aaa

## 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 -->
<!-- - -->

dashboard/package-lock.json

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

pyproject.toml

@@ -6,8 +6,6 @@ readme = "README.md"
 requires-python = ">=3.13"
 dependencies = [
     "aiofiles>=24.1.0",
-    "aiohttp>=3.12.14",
-    "types-aiofiles>=24.1.0.20250708",
     "pydantic>=2.11.7",
     "fastapi>=0.116.1",
     "filelock>=3.18.0",
@@ -17,8 +15,8 @@ dependencies = [
     "loguru>=0.7.3",
     "exo_pyo3_bindings", # rust bindings
     "anyio==4.11.0",
-    "mlx==0.30.1; sys_platform == 'darwin'",
-    "mlx[cpu]==0.30.1; sys_platform == 'linux'",
+    "mlx==0.30.3; sys_platform == 'darwin'",
+    "mlx[cpu]==0.30.3; sys_platform == 'linux'",
     "mlx-lm @ git+https://github.com/AlexCheema/mlx-lm.git@fix-transformers-5.0.0rc2",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",

src/exo/master/tests/test_placement.py

@@ -276,9 +276,7 @@ def test_placement_selects_leaf_nodes(
     # arrange
     topology = Topology()
 
-    # Model requires more than any single node but fits within a 3-node cycle
-    model_card.storage_size.in_bytes = 1500
-    model_card.n_layers = 12
+    model_card.storage_size = Memory.from_bytes(1000)
 
     node_id_a = NodeId()
     node_id_b = NodeId()

src/exo/worker/download/download_utils.py

@@ -8,13 +8,13 @@ import traceback
 from collections.abc import Awaitable
 from datetime import timedelta
 from pathlib import Path
-from typing import Callable, Literal
+from typing import Callable, Literal, cast
 from urllib.parse import urljoin
 
 import aiofiles
 import aiofiles.os as aios
-import aiohttp
 import certifi
+import httpx
 from loguru import logger
 from pydantic import (
     BaseModel,
@@ -208,23 +208,22 @@ async def _fetch_file_list(
     headers = await get_download_headers()
     async with (
         create_http_session(timeout_profile="short") as session,
-        session.get(url, headers=headers) as response,
     ):
-        if response.status == 200:
-            data_json = await response.text()
-            data = TypeAdapter(list[FileListEntry]).validate_json(data_json)
-            files: list[FileListEntry] = []
-            for item in data:
-                if item.type == "file":
-                    files.append(FileListEntry.model_validate(item))
-                elif item.type == "directory" and recursive:
-                    subfiles = await _fetch_file_list(
-                        repo_id, revision, item.path, recursive
-                    )
-                    files.extend(subfiles)
-            return files
-        else:
-            raise Exception(f"Failed to fetch file list: {response.status}")
+        response = await session.get(url, headers=headers)
+        if response.status_code != 200:
+            raise Exception(f"Failed to fetch file list: {response.status_code}")
+
+        data = TypeAdapter(list[FileListEntry]).validate_json(response.text)
+        files: list[FileListEntry] = []
+        for item in data:
+            if item.type == "file":
+                files.append(FileListEntry.model_validate(item))
+            elif item.type == "directory" and recursive:
+                subfiles = await _fetch_file_list(
+                    repo_id, revision, item.path, recursive
+                )
+                files.extend(subfiles)
+        return files
 
 
 async def get_download_headers() -> dict[str, str]:
@@ -232,34 +231,28 @@ async def get_download_headers() -> dict[str, str]:
 
 
 def create_http_session(
-    auto_decompress: bool = False,
     timeout_profile: Literal["short", "long"] = "long",
-) -> aiohttp.ClientSession:
+) -> httpx.AsyncClient:
     if timeout_profile == "short":
         total_timeout = 30
         connect_timeout = 10
-        sock_read_timeout = 30
-        sock_connect_timeout = 10
+        read_timeout = 30
     else:
         total_timeout = 1800
         connect_timeout = 60
-        sock_read_timeout = 1800
-        sock_connect_timeout = 60
+        read_timeout = 1800
 
-    ssl_context = ssl.create_default_context(
-        cafile=os.getenv("SSL_CERT_FILE") or certifi.where()
-    )
-    connector = aiohttp.TCPConnector(ssl=ssl_context)
+    ssl_context = ssl.create_default_context(cafile=os.getenv("SSL_CERT_FILE") or certifi.where())
 
-    return aiohttp.ClientSession(
-        auto_decompress=auto_decompress,
-        connector=connector,
+    return httpx.AsyncClient(
+        verify=ssl_context,
         proxy=os.getenv("HTTPS_PROXY") or os.getenv("HTTP_PROXY") or None,
-        timeout=aiohttp.ClientTimeout(
-            total=total_timeout,
+        follow_redirects=True,
+        timeout=httpx.Timeout(
             connect=connect_timeout,
-            sock_read=sock_read_timeout,
-            sock_connect=sock_connect_timeout,
+            read=read_timeout,
+            write=total_timeout,
+            pool=total_timeout,
         ),
     )
 
@@ -286,23 +279,14 @@ async def file_meta(
     headers = await get_download_headers()
     async with (
         create_http_session(timeout_profile="short") as session,
-        session.head(url, headers=headers) as r,
     ):
-        if r.status == 307:
-            # On redirect, only trust Hugging Face's x-linked-* headers.
-            x_linked_size = r.headers.get("x-linked-size")
-            x_linked_etag = r.headers.get("x-linked-etag")
-            if x_linked_size and x_linked_etag:
-                content_length = int(x_linked_size)
-                etag = trim_etag(x_linked_etag)
-                return content_length, etag
-            # Otherwise, follow the redirect to get authoritative size/hash
-            redirected_location = r.headers.get("location")
-            return await file_meta(repo_id, revision, path, redirected_location)
-        content_length = int(
-            r.headers.get("x-linked-size") or r.headers.get("content-length") or 0
+        r = await session.head(url, headers=headers)
+        content_length = cast(
+            str | None,
+            r.headers.get("x-linked-size") or r.headers.get("content-length"),
         )
-        etag = r.headers.get("x-linked-etag") or r.headers.get("etag")
+        content_length = 0 if content_length is None else int(content_length)
+        etag = cast(str | None, r.headers.get("x-linked-etag") or r.headers.get("etag"))
         assert content_length > 0, f"No content length for {url}"
         assert etag is not None, f"No remote hash for {url}"
         etag = trim_etag(etag)
@@ -361,17 +345,17 @@ async def _download_file(
         n_read = resume_byte_pos or 0
         async with (
             create_http_session(timeout_profile="long") as session,
-            session.get(url, headers=headers) as r,
         ):
-            if r.status == 404:
+            r = await session.get(url, headers=headers)
+            if r.status_code == 404:
                 raise FileNotFoundError(f"File not found: {url}")
-            assert r.status in [200, 206], (
-                f"Failed to download {path} from {url}: {r.status}"
+            assert r.status_code in [200, 206], (
+                f"Failed to download {path} from {url}: {r.status_code}"
             )
             async with aiofiles.open(
                 partial_path, "ab" if resume_byte_pos else "wb"
             ) as f:
-                while chunk := await r.content.read(8 * 1024 * 1024):
+                async for chunk in r.aiter_bytes(8 * 1024 * 1024):
                     n_read = n_read + (await f.write(chunk))
                     on_progress(n_read, length, False)
 

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

@@ -566,7 +566,7 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
                 layer.block_sparse_moe.switch_mlp.up_proj
             )
             layer.block_sparse_moe = ShardedQwenMoE(layer.block_sparse_moe)  # pyright: ignore[reportAttributeAccessIssue, reportArgumentType]
-            layer.block_sparse_moe.sharding_group = self.group
+            layer.block_sparse_moe.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
 
         return model
 
@@ -661,7 +661,7 @@ class GptOssShardingStrategy(TensorParallelShardingStrategy):
             self.all_to_sharded_linear_in_place(layer.mlp.experts.up_proj)
 
             layer.mlp = ShardedGptOssMoE(layer.mlp)  # type: ignore
-            layer.mlp.sharding_group = self.group
+            layer.mlp.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
 
         return model
 

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

@@ -169,10 +169,10 @@ def mlx_distributed_init(
 
                 # TODO: update once upstream fixes
                 logger.info(
-                    f"rank {rank} MLX_IBV_DEVICES: {coordination_file} with devices: {jaccl_devices_json}"
+                    f"rank {rank} MLX_JACCL_DEVICES: {coordination_file} with devices: {jaccl_devices_json}"
                 )
                 logger.info(f"rank {rank} MLX_JACCL_COORDINATOR: {jaccl_coordinator}")
-                os.environ["MLX_IBV_DEVICES"] = coordination_file
+                os.environ["MLX_JACCL_DEVICES"] = coordination_file
                 os.environ["MLX_RANK"] = str(rank)
                 os.environ["MLX_JACCL_COORDINATOR"] = jaccl_coordinator
                 group = mx.distributed.init(backend="jaccl", strict=True)
@@ -312,6 +312,9 @@ def get_eos_token_ids_for_model(model_id: str) -> list[int] | None:
     model_id_lower = model_id.lower()
     if "kimi-k2" in model_id_lower:
         return [163586]
+    elif "glm-4.7-flash" in model_id_lower:
+        # 154820: <|endoftext|>, 154827: <|user|>, 154829: <|observation|>
+        return [154820, 154827, 154829]
     elif "glm" in model_id_lower:
         return [151336, 151329, 151338]
     return None

src/exo/worker/main.py

@@ -413,11 +413,6 @@ class Worker:
             )
             for nid in conns:
                 for ip in conns[nid]:
-                    if "127.0.0.1" in ip or "localhost" in ip:
-                        logger.warning(
-                            f"Loopback connection should not happen: {ip=} for {nid=}"
-                        )
-
                     edge = SocketConnection(
                         # nonsense multiaddr
                         sink_multiaddr=Multiaddr(address=f"/ip4/{ip}/tcp/52415")
@@ -438,6 +433,9 @@ class Worker:
             for conn in self.state.topology.out_edges(self.node_id):
                 if not isinstance(conn.edge, SocketConnection):
                     continue
+                # ignore mDNS discovered connections
+                if conn.edge.sink_multiaddr.port != 52415:
+                    continue
                 if (
                     conn.sink not in conns
                     or conn.edge.sink_multiaddr.ip_address

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

@@ -18,6 +18,7 @@ def _check_model_exists() -> bool:
 
 
 pytestmark = [
+    pytest.mark.slow,
     pytest.mark.skipif(
         not _check_model_exists(),
         reason=f"GPT-OSS model not found at {DEFAULT_GPT_OSS_CONFIG.model_path}",

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

@@ -89,6 +89,8 @@ def get_test_models() -> list[tuple[str, ModelCard]]:
 
 TEST_MODELS: list[tuple[str, ModelCard]] = get_test_models()
 
+pytestmark = pytest.mark.slow
+
 
 @pytest.fixture(scope="module")
 def event_loop():