cleaning up the todos

auto bench (#1405 )
runs exo_bench remotely with some nice git QoL ## usage run tests/auto_bench.sh host1 [host2] exo bench will be run on those hosts and its output saved to bench/commit_hash/*.json on all models currently downloaded
2026-02-06 20:21:39 -05:00 · 2026-02-06 15:38:56 +00:00 · 2026-02-06 15:35:46 +00:00 · 2026-02-06 15:28:10 +00:00 · 2026-02-06 13:26:59 +00:00 · 2026-02-06 13:06:59 +00:00
64 changed files with 5162 additions and 1956 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -32,3 +32,6 @@ dashboard/.svelte-kit/
 # host config snapshots
 hosts_*.json
 .swp
+
+# bench files
+bench/**/*.json
--- a/.mlx_typings/mlx/core/init.pyi
+++ b/.mlx_typings/mlx/core/init.pyi
@@ -1139,7 +1139,7 @@ class array:
    ) -> array:
        """See :func:`flatten`."""

-    def reshape(self, *shape, stream: Stream | Device | None = ...) -> array:
+    def reshape(self, *shape: int, stream: Stream | Device | None = ...) -> array:
        """
        Equivalent to :func:`reshape` but the shape can be passed either as a
        :obj:`tuple` or as separate arguments.
@@ -1222,7 +1222,7 @@ class array:
    ) -> array:
        """See :func:`swapaxes`."""

-    def transpose(self, *axes, stream: Stream | Device | None = ...) -> array:
+    def transpose(self, *axes: int, stream: Stream | Device | None = ...) -> array:
        """
        Equivalent to :func:`transpose` but the axes can be passed either as
        a tuple or as separate arguments.
--- a/.mlx_typings/mlx/nn/layers/convolution.pyi
+++ b/.mlx_typings/mlx/nn/layers/convolution.pyi
@@ -30,6 +30,9 @@ class Conv1d(Module):
        bias (bool, optional): If ``True`` add a learnable bias to the output.
            Default: ``True``
    """
+
+    weight: mx.array
+    groups: int
    def __init__(
        self,
        in_channels: int,
--- a/.mlx_typings/mlx_lm/models/cache.pyi
+++ b/.mlx_typings/mlx_lm/models/cache.pyi
@@ -11,7 +11,10 @@ import mlx.core as mx
 class Cache(Protocol):
    keys: mx.array
    values: mx.array
-    def update_and_fetch(self, keys: mx.array, values: mx.array) -> None: ...
+    offset: int
+    def update_and_fetch(
+        self, keys: mx.array, values: mx.array
+    ) -> tuple[mx.array, mx.array]: ...
    @property
    def state(self) -> tuple[mx.array, mx.array]: ...
    @state.setter
@@ -87,6 +90,7 @@ def create_attention_mask(
 class _BaseCache(Cache):
    keys: mx.array
    values: mx.array
+    offset: int
    @property
    def state(self) -> tuple[mx.array, mx.array]: ...
    @state.setter
--- a/.mlx_typings/mlx_lm/models/deepseek_v3.pyi
+++ b/.mlx_typings/mlx_lm/models/deepseek_v3.pyi
@@ -5,6 +5,7 @@ from typing import Any, Dict, Optional

 import mlx.core as mx
 import mlx.nn as nn
+from mlx_lm.models.mla import MultiLinear

 from .base import BaseModelArgs
 from .switch_layers import SwitchGLU
@@ -60,7 +61,10 @@ class DeepseekV3Attention(nn.Module):
    q_b_proj: nn.Linear
    kv_a_proj_with_mqa: nn.Linear
    kv_a_layernorm: nn.RMSNorm
-    kv_b_proj: nn.Linear
+    # kv_b_proj: nn.Linear
+    embed_q: MultiLinear
+    unembed_out: MultiLinear
+
    o_proj: nn.Linear
    rope: Any

--- a/.mlx_typings/mlx_lm/models/qwen3_next.pyi
+++ b/.mlx_typings/mlx_lm/models/qwen3_next.pyi
@@ -0,0 +1,114 @@
+"""Type stubs for mlx_lm.models.qwen3_next"""
+
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .switch_layers import SwitchGLU
+
+class Qwen3NextMLP(nn.Module):
+    gate_proj: nn.Linear
+    down_proj: nn.Linear
+    up_proj: nn.Linear
+
+    def __init__(self, dim: int, hidden_dim: int) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Qwen3NextGatedDeltaNet(nn.Module):
+    hidden_size: int
+    num_v_heads: int
+    num_k_heads: int
+    head_k_dim: int
+    head_v_dim: int
+    key_dim: int
+    value_dim: int
+    conv_kernel_size: int
+    conv_dim: int
+    conv1d: nn.Conv1d
+    in_proj_qkvz: nn.Linear
+    in_proj_ba: nn.Linear
+    dt_bias: mx.array
+    A_log: mx.array
+    out_proj: nn.Linear
+
+    def __init__(self, config: Any) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Qwen3NextAttention(nn.Module):
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: int
+    scale: float
+    q_proj: nn.Linear
+    k_proj: nn.Linear
+    v_proj: nn.Linear
+    o_proj: nn.Linear
+
+    def __init__(self, args: Any) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Qwen3NextSparseMoeBlock(nn.Module):
+    norm_topk_prob: bool
+    num_experts: int
+    top_k: int
+    gate: nn.Linear
+    switch_mlp: SwitchGLU
+    shared_expert: Qwen3NextMLP
+    shared_expert_gate: nn.Linear
+
+    def __init__(self, args: Any) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Qwen3NextDecoderLayer(nn.Module):
+    is_linear: bool
+    linear_attn: Qwen3NextGatedDeltaNet
+    self_attn: Qwen3NextAttention
+    input_layernorm: nn.RMSNorm
+    post_attention_layernorm: nn.RMSNorm
+    mlp: Qwen3NextMLP | Qwen3NextSparseMoeBlock
+
+    def __init__(self, args: Any, layer_idx: int) -> None: ...
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Qwen3NextModel(nn.Module):
+    embed_tokens: nn.Embedding
+    layers: list[Qwen3NextDecoderLayer]
+    norm: nn.RMSNorm
+
+    def __init__(self, args: Any) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+
+class Model(nn.Module):
+    model_type: str
+    model: Qwen3NextModel
+    lm_head: nn.Linear
+
+    def __init__(self, args: Any) -> None: ...
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache: Optional[Any] = None,
+    ) -> mx.array: ...
+    def sanitize(self, weights: dict[str, Any]) -> dict[str, Any]: ...
+    @property
+    def layers(self) -> list[Qwen3NextDecoderLayer]: ...
--- a/.mlx_typings/mlx_lm/tokenizer_utils.pyi
+++ b/.mlx_typings/mlx_lm/tokenizer_utils.pyi
@@ -113,6 +113,10 @@ class TokenizerWrapper:
    bos_token: str | None
    vocab_size: int
    all_special_tokens: list[str]
+    think_start: str | None
+    think_end: str | None
+    think_start_id: int | None
+    think_end_id: int | None

    def __init__(
        self,
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -141,6 +141,12 @@ version = "0.3.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "76a2e8124351fda1ef8aaaa3bbd7ebbcb486bbcd4225aca0aa0d84bb2db8fecb"

+[[package]]
+name = "arrayvec"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
+
 [[package]]
 name = "asn1-rs"
 version = "0.7.1"
@@ -298,6 +304,19 @@ version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "55248b47b0caf0546f7988906588779981c43bb1bc9d0c44087278f80cdb44ba"

+[[package]]
+name = "bigdecimal"
+version = "0.4.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "560f42649de9fa436b73517378a147ec21f6c997a546581df4b4b31677828934"
+dependencies = [
+ "autocfg",
+ "libm",
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "bimap"
 version = "0.6.3"
@@ -334,6 +353,31 @@ dependencies = [
 "generic-array",
 ]

+[[package]]
+name = "bon"
+version = "3.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebeb9aaf9329dff6ceb65c689ca3db33dbf15f324909c60e4e5eef5701ce31b1"
+dependencies = [
+ "bon-macros",
+ "rustversion",
+]
+
+[[package]]
+name = "bon-macros"
+version = "3.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "77e9d642a7e3a318e37c2c9427b5a6a48aa1ad55dcd986f3034ab2239045a645"
+dependencies = [
+ "darling",
+ "ident_case",
+ "prettyplease",
+ "proc-macro2",
+ "quote",
+ "rustversion",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "bs58"
 version = "0.5.1"
@@ -497,6 +541,15 @@ version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2f421161cb492475f1661ddc9815a745a1c894592070661180fdec3d4872e9c3"

+[[package]]
+name = "convert_case"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "633458d4ef8c78b72454de2d54fd6ab2e60f9e02be22f3c6104cdc8a4e0fceb9"
+dependencies = [
+ "unicode-segmentation",
+]
+
 [[package]]
 name = "core-foundation"
 version = "0.9.4"
@@ -647,6 +700,41 @@ dependencies = [
 "syn 2.0.111",
 ]

+[[package]]
+name = "darling"
+version = "0.21.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9cdf337090841a411e2a7f3deb9187445851f91b309c0c0a29e05f74a00a48c0"
+dependencies = [
+ "darling_core",
+ "darling_macro",
+]
+
+[[package]]
+name = "darling_core"
+version = "0.21.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1247195ecd7e3c85f83c8d2a366e4210d588e802133e1e355180a9870b517ea4"
+dependencies = [
+ "fnv",
+ "ident_case",
+ "proc-macro2",
+ "quote",
+ "strsim",
+ "syn 2.0.111",
+]
+
+[[package]]
+name = "darling_macro"
+version = "0.21.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d38308df82d1080de0afee5d069fa14b0326a88c14f15c5ccda35b4a6c414c81"
+dependencies = [
+ "darling_core",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "data-encoding"
 version = "2.9.0"
@@ -673,6 +761,17 @@ dependencies = [
 "syn 2.0.111",
 ]

+[[package]]
+name = "delegate"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "780eb241654bf097afb00fc5f054a09b687dad862e485fdcf8399bb056565370"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "der"
 version = "0.7.10"
@@ -707,6 +806,29 @@ dependencies = [
 "powerfmt",
 ]

+[[package]]
+name = "derive_more"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "10b768e943bed7bf2cab53df09f4bc34bfd217cdb57d971e769874c9a6710618"
+dependencies = [
+ "derive_more-impl",
+]
+
+[[package]]
+name = "derive_more-impl"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d286bfdaf75e988b4a78e013ecd79c581e06399ab53fbacd2d916c2f904f30b"
+dependencies = [
+ "convert_case",
+ "proc-macro2",
+ "quote",
+ "rustc_version",
+ "syn 2.0.111",
+ "unicode-xid",
+]
+
 [[package]]
 name = "digest"
 version = "0.10.7"
@@ -876,23 +998,37 @@ dependencies = [
 name = "exo_pyo3_bindings"
 version = "0.0.1"
 dependencies = [
+ "delegate",
+ "derive_more",
 "env_logger",
- "futures-lite",
+ "extend",
+ "futures",
+ "impl-trait-for-tuples",
 "libp2p",
 "log",
 "networking",
+ "once_cell",
+ "pin-project",
 "pyo3",
 "pyo3-async-runtimes",
 "pyo3-log",
 "pyo3-stub-gen",
+ "thiserror 2.0.17",
+ "thread_local",
 "tokio",
+ "util",
 ]

 [[package]]
-name = "fastrand"
-version = "2.3.0"
+name = "extend"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
+checksum = "311a6d2f1f9d60bff73d2c78a0af97ed27f79672f15c238192a5bbb64db56d00"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]

 [[package]]
 name = "ff"
@@ -1002,10 +1138,7 @@ version = "2.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f78e10609fe0e0b3f4157ffab1876319b5b0db102a2c60dc4626306dc46b44ad"
 dependencies = [
- "fastrand",
 "futures-core",
- "futures-io",
- "parking",
 "pin-project-lite",
 ]

@@ -1492,6 +1625,12 @@ dependencies = [
 "zerovec",
 ]

+[[package]]
+name = "ident_case"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9e0384b61958566e926dc50660321d12159025e767c18e043daf26b70104c39"
+
 [[package]]
 name = "idna"
 version = "1.1.0"
@@ -1567,6 +1706,17 @@ dependencies = [
 "xmltree",
 ]

+[[package]]
+name = "impl-trait-for-tuples"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a0eb5a3343abf848c0984fe4604b2b105da9539376e24fc0a3b0007411ae4fd9"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "indexmap"
 version = "2.12.1"
@@ -1595,6 +1745,15 @@ dependencies = [
 "generic-array",
 ]

+[[package]]
+name = "internment"
+version = "0.8.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "636d4b0f6a39fd684effe2a73f5310df16a3fa7954c26d36833e98f44d1977a2"
+dependencies = [
+ "hashbrown 0.15.5",
+]
+
 [[package]]
 name = "inventory"
 version = "0.3.21"
@@ -1721,6 +1880,12 @@ dependencies = [
 "cpufeatures",
 ]

+[[package]]
+name = "keccak-const"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "57d8d8ce877200136358e0bbff3a77965875db3af755a11e1fa6b1b3e2df13ea"
+
 [[package]]
 name = "lalrpop-util"
 version = "0.20.2"
@@ -1739,6 +1904,12 @@ version = "0.2.178"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37c93d8daa9d8a012fd8ab92f088405fb202ea0b6ab73ee2482ae66af4f42091"

+[[package]]
+name = "libm"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f9fbbcab51052fe104eb5e5d351cf728d30a5be1fe14d9be8a3b097481fb97de"
+
 [[package]]
 name = "libp2p"
 version = "0.56.0"
@@ -2727,10 +2898,20 @@ dependencies = [
 name = "networking"
 version = "0.0.1"
 dependencies = [
+ "delegate",
+ "derive_more",
+ "either",
+ "extend",
+ "futures",
+ "futures-timer",
+ "impl-trait-for-tuples",
+ "keccak-const",
 "libp2p",
 "log",
+ "thiserror 2.0.17",
 "tokio",
 "tracing-subscriber",
+ "util",
 ]

 [[package]]
@@ -2812,6 +2993,17 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "num-rational"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f83d14da390562dca69fc84082e73e548e1ad308d24accdedd2720017cb37824"
+dependencies = [
+ "num-bigint",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "num-traits"
 version = "0.2.19"
@@ -3115,6 +3307,16 @@ dependencies = [
 "zerocopy",
 ]

+[[package]]
+name = "prettyplease"
+version = "0.2.37"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
+dependencies = [
+ "proc-macro2",
+ "syn 2.0.111",
+]
+
 [[package]]
 name = "primeorder"
 version = "0.13.6"
@@ -3162,14 +3364,28 @@ version = "0.27.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab53c047fcd1a1d2a8820fe84f05d6be69e9526be40cb03b73f86b6b03e6d87d"
 dependencies = [
+ "bigdecimal",
+ "either",
+ "hashbrown 0.16.1",
+ "indexmap",
 "indoc",
+ "inventory",
 "libc",
+ "lock_api",
 "memoffset",
+ "num-bigint",
+ "num-complex",
+ "num-rational",
+ "num-traits",
 "once_cell",
+ "ordered-float",
+ "parking_lot",
 "portable-atomic",
 "pyo3-build-config",
 "pyo3-ffi",
 "pyo3-macros",
+ "rust_decimal",
+ "smallvec",
 "unindent",
 ]

@@ -3524,6 +3740,12 @@ dependencies = [
 "yasna",
 ]

+[[package]]
+name = "recursion"
+version = "0.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9dba2197bf7b1d87b4dd460c195f4edeb45a94e82e8054f8d5f317c1f0e93ca1"
+
 [[package]]
 name = "redox_syscall"
 version = "0.5.18"
@@ -3610,6 +3832,16 @@ dependencies = [
 "tokio",
 ]

+[[package]]
+name = "rust_decimal"
+version = "1.39.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "35affe401787a9bd846712274d97654355d21b2a2c092a3139aabe31e9022282"
+dependencies = [
+ "arrayvec",
+ "num-traits",
+]
+
 [[package]]
 name = "rustc-hash"
 version = "1.1.0"
@@ -4474,12 +4706,24 @@ version = "1.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9312f7c4f6ff9069b165498234ce8be658059c6728633667c526e27dc2cf1df5"

+[[package]]
+name = "unicode-segmentation"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6ccf251212114b54433ec949fd6a7841275f9ada20dddd2f29e9ceea4501493"
+
 [[package]]
 name = "unicode-width"
 version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b4ac048d71ede7ee76d585517add45da530660ef4390e49b098733c6e897f254"

+[[package]]
+name = "unicode-xid"
+version = "0.2.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
+
 [[package]]
 name = "unicode_names2"
 version = "1.3.0"
@@ -4560,6 +4804,19 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"

+[[package]]
+name = "util"
+version = "0.0.1"
+dependencies = [
+ "bon",
+ "derive_more",
+ "extend",
+ "internment",
+ "once_cell",
+ "recursion",
+ "thiserror 2.0.17",
+]
+
 [[package]]
 name = "uuid"
 version = "1.19.0"
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,6 +3,7 @@ resolver = "3"
 members = [
    "rust/networking",
    "rust/exo_pyo3_bindings",
+    "rust/util",
 ]

 [workspace.package]
@@ -23,18 +24,62 @@ opt-level = 3
 [workspace.dependencies]
 ## Crate members as common dependencies
 networking = { path = "rust/networking" }
+util = { path = "rust/util" }
+
+# Proc-macro authoring tools
+syn = "2.0"
+quote = "1.0"
+proc-macro2 = "1.0"
+darling = "0.20"
+
+# Macro dependecies
+extend = "1.2"
+delegate = "0.13"
+impl-trait-for-tuples = "0.2"
+clap = "4.5"
+derive_more = { version = "2.0.1", features = ["display"] }
+pin-project = "1"
+
+# Utility dependencies
+itertools = "0.14"
+thiserror = "2"
+internment = "0.8"
+recursion = "0.5"
+regex = "1.11"
+once_cell = "1.21"
+thread_local = "1.1"
+bon = "3.4"
+generativity = "1.1"
+anyhow = "1.0"
+keccak-const = "0.2"
+
+# Functional generics/lenses frameworks
+frunk_core = "0.4"
+frunk = "0.4"
+frunk_utils = "0.2"
+frunk-enum-core = "0.3"

 # Async dependencies
 tokio = "1.46"
+futures = "0.3"
+futures-util = "0.3"
+futures-timer = "3.0"
+
+# Data structures
+either = "1.15"
+ordered-float = "5.0"
+ahash = "0.8"

 # Tracing/logging
 log = "0.4"

 # networking
 libp2p = "0.56"
+libp2p-tcp = "0.44"

 [workspace.lints.rust]
-static_mut_refs = "warn"
+static_mut_refs = "warn"      # Or use "warn" instead of deny
+incomplete_features = "allow"

 # Clippy's lint category level configurations;
 # every member crate needs to inherit these by adding
@@ -55,3 +100,64 @@ perf = { level = "warn", priority = -1 }
 pedantic = { level = "warn", priority = -1 }
 nursery = { level = "warn", priority = -1 }
 cargo = { level = "warn", priority = -1 }
+
+# Individual Clippy lints from the `restriction` category
+arithmetic_side_effects = "warn"
+as_conversions = "warn"
+assertions_on_result_states = "warn"
+clone_on_ref_ptr = "warn"
+decimal_literal_representation = "warn"
+default_union_representation = "warn"
+deref_by_slicing = "warn"
+disallowed_script_idents = "deny"
+else_if_without_else = "warn"
+empty_enum_variants_with_brackets = "warn"
+empty_structs_with_brackets = "warn"
+error_impl_error = "warn"
+exit = "deny"
+expect_used = "warn"
+float_cmp_const = "warn"
+get_unwrap = "warn"
+if_then_some_else_none = "warn"
+impl_trait_in_params = "warn"
+indexing_slicing = "warn"
+infinite_loop = "warn"
+let_underscore_must_use = "warn"
+let_underscore_untyped = "warn"
+lossy_float_literal = "warn"
+mem_forget = "warn"
+missing_inline_in_public_items = "warn"
+multiple_inherent_impl = "warn"
+multiple_unsafe_ops_per_block = "warn"
+mutex_atomic = "warn"
+non_zero_suggestions = "warn"
+panic = "warn"
+partial_pub_fields = "warn"
+pattern_type_mismatch = "warn"
+pub_without_shorthand = "warn"
+rc_buffer = "warn"
+rc_mutex = "warn"
+redundant_type_annotations = "warn"
+renamed_function_params = "warn"
+rest_pat_in_fully_bound_structs = "warn"
+same_name_method = "warn"
+self_named_module_files = "deny"
+semicolon_inside_block = "warn"
+shadow_same = "warn"
+shadow_unrelated = "warn"
+str_to_string = "warn"
+string_add = "warn"
+string_lit_chars_any = "warn"
+string_to_string = "warn"
+tests_outside_test_module = "warn"
+todo = "warn"
+try_err = "warn"
+undocumented_unsafe_blocks = "warn"
+unnecessary_safety_comment = "warn"
+unnecessary_safety_doc = "warn"
+unneeded_field_pattern = "warn"
+unseparated_literal_suffix = "warn"
+unused_result_ok = "warn"
+unused_trait_names = "warn"
+unwrap_used = "warn"
+verbose_file_reads = "warn"
--- a/TODO.md
+++ b/TODO.md
@@ -1,28 +1,14 @@
-2. Currently a lot of requests from the API are timing out, but we still process those requests internally. If an API request times out, we should cancel all corresponding tasks to that API request (why process a request with nobody listening).
 3. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 4. I'd like to see profiled network latency / bandwidth.
 5. I'd like to see how much bandwidth each link is using.
-6. We should handle the case where one machine doesn't have the model downloaded and then other machines are waiting on it. In this case we get loads of timeout errors because the others are waiting for the one that needs to download the model.
 7. Solve the problem of in continuous batching when a new prompt comes in, it will block decode of the current batch until the prefill is complete.
 8. We want people to be able to copy models over to a new device without ever connecting EXO to the internet. Right now EXO require internet connection once to cache some files to check if a download is complete. Instead, we should simply check if there is a non-empty model folder locally with no .partial files. This indicates it's a fully downloaded model that can be loaded.
-10. More granular control over how to deploy instances.
-12. Nix is great but installing it is a pain and we have ended up in a lot of cases having PATH issues or installation issues. For example, after rebooting mike it seemed to no longer have a nix installation and needed reinstalling. It has a bunch of broken symlinks left over from nix that caused ssh to fail, making it even harder to debug. We need consistent environments (perhaps MDM) so we can guarantee nix is installed properly on each machine.
 13. Memory pressure instead of memory used.
 14. Show the type of each connection (TB5, Ethernet, etc.) in the UI. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 15. Prioritise certain connection types (or by latency). TB5 > Ethernet > WiFi. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
 16. Dynamically switch to higher priority connection when it becomes available. Probably bring back InstanceReplacedAtomically.
 17. Faster model loads by streaming model from other devices in cluster.
 18. Add support for specifying the type of network connection to use in a test. Depends on 15/16.
-20. Add chat completion cancellations (e.g OpenWebUI has something for cancelling an ongoing request).
-23. Do we need cache_limit? We went back and forth on that a lot because we thought it might be causing issues. One problem is it sets it relative to model size. So if you have multiple models loaded in it will take the most recent model size for the cache_limit. This is problematic if you launch DeepSeek -> Llama for example.
-24. further openai/lmstudio api compatibility
 25. Rethink retry logic
-26. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
 27. Log cleanup - per-module log filters and default to DEBUG log levels
 28. Validate RDMA connections with ibv_devinfo in the info gatherer
-
-Potential refactors:
-
-2. Topology can be simplified
-
-Random errors we've run into:
--- a/bench/exo_bench.py
+++ b/bench/exo_bench.py
@@ -431,7 +431,12 @@ def main() -> int:
    ap.add_argument(
        "--skip-pipeline-jaccl",
        action="store_true",
-        help="Pipeline jaccl is often pointless, skip by default",
+        help="Skip pipeline+jaccl placements, as it's often pointless.",
+    )
+    ap.add_argument(
+        "--skip-tensor-ring",
+        action="store_true",
+        help="Skip tensor+ring placements, as it's so slow.",
    )
    ap.add_argument(
        "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
@@ -450,6 +455,7 @@ def main() -> int:
        default="bench/results.json",
        help="Write raw per-run results JSON to this path.",
    )
+    ap.add_argument("--stdout", action="store_true", help="Write results to stdout")
    ap.add_argument(
        "--dry-run", action="store_true", help="List selected placements and exit."
    )
@@ -533,6 +539,16 @@ def main() -> int:
        ):
            continue

+        if (
+            args.skip_tensor_ring
+            and (
+                args.instance_meta == "both"
+                and "ring" in p.get("instance_meta", "").lower()
+            )
+            and (args.sharding == "both" and "tensor" in p.get("sharding", "").lower())
+        ):
+            continue
+
        if args.min_nodes <= n <= args.max_nodes:
            selected.append(p)

@@ -652,7 +668,9 @@ def main() -> int:

            time.sleep(5)

-    if args.json_out:
+    if args.stdout:
+        json.dump(all_rows, sys.stdout, indent=2, ensure_ascii=False)
+    elif args.json_out:
        with open(args.json_out, "w", encoding="utf-8") as f:
            json.dump(all_rows, f, indent=2, ensure_ascii=False)
        logger.debug(f"\nWrote results JSON: {args.json_out}")
--- a/2
+++ b/2
@@ -20,7 +20,7 @@ sync-clean:

 rust-rebuild:
    cargo run --bin stub_gen
-    just sync-clean
+    uv sync --reinstall-package exo_pyo3_bindings

 build-dashboard:
    #!/usr/bin/env bash
--- a/nix/mlx.nix
+++ b/nix/mlx.nix
@@ -41,7 +41,7 @@ let

  mlx = stdenv.mkDerivation rec {
    pname = "mlx";
-    version = let v = "0.30.4"; in
+    version = let v = "0.30.5"; in
      assert v == uvLockMlxVersion || throw "MLX version mismatch: nix/mlx.nix has ${v} but uv.lock has ${uvLockMlxVersion}. Update both the version and hash in nix/mlx.nix.";
      v;
    pyproject = true;
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -17,9 +17,9 @@ dependencies = [
    "loguru>=0.7.3",
    "exo_pyo3_bindings", # rust bindings
    "anyio==4.11.0",
-    "mlx==0.30.4; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.4; sys_platform == 'linux'",
-    "mlx-lm",
+    "mlx==0.30.5; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.5; sys_platform == 'linux'",
+    "mlx-lm==0.30.6",
    "tiktoken>=0.12.0", # required for kimi k2 tokenizer
    "hypercorn>=0.18.0",
    "openai-harmony>=0.0.8",
@@ -31,8 +31,6 @@ dependencies = [
 ]

 [project.scripts]
-exo-master = "exo.master.main:main"
-exo-worker = "exo.worker.main:main"
 exo = "exo.main:main"

 # dependencies only required for development
@@ -63,7 +61,7 @@ members = [

 [tool.uv.sources]
 exo_pyo3_bindings = { workspace = true }
-mlx-lm = { git = "https://github.com/ml-explore/mlx-lm", branch = "main" }
+#mlx-lm = { git = "https://github.com/davidmcc73/mlx-lm", branch = "stable" }
 # Uncomment to use local mlx/mlx-lm development versions:
 # mlx = { path = "/Users/Shared/mlx", editable=true }
 # mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }
@@ -105,6 +103,7 @@ root = "src"

 # supported platforms for this project
 [tool.uv]
+required-version = ">=0.8.6"
 prerelease = "allow"
 environments = [
    "sys_platform == 'darwin'",
--- a/python/parts.nix
+++ b/python/parts.nix
@@ -59,6 +59,22 @@
        }
      );

+      mkPythonScript = name: path: pkgs.writeShellApplication {
+        inherit name;
+        runtimeInputs = [ exoVenv ];
+        runtimeEnv = {
+          EXO_DASHBOARD_DIR = self'.packages.dashboard;
+          EXO_RESOURCES_DIR = inputs.self + /resources;
+        };
+        text = ''exec python ${path} "$@"'';
+      };
+
+      mkSimplePythonScript = name: path: pkgs.writeShellApplication {
+        inherit name;
+        runtimeInputs = [ pkgs.python313 ];
+        text = ''exec python ${path} "$@"'';
+      };
+
      exoPackage = pkgs.runCommand "exo"
        {
          nativeBuildInputs = [ pkgs.makeWrapper ];
@@ -66,28 +82,30 @@
        ''
          mkdir -p $out/bin

-          # Create wrapper scripts
-          for script in exo exo-master exo-worker; do
-            makeWrapper ${exoVenv}/bin/$script $out/bin/$script \
-              --set EXO_DASHBOARD_DIR ${self'.packages.dashboard} \
-              --set EXO_RESOURCES_DIR ${inputs.self + "/resources"} \
-              ${lib.optionalString pkgs.stdenv.isDarwin "--prefix PATH : ${pkgs.macmon}/bin"}
-          done
+          # Create wrapper script
+          makeWrapper ${exoVenv}/bin/exo $out/bin/exo \
+            --set EXO_DASHBOARD_DIR ${self'.packages.dashboard} \
+            --set EXO_RESOURCES_DIR ${inputs.self + /resources} \
+            ${lib.optionalString pkgs.stdenv.hostPlatform.isDarwin "--prefix PATH : ${pkgs.macmon}/bin"}
        '';
    in
    {
      # Python package only available on macOS (requires MLX/Metal)
-      packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
-        exo = exoPackage;
-        # Test environment for running pytest outside of Nix sandbox (needs GPU access)
-        exo-test-env = testVenv;
+      packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin
+        {
+          exo = exoPackage;
+          # Test environment for running pytest outside of Nix sandbox (needs GPU access)
+          exo-test-env = testVenv;
+          exo-bench = mkPythonScript "exo-bench" (inputs.self + /bench/exo_bench.py);
+        } // {
+        exo-get-all-models-on-cluster = mkSimplePythonScript "exo-get-all-models-on-cluster" (inputs.self + /tests/get_all_models_on_cluster.py);
      };

      checks = {
        # Ruff linting (works on all platforms)
        lint = pkgs.runCommand "ruff-lint" { } ''
          export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
-          ${pkgs.ruff}/bin/ruff check ${inputs.self}/
+          ${pkgs.ruff}/bin/ruff check ${inputs.self}
          touch $out
        '';
      };
--- a/rust/clippy.toml
+++ b/rust/clippy.toml
@@ -0,0 +1,2 @@
+# we can manually exclude false-positive lint errors for dual packages (if in dependencies)
+#allowed-duplicate-crates = ["hashbrown"]
--- a/rust/exo_pyo3_bindings/Cargo.toml
+++ b/rust/exo_pyo3_bindings/Cargo.toml
@@ -25,25 +25,44 @@ workspace = true
 networking = { workspace = true }

 # interop
-pyo3 = { version = "0.27.2", features = [
-    "abi3-py313", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.11
-    # "nightly", # enables better-supported GIL integration
-    "experimental-async" # async support in #[pyfunction] & #[pymethods]
-    # "experimental-inspect", # inspection of generated binary => easier to automate type-hint generation
-    # "py-clone", # adding Clone-ing of `Py<T>` without GIL (may cause panics - remove if panics happen)
-    # "multiple-pymethods", # allows multiple #[pymethods] sections per class
+pyo3 = { version = "0.27.1", features = [
+    # "abi3-py311", # tells pyo3 (and maturin) to build using the stable ABI with minimum Python version 3.11
+    "nightly", # enables better-supported GIL integration
+    "experimental-async", # async support in #[pyfunction] & #[pymethods]
+    #"experimental-inspect", # inspection of generated binary => easier to automate type-hint generation
+    #"py-clone", # adding Clone-ing of `Py<T>` without GIL (may cause panics - remove if panics happen)
+    "multiple-pymethods", # allows multiple #[pymethods] sections per class

    # integrations with other libraries
-    # "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
-    # "ordered-float", "rust_decimal", "smallvec",
+    "arc_lock", "bigdecimal", "either", "hashbrown", "indexmap", "num-bigint", "num-complex", "num-rational",
+    "ordered-float", "rust_decimal", "smallvec",
    # "anyhow", "chrono", "chrono-local", "chrono-tz", "eyre", "jiff-02", "lock_api", "parking-lot", "time",  "serde",
 ] }
 pyo3-stub-gen = { version = "0.17.2" }
 pyo3-async-runtimes = { version = "0.27.0", features = ["attributes", "tokio-runtime", "testing"] }
 pyo3-log = "0.13.2"

+# macro dependencies
+extend = { workspace = true }
+delegate = { workspace = true }
+impl-trait-for-tuples = { workspace = true }
+derive_more = { workspace = true }
+pin-project = { workspace = true }
+
 # async runtime
 tokio = { workspace = true, features = ["full", "tracing"] }
+futures = { workspace = true }
+
+# utility dependencies
+once_cell = "1.21.3"
+thread_local = "1.1.9"
+util = { workspace = true }
+thiserror = { workspace = true }
+#internment = { workspace = true }
+#recursion = { workspace = true }
+#generativity = { workspace = true }
+#itertools = { workspace = true }
+

 # Tracing
 #tracing = "0.1"
@@ -56,4 +75,3 @@ env_logger = "0.11"

 # Networking
 libp2p = { workspace = true, features = ["full"] }
-futures-lite = "2.6.1"
--- a/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi
+++ b/rust/exo_pyo3_bindings/exo_pyo3_bindings.pyi
@@ -2,39 +2,220 @@
 # ruff: noqa: E501, F401

 import builtins
+import enum
 import typing

@typing.final
-class Keypair:
-    @staticmethod
-    def generate() -> Keypair:
+class AllQueuesFullError(builtins.Exception):
+    def __new__(cls, *args: typing.Any) -> AllQueuesFullError: ...
+    def __repr__(self) -> builtins.str: ...
+    def __str__(self) -> builtins.str: ...
+
+@typing.final
+class ConnectionUpdate:
+    @property
+    def update_type(self) -> ConnectionUpdateType:
        r"""
-        Generate a new ed25519 keypair
+        Whether this is a connection or disconnection event
+        """
+    @property
+    def peer_id(self) -> PeerId:
+        r"""
+        Identity of the peer that we have connected to or disconnected from.
+        """
+    @property
+    def remote_ipv4(self) -> builtins.str:
+        r"""
+        Remote connection's IPv4 address.
+        """
+    @property
+    def remote_tcp_port(self) -> builtins.int:
+        r"""
+        Remote connection's TCP port.
+        """
+
+@typing.final
+class Keypair:
+    r"""
+    Identity keypair of a node.
+    """
+    @staticmethod
+    def generate_ed25519() -> Keypair:
+        r"""
+        Generate a new Ed25519 keypair.
+        """
+    @staticmethod
+    def generate_ecdsa() -> Keypair:
+        r"""
+        Generate a new ECDSA keypair.
+        """
+    @staticmethod
+    def generate_secp256k1() -> Keypair:
+        r"""
+        Generate a new Secp256k1 keypair.
        """
    @staticmethod
    def from_protobuf_encoding(bytes: bytes) -> Keypair:
        r"""
        Decode a private key from a protobuf structure and parse it as a `Keypair`.
        """
+    @staticmethod
+    def rsa_from_pkcs8(bytes: bytes) -> Keypair:
+        r"""
+        Decode an keypair from a DER-encoded secret key in PKCS#8 `PrivateKeyInfo`
+        format (i.e. unencrypted) as defined in [RFC5208].
+        
+        [RFC5208]: https://tools.ietf.org/html/rfc5208#section-5
+        """
+    @staticmethod
+    def secp256k1_from_der(bytes: bytes) -> Keypair:
+        r"""
+        Decode a keypair from a DER-encoded Secp256k1 secret key in an `ECPrivateKey`
+        structure as defined in [RFC5915].
+        
+        [RFC5915]: https://tools.ietf.org/html/rfc5915
+        """
+    @staticmethod
+    def ed25519_from_bytes(bytes: bytes) -> Keypair: ...
    def to_protobuf_encoding(self) -> bytes:
        r"""
-        Encode a private key to a protobuf structure.
+        Encode a private key as protobuf structure.
+        """
+    def to_peer_id(self) -> PeerId:
+        r"""
+        Convert the `Keypair` into the corresponding `PeerId`.
        """
-    def to_string(self) -> builtins.str: ...

@typing.final
-class PyPeer:
+class Multiaddr:
+    r"""
+    Representation of a Multiaddr.
+    """
    @staticmethod
-    def new(kp: Keypair, namespace: builtins.str) -> PyPeer: ...
-    async def subscribe(self, topic: builtins.str) -> None: ...
-    async def unsubscribe(self, topic: builtins.str) -> None: ...
-    async def send(self, topic: builtins.str, payload: bytes) -> None: ...
-    async def run(self) -> None: ...
-    async def recv(self) -> PySwarmEvent: ...
+    def empty() -> Multiaddr:
+        r"""
+        Create a new, empty multiaddress.
+        """
+    @staticmethod
+    def with_capacity(n: builtins.int) -> Multiaddr:
+        r"""
+        Create a new, empty multiaddress with the given capacity.
+        """
+    @staticmethod
+    def from_bytes(bytes: bytes) -> Multiaddr:
+        r"""
+        Parse a `Multiaddr` value from its byte slice representation.
+        """
+    @staticmethod
+    def from_string(string: builtins.str) -> Multiaddr:
+        r"""
+        Parse a `Multiaddr` value from its string representation.
+        """
+    def len(self) -> builtins.int:
+        r"""
+        Return the length in bytes of this multiaddress.
+        """
+    def is_empty(self) -> builtins.bool:
+        r"""
+        Returns true if the length of this multiaddress is 0.
+        """
+    def to_bytes(self) -> bytes:
+        r"""
+        Return a copy of this [`Multiaddr`]'s byte representation.
+        """
+    def to_string(self) -> builtins.str:
+        r"""
+        Convert a Multiaddr to a string.
+        """

@typing.final
-class PySwarmEvent:
-    def downcast_discovered(self) -> typing.Optional[builtins.str]: ...
-    def downcast_expired(self) -> typing.Optional[builtins.str]: ...
-    def downcast_message(self) -> typing.Optional[tuple[builtins.str, builtins.str, bytes]]: ...
+class NetworkingHandle:
+    def __new__(cls, identity: Keypair) -> NetworkingHandle: ...
+    async def connection_update_recv(self) -> ConnectionUpdate:
+        r"""
+        Receives the next `ConnectionUpdate` from networking.
+        """
+    async def connection_update_recv_many(self, limit: builtins.int) -> builtins.list[ConnectionUpdate]:
+        r"""
+        Receives at most `limit` `ConnectionUpdate`s from networking and returns them.
+        
+        For `limit = 0`, an empty collection of `ConnectionUpdate`s will be returned immediately.
+        For `limit > 0`, if there are no `ConnectionUpdate`s in the channel's queue this method
+        will sleep until a `ConnectionUpdate`s is sent.
+        """
+    async def gossipsub_subscribe(self, topic: builtins.str) -> builtins.bool:
+        r"""
+        Subscribe to a `GossipSub` topic.
+        
+        Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
+        """
+    async def gossipsub_unsubscribe(self, topic: builtins.str) -> builtins.bool:
+        r"""
+        Unsubscribes from a `GossipSub` topic.
+        
+        Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
+        """
+    async def gossipsub_publish(self, topic: builtins.str, data: bytes) -> None:
+        r"""
+        Publishes a message with multiple topics to the `GossipSub` network.
+        
+        If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
+        """
+    async def gossipsub_recv(self) -> tuple[builtins.str, bytes]:
+        r"""
+        Receives the next message from the `GossipSub` network.
+        """
+    async def gossipsub_recv_many(self, limit: builtins.int) -> builtins.list[tuple[builtins.str, bytes]]:
+        r"""
+        Receives at most `limit` messages from the `GossipSub` network and returns them.
+        
+        For `limit = 0`, an empty collection of messages will be returned immediately.
+        For `limit > 0`, if there are no messages in the channel's queue this method
+        will sleep until a message is sent.
+        """
+
+@typing.final
+class NoPeersSubscribedToTopicError(builtins.Exception):
+    def __new__(cls, *args: typing.Any) -> NoPeersSubscribedToTopicError: ...
+    def __repr__(self) -> builtins.str: ...
+    def __str__(self) -> builtins.str: ...
+
+@typing.final
+class PeerId:
+    r"""
+    Identifier of a peer of the network.
+    
+    The data is a `CIDv0` compatible multihash of the protobuf encoded public key of the peer
+    as specified in [specs/peer-ids](https://github.com/libp2p/specs/blob/master/peer-ids/peer-ids.md).
+    """
+    @staticmethod
+    def random() -> PeerId:
+        r"""
+        Generates a random peer ID from a cryptographically secure PRNG.
+        
+        This is useful for randomly walking on a DHT, or for testing purposes.
+        """
+    @staticmethod
+    def from_bytes(bytes: bytes) -> PeerId:
+        r"""
+        Parses a `PeerId` from bytes.
+        """
+    def to_bytes(self) -> bytes:
+        r"""
+        Returns a raw bytes representation of this `PeerId`.
+        """
+    def to_base58(self) -> builtins.str:
+        r"""
+        Returns a base-58 encoded string of this `PeerId`.
+        """
+    def __repr__(self) -> builtins.str: ...
+    def __str__(self) -> builtins.str: ...
+
+@typing.final
+class ConnectionUpdateType(enum.Enum):
+    r"""
+    Connection or disconnection event discriminant type.
+    """
+    Connected = ...
+    Disconnected = ...

--- a/rust/exo_pyo3_bindings/src/allow_threading.rs
+++ b/rust/exo_pyo3_bindings/src/allow_threading.rs
@@ -1,4 +1,8 @@
-//! See: <https://pyo3.rs/v0.27.2/async-await.html#detaching-from-the-interpreter-across-await>
+//! SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
+//!
+
+use pin_project::pin_project;
+use pyo3::marker::Ungil;
 use pyo3::prelude::*;
 use std::{
    future::Future,
@@ -6,17 +10,31 @@ use std::{
    task::{Context, Poll},
 };

-pub struct AllowThreads<F>(pub(crate) F);
+/// SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
+#[pin_project]
+#[repr(transparent)]
+pub(crate) struct AllowThreads<F>(#[pin] F);
+
+impl<F> AllowThreads<F>
+where
+    Self: Future,
+{
+    pub fn new(f: F) -> Self {
+        Self(f)
+    }
+}

 impl<F> Future for AllowThreads<F>
 where
-    F: Future + Unpin + Send,
-    F::Output: Send,
+    F: Future + Ungil,
+    F::Output: Ungil,
 {
    type Output = F::Output;

-    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        let waker = cx.waker();
-        Python::attach(|py| py.detach(|| pin!(&mut self.0).poll(&mut Context::from_waker(waker))))
+        Python::with_gil(|py| {
+            py.allow_threads(|| self.project().0.poll(&mut Context::from_waker(waker)))
+        })
    }
 }
--- a/rust/exo_pyo3_bindings/src/examples/mod.rs
+++ b/rust/exo_pyo3_bindings/src/examples/mod.rs
@@ -0,0 +1,240 @@
+//! This module exists to hold examples of some pyo3 patterns that may be too complex to
+//! re-create from scratch, but too inhomogenous to create an abstraction/wrapper around.
+//!
+//! Pattern examples include:
+//!  - Async task handles: with GC-integrated cleanup
+//!  - Sync/async callbacks from python: with propper eventloop handling
+//!
+//! Mutability pattern: https://pyo3.rs/v0.26.0/async-await.html#send--static-constraint
+//!  - Store mutable fields in tokio's `Mutex<T>`
+//!  - For async code: take `&self` and `.lock().await`
+//!  - For sync code: take `&mut self` and `.get_mut()`
+
+use crate::ext::{PyResultExt as _, ResultExt as _, TokioRuntimeExt as _};
+use futures::FutureExt as _;
+use futures::future::BoxFuture;
+use pyo3::exceptions::PyRuntimeError;
+use pyo3::prelude::{PyModule, PyModuleMethods as _};
+use pyo3::{
+    Bound, Py, PyAny, PyErr, PyResult, PyTraverseError, PyVisit, Python, pyclass, pymethods,
+};
+use std::time::Duration;
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::TryRecvError;
+
+fn needs_tokio_runtime() {
+    tokio::runtime::Handle::current();
+}
+
+type SyncCallback = Box<dyn Fn() + Send + Sync>;
+type AsyncCallback = Box<dyn Fn() -> BoxFuture<'static, ()> + Send + Sync>;
+
+enum AsyncTaskMessage {
+    SyncCallback(SyncCallback),
+    AsyncCallback(AsyncCallback),
+}
+
+async fn async_task(
+    sender: mpsc::UnboundedSender<()>,
+    mut receiver: mpsc::UnboundedReceiver<AsyncTaskMessage>,
+) {
+    log::info!("RUST: async task started");
+
+    // task state
+    let mut interval = tokio::time::interval(Duration::from_secs(1));
+
+    let mut sync_cbs: Vec<SyncCallback> = vec![];
+    let mut async_cbs: Vec<AsyncCallback> = vec![];
+
+    loop {
+        tokio::select! {
+            // handle incoming messages from task-handle
+            message = receiver.recv() => {
+                // handle closed channel by exiting
+                let Some(message) = message else {
+                    log::info!("RUST: channel closed");
+                    break;
+                };
+
+                // dispatch incoming event
+                match message {
+                    AsyncTaskMessage::SyncCallback(cb) => {
+                        sync_cbs.push(cb);
+                    }
+                    AsyncTaskMessage::AsyncCallback(cb) => {
+                        async_cbs.push(cb);
+                    }
+                }
+            }
+
+            // handle all other events
+            _ = interval.tick() => {
+                log::info!("RUST: async task tick");
+
+                // call back all sync callbacks
+                for cb in &sync_cbs {
+                    cb();
+                }
+
+                // call back all async callbacks
+                for cb in &async_cbs {
+                    cb().await;
+                }
+
+                // send event on unbounded channel
+                sender.send(()).expect("handle receiver cannot be closed/dropped");
+            }
+        }
+    }
+
+    log::info!("RUST: async task stopped");
+}
+
+// #[gen_stub_pyclass]
+#[pyclass(name = "AsyncTaskHandle")]
+#[derive(Debug)]
+struct PyAsyncTaskHandle {
+    sender: Option<mpsc::UnboundedSender<AsyncTaskMessage>>,
+    receiver: mpsc::UnboundedReceiver<()>,
+}
+
+#[allow(clippy::expect_used)]
+impl PyAsyncTaskHandle {
+    const fn sender(&self) -> &mpsc::UnboundedSender<AsyncTaskMessage> {
+        self.sender
+            .as_ref()
+            .expect("The sender should only be None after de-initialization.")
+    }
+
+    const fn sender_mut(&mut self) -> &mpsc::UnboundedSender<AsyncTaskMessage> {
+        self.sender
+            .as_mut()
+            .expect("The sender should only be None after de-initialization.")
+    }
+
+    const fn new(
+        sender: mpsc::UnboundedSender<AsyncTaskMessage>,
+        receiver: mpsc::UnboundedReceiver<()>,
+    ) -> Self {
+        Self {
+            sender: Some(sender),
+            receiver,
+        }
+    }
+}
+
+// #[gen_stub_pymethods]
+#[pymethods]
+impl PyAsyncTaskHandle {
+    #[new]
+    fn py_new(py: Python<'_>) -> PyResult<Self> {
+        use pyo3_async_runtimes::tokio::get_runtime;
+
+        // create communication channel TOWARDS our task
+        let (h_sender, t_receiver) = mpsc::unbounded_channel::<AsyncTaskMessage>();
+
+        // create communication channel FROM our task
+        let (t_sender, h_receiver) = mpsc::unbounded_channel::<()>();
+
+        // perform necessary setup within tokio context - or it crashes
+        let () = get_runtime().block_on(async { needs_tokio_runtime() });
+
+        // spawn tokio task with this thread's task-locals - without this, async callbacks on the new threads will not work!!
+        _ = get_runtime().spawn_with_scope(py, async move {
+            async_task(t_sender, t_receiver).await;
+        });
+        Ok(Self::new(h_sender, h_receiver))
+    }
+
+    /// NOTE: exceptions in callbacks are silently ignored until end of execution
+    fn add_sync_callback(
+        &self,
+        // #[gen_stub(override_type(
+        //     type_repr="collections.abc.Callable[[], None]",
+        //     imports=("collections.abc")
+        // ))]
+        callback: Py<PyAny>,
+    ) -> PyResult<()> {
+        // blocking call to async method -> can do non-blocking if needed
+        self.sender()
+            .send(AsyncTaskMessage::SyncCallback(Box::new(move || {
+                _ = Python::with_gil(|py| callback.call0(py).write_unraisable_with(py));
+            })))
+            .pyerr()?;
+        Ok(())
+    }
+
+    /// NOTE: exceptions in callbacks are silently ignored until end of execution
+    fn add_async_callback(
+        &self,
+        // #[gen_stub(override_type(
+        //     type_repr="collections.abc.Callable[[], collections.abc.Awaitable[None]]",
+        //     imports=("collections.abc")
+        // ))]
+        callback: Py<PyAny>,
+    ) -> PyResult<()> {
+        // blocking call to async method -> can do non-blocking if needed
+        self.sender()
+            .send(AsyncTaskMessage::AsyncCallback(Box::new(move || {
+                let c = Python::with_gil(|py| callback.clone_ref(py));
+                async move {
+                    if let Some(f) = Python::with_gil(|py| {
+                        let coroutine = c.call0(py).write_unraisable_with(py)?;
+                        pyo3_async_runtimes::tokio::into_future(coroutine.into_bound(py))
+                            .write_unraisable_with(py)
+                    }) {
+                        _ = f.await.write_unraisable();
+                    }
+                }
+                .boxed()
+            })))
+            .pyerr()?;
+        Ok(())
+    }
+
+    async fn receive_unit(&mut self) -> PyResult<()> {
+        self.receiver
+            .recv()
+            .await
+            .ok_or(PyErr::new::<PyRuntimeError, _>(
+                "cannot receive unit on closed channel",
+            ))
+    }
+
+    fn drain_units(&mut self) -> PyResult<i32> {
+        let mut cnt = 0;
+        loop {
+            match self.receiver.try_recv() {
+                Err(TryRecvError::Disconnected) => {
+                    return Err(PyErr::new::<PyRuntimeError, _>(
+                        "cannot receive unit on closed channel",
+                    ));
+                }
+                Err(TryRecvError::Empty) => return Ok(cnt),
+                Ok(()) => {
+                    cnt += 1;
+                    continue;
+                }
+            }
+        }
+    }
+
+    // #[gen_stub(skip)]
+    const fn __traverse__(&self, _visit: PyVisit<'_>) -> Result<(), PyTraverseError> {
+        Ok(()) // This is needed purely so `__clear__` can work
+    }
+
+    // #[gen_stub(skip)]
+    fn __clear__(&mut self) {
+        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
+        //       to ensure that the networking task is done BEFORE exiting the clear function...
+        //       but this may require GIL?? and it may not be safe to call GIL here??
+        self.sender = None; // Using Option<T> as a trick to force `sender` channel to be dropped
+    }
+}
+
+pub fn examples_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyAsyncTaskHandle>()?;
+
+    Ok(())
+}
--- a/rust/exo_pyo3_bindings/src/lib.rs
+++ b/rust/exo_pyo3_bindings/src/lib.rs
@@ -1,42 +1,216 @@
 //! TODO: crate documentation
-pub(crate) mod allow_threading;
+//!
+//! this is here as a placeholder documentation
+//!
+//!

+// enable Rust-unstable features for convenience
+#![feature(trait_alias)]
+#![feature(tuple_trait)]
+#![feature(unboxed_closures)]
+// #![feature(stmt_expr_attributes)]
+// #![feature(assert_matches)]
+// #![feature(async_fn_in_dyn_trait)]
+// #![feature(async_for_loop)]
+// #![feature(auto_traits)]
+// #![feature(negative_impls)]
+
+extern crate core;
+mod allow_threading;
+mod examples;
 pub(crate) mod networking;
-pub(crate) mod take_once {
-    use tokio::sync::Mutex;
+pub(crate) mod pylibp2p;

-    pub struct TakeOnce<T>(Mutex<Option<T>>);
-    impl<T> TakeOnce<T> {
-        pub fn new(t: T) -> Self {
-            Self(Mutex::new(Some(t)))
+use crate::networking::networking_submodule;
+use crate::pylibp2p::ident::ident_submodule;
+use crate::pylibp2p::multiaddr::multiaddr_submodule;
+use pyo3::prelude::PyModule;
+use pyo3::prelude::*;
+use pyo3::{Bound, PyResult, pyclass, pymodule};
+use pyo3_stub_gen::define_stub_info_gatherer;
+
+/// Namespace for all the constants used by this crate.
+pub(crate) mod r#const {
+    pub const MPSC_CHANNEL_SIZE: usize = 1024;
+}
+
+/// Namespace for all the type/trait aliases used by this crate.
+pub(crate) mod alias {
+    use std::error::Error;
+    use std::marker::Tuple;
+
+    pub trait SendFn<Args: Tuple + Send + 'static, Output> =
+        Fn<Args, Output = Output> + Send + 'static;
+
+    pub type AnyError = Box<dyn Error + Send + Sync + 'static>;
+    pub type AnyResult<T> = Result<T, AnyError>;
+}
+
+/// Namespace for crate-wide extension traits/methods
+pub(crate) mod ext {
+    use crate::allow_threading::AllowThreads;
+    use extend::ext;
+    use pyo3::exceptions::{PyConnectionError, PyRuntimeError};
+    use pyo3::marker::Ungil;
+    use pyo3::types::PyBytes;
+    use pyo3::{Py, PyErr, PyResult, Python};
+    use tokio::runtime::Runtime;
+    use tokio::sync::mpsc;
+    use tokio::sync::mpsc::error::TryRecvError;
+    use tokio::task::JoinHandle;
+
+    #[ext(pub, name = ByteArrayExt)]
+    impl [u8] {
+        fn pybytes(&self) -> Py<PyBytes> {
+            Python::with_gil(|py| PyBytes::new(py, self).unbind())
        }
-        pub fn take(&self) -> Option<T> {
-            match self.0.try_lock() {
-                Ok(mut o) => o.take(),
-                Err(_) => None,
+    }
+
+    #[ext(pub, name = ResultExt)]
+    impl<T, E> Result<T, E>
+    where
+        E: ToString,
+    {
+        fn pyerr(self) -> PyResult<T> {
+            self.map_err(|e| PyRuntimeError::new_err(e.to_string()))
+        }
+    }
+
+    pub trait FutureExt: Future + Sized {
+        /// SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
+        fn allow_threads_py(self) -> AllowThreads<Self>
+        where
+            AllowThreads<Self>: Future,
+        {
+            AllowThreads::new(self)
+        }
+    }
+
+    impl<T: Future> FutureExt for T {}
+
+    #[ext(pub, name = PyErrExt)]
+    impl PyErr {
+        fn receiver_channel_closed() -> Self {
+            PyConnectionError::new_err("Receiver channel closed unexpectedly")
+        }
+    }
+
+    #[ext(pub, name = PyResultExt)]
+    impl<T> PyResult<T> {
+        fn write_unraisable(self) -> Option<T> {
+            Python::with_gil(|py| self.write_unraisable_with(py))
+        }
+
+        fn write_unraisable_with(self, py: Python<'_>) -> Option<T> {
+            match self {
+                Ok(v) => Some(v),
+                Err(e) => {
+                    // write error back to python
+                    e.write_unraisable(py, None);
+                    None
+                }
+            }
+        }
+    }
+
+    #[ext(pub, name = TokioRuntimeExt)]
+    impl Runtime {
+        fn spawn_with_scope<F>(&self, py: Python<'_>, future: F) -> PyResult<JoinHandle<F::Output>>
+        where
+            F: Future + Send + 'static,
+            F::Output: Send + 'static,
+        {
+            let locals = pyo3_async_runtimes::tokio::get_current_locals(py)?;
+            Ok(self.spawn(pyo3_async_runtimes::tokio::scope(locals, future)))
+        }
+    }
+
+    #[ext(pub, name = TokioMpscSenderExt)]
+    impl<T> mpsc::Sender<T> {
+        /// Sends a value, waiting until there is capacity.
+        ///
+        /// A successful send occurs when it is determined that the other end of the
+        /// channel has not hung up already. An unsuccessful send would be one where
+        /// the corresponding receiver has already been closed.
+        async fn send_py(&self, value: T) -> PyResult<()> {
+            self.send(value)
+                .await
+                .map_err(|_| PyErr::receiver_channel_closed())
+        }
+    }
+
+    #[ext(pub, name = TokioMpscReceiverExt)]
+    impl<T> mpsc::Receiver<T> {
+        /// Receives the next value for this receiver.
+        async fn recv_py(&mut self) -> PyResult<T> {
+            self.recv().await.ok_or_else(PyErr::receiver_channel_closed)
+        }
+
+        /// Receives at most `limit` values for this receiver and returns them.
+        ///
+        /// For `limit = 0`, an empty collection of messages will be returned immediately.
+        /// For `limit > 0`, if there are no messages in the channel's queue this method
+        /// will sleep until a message is sent.
+        async fn recv_many_py(&mut self, limit: usize) -> PyResult<Vec<T>> {
+            // get updates from receiver channel
+            let mut updates = Vec::with_capacity(limit);
+            let received = self.recv_many(&mut updates, limit).await;
+
+            // if we received zero items, then the channel was unexpectedly closed
+            if limit != 0 && received == 0 {
+                return Err(PyErr::receiver_channel_closed());
+            }
+
+            Ok(updates)
+        }
+
+        /// Tries to receive the next value for this receiver.
+        fn try_recv_py(&mut self) -> PyResult<Option<T>> {
+            match self.try_recv() {
+                Ok(v) => Ok(Some(v)),
+                Err(TryRecvError::Empty) => Ok(None),
+                Err(TryRecvError::Disconnected) => Err(PyErr::receiver_channel_closed()),
            }
        }
    }
 }

-use pyo3::prelude::*;
+pub(crate) mod private {
+    use std::marker::Sized;

-use pyo3_stub_gen::define_stub_info_gatherer;
+    /// Sealed traits support
+    pub trait Sealed {}
+    impl<T: ?Sized> Sealed for T {}
+}
+
+/// A wrapper around [`Py`] that implements [`Clone`] using [`Python::with_gil`].
+#[repr(transparent)]
+pub(crate) struct ClonePy<T>(pub Py<T>);
+
+impl<T> Clone for ClonePy<T> {
+    fn clone(&self) -> Self {
+        Python::with_gil(|py| Self(self.0.clone_ref(py)))
+    }
+}

 /// A Python module implemented in Rust. The name of this function must match
 /// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
 /// import the module.
 #[pymodule(name = "exo_pyo3_bindings")]
-pub fn networking_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
+fn main_module(m: &Bound<'_, PyModule>) -> PyResult<()> {
    // install logger
    pyo3_log::init();
-    // setup runtime
-    let mut builder = tokio::runtime::Builder::new_multi_thread();
-    builder.enable_all();
-    pyo3_async_runtimes::tokio::init(builder);

-    m.add_class::<networking::PyPeer>()?;
-    m.add_class::<networking::PyKeypair>()?;
+    // 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...
+    ident_submodule(m)?;
+    multiaddr_submodule(m)?;
+    networking_submodule(m)?;
+
+    // top-level constructs
+    // TODO: ...
+
    Ok(())
 }

--- a/rust/exo_pyo3_bindings/src/networking.rs
+++ b/rust/exo_pyo3_bindings/src/networking.rs
@@ -1,214 +1,571 @@
-use crate::allow_threading::AllowThreads;
-use crate::take_once::TakeOnce;
+#![allow(
+    clippy::multiple_inherent_impl,
+    clippy::unnecessary_wraps,
+    clippy::unused_self,
+    clippy::needless_pass_by_value
+)]

-use std::pin::pin;
+use crate::r#const::MPSC_CHANNEL_SIZE;
+use crate::ext::{ByteArrayExt as _, FutureExt, PyErrExt as _};
+use crate::ext::{ResultExt as _, TokioMpscReceiverExt as _, TokioMpscSenderExt as _};
+use crate::pyclass;
+use crate::pylibp2p::ident::{PyKeypair, PyPeerId};
+use libp2p::futures::StreamExt as _;
+use libp2p::gossipsub::{IdentTopic, Message, MessageId, PublishError};
+use libp2p::swarm::SwarmEvent;
+use libp2p::{gossipsub, mdns};
+use networking::discovery;
+use networking::swarm::create_swarm;
+use pyo3::prelude::{PyModule, PyModuleMethods as _};
+use pyo3::types::PyBytes;
+use pyo3::{Bound, Py, PyErr, PyResult, PyTraverseError, PyVisit, Python, pymethods};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pyclass_enum, gen_stub_pymethods};
+use std::net::IpAddr;
+use tokio::sync::{Mutex, mpsc, oneshot};
+use util::ext::VecExt as _;

-use futures_lite::FutureExt;
-use libp2p::{gossipsub::PublishError, identity::Keypair};
-use networking::{FromSwarm, Peer, ToSwarm};
-use pyo3::{
-    coroutine::CancelHandle,
-    exceptions::{PyConnectionError, PyRuntimeError, PyValueError},
-    prelude::*,
-    types::PyBytes,
-};
-use pyo3_stub_gen::{
-    derive::{gen_methods_from_python, gen_stub_pyclass, gen_stub_pymethods},
-    inventory::submit,
-};
-use tokio::sync::{Mutex, mpsc};
+mod exception {
+    use pyo3::types::PyTuple;
+    use pyo3::{PyErrArguments, exceptions::PyException, prelude::*};
+    use pyo3_stub_gen::derive::*;

-#[gen_stub_pyclass]
-#[pyclass(name = "Keypair", frozen)]
-#[derive(Clone)]
-pub struct PyKeypair(Keypair);
+    #[gen_stub_pyclass]
+    #[pyclass(frozen, extends=PyException, name="NoPeersSubscribedToTopicError")]
+    pub struct PyNoPeersSubscribedToTopicError {}

-#[gen_stub_pymethods]
-#[pymethods]
-impl PyKeypair {
-    /// Generate a new ed25519 keypair
-    #[staticmethod]
-    fn generate() -> Self {
-        Self(Keypair::generate_ed25519())
-    }
+    impl PyNoPeersSubscribedToTopicError {
+        const MSG: &'static str = "\
+        No peers are currently subscribed to receive messages on this topic. \
+        Wait for peers to subscribe or check your network connectivity.";

-    /// Decode a private key from a protobuf structure and parse it as a `Keypair`.
-    #[staticmethod]
-    fn from_protobuf_encoding(bytes: &Bound<'_, PyBytes>) -> Self {
-        let bytes = Vec::from(bytes.as_bytes());
-        Self(Keypair::from_protobuf_encoding(&bytes).expect("todo"))
-    }
-
-    /// Encode a private key to a protobuf structure.
-    fn to_protobuf_encoding<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
-        match self.0.to_protobuf_encoding() {
-            Ok(bytes) => Ok(PyBytes::new(py, &bytes)),
-            Err(e) => Err(PyValueError::new_err(e.to_string())),
+        ///   Creates a new  [ `PyErr` ]  of this type.
+        ///
+        ///   [`PyErr`] :  https://docs.rs/pyo3/latest/pyo3/struct.PyErr.html   "PyErr in pyo3"
+        pub(crate) fn new_err() -> PyErr {
+            PyErr::new::<Self, _>(()) // TODO: check if this needs to be replaced???
        }
    }

-    fn to_string(&self) -> String {
-        self.0.public().to_peer_id().to_base58()
+    #[gen_stub_pymethods]
+    #[pymethods]
+    impl PyNoPeersSubscribedToTopicError {
+        #[new]
+        #[pyo3(signature = (*args))]
+        #[allow(unused_variables)]
+        pub(crate) fn new(args: &Bound<'_, PyTuple>) -> Self {
+            Self {}
+        }
+
+        fn __repr__(&self) -> String {
+            format!("PeerId(\"{}\")", Self::MSG)
+        }
+
+        fn __str__(&self) -> String {
+            Self::MSG.to_string()
+        }
+    }
+
+    #[gen_stub_pyclass]
+    #[pyclass(frozen, extends=PyException, name="AllQueuesFullError")]
+    pub struct PyAllQueuesFullError {}
+
+    impl PyAllQueuesFullError {
+        const MSG: &'static str =
+            "All libp2p peers are unresponsive, resend the message or reconnect.";
+
+        ///   Creates a new  [ `PyErr` ]  of this type.
+        ///
+        ///   [`PyErr`] :  https://docs.rs/pyo3/latest/pyo3/struct.PyErr.html   "PyErr in pyo3"
+        pub(crate) fn new_err() -> PyErr {
+            PyErr::new::<Self, _>(()) // TODO: check if this needs to be replaced???
+        }
+    }
+
+    #[gen_stub_pymethods]
+    #[pymethods]
+    impl PyAllQueuesFullError {
+        #[new]
+        #[pyo3(signature = (*args))]
+        #[allow(unused_variables)]
+        pub(crate) fn new(args: &Bound<'_, PyTuple>) -> Self {
+            Self {}
+        }
+
+        fn __repr__(&self) -> String {
+            format!("PeerId(\"{}\")", Self::MSG)
+        }
+
+        fn __str__(&self) -> String {
+            Self::MSG.to_string()
+        }
    }
 }

-struct PeerBuilder(
-    String,
-    Keypair,
-    mpsc::Sender<FromSwarm>,
-    mpsc::Receiver<ToSwarm>,
-);
+/// Connection or disconnection event discriminant type.
+#[gen_stub_pyclass_enum]
+#[pyclass(eq, eq_int, name = "ConnectionUpdateType")]
+#[derive(Debug, Clone, PartialEq)]
+enum PyConnectionUpdateType {
+    Connected = 0,
+    Disconnected,
+}

 #[gen_stub_pyclass]
-#[pyclass]
-pub struct PyPeer {
-    peer: TakeOnce<PeerBuilder>,
-    to_swarm: mpsc::Sender<ToSwarm>,
-    from_swarm: Mutex<mpsc::Receiver<FromSwarm>>,
+#[pyclass(frozen, name = "ConnectionUpdate")]
+#[derive(Debug, Clone)]
+struct PyConnectionUpdate {
+    /// Whether this is a connection or disconnection event
+    #[pyo3(get)]
+    update_type: PyConnectionUpdateType,
+
+    /// Identity of the peer that we have connected to or disconnected from.
+    #[pyo3(get)]
+    peer_id: PyPeerId,
+
+    /// Remote connection's IPv4 address.
+    #[pyo3(get)]
+    remote_ipv4: String,
+
+    /// Remote connection's TCP port.
+    #[pyo3(get)]
+    remote_tcp_port: u16,
+}
+
+enum ToTask {
+    GossipsubSubscribe {
+        topic: String,
+        result_tx: oneshot::Sender<PyResult<bool>>,
+    },
+    GossipsubUnsubscribe {
+        topic: String,
+        result_tx: oneshot::Sender<bool>,
+    },
+    GossipsubPublish {
+        topic: String,
+        data: Vec<u8>,
+        result_tx: oneshot::Sender<PyResult<MessageId>>,
+    },
+}
+
+#[allow(clippy::enum_glob_use)]
+async fn networking_task(
+    mut swarm: networking::swarm::Swarm,
+    mut to_task_rx: mpsc::Receiver<ToTask>,
+    connection_update_tx: mpsc::Sender<PyConnectionUpdate>,
+    gossipsub_message_tx: mpsc::Sender<(String, Vec<u8>)>,
+) {
+    use SwarmEvent::*;
+    use ToTask::*;
+    use mdns::Event::*;
+    use networking::swarm::BehaviourEvent::*;
+
+    log::info!("RUST: networking task started");
+
+    loop {
+        tokio::select! {
+            message = to_task_rx.recv() => {
+                // handle closed channel
+                let Some(message) = message else {
+                    log::info!("RUST: channel closed");
+                    break;
+                };
+
+                // dispatch incoming messages
+                match message {
+                    GossipsubSubscribe { topic, result_tx } => {
+                        // try to subscribe
+                        let result = swarm.behaviour_mut()
+                            .gossipsub.subscribe(&IdentTopic::new(topic));
+
+                        // send response oneshot
+                        if let Err(e) = result_tx.send(result.pyerr()) {
+                            log::error!("RUST: could not subscribe to gossipsub topic since channel already closed: {e:?}");
+                            continue;
+                        }
+                    }
+                    GossipsubUnsubscribe { topic, result_tx } => {
+                        // try to unsubscribe from the topic
+                        let result = swarm.behaviour_mut()
+                            .gossipsub.unsubscribe(&IdentTopic::new(topic));
+
+                        // send response oneshot (or exit if connection closed)
+                        if let Err(e) = result_tx.send(result) {
+                            log::error!("RUST: could not unsubscribe from gossipsub topic since channel already closed: {e:?}");
+                            continue;
+                        }
+                    }
+                    GossipsubPublish { topic, data, result_tx } => {
+                        // try to publish the data -> catch NoPeersSubscribedToTopic error & convert to correct exception
+                        let result = swarm.behaviour_mut().gossipsub.publish(
+                            IdentTopic::new(topic), data);
+                        let pyresult: PyResult<MessageId> = if let Err(PublishError::NoPeersSubscribedToTopic) = result {
+                            Err(exception::PyNoPeersSubscribedToTopicError::new_err())
+                        } else if let Err(PublishError::AllQueuesFull(_)) = result {
+                            Err(exception::PyAllQueuesFullError::new_err())
+                        } else {
+                            result.pyerr()
+                        };
+
+                        // send response oneshot (or exit if connection closed)
+                        if let Err(e) = result_tx.send(pyresult) {
+                            log::error!("RUST: could not publish gossipsub message since channel already closed: {e:?}");
+                            continue;
+                        }
+                    }
+                }
+            }
+
+            // architectural solution to this problem:
+            // create keep_alive behavior who's job it is to dial peers discovered by mDNS (and drop when expired)
+            //   -> it will emmit TRUE connected/disconnected events consumable elsewhere
+            //
+            // gossipsub will feed off-of dial attempts created by networking, and that will bootstrap its' peers list
+            // then for actual communication it will dial those peers if need-be
+            swarm_event = swarm.select_next_some() => {
+                match swarm_event {
+                    Behaviour(Gossipsub(gossipsub::Event::Message {
+                        message: Message {
+                            topic,
+                            data,
+                            ..
+                        },
+                        ..
+                    })) => {
+                        // topic-ID is just the topic hash!!! (since we used identity hasher)
+                        let message = (topic.into_string(), data);
+
+                        // send incoming message to channel (or exit if connection closed)
+                        if let Err(e) = gossipsub_message_tx.send(message).await {
+                            log::error!("RUST: could not send incoming gossipsub message since channel already closed: {e}");
+                            continue;
+                        }
+                    },
+                    Behaviour(Discovery(discovery::Event::ConnectionEstablished { peer_id, remote_ip, remote_tcp_port, .. })) => {
+                        // grab IPv4 string
+                        let remote_ipv4 = match remote_ip {
+                            IpAddr::V4(ip) => ip.to_string(),
+                            IpAddr::V6(ip) => {
+                                log::warn!("RUST: ignoring connection to IPv6 address: {ip}");
+                                continue;
+                            }
+                        };
+
+                        // send connection event to channel (or exit if connection closed)
+                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
+                            update_type: PyConnectionUpdateType::Connected,
+                            peer_id: PyPeerId(peer_id),
+                            remote_ipv4,
+                            remote_tcp_port,
+                        }).await {
+                            log::error!("RUST: could not send connection update since channel already closed: {e}");
+                            continue;
+                        }
+                    },
+                    Behaviour(Discovery(discovery::Event::ConnectionClosed { peer_id, remote_ip, remote_tcp_port, .. })) => {
+                        // grab IPv4 string
+                        let remote_ipv4 = match remote_ip {
+                            IpAddr::V4(ip) => ip.to_string(),
+                            IpAddr::V6(ip) => {
+                                log::warn!("RUST: ignoring disconnection from IPv6 address: {ip}");
+                                continue;
+                            }
+                        };
+
+                        // send disconnection event to channel (or exit if connection closed)
+                        if let Err(e) = connection_update_tx.send(PyConnectionUpdate {
+                            update_type: PyConnectionUpdateType::Disconnected,
+                            peer_id: PyPeerId(peer_id),
+                            remote_ipv4,
+                            remote_tcp_port,
+                        }).await {
+                            log::error!("RUST: could not send connection update since channel already closed: {e}");
+                            continue;
+                        }
+                    },
+                    e => {
+                        log::info!("RUST: other event {e:?}");
+                    }
+                }
+            }
+        }
+    }
+
+    log::info!("RUST: networking task stopped");
+}
+
+#[gen_stub_pyclass]
+#[pyclass(name = "NetworkingHandle")]
+#[derive(Debug)]
+struct PyNetworkingHandle {
+    // channels
+    to_task_tx: Option<mpsc::Sender<ToTask>>,
+    connection_update_rx: Mutex<mpsc::Receiver<PyConnectionUpdate>>,
+    gossipsub_message_rx: Mutex<mpsc::Receiver<(String, Vec<u8>)>>,
+}
+
+impl Drop for PyNetworkingHandle {
+    fn drop(&mut self) {
+        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
+        //       to ensure that the networking task is done BEFORE exiting the clear function...
+        //       but this may require GIL?? and it may not be safe to call GIL here??
+        self.to_task_tx = None; // Using Option<T> as a trick to force channel to be dropped
+    }
+}
+
+#[allow(clippy::expect_used)]
+impl PyNetworkingHandle {
+    fn new(
+        to_task_tx: mpsc::Sender<ToTask>,
+        connection_update_rx: mpsc::Receiver<PyConnectionUpdate>,
+        gossipsub_message_rx: mpsc::Receiver<(String, Vec<u8>)>,
+    ) -> Self {
+        Self {
+            to_task_tx: Some(to_task_tx),
+            connection_update_rx: Mutex::new(connection_update_rx),
+            gossipsub_message_rx: Mutex::new(gossipsub_message_rx),
+        }
+    }
+
+    const fn to_task_tx(&self) -> &mpsc::Sender<ToTask> {
+        self.to_task_tx
+            .as_ref()
+            .expect("The sender should only be None after de-initialization.")
+    }
 }

 #[gen_stub_pymethods]
 #[pymethods]
-impl PyPeer {
-    #[staticmethod]
-    fn new(kp: PyKeypair, namespace: String) -> PyResult<Self> {
-        let (to_client, from_swarm) = mpsc::channel(1024);
-        let (to_swarm, from_client) = mpsc::channel(1024);
-        Ok(Self {
-            peer: TakeOnce::new(PeerBuilder(namespace, kp.0, to_client, from_client)),
-            to_swarm,
-            from_swarm: Mutex::new(from_swarm),
-        })
+impl PyNetworkingHandle {
+    // NOTE: `async fn`s here that use `.await` will wrap the future in `.allow_threads_py()`
+    //       immediately beforehand to release the interpreter.
+    //       SEE: https://pyo3.rs/v0.26.0/async-await.html#detaching-from-the-interpreter-across-await
+
+    // ---- Lifecycle management methods ----
+
+    #[new]
+    fn py_new(identity: Bound<'_, PyKeypair>) -> PyResult<Self> {
+        use pyo3_async_runtimes::tokio::get_runtime;
+
+        // create communication channels
+        let (to_task_tx, to_task_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
+        let (connection_update_tx, connection_update_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
+        let (gossipsub_message_tx, gossipsub_message_rx) = mpsc::channel(MPSC_CHANNEL_SIZE);
+
+        // get identity
+        let identity = identity.borrow().0.clone();
+
+        // create networking swarm (within tokio context!! or it crashes)
+        let swarm = get_runtime()
+            .block_on(async { create_swarm(identity) })
+            .pyerr()?;
+
+        // spawn tokio task running the networking logic
+        get_runtime().spawn(async move {
+            networking_task(
+                swarm,
+                to_task_rx,
+                connection_update_tx,
+                gossipsub_message_tx,
+            )
+            .await;
+        });
+        Ok(Self::new(
+            to_task_tx,
+            connection_update_rx,
+            gossipsub_message_rx,
+        ))
    }

    #[gen_stub(skip)]
-    async fn run(&self, #[pyo3(cancel_handle)] mut cancel: CancelHandle) -> PyResult<()> {
-        let builder = self
-            .peer
-            .take()
-            .ok_or_else(|| PyRuntimeError::new_err("tried to run peer twice"))?;
-        let jh = pyo3_async_runtimes::tokio::get_runtime()
-            .spawn(async move {
-                let mut peer =
-                    Peer::new(builder.0, builder.1, builder.2, builder.3).map_err(|_| {
-                        PyConnectionError::new_err("peer failed to listen on default address")
-                    })?;
-                peer.run()
-                    .await
-                    .map_err(|()| PyConnectionError::new_err("peer communication closed"))
+    const fn __traverse__(&self, _visit: PyVisit<'_>) -> Result<(), PyTraverseError> {
+        Ok(()) // This is needed purely so `__clear__` can work
+    }
+
+    #[gen_stub(skip)]
+    fn __clear__(&mut self) {
+        // TODO: may or may not need to await a "kill-signal" oneshot channel message,
+        //       to ensure that the networking task is done BEFORE exiting the clear function...
+        //       but this may require GIL?? and it may not be safe to call GIL here??
+        self.to_task_tx = None; // Using Option<T> as a trick to force channel to be dropped
+    }
+
+    // ---- Connection update receiver methods ----
+
+    /// Receives the next `ConnectionUpdate` from networking.
+    async fn connection_update_recv(&self) -> PyResult<PyConnectionUpdate> {
+        self.connection_update_rx
+            .lock()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .recv_py()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+    }
+
+    /// Receives at most `limit` `ConnectionUpdate`s from networking and returns them.
+    ///
+    /// For `limit = 0`, an empty collection of `ConnectionUpdate`s will be returned immediately.
+    /// For `limit > 0`, if there are no `ConnectionUpdate`s in the channel's queue this method
+    /// will sleep until a `ConnectionUpdate`s is sent.
+    async fn connection_update_recv_many(&self, limit: usize) -> PyResult<Vec<PyConnectionUpdate>> {
+        self.connection_update_rx
+            .lock()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .recv_many_py(limit)
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+    }
+
+    // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)
+    //       so its too dangerous to expose just yet. figure out a better semantics for handling this,
+    //       so things don't randomly block
+    // /// Tries to receive the next `ConnectionUpdate` from networking.
+    // fn connection_update_try_recv(&self) -> PyResult<Option<PyConnectionUpdate>> {
+    //     self.connection_update_rx.blocking_lock().try_recv_py()
+    // }
+    //
+    // /// Checks if the `ConnectionUpdate` channel is empty.
+    // fn connection_update_is_empty(&self) -> bool {
+    //     self.connection_update_rx.blocking_lock().is_empty()
+    // }
+    //
+    // /// Returns the number of `ConnectionUpdate`s in the channel.
+    // fn connection_update_len(&self) -> usize {
+    //     self.connection_update_rx.blocking_lock().len()
+    // }
+
+    // ---- Gossipsub management methods ----
+
+    /// Subscribe to a `GossipSub` topic.
+    ///
+    /// Returns `True` if the subscription worked. Returns `False` if we were already subscribed.
+    async fn gossipsub_subscribe(&self, topic: String) -> PyResult<bool> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to subscribe
+        self.to_task_tx()
+            .send_py(ToTask::GossipsubSubscribe {
+                topic,
+                result_tx: tx,
            })
-            .or(async {
-                cancel.cancelled().await;
-                Ok(Ok(()))
-            });
-        match AllowThreads(pin!(jh)).await {
-            Err(e) if e.is_cancelled() => Ok(()),
-            Err(e) if e.is_panic() => Err(PyRuntimeError::new_err(format!("tokio panic {e}"))),
-            Err(_) => unreachable!(),
-            Ok(res) => res,
-        }
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & return any errors
+        rx.allow_threads_py() // allow-threads-aware async call
+            .await
+            .map_err(|_| PyErr::receiver_channel_closed())?
    }

-    async fn subscribe(&self, topic: String) -> PyResult<()> {
-        self.to_swarm
-            .send(ToSwarm::Subscribe(topic))
+    /// Unsubscribes from a `GossipSub` topic.
+    ///
+    /// Returns `True` if we were subscribed to this topic. Returns `False` if we were not subscribed.
+    async fn gossipsub_unsubscribe(&self, topic: String) -> PyResult<bool> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to unsubscribe
+        self.to_task_tx()
+            .send_py(ToTask::GossipsubUnsubscribe {
+                topic,
+                result_tx: tx,
+            })
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & convert any errors
+        rx.allow_threads_py() // allow-threads-aware async call
            .await
-            .map_err(|_| PyRuntimeError::new_err("swarm communication closed"))
-    }
-    async fn unsubscribe(&self, topic: String) -> PyResult<()> {
-        self.to_swarm
-            .send(ToSwarm::Unsubscribe(topic))
-            .await
-            .map_err(|_| PyRuntimeError::new_err("swarm communication closed"))
-    }
-    async fn send(&self, topic: String, payload: Py<PyBytes>) -> PyResult<()> {
-        // this function attaches to the python interpreter synchronously to avoid holding the GIL
-        let bytes = Python::attach(|py| Vec::from(payload.bind(py).as_bytes()));
-        self.to_swarm
-            .send(ToSwarm::Message(topic, bytes))
-            .await
-            .map_err(|_| PyRuntimeError::new_err("swarm communication closed"))
+            .map_err(|_| PyErr::receiver_channel_closed())
    }

-    #[gen_stub(skip)]
-    async fn recv(
-        &self,
-        #[pyo3(cancel_handle)] mut cancel: CancelHandle,
-    ) -> PyResult<PySwarmEvent> {
-        loop {
-            return match AllowThreads(pin!(
-                self.from_swarm
-                    .try_lock()
-                    .map_err(|_| PyRuntimeError::new_err("tried to recv twice"))?
-                    .recv()
-                    .or(async {
-                        cancel.cancelled().await;
-                        None
-                    })
-            ))
+    /// Publishes a message with multiple topics to the `GossipSub` network.
+    ///
+    /// If no peers are found that subscribe to this topic, throws `NoPeersSubscribedToTopicError` exception.
+    async fn gossipsub_publish(&self, topic: String, data: Py<PyBytes>) -> PyResult<()> {
+        let (tx, rx) = oneshot::channel();
+
+        // send off request to subscribe
+        let data = Python::with_gil(|py| Vec::from(data.as_bytes(py)));
+        self.to_task_tx()
+            .send_py(ToTask::GossipsubPublish {
+                topic,
+                data,
+                result_tx: tx,
+            })
+            .allow_threads_py() // allow-threads-aware async call
+            .await?;
+
+        // wait for response & return any errors => ignore messageID for now!!!
+        let _ = rx
+            .allow_threads_py() // allow-threads-aware async call
            .await
-            {
-                Some(FromSwarm::PublishError(p)) => match p {
-                    PublishError::AllQueuesFull(_) => {
-                        Err(PyConnectionError::new_err("swarm overloaded"))
-                    }
-                    PublishError::MessageTooLarge => {
-                        Err(PyValueError::new_err("message too large"))
-                    }
-                    PublishError::NoPeersSubscribedToTopic => {
-                        continue;
-                    }
-                    // TODO(evan): logs here
-                    _ => continue,
-                },
-                None => Err(PyRuntimeError::new_err("swarm communication closed")),
-                Some(fs) => Ok(PySwarmEvent(fs)),
-            };
-        }
+            .map_err(|_| PyErr::receiver_channel_closed())??;
+        Ok(())
    }
+
+    // ---- Gossipsub message receiver methods ----
+
+    /// Receives the next message from the `GossipSub` network.
+    async fn gossipsub_recv(&self) -> PyResult<(String, Py<PyBytes>)> {
+        self.gossipsub_message_rx
+            .lock()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .recv_py()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .map(|(t, d)| (t, d.pybytes()))
+    }
+
+    /// Receives at most `limit` messages from the `GossipSub` network and returns them.
+    ///
+    /// For `limit = 0`, an empty collection of messages will be returned immediately.
+    /// For `limit > 0`, if there are no messages in the channel's queue this method
+    /// will sleep until a message is sent.
+    async fn gossipsub_recv_many(&self, limit: usize) -> PyResult<Vec<(String, Py<PyBytes>)>> {
+        Ok(self
+            .gossipsub_message_rx
+            .lock()
+            .allow_threads_py() // allow-threads-aware async call
+            .await
+            .recv_many_py(limit)
+            .allow_threads_py() // allow-threads-aware async call
+            .await?
+            .map(|(t, d)| (t, d.pybytes())))
+    }
+
+    // TODO: rn this blocks main thread if anything else is awaiting the channel (bc its a mutex)
+    //       so its too dangerous to expose just yet. figure out a better semantics for handling this,
+    //       so things don't randomly block
+    // /// Tries to receive the next message from the `GossipSub` network.
+    // fn gossipsub_try_recv(&self) -> PyResult<Option<(String, Py<PyBytes>)>> {
+    //     Ok(self
+    //         .gossipsub_message_rx
+    //         .blocking_lock()
+    //         .try_recv_py()?
+    //         .map(|(t, d)| (t, d.pybytes())))
+    // }
+    //
+    // /// Checks if the `GossipSub` message channel is empty.
+    // fn gossipsub_is_empty(&self) -> bool {
+    //     self.gossipsub_message_rx.blocking_lock().is_empty()
+    // }
+    //
+    // /// Returns the number of `GossipSub` messages in the channel.
+    // fn gossipsub_len(&self) -> usize {
+    //     self.gossipsub_message_rx.blocking_lock().len()
+    // }
 }

-// Manually submit the run()/recv() stub because the cancelhandle is poorly understood
-submit! {
-    gen_methods_from_python! {
-        r#"
-        class PyPeer:
-            async def run(self): ...
-            async def recv(self) -> PySwarmEvent: ...
-        "#
-    }
-}
+pub fn networking_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<exception::PyNoPeersSubscribedToTopicError>()?;
+    m.add_class::<exception::PyAllQueuesFullError>()?;

-#[gen_stub_pyclass]
-#[pyclass]
-pub struct PySwarmEvent(FromSwarm);
+    m.add_class::<PyConnectionUpdateType>()?;
+    m.add_class::<PyConnectionUpdate>()?;
+    m.add_class::<PyConnectionUpdateType>()?;
+    m.add_class::<PyNetworkingHandle>()?;

-#[gen_stub_pymethods]
-#[pymethods]
-impl PySwarmEvent {
-    // probably a better way to do this, but...
-    fn downcast_discovered(&self) -> Option<String> {
-        if let FromSwarm::Discovered(peer_id) = self.0 {
-            Some(peer_id.to_base58())
-        } else {
-            None
-        }
-    }
-    fn downcast_expired(&self) -> Option<String> {
-        if let FromSwarm::Expired(peer_id) = self.0 {
-            Some(peer_id.to_base58())
-        } else {
-            None
-        }
-    }
-    fn downcast_message<'py>(
-        &self,
-        py: Python<'py>,
-    ) -> Option<(String, String, Bound<'py, PyBytes>)> {
-        if let FromSwarm::Message(peer_id, topic, data) = &self.0 {
-            Some((peer_id.to_base58(), topic.clone(), PyBytes::new(py, data)))
-        } else {
-            None
-        }
-    }
+    Ok(())
 }
--- a/rust/exo_pyo3_bindings/src/pylibp2p/ident.rs
+++ b/rust/exo_pyo3_bindings/src/pylibp2p/ident.rs
@@ -0,0 +1,159 @@
+use crate::ext::ResultExt as _;
+use libp2p::PeerId;
+use libp2p::identity::Keypair;
+use pyo3::prelude::{PyBytesMethods as _, PyModule, PyModuleMethods as _};
+use pyo3::types::PyBytes;
+use pyo3::{Bound, PyResult, Python, pyclass, pymethods};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+
+/// Identity keypair of a node.
+#[gen_stub_pyclass]
+#[pyclass(name = "Keypair", frozen)]
+#[repr(transparent)]
+pub struct PyKeypair(pub Keypair);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyKeypair {
+    /// Generate a new Ed25519 keypair.
+    #[staticmethod]
+    fn generate_ed25519() -> Self {
+        Self(Keypair::generate_ed25519())
+    }
+
+    /// Generate a new ECDSA keypair.
+    #[staticmethod]
+    fn generate_ecdsa() -> Self {
+        Self(Keypair::generate_ecdsa())
+    }
+
+    /// Generate a new Secp256k1 keypair.
+    #[staticmethod]
+    fn generate_secp256k1() -> Self {
+        Self(Keypair::generate_secp256k1())
+    }
+
+    /// Decode a private key from a protobuf structure and parse it as a `Keypair`.
+    #[staticmethod]
+    fn from_protobuf_encoding(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::from_protobuf_encoding(&bytes).pyerr()?))
+    }
+
+    /// Decode an keypair from a DER-encoded secret key in PKCS#8 `PrivateKeyInfo`
+    /// format (i.e. unencrypted) as defined in [RFC5208].
+    ///
+    /// [RFC5208]: https://tools.ietf.org/html/rfc5208#section-5
+    #[staticmethod]
+    fn rsa_from_pkcs8(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::rsa_from_pkcs8(&mut bytes).pyerr()?))
+    }
+
+    /// Decode a keypair from a DER-encoded Secp256k1 secret key in an `ECPrivateKey`
+    /// structure as defined in [RFC5915].
+    ///
+    /// [RFC5915]: https://tools.ietf.org/html/rfc5915
+    #[staticmethod]
+    fn secp256k1_from_der(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::secp256k1_from_der(&mut bytes).pyerr()?))
+    }
+
+    #[staticmethod]
+    fn ed25519_from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let mut bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Keypair::ed25519_from_bytes(&mut bytes).pyerr()?))
+    }
+
+    /// Encode a private key as protobuf structure.
+    fn to_protobuf_encoding<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
+        let bytes = self.0.to_protobuf_encoding().pyerr()?;
+        Ok(PyBytes::new(py, &bytes))
+    }
+
+    /// Convert the `Keypair` into the corresponding `PeerId`.
+    fn to_peer_id(&self) -> PyPeerId {
+        PyPeerId(self.0.public().to_peer_id())
+    }
+
+    // /// Hidden constructor for pickling support. TODO: figure out how to do pickling...
+    // #[gen_stub(skip)]
+    // #[new]
+    // fn py_new(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+    //     Self::from_protobuf_encoding(bytes)
+    // }
+    //
+    // #[gen_stub(skip)]
+    // fn __setstate__(&mut self, state: Bound<'_, PyBytes>) -> PyResult<()> {
+    //     *self = Self::from_protobuf_encoding(state)?;
+    //     Ok(())
+    // }
+    //
+    // #[gen_stub(skip)]
+    // fn __getstate__<'py>(&self, py: Python<'py>) -> PyResult<Bound<'py, PyBytes>> {
+    //     self.to_protobuf_encoding(py)
+    // }
+    //
+    // #[gen_stub(skip)]
+    // pub fn __getnewargs__<'py>(&self, py: Python<'py>) -> PyResult<(Bound<'py, PyBytes>,)> {
+    //     Ok((self.to_protobuf_encoding(py)?,))
+    // }
+}
+
+/// Identifier of a peer of the network.
+///
+/// The data is a `CIDv0` compatible multihash of the protobuf encoded public key of the peer
+/// as specified in [specs/peer-ids](https://github.com/libp2p/specs/blob/master/peer-ids/peer-ids.md).
+#[gen_stub_pyclass]
+#[pyclass(name = "PeerId", frozen)]
+#[derive(Debug, Clone)]
+#[repr(transparent)]
+pub struct PyPeerId(pub PeerId);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyPeerId {
+    /// Generates a random peer ID from a cryptographically secure PRNG.
+    ///
+    /// This is useful for randomly walking on a DHT, or for testing purposes.
+    #[staticmethod]
+    fn random() -> Self {
+        Self(PeerId::random())
+    }
+
+    /// Parses a `PeerId` from bytes.
+    #[staticmethod]
+    fn from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(PeerId::from_bytes(&bytes).pyerr()?))
+    }
+
+    /// Returns a raw bytes representation of this `PeerId`.
+    fn to_bytes<'py>(&self, py: Python<'py>) -> Bound<'py, PyBytes> {
+        let bytes = self.0.to_bytes();
+        PyBytes::new(py, &bytes)
+    }
+
+    /// Returns a base-58 encoded string of this `PeerId`.
+    fn to_base58(&self) -> String {
+        self.0.to_base58()
+    }
+
+    fn __repr__(&self) -> String {
+        format!("PeerId({})", self.to_base58())
+    }
+
+    fn __str__(&self) -> String {
+        self.to_base58()
+    }
+}
+
+pub fn ident_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyKeypair>()?;
+    m.add_class::<PyPeerId>()?;
+
+    Ok(())
+}
--- a/rust/exo_pyo3_bindings/src/pylibp2p/mod.rs
+++ b/rust/exo_pyo3_bindings/src/pylibp2p/mod.rs
@@ -0,0 +1,8 @@
+//! A module for exposing Rust's libp2p datatypes over Pyo3
+//!
+//! TODO: right now we are coupled to libp2p's identity, but eventually we want to create our own
+//!       independent identity type of some kind or another. This may require handshaking.
+//!
+
+pub mod ident;
+pub mod multiaddr;
--- a/rust/exo_pyo3_bindings/src/pylibp2p/multiaddr.rs
+++ b/rust/exo_pyo3_bindings/src/pylibp2p/multiaddr.rs
@@ -0,0 +1,81 @@
+use crate::ext::ResultExt as _;
+use libp2p::Multiaddr;
+use pyo3::prelude::{PyBytesMethods as _, PyModule, PyModuleMethods as _};
+use pyo3::types::PyBytes;
+use pyo3::{Bound, PyResult, Python, pyclass, pymethods};
+use pyo3_stub_gen::derive::{gen_stub_pyclass, gen_stub_pymethods};
+use std::str::FromStr as _;
+
+/// Representation of a Multiaddr.
+#[gen_stub_pyclass]
+#[pyclass(name = "Multiaddr", frozen)]
+#[derive(Debug, Clone)]
+#[repr(transparent)]
+pub struct PyMultiaddr(pub Multiaddr);
+
+#[gen_stub_pymethods]
+#[pymethods]
+#[allow(clippy::needless_pass_by_value)]
+impl PyMultiaddr {
+    /// Create a new, empty multiaddress.
+    #[staticmethod]
+    fn empty() -> Self {
+        Self(Multiaddr::empty())
+    }
+
+    /// Create a new, empty multiaddress with the given capacity.
+    #[staticmethod]
+    fn with_capacity(n: usize) -> Self {
+        Self(Multiaddr::with_capacity(n))
+    }
+
+    /// Parse a `Multiaddr` value from its byte slice representation.
+    #[staticmethod]
+    fn from_bytes(bytes: Bound<'_, PyBytes>) -> PyResult<Self> {
+        let bytes = Vec::from(bytes.as_bytes());
+        Ok(Self(Multiaddr::try_from(bytes).pyerr()?))
+    }
+
+    /// Parse a `Multiaddr` value from its string representation.
+    #[staticmethod]
+    fn from_string(string: String) -> PyResult<Self> {
+        Ok(Self(Multiaddr::from_str(&string).pyerr()?))
+    }
+
+    /// Return the length in bytes of this multiaddress.
+    fn len(&self) -> usize {
+        self.0.len()
+    }
+
+    /// Returns true if the length of this multiaddress is 0.
+    fn is_empty(&self) -> bool {
+        self.0.is_empty()
+    }
+
+    /// Return a copy of this [`Multiaddr`]'s byte representation.
+    fn to_bytes<'py>(&self, py: Python<'py>) -> Bound<'py, PyBytes> {
+        let bytes = self.0.to_vec();
+        PyBytes::new(py, &bytes)
+    }
+
+    /// Convert a Multiaddr to a string.
+    fn to_string(&self) -> String {
+        self.0.to_string()
+    }
+
+    #[gen_stub(skip)]
+    fn __repr__(&self) -> String {
+        format!("Multiaddr({})", self.0)
+    }
+
+    #[gen_stub(skip)]
+    fn __str__(&self) -> String {
+        self.to_string()
+    }
+}
+
+pub fn multiaddr_submodule(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyMultiaddr>()?;
+
+    Ok(())
+}
--- a/rust/exo_pyo3_bindings/tests/dummy.rs
+++ b/rust/exo_pyo3_bindings/tests/dummy.rs
@@ -0,0 +1,54 @@
+#[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;
+
+        let (tx, mut rx) = mpsc::channel::<Ping>(10);
+
+        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;
+                            }
+                        }
+                    }
+                    _ = tokio::time::sleep(Duration::from_secs_f32(0.1)) => {
+                        println!("TASK: heartbeat");
+                    }
+                }
+            }
+
+            println!("TASK: exited");
+        });
+
+        let tx2 = tx.clone();
+
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+
+        tx.send(Ping).await.expect("Should not fail");
+        drop(tx);
+
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+
+        tx2.send(Ping).await.expect("Should not fail");
+        drop(tx2);
+
+        tokio::time::sleep(Duration::from_secs_f32(0.11)).await;
+    }
+}
--- a/rust/networking/Cargo.toml
+++ b/rust/networking/Cargo.toml
@@ -13,14 +13,32 @@ path = "src/lib.rs"
 workspace = true

 [dependencies]
+# datastructures
+either = { workspace = true }
+
+# macro dependencies
+extend = { workspace = true }
+delegate = { workspace = true }
+impl-trait-for-tuples = { workspace = true }
+derive_more = { workspace = true }
+
 # async
 tokio = { workspace = true, features = ["full"] }
+futures = { workspace = true }
+futures-timer = { workspace = true }

 # utility dependencies
+util = { workspace = true }
+thiserror = { workspace = true }
+#internment = { workspace = true }
+#recursion = { workspace = true }
+#generativity = { workspace = true }
+#itertools = { workspace = true }
 tracing-subscriber = { version = "0.3.19", features = ["default", "env-filter"] }
+keccak-const = { workspace = true }

 # tracing/logging
 log = { workspace = true }

 # networking
-libp2p = { workspace = true, features = ["full"] }
+libp2p = { workspace = true, features = ["full"] }
--- a/rust/networking/examples/chatroom.rs
+++ b/rust/networking/examples/chatroom.rs
@@ -1,6 +1,6 @@
-use libp2p::identity;
-use networking::{self, FromSwarm, ToSwarm};
-use tokio::sync::mpsc;
+use futures::stream::StreamExt as _;
+use libp2p::{gossipsub, identity, swarm::SwarmEvent};
+use networking::{discovery, swarm};
 use tokio::{io, io::AsyncBufReadExt as _, select};
 use tracing_subscriber::EnvFilter;
 use tracing_subscriber::filter::LevelFilter;
@@ -12,51 +12,63 @@ async fn main() {
        .try_init();

    // Configure swarm
-    let (to_client, mut from_swarm) = mpsc::channel(20);
-    let (to_swarm, from_client) = mpsc::channel(20);
-    let mut peer = networking::Peer::new(
-        "chatroom!".to_string(),
-        identity::Keypair::generate_ed25519(),
-        to_client,
-        from_client,
-    )
-    .expect("listen error");
+    let mut swarm =
+        swarm::create_swarm(identity::Keypair::generate_ed25519()).expect("Swarm creation failed");

    // Create a Gossipsub topic & subscribe
+    let topic = gossipsub::IdentTopic::new("test-net");
+    swarm
+        .behaviour_mut()
+        .gossipsub
+        .subscribe(&topic)
+        .expect("Subscribing to topic failed");
+
    // Read full lines from stdin
    let mut stdin = io::BufReader::new(io::stdin()).lines();
    println!("Enter messages via STDIN and they will be sent to connected peers using Gossipsub");

-    let jh = tokio::spawn(async move { peer.run().await });
-    _ = to_swarm
-        .send(ToSwarm::Subscribe("chatting".to_string()))
-        .await;
-
    // Kick it off
    loop {
        select! {
            // on gossipsub outgoing
            Ok(Some(line)) = stdin.next_line() => {
-                _ = to_swarm.send(ToSwarm::Message("chatting".to_string(), line.into_bytes())).await;
+                if let Err(e) = swarm
+                    .behaviour_mut().gossipsub
+                    .publish(topic.clone(), line.as_bytes()) {
+                    println!("Publish error: {e:?}");
+                }
            }
-            event = from_swarm.recv() => match event {
+            event = swarm.select_next_some() => match event {
                // on gossipsub incoming
-                Some(FromSwarm::Message(peer_id,_, data)) => println!(
-                        "\n\nGot message: '{}' from peer: {peer_id}\n\n",
-                        String::from_utf8_lossy(&data),
+                SwarmEvent::Behaviour(swarm::BehaviourEvent::Gossipsub(gossipsub::Event::Message {
+                    propagation_source: peer_id,
+                    message_id: id,
+                    message,
+                })) => println!(
+                        "\n\nGot message: '{}' with id: {id} from peer: {peer_id}\n\n",
+                        String::from_utf8_lossy(&message.data),
                    ),

                // on discovery
-                Some(FromSwarm::Discovered(peer_id)) => {
-                    println!("\n\nConnected to: {peer_id}\n\n");
+                SwarmEvent::Behaviour(swarm::BehaviourEvent::Discovery(e)) => match e {
+                    discovery::Event::ConnectionEstablished {
+                        peer_id, connection_id, remote_ip, remote_tcp_port
+                    } => {
+                        println!("\n\nConnected to: {peer_id}; connection ID: {connection_id}; remote IP: {remote_ip}; remote TCP port: {remote_tcp_port}\n\n");
+                    }
+                    discovery::Event::ConnectionClosed {
+                        peer_id, connection_id, remote_ip, remote_tcp_port
+                    } => {
+                        eprintln!("\n\nDisconnected from: {peer_id}; connection ID: {connection_id}; remote IP: {remote_ip}; remote TCP port: {remote_tcp_port}\n\n");
+                    }
                }
-                Some(FromSwarm::Expired(peer_id)) => {
-                    println!("\n\nDisconnected from: {peer_id}\n\n");
-                }
-                Some(FromSwarm::PublishError(e)) => eprintln!("\n\nError {e:?}\n\n"),
-                None => break,
+
+                // ignore outgoing errors: those are normal
+                e@SwarmEvent::OutgoingConnectionError { .. } => { log::debug!("Outgoing connection error: {e:?}"); }
+
+                // otherwise log any other event
+                e => { log::info!("Other event {e:?}"); }
            }
        }
    }
-    _ = jh.await;
 }
--- a/rust/networking/examples/chatroom_manual.rs
+++ b/rust/networking/examples/chatroom_manual.rs
@@ -0,0 +1,127 @@
+// Copyright 2018 Parity Technologies (UK) Ltd.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal in the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in
+// all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+// OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+// DEALINGS IN THE SOFTWARE.
+
+use futures::stream::StreamExt;
+use libp2p::{
+    gossipsub, mdns, noise,
+    swarm::{NetworkBehaviour, SwarmEvent},
+    tcp, yamux,
+};
+use std::time::Duration;
+use std::{error::Error, hash::Hash};
+use tokio::{io, io::AsyncBufReadExt, select};
+use tracing_subscriber::EnvFilter;
+
+// We create a custom network behaviour that combines Gossipsub and Mdns.
+#[derive(NetworkBehaviour)]
+struct MyBehaviour {
+    gossipsub: gossipsub::Behaviour,
+    mdns: mdns::tokio::Behaviour,
+}
+
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn Error>> {
+    let _ = tracing_subscriber::fmt()
+        .with_env_filter(EnvFilter::from_default_env())
+        .try_init();
+
+    let mut swarm = libp2p::SwarmBuilder::with_new_identity()
+        .with_tokio()
+        .with_tcp(
+            tcp::Config::default(),
+            noise::Config::new,
+            yamux::Config::default,
+        )?
+        .with_behaviour(|key| {
+            // Set a custom gossipsub configuration
+            let gossipsub_config = gossipsub::ConfigBuilder::default()
+                .heartbeat_interval(Duration::from_secs(10))
+                .validation_mode(gossipsub::ValidationMode::Strict) // This sets the kind of message validation. The default is Strict (enforce message signing)
+                .build()
+                .map_err(io::Error::other)?; // Temporary hack because `build` does not return a proper `std::error::Error`.
+
+            // build a gossipsub network behaviour
+            let gossipsub = gossipsub::Behaviour::new(
+                gossipsub::MessageAuthenticity::Signed(key.clone()),
+                gossipsub_config,
+            )?;
+
+            let mdns =
+                mdns::tokio::Behaviour::new(mdns::Config::default(), key.public().to_peer_id())?;
+            Ok(MyBehaviour { gossipsub, mdns })
+        })?
+        .build();
+
+    println!("Running swarm with identity {}", swarm.local_peer_id());
+
+    // Create a Gossipsub topic
+    let topic = gossipsub::IdentTopic::new("test-net");
+    // subscribes to our topic
+    swarm.behaviour_mut().gossipsub.subscribe(&topic)?;
+
+    // Read full lines from stdin
+    let mut stdin = io::BufReader::new(io::stdin()).lines();
+
+    // Listen on all interfaces and whatever port the OS assigns
+    swarm.listen_on("/ip4/0.0.0.0/tcp/0".parse()?)?;
+
+    println!("Enter messages via STDIN and they will be sent to connected peers using Gossipsub");
+
+    // Kick it off
+    loop {
+        select! {
+            Ok(Some(line)) = stdin.next_line() => {
+                if let Err(e) = swarm
+                    .behaviour_mut().gossipsub
+                    .publish(topic.clone(), line.as_bytes()) {
+                    println!("Publish error: {e:?}");
+                }
+            }
+            event = swarm.select_next_some() => match event {
+                SwarmEvent::Behaviour(MyBehaviourEvent::Mdns(mdns::Event::Discovered(list))) => {
+                    for (peer_id, multiaddr) in list {
+                        println!("mDNS discovered a new peer: {peer_id} on {multiaddr}");
+                        swarm.behaviour_mut().gossipsub.add_explicit_peer(&peer_id);
+                    }
+                },
+                SwarmEvent::Behaviour(MyBehaviourEvent::Mdns(mdns::Event::Expired(list))) => {
+                    for (peer_id, multiaddr) in list {
+                        println!("mDNS discover peer has expired: {peer_id} on {multiaddr}");
+                        swarm.behaviour_mut().gossipsub.remove_explicit_peer(&peer_id);
+                    }
+                },
+                SwarmEvent::Behaviour(MyBehaviourEvent::Gossipsub(gossipsub::Event::Message {
+                    propagation_source: peer_id,
+                    message_id: id,
+                    message,
+                })) => println!(
+                        "Got message: '{}' with id: {id} from peer: {peer_id}",
+                        String::from_utf8_lossy(&message.data),
+                    ),
+                SwarmEvent::NewListenAddr { address, .. } => {
+                    println!("Local node is listening on {address}");
+                }
+                e => {
+                    println!("Other swarm event: {:?}", e);
+                }
+            }
+        }
+    }
+}
--- a/rust/networking/src/RESEARCH_NOTES.txt
+++ b/rust/networking/src/RESEARCH_NOTES.txt
@@ -0,0 +1,44 @@
+https://github.com/ml-explore/mlx/commit/3fe98bacc7640d857acf3539f1d21b47a32e5609
+^raw sockets distributed -> `<net/ndrv.h>` -> https://newosxbook.com/code/xnu-3247.1.106/bsd/net/ndrv.h.auto.html
+--> header file for a networking component found in the macOS kernel (XNU) that defines structures for network device driver registration, specifically the ndrv_demux_desc and ndrv_protocol_desc structures used for demultiplexing protocol data at the network interface level. It specifies how to describe protocol data, such as an Ethernet type or a SNAP header, and how to associate these descriptions with a specific protocol family to receive matching packets.
+--> Used to bind an NDRV socket so that packets that match given protocol demux descriptions can be received.
+--> An NDRV socket is a special kind of socket in the Darwin/macOS operating system's XNU kernel, used for low-level network packet manipulation and binding to specific protocols for packet processing. It allows user-space applications or drivers to directly write Layer 2 (L2) network packets or interact with the network stack at a lower level, often by binding to protocol descriptors like the ndrv_protocol_desc. This type of socket is used for functions such as capturing and injecting packets, especially in network infrastructure software like routers or for kernel-level network monitoring and security tools.
+--> also called PF_NDRV sockets --> https://newosxbook.com/bonus/vol1ch16.html
+----> they are conceptually similar to https://scapy.disruptivelabs.in/networking/socket-interface PF_RAW or PF_PACKET
+
+https://stackoverflow.com/questions/17169298/af-packet-on-osx
+^AF_PACKET duplicates the packets as soon as it receives them from the physical layer (for incoming packets) or just before sending them out to the physical layer (for outgoing packets). -> this is on Linux only
+^it doesn't exist on OS X so you can use /dev/bpfX (Berkeley Packet Filter) for sniffing
+
+https://www.unix.com/man_page/mojave/4/ip/
+^OS X manpages for IP
+
+https://developer.apple.com/documentation/kernel/implementing_drivers_system_extensions_and_kexts
+^driver kit, system extensions & kexts for macOS
+
+----
+
+To set up a Linux system to use a Thunderbolt connection as a network device, connect the two computers with a Thunderbolt cable, load the thunderbolt-net kernel module (usually automatic but modprobe is an option for manual loading), and then the operating system will create virtual Ethernet interfaces (e.g., thunderbolt0) for networking. You can then use standard tools like ifconfig or your desktop environment's network manager to configure these new interfaces for a link-local network.
+--> https://gist.github.com/geosp/80fbd39e617b7d1d9421683df4ea224a
+----> here is a guide on how to set up thunderbolt-ethernet on linux
+----> I may be able to steal the thunderbolt-net code ideas to implement a kernel module for MacOS
+
+https://chatgpt.com/s/t_68af8e41a8548191993281a014f846a7
+^GPT discussion about making socket interface
+
+https://chatgpt.com/s/t_68afb798a85c8191973c02a0fa7a48a3 --> link-local address,,??
+https://chatgpt.com/s/t_68afb02987e08191b2b0044d3667ece2
+^GPT discussion about accessing TB on MacOS low level interactions
+
+--------------------------------
+
+https://www.intel.com/content/www/us/en/support/articles/000098893/software.html
+^Thunderbolt Share & Thunderbolt Networking Mode => intel's equivalent of thunderbolt bridge
+
+
+---------------------------------
+
+https://www.zerotier.com/blog/how-zerotier-eliminated-kernel-extensions-on-macos/
+-->fake ethernet devices on MacOS -> omg??? we can detect thunderbolt bridge, then bind to it, then re-expose it as fake ethernet??
+-->ps: https://chatgpt.com/s/t_68afb2b25fb881919526763fb5d7359c, AF/PF_NDRV are one and the same!!!
+-->https://github.com/zerotier/ZeroTierOne/blob/dev/osdep/MacEthernetTapAgent.c
--- a/rust/networking/src/discovery.rs
+++ b/rust/networking/src/discovery.rs
@@ -0,0 +1,383 @@
+use crate::ext::MultiaddrExt;
+use crate::keep_alive;
+use delegate::delegate;
+use either::Either;
+use futures::FutureExt;
+use futures_timer::Delay;
+use libp2p::core::transport::PortUse;
+use libp2p::core::{ConnectedPoint, Endpoint};
+use libp2p::swarm::behaviour::ConnectionEstablished;
+use libp2p::swarm::dial_opts::DialOpts;
+use libp2p::swarm::{
+    CloseConnection, ConnectionClosed, ConnectionDenied, ConnectionHandler,
+    ConnectionHandlerSelect, ConnectionId, FromSwarm, NetworkBehaviour, THandler, THandlerInEvent,
+    THandlerOutEvent, ToSwarm, dummy,
+};
+use libp2p::{Multiaddr, PeerId, identity, mdns};
+use std::collections::{BTreeSet, HashMap};
+use std::convert::Infallible;
+use std::io;
+use std::net::IpAddr;
+use std::task::{Context, Poll};
+use std::time::Duration;
+use util::wakerdeque::WakerDeque;
+
+const RETRY_CONNECT_INTERVAL: Duration = Duration::from_secs(5);
+
+mod managed {
+    use libp2p::swarm::NetworkBehaviour;
+    use libp2p::{identity, mdns, ping};
+    use std::io;
+    use std::time::Duration;
+
+    const MDNS_RECORD_TTL: Duration = Duration::from_secs(2_500);
+    const MDNS_QUERY_INTERVAL: Duration = Duration::from_secs(1_500);
+    const PING_TIMEOUT: Duration = Duration::from_millis(2_500);
+    const PING_INTERVAL: Duration = Duration::from_millis(2_500);
+
+    #[derive(NetworkBehaviour)]
+    pub struct Behaviour {
+        mdns: mdns::tokio::Behaviour,
+        ping: ping::Behaviour,
+    }
+
+    impl Behaviour {
+        pub fn new(keypair: &identity::Keypair) -> io::Result<Self> {
+            Ok(Self {
+                mdns: mdns_behaviour(keypair)?,
+                ping: ping_behaviour(),
+            })
+        }
+    }
+
+    fn mdns_behaviour(keypair: &identity::Keypair) -> io::Result<mdns::tokio::Behaviour> {
+        use mdns::{Config, tokio};
+
+        // mDNS config => enable IPv6
+        let mdns_config = Config {
+            ttl: MDNS_RECORD_TTL,
+            query_interval: MDNS_QUERY_INTERVAL,
+
+            // enable_ipv6: true, // TODO: for some reason, TCP+mDNS don't work well with ipv6?? figure out how to make work
+            ..Default::default()
+        };
+
+        let mdns_behaviour = tokio::Behaviour::new(mdns_config, keypair.public().to_peer_id());
+        Ok(mdns_behaviour?)
+    }
+
+    fn ping_behaviour() -> ping::Behaviour {
+        ping::Behaviour::new(
+            ping::Config::new()
+                .with_timeout(PING_TIMEOUT)
+                .with_interval(PING_INTERVAL),
+        )
+    }
+}
+
+/// Events for when a listening connection is truly established and truly closed.
+#[derive(Debug, Clone)]
+pub enum Event {
+    ConnectionEstablished {
+        peer_id: PeerId,
+        connection_id: ConnectionId,
+        remote_ip: IpAddr,
+        remote_tcp_port: u16,
+    },
+    ConnectionClosed {
+        peer_id: PeerId,
+        connection_id: ConnectionId,
+        remote_ip: IpAddr,
+        remote_tcp_port: u16,
+    },
+}
+
+/// Discovery behavior that wraps mDNS to produce truly discovered durable peer-connections.
+///
+/// The behaviour operates as such:
+///  1) All true (listening) connections/disconnections are tracked, emitting corresponding events
+///     to the swarm.
+///  1) mDNS discovered/expired peers are tracked; discovered but not connected peers are dialed
+///     immediately, and expired but connected peers are disconnected from immediately.
+///  2) Every fixed interval: discovered but not connected peers are dialed, and expired but
+///     connected peers are disconnected from.
+pub struct Behaviour {
+    // state-tracking for managed behaviors & mDNS-discovered peers
+    managed: managed::Behaviour,
+    mdns_discovered: HashMap<PeerId, BTreeSet<Multiaddr>>,
+
+    retry_delay: Delay, // retry interval
+
+    // pending events to emmit => waker-backed Deque to control polling
+    pending_events: WakerDeque<ToSwarm<Event, Infallible>>,
+}
+
+impl Behaviour {
+    pub fn new(keypair: &identity::Keypair) -> io::Result<Self> {
+        Ok(Self {
+            managed: managed::Behaviour::new(keypair)?,
+            mdns_discovered: HashMap::new(),
+            retry_delay: Delay::new(RETRY_CONNECT_INTERVAL),
+            pending_events: WakerDeque::new(),
+        })
+    }
+
+    fn dial(&mut self, peer_id: PeerId, addr: Multiaddr) {
+        self.pending_events.push_back(ToSwarm::Dial {
+            opts: DialOpts::peer_id(peer_id).addresses(vec![addr]).build(),
+        })
+    }
+
+    fn close_connection(&mut self, peer_id: PeerId, connection: ConnectionId) {
+        // push front to make this IMMEDIATE
+        self.pending_events.push_front(ToSwarm::CloseConnection {
+            peer_id,
+            connection: CloseConnection::One(connection),
+        })
+    }
+
+    fn handle_mdns_discovered(&mut self, peers: Vec<(PeerId, Multiaddr)>) {
+        for (p, ma) in peers {
+            self.dial(p, ma.clone()); // always connect
+
+            // get peer's multi-addresses or insert if missing
+            let Some(mas) = self.mdns_discovered.get_mut(&p) else {
+                self.mdns_discovered.insert(p, BTreeSet::from([ma]));
+                continue;
+            };
+
+            // multiaddress should never already be present - else something has gone wrong
+            let is_new_addr = mas.insert(ma);
+            assert!(is_new_addr, "cannot discover a discovered peer");
+        }
+    }
+
+    fn handle_mdns_expired(&mut self, peers: Vec<(PeerId, Multiaddr)>) {
+        for (p, ma) in peers {
+            // at this point, we *must* have the peer
+            let mas = self
+                .mdns_discovered
+                .get_mut(&p)
+                .expect("nonexistent peer cannot expire");
+
+            // at this point, we *must* have the multiaddress
+            let was_present = mas.remove(&ma);
+            assert!(was_present, "nonexistent multiaddress cannot expire");
+
+            // if empty, remove the peer-id entirely
+            if mas.is_empty() {
+                self.mdns_discovered.remove(&p);
+            }
+        }
+    }
+
+    fn on_connection_established(
+        &mut self,
+        peer_id: PeerId,
+        connection_id: ConnectionId,
+        remote_ip: IpAddr,
+        remote_tcp_port: u16,
+    ) {
+        // send out connected event
+        self.pending_events
+            .push_back(ToSwarm::GenerateEvent(Event::ConnectionEstablished {
+                peer_id,
+                connection_id,
+                remote_ip,
+                remote_tcp_port,
+            }));
+    }
+
+    fn on_connection_closed(
+        &mut self,
+        peer_id: PeerId,
+        connection_id: ConnectionId,
+        remote_ip: IpAddr,
+        remote_tcp_port: u16,
+    ) {
+        // send out disconnected event
+        self.pending_events
+            .push_back(ToSwarm::GenerateEvent(Event::ConnectionClosed {
+                peer_id,
+                connection_id,
+                remote_ip,
+                remote_tcp_port,
+            }));
+    }
+}
+
+impl NetworkBehaviour for Behaviour {
+    type ConnectionHandler =
+        ConnectionHandlerSelect<dummy::ConnectionHandler, THandler<managed::Behaviour>>;
+    type ToSwarm = Event;
+
+    // simply delegate to underlying mDNS behaviour
+
+    delegate! {
+        to self.managed {
+            fn handle_pending_inbound_connection(&mut self, connection_id: ConnectionId, local_addr: &Multiaddr, remote_addr: &Multiaddr) -> Result<(), ConnectionDenied>;
+            fn handle_pending_outbound_connection(&mut self, connection_id: ConnectionId, maybe_peer: Option<PeerId>, addresses: &[Multiaddr], effective_role: Endpoint) -> Result<Vec<Multiaddr>, ConnectionDenied>;
+        }
+    }
+
+    fn handle_established_inbound_connection(
+        &mut self,
+        connection_id: ConnectionId,
+        peer: PeerId,
+        local_addr: &Multiaddr,
+        remote_addr: &Multiaddr,
+    ) -> Result<THandler<Self>, ConnectionDenied> {
+        Ok(ConnectionHandler::select(
+            dummy::ConnectionHandler,
+            self.managed.handle_established_inbound_connection(
+                connection_id,
+                peer,
+                local_addr,
+                remote_addr,
+            )?,
+        ))
+    }
+
+    #[allow(clippy::needless_question_mark)]
+    fn handle_established_outbound_connection(
+        &mut self,
+        connection_id: ConnectionId,
+        peer: PeerId,
+        addr: &Multiaddr,
+        role_override: Endpoint,
+        port_use: PortUse,
+    ) -> Result<THandler<Self>, ConnectionDenied> {
+        Ok(ConnectionHandler::select(
+            dummy::ConnectionHandler,
+            self.managed.handle_established_outbound_connection(
+                connection_id,
+                peer,
+                addr,
+                role_override,
+                port_use,
+            )?,
+        ))
+    }
+
+    fn on_connection_handler_event(
+        &mut self,
+        peer_id: PeerId,
+        connection_id: ConnectionId,
+        event: THandlerOutEvent<Self>,
+    ) {
+        match event {
+            Either::Left(ev) => libp2p::core::util::unreachable(ev),
+            Either::Right(ev) => {
+                self.managed
+                    .on_connection_handler_event(peer_id, connection_id, ev)
+            }
+        }
+    }
+
+    // hook into these methods to drive behavior
+
+    fn on_swarm_event(&mut self, event: FromSwarm) {
+        self.managed.on_swarm_event(event); // let mDNS handle swarm events
+
+        // handle swarm events to update internal state:
+        match event {
+            FromSwarm::ConnectionEstablished(ConnectionEstablished {
+                peer_id,
+                connection_id,
+                endpoint,
+                ..
+            }) => {
+                let remote_address = match endpoint {
+                    ConnectedPoint::Dialer { address, .. } => address,
+                    ConnectedPoint::Listener { send_back_addr, .. } => send_back_addr,
+                };
+
+                if let Some((ip, port)) = remote_address.try_to_tcp_addr() {
+                    // handle connection established event which is filtered correctly
+                    self.on_connection_established(peer_id, connection_id, ip, port)
+                }
+            }
+            FromSwarm::ConnectionClosed(ConnectionClosed {
+                peer_id,
+                connection_id,
+                endpoint,
+                ..
+            }) => {
+                let remote_address = match endpoint {
+                    ConnectedPoint::Dialer { address, .. } => address,
+                    ConnectedPoint::Listener { send_back_addr, .. } => send_back_addr,
+                };
+
+                if let Some((ip, port)) = remote_address.try_to_tcp_addr() {
+                    // handle connection closed event which is filtered correctly
+                    self.on_connection_closed(peer_id, connection_id, ip, port)
+                }
+            }
+
+            // since we are running TCP/IP transport layer, we are assuming that
+            // no address changes can occur, hence encountering one is a fatal error
+            FromSwarm::AddressChange(a) => {
+                unreachable!("unhandlable: address change encountered: {:?}", a)
+            }
+            _ => {}
+        }
+    }
+
+    fn poll(&mut self, cx: &mut Context) -> Poll<ToSwarm<Self::ToSwarm, THandlerInEvent<Self>>> {
+        // delegate to managed behaviors for any behaviors they need to perform
+        match self.managed.poll(cx) {
+            Poll::Ready(ToSwarm::GenerateEvent(e)) => {
+                match e {
+                    // handle discovered and expired events from mDNS
+                    managed::BehaviourEvent::Mdns(e) => match e.clone() {
+                        mdns::Event::Discovered(peers) => {
+                            self.handle_mdns_discovered(peers);
+                        }
+                        mdns::Event::Expired(peers) => {
+                            self.handle_mdns_expired(peers);
+                        }
+                    },
+
+                    // handle ping events => if error then disconnect
+                    managed::BehaviourEvent::Ping(e) => {
+                        if let Err(_) = e.result {
+                            self.close_connection(e.peer, e.connection.clone())
+                        }
+                    }
+                }
+
+                // since we just consumed an event, we should immediately wake just in case
+                // there are more events to come where that came from
+                cx.waker().wake_by_ref();
+            }
+
+            // forward any other mDNS event to the swarm or its connection handler(s)
+            Poll::Ready(e) => {
+                return Poll::Ready(
+                    e.map_out(|_| unreachable!("events returning to swarm already handled"))
+                        .map_in(Either::Right),
+                );
+            }
+
+            Poll::Pending => {}
+        }
+
+        // retry connecting to all mDNS peers periodically (fails safely if already connected)
+        if self.retry_delay.poll_unpin(cx).is_ready() {
+            for (p, mas) in self.mdns_discovered.clone() {
+                for ma in mas {
+                    self.dial(p, ma)
+                }
+            }
+            self.retry_delay.reset(RETRY_CONNECT_INTERVAL) // reset timeout
+        }
+
+        // send out any pending events from our own service
+        if let Some(e) = self.pending_events.pop_front(cx) {
+            return Poll::Ready(e.map_in(Either::Left));
+        }
+
+        // wait for pending events
+        Poll::Pending
+    }
+}
--- a/rust/networking/src/keep_alive.rs
+++ b/rust/networking/src/keep_alive.rs
@@ -0,0 +1,44 @@
+use delegate::delegate;
+use libp2p::swarm::handler::ConnectionEvent;
+use libp2p::swarm::{ConnectionHandlerEvent, SubstreamProtocol, dummy, handler};
+use std::task::{Context, Poll};
+
+/// An implementation of [`ConnectionHandler`] that doesn't handle any protocols, but it keeps
+/// the connection alive.
+#[derive(Clone)]
+#[repr(transparent)]
+pub struct ConnectionHandler(dummy::ConnectionHandler);
+
+impl ConnectionHandler {
+    pub fn new() -> Self {
+        ConnectionHandler(dummy::ConnectionHandler)
+    }
+}
+
+impl handler::ConnectionHandler for ConnectionHandler {
+    // delegate types and implementation mostly to dummy handler
+    type FromBehaviour = <dummy::ConnectionHandler as handler::ConnectionHandler>::FromBehaviour;
+    type ToBehaviour = <dummy::ConnectionHandler as handler::ConnectionHandler>::ToBehaviour;
+    type InboundProtocol =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::InboundProtocol;
+    type OutboundProtocol =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::OutboundProtocol;
+    type InboundOpenInfo =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::InboundOpenInfo;
+    type OutboundOpenInfo =
+        <dummy::ConnectionHandler as handler::ConnectionHandler>::OutboundOpenInfo;
+
+    delegate! {
+        to self.0 {
+            fn listen_protocol(&self) -> SubstreamProtocol<Self::InboundProtocol, Self::InboundOpenInfo>;
+            fn poll(&mut self, cx: &mut Context<'_>) -> Poll<ConnectionHandlerEvent<Self::OutboundProtocol, Self::OutboundOpenInfo, Self::ToBehaviour>>;
+            fn on_behaviour_event(&mut self, event: Self::FromBehaviour);
+            fn on_connection_event(&mut self, event: ConnectionEvent<Self::InboundProtocol, Self::OutboundProtocol, Self::InboundOpenInfo, Self::OutboundOpenInfo>);
+        }
+    }
+
+    // specifically override this to force connection to stay alive
+    fn connection_keep_alive(&self) -> bool {
+        true
+    }
+}
--- a/rust/networking/src/lib.rs
+++ b/rust/networking/src/lib.rs
@@ -1,299 +1,64 @@
-use libp2p::{
-    Multiaddr, PeerId,
-    futures::StreamExt,
-    gossipsub::{self, TopicHash},
-    identify,
-    identity::Keypair,
-    mdns,
-    swarm::{NetworkBehaviour, SwarmEvent, dial_opts::DialOpts},
-};
-use std::collections::HashMap;
-use tokio::sync::mpsc;
+//! TODO: crate documentation
+//!
+//! this is here as a placeholder documentation
+//!
+//!

-#[derive(Debug)]
-pub struct ListenError;
+// enable Rust-unstable features for convenience
+#![feature(trait_alias)]
+// #![feature(stmt_expr_attributes)]
+// #![feature(unboxed_closures)]
+// #![feature(assert_matches)]
+// #![feature(async_fn_in_dyn_trait)]
+// #![feature(async_for_loop)]
+// #![feature(auto_traits)]
+// #![feature(negative_impls)]

-pub enum FromSwarm {
-    PublishError(gossipsub::PublishError),
-    Discovered(PeerId),
-    Expired(PeerId),
-    Message(PeerId, String, Vec<u8>),
-}
-pub enum ToSwarm {
-    Message(String, Vec<u8>),
-    Subscribe(String),
-    Unsubscribe(String),
+pub mod discovery;
+pub mod keep_alive;
+pub mod swarm;
+
+/// Namespace for all the type/trait aliases used by this crate.
+pub(crate) mod alias {
+    use std::error::Error;
+
+    pub type AnyError = Box<dyn Error + Send + Sync + 'static>;
+    pub type AnyResult<T> = Result<T, AnyError>;
 }

-pub struct Peer {
-    pub swarm: libp2p::Swarm<Behaviour>,
-    to_client: mpsc::Sender<FromSwarm>,
-    from_client: mpsc::Receiver<ToSwarm>,
-    namespace: String,
-    known_peers: HashMap<PeerId, Vec<Multiaddr>>,
-}
-impl Peer {
-    pub fn new(
-        namespace: String,
-        kp: Keypair,
-        to_client: mpsc::Sender<FromSwarm>,
-        from_client: mpsc::Receiver<ToSwarm>,
-    ) -> Result<Self, ListenError> {
-        let mut swarm = libp2p::SwarmBuilder::with_existing_identity(kp)
-            .with_tokio()
-            .with_quic()
-            // TODO(evan) .with_bandwidth_metrics()
-            .with_behaviour(|kp| Behaviour::new(namespace.clone(), kp))
-            .expect("invalid swarm behaviour")
-            .build();
+/// Namespace for crate-wide extension traits/methods
+pub(crate) mod ext {
+    use extend::ext;
+    use libp2p::Multiaddr;
+    use libp2p::multiaddr::Protocol;
+    use std::net::IpAddr;

-        swarm
-            .listen_on("/ip6/::/udp/0/quic-v1".parse().expect("invalid multiaddr"))
-            .map_err(|_| ListenError)?;
-        swarm
-            .listen_on(
-                "/ip4/0.0.0.0/udp/0/quic-v1"
-                    .parse()
-                    .expect("invalid multiaddr"),
-            )
-            .map_err(|_| ListenError)?;
-        Ok(Self {
-            swarm,
-            to_client,
-            from_client,
-            namespace,
-            known_peers: HashMap::default(),
-        })
-    }
-    pub async fn run(&mut self) -> Result<(), ()> {
-        loop {
-            tokio::select! {
-                event = self.swarm.next() => self.handle_event(event.ok_or(())?).await?,
-                msg = self.from_client.recv() => self.handle_message(msg.ok_or(())?).await?,
-            }
-        }
-    }
-    async fn handle_message(&mut self, message: ToSwarm) -> Result<(), ()> {
-        match message {
-            ToSwarm::Message(topic, data) => {
-                if let Err(e) = self
-                    .swarm
-                    .behaviour_mut()
-                    .gossipsub
-                    .publish(TopicHash::from_raw(topic), data)
-                {
-                    self.to_client
-                        .send(FromSwarm::PublishError(e))
-                        .await
-                        .map_err(|_| ())?;
+    #[ext(pub, name = MultiaddrExt)]
+    impl Multiaddr {
+        /// If the multiaddress corresponds to a TCP address, extracts it
+        fn try_to_tcp_addr(&self) -> Option<(IpAddr, u16)> {
+            let mut ps = self.into_iter();
+            let ip = if let Some(p) = ps.next() {
+                match p {
+                    Protocol::Ip4(ip) => IpAddr::V4(ip),
+                    Protocol::Ip6(ip) => IpAddr::V6(ip),
+                    _ => return None,
                }
-            }
-            ToSwarm::Subscribe(topic) => {
-                match self
-                    .swarm
-                    .behaviour_mut()
-                    .gossipsub
-                    .subscribe(&gossipsub::IdentTopic::new(topic))
-                {
-                    Ok(_) => {}
-                    Err(gossipsub::SubscriptionError::NotAllowed) => {
-                        unreachable!("subscription filter hit")
-                    }
-                    Err(gossipsub::SubscriptionError::PublishError(e)) => self
-                        .to_client
-                        .send(FromSwarm::PublishError(e))
-                        .await
-                        .map_err(|_| ())?,
-                }
-            }
-            ToSwarm::Unsubscribe(topic) => {
-                self.swarm
-                    .behaviour_mut()
-                    .gossipsub
-                    .unsubscribe(&gossipsub::IdentTopic::new(topic));
-            }
-        }
-        Ok(())
-    }
-    async fn handle_event(&mut self, event: SwarmEvent<BehaviourEvent>) -> Result<(), ()> {
-        let SwarmEvent::Behaviour(event) = event else {
-            return Ok(());
-        };
-        match event {
-            BehaviourEvent::Gossipsub(gossipsub::Event::Message { message, .. }) => {
-                if let Some(source) = message.source {
-                    self.to_client
-                        .send(FromSwarm::Message(
-                            source,
-                            message.topic.into_string(),
-                            message.data,
-                        ))
-                        .await
-                        .map_err(|_| ())?;
-                }
-            }
-            BehaviourEvent::Identify(identify::Event::Received { peer_id, info, .. }) => {
-                log::debug!(
-                    "identify from {peer_id}: protocol_version='{}' agent_version='{}' (local namespace='{}')",
-                    info.protocol_version,
-                    info.agent_version,
-                    self.namespace
-                );
-                if info.protocol_version == self.namespace {
-                    self.passed_namespace(peer_id);
-                    self.to_client
-                        .send(FromSwarm::Discovered(peer_id))
-                        .await
-                        .map_err(|_| ())?;
-                } else {
-                    self.failed_namespace(peer_id);
-                }
-            }
-            BehaviourEvent::Mdns(mdns::Event::Discovered(v)) => {
-                for (peer_id, addr) in v {
-                    self.known_peers.entry(peer_id).or_default().push(addr);
-                }
-                for (peer_id, addrs) in &self.known_peers {
-                    // dialopts handles rate limiting, we should check errors if we want to blacklist earlier
-                    let _ = self
-                        .swarm
-                        .dial(DialOpts::peer_id(*peer_id).addresses(addrs.clone()).build());
-                }
-            }
-            BehaviourEvent::Mdns(mdns::Event::Expired(v)) => {
-                for (peer_id, addr) in v {
-                    let addrs = self.known_peers.entry(peer_id).or_default();
-                    addrs.retain(|a| *a != addr);
-                    if addrs.is_empty() {
-                        self.known_peers.remove(&peer_id);
-                        self.swarm
-                            .behaviour_mut()
-                            .gossipsub
-                            .remove_explicit_peer(&peer_id);
-                        self.to_client
-                            .send(FromSwarm::Expired(peer_id))
-                            .await
-                            .map_err(|_| ())?;
-                    }
-                }
-            }
-            _ => {}
-        }
-        Ok(())
-    }
-    fn passed_namespace(&mut self, peer_id: PeerId) {
-        self.swarm
-            .behaviour_mut()
-            .gossipsub
-            .remove_blacklisted_peer(&peer_id);
-        self.swarm
-            .behaviour_mut()
-            .gossipsub
-            .add_explicit_peer(&peer_id);
-    }
-    fn failed_namespace(&mut self, peer_id: PeerId) {
-        self.swarm
-            .behaviour_mut()
-            .gossipsub
-            .blacklist_peer(&peer_id);
-        self.swarm
-            .behaviour_mut()
-            .gossipsub
-            .remove_explicit_peer(&peer_id);
-    }
-}
-
-#[derive(NetworkBehaviour)]
-pub struct Behaviour {
-    gossipsub: gossipsub::Behaviour,
-    mdns: mdns::tokio::Behaviour,
-    identify: identify::Behaviour,
-}
-
-impl Behaviour {
-    fn new(namespace: String, kp: &Keypair) -> Self {
-        let mdns = mdns::Behaviour::new(mdns::Config::default(), kp.public().to_peer_id())
-            .expect("mdns behaviour failed to build");
-
-        let identify =
-            identify::Behaviour::new(identify::Config::new_with_signed_peer_record(namespace, kp));
-
-        let gossipsub = gossipsub::Behaviour::new(
-            gossipsub::MessageAuthenticity::Signed(kp.clone()),
-            gossipsub::ConfigBuilder::default()
-                .max_transmit_size(1024 * 1024)
-                .validation_mode(gossipsub::ValidationMode::Strict)
-                .build()
-                .expect("invalid gossipsub configuration"),
-        )
-        .expect("gossipsub behaviour failed ot build");
-
-        Self {
-            gossipsub,
-            mdns,
-            identify,
+            } else {
+                return None;
+            };
+            let Some(Protocol::Tcp(port)) = ps.next() else {
+                return None;
+            };
+            Some((ip, port))
        }
    }
 }

-// TODO: more tests
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use tokio::time::{Duration, timeout};
+pub(crate) mod private {
+    #![allow(dead_code)]

-    fn make_peer(namespace: &str) -> (Peer, mpsc::Receiver<FromSwarm>, mpsc::Sender<ToSwarm>) {
-        let kp = Keypair::generate_ed25519();
-
-        let (to_client_tx, to_client_rx) = mpsc::channel(64);
-        let (to_peer_tx, to_peer_rx) = mpsc::channel(64);
-
-        let peer = Peer::new(namespace.to_string(), kp, to_client_tx, to_peer_rx)
-            .expect("Peer::new should succeed in tests");
-
-        (peer, to_client_rx, to_peer_tx)
-    }
-
-    async fn next_listen_addr(peer: &mut Peer) -> Multiaddr {
-        loop {
-            match peer.swarm.next().await {
-                Some(SwarmEvent::NewListenAddr { address, .. }) => return address,
-                Some(_) => {}
-                None => panic!("swarm stream ended unexpectedly"),
-            }
-        }
-    }
-
-    #[tokio::test]
-    async fn subscribe_and_unsubscribe_do_not_error() {
-        let (mut peer, mut events_rx, commands_tx) = make_peer("ns-test");
-
-        // Drive the swarm just enough to get at least one listen address event,
-        // so the background run loop has something initialized.
-        let _addr = next_listen_addr(&mut peer).await;
-
-        // Run the peer loop in the background.
-        let handle = tokio::spawn(async move {
-            let _ = peer.run().await;
-        });
-
-        commands_tx
-            .send(ToSwarm::Subscribe("topic-a".to_string()))
-            .await
-            .unwrap();
-
-        commands_tx
-            .send(ToSwarm::Unsubscribe("topic-a".to_string()))
-            .await
-            .unwrap();
-
-        // We don't *require* any FromSwarm events here; this is mainly a
-        // smoke test that the message-handling path doesn't panic/hang.
-        // Still, poll briefly to ensure the task is alive.
-        let _ = timeout(Duration::from_millis(200), events_rx.recv()).await;
-
-        // Shut down: dropping the command sender closes the channel, causing run() to return Err.
-        drop(commands_tx);
-        let _ = handle.await;
-    }
+    /// Sealed traits support
+    pub trait Sealed {}
+    impl<T: ?Sized> Sealed for T {}
 }
--- a/rust/networking/src/swarm.rs
+++ b/rust/networking/src/swarm.rs
@@ -0,0 +1,145 @@
+use crate::alias;
+use crate::swarm::transport::tcp_transport;
+pub use behaviour::{Behaviour, BehaviourEvent};
+use libp2p::{SwarmBuilder, identity};
+
+pub type Swarm = libp2p::Swarm<Behaviour>;
+
+/// The current version of the network: this prevents devices running different versions of the
+/// software from interacting with each other.
+///
+/// TODO: right now this is a hardcoded constant; figure out what the versioning semantics should
+///       even be, and how to inject the right version into this config/initialization. E.g. should
+///       this be passed in as a parameter? What about rapidly changing versions in debug builds?
+///       this is all VERY very hard to figure out and needs to be mulled over as a team.
+pub const NETWORK_VERSION: &[u8] = b"v0.0.1";
+pub const OVERRIDE_VERSION_ENV_VAR: &str = "EXO_LIBP2P_NAMESPACE";
+
+/// Create and configure a swarm which listens to all ports on OS
+pub fn create_swarm(keypair: identity::Keypair) -> alias::AnyResult<Swarm> {
+    let mut swarm = SwarmBuilder::with_existing_identity(keypair)
+        .with_tokio()
+        .with_other_transport(tcp_transport)?
+        .with_behaviour(Behaviour::new)?
+        .build();
+
+    // Listen on all interfaces and whatever port the OS assigns
+    swarm.listen_on("/ip4/0.0.0.0/tcp/0".parse()?)?;
+    Ok(swarm)
+}
+
+mod transport {
+    use crate::alias;
+    use crate::swarm::{NETWORK_VERSION, OVERRIDE_VERSION_ENV_VAR};
+    use futures::{AsyncRead, AsyncWrite};
+    use keccak_const::Sha3_256;
+    use libp2p::core::muxing;
+    use libp2p::core::transport::Boxed;
+    use libp2p::pnet::{PnetError, PnetOutput};
+    use libp2p::{PeerId, Transport, identity, noise, pnet, yamux};
+    use std::{env, sync::LazyLock};
+
+    /// Key used for networking's private network; parametrized on the [`NETWORK_VERSION`].
+    /// See [`pnet_upgrade`] for more.
+    static PNET_PRESHARED_KEY: LazyLock<[u8; 32]> = LazyLock::new(|| {
+        let builder = Sha3_256::new().update(b"exo_discovery_network");
+
+        if let Ok(var) = env::var(OVERRIDE_VERSION_ENV_VAR) {
+            let bytes = var.into_bytes();
+            builder.update(&bytes)
+        } else {
+            builder.update(NETWORK_VERSION)
+        }
+        .finalize()
+    });
+
+    /// Make the Swarm run on a private network, as to not clash with public libp2p nodes and
+    /// also different-versioned instances of this same network.
+    /// This is implemented as an additional "upgrade" ontop of existing [`libp2p::Transport`] layers.
+    async fn pnet_upgrade<TSocket>(
+        socket: TSocket,
+        _: impl Sized,
+    ) -> Result<PnetOutput<TSocket>, PnetError>
+    where
+        TSocket: AsyncRead + AsyncWrite + Send + Unpin + 'static,
+    {
+        use pnet::{PnetConfig, PreSharedKey};
+        PnetConfig::new(PreSharedKey::new(*PNET_PRESHARED_KEY))
+            .handshake(socket)
+            .await
+    }
+
+    /// TCP/IP transport layer configuration.
+    pub fn tcp_transport(
+        keypair: &identity::Keypair,
+    ) -> alias::AnyResult<Boxed<(PeerId, muxing::StreamMuxerBox)>> {
+        use libp2p::{
+            core::upgrade::Version,
+            tcp::{Config, tokio},
+        };
+
+        // `TCP_NODELAY` enabled => avoid latency
+        let tcp_config = Config::default().nodelay(true);
+
+        // V1 + lazy flushing => 0-RTT negotiation
+        let upgrade_version = Version::V1Lazy;
+
+        // Noise is faster than TLS + we don't care much for security
+        let noise_config = noise::Config::new(keypair)?;
+
+        // Use default Yamux config for multiplexing
+        let yamux_config = yamux::Config::default();
+
+        // Create new Tokio-driven TCP/IP transport layer
+        let base_transport = tokio::Transport::new(tcp_config)
+            .and_then(pnet_upgrade)
+            .upgrade(upgrade_version)
+            .authenticate(noise_config)
+            .multiplex(yamux_config);
+
+        // Return boxed transport (to flatten complex type)
+        Ok(base_transport.boxed())
+    }
+}
+
+mod behaviour {
+    use crate::{alias, discovery};
+    use libp2p::swarm::NetworkBehaviour;
+    use libp2p::{gossipsub, identity};
+    use std::time::Duration;
+
+    /// Behavior of the Swarm which composes all desired behaviors:
+    /// Right now its just [`discovery::Behaviour`] and [`gossipsub::Behaviour`].
+    #[derive(NetworkBehaviour)]
+    pub struct Behaviour {
+        pub discovery: discovery::Behaviour,
+        pub gossipsub: gossipsub::Behaviour,
+    }
+
+    impl Behaviour {
+        pub fn new(keypair: &identity::Keypair) -> alias::AnyResult<Self> {
+            Ok(Self {
+                discovery: discovery::Behaviour::new(keypair)?,
+                gossipsub: gossipsub_behaviour(keypair),
+            })
+        }
+    }
+
+    fn gossipsub_behaviour(keypair: &identity::Keypair) -> gossipsub::Behaviour {
+        use gossipsub::{ConfigBuilder, MessageAuthenticity, ValidationMode};
+
+        // build a gossipsub network behaviour
+        //  => signed message authenticity + strict validation mode means the message-ID is
+        //     automatically provided by gossipsub w/out needing to provide custom message-ID function
+        gossipsub::Behaviour::new(
+            MessageAuthenticity::Signed(keypair.clone()),
+            ConfigBuilder::default()
+                .publish_queue_duration(Duration::from_secs(15))
+                .max_transmit_size(1024 * 1024)
+                .validation_mode(ValidationMode::Strict)
+                .build()
+                .expect("the configuration should always be valid"),
+        )
+        .expect("creating gossipsub behavior should always work")
+    }
+}
--- a/rust/networking/tests/dummy.rs
+++ b/rust/networking/tests/dummy.rs
@@ -0,0 +1,7 @@
+// maybe this will hold test in the future...??
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn does_nothing() {}
+}
--- a/rust/rust-toolchain.toml
+++ b/rust/rust-toolchain.toml
@@ -0,0 +1,2 @@
+[toolchain]
+channel = "nightly"
--- a/rust/util/Cargo.toml
+++ b/rust/util/Cargo.toml
@@ -0,0 +1,25 @@
+[package]
+name = "util"
+version = { workspace = true }
+edition = { workspace = true }
+publish = false
+
+[lib]
+doctest = false
+name = "util"
+path = "src/lib.rs"
+
+[lints]
+workspace = true
+
+[dependencies]
+# macro dependencies
+extend = { workspace = true }
+
+# utility dependencies
+thiserror = { workspace = true }
+once_cell = { workspace = true }
+internment = { workspace = true }
+derive_more = { workspace = true }
+bon = { workspace = true }
+recursion = { workspace = true }
--- a/rust/util/src/lib.rs
+++ b/rust/util/src/lib.rs
@@ -0,0 +1,53 @@
+//! TODO: crate documentation
+//!
+//! this is here as a placeholder documentation
+//!
+//!
+
+// enable Rust-unstable features for convenience
+#![feature(trait_alias)]
+#![feature(stmt_expr_attributes)]
+#![feature(type_alias_impl_trait)]
+#![feature(specialization)]
+#![feature(unboxed_closures)]
+#![feature(const_trait_impl)]
+#![feature(fn_traits)]
+
+pub mod nonempty;
+pub mod wakerdeque;
+
+pub(crate) mod private {
+    // sealed traits support
+    pub trait Sealed {}
+    impl<T: ?Sized> Sealed for T {}
+}
+
+/// Namespace for all the type/trait aliases used by this crate.
+pub(crate) mod alias {}
+
+/// Namespace for crate-wide extension traits/methods
+pub mod ext {
+    use extend::ext;
+
+    #[ext(pub, name = BoxedSliceExt)]
+    impl<T> Box<[T]> {
+        #[inline]
+        fn map<B, F>(self, f: F) -> Box<[B]>
+        where
+            F: FnMut(T) -> B,
+        {
+            self.into_iter().map(f).collect()
+        }
+    }
+
+    #[ext(pub, name = VecExt)]
+    impl<T> Vec<T> {
+        #[inline]
+        fn map<B, F>(self, f: F) -> Vec<B>
+        where
+            F: FnMut(T) -> B,
+        {
+            self.into_iter().map(f).collect()
+        }
+    }
+}
--- a/rust/util/src/nonempty.rs
+++ b/rust/util/src/nonempty.rs
@@ -0,0 +1,138 @@
+use std::slice::SliceIndex;
+use std::{ops, slice};
+use thiserror::Error;
+
+#[derive(Error, Debug)]
+#[error("Cannot create to `NonemptyArray` because the supplied slice is empty")]
+pub struct EmptySliceError;
+
+/// A pointer to a non-empty fixed-size slice allocated on the heap.
+#[derive(Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord)]
+#[repr(transparent)]
+pub struct NonemptyArray<T>(Box<[T]>);
+
+#[allow(clippy::arbitrary_source_item_ordering)]
+impl<T> NonemptyArray<T> {
+    #[inline]
+    pub fn singleton(value: T) -> Self {
+        Self(Box::new([value]))
+    }
+
+    #[allow(clippy::missing_errors_doc)]
+    #[inline]
+    pub fn try_from_boxed_slice<S: Into<Box<[T]>>>(
+        boxed_slice: S,
+    ) -> Result<Self, EmptySliceError> {
+        let boxed_slice = boxed_slice.into();
+        if boxed_slice.is_empty() {
+            Err(EmptySliceError)
+        } else {
+            Ok(Self(boxed_slice))
+        }
+    }
+
+    #[must_use]
+    #[inline]
+    pub fn into_boxed_slice(self) -> Box<[T]> {
+        self.0
+    }
+
+    #[must_use]
+    #[inline]
+    pub fn to_vec(&self) -> Vec<T>
+    where
+        T: Clone,
+    {
+        self.0.to_vec()
+    }
+
+    #[must_use]
+    #[inline]
+    pub const fn as_slice(&self) -> &[T] {
+        &self.0
+    }
+
+    #[allow(clippy::indexing_slicing)]
+    #[must_use]
+    #[inline]
+    pub fn first(&self) -> &T {
+        &self.0[0]
+    }
+
+    #[allow(clippy::indexing_slicing, clippy::arithmetic_side_effects)]
+    #[must_use]
+    #[inline]
+    pub fn last(&self) -> &T {
+        &self.0[self.0.len() - 1]
+    }
+
+    #[must_use]
+    #[inline]
+    pub fn get<I>(&self, index: I) -> Option<&I::Output>
+    where
+        I: SliceIndex<[T]>,
+    {
+        self.0.get(index)
+    }
+
+    #[allow(clippy::len_without_is_empty)]
+    #[must_use]
+    #[inline]
+    pub const fn len(&self) -> usize {
+        self.0.len()
+    }
+
+    #[allow(clippy::iter_without_into_iter)]
+    #[inline]
+    pub fn iter(&self) -> slice::Iter<'_, T> {
+        self.0.iter()
+    }
+
+    #[allow(clippy::iter_without_into_iter)]
+    #[inline]
+    pub fn iter_mut(&mut self) -> slice::IterMut<'_, T> {
+        self.0.iter_mut()
+    }
+
+    #[inline]
+    #[must_use]
+    pub fn map<U, F: FnMut(T) -> U>(self, f: F) -> NonemptyArray<U> {
+        NonemptyArray(self.0.into_iter().map(f).collect())
+    }
+}
+
+impl<T> From<NonemptyArray<T>> for Box<[T]> {
+    #[inline]
+    fn from(value: NonemptyArray<T>) -> Self {
+        value.into_boxed_slice()
+    }
+}
+
+impl<T> ops::Index<usize> for NonemptyArray<T> {
+    type Output = T;
+
+    #[inline]
+    fn index(&self, index: usize) -> &Self::Output {
+        self.0.index(index)
+    }
+}
+
+impl<T> IntoIterator for NonemptyArray<T> {
+    type Item = T;
+    type IntoIter = std::vec::IntoIter<T>;
+
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        self.into_boxed_slice().into_vec().into_iter()
+    }
+}
+
+impl<'a, T> IntoIterator for &'a NonemptyArray<T> {
+    type Item = &'a T;
+    type IntoIter = slice::Iter<'a, T>;
+
+    #[inline]
+    fn into_iter(self) -> Self::IntoIter {
+        self.iter()
+    }
+}
--- a/rust/util/src/wakerdeque.rs
+++ b/rust/util/src/wakerdeque.rs
@@ -0,0 +1,55 @@
+use std::collections::VecDeque;
+use std::fmt::{Debug, Formatter};
+use std::task::{Context, Waker};
+
+/// A wrapper around [`VecDeque`] which wakes (if it can) on any `push_*` methods,
+/// and updates the internally stored waker by consuming [`Context`] on any `pop_*` methods.
+pub struct WakerDeque<T> {
+    waker: Option<Waker>,
+    deque: VecDeque<T>,
+}
+
+impl<T: Debug> Debug for WakerDeque<T> {
+    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+        self.deque.fmt(f)
+    }
+}
+
+impl<T> WakerDeque<T> {
+    pub fn new() -> Self {
+        Self {
+            waker: None,
+            deque: VecDeque::new(),
+        }
+    }
+
+    fn update(&mut self, cx: &mut Context<'_>) {
+        self.waker = Some(cx.waker().clone());
+    }
+
+    fn wake(&mut self) {
+        let Some(ref mut w) = self.waker else { return };
+        w.wake_by_ref();
+        self.waker = None;
+    }
+
+    pub fn pop_front(&mut self, cx: &mut Context<'_>) -> Option<T> {
+        self.update(cx);
+        self.deque.pop_front()
+    }
+
+    pub fn pop_back(&mut self, cx: &mut Context<'_>) -> Option<T> {
+        self.update(cx);
+        self.deque.pop_back()
+    }
+
+    pub fn push_front(&mut self, value: T) {
+        self.wake();
+        self.deque.push_front(value);
+    }
+
+    pub fn push_back(&mut self, value: T) {
+        self.wake();
+        self.deque.push_back(value);
+    }
+}
--- a/src/exo/download/coordinator.py
+++ b/src/exo/download/coordinator.py
@@ -16,6 +16,7 @@ from exo.download.download_utils import (
 from exo.download.shard_downloader import ShardDownloader
 from exo.shared.models.model_cards import ModelId
 from exo.shared.types.commands import (
+    CancelDownload,
    DeleteDownload,
    ForwarderDownloadCommand,
    StartDownload,
@@ -107,6 +108,13 @@ class DownloadCoordinator:
                        await self._start_download(shard)
                    case DeleteDownload(model_id=model_id):
                        await self._delete_download(model_id)
+                    case CancelDownload(model_id=model_id):
+                        await self._cancel_download(model_id)
+
+    async def _cancel_download(self, model_id: ModelId) -> None:
+        if model_id in self.active_downloads and model_id in self.download_status:
+            logger.info(f"Cancelling download for {model_id}")
+            self.active_downloads.pop(model_id).cancel()

    async def _start_download(self, shard: ShardMetadata) -> None:
        model_id = shard.model_card.model_id
--- a/src/exo/download/download_utils.py
+++ b/src/exo/download/download_utils.py
@@ -158,6 +158,78 @@ async def seed_models(seed_dir: str | Path):
                    logger.error(traceback.format_exc())


+async def _build_file_list_from_local_directory(
+    model_id: ModelId,
+    recursive: bool = False,
+) -> list[FileListEntry] | None:
+    """Build a file list from locally existing model files.
+
+    We can only figure out the files we need from safetensors index, so
+    a local directory must contain a *.safetensors.index.json and
+    safetensors listed there.
+    """
+    model_dir = (await ensure_models_dir()) / model_id.normalize()
+    if not await aios.path.exists(model_dir):
+        return None
+
+    def _scan() -> list[FileListEntry] | None:
+        index_files = list(model_dir.glob("**/*.safetensors.index.json"))
+        if not index_files:
+            return None
+
+        entries_by_path: dict[str, FileListEntry] = {}
+
+        if recursive:
+            for dirpath, _, filenames in os.walk(model_dir):
+                for filename in filenames:
+                    if filename.endswith(".partial"):
+                        continue
+                    full_path = Path(dirpath) / filename
+                    rel_path = str(full_path.relative_to(model_dir))
+                    entries_by_path[rel_path] = FileListEntry(
+                        type="file",
+                        path=rel_path,
+                        size=full_path.stat().st_size,
+                    )
+        else:
+            for item in model_dir.iterdir():
+                if item.is_file() and not item.name.endswith(".partial"):
+                    entries_by_path[item.name] = FileListEntry(
+                        type="file",
+                        path=item.name,
+                        size=item.stat().st_size,
+                    )
+
+        # Add expected weight files from index that haven't been downloaded yet
+        for index_file in index_files:
+            try:
+                index_data = ModelSafetensorsIndex.model_validate_json(
+                    index_file.read_text()
+                )
+                relative_dir = index_file.parent.relative_to(model_dir)
+                for filename in set(index_data.weight_map.values()):
+                    rel_path = (
+                        str(relative_dir / filename)
+                        if relative_dir != Path(".")
+                        else filename
+                    )
+                    if rel_path not in entries_by_path:
+                        entries_by_path[rel_path] = FileListEntry(
+                            type="file",
+                            path=rel_path,
+                            size=None,
+                        )
+            except Exception:
+                continue
+
+        return list(entries_by_path.values())
+
+    file_list = await asyncio.to_thread(_scan)
+    if not file_list:
+        return None
+    return file_list
+
+
 _fetched_file_lists_this_session: set[str] = set()


@@ -183,6 +255,14 @@ async def fetch_file_list_with_cache(
        if await aios.path.exists(cache_file):
            async with aiofiles.open(cache_file, "r") as f:
                return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+        local_file_list = await _build_file_list_from_local_directory(
+            model_id, recursive
+        )
+        if local_file_list is not None:
+            logger.warning(
+                f"No internet and no cached file list for {model_id} - using local file list"
+            )
+            return local_file_list
        raise FileNotFoundError(
            f"No internet connection and no cached file list for {model_id}"
        )
@@ -203,10 +283,18 @@ async def fetch_file_list_with_cache(
    except Exception as e:
        if await aios.path.exists(cache_file):
            logger.warning(
-                f"Failed to fetch file list for {model_id}, using cached data: {e}"
+                f"No internet and no cached file list for {model_id} - using local file list"
            )
            async with aiofiles.open(cache_file, "r") as f:
                return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+        local_file_list = await _build_file_list_from_local_directory(
+            model_id, recursive
+        )
+        if local_file_list is not None:
+            logger.warning(
+                f"Failed to fetch file list for {model_id} and no cache exists, "
+            )
+            return local_file_list
        raise FileNotFoundError(f"Failed to fetch file list for {model_id}: {e}") from e


@@ -378,10 +466,14 @@ async def download_file_with_retry(
            logger.error(traceback.format_exc())
            await asyncio.sleep(2.0**attempt)
        except Exception as e:
-            on_connection_lost()
            if attempt == n_attempts - 1:
+                on_connection_lost()
                raise e
-            break
+            logger.error(
+                f"Download error on attempt {attempt + 1}/{n_attempts} for {model_id=} {revision=} {path=} {target_dir=}"
+            )
+            logger.error(traceback.format_exc())
+            await asyncio.sleep(2.0**attempt)
    raise Exception(
        f"Failed to download file {model_id=} {revision=} {path=} {target_dir=}"
    )
--- a/src/exo/download/impl_shard_downloader.py
+++ b/src/exo/download/impl_shard_downloader.py
@@ -195,6 +195,10 @@ class ResumableShardDownloader(ShardDownloader):
        self, shard: ShardMetadata
    ) -> RepoDownloadProgress:
        _, progress = await download_shard(
-            shard, self.on_progress_wrapper, skip_download=True
+            shard,
+            self.on_progress_wrapper,
+            skip_download=True,
+            skip_internet=not self.internet_connection,
+            on_connection_lost=lambda: self.set_internet_connection(False),
        )
        return progress
--- a/src/exo/main.py
+++ b/src/exo/main.py
@@ -1,5 +1,4 @@
 import argparse
-import importlib.metadata
 import itertools
 import multiprocessing as mp
 import os
@@ -45,9 +44,9 @@ class Node:
    @classmethod
    async def create(cls, args: "Args") -> "Self":
        keypair = get_node_id_keypair()
-        node_id = NodeId(keypair.to_string())
+        node_id = NodeId(keypair.to_peer_id().to_base58())
        session_id = SessionId(master_node_id=node_id, election_clock=0)
-        router = Router.create(keypair, namespace=args.namespace)
+        router = Router.create(keypair)
        await router.register_topic(topics.GLOBAL_EVENTS)
        await router.register_topic(topics.LOCAL_EVENTS)
        await router.register_topic(topics.COMMANDS)
@@ -73,7 +72,7 @@ class Node:
        else:
            download_coordinator = None

-        if not args.no_api:
+        if args.spawn_api:
            api = API(
                node_id,
                session_id,
@@ -106,6 +105,7 @@ class Node:
            global_event_sender=router.sender(topics.GLOBAL_EVENTS),
            local_event_receiver=router.receiver(topics.LOCAL_EVENTS),
            command_receiver=router.receiver(topics.COMMANDS),
+            download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
        )

        er_send, er_recv = channel[ElectionResult]()
@@ -189,6 +189,9 @@ class Node:
                        global_event_sender=self.router.sender(topics.GLOBAL_EVENTS),
                        local_event_receiver=self.router.receiver(topics.LOCAL_EVENTS),
                        command_receiver=self.router.receiver(topics.COMMANDS),
+                        download_command_sender=self.router.sender(
+                            topics.DOWNLOAD_COMMANDS
+                        ),
                    )
                    self._tg.start_soon(self.master.run)
                elif (
@@ -254,7 +257,7 @@ def main():
    # TODO: Refactor the current verbosity system
    logger_setup(EXO_LOG, args.verbosity)
    logger.info("Starting EXO")
-    logger.info(f"Namespace: {args.namespace}")
+    logger.info(f"EXO_LIBP2P_NAMESPACE: {os.getenv('EXO_LIBP2P_NAMESPACE')}")

    # Set FAST_SYNCH override env var for runner subprocesses
    if args.fast_synch is True:
@@ -271,13 +274,13 @@ def main():


 class Args(CamelCaseModel):
-    verbosity: int
-    force_master: bool
-    no_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
-    namespace: str
    fast_synch: bool | None = None  # None = auto, True = force on, False = force off

    @classmethod
@@ -307,15 +310,14 @@ class Args(CamelCaseModel):
        )
        parser.add_argument(
            "--no-api",
-            action="store_true",
-            help="Disable the API server for this node",
+            action="store_false",
+            dest="spawn_api",
        )
        parser.add_argument(
            "--api-port",
            type=int,
            dest="api_port",
            default=52415,
-            help="Which port the API server will be available on",
        )
        parser.add_argument(
            "--no-worker",
@@ -326,11 +328,6 @@ class Args(CamelCaseModel):
            action="store_true",
            help="Disable the download coordinator (node won't download models)",
        )
-        parser.add_argument(
-            "--namespace",
-            default=importlib.metadata.version("exo"),
-            help="Set the EXO namespace to run multiple isolated clusters",
-        )
        fast_synch_group = parser.add_mutually_exclusive_group()
        fast_synch_group.add_argument(
            "--fast-synch",
--- a/src/exo/master/main.py
+++ b/src/exo/master/main.py
@@ -6,6 +6,7 @@ from loguru import logger

 from exo.master.placement import (
    add_instance_to_placements,
+    cancel_unnecessary_downloads,
    delete_instance,
    get_transition_events,
    place_instance,
@@ -16,6 +17,7 @@ from exo.shared.types.commands import (
    CreateInstance,
    DeleteInstance,
    ForwarderCommand,
+    ForwarderDownloadCommand,
    ImageEdits,
    ImageGeneration,
    PlaceInstance,
@@ -66,12 +68,9 @@ class Master:
        session_id: SessionId,
        *,
        command_receiver: Receiver[ForwarderCommand],
-        # Receiving indexed events from the forwarder to be applied to state
-        # Ideally these would be WorkerForwarderEvents but type system says no :(
        local_event_receiver: Receiver[ForwarderEvent],
-        # Send events to the forwarder to be indexed (usually from command processing)
-        # Ideally these would be MasterForwarderEvents but type system says no :(
        global_event_sender: Sender[ForwarderEvent],
+        download_command_sender: Sender[ForwarderDownloadCommand],
    ):
        self.state = State()
        self._tg: TaskGroup = anyio.create_task_group()
@@ -81,6 +80,7 @@ class Master:
        self.command_receiver = command_receiver
        self.local_event_receiver = local_event_receiver
        self.global_event_sender = global_event_sender
+        self.download_command_sender = download_command_sender
        send, recv = channel[Event]()
        self.event_sender: Sender[Event] = send
        self._loopback_event_receiver: Receiver[Event] = recv
@@ -280,6 +280,14 @@ class Master:
                            transition_events = get_transition_events(
                                self.state.instances, placement
                            )
+                            for cmd in cancel_unnecessary_downloads(
+                                placement, self.state.downloads
+                            ):
+                                await self.download_command_sender.send(
+                                    ForwarderDownloadCommand(
+                                        origin=self.node_id, command=cmd
+                                    )
+                                )
                            generated_events.extend(transition_events)
                        case PlaceInstance():
                            placement = place_instance(
@@ -371,7 +379,7 @@ class Master:
                        await self._handle_traces_collected(event)
                        continue

-                    logger.trace(f"Master indexing event: {str(event)[:100]}")
+                    logger.debug(f"Master indexing event: {str(event)[:100]}")
                    indexed = IndexedEvent(event=event, idx=len(self._event_log))
                    self.state = apply(self.state, indexed)

--- a/src/exo/master/placement.py
+++ b/src/exo/master/placement.py
@@ -15,14 +15,20 @@ from exo.master.placement_utils import (
 from exo.shared.models.model_cards import ModelId
 from exo.shared.topology import Topology
 from exo.shared.types.commands import (
+    CancelDownload,
    CreateInstance,
    DeleteInstance,
+    DownloadCommand,
    PlaceInstance,
 )
 from exo.shared.types.common import NodeId
 from exo.shared.types.events import Event, InstanceCreated, InstanceDeleted
 from exo.shared.types.memory import Memory
 from exo.shared.types.profiling import MemoryUsage, NodeNetworkInfo
+from exo.shared.types.worker.downloads import (
+    DownloadOngoing,
+    DownloadProgress,
+)
 from exo.shared.types.worker.instances import (
    Instance,
    InstanceId,
@@ -202,3 +208,29 @@ def get_transition_events(
            )

    return events
+
+
+def cancel_unnecessary_downloads(
+    instances: Mapping[InstanceId, Instance],
+    download_status: Mapping[NodeId, Sequence[DownloadProgress]],
+) -> Sequence[DownloadCommand]:
+    commands: list[DownloadCommand] = []
+    currently_downloading = [
+        (k, v.shard_metadata.model_card.model_id)
+        for k, vs in download_status.items()
+        for v in vs
+        if isinstance(v, (DownloadOngoing))
+    ]
+    active_models = set(
+        (
+            node_id,
+            instance.shard_assignments.runner_to_shard[runner_id].model_card.model_id,
+        )
+        for instance in instances.values()
+        for node_id, runner_id in instance.shard_assignments.node_to_runner.items()
+    )
+    for pair in currently_downloading:
+        if pair not in active_models:
+            commands.append(CancelDownload(target_node_id=pair[0], model_id=pair[1]))
+
+    return commands
--- a/src/exo/master/tests/test_master.py
+++ b/src/exo/master/tests/test_master.py
@@ -11,6 +11,7 @@ from exo.shared.models.model_cards import ModelCard, ModelTask
 from exo.shared.types.commands import (
    CommandId,
    ForwarderCommand,
+    ForwarderDownloadCommand,
    PlaceInstance,
    TextGeneration,
 )
@@ -41,12 +42,13 @@ from exo.utils.channels import channel
@pytest.mark.asyncio
 async def test_master():
    keypair = get_node_id_keypair()
-    node_id = NodeId(keypair.to_string())
+    node_id = NodeId(keypair.to_peer_id().to_base58())
    session_id = SessionId(master_node_id=node_id, election_clock=0)

    ge_sender, global_event_receiver = channel[ForwarderEvent]()
    command_sender, co_receiver = channel[ForwarderCommand]()
    local_event_sender, le_receiver = channel[ForwarderEvent]()
+    fcds, _fcdr = channel[ForwarderDownloadCommand]()

    all_events: list[IndexedEvent] = []

@@ -67,12 +69,13 @@ async def test_master():
        global_event_sender=ge_sender,
        local_event_receiver=le_receiver,
        command_receiver=co_receiver,
+        download_command_sender=fcds,
    )
    logger.info("run the master")
    async with anyio.create_task_group() as tg:
        tg.start_soon(master.run)

-        sender_node_id = NodeId(f"{keypair.to_string()}_sender")
+        sender_node_id = NodeId(f"{keypair.to_peer_id().to_base58()}_sender")
        # inject a NodeGatheredInfo event
        logger.info("inject a NodeGatheredInfo event")
        await local_event_sender.send(
--- a/src/exo/routing/connection_message.py
+++ b/src/exo/routing/connection_message.py
@@ -1,9 +1,37 @@
+from enum import Enum
+
+from exo_pyo3_bindings import ConnectionUpdate, ConnectionUpdateType
+
 from exo.shared.types.common import NodeId
 from exo.utils.pydantic_ext import CamelCaseModel

 """Serialisable types for Connection Updates/Messages"""


+class ConnectionMessageType(Enum):
+    Connected = 0
+    Disconnected = 1
+
+    @staticmethod
+    def from_update_type(update_type: ConnectionUpdateType):
+        match update_type:
+            case ConnectionUpdateType.Connected:
+                return ConnectionMessageType.Connected
+            case ConnectionUpdateType.Disconnected:
+                return ConnectionMessageType.Disconnected
+
+
 class ConnectionMessage(CamelCaseModel):
    node_id: NodeId
-    expired: bool
+    connection_type: ConnectionMessageType
+    remote_ipv4: str
+    remote_tcp_port: int
+
+    @classmethod
+    def from_update(cls, update: ConnectionUpdate) -> "ConnectionMessage":
+        return cls(
+            node_id=NodeId(update.peer_id.to_base58()),
+            connection_type=ConnectionMessageType.from_update_type(update.update_type),
+            remote_ipv4=update.remote_ipv4,
+            remote_tcp_port=update.remote_tcp_port,
+        )
--- a/src/exo/routing/router.py
+++ b/src/exo/routing/router.py
@@ -1,5 +1,5 @@
 from copy import copy
-from dataclasses import dataclass, field
+from itertools import count
 from math import inf
 from os import PathLike
 from pathlib import Path
@@ -14,14 +14,15 @@ from anyio import (
 )
 from anyio.abc import TaskGroup
 from exo_pyo3_bindings import (
+    AllQueuesFullError,
    Keypair,
-    PyPeer,
+    NetworkingHandle,
+    NoPeersSubscribedToTopicError,
 )
 from filelock import FileLock
 from loguru import logger

 from exo.shared.constants import EXO_NODE_ID_KEYPAIR
-from exo.shared.types.common import NodeId
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.pydantic_ext import CamelCaseModel

@@ -98,32 +99,28 @@ class TopicRouter[T: CamelCaseModel]:
        )


-@dataclass
 class Router:
-    _peer: PyPeer
-    topic_routers: dict[str, TopicRouter[CamelCaseModel]] = field(
-        init=False, default_factory=dict
-    )
-    networking_receiver: Receiver[tuple[str, bytes]] = field(init=False)
-    _tmp_networking_sender: Sender[tuple[str, bytes]] | None = field(init=False)
-    _tg: TaskGroup | None = None
-
-    def __post_init__(self):
-        self._tmp_networking_sender, self.networking_receiver = channel()
-
    @classmethod
-    def create(cls, identity: Keypair, namespace: str) -> "Router":
-        return cls(_peer=PyPeer.new(identity, namespace))
+    def create(cls, identity: Keypair) -> "Router":
+        return cls(handle=NetworkingHandle(identity))
+
+    def __init__(self, handle: NetworkingHandle):
+        self.topic_routers: dict[str, TopicRouter[CamelCaseModel]] = {}
+        send, recv = channel[tuple[str, bytes]]()
+        self.networking_receiver: Receiver[tuple[str, bytes]] = recv
+        self._net: NetworkingHandle = handle
+        self._tmp_networking_sender: Sender[tuple[str, bytes]] | None = send
+        self._id_count = count()
+        self._tg: TaskGroup | None = None

    async def register_topic[T: CamelCaseModel](self, topic: TypedTopic[T]):
+        assert self._tg is None, "Attempted to register topic after setup time"
        send = self._tmp_networking_sender
        if send:
            self._tmp_networking_sender = None
        else:
            send = self.networking_receiver.clone_sender()
        router = TopicRouter[T](topic, send)
-        if self._tg is not None:
-            self._tg.start_soon(router.run)
        self.topic_routers[topic.topic] = cast(TopicRouter[CamelCaseModel], router)
        await self._networking_subscribe(str(topic.topic))

@@ -151,18 +148,14 @@ class Router:
    async def run(self):
        logger.debug("Starting Router")
        try:
-
-            async def _peer_run():
-                await self._peer.run()
-
            async with create_task_group() as tg:
                self._tg = tg
                for topic in self.topic_routers:
                    router = self.topic_routers[topic]
                    tg.start_soon(router.run)
                tg.start_soon(self._networking_recv)
+                tg.start_soon(self._networking_recv_connection_messages)
                tg.start_soon(self._networking_publish)
-                tg.start_soon(_peer_run)
                # Router only shuts down if you cancel it.
                await sleep_forever()
        finally:
@@ -177,58 +170,47 @@ class Router:
        self._tg.cancel_scope.cancel()

    async def _networking_subscribe(self, topic: str):
-        await self._peer.subscribe(topic)
+        await self._net.gossipsub_subscribe(topic)
        logger.info(f"Subscribed to {topic}")

    async def _networking_unsubscribe(self, topic: str):
-        await self._peer.unsubscribe(topic)
+        await self._net.gossipsub_unsubscribe(topic)
        logger.info(f"Unsubscribed from {topic}")

    async def _networking_recv(self):
        while True:
-            try:
-                swarm_event = await self._peer.recv()
-            except ValueError:
-                logger.error("Message too large for gossipsub, dropped")
-                continue
-            except ConnectionError:
-                logger.error("All peer queues full, network overloaded")
-                continue
-            except RuntimeError:
-                break
-
-            cm = None
-            if (peer_id := swarm_event.downcast_discovered()) is not None:
-                cm = ConnectionMessage(node_id=NodeId(peer_id), expired=False)
-            if (peer_id := swarm_event.downcast_expired()) is not None:
-                cm = ConnectionMessage(node_id=NodeId(peer_id), expired=True)
-
-            if cm is not None:
-                if CONNECTION_MESSAGES.topic in self.topic_routers:
-                    router = self.topic_routers[CONNECTION_MESSAGES.topic]
-                    assert router.topic.model_type == ConnectionMessage
-                    router = cast(TopicRouter[ConnectionMessage], router)
-                    await router.publish(cm)
-                continue
-
-            assert (msg := swarm_event.downcast_message()) is not None
-            _origin, topic, payload = msg
-            logger.debug(f"Received message on {topic} with payload {payload}")
+            topic, data = await self._net.gossipsub_recv()
+            logger.trace(f"Received message on {topic} with payload {data}")
            if topic not in self.topic_routers:
                logger.warning(f"Received message on unknown or inactive topic {topic}")
                continue

            router = self.topic_routers[topic]
-            await router.publish_bytes(payload)
+            await router.publish_bytes(data)
+
+    async def _networking_recv_connection_messages(self):
+        while True:
+            update = await self._net.connection_update_recv()
+            message = ConnectionMessage.from_update(update)
+            logger.trace(
+                f"Received message on connection_messages with payload {message}"
+            )
+            if CONNECTION_MESSAGES.topic in self.topic_routers:
+                router = self.topic_routers[CONNECTION_MESSAGES.topic]
+                assert router.topic.model_type == ConnectionMessage
+                router = cast(TopicRouter[ConnectionMessage], router)
+                await router.publish(message)

    async def _networking_publish(self):
        with self.networking_receiver as networked_items:
            async for topic, data in networked_items:
                try:
                    logger.trace(f"Sending message on {topic} with payload {data}")
-                    await self._peer.send(topic, data)
-                except RuntimeError:
-                    break
+                    await self._net.gossipsub_publish(topic, data)
+                # As a hack, this also catches AllQueuesFull
+                # Need to fix that ASAP.
+                except (NoPeersSubscribedToTopicError, AllQueuesFullError):
+                    pass


 def get_node_id_keypair(
@@ -239,7 +221,7 @@ def get_node_id_keypair(
    Obtain the :class:`PeerId` by from it.
    """
    # TODO(evan): bring back node id persistence once we figure out how to deal with duplicates
-    return Keypair.generate()
+    return Keypair.generate_ed25519()

    def lock_path(path: str | bytes | PathLike[str] | PathLike[bytes]) -> Path:
        return Path(str(path) + ".lock")
--- a/src/exo/shared/tests/test_election.py
+++ b/src/exo/shared/tests/test_election.py
@@ -1,7 +1,7 @@
 import pytest
 from anyio import create_task_group, fail_after, move_on_after

-from exo.routing.connection_message import ConnectionMessage
+from exo.routing.connection_message import ConnectionMessage, ConnectionMessageType
 from exo.shared.election import Election, ElectionMessage, ElectionResult
 from exo.shared.types.commands import ForwarderCommand, TestCommand
 from exo.shared.types.common import NodeId, SessionId
@@ -330,7 +330,9 @@ async def test_connection_message_triggers_new_round_broadcast() -> None:
            await cm_tx.send(
                ConnectionMessage(
                    node_id=NodeId(),
-                    expired=False,
+                    connection_type=ConnectionMessageType.Connected,
+                    remote_ipv4="",
+                    remote_tcp_port=0,
                )
            )

--- a/src/exo/shared/types/commands.py
+++ b/src/exo/shared/types/commands.py
@@ -72,7 +72,12 @@ class DeleteDownload(BaseCommand):
    model_id: ModelId


-DownloadCommand = StartDownload | DeleteDownload
+class CancelDownload(BaseCommand):
+    target_node_id: NodeId
+    model_id: ModelId
+
+
+DownloadCommand = StartDownload | DeleteDownload | CancelDownload


 Command = (
--- a/src/exo/shared/types/mlx.py
+++ b/src/exo/shared/types/mlx.py
@@ -3,10 +3,11 @@
 from collections.abc import Sequence

 from mlx_lm.models.cache import (
+    ArraysCache,
    KVCache,
    QuantizedKVCache,
    RotatingKVCache,
 )

 # This list contains one cache entry per transformer layer
-KVCacheType = Sequence[KVCache | RotatingKVCache | QuantizedKVCache]
+KVCacheType = Sequence[KVCache | RotatingKVCache | QuantizedKVCache | ArraysCache]
--- a/src/exo/worker/engines/mlx/auto_parallel.py
+++ b/src/exo/worker/engines/mlx/auto_parallel.py
@@ -13,6 +13,9 @@ from mlx.nn.layers.distributed import (
    shard_linear,
    sum_gradients,
 )
+from mlx_lm.models.base import (
+    scaled_dot_product_attention,  # pyright: ignore[reportUnknownVariableType]
+)
 from mlx_lm.models.deepseek_v3 import DeepseekV3MLP
 from mlx_lm.models.deepseek_v3 import Model as DeepseekV3Model
 from mlx_lm.models.deepseek_v32 import DeepseekV32MLP
@@ -25,16 +28,21 @@ from mlx_lm.models.gpt_oss import GptOssMoeModel
 from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.models.kimi_k25 import Model as KimiK25Model
 from mlx_lm.models.llama import Model as LlamaModel
+from mlx_lm.models.minimax import MiniMaxAttention
 from mlx_lm.models.minimax import Model as MiniMaxModel
 from mlx_lm.models.ministral3 import Model as Ministral3Model
 from mlx_lm.models.qwen3_moe import Model as Qwen3MoeModel
 from mlx_lm.models.qwen3_moe import Qwen3MoeSparseMoeBlock
 from mlx_lm.models.qwen3_next import Model as Qwen3NextModel
-from mlx_lm.models.qwen3_next import Qwen3NextSparseMoeBlock
+from mlx_lm.models.qwen3_next import Qwen3NextDecoderLayer, Qwen3NextSparseMoeBlock
+from transformers.models.qwen3.modeling_qwen3 import Qwen3DecoderLayer

 from exo.shared.logging import logger
 from exo.shared.types.worker.shards import PipelineShardMetadata

+if TYPE_CHECKING:
+    from mlx_lm.models.cache import Cache
+
 TimeoutCallback = Callable[[], None]


@@ -503,12 +511,24 @@ class DeepSeekShardingStrategy(TensorParallelShardingStrategy):
                layer.self_attn.q_b_proj = self.all_to_sharded_linear(
                    layer.self_attn.q_b_proj
                )
-            layer.self_attn.kv_b_proj = self.all_to_sharded_linear(
-                layer.self_attn.kv_b_proj
-            )
+
+            # layer.self_attn.kv_b_proj = self.all_to_sharded_linear(
+            #     layer.self_attn.kv_b_proj
+            # )
            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
            layer.self_attn.num_heads //= self.N

+            # Logic from upstream mlx
+            num_heads = layer.self_attn.num_heads
+            sh = self.group.rank() * num_heads
+            eh = sh + num_heads
+
+            def shard_heads(w: mx.array, sh: int = sh, eh: int = eh) -> mx.array:
+                return w[sh:eh]
+
+            layer.self_attn.embed_q.apply(shard_heads)
+            layer.self_attn.unembed_out.apply(shard_heads)
+
            # Shard the MLP
            if isinstance(layer.mlp, (DeepseekV3MLP, DeepseekV32MLP)):
                layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
@@ -624,6 +644,84 @@ class ShardedGLM4MoeLiteMoE(CustomMlxLayer):
        return y


+class WrappedMiniMaxAttention(CustomMlxLayer):
+    def __init__(self, layer: _LayerCallable, group: mx.distributed.Group):
+        super().__init__(layer)
+        self.group = group
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: mx.array | None = None,
+        cache: "Cache | None" = None,
+    ) -> mx.array:
+        batch_dim, seq_dim, _ = x.shape
+
+        self._original_layer = cast(MiniMaxAttention, self.original_layer)  # type: ignore
+
+        queries: mx.array = self._original_layer.q_proj(x)
+        keys: mx.array = self._original_layer.k_proj(x)
+        values: mx.array = self._original_layer.v_proj(x)
+
+        if getattr(self, "use_qk_norm", False):
+            q_dim = queries.shape[-1]
+            k_dim = keys.shape[-1]
+            n = self.group.size()
+
+            qk = mx.concatenate(
+                [queries, keys], axis=-1
+            )  # (batch_dim, seq_dim, q_dim + k_dim)
+            qk = mx.distributed.all_gather(
+                qk, group=self.group
+            )  # (n*batch_dim, seq_dim, q_dim + k_dim)
+
+            qk = qk.reshape(n, batch_dim, seq_dim, q_dim + k_dim).transpose(1, 2, 0, 3)
+            queries = qk[..., :q_dim].reshape(
+                batch_dim, seq_dim, -1
+            )  # (batch_dim, seq_dim, n * q_dim)
+            keys = qk[..., q_dim:].reshape(
+                batch_dim, seq_dim, -1
+            )  # (batch_dim, seq_dim, n * k_dim)
+
+            queries = self._original_layer.q_norm(queries)
+            keys = self._original_layer.k_norm(keys)
+
+            # Split back and take this rank's portion
+            queries = mx.split(queries, n, axis=-1)[self.group.rank()]
+            keys = mx.split(keys, n, axis=-1)[self.group.rank()]
+
+        queries = queries.reshape(
+            batch_dim, seq_dim, self._original_layer.num_attention_heads, -1
+        ).transpose(0, 2, 1, 3)
+        keys = keys.reshape(
+            batch_dim, seq_dim, self._original_layer.num_key_value_heads, -1
+        ).transpose(0, 2, 1, 3)
+        values = values.reshape(
+            batch_dim, seq_dim, self._original_layer.num_key_value_heads, -1
+        ).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            queries = self._original_layer.rope(queries, offset=cache.offset)
+            keys = self._original_layer.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self._original_layer.rope(queries)
+            keys = self._original_layer.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries,
+            keys,
+            values,
+            cache=cache,
+            scale=self._original_layer.scale,  # type: ignore
+            mask=mask,
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(batch_dim, seq_dim, -1)
+
+        return self._original_layer.o_proj(output)
+
+
 class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
    def shard_model(
        self,
@@ -632,7 +730,6 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
        on_timeout: TimeoutCallback | None,
    ) -> nn.Module:
        model = cast(MiniMaxModel, model)
-        rank = self.group.rank()
        for layer in model.layers:
            eval_with_timeout(
                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
@@ -643,18 +740,11 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)

-            # Shard qk_norm weights if present (must match sharded head count)
-            if getattr(layer.self_attn, "use_qk_norm", False):
-                layer.self_attn.q_norm.weight = layer.self_attn.q_norm.weight.split(  # type: ignore
-                    self.N, axis=-1
-                )[rank]
-                layer.self_attn.k_norm.weight = layer.self_attn.k_norm.weight.split(  # type: ignore
-                    self.N, axis=-1
-                )[rank]
-
            layer.self_attn.num_attention_heads //= self.N
            layer.self_attn.num_key_value_heads //= self.N

+            layer.self_attn = WrappedMiniMaxAttention(layer.self_attn, self.group)  # pyright: ignore[reportAttributeAccessIssue,reportArgumentType]
+
            # Shard the MoE. Shard in place since the MoE should be responsible
            # for aggregating the results.
            self.all_to_sharded_linear_in_place(
@@ -679,18 +769,95 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
        timeout_seconds: float,
        on_timeout: TimeoutCallback | None,
    ) -> nn.Module:
-        model = cast(Qwen3MoeModel, model)
+        model = cast(Qwen3MoeModel | Qwen3NextModel, model)
        for layer in model.layers:
            eval_with_timeout(
                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
            )
            # Shard the self attention
-            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
-            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
-            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
-            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
-            layer.self_attn.n_heads //= self.N
-            layer.self_attn.n_kv_heads //= self.N
+            if isinstance(layer, Qwen3DecoderLayer):
+                layer.self_attn.q_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_proj
+                )
+                layer.self_attn.k_proj = self.all_to_sharded_linear(
+                    layer.self_attn.k_proj
+                )
+                layer.self_attn.v_proj = self.all_to_sharded_linear(
+                    layer.self_attn.v_proj
+                )
+                layer.self_attn.o_proj = self.sharded_to_all_linear(
+                    layer.self_attn.o_proj
+                )
+            else:
+                assert isinstance(layer, Qwen3NextDecoderLayer)
+                if hasattr(layer, "linear_attn"):
+                    linear_attn = layer.linear_attn
+
+                    linear_attn.in_proj_qkvz = self.all_to_sharded_linear(
+                        linear_attn.in_proj_qkvz
+                    )
+                    linear_attn.in_proj_ba = self.all_to_sharded_linear(
+                        linear_attn.in_proj_ba
+                    )
+                    linear_attn.out_proj = self.sharded_to_all_linear(
+                        linear_attn.out_proj
+                    )
+
+                    # Shard conv1d: depthwise conv with non-contiguous channel slicing.
+                    # Channel layout is [q(key_dim), k(key_dim), v(value_dim)].
+                    # Each rank takes its head-slice from each of the three sections.
+                    rank = self.group.rank()
+                    key_dim = linear_attn.key_dim
+                    value_dim = linear_attn.value_dim
+                    key_dim_shard = key_dim // self.N
+                    value_dim_shard = value_dim // self.N
+
+                    q_idx = mx.arange(rank * key_dim_shard, (rank + 1) * key_dim_shard)
+                    k_idx = mx.arange(
+                        key_dim + rank * key_dim_shard,
+                        key_dim + (rank + 1) * key_dim_shard,
+                    )
+                    v_idx = mx.arange(
+                        2 * key_dim + rank * value_dim_shard,
+                        2 * key_dim + (rank + 1) * value_dim_shard,
+                    )
+                    conv_indices = mx.concatenate([q_idx, k_idx, v_idx])
+                    linear_attn.conv1d.weight = linear_attn.conv1d.weight[conv_indices]
+                    new_conv_dim = key_dim_shard * 2 + value_dim_shard
+                    linear_attn.conv1d.groups = new_conv_dim
+
+                    num_v_shard = linear_attn.num_v_heads // self.N
+                    v_start = rank * num_v_shard
+                    v_end = v_start + num_v_shard
+                    linear_attn.A_log = linear_attn.A_log[v_start:v_end]
+                    linear_attn.dt_bias = linear_attn.dt_bias[v_start:v_end]
+
+                    linear_attn.num_k_heads //= self.N
+                    linear_attn.num_v_heads //= self.N
+                    linear_attn.key_dim = (
+                        linear_attn.head_k_dim * linear_attn.num_k_heads
+                    )
+                    linear_attn.value_dim = (
+                        linear_attn.head_v_dim * linear_attn.num_v_heads
+                    )
+                    linear_attn.conv_dim = (
+                        linear_attn.key_dim * 2 + linear_attn.value_dim
+                    )
+                else:
+                    layer.self_attn.q_proj = self.all_to_sharded_linear(
+                        layer.self_attn.q_proj
+                    )
+                    layer.self_attn.k_proj = self.all_to_sharded_linear(
+                        layer.self_attn.k_proj
+                    )
+                    layer.self_attn.v_proj = self.all_to_sharded_linear(
+                        layer.self_attn.v_proj
+                    )
+                    layer.self_attn.o_proj = self.sharded_to_all_linear(
+                        layer.self_attn.o_proj
+                    )
+                    layer.self_attn.num_attention_heads //= self.N
+                    layer.self_attn.num_key_value_heads //= self.N

            # Shard the MoE. Shard in place since the MoE should be responsible
            # for aggregating the results.
@@ -700,6 +867,14 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.gate_proj)
                self.sharded_to_all_linear_in_place(layer.mlp.switch_mlp.down_proj)
                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.up_proj)
+                if isinstance(layer.mlp, Qwen3NextSparseMoeBlock):
+                    self.all_to_sharded_linear_in_place(
+                        layer.mlp.shared_expert.gate_proj
+                    )
+                    self.sharded_to_all_linear_in_place(
+                        layer.mlp.shared_expert.down_proj
+                    )
+                    self.all_to_sharded_linear_in_place(layer.mlp.shared_expert.up_proj)
                layer.mlp = ShardedQwenMoE(layer.mlp)  # pyright: ignore[reportAttributeAccessIssue, reportArgumentType]
                layer.mlp.sharding_group = self.group

--- a/src/exo/worker/engines/mlx/cache.py
+++ b/src/exo/worker/engines/mlx/cache.py
@@ -1,16 +1,14 @@
 import os
 from copy import deepcopy
-from typing import Any, cast

 import mlx.core as mx
 import psutil
 from mlx_lm.models.cache import (
+    ArraysCache,
    KVCache,
    QuantizedKVCache,
    RotatingKVCache,
-    trim_prompt_cache,
 )
-from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper

 from exo.shared.types.memory import Memory
@@ -26,51 +24,119 @@ _MEMORY_THRESHOLD = float(
 )


-class KVPrefixCache:
+class CacheSnapshot:
+    """Snapshot of states at a known token position."""
+
    def __init__(
-        self, tokenizer: TokenizerWrapper, group: mx.distributed.Group | None = None
+        self, states: list[RotatingKVCache | ArraysCache | None], token_count: int
    ):
+        self.states = states
+        self.token_count = token_count
+
+
+def snapshot_ssm_states(cache: KVCacheType) -> CacheSnapshot:
+    states: list[ArraysCache | RotatingKVCache | None] = []
+    for c in cache:
+        if isinstance(c, (ArraysCache, RotatingKVCache)):
+            states.append(deepcopy(c))
+        else:
+            states.append(None)
+    token_count = cache_length(cache)
+    return CacheSnapshot(states=states, token_count=token_count)
+
+
+def _find_nearest_snapshot(
+    snapshots: list[CacheSnapshot],
+    target_token_count: int,
+) -> CacheSnapshot | None:
+    best: CacheSnapshot | None = None
+    for snap in snapshots:
+        if snap.token_count <= target_token_count and (
+            best is None or snap.token_count > best.token_count
+        ):
+            best = snap
+    return best
+
+
+def has_non_kv_caches(cache: KVCacheType) -> bool:
+    """Check if a cache contains any ArraysCache (SSM) entries."""
+    return any(isinstance(c, (ArraysCache, RotatingKVCache)) for c in cache)
+
+
+class KVPrefixCache:
+    def __init__(self, group: mx.distributed.Group | None = None):
        self.prompts: list[mx.array] = []  # mx array of tokens (ints)
        self.caches: list[KVCacheType] = []
+        self._snapshots: list[list[CacheSnapshot] | None] = []
        self._last_used: list[int] = []  # monotonic counter of last access per entry
        self._access_counter: int = 0
-        self._tokenizer: TokenizerWrapper = tokenizer
        self._group = group

    def clear(self):
        """Clear all cached prompts and caches."""
        self.prompts.clear()
        self.caches.clear()
+        self._snapshots.clear()
        self._last_used.clear()

-    def add_kv_cache(self, prompt: str, cache: KVCacheType):
+    def add_kv_cache(
+        self,
+        prompt_tokens: mx.array,
+        cache: KVCacheType,
+        ssm_snapshots: list[CacheSnapshot] | None = None,
+    ):
        """Add a new cache entry. Evicts LRU entries if memory is high."""
        self._evict_if_needed()
-        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
-        self.prompts.append(tokenized_prompt)
+        self.prompts.append(prompt_tokens)
        self.caches.append(deepcopy(cache))
+        self._snapshots.append(ssm_snapshots)
        self._access_counter += 1
        self._last_used.append(self._access_counter)
-        logger.info(f"KV cache added: {len(tokenized_prompt)} tokens")
+        logger.info(f"KV cache added: {len(prompt_tokens)} tokens")

    def update_kv_cache(
        self,
        index: int,
-        prompt: str,
+        prompt_tokens: mx.array,
        cache: KVCacheType,
+        snapshots: list[CacheSnapshot] | None,
+        restore_pos: int,
    ):
        """Update an existing cache entry in-place."""
-        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
-        self.prompts[index] = tokenized_prompt
+        old_snapshots = self._snapshots[index]
+        merged: list[CacheSnapshot] = []
+        if old_snapshots:
+            merged = [s for s in old_snapshots if s.token_count <= restore_pos]
+        if snapshots:
+            merged.extend(snapshots)
+
+        self.prompts[index] = prompt_tokens
        self.caches[index] = deepcopy(cache)
+        self._snapshots[index] = merged or None
        self._access_counter += 1
        self._last_used[index] = self._access_counter
-        logger.info(f"KV cache updated (index {index}): {len(tokenized_prompt)} tokens")
+        logger.info(f"KV cache updated (index {index}): {len(prompt_tokens)} tokens")
+
+    def _get_snapshot(
+        self, entry_index: int, target_token_count: int
+    ) -> tuple[int, CacheSnapshot | None]:
+        if not has_non_kv_caches(self.caches[entry_index]):
+            return target_token_count, None
+
+        snapshots = self._snapshots[entry_index]
+        if not snapshots:
+            return 0, None
+
+        snap = _find_nearest_snapshot(snapshots, target_token_count)
+        if snap is not None:
+            return snap.token_count, snap
+
+        return 0, None

    def get_kv_cache(
        self,
        model: Model,
-        prompt: str,
+        prompt_tokens: mx.array,
    ) -> tuple[KVCacheType, mx.array, int | None]:
        """Get KV cache for prompt, returning remaining tokens to prefill.

@@ -79,76 +145,71 @@ class KVPrefixCache:
            - cache: KV cache to use for generation
            - remaining_tokens: tokens that still need prefilling
            - matched_index: index of the matched entry (None if no match)
+
+        For models with SSM layers (which are ArraysCache in mlx), the cache is trimmed to the
+        nearest SSM snapshot position at or before the match point for correctness.
+        Same for rotating KV Cache.
        """
-        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
-        max_length = len(tokenized_prompt)
+        max_length = len(prompt_tokens)

-        best_snapshot_index, best_snapshot_length = None, 0
+        best_index: int | None = None
+        best_length = 0
+        is_exact = False

+        # Find best cache
        for i, cached_prompt in enumerate(self.prompts):
-            length = get_prefix_length(tokenized_prompt, cached_prompt)
-
+            length = get_prefix_length(prompt_tokens, cached_prompt)
+            if length > best_length:
+                best_index, best_length = i, length
            if length == max_length:
-                # Exact match - cached prompt starts with our entire prompt
-                # Trim cache to prompt length - 1, return last token for stream_generate
-                prompt_cache = deepcopy(self.caches[i])
-                cached_length = cache_length(self.caches[i])
-                tokens_to_trim = cached_length - (max_length - 1)
-                if tokens_to_trim > 0:
-                    trim_prompt_cache(cast(list[Any], prompt_cache), tokens_to_trim)
-                self._access_counter += 1
-                self._last_used[i] = self._access_counter
-                logger.info(f"KV cache exact match: {max_length} tokens (instant)")
-                return prompt_cache, tokenized_prompt[-1:], i
+                is_exact = True
+                best_index, best_length = i, length
+                break

-            if length > best_snapshot_length:
-                best_snapshot_index, best_snapshot_length = i, length
+        if best_index is None:
+            return make_kv_cache(model), prompt_tokens, None

-        if best_snapshot_index is not None:
-            new_tokens = max_length - best_snapshot_length
-            logger.info(
-                f"KV cache prefix match: {best_snapshot_length}/{max_length} tokens "
-                f"(reusing {best_snapshot_length}, need to prefill {new_tokens})"
-            )
+        # For exact match: trim to max_length-1 so remaining has the last token
+        # For partial match: trim to best_length, remaining has suffix to prefill
+        # This ensures stream_generate always has at least one token to start with
+        target = (max_length - 1) if is_exact else best_length
+        restore_pos, restore_snap = self._get_snapshot(best_index, target)

-            prompt_cache = deepcopy(self.caches[best_snapshot_index])
+        # No usable snapshot — need fresh cache
+        if restore_snap is None and has_non_kv_caches(self.caches[best_index]):
+            return make_kv_cache(model), prompt_tokens, None

-            # Trim removes tokens from the end, so we trim (cached_length - prefix_length) to keep the prefix
-            cached_length = cache_length(self.caches[best_snapshot_index])
-            tokens_to_trim = cached_length - best_snapshot_length
-            if tokens_to_trim > 0:
-                trim_prompt_cache(cast(list[Any], prompt_cache), tokens_to_trim)
+        prompt_cache = deepcopy(self.caches[best_index])
+        cached_length = cache_length(self.caches[best_index])
+        tokens_to_trim = cached_length - restore_pos
+        if tokens_to_trim > 0:
+            trim_cache(prompt_cache, tokens_to_trim, restore_snap)
+            # Reset cache offset to match trimmed length
+            for c in prompt_cache:
+                if hasattr(c, "offset"):
+                    c.offset = restore_pos

-            self._access_counter += 1
-            self._last_used[best_snapshot_index] = self._access_counter
-            remaining_tokens = tokenized_prompt[best_snapshot_length:]
-            return prompt_cache, remaining_tokens, best_snapshot_index
+        self._access_counter += 1
+        self._last_used[best_index] = self._access_counter
+        remaining = prompt_tokens[restore_pos:]

-        else:
-            prompt_cache = make_kv_cache(model)
-            if len(self.prompts) == 0:
-                logger.info(f"KV cache empty, need to prefill {max_length} tokens")
-            else:
-                logger.info(
-                    f"KV cache no prefix match, need to prefill {max_length} tokens"
-                )
-
-            return prompt_cache, tokenized_prompt, None
+        return prompt_cache, remaining, best_index

    def _evict_if_needed(self):
        """Evict least recently used entries while memory usage is high."""
        if len(self.caches) == 0:
            return

-        # Evict LRU entries until below threshold or only one entry left
+        # Evict LRU entries until below threshold
        while (
-            len(self.caches) > 1
+            len(self.caches) > 0
            and self.get_memory_used_percentage() > _MEMORY_THRESHOLD
        ):
            lru_index = self._last_used.index(min(self._last_used))
            evicted_tokens = len(self.prompts[lru_index])
            self.prompts.pop(lru_index)
            self.caches.pop(lru_index)
+            self._snapshots.pop(lru_index)
            self._last_used.pop(lru_index)
            logger.info(
                f"KV cache evicted LRU entry ({evicted_tokens} tokens) due to memory usage"
@@ -169,6 +230,21 @@ class KVPrefixCache:
        return max_pressure


+def trim_cache(
+    cache: KVCacheType,
+    num_tokens: int,
+    snapshot: CacheSnapshot | None = None,
+) -> None:
+    for i, c in enumerate(cache):
+        if isinstance(c, (ArraysCache, RotatingKVCache)):
+            if snapshot is not None and snapshot.states[i] is not None:
+                cache[i] = deepcopy(snapshot.states[i])  # type: ignore
+            else:
+                c.state = [None] * len(c.state)  # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType]
+        else:
+            c.trim(num_tokens)  # pyright: ignore[reportUnknownMemberType]
+
+
 def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
    """Encode a prompt string to token array.

@@ -177,14 +253,14 @@ def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
    that would corrupt the prompt structure.
    """
    # Chat templates define their own structure - don't add BOS/EOS
-    tokenized_prompt = tokenizer.encode(prompt, add_special_tokens=False)
-    return mx.array(tokenized_prompt)
+    prompt_tokens = tokenizer.encode(prompt, add_special_tokens=False)
+    return mx.array(prompt_tokens)


 def cache_length(cache: KVCacheType) -> int:
    """Get the number of tokens in a KV cache."""
-    # Use .offset attribute which all cache types have (len() not implemented in older QuantizedKVCache)
-    return max(c.offset for c in cache)  # type: ignore
+    # Use .offset attribute which KVCache types have (len() not implemented in older QuantizedKVCache).
+    return max(getattr(c, "offset", 0) for c in cache)


 def get_prefix_length(prompt: mx.array, cached_prompt: mx.array) -> int:
@@ -215,7 +291,7 @@ def make_kv_cache(
    assert hasattr(model, "layers")

    # TODO: Do this for all models
-    if hasattr(model, "make_cache") and isinstance(model, GptOssModel):
+    if hasattr(model, "make_cache"):
        logger.info("Using MLX LM's make cache")
        return model.make_cache()  # type: ignore

--- a/src/exo/worker/engines/mlx/generator/generate.py
+++ b/src/exo/worker/engines/mlx/generator/generate.py
@@ -1,9 +1,10 @@
 import time
-from typing import Any, Callable, Generator, cast, get_args
+from copy import deepcopy
+from typing import Callable, Generator, cast, get_args

 import mlx.core as mx
 from mlx_lm.generate import stream_generate
-from mlx_lm.models.cache import trim_prompt_cache
+from mlx_lm.models.cache import ArraysCache, RotatingKVCache
 from mlx_lm.sample_utils import make_sampler
 from mlx_lm.tokenizer_utils import TokenizerWrapper

@@ -23,7 +24,14 @@ from exo.shared.types.worker.runner_response import (
    GenerationResponse,
 )
 from exo.worker.engines.mlx import Model
-from exo.worker.engines.mlx.cache import KVPrefixCache, encode_prompt, make_kv_cache
+from exo.worker.engines.mlx.cache import (
+    CacheSnapshot,
+    KVPrefixCache,
+    encode_prompt,
+    has_non_kv_caches,
+    make_kv_cache,
+    snapshot_ssm_states,
+)
 from exo.worker.engines.mlx.constants import (
    DEFAULT_TOP_LOGPROBS,
    KV_BITS,
@@ -32,6 +40,7 @@ from exo.worker.engines.mlx.constants import (
 )
 from exo.worker.engines.mlx.utils_mlx import (
    apply_chat_template,
+    fix_unmatched_think_end_tokens,
    mx_barrier,
 )
 from exo.worker.runner.bootstrap import logger
@@ -47,7 +56,7 @@ def prefill(
    sampler: Callable[[mx.array], mx.array],
    prompt_tokens: mx.array,
    cache: KVCacheType,
-) -> tuple[float, int]:
+) -> tuple[float, int, list[CacheSnapshot]]:
    """Prefill the KV cache with prompt tokens.

    This runs the model over the prompt tokens to populate the cache,
@@ -58,17 +67,21 @@ def prefill(
    """
    num_tokens = len(prompt_tokens)
    if num_tokens == 0:
-        return 0.0, 0
+        return 0.0, 0, []

    logger.debug(f"Prefilling {num_tokens} tokens...")
    start_time = time.perf_counter()
+    has_ssm = has_non_kv_caches(cache)
+    snapshots: list[CacheSnapshot] = []

    def progress_callback(processed: int, total: int) -> None:
-        elapsed = time.time() - start_time
+        elapsed = time.perf_counter() - start_time
        tok_per_sec = processed / elapsed if elapsed > 0 else 0
        logger.debug(
            f"Prefill progress: {processed}/{total} tokens ({tok_per_sec:.1f} tok/s)"
        )
+        if has_ssm:
+            snapshots.append(snapshot_ssm_states(cache))

    # Use max_tokens=1 because max_tokens=0 does not work.
    # We just throw away the generated token - we only care about filling the cache
@@ -85,7 +98,18 @@ def prefill(
        prompt_progress_callback=progress_callback,
    ):
        break  # Stop after first iteration - cache is now filled
-    trim_prompt_cache(cast(list[Any], cache), 1)
+
+    # stream_generate added 1 extra generated token to the cache, so we should trim it.
+    # Because of needing to roll back arrays cache, we will generate on 2 tokens so trim 1 more.
+    pre_gen = deepcopy(snapshots[-2]) if has_ssm else None
+    for i, c in enumerate(cache):
+        if has_ssm and isinstance(c, (ArraysCache, RotatingKVCache)):
+            assert pre_gen is not None
+            if pre_gen.states[i] is not None:
+                cache[i] = deepcopy(pre_gen.states[i])  # type: ignore
+        else:
+            assert not isinstance(c, (ArraysCache, RotatingKVCache))
+            c.trim(2)  # pyright: ignore[reportUnknownMemberType]

    elapsed = time.perf_counter() - start_time
    tokens_per_sec = num_tokens / elapsed if elapsed > 0 else 0.0
@@ -93,12 +117,14 @@ def prefill(
        f"Prefill complete: {num_tokens} tokens in {elapsed:.2f}s "
        f"({tokens_per_sec:.1f} tok/s)"
    )
-    return tokens_per_sec, num_tokens
+    # Exclude the last snapshot
+    return tokens_per_sec, num_tokens, snapshots[:-1] if snapshots else []


 def warmup_inference(
    model: Model,
    tokenizer: TokenizerWrapper,
+    group: mx.distributed.Group | None = None,
 ) -> int:
    content = "Prompt to warm up the inference engine. Repeat this."

@@ -117,7 +143,7 @@ def warmup_inference(
    )

    # Use a default sampler for warmup
-    sampler = make_sampler(temp=0.7)
+    sampler = make_sampler(temp=0.0)

    logger.info("Generating warmup tokens")
    for _r in stream_generate(
@@ -136,9 +162,7 @@ def warmup_inference(

    logger.info("Generated ALL warmup tokens")

-    # TODO: Do we want an mx_barrier?
-    #  At least this version is actively incorrect, as it should use mx_barrier(group)
-    mx_barrier()
+    mx_barrier(group)

    return tokens_generated

@@ -221,11 +245,17 @@ def mlx_generate(
    task: TextGenerationTaskParams,
    prompt: str,
    kv_prefix_cache: KVPrefixCache | None = None,
+    group: mx.distributed.Group | None = None,
 ) -> Generator[GenerationResponse]:
    # Ensure that generation stats only contains peak memory for this generation
    mx.reset_peak_memory()
-    if task.seed is not None:
-        mx.random.seed(task.seed)
+    # TODO: Randomise task seed and set in taskparams, instead of hard coding as 42.
+    seed = task.seed or 42
+    mx.random.seed(seed)
+
+    # Encode prompt once at the top and fix unmatched think tags
+    all_prompt_tokens = encode_prompt(tokenizer, prompt)
+    all_prompt_tokens = fix_unmatched_think_end_tokens(all_prompt_tokens, tokenizer)

    # Do not use the prefix cache if we are trying to do benchmarks.
    is_bench = task.bench
@@ -237,13 +267,16 @@ def mlx_generate(
    matched_index: int | None = None
    if kv_prefix_cache is None:
        caches = make_kv_cache(model=model)
-        prompt_tokens = encode_prompt(tokenizer, prompt)
+        prompt_tokens = all_prompt_tokens
    else:
        caches, prompt_tokens, matched_index = kv_prefix_cache.get_kv_cache(
-            model, prompt
+            model, all_prompt_tokens
        )
-        all_prompt_tokens = encode_prompt(tokenizer, prompt)
        prefix_hit_length = len(all_prompt_tokens) - len(prompt_tokens)
+        if prefix_hit_length > 0:
+            logger.info(
+                f"KV cache hit: {prefix_hit_length}/{len(all_prompt_tokens)} tokens cached ({100 * prefix_hit_length / len(all_prompt_tokens):.1f}%)"
+            )

    logits_processors: list[Callable[[mx.array, mx.array], mx.array]] = []
    if is_bench:
@@ -266,12 +299,17 @@ def mlx_generate(
    max_stop_len = max((len(s) for s in stop_sequences), default=0)

    # Prefill cache with all tokens except the last one
-    prefill_tps, prefill_tokens = prefill(
-        model, tokenizer, sampler, prompt_tokens[:-1], caches
+    prefill_tps, prefill_tokens, ssm_snapshots_list = prefill(
+        model,
+        tokenizer,
+        sampler,
+        prompt_tokens[:-1],
+        caches,
    )
+    cache_snapshots: list[CacheSnapshot] | None = ssm_snapshots_list or None

    # stream_generate starts from the last token
-    last_token = prompt_tokens[-1:]
+    last_token = prompt_tokens[-2:]

    max_tokens = task.max_output_tokens or MAX_TOKENS
    accumulated_text = ""
@@ -299,7 +337,6 @@ def mlx_generate(
        start=1,
    ):
        generated_text_parts.append(out.text)
-        logger.info(out.text)
        accumulated_text += out.text

        if think_start is not None and out.text == think_start:
@@ -367,16 +404,6 @@ def mlx_generate(
                selected_token=out.token,
            )

-        yield GenerationResponse(
-            text=text,
-            token=out.token,
-            logprob=logprob,
-            top_logprobs=top_logprobs,
-            finish_reason=finish_reason,
-            stats=stats,
-            usage=usage,
-        )
-
        if is_done:
            # Log generation stats
            generation_elapsed = time.perf_counter() - generation_start_time
@@ -390,18 +417,44 @@ def mlx_generate(
                f"{generation_tps:.1f} tok/s"
            )
            if kv_prefix_cache is not None:
-                full_prompt = prompt + "".join(generated_text_parts)
+                generated_tokens_array = mx.array(
+                    tokenizer.encode(
+                        "".join(generated_text_parts), add_special_tokens=False
+                    )
+                )
+                full_prompt_tokens = mx.concatenate(
+                    [all_prompt_tokens, generated_tokens_array]
+                )
                if (
                    matched_index is not None
                    and prefix_hit_length >= _MIN_PREFIX_HIT_TO_UPDATE
                ):
-                    kv_prefix_cache.update_kv_cache(matched_index, full_prompt, caches)
+                    kv_prefix_cache.update_kv_cache(
+                        matched_index,
+                        full_prompt_tokens,
+                        caches,
+                        cache_snapshots,
+                        restore_pos=prefix_hit_length,
+                    )
                else:
-                    kv_prefix_cache.add_kv_cache(full_prompt, caches)
+                    kv_prefix_cache.add_kv_cache(
+                        full_prompt_tokens, caches, cache_snapshots
+                    )
+
+        yield GenerationResponse(
+            text=text,
+            token=out.token,
+            logprob=logprob,
+            top_logprobs=top_logprobs,
+            finish_reason=finish_reason,
+            stats=stats,
+            usage=usage,
+        )
+
+        if is_done:
+            mx_barrier(group)
            break

        # Limit accumulated_text to what's needed for stop sequence detection
        if max_stop_len > 0 and len(accumulated_text) > max_stop_len:
            accumulated_text = accumulated_text[-max_stop_len:]
-
-        # TODO: Do we want an mx_barrier?
--- a/src/exo/worker/engines/mlx/utils_mlx.py
+++ b/src/exo/worker/engines/mlx/utils_mlx.py
@@ -490,6 +490,30 @@ def detect_thinking_prompt_suffix(prompt: str, tokenizer: TokenizerWrapper) -> b
    return think_token is not None and prompt.rstrip().endswith(think_token)


+def fix_unmatched_think_end_tokens(
+    tokens: mx.array, tokenizer: TokenizerWrapper
+) -> mx.array:
+    if not tokenizer.has_thinking:
+        return tokens
+    assert tokenizer.think_start_id
+    assert tokenizer.think_end_id
+    think_start_id: int = tokenizer.think_start_id
+    think_end_id: int = tokenizer.think_end_id
+    token_list: list[int] = cast(list[int], tokens.tolist())
+    result: list[int] = []
+    depth = 0
+    for token in token_list:
+        if token == think_start_id:
+            depth += 1
+        elif token == think_end_id:
+            if depth == 0:
+                result.append(think_start_id)
+            else:
+                depth -= 1
+        result.append(token)
+    return mx.array(result)
+
+
 class NullKVCache(KVCache):
    """
    A KVCache that pretends to exist but holds zero tokens.
--- a/src/exo/worker/main.py
+++ b/src/exo/worker/main.py
@@ -336,7 +336,7 @@ class Worker:
                    session=self.session_id,
                    event=event,
                )
-                logger.trace(f"Worker published event {idx}: {str(event)[:100]}")
+                logger.debug(f"Worker published event {idx}: {str(event)[:100]}")
                await self.local_event_sender.send(fe)
                self.out_for_delivery[event.event_id] = fe

--- a/src/exo/worker/runner/runner.py
+++ b/src/exo/worker/runner/runner.py
@@ -193,7 +193,7 @@ def main(
                        logger.info(
                            f"model has_tool_calling={tokenizer.has_tool_calling}"
                        )
-                        kv_prefix_cache = KVPrefixCache(tokenizer, group)
+                        kv_prefix_cache = KVPrefixCache(group)

                    elif (
                        ModelTask.TextToImage in shard_metadata.model_card.tasks
@@ -226,6 +226,7 @@ def main(
                        toks = warmup_inference(
                            model=model,
                            tokenizer=tokenizer,
+                            group=group,
                            # kv_prefix_cache=kv_prefix_cache,  # supply for warmup-time prefix caching
                        )
                        logger.info(f"warmed up by generating {toks} tokens")
@@ -274,6 +275,7 @@ def main(
                            task=task_params,
                            prompt=prompt,
                            kv_prefix_cache=kv_prefix_cache,
+                            group=group,
                        )

                        # For other thinking models (GLM, etc.), check if we need to
@@ -627,7 +629,7 @@ def parse_thinking_models(
            yield response.model_copy(
                update={
                    "text": tokenizer.think_start,
-                    "token": tokenizer.think_start_id,  # type: ignore
+                    "token": tokenizer.think_start_id,
                }
            )
        yield response
--- a/src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py
+++ b/src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py
@@ -88,12 +88,12 @@ class TestKVPrefix:
        return tokenizer

    def test_starts_empty(self, mock_tokenizer):
-        cache = KVPrefixCache(mock_tokenizer)
+        cache = KVPrefixCache()
        assert len(cache.prompts) == 0
        assert len(cache.caches) == 0

    def test_clear_empties_cache(self, mock_tokenizer):
-        cache = KVPrefixCache(mock_tokenizer)
+        cache = KVPrefixCache()
        cache.prompts.append(mx.array([1, 2, 3]))
        cache.caches.append([KVCache()])
        cache.clear()
@@ -101,7 +101,7 @@ class TestKVPrefix:
        assert len(cache.caches) == 0

    def test_clear_on_empty_cache(self, mock_tokenizer):
-        cache = KVPrefixCache(mock_tokenizer)
+        cache = KVPrefixCache()
        cache.clear()
        assert len(cache.prompts) == 0

@@ -142,10 +142,12 @@ class TestKVPrefixCacheWithModel:
        tokens = encode_prompt(tokenizer, prompt)
        cache = make_kv_cache(model)

-        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+        _, _, snapshots = prefill(model, tokenizer, make_sampler(0.0), tokens, cache)

-        # Cache should now hold the prompt tokens
-        assert cache_length(cache) == len(tokens)
+        # Cache should now hold the prompt tokens minus one
+        assert cache_length(cache) == len(tokens) - 1
+        # Snapshots should be available for models with non-KV caches
+        assert len(snapshots) > 0

    def test_add_and_get_exact_match(self, model_and_tokenizer):
        model, tokenizer = model_and_tokenizer
@@ -159,10 +161,10 @@ class TestKVPrefixCacheWithModel:
        tokens = encode_prompt(tokenizer, prompt)
        cache = make_kv_cache(model)

-        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+        _, _, snapshots = prefill(model, tokenizer, make_sampler(0.0), tokens, cache)

-        kv_prefix_cache = KVPrefixCache(tokenizer)
-        kv_prefix_cache.add_kv_cache(prompt, cache)
+        kv_prefix_cache = KVPrefixCache()
+        kv_prefix_cache.add_kv_cache(tokens, cache, snapshots)

        assert len(kv_prefix_cache.prompts) == 1
        stored_length = cache_length(kv_prefix_cache.caches[0])
@@ -170,7 +172,7 @@ class TestKVPrefixCacheWithModel:

        # Retrieve with same prompt: exact match
        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
-            model, prompt
+            model, tokens
        )
        assert matched_index == 0

@@ -191,10 +193,12 @@ class TestKVPrefixCacheWithModel:
        short_tokens = encode_prompt(tokenizer, short_prompt)
        cache = make_kv_cache(model)

-        prefill(model, tokenizer, make_sampler(0.0), short_tokens, cache)
+        _, _, snapshots = prefill(
+            model, tokenizer, make_sampler(0.0), short_tokens, cache
+        )

-        kv_prefix_cache = KVPrefixCache(tokenizer)
-        kv_prefix_cache.add_kv_cache(short_prompt, cache)
+        kv_prefix_cache = KVPrefixCache()
+        kv_prefix_cache.add_kv_cache(short_tokens, cache, snapshots)

        # Query with longer prompt that shares the chat template prefix
        long_task = TextGenerationTaskParams(
@@ -212,13 +216,12 @@ class TestKVPrefixCacheWithModel:
        )

        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
-            model, long_prompt
+            model, long_tokens
        )
        assert matched_index == 0

-        # remaining_tokens should be the suffix after the shared prefix
-        assert len(remaining_tokens) == len(long_tokens) - expected_prefix
-        assert mx.array_equal(remaining_tokens, long_tokens[expected_prefix:])
+        # remaining_tokens covers from snapshot restore position to end
+        assert len(remaining_tokens) >= len(long_tokens) - expected_prefix

    def test_stored_cache_not_mutated_after_get_and_generation(
        self, model_and_tokenizer
@@ -235,15 +238,15 @@ class TestKVPrefixCacheWithModel:
        tokens = encode_prompt(tokenizer, prompt)
        cache = make_kv_cache(model)

-        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+        _, _, snapshots = prefill(model, tokenizer, make_sampler(0.0), tokens, cache)

-        kv_prefix_cache = KVPrefixCache(tokenizer)
-        kv_prefix_cache.add_kv_cache(prompt, cache)
+        kv_prefix_cache = KVPrefixCache()
+        kv_prefix_cache.add_kv_cache(tokens, cache, snapshots)

        stored_length = cache_length(kv_prefix_cache.caches[0])

        # Get cache and mutate it (simulating what generation does)
-        result_cache, _, matched_index = kv_prefix_cache.get_kv_cache(model, prompt)
+        result_cache, _, matched_index = kv_prefix_cache.get_kv_cache(model, tokens)
        assert matched_index == 0

        # Simulate generation: feed many additional tokens through the cache
@@ -273,15 +276,15 @@ class TestKVPrefixCacheWithModel:
        tokens = encode_prompt(tokenizer, prompt)
        cache = make_kv_cache(model)

-        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+        _, _, snapshots = prefill(model, tokenizer, make_sampler(0.0), tokens, cache)

-        kv_prefix_cache = KVPrefixCache(tokenizer)
-        kv_prefix_cache.add_kv_cache(prompt, cache)
+        kv_prefix_cache = KVPrefixCache()
+        kv_prefix_cache.add_kv_cache(tokens, cache, snapshots)

        stored_length = cache_length(kv_prefix_cache.caches[0])

        for i in range(3):
-            result_cache, _, _ = kv_prefix_cache.get_kv_cache(model, prompt)
+            result_cache, _, _ = kv_prefix_cache.get_kv_cache(model, tokens)

            head_dim = result_cache[0].keys.shape[-1]
            num_heads = result_cache[0].keys.shape[1]
@@ -298,7 +301,7 @@ class TestKVPrefixCacheWithModel:
        """mlx_generate should save the cache after generation completes."""
        model, tokenizer = model_and_tokenizer

-        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache = KVPrefixCache()
        task = TextGenerationTaskParams(
            model=DEFAULT_GPT_OSS_MODEL_ID,
            input=[InputMessage(role="user", content="Hello")],
@@ -328,7 +331,7 @@ class TestKVPrefixCacheWithModel:
        """Second mlx_generate call with same prompt should get a prefix hit from stored cache."""
        model, tokenizer = model_and_tokenizer

-        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache = KVPrefixCache()
        task = TextGenerationTaskParams(
            model=DEFAULT_GPT_OSS_MODEL_ID,
            input=[InputMessage(role="user", content="Reuse test")],
@@ -352,20 +355,20 @@ class TestKVPrefixCacheWithModel:
        # Second call should find a prefix match (the stored cache contains
        # prompt + generated tokens, which shares the prompt prefix)
        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
-            model, prompt
+            model, prompt_tokens
        )
        # The stored cache is longer than the prompt (it includes generated tokens),
        # so this is a prefix match where our prompt is fully contained
        assert matched_index == 0
-        # Exact match: remaining_tokens is just the last token
-        assert len(remaining_tokens) == 1
-        assert mx.array_equal(remaining_tokens, prompt_tokens[-1:])
+        # Exact match: remaining_tokens is just the last token and the one before
+        assert len(remaining_tokens) == 2
+        assert mx.array_equal(remaining_tokens, prompt_tokens[-2:])

    def test_mlx_generate_long_prompt_updates_cache_in_place(self, model_and_tokenizer):
        """With a prompt > 1000 tokens, second generation should update the cache entry in-place."""
        model, tokenizer = model_and_tokenizer

-        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache = KVPrefixCache()

        # Build a long user message (> 1000 tokens) to exceed _MIN_PREFIX_HIT_TO_UPDATE
        base_text = "The quick brown fox jumps over the lazy dog. "
@@ -444,7 +447,7 @@ class TestKVPrefixCacheWithModel:
        """After mlx_generate saves a cache, a second generation must not corrupt the stored copy."""
        model, tokenizer = model_and_tokenizer

-        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache = KVPrefixCache()
        task = TextGenerationTaskParams(
            model=DEFAULT_GPT_OSS_MODEL_ID,
            input=[InputMessage(role="user", content="Immutable test")],
@@ -481,7 +484,7 @@ class TestKVPrefixCacheWithModel:
        """Under memory pressure, adding a new cache entry evicts the least recently used one."""
        model, tokenizer = model_and_tokenizer

-        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache = KVPrefixCache()

        # Add three cache entries with different prompts
        prompts = ["First entry", "Second entry", "Third entry"]
@@ -495,7 +498,7 @@ class TestKVPrefixCacheWithModel:
            tokens = encode_prompt(tokenizer, prompt)
            cache = make_kv_cache(model)
            prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
-            kv_prefix_cache.add_kv_cache(prompt, cache)
+            kv_prefix_cache.add_kv_cache(tokens, cache)
            # Stagger _last_used so LRU order is deterministic
            kv_prefix_cache._last_used[i] = float(i)

@@ -505,19 +508,10 @@ class TestKVPrefixCacheWithModel:
        kv_prefix_cache._last_used[2] = 100.0
        # Entry 0 (_last_used=0.0) is LRU, entry 1 (_last_used=1.0) is next

-        # Simulate memory pressure: active memory exceeds threshold
-        fake_limit = 1000
-        fake_active = int(fake_limit * 0.90)  # Above _MEMORY_THRESHOLD (0.85)
-
-        with (
-            patch(
-                "exo.worker.engines.mlx.cache.mx.metal.get_active_memory",
-                return_value=fake_active,
-            ),
-            patch(
-                "exo.worker.engines.mlx.cache.mx.metal.device_info",
-                return_value={"max_recommended_working_set_size": fake_limit},
-            ),
+        # Simulate memory pressure: return usage above _MEMORY_THRESHOLD (0.9)
+        with patch(
+            "exo.worker.engines.mlx.cache.get_memory_used_percentage",
+            return_value=0.95,
        ):
            # Trigger eviction by adding a new entry
            task = TextGenerationTaskParams(
@@ -529,14 +523,11 @@ class TestKVPrefixCacheWithModel:
            tokens = encode_prompt(tokenizer, prompt)
            cache = make_kv_cache(model)
            prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
-            kv_prefix_cache.add_kv_cache(prompt, cache)
+            kv_prefix_cache.add_kv_cache(tokens, cache)

        # LRU entries should have been evicted (entries 0, 1, 2 in order of _last_used)
        # Since fake_active stays above threshold after each eviction (we don't change it),
        # all old entries get evicted, leaving only the newly added one
        assert len(kv_prefix_cache.prompts) == 1
        # The surviving entry should be the newly added one
-        new_tokens = encode_prompt(tokenizer, prompt)
-        assert get_prefix_length(kv_prefix_cache.prompts[0], new_tokens) == len(
-            new_tokens
-        )
+        assert get_prefix_length(kv_prefix_cache.prompts[0], tokens) == len(tokens)
--- a/src/exo/worker/tests/unittests/test_mlx/test_tokenizers.py
+++ b/src/exo/worker/tests/unittests/test_mlx/test_tokenizers.py
@@ -34,6 +34,7 @@ TOKENIZER_FILE_PATTERNS = [
    "added_tokens.json",
    "tokenizer.model",
    "tokenization_*.py",  # Custom tokenizer implementations
+    "tool_declaration_ts.py",  # Dependency of tokenization_kimi.py
 ]


--- a/tests/auto_bench.sh
+++ b/tests/auto_bench.sh
@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+
+[ $# -lt 1 ] && {
+  echo "Usage: $0 host1 [host2 ...]"
+  exit 1
+}
+
+[ -z "$(git status --porcelain)" ] || {
+  echo "Uncommitted changes"
+  exit 1
+}
+
+commit=$(git rev-parse HEAD)
+git fetch -q origin
+git branch -r --contains "$commit" | grep -qE '^\s*origin/' || {
+  echo "Not pushed to origin"
+  exit 1
+}
+hosts=("$@")
+cleanup() {
+  for host in "${hosts[@]}"; do
+    ssh -T -o BatchMode=yes "$host@$host" "pkill -f bin/exo" &
+  done
+  sleep 1
+  jobs -pr | xargs -r kill 2>/dev/null || true
+}
+trap 'cleanup' EXIT INT TERM
+
+for host; do
+  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
+    "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix build github:exo-explore/exo/$commit" &
+done
+wait
+for host; do
+  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
+    "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix run github:exo-explore/exo/$commit" &>/dev/null &
+done
+
+for host; do
+  echo "Waiting for $host..." 1>&2
+  until curl -sf "http://$host:52415/models" &>/dev/null; do sleep 1; done
+done
+
+echo "Waiting 30s for cluster setup" 1>&2
+sleep 30
+echo "EXO loaded" 1>&2
+bench_runner="${hosts[0]}"
+mkdir -p "./bench/$commit"
+nix run .#exo-get-all-models-on-cluster -- "$bench_runner" | while IFS= read -r model; do
+  echo "running bench for $model" 1>&2
+  ssh -Tn -o BatchMode=yes -o ServerAliveInterval=30 "$bench_runner@$bench_runner" "/nix/var/nix/profiles/default/bin/nix run github:exo-explore/exo/$commit#exo-bench -- --model $model --pp 128 4096 --tg 128 --stdout --skip-tensor-ring" >>"./bench/$commit/${model//\//--}.json"
+  echo
+done
--- a/tests/get_all_models_on_cluster.py
+++ b/tests/get_all_models_on_cluster.py
@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+# pyright: reportAny=false
+import json
+import subprocess
+import sys
+from typing import Any, cast
+from urllib.request import urlopen
+
+h = sys.argv[1] if len(sys.argv) > 1 else sys.exit(f"USAGE: {sys.argv[0]} host")
+ts = subprocess.run(
+    ["tailscale", "status"], check=True, text=True, capture_output=True
+).stdout.splitlines()
+ip = next(
+    (sl[0] for line in ts if len(sl := line.split()) >= 2 if sl[1] == h), None
+) or sys.exit(f"{h} not found in tailscale")
+with urlopen(f"http://{ip}:52415/state", timeout=5) as r:
+    data = json.loads(r.read()).get("downloads", {})
+
+
+def mid(x: dict[str, Any]) -> str | None:
+    for k in (
+        "DownloadCompleted",
+        "shardMetadata",
+        "PipelineShardMetadata",
+        "modelCard",
+        "modelId",
+    ):
+        x = x.get(k, {})
+    return cast(str | None, x if x != {} else None)
+
+
+common = set[str].intersection(
+    *[{m for d in nid if (m := mid(d))} for nid in data.values()]
+)
+for c in common:
+    print(c)
--- a/tests/run_exo_on.sh
+++ b/tests/run_exo_on.sh
@@ -35,7 +35,7 @@ i=0
 for host; do
  colour=${colours[i++ % 4]}
  ssh -T -o BatchMode=yes -o ServerAliveInterval=30 "$host@$host" \
-    "/nix/var/nix/profiles/default/bin/nix run github:exo-explore/exo/$commit" |&
+    "EXO_LIBP2P_NAMESPACE=$commit /nix/var/nix/profiles/default/bin/nix run github:exo-explore/exo/$commit" |&
    awk -v p="${colour}[${host}]${reset}" '{ print p $0; fflush() }' &
 done

--- a/uv.lock
+++ b/uv.lock
Author	SHA1	Message	Date
Evan	3cc994c19a	cleaning up the todos	2026-02-06 15:38:56 +00:00
Evan Quiney	9b5cae3db6	auto bench (#1405 ) runs exo_bench remotely with some nice git QoL ## usage run tests/auto_bench.sh host1 [host2] exo bench will be run on those hosts and its output saved to bench/commit_hash/*.json on all models currently downloaded	2026-02-06 15:35:46 +00:00
Jake Hillion	cf7201f91e	pyproject: set minimum uv version The uv.lock is churning constantly as different UV versions bounce it between revisions. This is made worse by GitHub automatically hiding the uv.lock changes, meaning it's hard to notice when this went wrong. Set a minimum version for `uv` in pyproject.toml to fix this. I tried quite a few versions (not all) and found 0.8.6 sets the revision to 3, which I believe is the latest. This is from August 2025 so has been around for a while. Test plan: ``` jake@maverick:/data/users/jake/repos/exo/ > git checkout main uv.lock jake@maverick:/data/users/jake/repos/exo/ > nix shell github:nixos/nixpkgs/3dce7f4a77812afd69efcbfe15e5223f98c5c69e#uv --command sh -c 'uv add pip --frozen && uv lock && uv remove pip --frozen && uv lock && uv --version' Resolved 140 packages in 147ms Added pip v26.0.1 Resolved 139 packages in 48ms Removed pip v26.0.1 uv 0.8.6 ```	2026-02-06 15:28:10 +00:00
rltakashige	b315035ae0	Add minimax and fix qwen sharding strategies (#1318 ) ## Motivation MiniMax tensor sharding does not provide equivalent outputs to running it as a single node because RMSNorm weights cannot be split without affecting the output. Qwen3Next sharding was broken, and something with Qwen3MoE was likely changed upstream, as several variables no longer exist. This also ballooned into fixing prefix caching for non-standard models as Qwen3Next was behaving weirdly. ## Changes <!-- Describe what you changed in detail --> ## Why It Works <!-- Explain why your approach solves the problem --> ## Test Plan ### Manual Testing Worked for a 8 hour long eval at the same performance and a more similar completion/reasoning token distribution. --------- Co-authored-by: Alex Cheema <41707476+AlexCheema@users.noreply.github.com> Co-authored-by: Alex Cheema <alexcheema123@gmail.com> Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com> Co-authored-by: Evan <evanev7@gmail.com>	2026-02-06 13:26:59 +00:00
rltakashige	c8dbbee27b	skip tensor ring on bench (#1403 ) ## Motivation <!-- Why is this change needed? What problem does it solve? --> <!-- If it fixes an open issue, please link to the issue here --> ## Changes <!-- Describe what you changed in detail --> ## Why It Works <!-- Explain why your approach solves the problem --> ## Test Plan ### Manual Testing <!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB, connected via Thunderbolt 4) --> <!-- What you did: --> <!-- - --> ### Automated Testing <!-- Describe changes to automated tests, or how existing tests cover this change --> <!-- - -->	2026-02-06 13:06:59 +00:00
rltakashige	f0107e9670	Fix offline no cache (#1402 ) ## Motivation In offline mode, exo complains if there is no caches directory, even if the files are there. ## Changes Check safetensors index and the directory structure to build caches directory. ## Test Plan ### Manual Testing <img width="2338" height="1102" alt="image" src="https://github.com/user-attachments/assets/ad769911-399b-4fca-ac80-aeaa046af06b" /> <img width="656" height="1668" alt="image" src="https://github.com/user-attachments/assets/6080986c-3904-4600-a340-8c70f1b33266" />	2026-02-06 12:57:01 +00:00
Hunter Bown	9f502793c1	fix: retry downloads on transient errors instead of breaking (#1398 ) ## Motivation `download_file_with_retry()` has a `break` in the generic exception handler that exits the retry loop after the first transient failure. This means network timeouts, connection resets, and server errors all cause an immediate download failure — the two remaining retry attempts never run. ## Changes download_utils.py: Replaced `break` with logging and exponential backoff in the generic exception handler, matching the existing rate-limit handler behavior. Before: ```python except Exception as e: on_connection_lost() if attempt == n_attempts - 1: raise e break # exits loop immediately ``` After: ```python except Exception as e: on_connection_lost() if attempt == n_attempts - 1: raise e logger.error(f"Download error on attempt {attempt + 1}/{n_attempts} ...") logger.error(traceback.format_exc()) await asyncio.sleep(2.0**attempt) ``` ## Why It Works The `break` statement was bypassing the retry mechanism entirely. Replacing it with the same log-and-backoff pattern used by the `HuggingFaceRateLimitError` handler means all 3 attempts are actually used before giving up. The exponential backoff (1s, 2s) gives transient issues time to resolve between attempts. ## Test Plan ### Manual Testing - Downloads that hit transient network errors now retry instead of failing immediately ### Automated Testing - `uv run basedpyright` — 0 errors - `uv run ruff check` — passes - `uv run pytest src/exo/download/tests/ -v` — 11 tests pass --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: rltakashige <rl.takashige@gmail.com>	2026-02-06 11:51:54 +00:00
Evan Quiney	c8371349d5	add scripts (#1401 ) allow running exo-bench and the headless runner from nix	2026-02-06 11:06:40 +00:00
Evan Quiney	6b907398a4	cancel downloads for deleted instances (#1393 ) after deleting an instance, if a given (node_id, model_id) pair doesn't exist in the left over instances, cancel the download of model_id on node_id.	2026-02-05 18:16:43 +00:00