1.state.put("aa").await.expect("publish aa");
-
-    let pub2: LVPublisher = session
-        .last_value_publisher(key_b.clone())
-        .expect("declare LV publisher b");
-
-    pub2.state.put("bb").await.expect("publish bb");
-
-    // Let publisher detection / cache metadata settle before the late subscriber joins.
-    tokio::time::sleep(Duration::from_millis(250)).await;
-
-    let sub: LVSubscriber = session
-        .last_value_subscriber(&*sub_key)
-        .expect("declare LV subscriber");
-
-    expect_two_values(&sub.subscriber, &*key_a, "aa", &*key_b, "bb").await
-}
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn wildcard_advanced_subscriber_receives_history_from_both_publishers() {
-    let cfg =
-        networking::cfg(&format!("{:x}", rand::random::<u128>()), 52412).expect("create config");
-    let n_session = networking::open(cfg, "exo", 52412, 52411)
-        .await
-        .expect("open session");
-    let session = n_session.z.clone();
-
-    // Unique prefix so the wildcard subscriber cannot accidentally see unrelated traffic.
-    let run_id = SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .unwrap()
-        .as_nanos();
-
-    let base = format!("zenoh_advanced_history_test/{run_id}");
-    let key_a = format!("{base}/a");
-    let key_b = format!("{base}/b");
-    let sub_key = format!("{base}/*");
-
-    let pub1 = session
-        .declare_publisher(key_a.clone())
-        .advanced()
-        .publisher_detection()
-        .sample_miss_detection(MissDetectionConfig::default())
-        .cache(CacheConfig::default().max_samples(1))
-        .await
-        .expect("declare advanced publisher a");
-
-    pub1.put("aa").await.expect("publish aa");
-
-    let pub2 = session
-        .declare_publisher(key_b.clone())
-        .advanced()
-        .sample_miss_detection(MissDetectionConfig::default())
-        .publisher_detection()
-        .cache(CacheConfig::default().max_samples(1))
-        .await
-        .expect("declare advanced publisher b");
-
-    pub2.put("bb").await.expect("publish bb");
-
-    // Give liveliness/cache declarations a brief chance to settle before declaring
-    // the late-joining advanced subscriber.
-    tokio::time::sleep(Duration::from_millis(250)).await;
-
-    let sub = session
-        .declare_subscriber(sub_key)
-        .advanced()
-        .history(
-            HistoryConfig::default()
-                .max_samples(1)
-                .detect_late_publishers(),
-        )
-        .await
-        .expect("declare advanced subscriber");
-
-    expect_two_values(&sub, &*key_a, "aa", &*key_b, "bb").await
-}
@@ -1,51 +1,54 @@
-use core::mem::drop;
-use core::option::Option::Some;
-use core::time::Duration;
-use tokio;
-use tokio::sync::mpsc;
+#[cfg(test)]
+mod tests {
+    use core::mem::drop;
+    use core::option::Option::Some;
+    use core::time::Duration;
+    use tokio;
+    use tokio::sync::mpsc;

-#[tokio::test]
-async fn test_drop_channel() {
-    struct Ping;
+    #[tokio::test]
+    async fn test_drop_channel() {
+        struct Ping;

-    let (tx, mut rx) = mpsc::channel::<Ping>(10);
+        let (tx, mut rx) = mpsc::channel::<Ping>(10);

-    let _ = tokio::spawn(async move {
-        println!("TASK: entered");
+        let _ = tokio::spawn(async move {
+            println!("TASK: entered");

-        loop {
-            tokio::select! {
-                result = rx.recv() => {
-                    match result {
-                        Some(_) => {
-                            println!("TASK: pinged");
-                        }
-                        None => {
-                            println!("TASK: closing channel");
-                            break;
+            loop {
+                tokio::select! {
+                    result = rx.recv() => {
+                        match result {
+                            Some(_) => {
+                                println!("TASK: pinged");
+                            }
+                            None => {
+                                println!("TASK: closing channel");
+                                break;
+                            }
                        }
                    }
-                }
-                _ = tokio::time::sleep(Duration::from_secs_f32(0.1)) => {
-                    println!("TASK: heartbeat");
+                    _ = tokio::time::sleep(Duration::from_secs_f32(0.1)) => {
+                        println!("TASK: heartbeat");
+                    }
                }
            }
-        }

-        println!("TASK: exited");
-    });
+            println!("TASK: exited");
+        });

-    let tx2 = tx.clone();
+        let tx2 = tx.clone();

-    tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;

-    tx.send(Ping).await.expect("Should not fail");
-    drop(tx);
+        tx.send(Ping).await.expect("Should not fail");
+        drop(tx);

-    tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;

-    tx2.send(Ping).await.expect("Should not fail");
-    drop(tx2);
+        tx2.send(Ping).await.expect("Should not fail");
+        drop(tx2);

-    tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+    }
 }
@@ -1,69 +0,0 @@
-import os
-import uuid
-
-import pytest
-from exo_rs import SessionHandle
-
-
-ZENOH_PORT = 52414
-DISCOVERY_PORT = 52413
-
-
-@pytest.fixture(scope="module")
-def storage():
-    node_id = os.urandom(16).hex().rstrip("0")
-
-    session_handle, _nh = SessionHandle.new(
-        node_id,
-        ZENOH_PORT,
-        DISCOVERY_PORT,
-    )
-
-    return session_handle.storage_interface()
-
-
-@pytest.mark.asyncio
-async def test_storage_get_missing_key_returns_none(storage):
-    key = f"tests/storage/{uuid.uuid4().hex}/missing"
-
-    value = await storage.get(key)
-    assert value is None
-
-
-@pytest.mark.asyncio
-async def test_storage_put_then_get_returns_value(storage):
-    key = f"tests/storage/{uuid.uuid4().hex}/value"
-    expected = "hello storage"
-
-    await storage.put(key, expected)
-    assert await storage.get(key) == expected
-
-
-@pytest.mark.asyncio
-async def test_storage_put_overwrites_value(storage):
-    key = f"tests/storage/{uuid.uuid4().hex}/overwrite"
-
-    await storage.put(key, "old")
-    await storage.put(key, "new")
-    assert await storage.get(key) == "new"
-
-
-@pytest.mark.asyncio
-async def test_storage_put_overwrites_value(storage):
-    key = f"tests/storage/{uuid.uuid4().hex}/overwrite"
-
-    await storage.put(key, "old")
-    await storage.delete(key)
-    assert await storage.get(key) == None
-
-
-@pytest.mark.asyncio
-async def test_storage_get_rejects_wildcard_key(storage):
-    with pytest.raises(ValueError, match="only supports fixed keys"):
-        await storage.get("tests/storage/*")
-
-
-@pytest.mark.asyncio
-async def test_storage_put_rejects_wildcard_key(storage):
-    with pytest.raises(ValueError, match="only supports fixed keys"):
-        await storage.put("tests/storage/*", "value")
@@ -7,16 +7,9 @@ use zenoh_plugin_trait::PluginsManager;

 pub use zenoh::{Config, config::ZenohId};

-use crate::{
-    discovery::Discovery,
-    liveliness_aggregator::{LivelinessAggregator, spawn_liveliness_aggregator},
-};
-
-pub use zenoh_plugin_storage_manager::read_raw_memory_storage;
-pub const STORAGE_PREFIX: &str = "storage/mem1";
+use crate::discovery::Discovery;

 pub mod discovery;
-pub mod liveliness_aggregator;
 pub mod swarm;

 pub fn is_valid_zid(identity: &str) -> bool {
@@ -40,22 +33,20 @@ pub fn cfg(identity: &str, listen_port: u16) -> Result<zenoh::Config> {
    cfg.insert_json5("scouting/multicast/autoconnect", "[]")?;
    cfg.insert_json5("scouting/gossip/multihop", "true")?;
    cfg.insert_json5("adminspace/enabled", "true")?;
-    cfg.insert_json5("transport/link/tx/batch_size", "9216")?;
+    //cfg.insert_json5("transport/link/tx/batch_size", "9216")?;
    cfg.insert_json5("transport/link/rx/buffer_size", "16777216")?;
-    cfg.insert_json5("timestamping/enabled", "true")?;
+    //cfg.insert_json5("timestamping/enabled", "true")?;
    cfg.insert_json5("plugins/storage_manager/__required__", "true")?;
    cfg.insert_json5(
-        "plugins/storage_manager/storages/mem1/key_expr",
-        &format!("\"{STORAGE_PREFIX}/**\""),
-    )?;
-    cfg.insert_json5(
-        "plugins/storage_manager/storages/mem1/strip_prefix",
-        &format!("\"{STORAGE_PREFIX}\""),
-    )?;
-    cfg.insert_json5("plugins/storage_manager/storages/mem1/volume", "\"memory\"")?;
-    cfg.insert_json5(
-        "plugins/storage_manager/storages/mem1/replication/interval",
-        "2",
+        "plugins/storage_manager/storages/mem1",
+        r#"{
+            key_expr: "storage/mem1/**",
+            strip_prefix: "storage/mem1",
+            volume: "memory",
+            replication: {
+                interval: 2,
+            }
+        }"#,
    )?;
    Ok(cfg)
 }
@@ -108,15 +99,10 @@ pub async fn open(
                .await;
        }
    })));
-    let liveliness_aggregator = spawn_liveliness_aggregator(&z)?;
-    Ok(Session {
-        z,
-        liveliness_aggregator,
-        _jh,
-    })
+    Ok(Session { z, _jh })
 }

-pub struct AbortOnDrop(pub JoinHandle<()>);
+struct AbortOnDrop(JoinHandle<()>);
 impl Drop for AbortOnDrop {
    fn drop(&mut self) {
        self.0.abort();
@@ -126,6 +112,5 @@ impl Drop for AbortOnDrop {
 #[derive(Clone)]
 pub struct Session {
    pub z: ZSession,
-    pub liveliness_aggregator: LivelinessAggregator,
    _jh: Arc<AbortOnDrop>,
 }
@@ -1,44 +0,0 @@
-use std::{collections::HashSet, sync::Arc};
-
-use parking_lot::Mutex;
-use zenoh::{Result, Session, Wait, sample::SampleKind};
-
-pub fn spawn_liveliness_aggregator(session: &Session) -> Result<LivelinessAggregator> {
-    let store = Arc::new(Mutex::new(HashSet::default()));
-    session
-        .liveliness()
-        .declare_subscriber("live/*")
-        .history(true)
-        .callback({
-            let store = Arc::clone(&store);
-            move |sample| {
-                let Some(nid) = sample
-                    .key_expr()
-                    .to_string()
-                    .strip_prefix("live/")
-                    .map(str::to_owned)
-                else {
-                    return;
-                };
-                let mut mg = store.lock();
-                match sample.kind() {
-                    SampleKind::Put => mg.insert(nid),
-                    SampleKind::Delete => mg.remove(&nid),
-                };
-            }
-        })
-        .background()
-        .wait()?;
-    Ok(LivelinessAggregator { store })
-}
-
-#[derive(Clone)]
-pub struct LivelinessAggregator {
-    // need two arcs as the sub owns an arc to the store.
-    store: Arc<Mutex<HashSet<String>>>,
-}
-impl LivelinessAggregator {
-    pub fn dump(&self) -> HashSet<String> {
-        self.store.lock().clone()
-    }
-}
@@ -1,99 +0,0 @@
-use std::time::Duration;
-
-use zenoh::Wait;
-
-fn unique_key(name: &str) -> String {
-    format!("test/zenoh-runtime-polling/{}/{}", std::process::id(), name)
-}
-
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn get_and_recv_work_on_tokio_baseline() {
-    let session = zenoh::open(zenoh::Config::default())
-        .await
-        .expect("open session");
-
-    let key = unique_key("tokio-baseline");
-    let reply_key = key.clone();
-
-    let _queryable = session
-        .declare_queryable(key.clone())
-        .callback(move |query| {
-            query
-                .reply(reply_key.clone(), "hello-from-queryable")
-                .wait()
-                .expect("reply from queryable");
-        })
-        .await
-        .expect("declare queryable");
-
-    let replies = session.get(key).await.expect("get");
-
-    let reply = tokio::time::timeout(Duration::from_secs(5), replies.recv_async())
-        .await
-        .expect("timed out waiting for reply")
-        .expect("reply channel closed");
-
-    let sample = reply.result().expect("reply result was error");
-    let payload = sample
-        .payload()
-        .try_to_string()
-        .expect("payload should be utf8");
-
-    assert_eq!(payload.as_ref(), "hello-from-queryable");
-}
-
-#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
-async fn get_and_recv_work_when_polled_by_smol_without_tokio_context() {
-    let session = zenoh::open(zenoh::Config::default())
-        .await
-        .expect("open session under tokio");
-
-    let key = unique_key("smol-no-tokio-context");
-    let reply_key = key.clone();
-
-    let _queryable = session
-        .declare_queryable(key.clone())
-        .callback(move |query| {
-            query
-                .reply(reply_key.clone(), "hello-from-queryable")
-                .wait()
-                .expect("reply from queryable");
-        })
-        .await
-        .expect("declare queryable under tokio");
-
-    let session_for_smol = session.clone();
-
-    let (tx, rx) = std::sync::mpsc::channel();
-
-    std::thread::spawn(move || {
-        // This thread was not entered by Tokio.
-        // If Zenoh's get/recv path requires an ambient Tokio Handle in the polling thread,
-        // this is where it should panic, hang, or error.
-        let result = {
-            smol::block_on(async move {
-                let replies = session_for_smol.get(key).await.expect("get under smol");
-
-                let reply = replies.recv_async().await.expect("reply channel closed");
-
-                let sample = reply.result().expect("reply result was error");
-                let payload = sample
-                    .payload()
-                    .try_to_string()
-                    .expect("payload should be utf8");
-
-                payload.to_string()
-            })
-        };
-
-        tx.send(result).expect("send test result");
-    });
-
-    let result = rx
-        .recv_timeout(Duration::from_secs(5))
-        .expect("smol thread timed out; likely hung waiting for get/reply");
-
-    let payload = result;
-
-    assert_eq!(payload, "hello-from-queryable");
-}
@@ -13,7 +13,6 @@ from uuid import uuid4

 import anyio
 from anyio import BrokenResourceError, ClosedResourceError
-from exo_rs import SessionHandle
 from fastapi import FastAPI, File, Form, HTTPException, Query, Request, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, JSONResponse, StreamingResponse
@@ -23,7 +22,6 @@ from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from hypercorn.utils import LifespanTimeoutError, ShutdownError
 from loguru import logger
-from pydantic import ValidationError

 from exo.api.adapters.chat_completions import (
    chat_request_to_text_generation,
@@ -155,11 +153,14 @@ from exo.shared.types.chunks import (
    ToolCallChunk,
 )
 from exo.shared.types.commands import (
+    AddCustomModelCard,
    CancelDownload,
    Command,
    CreateInstance,
+    DeleteCustomModelCard,
    DeleteDownload,
    DeleteInstance,
+    DeleteInstanceLink,
    DownloadCommand,
    ForwarderCommand,
    ForwarderDownloadCommand,
@@ -167,6 +168,7 @@ from exo.shared.types.commands import (
    ImageGeneration,
    PlaceInstance,
    SendInputChunk,
+    SetInstanceLink,
    StartDownload,
    TaskCancelled,
    TaskFinished,
@@ -244,7 +246,6 @@ class API:
        download_command_sender: Sender[ForwarderDownloadCommand],
        # This lets us pause the API if an election is running
        election_receiver: Receiver[ElectionMessage],
-        session_handle: SessionHandle,
    ) -> None:
        self.state = State()
        self._event_log = DiskEventLog(_API_EVENT_LOG_DIR)
@@ -257,8 +258,6 @@ class API:
        self.last_completed_election: int = 0
        self.port = port
        self._sent_image_hashes: set[str] = set()
-        self.aggregator = session_handle.last_value_aggregator("metrics")
-        self.storage = session_handle.storage_interface()

        self.paused: bool = False
        self.paused_ev: anyio.Event = anyio.Event()
@@ -410,12 +409,10 @@ class API:
        self.app.post("/onboarding")(self.complete_onboarding)

    def get_state(self, path: str = ""):
-        state = self.state.with_aggregator(self.aggregator)
-
        if path == "":
-            return state
+            return self.state
        try:
-            x: Any = state.model_dump(by_alias=True)
+            x = self.state.model_dump(by_alias=True)
            for attr in path.split("/"):
                if attr != "":
                    if isinstance(x, dict):
@@ -479,7 +476,6 @@ class API:
        model_card = await ModelCard.load(model_id)

        try:
-            state = self.state.with_aggregator(self.aggregator)
            placements = get_instance_placements(
                PlaceInstance(
                    model_card=model_card,
@@ -487,13 +483,13 @@ class API:
                    instance_meta=instance_meta,
                    min_nodes=min_nodes,
                ),
-                node_memory=state.node_memory,
-                node_network=state.node_network,
-                node_backends=state.node_backends,
-                topology=state.topology,
-                current_instances=state.instances,
-                download_status=state.downloads,
-                node_rdma_ctl=state.node_rdma_ctl,
+                node_memory=self.state.node_memory,
+                node_network=self.state.node_network,
+                node_backends=self.state.node_backends,
+                topology=self.state.topology,
+                current_instances=self.state.instances,
+                download_status=self.state.downloads,
+                node_rdma_ctl=self.state.node_rdma_ctl,
            )
        except ValueError as exc:
            raise HTTPException(status_code=400, detail=str(exc)) from exc
@@ -518,9 +514,8 @@ class API:
        seen: set[tuple[ModelId, Sharding, InstanceMeta, int]] = set()
        previews: list[PlacementPreview] = []
        required_nodes = set(node_ids) if node_ids else None
-        state = self.state.with_aggregator(self.aggregator)

-        if len(list(state.topology.list_nodes())) == 0:
+        if len(list(self.state.topology.list_nodes())) == 0:
            return PlacementPreviewResponse(previews=[])

        try:
@@ -535,7 +530,9 @@ class API:
                instance_combinations.extend(
                    [
                        (sharding, instance_meta, i)
-                        for i in range(1, len(list(state.topology.list_nodes())) + 1)
+                        for i in range(
+                            1, len(list(self.state.topology.list_nodes())) + 1
+                        )
                    ]
                )
        # TODO: PDD
@@ -550,14 +547,14 @@ class API:
                        instance_meta=instance_meta,
                        min_nodes=min_nodes,
                    ),
-                    node_memory=state.node_memory,
-                    node_network=state.node_network,
-                    node_backends=state.node_backends,
-                    topology=state.topology,
-                    current_instances=state.instances,
+                    node_memory=self.state.node_memory,
+                    node_network=self.state.node_network,
+                    node_backends=self.state.node_backends,
+                    topology=self.state.topology,
+                    current_instances=self.state.instances,
                    required_nodes=required_nodes,
-                    download_status=state.downloads,
-                    node_rdma_ctl=state.node_rdma_ctl,
+                    download_status=self.state.downloads,
+                    node_rdma_ctl=self.state.node_rdma_ctl,
                )
            except ValueError as exc:
                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
@@ -699,16 +696,9 @@ class API:
        return {"disaggregation": ENABLE_DISAGGREGATION}

    async def list_instance_links(self) -> list[InstanceLink]:
-        links: list[InstanceLink] = []
        if not ENABLE_DISAGGREGATION:
-            return links
-        for _, value in (await self.storage.dump("custom_model_cards/")).items():
-            try:
-                link = InstanceLink.model_validate_json(value)
-            except ValidationError:
-                continue
-            links.append(link)
-        return links
+            return []
+        return list(self.state.instance_links.values())

    async def create_instance_link(
        self, body: InstanceLinkBody
@@ -725,22 +715,25 @@ class API:
    async def _set_instance_link(
        self, link_id: InstanceLinkId, body: InstanceLinkBody
    ) -> InstanceLinkResponse:
-        await self.storage.put(
-            f"instance_links/{link_id}",
-            InstanceLink(
-                link_id=link_id,
-                prefill_instances=body.prefill_instances,
-                decode_instances=body.decode_instances,
-            ).model_dump_json(),
+        command = SetInstanceLink(
+            link_id=link_id,
+            prefill_instances=list(body.prefill_instances),
+            decode_instances=list(body.decode_instances),
+        )
+        await self._send(command)
+        return InstanceLinkResponse(
+            message="Command received.", command_id=command.command_id
        )
-        return InstanceLinkResponse(message="Command received.")

    async def delete_instance_link(
        self, link_id: InstanceLinkId
    ) -> InstanceLinkResponse:
        _require_disaggregation_enabled()
-        await self.storage.delete(f"instance_links/{link_id}")
-        return InstanceLinkResponse(message="Command received.")
+        command = DeleteInstanceLink(link_id=link_id)
+        await self._send(command)
+        return InstanceLinkResponse(
+            message="Command received.", command_id=command.command_id
+        )

    async def cancel_command(self, command_id: CommandId) -> CancelCommandResponse:
        """Cancel an active command by closing its stream and notifying workers."""
@@ -798,11 +791,7 @@ class API:
    async def _collect_text_generation_with_stats(
        self, command_id: CommandId
    ) -> BenchChatCompletionResponse:
-        sampler = PowerSampler(
-            get_node_system=lambda: self.state.with_aggregator(
-                self.aggregator
-            ).node_system
-        )
+        sampler = PowerSampler(get_node_system=lambda: self.state.node_system)
        text_parts: list[str] = []
        tool_calls: list[ToolCall] = []
        model: ModelId | None = None
@@ -1325,11 +1314,7 @@ class API:
        num_images: int,
        response_format: str,
    ) -> BenchImageGenerationResponse:
-        sampler = PowerSampler(
-            get_node_system=lambda: self.state.with_aggregator(
-                self.aggregator
-            ).node_system
-        )
+        sampler = PowerSampler(get_node_system=lambda: self.state.node_system)
        images: list[ImageData] = []
        stats: ImageGenerationStats | None = None
        async with anyio.create_task_group() as tg:
@@ -1793,7 +1778,7 @@ class API:
        """Calculate total available memory across all nodes in bytes."""
        total_available = Memory()

-        for memory in self.state.with_aggregator(self.aggregator).node_memory.values():
+        for memory in self.state.node_memory.values():
            total_available += memory.ram_available

        return total_available
@@ -1842,8 +1827,11 @@ class API:
                status_code=400, detail=f"Failed to fetch model: {exc}"
            ) from exc

-        await self.storage.put(
-            f"custom_model_cards/{card.model_id.normalize()}", card.model_dump_json()
+        await self.command_sender.send(
+            ForwarderCommand(
+                origin=self._system_id,
+                command=AddCustomModelCard(model_card=card),
+            )
        )

        # Immediately update the local cache so the subsequent GET /models
@@ -1868,7 +1856,12 @@ class API:
        if card is None or not card.is_custom:
            raise HTTPException(status_code=404, detail="Custom model card not found")

-        await self.storage.delete(f"custom_model_cards/{card.model_id.normalize()}")
+        await self.command_sender.send(
+            ForwarderCommand(
+                origin=self._system_id,
+                command=DeleteCustomModelCard(model_id=model_id),
+            )
+        )

        return JSONResponse(
            {"message": "Model card deleted", "model_id": str(model_id)}
@@ -329,6 +329,7 @@ class InstanceLinkBody(BaseModel):

 class InstanceLinkResponse(BaseModel):
    message: str
+    command_id: CommandId


 ImageSize = Literal[
@@ -10,7 +10,7 @@ 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, SessionHandle
+from exo_rs import Pidfile, PidfileError
 from loguru import logger
 from pydantic import PositiveInt

@@ -47,21 +47,18 @@ class Node:
    node_id: NodeId
    offline: bool
    _api_port: int
-    _sh: SessionHandle
    _tg: TaskGroup = field(init=False, default_factory=TaskGroup)

    @classmethod
    async def create(cls, args: "Args") -> Self:
        node_id = get_node_zid()
        session_id = SessionId(master_node_id=node_id, election_clock=0)
-        session_handle, _nh = SessionHandle.new(
+        router = Router.create(
            node_id,
            namespace=args.namespace,
            listen_port=args.zenoh_port,
            discovery_service_port=args.discovery_port,
        )
-        router = Router(_nh)
-
        await router.register_topic(topics.GLOBAL_EVENTS)
        await router.register_topic(topics.LOCAL_EVENTS)
        await router.register_topic(topics.COMMANDS)
@@ -100,7 +97,6 @@ class Node:
                command_sender=router.sender(topics.COMMANDS),
                download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                election_receiver=router.receiver(topics.ELECTION_MESSAGES),
-                session_handle=session_handle,
            )
        else:
            api = None
@@ -112,7 +108,6 @@ class Node:
                event_sender=event_router.sender(),
                command_sender=router.sender(topics.COMMANDS),
                download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
-                session_handle=session_handle,
                api_port=args.api_port,
            )
        else:
@@ -127,8 +122,6 @@ class Node:
            local_event_receiver=router.receiver(topics.LOCAL_EVENTS),
            command_receiver=router.receiver(topics.COMMANDS),
            download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
-            aggregator=session_handle.last_value_aggregator("metrics"),
-            storage=session_handle.storage_interface(),
        )

        er_send, er_recv = channel[ElectionResult]()
@@ -157,7 +150,6 @@ class Node:
            node_id,
            args.offline,
            args.api_port,
-            session_handle,
        )

    async def run(self):
@@ -233,8 +225,6 @@ class Node:
                        download_command_sender=self.router.sender(
                            topics.DOWNLOAD_COMMANDS
                        ),
-                        aggregator=self._sh.last_value_aggregator("metrics"),
-                        storage=self._sh.storage_interface(),
                    )
                    self._tg.start_soon(self.master.run)
                elif (
@@ -274,7 +264,6 @@ class Node:
                            download_command_sender=self.router.sender(
                                topics.DOWNLOAD_COMMANDS
                            ),
-                            session_handle=self._sh,
                            api_port=self._api_port,
                        )
                        self._tg.start_soon(self.worker.run)
@@ -389,16 +378,17 @@ def main_inner(args: "Args"):


 class Args(FrozenModel):
-    verbosity: int
-    force_master: bool
-    spawn_api: bool
-    api_port: PositiveInt
+    verbosity: int = 0
+    force_master: bool = False
+    spawn_api: bool = False
+    api_port: PositiveInt = 52415
+    tb_only: bool = False
    no_worker: bool = False
    no_downloads: bool = False
-    offline: bool
-    no_batch: bool
+    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
+    legacy_daemon: bool = False
    bootstrap_peers: list[str] = []
    namespace: str
    zenoh_port: int
@@ -486,7 +476,7 @@ class Args(FrozenModel):
            type=int,
            default=52414,
            dest="zenoh_port",
-            help="Fixed port for zenoh to listen on.",
+            help="Fixed TCP port for zenoh to listen.",
        )
        parser.add_argument(
            "--discovery-port",
@@ -1,9 +1,7 @@
 from datetime import datetime, timedelta, timezone

 import anyio
-from exo_rs import LVAggregator, Storage
 from loguru import logger
-from pydantic import ValidationError

 from exo.master.placement import (
    add_instance_to_placements,
@@ -20,8 +18,11 @@ from exo.routing.event_router import (
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_EVENT_LOG_DIR, EXO_TRACING_ENABLED
 from exo.shared.types.commands import (
+    AddCustomModelCard,
    CreateInstance,
+    DeleteCustomModelCard,
    DeleteInstance,
+    DeleteInstanceLink,
    ForwarderCommand,
    ForwarderDownloadCommand,
    ImageEdits,
@@ -29,6 +30,7 @@ from exo.shared.types.commands import (
    PlaceInstance,
    RequestEventLog,
    SendInputChunk,
+    SetInstanceLink,
    TaskCancelled,
    TaskFinished,
    TestCommand,
@@ -36,11 +38,15 @@ from exo.shared.types.commands import (
 )
 from exo.shared.types.common import CommandId, NodeId, SessionId, SystemId
 from exo.shared.types.events import (
+    CustomModelCardAdded,
+    CustomModelCardDeleted,
    Event,
    GlobalForwarderEvent,
    IndexedEvent,
    InputChunkReceived,
    InstanceDeleted,
+    InstanceLinkCreated,
+    InstanceLinkDeleted,
    LocalForwarderEvent,
    NodeGatheredInfo,
    NodeTimedOut,
@@ -73,9 +79,7 @@ from exo.utils.event_buffer import MultiSourceBuffer
 from exo.utils.task_group import TaskGroup


-def _prefill_endpoint_for(
-    state: State, instance_links: list[InstanceLink], decode_instance_id: InstanceId
-) -> str | None:
+def _prefill_endpoint_for(state: State, decode_instance_id: InstanceId) -> str | None:
    decode = state.instances.get(decode_instance_id)
    if decode is None:
        return None
@@ -84,7 +88,7 @@ def _prefill_endpoint_for(
        return None

    sources: set[InstanceId] = set()
-    for link in instance_links:
+    for link in state.instance_links.values():
        if decode_instance_id in link.decode_instances:
            sources.update(link.prefill_instances)
    sources.discard(decode_instance_id)
@@ -126,8 +130,6 @@ class Master:
        local_event_receiver: Receiver[LocalForwarderEvent],
        global_event_sender: Sender[GlobalForwarderEvent],
        download_command_sender: Sender[ForwarderDownloadCommand],
-        aggregator: LVAggregator,
-        storage: Storage,
    ):
        self.node_id = node_id
        self.session_id = session_id
@@ -144,8 +146,6 @@ class Master:
        self._event_log = DiskEventLog(EXO_EVENT_LOG_DIR / "master")
        self._pending_traces: dict[TaskId, dict[int, list[TraceEventData]]] = {}
        self._expected_ranks: dict[TaskId, set[int]] = {}
-        self.aggregator: LVAggregator = aggregator
-        self.storage: Storage = storage

    async def run(self):
        logger.info("Starting Master")
@@ -182,21 +182,10 @@ class Master:
                            pass
                        case TextGeneration():
                            # set-difference => prefill-only nodes
-                            instance_links: list[InstanceLink] = []
                            prefill_only: set[InstanceId] = set()
-                            for _, link in (
-                                await self.storage.dump("instance_links/")
-                            ).items():
-                                try:
-                                    instance_links.append(
-                                        InstanceLink.model_validate_json(link)
-                                    )
-                                except ValidationError:
-                                    continue
-
-                            for link in instance_links:
+                            for link in self.state.instance_links.values():
                                prefill_only.update(link.prefill_instances)
-                            for link in instance_links:
+                            for link in self.state.instance_links.values():
                                prefill_only.difference_update(link.decode_instances)

                            for instance in self.state.instances.values():
@@ -236,9 +225,7 @@ class Master:
                            params = command.task_params.model_copy(
                                update={
                                    "prefill_endpoint": _prefill_endpoint_for(
-                                        self.state.with_aggregator(self.aggregator),
-                                        instance_links,
-                                        decode_instance_id,
+                                        self.state, decode_instance_id
                                    ),
                                }
                            )
@@ -382,16 +369,15 @@ class Master:
                                )
                            generated_events.extend(transition_events)
                        case PlaceInstance():
-                            state = self.state.with_aggregator(self.aggregator)
                            placement = place_instance(
                                command,
-                                state.topology,
-                                state.instances,
-                                state.node_memory,
-                                state.node_network,
-                                state.node_backends,
-                                download_status=state.downloads,
-                                node_rdma_ctl=state.node_rdma_ctl,
+                                self.state.topology,
+                                self.state.instances,
+                                self.state.node_memory,
+                                self.state.node_network,
+                                self.state.node_backends,
+                                download_status=self.state.downloads,
+                                node_rdma_ctl=self.state.node_rdma_ctl,
                            )
                            transition_events = get_transition_events(
                                self.state.instances, placement, self.state.tasks
@@ -442,6 +428,29 @@ class Master:
                                    f"Finished command {command.finished_command_id} finished"
                                )

+                        case AddCustomModelCard():
+                            generated_events.append(
+                                CustomModelCardAdded(model_card=command.model_card)
+                            )
+                        case DeleteCustomModelCard():
+                            generated_events.append(
+                                CustomModelCardDeleted(model_id=command.model_id)
+                            )
+                        case SetInstanceLink():
+                            link = InstanceLink(
+                                link_id=command.link_id,
+                                prefill_instances=list(
+                                    dict.fromkeys(command.prefill_instances)
+                                ),
+                                decode_instances=list(
+                                    dict.fromkeys(command.decode_instances)
+                                ),
+                            )
+                            generated_events.append(InstanceLinkCreated(link=link))
+                        case DeleteInstanceLink():
+                            generated_events.append(
+                                InstanceLinkDeleted(link_id=command.link_id)
+                            )
                        case RequestEventLog():
                            # We should just be able to send everything, since other buffers will ignore old messages
                            # rate limit to 1000 at a time
@@ -462,9 +471,7 @@ class Master:
    async def _plan(self) -> None:
        while True:
            # kill broken instances
-            connected_node_ids = set(
-                self.state.with_aggregator(self.aggregator).topology.list_nodes()
-            )
+            connected_node_ids = set(self.state.topology.list_nodes())
            for instance_id, instance in self.state.instances.items():
                for node_id in instance.shard_assignments.node_to_runner:
                    if node_id not in connected_node_ids:
@@ -474,9 +481,7 @@ class Master:
                        break

            # time out dead nodes
-            for node_id, time in self.state.with_aggregator(
-                self.aggregator
-            ).last_seen.items():
+            for node_id, time in self.state.last_seen.items():
                now = datetime.now(tz=timezone.utc)
                if now - time > timedelta(seconds=30):
                    logger.info(f"Manually removing node {node_id} due to inactivity")
@@ -47,25 +47,6 @@ from exo.utils.channels import channel
 from exo.utils.info_gatherer.info_gatherer import NodeBackends


-class MockAggregator:
-    def dump(self) -> dict[str, str]:
-        return {}
-
-
-class MockStorage:
-    async def get(self, _: str) -> None:
-        return None
-
-    async def put(self, _1: str, _2: str) -> None:
-        return None
-
-    async def delete(self, _: str) -> None:
-        return None
-
-    async def dump(self, _: str) -> dict[str, str]:
-        return {}
-
-
@pytest.mark.asyncio
 async def test_master():
    node_id = get_node_zid()
@@ -113,8 +94,6 @@ async def test_master():
        local_event_receiver=le_receiver,
        command_receiver=co_receiver,
        download_command_sender=fcds,
-        aggregator=MockAggregator(),  # pyright: ignore[reportArgumentType]
-        storage=MockStorage(),  # pyright: ignore[reportArgumentType]
    )
    logger.info("run the master")
    async with anyio.create_task_group() as tg:
@@ -4,14 +4,19 @@ from datetime import datetime

 from loguru import logger

-from exo.shared.types.common import NodeId
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.types.common import ModelId, NodeId
 from exo.shared.types.events import (
    ChunkGenerated,
+    CustomModelCardAdded,
+    CustomModelCardDeleted,
    Event,
    IndexedEvent,
    InputChunkReceived,
    InstanceCreated,
    InstanceDeleted,
+    InstanceLinkCreated,
+    InstanceLinkDeleted,
    NodeDownloadProgress,
    NodeGatheredInfo,
    NodeTimedOut,
@@ -27,6 +32,7 @@ from exo.shared.types.events import (
    TracesCollected,
    TracesMerged,
 )
+from exo.shared.types.instance_link import InstanceLink, InstanceLinkId
 from exo.shared.types.profiling import (
    NodeIdentity,
    NodeNetworkInfo,
@@ -36,6 +42,7 @@ from exo.shared.types.profiling import (
 )
 from exo.shared.types.state import State
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
+from exo.shared.types.topology import Connection, RDMAConnection
 from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import (
@@ -60,6 +67,18 @@ from exo.utils.info_gatherer.info_gatherer import (
 )


+def _is_rdma_ctl_enabled(
+    node_id: NodeId, node_rdma_ctl: Mapping[NodeId, NodeRdmaCtlStatus]
+) -> bool:
+    """A node is RDMA-capable only if rdma_ctl status has been observed as enabled.
+
+    Missing entries default to ``False`` — if we have not yet observed (or the node
+    cannot run) ``rdma_ctl``, it must not participate in an RDMA-backed instance.
+    """
+    status = node_rdma_ctl.get(node_id)
+    return status is not None and status.enabled
+
+
 def event_apply(event: Event, state: State) -> State:
    """Apply an event to state."""
    match event:
@@ -72,6 +91,10 @@ def event_apply(event: Event, state: State) -> State:
            | TracesMerged()
        ):  # Pass-through events that don't modify state
            return state
+        case CustomModelCardAdded():
+            return apply_custom_model_card_added(event, state)
+        case CustomModelCardDeleted():
+            return apply_custom_model_card_deleted(event, state)
        case InstanceCreated():
            return apply_instance_created(event, state)
        case InstanceDeleted():
@@ -96,6 +119,10 @@ def event_apply(event: Event, state: State) -> State:
            return apply_topology_edge_created(event, state)
        case TopologyEdgeDeleted():
            return apply_topology_edge_deleted(event, state)
+        case InstanceLinkCreated():
+            return apply_instance_link_created(event, state)
+        case InstanceLinkDeleted():
+            return apply_instance_link_deleted(event, state)


 def apply(state: State, event: IndexedEvent) -> State:
@@ -195,7 +222,38 @@ def apply_instance_deleted(event: InstanceDeleted, state: State) -> State:
    new_instances: Mapping[InstanceId, Instance] = {
        iid: inst for iid, inst in state.instances.items() if iid != event.instance_id
    }
-    return state.model_copy(update={"instances": new_instances})
+    new_links: dict[InstanceLinkId, InstanceLink] = {}
+    for link_id, link in state.instance_links.items():
+        prefill = [i for i in link.prefill_instances if i != event.instance_id]
+        decode = [i for i in link.decode_instances if i != event.instance_id]
+        if not prefill or not decode:
+            continue
+        if prefill == list(link.prefill_instances) and decode == list(
+            link.decode_instances
+        ):
+            new_links[link_id] = link
+        else:
+            new_links[link_id] = link.model_copy(
+                update={"prefill_instances": prefill, "decode_instances": decode}
+            )
+    return state.model_copy(
+        update={"instances": new_instances, "instance_links": new_links}
+    )
+
+
+def apply_instance_link_created(event: InstanceLinkCreated, state: State) -> State:
+    new_links: Mapping[InstanceLinkId, InstanceLink] = {
+        **state.instance_links,
+        event.link.link_id: event.link,
+    }
+    return state.model_copy(update={"instance_links": new_links})
+
+
+def apply_instance_link_deleted(event: InstanceLinkDeleted, state: State) -> State:
+    new_links: Mapping[InstanceLinkId, InstanceLink] = {
+        lid: link for lid, link in state.instance_links.items() if lid != event.link_id
+    }
+    return state.model_copy(update={"instance_links": new_links})


 def apply_runner_status_updated(event: RunnerStatusUpdated, state: State) -> State:
@@ -350,26 +408,59 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:
                event.node_id: NodeThunderboltInfo(interfaces=info.idents),
            }
        case MacThunderboltConnections():
-            update["node_thunderbolt_connections"] = {
-                **state.node_thunderbolt_connections,
-                event.node_id: info,
+            conn_map = {
+                tb_ident.domain_uuid: (nid, tb_ident.rdma_interface)
+                for nid in state.node_thunderbolt
+                for tb_ident in state.node_thunderbolt[nid].interfaces
            }
+            source_is_rdma_enabled = _is_rdma_ctl_enabled(
+                event.node_id, state.node_rdma_ctl
+            )
+            as_rdma_conns = [
+                Connection(
+                    source=event.node_id,
+                    sink=conn_map[tb_conn.sink_uuid][0],
+                    edge=RDMAConnection(
+                        source_rdma_iface=conn_map[tb_conn.source_uuid][1],
+                        sink_rdma_iface=conn_map[tb_conn.sink_uuid][1],
+                    ),
+                )
+                for tb_conn in info.conns
+                if tb_conn.source_uuid in conn_map
+                if tb_conn.sink_uuid in conn_map
+                if source_is_rdma_enabled
+                and _is_rdma_ctl_enabled(
+                    conn_map[tb_conn.sink_uuid][0], state.node_rdma_ctl
+                )
+            ]
+            topology.replace_all_out_rdma_connections(event.node_id, as_rdma_conns)
        case ThunderboltBridgeInfo():
            new_tb_bridge: dict[NodeId, ThunderboltBridgeStatus] = {
                **state.node_thunderbolt_bridge,
                event.node_id: info.status,
            }
            update["node_thunderbolt_bridge"] = new_tb_bridge
-            update["thunderbolt_bridge_cycles"] = (
-                topology.get_thunderbolt_bridge_cycles(
-                    new_tb_bridge, state.node_network
+            # Only recompute cycles if the enabled status changed
+            old_status = state.node_thunderbolt_bridge.get(event.node_id)
+            old_enabled = old_status.enabled if old_status else False
+            new_enabled = info.status.enabled
+            if old_enabled != new_enabled:
+                update["thunderbolt_bridge_cycles"] = (
+                    topology.get_thunderbolt_bridge_cycles(
+                        new_tb_bridge, state.node_network
+                    )
                )
-            )
        case RdmaCtlStatus():
            update["node_rdma_ctl"] = {
                **state.node_rdma_ctl,
                event.node_id: NodeRdmaCtlStatus(enabled=info.enabled),
            }
+            # If RDMA just got disabled on this node, drop any RDMA edges touching it
+            # so placement / topology consumers cannot pick a disabled node for an
+            # RDMA-backed instance. (Edges will repopulate on the next
+            # MacThunderboltConnections poll once both endpoints are enabled again.)
+            if not info.enabled:
+                topology.remove_all_rdma_connections_touching(event.node_id)
        case NodeBackends():
            update["node_backends"] = {
                **state.node_backends,
@@ -380,45 +471,32 @@ def apply_node_gathered_info(event: NodeGatheredInfo, state: State) -> State:


 def apply_topology_edge_created(event: TopologyEdgeCreated, state: State) -> State:
-    source_connections = state.node_socket_connections.get(event.conn.source, {})
-    sink_connections = source_connections.get(event.conn.sink, [])
-
-    update = {
-        "node_socket_connections": {
-            **state.node_socket_connections,
-            event.conn.source: {
-                **source_connections,
-                event.conn.sink: sink_connections
-                if event.conn.edge in sink_connections
-                else [*sink_connections, event.conn.edge],
-            },
-        }
-    }
-    return state.model_copy(update=update)
+    topology = copy.deepcopy(state.topology)
+    topology.add_connection(event.conn)
+    return state.model_copy(update={"topology": topology})


 def apply_topology_edge_deleted(event: TopologyEdgeDeleted, state: State) -> State:
-    inner_update = {
-        sink: final_edges
-        for sink, edges in state.node_socket_connections.get(
-            event.conn.source, {}
-        ).items()
-        if (
-            final_edges := [
-                edge
-                for edge in edges
-                if (edge != event.conn.edge or sink != event.conn.sink)
-            ]
-        )
+    topology = copy.deepcopy(state.topology)
+    topology.remove_connection(event.conn)
+    # TODO: Clean up removing the reverse connection
+    return state.model_copy(update={"topology": topology})
+
+
+def apply_custom_model_card_added(event: CustomModelCardAdded, state: State) -> State:
+    new_cards: Mapping[ModelId, ModelCard] = {
+        **state.custom_model_cards,
+        event.model_card.model_id: event.model_card,
    }
-    update = {
-        "node_socket_connections": {
-            source: maps
-            for source, maps in {
-                **state.node_socket_connections,
-                event.conn.source: inner_update,
-            }.items()
-            if maps
-        }
+    return state.model_copy(update={"custom_model_cards": new_cards})
+
+
+def apply_custom_model_card_deleted(
+    event: CustomModelCardDeleted, state: State
+) -> State:
+    new_cards: Mapping[ModelId, ModelCard] = {
+        model_id: card
+        for model_id, card in state.custom_model_cards.items()
+        if model_id != event.model_id
    }
-    return state.model_copy(update=update)
+    return state.model_copy(update={"custom_model_cards": new_cards})
@@ -55,7 +55,7 @@ class _CardCache:
        except OSError as e:
            logger.warning(f"failed to save custom model card ({e.strerror})")

-    async def delete(self, model_id: ModelId) -> "ModelCard | None":
+    async def pop(self, model_id: ModelId) -> "ModelCard | None":
        """Delete a user-added custom model card. Returns True if deleted."""
        card_path = _custom_cards_dir / (ModelId(model_id).normalize() + ".toml")
        try:
@@ -0,0 +1,46 @@
+from exo.shared.apply import apply
+from exo.shared.models.model_cards import ModelCard, ModelTask
+from exo.shared.types.backends import Backend
+from exo.shared.types.common import ModelId
+from exo.shared.types.events import (
+    CustomModelCardAdded,
+    CustomModelCardDeleted,
+    IndexedEvent,
+)
+from exo.shared.types.memory import Memory
+from exo.shared.types.state import State
+
+
+def _model_card(model_id: ModelId) -> ModelCard:
+    return ModelCard(
+        model_id=model_id,
+        n_layers=1,
+        storage_size=Memory.from_bytes(1),
+        hidden_size=1,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+        backends=[Backend.MlxMetal],
+    )
+
+
+def test_custom_model_card_added_is_reduced_into_state() -> None:
+    card = _model_card(ModelId("custom/model"))
+
+    state = apply(
+        State(),
+        IndexedEvent(idx=0, event=CustomModelCardAdded(model_card=card)),
+    )
+
+    assert state.custom_model_cards == {card.model_id: card}
+
+
+def test_custom_model_card_deleted_removes_card_from_state() -> None:
+    card = _model_card(ModelId("custom/model"))
+    state = State(custom_model_cards={card.model_id: card}, last_event_applied_idx=0)
+
+    state = apply(
+        state,
+        IndexedEvent(idx=1, event=CustomModelCardDeleted(model_id=card.model_id)),
+    )
+
+    assert state.custom_model_cards == {}
@@ -0,0 +1,72 @@
+from exo.shared.apply import (
+    apply_instance_deleted,
+    apply_instance_link_created,
+    apply_instance_link_deleted,
+)
+from exo.shared.types.events import (
+    InstanceDeleted,
+    InstanceLinkCreated,
+    InstanceLinkDeleted,
+)
+from exo.shared.types.instance_link import InstanceLink, InstanceLinkId
+from exo.shared.types.state import State
+from exo.shared.types.worker.instances import InstanceId
+
+
+def _link(
+    prefill: list[InstanceId],
+    decode: list[InstanceId],
+    link_id: InstanceLinkId | None = None,
+) -> InstanceLink:
+    return InstanceLink(
+        link_id=link_id or InstanceLinkId(),
+        prefill_instances=prefill,
+        decode_instances=decode,
+    )
+
+
+def test_create_link() -> None:
+    state = State()
+    link = _link([InstanceId("a")], [InstanceId("b")])
+    new_state = apply_instance_link_created(InstanceLinkCreated(link=link), state)
+    assert new_state.instance_links == {link.link_id: link}
+
+
+def test_update_replaces_existing_link() -> None:
+    a, b, c = InstanceId("a"), InstanceId("b"), InstanceId("c")
+    link = _link([a], [b])
+    state = State(instance_links={link.link_id: link})
+
+    updated = link.model_copy(update={"decode_instances": [b, c]})
+    new_state = apply_instance_link_created(InstanceLinkCreated(link=updated), state)
+    assert set(new_state.instance_links[link.link_id].decode_instances) == {b, c}
+
+
+def test_delete_link() -> None:
+    link = _link([InstanceId("a")], [InstanceId("b")])
+    state = State(instance_links={link.link_id: link})
+
+    new_state = apply_instance_link_deleted(
+        InstanceLinkDeleted(link_id=link.link_id), state
+    )
+    assert new_state.instance_links == {}
+
+
+def test_instance_deleted_strips_from_links() -> None:
+    a, b, c = InstanceId("a"), InstanceId("b"), InstanceId("c")
+    link = _link([a, c], [b])
+    state = State(instance_links={link.link_id: link})
+
+    new_state = apply_instance_deleted(InstanceDeleted(instance_id=a), state)
+    remaining = new_state.instance_links[link.link_id]
+    assert remaining.prefill_instances == [c]
+    assert remaining.decode_instances == [b]
+
+
+def test_instance_deleted_drops_link_when_role_empties() -> None:
+    a, b = InstanceId("a"), InstanceId("b")
+    link = _link([a], [b])
+    state = State(instance_links={link.link_id: link})
+
+    new_state = apply_instance_deleted(InstanceDeleted(instance_id=a), state)
+    assert link.link_id not in new_state.instance_links
@@ -217,7 +217,7 @@ def test_topology_remove_all_rdma_connections_touching_keeps_socket_edges():
        )
    )
    socket_edge = SocketConnection(
-        sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/8000"),
+        sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/8000")
    )
    topology.add_connection(Connection(source=node_a, sink=node_b, edge=socket_edge))

@@ -0,0 +1,35 @@
+from exo.shared.types.common import NodeId
+from exo.shared.types.multiaddr import Multiaddr
+from exo.shared.types.state import State
+from exo.shared.types.topology import Connection, SocketConnection
+
+
+def test_state_serialization_roundtrip() -> None:
+    """Verify that State → JSON → State round-trip preserves topology."""
+
+    # --- build a simple state ------------------------------------------------
+    node_a = NodeId("node-a")
+    node_b = NodeId("node-b")
+
+    connection = Connection(
+        source=node_a,
+        sink=node_b,
+        edge=SocketConnection(
+            sink_multiaddr=Multiaddr(address="/ip4/127.0.0.1/tcp/10001"),
+        ),
+    )
+
+    state = State()
+    state.topology.add_connection(connection)
+
+    json_repr = state.model_dump_json()
+    restored_state = State.model_validate_json(json_repr)
+
+    assert (
+        state.topology.to_snapshot().nodes
+        == restored_state.topology.to_snapshot().nodes
+    )
+    assert set(state.topology.to_snapshot().connections) == set(
+        restored_state.topology.to_snapshot().connections
+    )
+    assert restored_state.model_dump_json() == json_repr
@@ -1,8 +1,10 @@
+import contextlib
 from collections.abc import Mapping, Sequence
 from dataclasses import dataclass, field
 from typing import Iterable

 import rustworkx as rx
+from pydantic import BaseModel, ConfigDict

 from exo.shared.types.common import NodeId
 from exo.shared.types.profiling import (
@@ -18,6 +20,15 @@ from exo.shared.types.topology import (
 )


+class TopologySnapshot(BaseModel):
+    nodes: Sequence[NodeId]
+    connections: Mapping[
+        NodeId, Mapping[NodeId, Sequence[SocketConnection | RDMAConnection]]
+    ]
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+
@dataclass
 class Topology:
    _graph: rx.PyDiGraph[NodeId, SocketConnection | RDMAConnection] = field(
@@ -25,6 +36,28 @@ class Topology:
    )
    _vertex_indices: dict[NodeId, int] = field(init=False, default_factory=dict)

+    def to_snapshot(self) -> TopologySnapshot:
+        return TopologySnapshot(
+            nodes=list(self.list_nodes()), connections=self.map_connections()
+        )
+
+    @classmethod
+    def from_snapshot(cls, snapshot: TopologySnapshot) -> "Topology":
+        topology = cls()
+
+        for node_id in snapshot.nodes:
+            with contextlib.suppress(ValueError):
+                topology.add_node(node_id)
+
+        for source in snapshot.connections:
+            for sink in snapshot.connections[source]:
+                for edge in snapshot.connections[source][sink]:
+                    topology.add_connection(
+                        Connection(source=source, sink=sink, edge=edge)
+                    )
+
+        return topology
+
    def add_node(self, node_id: NodeId) -> None:
        if node_id in self._vertex_indices:
            return
@@ -7,6 +7,7 @@ from exo.api.types import (
 from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.common import CommandId, NodeId, SystemId
+from exo.shared.types.instance_link import InstanceLinkId
 from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
@@ -81,6 +82,24 @@ class CancelDownload(BaseCommand):
    model_id: ModelId


+class AddCustomModelCard(BaseCommand):
+    model_card: ModelCard
+
+
+class DeleteCustomModelCard(BaseCommand):
+    model_id: ModelId
+
+
+class SetInstanceLink(BaseCommand):
+    link_id: InstanceLinkId
+    prefill_instances: list[InstanceId]
+    decode_instances: list[InstanceId]
+
+
+class DeleteInstanceLink(BaseCommand):
+    link_id: InstanceLinkId
+
+
 DownloadCommand = StartDownload | DeleteDownload | CancelDownload


@@ -96,6 +115,10 @@ Command = (
    | TaskCancelled
    | TaskFinished
    | SendInputChunk
+    | AddCustomModelCard
+    | DeleteCustomModelCard
+    | SetInstanceLink
+    | DeleteInstanceLink
 )


@@ -3,9 +3,11 @@ from typing import final

 from pydantic import Field

+from exo.shared.models.model_cards import ModelCard
 from exo.shared.topology import Connection
 from exo.shared.types.chunks import Chunk, InputImageChunk
-from exo.shared.types.common import CommandId, Id, NodeId, SessionId, SystemId
+from exo.shared.types.common import CommandId, Id, ModelId, NodeId, SessionId, SystemId
+from exo.shared.types.instance_link import InstanceLink, InstanceLinkId
 from exo.shared.types.tasks import Task, TaskId, TaskStatus
 from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
@@ -106,6 +108,14 @@ class TopologyEdgeDeleted(BaseEvent):
    conn: Connection


+class CustomModelCardAdded(BaseEvent):
+    model_card: ModelCard
+
+
+class CustomModelCardDeleted(BaseEvent):
+    model_id: ModelId
+
+
@final
 class TraceEventData(FrozenModel):
    name: str
@@ -128,6 +138,14 @@ class TracesMerged(BaseEvent):
    traces: list[TraceEventData]


+class InstanceLinkCreated(BaseEvent):
+    link: InstanceLink
+
+
+class InstanceLinkDeleted(BaseEvent):
+    link_id: InstanceLinkId
+
+
 Event = (
    TestEvent
    | TaskCreated
@@ -147,6 +165,10 @@ Event = (
    | TopologyEdgeDeleted
    | TracesCollected
    | TracesMerged
+    | CustomModelCardAdded
+    | CustomModelCardDeleted
+    | InstanceLinkCreated
+    | InstanceLinkDeleted
 )


@@ -16,10 +16,6 @@ class MemoryUsage(FrozenModel):
    swap_total: Memory
    swap_available: Memory

-    @classmethod
-    def tag(cls) -> str:
-        return cls.__name__
-
    @classmethod
    def from_bytes(
        cls, *, ram_total: int, ram_available: int, swap_total: int, swap_available: int
@@ -1,15 +1,15 @@
 from collections.abc import Mapping, Sequence
 from datetime import datetime
-from typing import Any
+from typing import Any, cast

-from exo_rs import LVAggregator
-from pydantic import ConfigDict, Field, model_serializer
+from pydantic import ConfigDict, Field, field_serializer, field_validator
 from pydantic.alias_generators import to_camel
-from pydantic_core.core_schema import SerializerFunctionWrapHandler

-from exo.shared.topology import Topology
+from exo.shared.models.model_cards import ModelCard
+from exo.shared.topology import Topology, TopologySnapshot
 from exo.shared.types.backends import Backend
-from exo.shared.types.common import NodeId
+from exo.shared.types.common import ModelId, NodeId
+from exo.shared.types.instance_link import InstanceLink, InstanceLinkId
 from exo.shared.types.profiling import (
    DiskUsage,
    MemoryUsage,
@@ -21,15 +21,9 @@ from exo.shared.types.profiling import (
    ThunderboltBridgeStatus,
 )
 from exo.shared.types.tasks import Task, TaskId
-from exo.shared.types.topology import (
-    Connection,
-    RDMAConnection,
-    SocketConnection,
-)
 from exo.shared.types.worker.downloads import DownloadProgress
 from exo.shared.types.worker.instances import Instance, InstanceId
 from exo.shared.types.worker.runners import RunnerId, RunnerStatus
-from exo.utils.info_gatherer.info_gatherer import MacThunderboltConnections
 from exo.utils.pydantic_ext import FrozenModel


@@ -45,6 +39,7 @@ class State(FrozenModel):
        alias_generator=to_camel,
        validate_by_name=True,
        extra="forbid",
+        # I want to reenable this ASAP, but it's causing an issue with TaskStatus
        strict=True,
        arbitrary_types_allowed=True,
    )
@@ -53,6 +48,7 @@ class State(FrozenModel):
    downloads: Mapping[NodeId, Sequence[DownloadProgress]] = {}
    tasks: Mapping[TaskId, Task] = {}
    last_seen: Mapping[NodeId, datetime] = {}
+    topology: Topology = Field(default_factory=Topology)
    last_event_applied_idx: int = Field(default=-1, ge=-1)

    # Granular node state mappings (update independently at different frequencies)
@@ -65,94 +61,34 @@ class State(FrozenModel):
    node_thunderbolt_bridge: Mapping[NodeId, ThunderboltBridgeStatus] = {}
    node_rdma_ctl: Mapping[NodeId, NodeRdmaCtlStatus] = {}
    node_backends: Mapping[NodeId, list[Backend]] = {}
-    node_socket_connections: Mapping[
-        NodeId, Mapping[NodeId, Sequence[SocketConnection]]
-    ] = {}
-    node_thunderbolt_connections: Mapping[NodeId, MacThunderboltConnections] = {}

    # Detected cycles where all nodes have Thunderbolt bridge enabled (>2 nodes)
    thunderbolt_bridge_cycles: Sequence[Sequence[NodeId]] = []

+    instance_links: Mapping[InstanceLinkId, InstanceLink] = {}
    prefill_server_ports: Mapping[RunnerId, int] = {}

-    @model_serializer(mode="wrap")
-    def _serialize(self, handler: SerializerFunctionWrapHandler) -> dict[str, Any]:
-        data = handler(self)  # pyright: ignore[reportAny]
-        data["topology"] = {
-            "nodes": list(self.node_identities.keys()),
-            "connections": self.topology.map_connections(),
-        }
-        return data  # pyright: ignore[reportAny]
+    # User-added model cards. Workers can reconcile their on-disk custom card cache
+    custom_model_cards: Mapping[ModelId, ModelCard] = {}

-    @property
-    def topology(self) -> Topology:
-        topology = Topology()
-        thunderbolt_by_uuid = {
-            ident.domain_uuid: (node_id, ident.rdma_interface)
-            for node_id, info in self.node_thunderbolt.items()
-            for ident in info.interfaces
-        }
-        for node_id in self.node_identities:
-            topology.add_node(node_id)
+    @field_serializer("topology", mode="plain")
+    def _encode_topology(self, value: Topology) -> TopologySnapshot:
+        return value.to_snapshot()

-        for source, data in self.node_socket_connections.items():
-            for sink, conns in data.items():
-                for conn in conns:
-                    topology.add_connection(
-                        Connection(source=source, sink=sink, edge=conn)
-                    )
+    @field_validator("topology", mode="before")
+    @classmethod
+    def _deserialize_topology(cls, value: object) -> Topology:  # noqa: D401 – Pydantic validator signature
+        """Convert an incoming *value* into a :class:`Topology` instance.

-        for source, connections in self.node_thunderbolt_connections.items():
-            if not self.node_rdma_ctl.get(
-                source, NodeRdmaCtlStatus(enabled=False)
-            ).enabled:
-                continue
-            for connection in connections.conns:
-                if (
-                    source_iface := thunderbolt_by_uuid.get(connection.source_uuid)
-                ) is None or (
-                    sink_iface := thunderbolt_by_uuid.get(connection.sink_uuid)
-                ) is None:
-                    continue
-                if not self.node_rdma_ctl.get(
-                    sink_iface[0], NodeRdmaCtlStatus(enabled=False)
-                ).enabled:
-                    continue
-                assert source_iface[0] == source, "registered invalid source uuid"
-                topology.add_connection(
-                    Connection(
-                        source=source_iface[0],
-                        sink=sink_iface[0],
-                        edge=RDMAConnection(
-                            source_rdma_iface=source_iface[1],
-                            sink_rdma_iface=sink_iface[1],
-                        ),
-                    )
-                )
+        Accepts either an already constructed :class:`Topology` or a mapping
+        representing :class:`~shared.topology.TopologySnapshot`.
+        """

-        return topology
+        if isinstance(value, Topology):
+            return value

-    def with_aggregator(self, aggregator: LVAggregator) -> "State":
-        from datetime import datetime, timezone
+        if isinstance(value, Mapping):  # likely a snapshot-dict coming from JSON
+            snapshot = TopologySnapshot(**cast(dict[str, Any], value))  # type: ignore[arg-type]
+            return Topology.from_snapshot(snapshot)

-        from pydantic import TypeAdapter
-
-        from exo.shared.apply import event_apply
-        from exo.shared.types.events import NodeGatheredInfo
-        from exo.utils.info_gatherer.info_gatherer import GatheredInfo
-
-        state = self.model_copy()
-        for key, value in aggregator.dump().items():
-            try:
-                data = TypeAdapter[GatheredInfo](GatheredInfo).validate_json(value)
-                node_id = NodeId(key.split("/")[0])
-                event = NodeGatheredInfo(
-                    node_id=node_id, when=str(datetime.now(tz=timezone.utc)), info=data
-                )
-                state = event_apply(event, state)
-            except Exception as e:
-                print(
-                    f"\n{'=' * 10}key: {key} with exception {str(e)}\nvalue: {value}{'=' * 10}\n"
-                )
-
-        return state
+        raise TypeError("Invalid representation for Topology field in State")
@@ -10,13 +10,11 @@ from typing import Self, cast
 import anyio
 from anyio import fail_after, open_process, to_thread
 from anyio.streams.buffered import BufferedByteReceiveStream
-from exo_rs import LVPublisher, SessionHandle
 from loguru import logger
 from pydantic import ValidationError

 from exo.shared.constants import EXO_CONFIG_FILE, EXO_DEFAULT_MODELS_DIR
 from exo.shared.types.backends import Backend
-from exo.shared.types.common import NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import (
    DiskUsage,
@@ -29,6 +27,7 @@ from exo.shared.types.thunderbolt import (
    ThunderboltConnectivity,
    ThunderboltIdentifier,
 )
+from exo.utils.channels import Sender
 from exo.utils.pydantic_ext import TaggedModel
 from exo.utils.task_group import TaskGroup

@@ -402,42 +401,10 @@ GatheredInfo = (

@dataclass
 class InfoGatherer:
-    session_handle: SessionHandle
-    node_id: NodeId
-    info_senders: dict[str, LVPublisher] = field(init=False, default_factory=dict)
+    info_sender: Sender[GatheredInfo]
    _tg: TaskGroup = field(init=False, default_factory=TaskGroup)
    _psutil_enabled: bool = field(init=False, default=False)

-    async def send(self, info: GatheredInfo):
-        if (tag := info.tag()) not in self.info_senders:
-            self.info_senders[tag] = self.session_handle.last_value_publisher(
-                f"metrics/{self.node_id}/{tag}"
-            )
-        await self.info_senders[tag].put(info.model_dump_json())
-
-    async def run(self):
-        async with self._tg as tg:
-            if IS_DARWIN:
-                tg.start_soon(self._monitor_macmon, 1)
-                tg.start_soon(self._monitor_system_profiler_thunderbolt_data, 5)
-                tg.start_soon(self._monitor_thunderbolt_bridge_status, 10)
-                tg.start_soon(self._monitor_rdma_ctl_status, 10)
-            if not IS_DARWIN:
-                tg.start_soon(self._monitor_memory_usage, 1)
-            tg.start_soon(self._watch_system_info, 10)
-            tg.start_soon(self._monitor_misc, 60)
-            tg.start_soon(self._monitor_static_info, 60)
-            tg.start_soon(self._monitor_disk_usage, 30)
-
-            nc = await NodeConfig.gather()
-            if nc is not None:
-                await self.send(nc)
-
-            await self.send(await NodeBackends.gather())
-
-    def shutdown(self):
-        self._tg.cancel_tasks()
-
    async def _can_read_macmon_metrics(self, macmon_path: str) -> bool:
        try:
            with fail_after(5):
@@ -474,11 +441,34 @@ class InfoGatherer:

        return True

+    async def run(self):
+        async with self._tg as tg:
+            if IS_DARWIN:
+                tg.start_soon(self._monitor_macmon, 1)
+                tg.start_soon(self._monitor_system_profiler_thunderbolt_data, 5)
+                tg.start_soon(self._monitor_thunderbolt_bridge_status, 10)
+                tg.start_soon(self._monitor_rdma_ctl_status, 10)
+            if not IS_DARWIN:
+                tg.start_soon(self._monitor_memory_usage, 1)
+            tg.start_soon(self._watch_system_info, 10)
+            tg.start_soon(self._monitor_misc, 60)
+            tg.start_soon(self._monitor_static_info, 60)
+            tg.start_soon(self._monitor_disk_usage, 30)
+
+            nc = await NodeConfig.gather()
+            if nc is not None:
+                await self.info_sender.send(nc)
+
+            await self.info_sender.send(await NodeBackends.gather())
+
+    def shutdown(self):
+        self._tg.cancel_tasks()
+
    async def _monitor_static_info(self, static_info_poll_interval: float):
        while True:
            try:
                with fail_after(30):
-                    await self.send(await StaticNodeInformation.gather())
+                    await self.info_sender.send(await StaticNodeInformation.gather())
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering static node info")
            await anyio.sleep(static_info_poll_interval)
@@ -487,7 +477,7 @@ class InfoGatherer:
        while True:
            try:
                with fail_after(10):
-                    await self.send(await MiscData.gather())
+                    await self.info_sender.send(await MiscData.gather())
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering misc data")
            await anyio.sleep(misc_poll_interval)
@@ -508,10 +498,12 @@ class InfoGatherer:
                    idents = [
                        it for i in data if (it := i.ident(iface_map)) is not None
                    ]
-                    await self.send(MacThunderboltIdentifiers(idents=idents))
+                    await self.info_sender.send(
+                        MacThunderboltIdentifiers(idents=idents)
+                    )

                    conns = [it for i in data if (it := i.conn()) is not None]
-                    await self.send(MacThunderboltConnections(conns=conns))
+                    await self.info_sender.send(MacThunderboltConnections(conns=conns))
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering Thunderbolt data")
            await anyio.sleep(system_profiler_interval)
@@ -528,7 +520,7 @@ class InfoGatherer:
        )
        while True:
            try:
-                await self.send(
+                await self.info_sender.send(
                    MemoryUsage.from_psutil(override_memory=override_memory)
                )
            except Exception as e:
@@ -540,7 +532,7 @@ class InfoGatherer:
            try:
                with fail_after(10):
                    nics = await get_network_interfaces()
-                    await self.send(NodeNetworkInterfaces(ifaces=nics))
+                    await self.info_sender.send(NodeNetworkInterfaces(ifaces=nics))
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering network interfaces")
            await anyio.sleep(interface_watcher_interval)
@@ -553,7 +545,7 @@ class InfoGatherer:
                with fail_after(30):
                    curr = await ThunderboltBridgeInfo.gather()
                    if curr is not None:
-                        await self.send(curr)
+                        await self.info_sender.send(curr)
            except Exception as e:
                logger.opt(exception=e).warning(
                    "Error gathering Thunderbolt Bridge status"
@@ -565,7 +557,7 @@ class InfoGatherer:
            try:
                curr = await RdmaCtlStatus.gather()
                if curr is not None:
-                    await self.send(curr)
+                    await self.info_sender.send(curr)
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering RDMA ctl status")
            await anyio.sleep(rdma_ctl_poll_interval)
@@ -574,7 +566,7 @@ class InfoGatherer:
        while True:
            try:
                with fail_after(5):
-                    await self.send(await NodeDiskUsage.gather())
+                    await self.info_sender.send(await NodeDiskUsage.gather())
            except Exception as e:
                logger.opt(exception=e).warning("Error gathering disk usage")
            await anyio.sleep(disk_poll_interval)
@@ -619,7 +611,7 @@ class InfoGatherer:
                            )
                            text = data.decode("utf-8", errors="replace").strip()
                            metrics = MacmonMetrics.from_raw_json(text)
-                        await self.send(metrics)
+                        await self.info_sender.send(metrics)
            except TimeoutError:
                logger.warning(
                    f"MacMon produced no output for {read_timeout}s, restarting"
@@ -19,22 +19,18 @@ class FrozenModel(BaseModel):


 class TaggedModel(FrozenModel):
-    @classmethod
-    def tag(cls) -> str:
-        return cls.__name__
-
    @model_serializer(mode="wrap")
    def _serialize(self, handler: SerializerFunctionWrapHandler):
        inner = handler(self)  # pyright: ignore[reportAny]
-        return {self.tag(): inner}
+        return {self.__class__.__name__: inner}

    @model_validator(mode="wrap")
    @classmethod
    def _validate(cls, v: Any, handler: ValidatorFunctionWrapHandler) -> Self:  # pyright: ignore[reportAny]
-        if isinstance(v, dict) and len(v) == 1 and cls.tag() in v:  # pyright: ignore[reportUnknownArgumentType]
-            return handler(v[cls.tag()])  # pyright: ignore[reportAny]
+        if isinstance(v, dict) and len(v) == 1 and cls.__name__ in v:  # pyright: ignore[reportUnknownArgumentType]
+            return handler(v[cls.__name__])  # pyright: ignore[reportAny]

        return handler(v)  # pyright: ignore[reportAny]

    def __str__(self) -> str:
-        return f"{self.tag()}({super().__str__()})"
+        return f"{self.__class__.__name__}({super().__str__()})"
@@ -4,9 +4,7 @@ from datetime import datetime, timezone

 import anyio
 from anyio import fail_after, to_thread
-from exo_rs import LVAggregator, SessionHandle
 from loguru import logger
-from pydantic import ValidationError

 from exo.api.types import ImageEditsTaskParams
 from exo.download.download_utils import is_read_only_model_dir, resolve_existing_model
@@ -16,8 +14,7 @@ from exo.routing.event_router import (
 )
 from exo.shared.apply import apply
 from exo.shared.constants import EXO_MAX_INSTANCE_RETRIES
-from exo.shared.models import model_cards
-from exo.shared.models.model_cards import ModelCard, ModelId
+from exo.shared.models.model_cards import ModelId, card_cache
 from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.commands import (
    DeleteInstance,
@@ -56,7 +53,7 @@ from exo.shared.types.topology import Connection, SocketConnection
 from exo.shared.types.worker.downloads import DownloadCompleted
 from exo.shared.types.worker.instances import InstanceId
 from exo.shared.types.worker.runners import RunnerId
-from exo.utils.channels import Receiver, Sender
+from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo, InfoGatherer
 from exo.utils.info_gatherer.net_profile import check_reachable
 from exo.utils.keyed_backoff import KeyedBackoff
@@ -76,7 +73,6 @@ class Worker:
        # but I think it's the correct way to be thinking about commands
        command_sender: Sender[ForwarderCommand],
        download_command_sender: Sender[ForwarderDownloadCommand],
-        session_handle: SessionHandle,
        api_port: int,
    ):
        self.node_id: NodeId = node_id
@@ -102,17 +98,17 @@ class Worker:
            base=0.5, cap=10.0
        )
        self._stopped: anyio.Event = anyio.Event()
-        self._sh: SessionHandle = session_handle
-        self.aggregator: LVAggregator = session_handle.last_value_aggregator("metrics")

    async def run(self):
        logger.info("Starting Worker")

-        info_gatherer: InfoGatherer = InfoGatherer(self._sh, self.node_id)
+        info_send, info_recv = channel[GatheredInfo]()
+        info_gatherer: InfoGatherer = InfoGatherer(info_send)

        try:
            async with self._tg as tg:
                tg.start_soon(info_gatherer.run)
+                tg.start_soon(self._forward_info, info_recv)
                tg.start_soon(self.plan_step)
                tg.start_soon(self._event_applier)
                tg.start_soon(self._poll_connection_updates)
@@ -182,24 +178,17 @@ class Worker:
                            ] = img

    async def _reconcile_custom_cards(self) -> None:
-        storage = self._sh.storage_interface()
        while True:
-            await anyio.sleep(10)
-            target: list[ModelId] = []
-            for _, value in (await storage.dump("custom_model_cards/")).items():
-                try:
-                    card = ModelCard.model_validate_json(value)
-                except ValidationError:
+            await anyio.sleep(1)
+            target = dict(self.state.custom_model_cards)
+            for model_id, card in target.items():
+                if card_cache.get(model_id) == card:
                    continue
-                target.append(card.model_id)
-                if model_cards.card_cache.get(card.model_id) == card:
-                    continue
-                logger.info(f"Registered new custom model card for {card.model_id}")
-                await model_cards.card_cache.save(card)
+                await card_cache.save(card)

-            for card in await model_cards.card_cache.list_all():
+            for card in await card_cache.list_all():
                if card.model_id not in target:
-                    await model_cards.card_cache.delete(card.model_id)
+                    await card_cache.pop(card.model_id)

    async def plan_step(self):
        while True:
@@ -399,13 +388,14 @@ class Worker:

    async def _poll_connection_updates(self):
        while True:
-            state = self.state.with_aggregator(self.aggregator)
-            edges = set(conn.edge for conn in state.topology.out_edges(self.node_id))
+            edges = set(
+                conn.edge for conn in self.state.topology.out_edges(self.node_id)
+            )
            conns: defaultdict[NodeId, set[str]] = defaultdict(set)
            async for ip, nid in check_reachable(
-                state.topology,
+                self.state.topology,
                self.node_id,
-                state.node_network,
+                self.state.node_network,
                api_port=self.api_port,
            ):
                if ip in conns[nid]:
@@ -426,7 +416,7 @@ class Worker:
                        )
                    )

-            for conn in state.topology.out_edges(self.node_id):
+            for conn in self.state.topology.out_edges(self.node_id):
                if not isinstance(conn.edge, SocketConnection):
                    continue
                # ignore mDNS discovered connections
@@ -42,7 +42,7 @@ i=0
 for host; do
  colour=${colours[i++ % 4]}
  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
-    "ENABLE_DISAGGREGATION=true EXO_ZENOH_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix run $remote_installable" 2>&1 |
+    "EXO_ZENOH_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix run $remote_installable" 2>&1 |
    awk -v p="${colour}[${host}]${reset}" '{ print p $0; fflush() }' &
 done

@@ -450,7 +450,7 @@ 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.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.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", 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 +461,7 @@ mlx = [
 ]
 mlx-cpu = [
    { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { 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.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", 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 +473,7 @@ mlx-cpu = [
 ]
 mlx-cuda12 = [
    { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { 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.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", 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 +487,7 @@ mlx-cuda12 = [
 ]
 mlx-cuda13 = [
    { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { 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.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", 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')" },
@@ -1343,7 +1343,7 @@ 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.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.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", 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 +1374,7 @@ sdist = { url = "https://files.pythonhosted.org/packages/55/fa/96d4cc7ada2833571

 [[package]]
 name = "mlx"
-version = "0.32.0.dev20260512+cc3f3e60"
+version = "0.32.0.dev20260506+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 +1549,7 @@ 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.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.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 = "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 +1565,7 @@ 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.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.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", 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')" },