size tests...

the worker plan function had some skipping logic missing, leading to
double-submitting tasks.

.github/actions/typecheck/action.yml

@@ -1,12 +0,0 @@
-name: Type Check
-
-description: "Run type checker"
-
-runs:
-  using: "composite"
-  steps:
-    - name: Run type checker
-      run: |
-        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just sync
-        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just check
-      shell: bash

.github/workflows/pipeline.yml

@@ -26,73 +26,14 @@ jobs:
           name: exo
           authToken: "${{ secrets.CACHIX_AUTH_TOKEN }}"
 
-      - name: Configure git user
-        run: |
-          git config --local user.email "github-actions@users.noreply.github.com"
-          git config --local user.name  "github-actions bot"
-        shell: bash
+      - name: Load nix develop environment
+        run: nix run github:nicknovitski/nix-develop/v1
 
-      - name: Pull LFS files
-        run: |
-          echo "Pulling Git LFS files..."
-          git lfs pull
-        shell: bash
+      - name: Sync dependencies
+        run: uv sync --all-packages
 
-      - name: Setup Nix Environment
-        run: |
-          echo "Checking for nix installation..."
-          
-          # Check if nix binary exists directly
-          if [ -f /nix/var/nix/profiles/default/bin/nix ]; then
-            echo "Found nix binary at /nix/var/nix/profiles/default/bin/nix"
-            export PATH="/nix/var/nix/profiles/default/bin:$PATH"
-            echo "PATH=$PATH" >> $GITHUB_ENV
-            nix --version
-          elif [ -f /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh ]; then
-            echo "Found nix profile script, sourcing..."
-            source /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh
-            nix --version
-          elif command -v nix >/dev/null 2>&1; then
-            echo "Nix already in PATH"
-            nix --version
-          else
-            echo "Nix not found. Debugging info:"
-            echo "Contents of /nix/var/nix/profiles/default/:"
-            ls -la /nix/var/nix/profiles/default/ 2>/dev/null || echo "Directory not found"
-            echo "Contents of /nix/var/nix/profiles/default/bin/:"
-            ls -la /nix/var/nix/profiles/default/bin/ 2>/dev/null || echo "Directory not found"
-            exit 1
-          fi
-        shell: bash
-
-      - name: Configure basedpyright include for local MLX
-        run: |
-          RUNNER_LABELS='${{ toJSON(runner.labels) }}'
-          if echo "$RUNNER_LABELS" | grep -q "local_mlx"; then
-            if [ -d "/Users/Shared/mlx" ]; then
-              echo "Updating [tool.basedpyright].include to use /Users/Shared/mlx"
-              awk '
-                BEGIN { in=0 }
-                /^\[tool\.basedpyright\]/ { in=1; print; next }
-                in && /^\[/ { in=0 }  # next section
-                in && /^[ \t]*include[ \t]*=/ {
-                  print "include = [\"/Users/Shared/mlx\"]"
-                  next
-                }
-                { print }
-              ' pyproject.toml > pyproject.toml.tmp && mv pyproject.toml.tmp pyproject.toml
-
-              echo "New [tool.basedpyright] section:"
-              sed -n '/^\[tool\.basedpyright\]/,/^\[/p' pyproject.toml | sed '$d' || true
-            else
-              echo "local_mlx tag present but /Users/Shared/mlx not found; leaving pyproject unchanged."
-            fi
-          else
-            echo "Runner does not have 'local_mlx' tag; leaving pyproject unchanged."
-          fi
-        shell: bash
-
-      - uses: ./.github/actions/typecheck
+      - name: Run type checker
+        run: uv run basedpyright --project pyproject.toml
 
   nix:
     name: Build and check (${{ matrix.system }})
@@ -191,3 +132,14 @@ jobs:
 
       - name: Run nix flake check
         run: nix flake check
+
+      - name: Run pytest (macOS only)
+        if: runner.os == 'macOS'
+        run: |
+          # Build the test environment (requires relaxed sandbox for uv2nix on macOS)
+          TEST_ENV=$(nix build '.#exo-test-env' --option sandbox relaxed --print-out-paths)
+
+          # Run pytest outside sandbox (needs GPU access for MLX)
+          export HOME="$RUNNER_TEMP"
+          export EXO_TESTS=1
+          $TEST_ENV/bin/python -m pytest src -m "not slow" --import-mode=importlib

app/EXO/EXO.xcodeproj/project.pbxproj

@@ -342,6 +342,8 @@
 				SDKROOT = macosx;
 				SWIFT_ACTIVE_COMPILATION_CONDITIONS = "DEBUG $(inherited)";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+				SWIFT_TREAT_WARNINGS_AS_ERRORS = YES;
+				GCC_TREAT_WARNINGS_AS_ERRORS = YES;
 			};
 			name = Debug;
 		};
@@ -397,6 +399,8 @@
 				MTL_FAST_MATH = YES;
 				SDKROOT = macosx;
 				SWIFT_COMPILATION_MODE = wholemodule;
+				SWIFT_TREAT_WARNINGS_AS_ERRORS = YES;
+				GCC_TREAT_WARNINGS_AS_ERRORS = YES;
 			};
 			name = Release;
 		};

dashboard/package-lock.json

@@ -865,7 +865,6 @@
 			"integrity": "sha512-oH8tXw7EZnie8FdOWYrF7Yn4IKrqTFHhXvl8YxXxbKwTMcD/5NNCryUSEXRk2ZR4ojnub0P8rNrsVGHXWqIDtA==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"@standard-schema/spec": "^1.0.0",
 				"@sveltejs/acorn-typescript": "^1.0.5",
@@ -905,7 +904,6 @@
 			"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",
@@ -1522,7 +1520,6 @@
 			"integrity": "sha512-LCCV0HdSZZZb34qifBsyWlUmok6W7ouER+oQIGBScS8EsZsQbrtFTUrDX4hOl+CS6p7cnNC4td+qrSVGSCTUfQ==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"undici-types": "~6.21.0"
 			}
@@ -1532,7 +1529,6 @@
 			"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"
 			},
@@ -1945,7 +1941,6 @@
 			"integrity": "sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==",
 			"dev": true,
 			"license": "ISC",
-			"peer": true,
 			"engines": {
 				"node": ">=12"
 			}
@@ -2653,7 +2648,6 @@
 			"integrity": "sha512-5gTmgEY/sqK6gFXLIsQNH19lWb4ebPDLA4SdLP7dsWkIXHWlG66oPuVvXSGFPppYZz8ZDZq0dYYrbHfBCVUb1Q==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"engines": {
 				"node": ">=12"
 			},
@@ -2696,7 +2690,6 @@
 			"integrity": "sha512-UOnG6LftzbdaHZcKoPFtOcCKztrQ57WkHDeRD9t/PTQtmT0NHSeWWepj6pS0z/N7+08BHFDQVUrfmfMRcZwbMg==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"bin": {
 				"prettier": "bin/prettier.cjs"
 			},
@@ -2869,7 +2862,6 @@
 			"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",
@@ -3014,7 +3006,6 @@
 			"integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==",
 			"dev": true,
 			"license": "Apache-2.0",
-			"peer": true,
 			"bin": {
 				"tsc": "bin/tsc",
 				"tsserver": "bin/tsserver"
@@ -3036,7 +3027,6 @@
 			"integrity": "sha512-+Oxm7q9hDoLMyJOYfUYBuHQo+dkAloi33apOPP56pzj+vsdJDzr+j1NISE5pyaAuKL4A3UD34qd0lx5+kfKp2g==",
 			"dev": true,
 			"license": "MIT",
-			"peer": true,
 			"dependencies": {
 				"esbuild": "^0.25.0",
 				"fdir": "^6.4.4",

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

@@ -173,6 +173,11 @@ export interface PlacementPreviewResponse {
   previews: PlacementPreview[];
 }
 
+interface ImageApiResponse {
+  created: number;
+  data: Array<{ b64_json?: string; url?: string }>;
+}
+
 interface RawStateResponse {
   topology?: RawTopology;
   instances?: Record<
@@ -2095,107 +2100,137 @@ class AppStore {
         throw new Error(`API error: ${response.status} - ${errorText}`);
       }
 
-      const reader = response.body?.getReader();
-      if (!reader) {
-        throw new Error("No response body");
-      }
+      // Streaming requires both stream=true AND partialImages > 0
+      const isStreaming = params.stream && params.partialImages > 0;
 
-      interface ImageGenerationChunk {
-        data?: { b64_json?: string };
-        format?: string;
-        type?: "partial" | "final";
-        image_index?: number;
-        partial_index?: number;
-        total_partials?: number;
-      }
+      if (!isStreaming) {
+        // Non-streaming: parse JSON response directly
+        const jsonResponse = (await response.json()) as ImageApiResponse;
+        const format = params.outputFormat || "png";
+        const mimeType = `image/${format}`;
 
-      const numImages = params.numImages;
+        const attachments: MessageAttachment[] = jsonResponse.data
+          .filter((img) => img.b64_json)
+          .map((img, index) => ({
+            type: "generated-image" as const,
+            name: `generated-image-${index + 1}.${format}`,
+            preview: `data:${mimeType};base64,${img.b64_json}`,
+            mimeType,
+          }));
 
-      await this.parseSSEStream<ImageGenerationChunk>(
-        reader,
-        targetConversationId,
-        (parsed) => {
-          const imageData = parsed.data?.b64_json;
+        this.updateConversationMessage(
+          targetConversationId,
+          assistantMessage.id,
+          (msg) => {
+            msg.content = "";
+            msg.attachments = attachments;
+          },
+        );
+        this.syncActiveMessagesIfNeeded(targetConversationId);
+      } else {
+        // Streaming mode: use SSE parser
+        const reader = response.body?.getReader();
+        if (!reader) {
+          throw new Error("No response body");
+        }
 
-          if (imageData) {
-            const format = parsed.format || "png";
-            const mimeType = `image/${format}`;
-            const imageIndex = parsed.image_index ?? 0;
+        interface ImageGenerationChunk {
+          data?: { b64_json?: string };
+          format?: string;
+          type?: "partial" | "final";
+          image_index?: number;
+          partial_index?: number;
+          total_partials?: number;
+        }
 
-            if (parsed.type === "partial") {
-              // Update with partial image and progress
-              const partialNum = (parsed.partial_index ?? 0) + 1;
-              const totalPartials = parsed.total_partials ?? 3;
-              const progressText =
-                numImages > 1
-                  ? `Generating image ${imageIndex + 1}/${numImages}... ${partialNum}/${totalPartials}`
-                  : `Generating... ${partialNum}/${totalPartials}`;
+        const numImages = params.numImages;
 
-              const partialAttachment: MessageAttachment = {
-                type: "generated-image",
-                name: `generated-image.${format}`,
-                preview: `data:${mimeType};base64,${imageData}`,
-                mimeType,
-              };
+        await this.parseSSEStream<ImageGenerationChunk>(
+          reader,
+          targetConversationId,
+          (parsed) => {
+            const imageData = parsed.data?.b64_json;
 
-              this.updateConversationMessage(
-                targetConversationId,
-                assistantMessage.id,
-                (msg) => {
-                  msg.content = progressText;
-                  if (imageIndex === 0) {
-                    // First image - safe to replace attachments with partial preview
-                    msg.attachments = [partialAttachment];
-                  } else {
-                    // Subsequent images - keep existing finals, show partial at current position
-                    const existingAttachments = msg.attachments || [];
-                    // Keep only the completed final images (up to current imageIndex)
-                    const finals = existingAttachments.slice(0, imageIndex);
-                    msg.attachments = [...finals, partialAttachment];
-                  }
-                },
-              );
-            } else if (parsed.type === "final") {
-              // Final image - replace partial at this position
-              const newAttachment: MessageAttachment = {
-                type: "generated-image",
-                name: `generated-image-${imageIndex + 1}.${format}`,
-                preview: `data:${mimeType};base64,${imageData}`,
-                mimeType,
-              };
+            if (imageData) {
+              const format = parsed.format || "png";
+              const mimeType = `image/${format}`;
+              const imageIndex = parsed.image_index ?? 0;
 
-              this.updateConversationMessage(
-                targetConversationId,
-                assistantMessage.id,
-                (msg) => {
-                  if (imageIndex === 0) {
-                    // First final image - replace any partial preview
-                    msg.attachments = [newAttachment];
-                  } else {
-                    // Subsequent images - keep previous finals, replace partial at current position
-                    const existingAttachments = msg.attachments || [];
-                    // Slice keeps indices 0 to imageIndex-1 (the previous final images)
-                    const previousFinals = existingAttachments.slice(
-                      0,
-                      imageIndex,
-                    );
-                    msg.attachments = [...previousFinals, newAttachment];
-                  }
+              if (parsed.type === "partial") {
+                // Update with partial image and progress
+                const partialNum = (parsed.partial_index ?? 0) + 1;
+                const totalPartials = parsed.total_partials ?? 3;
+                const progressText =
+                  numImages > 1
+                    ? `Generating image ${imageIndex + 1}/${numImages}... ${partialNum}/${totalPartials}`
+                    : `Generating... ${partialNum}/${totalPartials}`;
 
-                  // Update progress message for multiple images
-                  if (numImages > 1 && imageIndex < numImages - 1) {
-                    msg.content = `Generating image ${imageIndex + 2}/${numImages}...`;
-                  } else {
-                    msg.content = "";
-                  }
-                },
-              );
+                const partialAttachment: MessageAttachment = {
+                  type: "generated-image",
+                  name: `generated-image.${format}`,
+                  preview: `data:${mimeType};base64,${imageData}`,
+                  mimeType,
+                };
+
+                this.updateConversationMessage(
+                  targetConversationId,
+                  assistantMessage.id,
+                  (msg) => {
+                    msg.content = progressText;
+                    if (imageIndex === 0) {
+                      // First image - safe to replace attachments with partial preview
+                      msg.attachments = [partialAttachment];
+                    } else {
+                      // Subsequent images - keep existing finals, show partial at current position
+                      const existingAttachments = msg.attachments || [];
+                      // Keep only the completed final images (up to current imageIndex)
+                      const finals = existingAttachments.slice(0, imageIndex);
+                      msg.attachments = [...finals, partialAttachment];
+                    }
+                  },
+                );
+              } else if (parsed.type === "final") {
+                // Final image - replace partial at this position
+                const newAttachment: MessageAttachment = {
+                  type: "generated-image",
+                  name: `generated-image-${imageIndex + 1}.${format}`,
+                  preview: `data:${mimeType};base64,${imageData}`,
+                  mimeType,
+                };
+
+                this.updateConversationMessage(
+                  targetConversationId,
+                  assistantMessage.id,
+                  (msg) => {
+                    if (imageIndex === 0) {
+                      // First final image - replace any partial preview
+                      msg.attachments = [newAttachment];
+                    } else {
+                      // Subsequent images - keep previous finals, replace partial at current position
+                      const existingAttachments = msg.attachments || [];
+                      // Slice keeps indices 0 to imageIndex-1 (the previous final images)
+                      const previousFinals = existingAttachments.slice(
+                        0,
+                        imageIndex,
+                      );
+                      msg.attachments = [...previousFinals, newAttachment];
+                    }
+
+                    // Update progress message for multiple images
+                    if (numImages > 1 && imageIndex < numImages - 1) {
+                      msg.content = `Generating image ${imageIndex + 2}/${numImages}...`;
+                    } else {
+                      msg.content = "";
+                    }
+                  },
+                );
+              }
+
+              this.syncActiveMessagesIfNeeded(targetConversationId);
             }
-
-            this.syncActiveMessagesIfNeeded(targetConversationId);
-          }
-        },
-      );
+          },
+        );
+      }
     } catch (error) {
       console.error("Error generating image:", error);
       this.handleStreamingError(
@@ -2343,69 +2378,98 @@ class AppStore {
         throw new Error(`API error: ${apiResponse.status} - ${errorText}`);
       }
 
-      const reader = apiResponse.body?.getReader();
-      if (!reader) {
-        throw new Error("No response body");
-      }
+      // Streaming requires both stream=true AND partialImages > 0
+      const isStreaming = params.stream && params.partialImages > 0;
 
-      interface ImageEditChunk {
-        data?: { b64_json?: string };
-        format?: string;
-        type?: "partial" | "final";
-        partial_index?: number;
-        total_partials?: number;
-      }
+      if (!isStreaming) {
+        // Non-streaming: parse JSON response directly
+        const jsonResponse = (await apiResponse.json()) as ImageApiResponse;
+        const format = params.outputFormat || "png";
+        const mimeType = `image/${format}`;
+        const attachments: MessageAttachment[] = jsonResponse.data
+          .filter((img) => img.b64_json)
+          .map((img) => ({
+            type: "generated-image" as const,
+            name: `edited-image.${format}`,
+            preview: `data:${mimeType};base64,${img.b64_json}`,
+            mimeType,
+          }));
 
-      await this.parseSSEStream<ImageEditChunk>(
-        reader,
-        targetConversationId,
-        (parsed) => {
-          const imageData = parsed.data?.b64_json;
+        this.updateConversationMessage(
+          targetConversationId,
+          assistantMessage.id,
+          (msg) => {
+            msg.content = "";
+            msg.attachments = attachments;
+          },
+        );
+        this.syncActiveMessagesIfNeeded(targetConversationId);
+      } else {
+        // Streaming mode: use SSE parser
+        const reader = apiResponse.body?.getReader();
+        if (!reader) {
+          throw new Error("No response body");
+        }
 
-          if (imageData) {
-            const format = parsed.format || "png";
-            const mimeType = `image/${format}`;
-            if (parsed.type === "partial") {
-              // Update with partial image and progress
-              const partialNum = (parsed.partial_index ?? 0) + 1;
-              const totalPartials = parsed.total_partials ?? 3;
-              this.updateConversationMessage(
-                targetConversationId,
-                assistantMessage.id,
-                (msg) => {
-                  msg.content = `Editing... ${partialNum}/${totalPartials}`;
-                  msg.attachments = [
-                    {
-                      type: "generated-image",
-                      name: `edited-image.${format}`,
-                      preview: `data:${mimeType};base64,${imageData}`,
-                      mimeType,
-                    },
-                  ];
-                },
-              );
-            } else if (parsed.type === "final") {
-              // Final image
-              this.updateConversationMessage(
-                targetConversationId,
-                assistantMessage.id,
-                (msg) => {
-                  msg.content = "";
-                  msg.attachments = [
-                    {
-                      type: "generated-image",
-                      name: `edited-image.${format}`,
-                      preview: `data:${mimeType};base64,${imageData}`,
-                      mimeType,
-                    },
-                  ];
-                },
-              );
+        interface ImageEditChunk {
+          data?: { b64_json?: string };
+          format?: string;
+          type?: "partial" | "final";
+          partial_index?: number;
+          total_partials?: number;
+        }
+
+        await this.parseSSEStream<ImageEditChunk>(
+          reader,
+          targetConversationId,
+          (parsed) => {
+            const imageData = parsed.data?.b64_json;
+
+            if (imageData) {
+              const format = parsed.format || "png";
+              const mimeType = `image/${format}`;
+              if (parsed.type === "partial") {
+                // Update with partial image and progress
+                const partialNum = (parsed.partial_index ?? 0) + 1;
+                const totalPartials = parsed.total_partials ?? 3;
+                this.updateConversationMessage(
+                  targetConversationId,
+                  assistantMessage.id,
+                  (msg) => {
+                    msg.content = `Editing... ${partialNum}/${totalPartials}`;
+                    msg.attachments = [
+                      {
+                        type: "generated-image",
+                        name: `edited-image.${format}`,
+                        preview: `data:${mimeType};base64,${imageData}`,
+                        mimeType,
+                      },
+                    ];
+                  },
+                );
+              } else if (parsed.type === "final") {
+                // Final image
+                this.updateConversationMessage(
+                  targetConversationId,
+                  assistantMessage.id,
+                  (msg) => {
+                    msg.content = "";
+                    msg.attachments = [
+                      {
+                        type: "generated-image",
+                        name: `edited-image.${format}`,
+                        preview: `data:${mimeType};base64,${imageData}`,
+                        mimeType,
+                      },
+                    ];
+                  },
+                );
+              }
+              this.syncActiveMessagesIfNeeded(targetConversationId);
             }
-            this.syncActiveMessagesIfNeeded(targetConversationId);
-          }
-        },
-      );
+          },
+        );
+      }
     } catch (error) {
       console.error("Error editing image:", error);
       this.handleStreamingError(

flake.lock

@@ -21,7 +21,9 @@
           "nixpkgs"
         ],
         "purescript-overlay": "purescript-overlay",
-        "pyproject-nix": "pyproject-nix"
+        "pyproject-nix": [
+          "pyproject-nix"
+        ]
       },
       "locked": {
         "lastModified": 1765953015,
@@ -149,19 +151,44 @@
         "type": "github"
       }
     },
+    "pyproject-build-systems": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": [
+          "pyproject-nix"
+        ],
+        "uv2nix": [
+          "uv2nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1763662255,
+        "narHash": "sha256-4bocaOyLa3AfiS8KrWjZQYu+IAta05u3gYZzZ6zXbT0=",
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "rev": "042904167604c681a090c07eb6967b4dd4dae88c",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "build-system-pkgs",
+        "type": "github"
+      }
+    },
     "pyproject-nix": {
       "inputs": {
         "nixpkgs": [
-          "dream2nix",
           "nixpkgs"
         ]
       },
       "locked": {
-        "lastModified": 1763017646,
-        "narHash": "sha256-Z+R2lveIp6Skn1VPH3taQIuMhABg1IizJd8oVdmdHsQ=",
+        "lastModified": 1764134915,
+        "narHash": "sha256-xaKvtPx6YAnA3HQVp5LwyYG1MaN4LLehpQI8xEdBvBY=",
         "owner": "pyproject-nix",
         "repo": "pyproject.nix",
-        "rev": "47bd6f296502842643078d66128f7b5e5370790c",
+        "rev": "2c8df1383b32e5443c921f61224b198a2282a657",
         "type": "github"
       },
       "original": {
@@ -178,7 +205,10 @@
         "flake-parts": "flake-parts",
         "nixpkgs": "nixpkgs",
         "nixpkgs-swift": "nixpkgs-swift",
-        "treefmt-nix": "treefmt-nix"
+        "pyproject-build-systems": "pyproject-build-systems",
+        "pyproject-nix": "pyproject-nix",
+        "treefmt-nix": "treefmt-nix",
+        "uv2nix": "uv2nix"
       }
     },
     "rust-analyzer-src": {
@@ -239,6 +269,29 @@
         "repo": "treefmt-nix",
         "type": "github"
       }
+    },
+    "uv2nix": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "pyproject-nix": [
+          "pyproject-nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1767701098,
+        "narHash": "sha256-CJhKZnWb3gumR9oTRjFvCg/6lYTGbZRU7xtvcyWIRwU=",
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "rev": "9d357f0d2ce6f5f35ec7959d7e704452352eb4da",
+        "type": "github"
+      },
+      "original": {
+        "owner": "pyproject-nix",
+        "repo": "uv2nix",
+        "type": "github"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -24,6 +24,26 @@
     dream2nix = {
       url = "github:nix-community/dream2nix";
       inputs.nixpkgs.follows = "nixpkgs";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+    };
+
+    # Python packaging with uv2nix
+    pyproject-nix = {
+      url = "github:pyproject-nix/pyproject.nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+
+    uv2nix = {
+      url = "github:pyproject-nix/uv2nix";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+
+    pyproject-build-systems = {
+      url = "github:pyproject-nix/build-system-pkgs";
+      inputs.pyproject-nix.follows = "pyproject-nix";
+      inputs.uv2nix.follows = "uv2nix";
+      inputs.nixpkgs.follows = "nixpkgs";
     };
 
     # Pinned nixpkgs for swift-format (swift is broken on x86_64-linux in newer nixpkgs)
@@ -48,6 +68,7 @@
         inputs.treefmt-nix.flakeModule
         ./dashboard/parts.nix
         ./rust/parts.nix
+        ./python/parts.nix
       ];
 
       perSystem =
@@ -88,12 +109,6 @@
             };
           };
 
-          checks.lint = pkgs.runCommand "lint-check" { } ''
-            export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
-            ${pkgs.ruff}/bin/ruff check ${inputs.self}/
-            touch $out
-          '';
-
           packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin (
             let
               uvLock = builtins.fromTOML (builtins.readFile ./uv.lock);

packaging/pyinstaller/exo.spec

@@ -10,7 +10,6 @@ PROJECT_ROOT = Path.cwd()
 SOURCE_ROOT = PROJECT_ROOT / "src"
 ENTRYPOINT = SOURCE_ROOT / "exo" / "__main__.py"
 DASHBOARD_DIR = PROJECT_ROOT / "dashboard" / "build"
-RESOURCES_DIR = PROJECT_ROOT / "resources"
 EXO_SHARED_MODELS_DIR = SOURCE_ROOT / "exo" / "shared" / "models"
 
 if not ENTRYPOINT.is_file():
@@ -19,9 +18,6 @@ if not ENTRYPOINT.is_file():
 if not DASHBOARD_DIR.is_dir():
     raise SystemExit(f"Dashboard assets are missing: {DASHBOARD_DIR}")
 
-if not RESOURCES_DIR.is_dir():
-    raise SystemExit(f"Resource assets are missing: {RESOURCES_DIR}")
-
 if not EXO_SHARED_MODELS_DIR.is_dir():
     raise SystemExit(f"Shared model assets are missing: {EXO_SHARED_MODELS_DIR}")
 
@@ -62,7 +58,6 @@ HIDDEN_IMPORTS = sorted(
 
 DATAS: list[tuple[str, str]] = [
     (str(DASHBOARD_DIR), "dashboard"),
-    (str(RESOURCES_DIR), "resources"),
     (str(MLX_LIB_DIR), "mlx/lib"),
     (str(EXO_SHARED_MODELS_DIR), "exo/shared/models"),
 ]

python/parts.nix

@@ -0,0 +1,93 @@
+{ inputs, ... }:
+{
+  perSystem =
+    { config, self', pkgs, lib, system, ... }:
+    let
+      # Load workspace from uv.lock
+      workspace = inputs.uv2nix.lib.workspace.loadWorkspace {
+        workspaceRoot = inputs.self;
+      };
+
+      # Create overlay from workspace
+      # Use wheels from PyPI for most packages; we override mlx with our pure Nix Metal build
+      overlay = workspace.mkPyprojectOverlay { sourcePreference = "wheel"; };
+
+      # Override overlay to inject Nix-built components
+      exoOverlay = final: prev: {
+        # Replace workspace exo_pyo3_bindings with Nix-built wheel
+        exo-pyo3-bindings = pkgs.stdenv.mkDerivation {
+          pname = "exo-pyo3-bindings";
+          version = "0.1.0";
+          src = self'.packages.exo_pyo3_bindings;
+          # Install from pre-built wheel
+          nativeBuildInputs = [ final.pyprojectWheelHook ];
+          dontStrip = true;
+        };
+      };
+
+      python = pkgs.python313;
+
+      # Overlay to provide build systems and custom packages
+      buildSystemsOverlay = final: prev: {
+        # Use our pure Nix-built MLX with Metal support
+        mlx = self'.packages.mlx;
+
+        # mlx-lm is a git dependency that needs setuptools
+        mlx-lm = prev.mlx-lm.overrideAttrs (old: {
+          nativeBuildInputs = (old.nativeBuildInputs or [ ]) ++ [
+            final.setuptools
+          ];
+        });
+      };
+
+      pythonSet = (pkgs.callPackage inputs.pyproject-nix.build.packages {
+        inherit python;
+      }).overrideScope (
+        lib.composeManyExtensions [
+          inputs.pyproject-build-systems.overlays.default
+          overlay
+          exoOverlay
+          buildSystemsOverlay
+        ]
+      );
+      exoVenv = pythonSet.mkVirtualEnv "exo-env" workspace.deps.default;
+
+      # Virtual environment with dev dependencies for testing
+      testVenv = pythonSet.mkVirtualEnv "exo-test-env" (
+        workspace.deps.default // {
+          exo = [ "dev" ]; # Include pytest, pytest-asyncio, pytest-env
+        }
+      );
+
+      exoPackage = pkgs.runCommand "exo"
+        {
+          nativeBuildInputs = [ pkgs.makeWrapper ];
+        }
+        ''
+          mkdir -p $out/bin
+
+          # Create wrapper scripts
+          for script in exo exo-master exo-worker; do
+            makeWrapper ${exoVenv}/bin/$script $out/bin/$script \
+              --set DASHBOARD_DIR ${self'.packages.dashboard}
+          done
+        '';
+    in
+    {
+      # Python package only available on macOS (requires MLX/Metal)
+      packages = lib.optionalAttrs pkgs.stdenv.hostPlatform.isDarwin {
+        exo = exoPackage;
+        # Test environment for running pytest outside of Nix sandbox (needs GPU access)
+        exo-test-env = testVenv;
+      };
+
+      checks = {
+        # Ruff linting (works on all platforms)
+        lint = pkgs.runCommand "ruff-lint" { } ''
+          export RUFF_CACHE_DIR="$TMPDIR/ruff-cache"
+          ${pkgs.ruff}/bin/ruff check ${inputs.self}/
+          touch $out
+        '';
+      };
+    };
+}

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-4bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-Krea-dev-4bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 15475325472
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 5950704160
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-Krea-dev-8bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-Krea-dev-8bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 21426029632
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 11901408320
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-Krea-dev.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-Krea-dev"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 33327437952
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 23802816640
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev-4bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-dev-4bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 15475325472
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 5950704160
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev-8bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-dev-8bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 21426029632
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 11901408320
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-dev.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-dev"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 33327437952
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 23802816640
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell-4bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-schnell-4bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 15470210592
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 5945589280
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell-8bit.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-schnell-8bit"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 21415799872
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 11891178560
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--FLUX.1-schnell.toml

@@ -1,45 +0,0 @@
-model_id = "exolabs/FLUX.1-schnell"
-n_layers = 57
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 33306978432
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0
-
-[[components]]
-component_name = "text_encoder_2"
-component_path = "text_encoder_2/"
-n_layers = 24
-can_shard = false
-safetensors_index_filename = "model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 9524621312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 57
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 23782357120
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-4bit.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image-4bit"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 26799533856
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 10215200544
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-8bit.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image-8bit"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 37014734400
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 20430401088
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-4bit.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image-Edit-2509-4bit"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["ImageToImage"]
-
-[storage_size]
-in_bytes = 26799533856
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 10215200544
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509-8bit.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image-Edit-2509-8bit"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["ImageToImage"]
-
-[storage_size]
-in_bytes = 37014734400
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 20430401088
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image-Edit-2509.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image-Edit-2509"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["ImageToImage"]
-
-[storage_size]
-in_bytes = 57445135488
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 40860802176
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/image_model_cards/exolabs--Qwen-Image.toml

@@ -1,35 +0,0 @@
-model_id = "exolabs/Qwen-Image"
-n_layers = 60
-hidden_size = 1
-supports_tensor = false
-tasks = ["TextToImage"]
-
-[storage_size]
-in_bytes = 57445135488
-
-[[components]]
-component_name = "text_encoder"
-component_path = "text_encoder/"
-n_layers = 12
-can_shard = false
-
-[components.storage_size]
-in_bytes = 16584333312
-
-[[components]]
-component_name = "transformer"
-component_path = "transformer/"
-n_layers = 60
-can_shard = true
-safetensors_index_filename = "diffusion_pytorch_model.safetensors.index.json"
-
-[components.storage_size]
-in_bytes = 40860802176
-
-[[components]]
-component_name = "vae"
-component_path = "vae/"
-can_shard = false
-
-[components.storage_size]
-in_bytes = 0

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/DeepSeek-V3.1-4bit"
-n_layers = 61
-hidden_size = 7168
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 405874409472

resources/inference_model_cards/mlx-community--DeepSeek-V3.1-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/DeepSeek-V3.1-8bit"
-n_layers = 61
-hidden_size = 7168
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 765577920512

resources/inference_model_cards/mlx-community--GLM-4.5-Air-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.5-Air-8bit"
-n_layers = 46
-hidden_size = 4096
-supports_tensor = false
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 122406567936

resources/inference_model_cards/mlx-community--GLM-4.5-Air-bf16.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.5-Air-bf16"
-n_layers = 46
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 229780750336

resources/inference_model_cards/mlx-community--GLM-4.7-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-4bit"
-n_layers = 91
-hidden_size = 5120
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 198556925568

resources/inference_model_cards/mlx-community--GLM-4.7-6bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-6bit"
-n_layers = 91
-hidden_size = 5120
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 286737579648

resources/inference_model_cards/mlx-community--GLM-4.7-8bit-gs32.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-8bit-gs32"
-n_layers = 91
-hidden_size = 5120
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 396963397248

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-Flash-4bit"
-n_layers = 47
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 19327352832

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-5bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-Flash-5bit"
-n_layers = 47
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 22548578304

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-6bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-Flash-6bit"
-n_layers = 47
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 26843545600

resources/inference_model_cards/mlx-community--GLM-4.7-Flash-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/GLM-4.7-Flash-8bit"
-n_layers = 47
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 34359738368

resources/inference_model_cards/mlx-community--Kimi-K2-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Kimi-K2-Instruct-4bit"
-n_layers = 61
-hidden_size = 7168
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 620622774272

resources/inference_model_cards/mlx-community--Kimi-K2-Thinking.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Kimi-K2-Thinking"
-n_layers = 61
-hidden_size = 7168
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 706522120192

resources/inference_model_cards/mlx-community--Kimi-K2.5.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Kimi-K2.5"
-n_layers = 61
-hidden_size = 7168
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 662498705408

resources/inference_model_cards/mlx-community--Llama-3.2-1B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"
-n_layers = 16
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 729808896

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Llama-3.2-3B-Instruct-4bit"
-n_layers = 28
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 1863319552

resources/inference_model_cards/mlx-community--Llama-3.2-3B-Instruct-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Llama-3.2-3B-Instruct-8bit"
-n_layers = 28
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 3501195264

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Llama-3.3-70B-Instruct-4bit"
-n_layers = 80
-hidden_size = 8192
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 40652242944

resources/inference_model_cards/mlx-community--Llama-3.3-70B-Instruct-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Llama-3.3-70B-Instruct-8bit"
-n_layers = 80
-hidden_size = 8192
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 76799803392

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-70B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"
-n_layers = 80
-hidden_size = 8192
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 40652242944

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"
-n_layers = 32
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 4637851648

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"
-n_layers = 32
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 8954839040

resources/inference_model_cards/mlx-community--Meta-Llama-3.1-8B-Instruct-bf16.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"
-n_layers = 32
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 16882073600

resources/inference_model_cards/mlx-community--MiniMax-M2.1-3bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/MiniMax-M2.1-3bit"
-n_layers = 61
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 100086644736

resources/inference_model_cards/mlx-community--MiniMax-M2.1-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/MiniMax-M2.1-8bit"
-n_layers = 61
-hidden_size = 3072
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 242986745856

resources/inference_model_cards/mlx-community--Qwen3-0.6B-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-0.6B-4bit"
-n_layers = 28
-hidden_size = 1024
-supports_tensor = false
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 342884352

resources/inference_model_cards/mlx-community--Qwen3-0.6B-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-0.6B-8bit"
-n_layers = 28
-hidden_size = 1024
-supports_tensor = false
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 698351616

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"
-n_layers = 94
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 141733920768

resources/inference_model_cards/mlx-community--Qwen3-235B-A22B-Instruct-2507-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"
-n_layers = 94
-hidden_size = 4096
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 268435456000

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-30B-A3B-4bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 17612931072

resources/inference_model_cards/mlx-community--Qwen3-30B-A3B-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-30B-A3B-8bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 33279705088

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"
-n_layers = 62
-hidden_size = 6144
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 289910292480

resources/inference_model_cards/mlx-community--Qwen3-Coder-480B-A35B-Instruct-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"
-n_layers = 62
-hidden_size = 6144
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 579820584960

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Instruct-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 46976204800

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Instruct-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 88814387200

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Thinking-4bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 88814387200

resources/inference_model_cards/mlx-community--Qwen3-Next-80B-A3B-Thinking-8bit.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"
-n_layers = 48
-hidden_size = 2048
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 88814387200

resources/inference_model_cards/mlx-community--gpt-oss-120b-MXFP4-Q8.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/gpt-oss-120b-MXFP4-Q8"
-n_layers = 36
-hidden_size = 2880
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 70652212224

resources/inference_model_cards/mlx-community--gpt-oss-20b-MXFP4-Q8.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/gpt-oss-20b-MXFP4-Q8"
-n_layers = 24
-hidden_size = 2880
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 12025908224

resources/inference_model_cards/mlx-community--llama-3.3-70b-instruct-fp16.toml

@@ -1,8 +0,0 @@
-model_id = "mlx-community/llama-3.3-70b-instruct-fp16"
-n_layers = 80
-hidden_size = 8192
-supports_tensor = true
-tasks = ["TextGeneration"]
-
-[storage_size]
-in_bytes = 144383672320

src/exo/download/impl_shard_downloader.py

@@ -7,7 +7,7 @@ from loguru import logger
 
 from exo.download.download_utils import RepoDownloadProgress, download_shard
 from exo.download.shard_downloader import ShardDownloader
-from exo.shared.models.model_cards import ModelCard, ModelId, get_model_cards
+from exo.shared.models.model_cards import MODEL_CARDS, ModelCard, ModelId
 from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
@@ -21,7 +21,7 @@ def exo_shard_downloader(max_parallel_downloads: int = 8) -> ShardDownloader:
 
 
 async def build_base_shard(model_id: ModelId) -> ShardMetadata:
-    model_card = await ModelCard.fetch_from_hf(model_id)
+    model_card = await ModelCard.load(model_id)
     return PipelineShardMetadata(
         model_card=model_card,
         device_rank=0,
@@ -160,15 +160,14 @@ class ResumableShardDownloader(ShardDownloader):
         # Kick off download status coroutines concurrently
         tasks = [
             asyncio.create_task(_status_for_model(model_card.model_id))
-            for model_card in await get_model_cards()
+            for model_card in MODEL_CARDS.values()
         ]
 
         for task in asyncio.as_completed(tasks):
             try:
                 yield await task
-            # TODO: except Exception
             except Exception as e:
-                logger.error("Error downloading shard:", e)
+                logger.warning(f"Error downloading shard: {type(e).__name__}")
 
     async def get_shard_download_status_for_shard(
         self, shard: ShardMetadata

src/exo/main.py

@@ -40,6 +40,7 @@ class Node:
 
     node_id: NodeId
     event_index_counter: Iterator[int]
+    _profiling_in_progress: set[NodeId] = field(default_factory=set)
     _tg: TaskGroup = field(init=False, default_factory=anyio.create_task_group)
 
     @classmethod
@@ -86,6 +87,7 @@ class Node:
         else:
             api = None
 
+        profiling_in_progress: set[NodeId] = set()
         if not args.no_worker:
             worker = Worker(
                 node_id,
@@ -96,6 +98,7 @@ class Node:
                 command_sender=router.sender(topics.COMMANDS),
                 download_command_sender=router.sender(topics.DOWNLOAD_COMMANDS),
                 event_index_counter=event_index_counter,
+                profiling_in_progress=profiling_in_progress,
             )
         else:
             worker = None
@@ -133,6 +136,7 @@ class Node:
             api,
             node_id,
             event_index_counter,
+            profiling_in_progress,
         )
 
     async def run(self):
@@ -239,6 +243,7 @@ class Node:
                                 topics.DOWNLOAD_COMMANDS
                             ),
                             event_index_counter=self.event_index_counter,
+                            profiling_in_progress=self._profiling_in_progress,
                         )
                         self._tg.start_soon(self.worker.run)
                     if self.api:

src/exo/master/api.py

@@ -8,8 +8,8 @@ from typing import Annotated, Literal, cast
 from uuid import uuid4
 
 import anyio
-from anyio import BrokenResourceError, create_task_group
-from anyio.abc import TaskGroup
+from anyio import BrokenResourceError, ClosedResourceError, EndOfStream, create_task_group
+from anyio.abc import SocketStream, TaskGroup
 from fastapi import FastAPI, File, Form, HTTPException, Query, Request, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, JSONResponse, StreamingResponse
@@ -22,13 +22,16 @@ from loguru import logger
 from exo.master.image_store import ImageStore
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
-from exo.shared.constants import DASHBOARD_DIR, EXO_IMAGE_CACHE_DIR, EXO_MAX_CHUNK_SIZE
+from exo.shared.constants import (
+    EXO_IMAGE_CACHE_DIR,
+    EXO_MAX_CHUNK_SIZE,
+)
 from exo.shared.election import ElectionMessage
 from exo.shared.logging import InterceptLogger
 from exo.shared.models.model_cards import (
+    MODEL_CARDS,
     ModelCard,
     ModelId,
-    get_model_cards,
 )
 from exo.shared.types.api import (
     AdvancedImageParams,
@@ -62,7 +65,9 @@ from exo.shared.types.api import (
     StartDownloadParams,
     StartDownloadResponse,
     StreamingChoiceResponse,
+    StreamOptions,
     ToolCall,
+    Usage,
 )
 from exo.shared.types.chunks import (
     ErrorChunk,
@@ -101,6 +106,7 @@ from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
 from exo.utils.channels import Receiver, Sender, channel
+from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
 
@@ -109,7 +115,9 @@ def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None)
 
 
 def chunk_to_response(
-    chunk: TokenChunk | ToolCallChunk, command_id: CommandId
+    chunk: TokenChunk | ToolCallChunk,
+    command_id: CommandId,
+    usage: Usage | None,
 ) -> ChatCompletionResponse:
     return ChatCompletionResponse(
         id=command_id,
@@ -134,6 +142,7 @@ def chunk_to_response(
                 finish_reason=chunk.finish_reason,
             )
         ],
+        usage=usage,
     )
 
 
@@ -174,7 +183,7 @@ class API:
         self.app.mount(
             "/",
             StaticFiles(
-                directory=DASHBOARD_DIR,
+                directory=find_dashboard(),
                 html=True,
             ),
             name="dashboard",
@@ -345,7 +354,10 @@ class API:
         if len(list(self.state.topology.list_nodes())) == 0:
             return PlacementPreviewResponse(previews=[])
 
-        model_card = await ModelCard.load(model_id)
+        cards = [card for card in MODEL_CARDS.values() if card.model_id == model_id]
+        if not cards:
+            raise HTTPException(status_code=404, detail=f"Model {model_id} not found")
+
         instance_combinations: list[tuple[Sharding, InstanceMeta, int]] = []
         for sharding in (Sharding.Pipeline, Sharding.Tensor):
             for instance_meta in (InstanceMeta.MlxRing, InstanceMeta.MlxJaccl):
@@ -360,93 +372,96 @@ class API:
         # TODO: PDD
         # instance_combinations.append((Sharding.PrefillDecodeDisaggregation, InstanceMeta.MlxRing, 1))
 
-        for sharding, instance_meta, min_nodes in instance_combinations:
-            try:
-                placements = get_instance_placements(
-                    PlaceInstance(
-                        model_card=model_card,
-                        sharding=sharding,
-                        instance_meta=instance_meta,
-                        min_nodes=min_nodes,
-                    ),
-                    node_memory=self.state.node_memory,
-                    node_network=self.state.node_network,
-                    topology=self.state.topology,
-                    current_instances=self.state.instances,
-                    required_nodes=required_nodes,
-                )
-            except ValueError as exc:
-                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
-                    previews.append(
-                        PlacementPreview(
-                            model_id=model_card.model_id,
+        for model_card in cards:
+            for sharding, instance_meta, min_nodes in instance_combinations:
+                try:
+                    placements = get_instance_placements(
+                        PlaceInstance(
+                            model_card=model_card,
                             sharding=sharding,
                             instance_meta=instance_meta,
-                            instance=None,
-                            error=str(exc),
+                            min_nodes=min_nodes,
+                        ),
+                        node_memory=self.state.node_memory,
+                        node_network=self.state.node_network,
+                        topology=self.state.topology,
+                        current_instances=self.state.instances,
+                        required_nodes=required_nodes,
+                    )
+                except ValueError as exc:
+                    if (model_card.model_id, sharding, instance_meta, 0) not in seen:
+                        previews.append(
+                            PlacementPreview(
+                                model_id=model_card.model_id,
+                                sharding=sharding,
+                                instance_meta=instance_meta,
+                                instance=None,
+                                error=str(exc),
+                            )
                         )
-                    )
-                seen.add((model_card.model_id, sharding, instance_meta, 0))
-                continue
+                    seen.add((model_card.model_id, sharding, instance_meta, 0))
+                    continue
 
-            current_ids = set(self.state.instances.keys())
-            new_instances = [
-                instance
-                for instance_id, instance in placements.items()
-                if instance_id not in current_ids
-            ]
+                current_ids = set(self.state.instances.keys())
+                new_instances = [
+                    instance
+                    for instance_id, instance in placements.items()
+                    if instance_id not in current_ids
+                ]
 
-            if len(new_instances) != 1:
-                if (model_card.model_id, sharding, instance_meta, 0) not in seen:
-                    previews.append(
-                        PlacementPreview(
-                            model_id=model_card.model_id,
-                            sharding=sharding,
-                            instance_meta=instance_meta,
-                            instance=None,
-                            error="Expected exactly one new instance from placement",
+                if len(new_instances) != 1:
+                    if (model_card.model_id, sharding, instance_meta, 0) not in seen:
+                        previews.append(
+                            PlacementPreview(
+                                model_id=model_card.model_id,
+                                sharding=sharding,
+                                instance_meta=instance_meta,
+                                instance=None,
+                                error="Expected exactly one new instance from placement",
+                            )
                         )
-                    )
-                seen.add((model_card.model_id, sharding, instance_meta, 0))
-                continue
+                    seen.add((model_card.model_id, sharding, instance_meta, 0))
+                    continue
 
-            instance = new_instances[0]
-            shard_assignments = instance.shard_assignments
-            placement_node_ids = list(shard_assignments.node_to_runner.keys())
+                instance = new_instances[0]
+                shard_assignments = instance.shard_assignments
+                placement_node_ids = list(shard_assignments.node_to_runner.keys())
 
-            memory_delta_by_node: dict[str, int] = {}
-            if placement_node_ids:
-                total_bytes = model_card.storage_size.in_bytes
-                per_node = total_bytes // len(placement_node_ids)
-                remainder = total_bytes % len(placement_node_ids)
-                for index, node_id in enumerate(sorted(placement_node_ids, key=str)):
-                    extra = 1 if index < remainder else 0
-                    memory_delta_by_node[str(node_id)] = per_node + extra
+                memory_delta_by_node: dict[str, int] = {}
+                if placement_node_ids:
+                    total_bytes = model_card.storage_size.in_bytes
+                    per_node = total_bytes // len(placement_node_ids)
+                    remainder = total_bytes % len(placement_node_ids)
+                    for index, node_id in enumerate(
+                        sorted(placement_node_ids, key=str)
+                    ):
+                        extra = 1 if index < remainder else 0
+                        memory_delta_by_node[str(node_id)] = per_node + extra
 
-            if (
-                model_card.model_id,
-                sharding,
-                instance_meta,
-                len(placement_node_ids),
-            ) not in seen:
-                previews.append(
-                    PlacementPreview(
-                        model_id=model_card.model_id,
-                        sharding=sharding,
-                        instance_meta=instance_meta,
-                        instance=instance,
-                        memory_delta_by_node=memory_delta_by_node or None,
-                        error=None,
-                    )
-                )
-            seen.add(
-                (
+                if (
                     model_card.model_id,
                     sharding,
                     instance_meta,
                     len(placement_node_ids),
+                ) not in seen:
+                    previews.append(
+                        PlacementPreview(
+                            model_id=model_card.model_id,
+                            sharding=sharding,
+                            instance_meta=instance_meta,
+                            instance=instance,
+                            memory_delta_by_node=memory_delta_by_node or None,
+                            error=None,
+                        )
+                    )
+                seen.add(
+                    (
+                        model_card.model_id,
+                        sharding,
+                        instance_meta,
+                        len(placement_node_ids),
+                    )
                 )
-            )
 
         return PlacementPreviewResponse(previews=previews)
 
@@ -500,9 +515,10 @@ class API:
                 del self._chat_completion_queues[command_id]
 
     async def _generate_chat_stream(
-        self, command_id: CommandId
+        self, command_id: CommandId, stream_options: StreamOptions | None = None
     ) -> AsyncGenerator[str, None]:
         """Generate chat completion stream as JSON strings."""
+        include_usage = stream_options.include_usage if stream_options else False
 
         async for chunk in self._chat_chunk_stream(command_id):
             assert not isinstance(chunk, ImageChunk)
@@ -518,8 +534,10 @@ class API:
                 yield "data: [DONE]\n\n"
                 return
 
+            usage = chunk.usage if include_usage else None
+
             chunk_response: ChatCompletionResponse = chunk_to_response(
-                chunk, command_id
+                chunk, command_id, usage=usage
             )
             logger.debug(f"chunk_response: {chunk_response}")
 
@@ -535,8 +553,9 @@ class API:
 
         text_parts: list[str] = []
         tool_calls: list[ToolCall] = []
-        model: str | None = None
+        model: ModelId | None = None
         finish_reason: FinishReason | None = None
+        usage: Usage | None = None
 
         async for chunk in self._chat_chunk_stream(command_id):
             if isinstance(chunk, ErrorChunk):
@@ -561,6 +580,9 @@ class API:
                     for i, tool in enumerate(chunk.tool_calls)
                 )
 
+            if chunk.usage is not None:
+                usage = chunk.usage
+
             if chunk.finish_reason is not None:
                 finish_reason = chunk.finish_reason
 
@@ -582,6 +604,7 @@ class API:
                     finish_reason=finish_reason,
                 )
             ],
+            usage=usage,
         )
 
     async def _collect_chat_completion_with_stats(
@@ -589,7 +612,7 @@ class API:
     ) -> BenchChatCompletionResponse:
         text_parts: list[str] = []
         tool_calls: list[ToolCall] = []
-        model: str | None = None
+        model: ModelId | None = None
         finish_reason: FinishReason | None = None
 
         stats: GenerationStats | None = None
@@ -642,7 +665,7 @@ class API:
         )
         return resp
 
-    async def _trigger_notify_user_to_download_model(self, model_id: str) -> None:
+    async def _trigger_notify_user_to_download_model(self, model_id: ModelId) -> None:
         logger.warning(
             "TODO: we should send a notification to the user to download the model"
         )
@@ -669,7 +692,7 @@ class API:
         await self._send(command)
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id),
+                self._generate_chat_stream(command.command_id, payload.stream_options),
                 media_type="text/event-stream",
             )
 
@@ -698,12 +721,12 @@ class API:
         response = await self._collect_chat_completion_with_stats(command.command_id)
         return response
 
-    async def _validate_image_model(self, model: str) -> ModelId:
+    async def _validate_image_model(self, model: ModelId) -> ModelId:
         """Validate model exists and return resolved model ID.
 
         Raises HTTPException 404 if no instance is found for the model.
         """
-        model_card = await ModelCard.load(ModelId(model))
+        model_card = await ModelCard.load(model)
         resolved_model = model_card.model_id
         if not any(
             instance.shard_assignments.model_id == resolved_model
@@ -749,7 +772,7 @@ class API:
         When stream=True and partial_images > 0, returns a StreamingResponse
         with SSE-formatted events for partial and final images.
         """
-        payload.model = await self._validate_image_model(payload.model)
+        payload.model = await self._validate_image_model(ModelId(payload.model))
 
         command = ImageGeneration(
             request_params=payload,
@@ -994,7 +1017,7 @@ class API:
     async def bench_image_generations(
         self, request: Request, payload: BenchImageGenerationTaskParams
     ) -> BenchImageGenerationResponse:
-        payload.model = await self._validate_image_model(payload.model)
+        payload.model = await self._validate_image_model(ModelId(payload.model))
 
         payload.stream = False
         payload.partial_images = 0
@@ -1015,7 +1038,7 @@ class API:
         self,
         image: UploadFile,
         prompt: str,
-        model: str,
+        model: ModelId,
         n: int,
         size: str,
         response_format: Literal["url", "b64_json"],
@@ -1110,7 +1133,7 @@ class API:
         command = await self._send_image_edits_command(
             image=image,
             prompt=prompt,
-            model=model,
+            model=ModelId(model),
             n=n,
             size=size,
             response_format=response_format,
@@ -1166,7 +1189,7 @@ class API:
         command = await self._send_image_edits_command(
             image=image,
             prompt=prompt,
-            model=model,
+            model=ModelId(model),
             n=n,
             size=size,
             response_format=response_format,
@@ -1209,7 +1232,7 @@ class API:
                     supports_tensor=card.supports_tensor,
                     tasks=[task.value for task in card.tasks],
                 )
-                for card in await get_model_cards()
+                for card in MODEL_CARDS.values()
             ]
         )
 
@@ -1227,6 +1250,7 @@ class API:
             tg.start_soon(self._apply_state)
             tg.start_soon(self._pause_on_new_election)
             tg.start_soon(self._cleanup_expired_images)
+            tg.start_soon(self._run_bandwidth_server)
             print_startup_banner(self.port)
             await serve(
                 cast(ASGIFramework, self.app),
@@ -1234,6 +1258,43 @@ class API:
                 shutdown_trigger=lambda: anyio.sleep_forever(),
             )
 
+    async def _run_bandwidth_server(self):
+        """TCP server for iperf-like bandwidth testing."""
+        bandwidth_port = self.port + 1
+        listener = await anyio.create_tcp_listener(local_port=bandwidth_port)
+        logger.info(f"Bandwidth test server listening on port {bandwidth_port}")
+        await listener.serve(self._handle_bandwidth_connection)
+
+    async def _handle_bandwidth_connection(self, stream: SocketStream) -> None:
+        """Handle a single bandwidth test connection."""
+        try:
+            mode = await stream.receive(1)
+            if mode == b"U":
+                # Upload test: client sends, we receive
+                bytes_received = 0
+                start = time.perf_counter()
+                while True:
+                    try:
+                        data = await stream.receive(1024 * 1024)
+                        if not data or data == b"DONE" or b"DONE" in data:
+                            break
+                        bytes_received += len(data)
+                    except EndOfStream:
+                        break
+                elapsed = time.perf_counter() - start
+                logger.debug(f"Bandwidth upload: {bytes_received} bytes in {elapsed:.3f}s")
+            elif mode == b"D":
+                # Download test: we send, client receives
+                chunk = b"X" * (1024 * 1024)
+                start = time.perf_counter()
+                while time.perf_counter() - start < 1.0:  # Send for 1s
+                    try:
+                        await stream.send(chunk)
+                    except (BrokenResourceError, ClosedResourceError):
+                        break
+        except Exception as e:
+            logger.debug(f"Bandwidth connection error: {e}")
+
         self.command_sender.close()
         self.global_event_receiver.close()
 

src/exo/master/tests/conftest.py

@@ -28,6 +28,9 @@ def create_node_network() -> NodeNetworkInfo:
 def create_socket_connection(ip: int, sink_port: int = 1234) -> SocketConnection:
     return SocketConnection(
         sink_multiaddr=Multiaddr(address=f"/ip4/169.254.0.{ip}/tcp/{sink_port}"),
+        latency_ms=1.0,
+        other_to_sink_bandwidth_mbps=1000.0,
+        sink_to_other_bandwidth_mbps=1000.0,
     )
 
 

src/exo/master/tests/test_placement.py

@@ -366,7 +366,10 @@ def test_tensor_rdma_backend_connectivity_matrix(
         ip_address="10.0.0.1",
     )
     ethernet_conn = SocketConnection(
-        sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/8000")
+        sink_multiaddr=Multiaddr(address="/ip4/10.0.0.1/tcp/8000"),
+        latency_ms=1.0,
+        other_to_sink_bandwidth_mbps=1000.0,
+        sink_to_other_bandwidth_mbps=1000.0,
     )
 
     node_network = {

src/exo/master/tests/test_topology.py

@@ -15,6 +15,9 @@ def topology() -> Topology:
 def socket_connection() -> SocketConnection:
     return SocketConnection(
         sink_multiaddr=Multiaddr(address="/ip4/127.0.0.1/tcp/1235"),
+        latency_ms=1.0,
+        sink_to_other_bandwidth_mbps=1000.0,
+        other_to_sink_bandwidth_mbps=1000.0,
     )
 
 

src/exo/routing/router.py

@@ -216,6 +216,8 @@ def get_node_id_keypair(
     Obtains the :class:`Keypair` associated with this node-ID.
     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_ed25519()
 
     def lock_path(path: str | bytes | PathLike[str] | PathLike[bytes]) -> Path:
         return Path(str(path) + ".lock")

src/exo/shared/constants.py

@@ -2,8 +2,6 @@ import os
 import sys
 from pathlib import Path
 
-from exo.utils.dashboard_path import find_dashboard, find_resources
-
 _EXO_HOME_ENV = os.environ.get("EXO_HOME", None)
 
 
@@ -33,14 +31,6 @@ EXO_MODELS_DIR = (
     if _EXO_MODELS_DIR_ENV is None
     else Path.home() / _EXO_MODELS_DIR_ENV
 )
-_RESOURCES_DIR_ENV = os.environ.get("EXO_RESOURCES_DIR", None)
-RESOURCES_DIR = (
-    find_resources() if _RESOURCES_DIR_ENV is None else Path.home() / _RESOURCES_DIR_ENV
-)
-_DASHBOARD_DIR_ENV = os.environ.get("EXO_DASHBOARD_DIR", None)
-DASHBOARD_DIR = (
-    find_dashboard() if _RESOURCES_DIR_ENV is None else Path.home() / _RESOURCES_DIR_ENV
-)
 
 # Log files (data/logs or cache)
 EXO_LOG = EXO_CACHE_HOME / "exo.log"

src/exo/shared/models/model_cards.py

@@ -12,42 +12,16 @@ from pydantic import (
     BaseModel,
     Field,
     PositiveInt,
-    ValidationError,
     field_validator,
     model_validator,
 )
-from tomlkit.exceptions import TOMLKitError
 
-from exo.shared.constants import EXO_ENABLE_IMAGE_MODELS, RESOURCES_DIR
+from exo.shared.constants import EXO_ENABLE_IMAGE_MODELS
 from exo.shared.types.common import ModelId
 from exo.shared.types.memory import Memory
 from exo.utils.pydantic_ext import CamelCaseModel
 
-# kinda ugly...
-# TODO: load search path from config.toml
-_csp = [Path(RESOURCES_DIR)]
-if EXO_ENABLE_IMAGE_MODELS:
-    _csp.append(Path(RESOURCES_DIR) / "image_models")
-
-CARD_SEARCH_PATH = _csp
-
-_card_cache: dict[ModelId, "ModelCard"] = {}
-
-
-async def _populate_card_cache():
-    for path in CARD_SEARCH_PATH:
-        async for toml_file in path.rglob("*.toml"):
-            try:
-                card = await ModelCard.load_from_path(toml_file)
-                _card_cache[card.model_id] = card
-            except (ValidationError, TOMLKitError):
-                pass
-
-
-async def get_model_cards() -> list["ModelCard"]:
-    if len(_card_cache) == 0:
-        await _populate_card_cache()
-    return list(_card_cache.values())
+_card_cache: dict[str, "ModelCard"] = {}
 
 
 class ModelTask(str, Enum):
@@ -81,37 +55,28 @@ class ModelCard(CamelCaseModel):
 
     async def save(self, path: Path) -> None:
         async with await open_file(path, "w") as f:
-            py = self.model_dump(exclude_none=True)
+            py = self.model_dump()
             data = tomlkit.dumps(py)  # pyright: ignore[reportUnknownMemberType]
             await f.write(data)
 
-    async def save_to_default_path(self):
-        await self.save(Path(RESOURCES_DIR) / (self.model_id.normalize() + ".toml"))
-
     @staticmethod
     async def load_from_path(path: Path) -> "ModelCard":
         async with await open_file(path, "r") as f:
             py = tomlkit.loads(await f.read())
             return ModelCard.model_validate(py)
 
-    # Is it okay that model card.load defaults to network access if the card doesn't exist? do we want to be more explicit here?
     @staticmethod
     async def load(model_id: ModelId) -> "ModelCard":
-        if len(_card_cache) == 0:
-            await _populate_card_cache()
-        if (mc := _card_cache.get(model_id)) is not None:
-            return mc
-
-        return await ModelCard.fetch_from_hf(model_id)
+        for card in MODEL_CARDS.values():
+            if card.model_id == model_id:
+                return card
+        return await ModelCard.from_hf(model_id)
 
     @staticmethod
-    async def fetch_from_hf(model_id: ModelId) -> "ModelCard":
+    async def from_hf(model_id: ModelId) -> "ModelCard":
         """Fetches storage size and number of layers for a Hugging Face model, returns Pydantic ModelMeta."""
-        if len(_card_cache) == 0:
-            await _populate_card_cache()
         if (mc := _card_cache.get(model_id)) is not None:
             return mc
-        # TODO: failure if files do not exist
         config_data = await get_config_data(model_id)
         num_layers = config_data.layer_count
         mem_size_bytes = await get_safetensors_size(model_id)
@@ -124,13 +89,544 @@ class ModelCard(CamelCaseModel):
             supports_tensor=config_data.supports_tensor,
             tasks=[ModelTask.TextGeneration],
         )
-        await mc.save_to_default_path()
         _card_cache[model_id] = mc
         return mc
 
 
-# TODO: quantizing and dynamically creating model cards
-def _generate_image_model_quant_variants(  # pyright: ignore[reportUnusedFunction]
+MODEL_CARDS: dict[str, ModelCard] = {
+    # deepseek v3
+    "deepseek-v3.1-4bit": ModelCard(
+        model_id=ModelId("mlx-community/DeepSeek-V3.1-4bit"),
+        storage_size=Memory.from_gb(378),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "deepseek-v3.1-8bit": ModelCard(
+        model_id=ModelId("mlx-community/DeepSeek-V3.1-8bit"),
+        storage_size=Memory.from_gb(713),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # kimi k2
+    "kimi-k2-instruct-4bit": ModelCard(
+        model_id=ModelId("mlx-community/Kimi-K2-Instruct-4bit"),
+        storage_size=Memory.from_gb(578),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "kimi-k2-thinking": ModelCard(
+        model_id=ModelId("mlx-community/Kimi-K2-Thinking"),
+        storage_size=Memory.from_gb(658),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "kimi-k2.5": ModelCard(
+        model_id=ModelId("mlx-community/Kimi-K2.5"),
+        storage_size=Memory.from_gb(617),
+        n_layers=61,
+        hidden_size=7168,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # llama-3.1
+    "llama-3.1-8b": ModelCard(
+        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-4bit"),
+        storage_size=Memory.from_mb(4423),
+        n_layers=32,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.1-8b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-8bit"),
+        storage_size=Memory.from_mb(8540),
+        n_layers=32,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.1-8b-bf16": ModelCard(
+        model_id=ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"),
+        storage_size=Memory.from_mb(16100),
+        n_layers=32,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.1-70b": ModelCard(
+        model_id=ModelId("mlx-community/Meta-Llama-3.1-70B-Instruct-4bit"),
+        storage_size=Memory.from_mb(38769),
+        n_layers=80,
+        hidden_size=8192,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # llama-3.2
+    "llama-3.2-1b": ModelCard(
+        model_id=ModelId("mlx-community/Llama-3.2-1B-Instruct-4bit"),
+        storage_size=Memory.from_mb(696),
+        n_layers=16,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.2-3b": ModelCard(
+        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-4bit"),
+        storage_size=Memory.from_mb(1777),
+        n_layers=28,
+        hidden_size=3072,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.2-3b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Llama-3.2-3B-Instruct-8bit"),
+        storage_size=Memory.from_mb(3339),
+        n_layers=28,
+        hidden_size=3072,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # llama-3.3
+    "llama-3.3-70b": ModelCard(
+        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-4bit"),
+        storage_size=Memory.from_mb(38769),
+        n_layers=80,
+        hidden_size=8192,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.3-70b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Llama-3.3-70B-Instruct-8bit"),
+        storage_size=Memory.from_mb(73242),
+        n_layers=80,
+        hidden_size=8192,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "llama-3.3-70b-fp16": ModelCard(
+        model_id=ModelId("mlx-community/llama-3.3-70b-instruct-fp16"),
+        storage_size=Memory.from_mb(137695),
+        n_layers=80,
+        hidden_size=8192,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # qwen3
+    "qwen3-0.6b": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-0.6B-4bit"),
+        storage_size=Memory.from_mb(327),
+        n_layers=28,
+        hidden_size=1024,
+        supports_tensor=False,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-0.6b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-0.6B-8bit"),
+        storage_size=Memory.from_mb(666),
+        n_layers=28,
+        hidden_size=1024,
+        supports_tensor=False,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-30b": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-30B-A3B-4bit"),
+        storage_size=Memory.from_mb(16797),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-30b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-30B-A3B-8bit"),
+        storage_size=Memory.from_mb(31738),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-80b-a3B-4bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit"),
+        storage_size=Memory.from_mb(44800),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-80b-a3B-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Instruct-8bit"),
+        storage_size=Memory.from_mb(84700),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-80b-a3B-thinking-4bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit"),
+        storage_size=Memory.from_mb(84700),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-80b-a3B-thinking-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Next-80B-A3B-Thinking-8bit"),
+        storage_size=Memory.from_mb(84700),
+        n_layers=48,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-235b-a22b-4bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit"),
+        storage_size=Memory.from_gb(132),
+        n_layers=94,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-235b-a22b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-235B-A22B-Instruct-2507-8bit"),
+        storage_size=Memory.from_gb(250),
+        n_layers=94,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-coder-480b-a35b-4bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit"),
+        storage_size=Memory.from_gb(270),
+        n_layers=62,
+        hidden_size=6144,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "qwen3-coder-480b-a35b-8bit": ModelCard(
+        model_id=ModelId("mlx-community/Qwen3-Coder-480B-A35B-Instruct-8bit"),
+        storage_size=Memory.from_gb(540),
+        n_layers=62,
+        hidden_size=6144,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # gpt-oss
+    "gpt-oss-120b-MXFP4-Q8": ModelCard(
+        model_id=ModelId("mlx-community/gpt-oss-120b-MXFP4-Q8"),
+        storage_size=Memory.from_kb(68_996_301),
+        n_layers=36,
+        hidden_size=2880,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "gpt-oss-20b-MXFP4-Q8": ModelCard(
+        model_id=ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"),
+        storage_size=Memory.from_kb(11_744_051),
+        n_layers=24,
+        hidden_size=2880,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # glm 4.5
+    "glm-4.5-air-8bit": ModelCard(
+        # Needs to be quantized g32 or g16 to work with tensor parallel
+        model_id=ModelId("mlx-community/GLM-4.5-Air-8bit"),
+        storage_size=Memory.from_gb(114),
+        n_layers=46,
+        hidden_size=4096,
+        supports_tensor=False,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.5-air-bf16": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.5-Air-bf16"),
+        storage_size=Memory.from_gb(214),
+        n_layers=46,
+        hidden_size=4096,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # glm 4.7
+    "glm-4.7-4bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-4bit"),
+        storage_size=Memory.from_bytes(198556925568),
+        n_layers=91,
+        hidden_size=5120,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.7-6bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-6bit"),
+        storage_size=Memory.from_bytes(286737579648),
+        n_layers=91,
+        hidden_size=5120,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.7-8bit-gs32": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-8bit-gs32"),
+        storage_size=Memory.from_bytes(396963397248),
+        n_layers=91,
+        hidden_size=5120,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # glm 4.7 flash
+    "glm-4.7-flash-4bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-Flash-4bit"),
+        storage_size=Memory.from_gb(18),
+        n_layers=47,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.7-flash-5bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-Flash-5bit"),
+        storage_size=Memory.from_gb(21),
+        n_layers=47,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.7-flash-6bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-Flash-6bit"),
+        storage_size=Memory.from_gb(25),
+        n_layers=47,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "glm-4.7-flash-8bit": ModelCard(
+        model_id=ModelId("mlx-community/GLM-4.7-Flash-8bit"),
+        storage_size=Memory.from_gb(32),
+        n_layers=47,
+        hidden_size=2048,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    # minimax-m2
+    "minimax-m2.1-8bit": ModelCard(
+        model_id=ModelId("mlx-community/MiniMax-M2.1-8bit"),
+        storage_size=Memory.from_bytes(242986745856),
+        n_layers=61,
+        hidden_size=3072,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+    "minimax-m2.1-3bit": ModelCard(
+        model_id=ModelId("mlx-community/MiniMax-M2.1-3bit"),
+        storage_size=Memory.from_bytes(100086644736),
+        n_layers=61,
+        hidden_size=3072,
+        supports_tensor=True,
+        tasks=[ModelTask.TextGeneration],
+    ),
+}
+
+_IMAGE_BASE_MODEL_CARDS: dict[str, ModelCard] = {
+    "flux1-schnell": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-schnell"),
+        storage_size=Memory.from_bytes(23782357120 + 9524621312),
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23782357120),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "flux1-dev": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-dev"),
+        storage_size=Memory.from_bytes(23782357120 + 9524621312),
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23802816640),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "flux1-krea-dev": ModelCard(
+        model_id=ModelId("exolabs/FLUX.1-Krea-dev"),
+        storage_size=Memory.from_bytes(23802816640 + 9524621312),  # Same as dev
+        n_layers=57,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_kb(0),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="text_encoder_2",
+                component_path="text_encoder_2/",
+                storage_size=Memory.from_bytes(9524621312),
+                n_layers=24,
+                can_shard=False,
+                safetensors_index_filename="model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(23802816640),
+                n_layers=57,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "qwen-image": ModelCard(
+        model_id=ModelId("exolabs/Qwen-Image"),
+        storage_size=Memory.from_bytes(16584333312 + 40860802176),
+        n_layers=60,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.TextToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_bytes(16584333312),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(40860802176),
+                n_layers=60,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+    "qwen-image-edit-2509": ModelCard(
+        model_id=ModelId("exolabs/Qwen-Image-Edit-2509"),
+        storage_size=Memory.from_bytes(16584333312 + 40860802176),
+        n_layers=60,
+        hidden_size=1,
+        supports_tensor=False,
+        tasks=[ModelTask.ImageToImage],
+        components=[
+            ComponentInfo(
+                component_name="text_encoder",
+                component_path="text_encoder/",
+                storage_size=Memory.from_bytes(16584333312),
+                n_layers=12,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+            ComponentInfo(
+                component_name="transformer",
+                component_path="transformer/",
+                storage_size=Memory.from_bytes(40860802176),
+                n_layers=60,
+                can_shard=True,
+                safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
+            ),
+            ComponentInfo(
+                component_name="vae",
+                component_path="vae/",
+                storage_size=Memory.from_kb(0),
+                n_layers=None,
+                can_shard=False,
+                safetensors_index_filename=None,
+            ),
+        ],
+    ),
+}
+
+
+def _generate_image_model_quant_variants(
     base_name: str,
     base_card: ModelCard,
 ) -> dict[str, ModelCard]:
@@ -210,6 +706,15 @@ def _generate_image_model_quant_variants(  # pyright: ignore[reportUnusedFunctio
     return variants
 
 
+_image_model_cards: dict[str, ModelCard] = {}
+for _base_name, _base_card in _IMAGE_BASE_MODEL_CARDS.items():
+    _image_model_cards |= _generate_image_model_quant_variants(_base_name, _base_card)
+_IMAGE_MODEL_CARDS = _image_model_cards
+
+if EXO_ENABLE_IMAGE_MODELS:
+    MODEL_CARDS.update(_IMAGE_MODEL_CARDS)
+
+
 class ConfigData(BaseModel):
     model_config = {"extra": "ignore"}  # Allow unknown fields
 

src/exo/shared/tests/test_node_id_persistence.py

@@ -8,7 +8,7 @@ from multiprocessing.synchronize import Event as EventT
 from multiprocessing.synchronize import Semaphore as SemaphoreT
 
 from loguru import logger
-from pytest import LogCaptureFixture
+from pytest import LogCaptureFixture, mark
 
 from exo.routing.router import get_node_id_keypair
 from exo.shared.constants import EXO_NODE_ID_KEYPAIR
@@ -74,6 +74,7 @@ def _delete_if_exists(p: str | bytes | os.PathLike[str] | os.PathLike[bytes]):
         os.remove(p)
 
 
+@mark.skip(reason="this functionality is currently disabled but may return in future")
 def test_node_id_fetching(caplog: LogCaptureFixture):
     reps = 10
 

src/exo/shared/tests/test_state_serialization.py

@@ -16,6 +16,9 @@ def test_state_serialization_roundtrip() -> None:
         sink=node_b,
         edge=SocketConnection(
             sink_multiaddr=Multiaddr(address="/ip4/127.0.0.1/tcp/10001"),
+            latency_ms=1.5,
+            sink_to_other_bandwidth_mbps=1000.0,
+            other_to_sink_bandwidth_mbps=1100.0,
         ),
     )
 

src/exo/shared/types/api.py

@@ -11,7 +11,7 @@ from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.memory import Memory
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding, ShardMetadata
-from exo.utils.pydantic_ext import CamelCaseModel
+from exo.utils.pydantic_ext import CamelCaseModel, ConfigDict, TaggedModel
 
 FinishReason = Literal[
     "stop", "length", "tool_calls", "content_filter", "function_call", "error"
@@ -116,8 +116,8 @@ class Usage(BaseModel):
     prompt_tokens: int
     completion_tokens: int
     total_tokens: int
-    prompt_tokens_details: PromptTokensDetails | None = None
-    completion_tokens_details: CompletionTokensDetails | None = None
+    prompt_tokens_details: PromptTokensDetails
+    completion_tokens_details: CompletionTokensDetails
 
 
 class StreamingChoiceResponse(BaseModel):
@@ -170,7 +170,13 @@ class BenchChatCompletionResponse(ChatCompletionResponse):
     generation_stats: GenerationStats | None = None
 
 
-class ChatCompletionTaskParams(BaseModel):
+class StreamOptions(BaseModel):
+    include_usage: bool = False
+
+
+class ChatCompletionTaskParams(TaggedModel):
+    model_config = ConfigDict(extra="ignore")
+
     model: str
     frequency_penalty: float | None = None
     messages: list[ChatCompletionMessage]
@@ -184,6 +190,7 @@ class ChatCompletionTaskParams(BaseModel):
     seed: int | None = None
     stop: str | list[str] | None = None
     stream: bool = False
+    stream_options: StreamOptions | None = None
     temperature: float | None = None
     top_p: float | None = None
     tools: list[dict[str, Any]] | None = None

src/exo/shared/types/chunks.py

@@ -2,7 +2,7 @@ from collections.abc import Generator
 from typing import Any, Literal
 
 from exo.shared.models.model_cards import ModelId
-from exo.shared.types.api import GenerationStats, ImageGenerationStats
+from exo.shared.types.api import GenerationStats, ImageGenerationStats, Usage
 from exo.utils.pydantic_ext import TaggedModel
 
 from .api import FinishReason
@@ -17,6 +17,7 @@ class BaseChunk(TaggedModel):
 class TokenChunk(BaseChunk):
     text: str
     token_id: int
+    usage: Usage | None
     finish_reason: Literal["stop", "length", "content_filter"] | None = None
     stats: GenerationStats | None = None
 
@@ -28,6 +29,7 @@ class ErrorChunk(BaseChunk):
 
 class ToolCallChunk(BaseChunk):
     tool_calls: list[ToolCallItem]
+    usage: Usage | None
     finish_reason: Literal["tool_calls"] = "tool_calls"
     stats: GenerationStats | None = None
 

src/exo/shared/types/commands.py

@@ -2,6 +2,7 @@ from pydantic import Field
 
 from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.api import (
+    BenchChatCompletionTaskParams,
     ChatCompletionTaskParams,
     ImageEditsInternalParams,
     ImageGenerationTaskParams,
@@ -22,7 +23,7 @@ class TestCommand(BaseCommand):
 
 
 class ChatCompletion(BaseCommand):
-    request_params: ChatCompletionTaskParams
+    request_params: ChatCompletionTaskParams | BenchChatCompletionTaskParams
 
 
 class ImageGeneration(BaseCommand):

src/exo/shared/types/tasks.py

@@ -3,6 +3,7 @@ from enum import Enum
 from pydantic import Field
 
 from exo.shared.types.api import (
+    BenchChatCompletionTaskParams,
     ChatCompletionTaskParams,
     ImageEditsInternalParams,
     ImageGenerationTaskParams,
@@ -54,7 +55,7 @@ class StartWarmup(BaseTask):  # emitted by Worker
 
 class ChatCompletion(BaseTask):  # emitted by Master
     command_id: CommandId
-    task_params: ChatCompletionTaskParams
+    task_params: ChatCompletionTaskParams | BenchChatCompletionTaskParams
 
     error_type: str | None = Field(default=None)
     error_message: str | None = Field(default=None)

src/exo/shared/types/topology.py

@@ -24,10 +24,22 @@ class RDMAConnection(FrozenModel):
 
 class SocketConnection(FrozenModel):
     sink_multiaddr: Multiaddr
+    latency_ms: float
+    sink_to_other_bandwidth_mbps: float
+    other_to_sink_bandwidth_mbps: float
+
+    @property
+    def bandwidth_mbps(self):
+        return min(self.sink_to_other_bandwidth_mbps, self.other_to_sink_bandwidth_mbps)
 
     def __hash__(self):
         return hash(self.sink_multiaddr.ip_address)
 
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, SocketConnection):
+            return NotImplemented
+        return self.sink_multiaddr == other.sink_multiaddr
+
 
 class Connection(FrozenModel):
     source: NodeId

src/exo/shared/types/worker/runner_response.py

@@ -6,6 +6,7 @@ from exo.shared.types.api import (
     GenerationStats,
     ImageGenerationStats,
     ToolCallItem,
+    Usage,
 )
 from exo.utils.pydantic_ext import TaggedModel
 
@@ -24,6 +25,7 @@ class GenerationResponse(BaseRunnerResponse):
     # logprobs: list[float] | None = None # too big. we can change to be top-k
     finish_reason: FinishReason | None = None
     stats: GenerationStats | None = None
+    usage: Usage | None
 
 
 class ImageGenerationResponse(BaseRunnerResponse):
@@ -57,6 +59,7 @@ class PartialImageResponse(BaseRunnerResponse):
 
 class ToolCallResponse(BaseRunnerResponse):
     tool_calls: list[ToolCallItem]
+    usage: Usage | None
 
 
 class FinishedResponse(BaseRunnerResponse):

src/exo/utils/dashboard_path.py

@@ -1,45 +1,31 @@
+import os
 import sys
 from pathlib import Path
 from typing import cast
 
 
-def find_resources() -> Path:
-    resources = _find_resources_in_repo() or _find_resources_in_bundle()
-    if resources is None:
-        raise FileNotFoundError(
-            "Unable to locate resources. Did you clone the repo properly?"
-        )
-    return resources
-
-
-def _find_resources_in_repo() -> Path | None:
-    current_module = Path(__file__).resolve()
-    for parent in current_module.parents:
-        build = parent / "resources"
-        if build.is_dir():
-            return build
-    return None
-
-
-def _find_resources_in_bundle() -> Path | None:
-    frozen_root = cast(str | None, getattr(sys, "_MEIPASS", None))
-    if frozen_root is None:
-        return None
-    candidate = Path(frozen_root) / "resources"
-    if candidate.is_dir():
-        return candidate
-    return None
-
-
 def find_dashboard() -> Path:
-    dashboard = _find_dashboard_in_repo() or _find_dashboard_in_bundle()
+    dashboard = (
+        _find_dashboard_in_env()
+        or _find_dashboard_in_repo()
+        or _find_dashboard_in_bundle()
+    )
     if not dashboard:
         raise FileNotFoundError(
-            "Unable to locate dashboard assets - you probably forgot to run `cd dashboard && npm install && npm run build && cd ..`"
+            "Unable to locate dashboard assets - make sure the dashboard has been built, or export DASHBOARD_DIR if you've built the dashboard elsewhere."
         )
     return dashboard
 
 
+def _find_dashboard_in_env() -> Path | None:
+    env = os.environ.get("DASHBOARD_DIR")
+    if not env:
+        return None
+    resolved_env = Path(env).expanduser().resolve()
+
+    return resolved_env
+
+
 def _find_dashboard_in_repo() -> Path | None:
     current_module = Path(__file__).resolve()
     for parent in current_module.parents:

src/exo/utils/info_gatherer/connection_profiler.py

@@ -0,0 +1,125 @@
+import time
+
+import anyio
+import httpx
+from anyio.abc import SocketStream
+from exo.shared.logging import logger
+from pydantic.v1 import BaseModel
+
+LATENCY_PING_COUNT = 5
+BANDWIDTH_TEST_DURATION_S = 0.5
+BANDWIDTH_TEST_PORT_OFFSET = 1  # API port + 1
+
+
+class ConnectionProfile(BaseModel):
+    latency_ms: float
+    upload_mbps: float
+    download_mbps: float
+
+
+async def measure_latency(target_ip: str, port: int = 52415) -> float:
+    if ":" in target_ip:
+        url = f"http://[{target_ip}]:{port}/node_id"
+    else:
+        url = f"http://{target_ip}:{port}/node_id"
+
+    rtts: list[float] = []
+
+    async with httpx.AsyncClient(timeout=10.0) as client:
+        for _ in range(LATENCY_PING_COUNT):
+            try:
+                start = time.perf_counter()
+                response = await client.get(url)
+                end = time.perf_counter()
+
+                if response.status_code == 200:
+                    rtts.append((end - start) * 1000)
+            except (httpx.TimeoutException, httpx.NetworkError, httpx.RemoteProtocolError) as e:
+                logger.debug(f"Latency ping failed: {e}")
+
+    if not rtts:
+        raise ConnectionError(f"Failed to measure latency to {target_ip}:{port}")
+
+    return sum(rtts) / len(rtts)
+
+
+async def _measure_upload_tcp(stream: SocketStream, duration: float) -> float:
+    """Send data for duration seconds, return Mbps."""
+    chunk = b"X" * (1024 * 1024)  # 1MB
+    bytes_sent = 0
+    start = time.perf_counter()
+    deadline = start + duration
+
+    while time.perf_counter() < deadline:
+        await stream.send(chunk)
+        bytes_sent += len(chunk)
+
+    elapsed = time.perf_counter() - start
+    return (bytes_sent * 8 / elapsed) / 1_000_000 if elapsed > 0 else 0.0
+
+
+async def _measure_download_tcp(stream: SocketStream, duration: float) -> float:
+    """Receive data for duration seconds, return Mbps."""
+    bytes_received = 0
+    start = time.perf_counter()
+
+    with anyio.move_on_after(duration):
+        while True:
+            data = await stream.receive(1024 * 1024)
+            if not data:
+                break
+            bytes_received += len(data)
+
+    elapsed = time.perf_counter() - start
+    return (bytes_received * 8 / elapsed) / 1_000_000 if elapsed > 0 else 0.0
+
+
+async def measure_bandwidth_tcp(target_ip: str, port: int) -> tuple[float, float]:
+    """Measure bandwidth using raw TCP like iperf."""
+    upload_mbps = 0.0
+    download_mbps = 0.0
+
+    try:
+        async with await anyio.connect_tcp(target_ip, port) as stream:
+            # Protocol: send 'U' for upload test, 'D' for download test
+            # Upload: client sends, server receives
+            await stream.send(b"U")
+            upload_mbps = await _measure_upload_tcp(stream, BANDWIDTH_TEST_DURATION_S)
+            await stream.send(b"DONE")
+            logger.debug(f"Upload: {upload_mbps:.1f} Mbps")
+    except Exception as e:
+        logger.debug(f"Upload TCP test failed: {e}")
+
+    try:
+        async with await anyio.connect_tcp(target_ip, port) as stream:
+            # Download: client receives, server sends
+            await stream.send(b"D")
+            download_mbps = await _measure_download_tcp(stream, BANDWIDTH_TEST_DURATION_S)
+            logger.debug(f"Download: {download_mbps:.1f} Mbps")
+    except Exception as e:
+        logger.debug(f"Download TCP test failed: {e}")
+
+    return upload_mbps, download_mbps
+
+
+async def profile_connection(target_ip: str, port: int = 52415) -> ConnectionProfile:
+    logger.debug(f"Profiling connection to {target_ip}:{port}")
+
+    latency_ms = await measure_latency(target_ip, port)
+    logger.debug(f"Measured latency to {target_ip}: {latency_ms:.2f}ms")
+
+    bandwidth_port = port + BANDWIDTH_TEST_PORT_OFFSET
+    upload_mbps, download_mbps = await measure_bandwidth_tcp(target_ip, bandwidth_port)
+    logger.debug(
+        f"Measured bandwidth to {target_ip}: "
+        f"upload={upload_mbps:.1f}Mbps, download={download_mbps:.1f}Mbps"
+    )
+
+    if upload_mbps == 0.0 and download_mbps == 0.0:
+        raise ConnectionError(f"Failed to measure bandwidth to {target_ip}:{bandwidth_port}")
+
+    return ConnectionProfile(
+        latency_ms=latency_ms,
+        upload_mbps=upload_mbps,
+        download_mbps=download_mbps,
+    )

src/exo/worker/engines/image/generate.py

@@ -98,8 +98,8 @@ def generate_image(
 
     partial_images = (
         task.partial_images
-        if task.partial_images is not None
-        else (3 if task.stream else 0)
+        if task.partial_images is not None and task.stream is not None and task.stream
+        else 0
     )
 
     image_path: Path | None = None

src/exo/worker/engines/image/pipeline/runner.py

@@ -348,6 +348,7 @@ class DiffusionRunner:
                 ctx.in_loop(  # pyright: ignore[reportAny]
                     t=t,
                     latents=latents,
+                    time_steps=time_steps,
                 )
 
                 mx.eval(latents)

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

@@ -201,6 +201,9 @@ def pipeline_auto_parallel(
     device_rank, world_size = model_shard_meta.device_rank, model_shard_meta.world_size
 
     layers = layers[start_layer:end_layer]
+    for layer in layers:
+        mx.eval(layer)  # type: ignore
+
     layers[0] = PipelineFirstLayer(layers[0], device_rank, group=group)
     layers[-1] = PipelineLastLayer(
         layers[-1],

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

@@ -3,6 +3,7 @@ from copy import deepcopy
 from typing import Any, cast
 
 import mlx.core as mx
+import psutil
 from mlx_lm.models.cache import (
     KVCache,
     QuantizedKVCache,
@@ -12,25 +13,29 @@ from mlx_lm.models.cache import (
 from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
+from exo.shared.types.memory import Memory
 from exo.shared.types.mlx import KVCacheType
 from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.constants import CACHE_GROUP_SIZE, KV_CACHE_BITS
 from exo.worker.runner.bootstrap import logger
 
 # Fraction of device memory above which LRU eviction kicks in
-_DEFAULT_MEMORY_THRESHOLD = 0.85
+_DEFAULT_MEMORY_THRESHOLD = 0.9
 _MEMORY_THRESHOLD = float(
     os.environ.get("EXO_MEMORY_THRESHOLD", _DEFAULT_MEMORY_THRESHOLD)
 )
 
 
 class KVPrefixCache:
-    def __init__(self, tokenizer: TokenizerWrapper):
+    def __init__(
+        self, tokenizer: TokenizerWrapper, group: mx.distributed.Group | None = None
+    ):
         self.prompts: list[mx.array] = []  # mx array of tokens (ints)
         self.caches: list[KVCacheType] = []
         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."""
@@ -81,13 +86,13 @@ class KVPrefixCache:
         best_snapshot_index, best_snapshot_length = None, 0
 
         for i, cached_prompt in enumerate(self.prompts):
-            length = _get_prefix_length(tokenized_prompt, cached_prompt)
+            length = get_prefix_length(tokenized_prompt, cached_prompt)
 
             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])
+                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)
@@ -109,7 +114,7 @@ class KVPrefixCache:
             prompt_cache = deepcopy(self.caches[best_snapshot_index])
 
             # 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])
+            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)
@@ -131,29 +136,37 @@ class KVPrefixCache:
             return prompt_cache, tokenized_prompt, None
 
     def _evict_if_needed(self):
-        """Evict least recently used entries while memory pressure is high."""
+        """Evict least recently used entries while memory usage is high."""
         if len(self.caches) == 0:
             return
 
-        active: int = mx.metal.get_active_memory()
-        limit = int(mx.metal.device_info()["max_recommended_working_set_size"])
-        if active < limit * _MEMORY_THRESHOLD:
-            return
-
         # Evict LRU entries until below threshold or only one entry left
-        while len(self.caches) > 0:
+        while (
+            len(self.caches) > 1
+            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._last_used.pop(lru_index)
             logger.info(
-                f"KV cache evicted LRU entry ({evicted_tokens} tokens) due to memory pressure"
+                f"KV cache evicted LRU entry ({evicted_tokens} tokens) due to memory usage"
             )
 
-            active = mx.metal.get_active_memory()
-            if active < limit * _MEMORY_THRESHOLD:
-                break
+    def get_memory_used_percentage(self) -> float:
+        local_pressure: float = get_memory_used_percentage()
+
+        if self._group is None:
+            return local_pressure
+
+        all_pressure = mx.distributed.all_gather(
+            mx.array([local_pressure], dtype=mx.float32),
+            group=self._group,
+        )
+        # .item() evals.
+        max_pressure = float(mx.max(all_pressure).item())
+        return max_pressure
 
 
 def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
@@ -168,13 +181,13 @@ def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
     return mx.array(tokenized_prompt)
 
 
-def _cache_length(cache: KVCacheType) -> int:
+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
 
 
-def _get_prefix_length(prompt: mx.array, cached_prompt: mx.array) -> int:
+def get_prefix_length(prompt: mx.array, cached_prompt: mx.array) -> int:
     """Find the length of the common prefix between two token arrays."""
     n = min(int(prompt.shape[0]), int(cached_prompt.shape[0]))
     if n == 0:
@@ -185,6 +198,17 @@ def _get_prefix_length(prompt: mx.array, cached_prompt: mx.array) -> int:
     return int(mx.sum(prefix_mask).item())
 
 
+def get_available_memory() -> Memory:
+    mem: int = psutil.virtual_memory().available
+    return Memory.from_bytes(mem)
+
+
+def get_memory_used_percentage() -> float:
+    mem = psutil.virtual_memory()
+    # percent is 0-100
+    return float(mem.percent / 100)
+
+
 def make_kv_cache(
     model: Model, max_kv_size: int | None = None, keep: int = 0
 ) -> KVCacheType:

src/exo/worker/engines/mlx/generator/generate.py

@@ -10,8 +10,11 @@ from mlx_lm.tokenizer_utils import TokenizerWrapper
 from exo.shared.types.api import (
     BenchChatCompletionTaskParams,
     ChatCompletionMessage,
+    CompletionTokensDetails,
     FinishReason,
     GenerationStats,
+    PromptTokensDetails,
+    Usage,
 )
 from exo.shared.types.memory import Memory
 from exo.shared.types.mlx import KVCacheType
@@ -39,7 +42,7 @@ def prefill(
     sampler: Callable[[mx.array], mx.array],
     prompt_tokens: mx.array,
     cache: KVCacheType,
-) -> float:
+) -> tuple[float, int]:
     """Prefill the KV cache with prompt tokens.
 
     This runs the model over the prompt tokens to populate the cache,
@@ -50,7 +53,7 @@ def prefill(
     """
     num_tokens = len(prompt_tokens)
     if num_tokens == 0:
-        return 0.0
+        return 0.0, 0
 
     logger.debug(f"Prefilling {num_tokens} tokens...")
     start_time = time.perf_counter()
@@ -85,7 +88,7 @@ def prefill(
         f"Prefill complete: {num_tokens} tokens in {elapsed:.2f}s "
         f"({tokens_per_sec:.1f} tok/s)"
     )
-    return tokens_per_sec
+    return tokens_per_sec, num_tokens
 
 
 def warmup_inference(
@@ -169,6 +172,8 @@ def mlx_generate(
     mx.reset_peak_memory()
     is_bench: bool = isinstance(task, BenchChatCompletionTaskParams)
 
+    logger.info(f"{is_bench=}")
+
     # Currently we support chat-completion tasks only.
     logger.debug(f"task_params: {task}")
 
@@ -204,7 +209,9 @@ def mlx_generate(
     )
 
     # Prefill cache with all tokens except the last one
-    prefill_tps = prefill(model, tokenizer, sampler, prompt_tokens[:-1], caches)
+    prefill_tps, prefill_tokens = prefill(
+        model, tokenizer, sampler, prompt_tokens[:-1], caches
+    )
 
     # stream_generate starts from the last token
     last_token = prompt_tokens[-1:]
@@ -212,28 +219,43 @@ def mlx_generate(
     max_tokens = task.max_tokens or MAX_TOKENS
     generated_text_parts: list[str] = []
     generation_start_time = time.perf_counter()
-    for out in stream_generate(
-        model=model,
-        tokenizer=tokenizer,
-        prompt=last_token,
-        max_tokens=max_tokens,
-        sampler=sampler,
-        logits_processors=logits_processors,
-        prompt_cache=caches,
-        # TODO: Dynamically change prefill step size to be the maximum possible without timing out.
-        prefill_step_size=2048,
-        kv_group_size=KV_GROUP_SIZE,
-        kv_bits=KV_BITS,
+    usage: Usage | None = None
+    in_thinking = False
+    reasoning_tokens = 0
+    think_start = tokenizer.think_start
+    think_end = tokenizer.think_end
+    for completion_tokens, out in enumerate(
+        stream_generate(
+            model=model,
+            tokenizer=tokenizer,
+            prompt=last_token,
+            max_tokens=max_tokens,
+            sampler=sampler,
+            logits_processors=logits_processors,
+            prompt_cache=caches,
+            # TODO: Dynamically change prefill step size to be the maximum possible without timing out.
+            prefill_step_size=2048,
+            kv_group_size=KV_GROUP_SIZE,
+            kv_bits=KV_BITS,
+        ),
+        start=1,
     ):
         generated_text_parts.append(out.text)
         logger.info(out.text)
 
+        if think_start is not None and out.text == think_start:
+            in_thinking = True
+        elif think_end is not None and out.text == think_end:
+            in_thinking = False
+        if in_thinking:
+            reasoning_tokens += 1
+
         stats: GenerationStats | None = None
         if out.finish_reason is not None:
             stats = GenerationStats(
                 prompt_tps=float(prefill_tps or out.prompt_tps),
                 generation_tps=float(out.generation_tps),
-                prompt_tokens=int(out.prompt_tokens),
+                prompt_tokens=int(prefill_tokens + out.prompt_tokens),
                 generation_tokens=int(out.generation_tokens),
                 peak_memory_usage=Memory.from_gb(out.peak_memory),
             )
@@ -245,11 +267,24 @@ def mlx_generate(
                     f"Model generated unexpected finish_reason: {out.finish_reason}"
                 )
 
+            usage = Usage(
+                prompt_tokens=int(out.prompt_tokens),
+                completion_tokens=completion_tokens,
+                total_tokens=int(out.prompt_tokens) + completion_tokens,
+                prompt_tokens_details=PromptTokensDetails(
+                    cached_tokens=prefix_hit_length
+                ),
+                completion_tokens_details=CompletionTokensDetails(
+                    reasoning_tokens=reasoning_tokens
+                ),
+            )
+
         yield GenerationResponse(
             text=out.text,
             token=out.token,
             finish_reason=cast(FinishReason | None, out.finish_reason),
             stats=stats,
+            usage=usage,
         )
 
         if out.finish_reason is not None:

src/exo/worker/main.py

@@ -44,6 +44,7 @@ from exo.shared.types.topology import Connection, SocketConnection
 from exo.shared.types.worker.runners import RunnerId
 from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.event_buffer import OrderedBuffer
+from exo.utils.info_gatherer.connection_profiler import profile_connection
 from exo.utils.info_gatherer.info_gatherer import GatheredInfo, InfoGatherer
 from exo.utils.info_gatherer.net_profile import check_reachable
 from exo.utils.keyed_backoff import KeyedBackoff
@@ -65,6 +66,7 @@ class Worker:
         command_sender: Sender[ForwarderCommand],
         download_command_sender: Sender[ForwarderDownloadCommand],
         event_index_counter: Iterator[int],
+        profiling_in_progress: set[NodeId] | None = None,
     ):
         self.node_id: NodeId = node_id
         self.session_id: SessionId = session_id
@@ -81,6 +83,8 @@ class Worker:
         self.state: State = State()
         self.runners: dict[RunnerId, RunnerSupervisor] = {}
         self._tg: TaskGroup = create_task_group()
+        self._profiling_in_progress: set[NodeId] = profiling_in_progress if profiling_in_progress is not None else set()
+        self._profiling_lock = anyio.Lock()
 
         self._nack_cancel_scope: CancelScope | None = None
         self._nack_attempts: int = 0
@@ -282,37 +286,73 @@ class Worker:
     async def _connection_message_event_writer(self):
         with self.connection_message_receiver as connection_messages:
             async for msg in connection_messages:
-                await self.event_sender.send(
-                    self._convert_connection_message_to_event(msg)
-                )
+                event = await self._convert_connection_message_to_event(msg)
+                if event:
+                    await self.event_sender.send(event)
 
-    def _convert_connection_message_to_event(self, msg: ConnectionMessage):
-        match msg.connection_type:
-            case ConnectionMessageType.Connected:
-                return TopologyEdgeCreated(
-                    conn=Connection(
-                        source=self.node_id,
-                        sink=msg.node_id,
-                        edge=SocketConnection(
-                            sink_multiaddr=Multiaddr(
-                                address=f"/ip4/{msg.remote_ipv4}/tcp/{msg.remote_tcp_port}"
-                            ),
-                        ),
-                    ),
-                )
+    async def _convert_connection_message_to_event(self, msg: ConnectionMessage):
+        if msg.connection_type == ConnectionMessageType.Connected:
+            async with self._profiling_lock:
+                if msg.node_id in self._profiling_in_progress:
+                    return None
+                self._profiling_in_progress.add(msg.node_id)
+            return await self._profile_and_emit_connection(
+                msg.node_id,
+                msg.remote_ipv4,
+                msg.remote_tcp_port,
+            )
+        elif msg.connection_type == ConnectionMessageType.Disconnected:
+            self._profiling_in_progress.discard(msg.node_id)
+            target_ip = msg.remote_ipv4
+            for connection in self.state.topology.list_connections():
+                if (
+                    isinstance(connection.edge, SocketConnection)
+                    and connection.edge.sink_multiaddr.ip_address == target_ip
+                ):
+                    return TopologyEdgeDeleted(conn=connection)
 
-            case ConnectionMessageType.Disconnected:
-                return TopologyEdgeDeleted(
-                    conn=Connection(
-                        source=self.node_id,
-                        sink=msg.node_id,
-                        edge=SocketConnection(
-                            sink_multiaddr=Multiaddr(
-                                address=f"/ip4/{msg.remote_ipv4}/tcp/{msg.remote_tcp_port}"
-                            ),
-                        ),
+    async def _profile_and_emit_connection(
+        self,
+        sink_node_id: NodeId,
+        remote_ip: str,
+        remote_port: int,
+    ):
+        try:
+            profile = await profile_connection(remote_ip)
+        except ConnectionError as e:
+            logger.warning(f"Failed to profile connection to {sink_node_id}: {e}")
+            profile = None
+
+        if profile:
+            latency_ms = profile.latency_ms
+            other_to_sink_mbps = profile.upload_mbps
+            sink_to_other_mbps = profile.download_mbps
+        else:
+            latency_ms = 0.0
+            other_to_sink_mbps = 0.0
+            sink_to_other_mbps = 0.0
+
+        logger.info(
+            f"Connection to {sink_node_id} profiled: "
+            f"latency={latency_ms:.2f}ms, "
+            f"upload={other_to_sink_mbps:.1f}Mbps, "
+            f"download={sink_to_other_mbps:.1f}Mbps"
+        )
+
+        return TopologyEdgeCreated(
+            conn=Connection(
+                source=self.node_id,
+                sink=sink_node_id,
+                edge=SocketConnection(
+                    sink_multiaddr=Multiaddr(
+                        address=f"/ip4/{remote_ip}/tcp/{remote_port}"
                     ),
-                )
+                    latency_ms=latency_ms,
+                    sink_to_other_bandwidth_mbps=sink_to_other_mbps,
+                    other_to_sink_bandwidth_mbps=other_to_sink_mbps,
+                ),
+            ),
+        )
 
     async def _nack_request(self, since_idx: int) -> None:
         # We request all events after (and including) the missing index.
@@ -389,21 +429,21 @@ class Worker:
             )
             for nid in conns:
                 for ip in conns[nid]:
-                    edge = SocketConnection(
-                        # nonsense multiaddr
+                    temp_edge = SocketConnection(
                         sink_multiaddr=Multiaddr(address=f"/ip4/{ip}/tcp/52415")
                         if "." in ip
-                        # nonsense multiaddr
                         else Multiaddr(address=f"/ip6/{ip}/tcp/52415"),
+                        latency_ms=0.0,
+                        sink_to_other_bandwidth_mbps=0.0,
+                        other_to_sink_bandwidth_mbps=0.0,
                     )
-                    if edge not in edges:
-                        logger.debug(f"ping discovered {edge=}")
-                        await self.event_sender.send(
-                            TopologyEdgeCreated(
-                                conn=Connection(
-                                    source=self.node_id, sink=nid, edge=edge
-                                )
-                            )
+                    if temp_edge not in edges:
+                        logger.debug(f"ping discovered new connection to {nid} at {ip}")
+                        self._tg.start_soon(
+                            self._profile_and_emit_connection,
+                            nid,
+                            ip,
+                            52415,
                         )
 
             for conn in self.state.topology.out_edges(self.node_id):

src/exo/worker/runner/runner.py

@@ -37,6 +37,7 @@ from exo.shared.types.tasks import (
     Shutdown,
     StartWarmup,
     Task,
+    TaskId,
     TaskStatus,
 )
 from exo.shared.types.worker.instances import BoundInstance
@@ -111,8 +112,12 @@ def main(
     event_sender.send(
         RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
     )
+    seen = set[TaskId]()
     with task_receiver as tasks:
         for task in tasks:
+            if task.task_id in seen:
+                logger.warning("repeat task - potential error")
+            seen.add(task.task_id)
             event_sender.send(
                 TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Running)
             )
@@ -163,7 +168,7 @@ def main(
                         logger.info(
                             f"model has_tool_calling={tokenizer.has_tool_calling}"
                         )
-                        kv_prefix_cache = KVPrefixCache(tokenizer)
+                        kv_prefix_cache = KVPrefixCache(tokenizer, group)
 
                     elif (
                         ModelTask.TextToImage in shard_metadata.model_card.tasks
@@ -277,9 +282,11 @@ def main(
                                 tokenizer.tool_parser,  # pyright: ignore[reportAny]
                             )
 
+                        completion_tokens = 0
                         for response in mlx_generator:
                             match response:
                                 case GenerationResponse():
+                                    completion_tokens += 1
                                     if (
                                         device_rank == 0
                                         and response.finish_reason == "error"
@@ -307,6 +314,7 @@ def main(
                                                     model=shard_metadata.model_card.model_id,
                                                     text=response.text,
                                                     token_id=response.token,
+                                                    usage=response.usage,
                                                     finish_reason=response.finish_reason,
                                                     stats=response.stats,
                                                 ),
@@ -320,6 +328,7 @@ def main(
                                                 chunk=ToolCallChunk(
                                                     tool_calls=response.tool_calls,
                                                     model=shard_metadata.model_card.model_id,
+                                                    usage=response.usage,
                                                 ),
                                             )
                                         )
@@ -535,10 +544,10 @@ def parse_gpt_oss(
                             name=current_tool_name,
                             arguments="".join(tool_arg_parts).strip(),
                         )
-                    ]
+                    ],
+                    usage=response.usage,
                 )
                 tool_arg_parts = []
-                break
             current_tool_name = recipient
 
         # If inside a tool call, accumulate arguments
@@ -684,7 +693,7 @@ def parse_tool_calls(
                     tools = [_validate_single_tool(tool) for tool in parsed]
                 else:
                     tools = [_validate_single_tool(parsed)]
-                yield ToolCallResponse(tool_calls=tools)
+                yield ToolCallResponse(tool_calls=tools, usage=response.usage)
 
             except (
                 json.JSONDecodeError,

src/exo/worker/runner/runner_supervisor.py

@@ -127,20 +127,25 @@ class RunnerSupervisor:
         self._tg.cancel_scope.cancel()
 
     async def start_task(self, task: Task):
+        if task.task_id in self.pending:
+            logger.warning(
+                f"Skipping invalid task {task} as it has already been submitted"
+            )
+            return
         if task.task_id in self.completed:
-            logger.info(
+            logger.warning(
                 f"Skipping invalid task {task} as it has already been completed"
             )
+            return
         logger.info(f"Starting task {task}")
         event = anyio.Event()
         self.pending[task.task_id] = event
         try:
-            self._task_sender.send(task)
+            await self._task_sender.send_async(task)
         except ClosedResourceError:
             logger.warning(f"Task {task} dropped, runner closed communication.")
             return
         await event.wait()
-        logger.info(f"Finished task {task}")
 
     async def _forward_events(self):
         with self._ev_recv as events:

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

@@ -14,9 +14,9 @@ from exo.shared.types.tasks import ChatCompletionTaskParams
 from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.cache import (
     KVPrefixCache,
-    _cache_length,
-    _get_prefix_length,
+    cache_length,
     encode_prompt,
+    get_prefix_length,
     make_kv_cache,
 )
 from exo.worker.engines.mlx.generator.generate import mlx_generate, prefill
@@ -35,47 +35,47 @@ class TestGetPrefixLength:
     def test_identical_arrays(self):
         a = mx.array([1, 2, 3, 4, 5])
         b = mx.array([1, 2, 3, 4, 5])
-        assert _get_prefix_length(a, b) == 5
+        assert get_prefix_length(a, b) == 5
 
     def test_no_common_prefix(self):
         a = mx.array([1, 2, 3])
         b = mx.array([4, 5, 6])
-        assert _get_prefix_length(a, b) == 0
+        assert get_prefix_length(a, b) == 0
 
     def test_partial_prefix(self):
         a = mx.array([1, 2, 3, 4, 5])
         b = mx.array([1, 2, 3, 7, 8])
-        assert _get_prefix_length(a, b) == 3
+        assert get_prefix_length(a, b) == 3
 
     def test_prompt_longer_than_cached(self):
         a = mx.array([1, 2, 3, 4, 5])
         b = mx.array([1, 2, 3])
-        assert _get_prefix_length(a, b) == 3
+        assert get_prefix_length(a, b) == 3
 
     def test_cached_longer_than_prompt(self):
         a = mx.array([1, 2, 3])
         b = mx.array([1, 2, 3, 4, 5])
-        assert _get_prefix_length(a, b) == 3
+        assert get_prefix_length(a, b) == 3
 
     def test_single_token_match(self):
         a = mx.array([1, 2, 3])
         b = mx.array([1, 5, 6])
-        assert _get_prefix_length(a, b) == 1
+        assert get_prefix_length(a, b) == 1
 
     def test_empty_prompt(self):
         a = mx.array([]).astype(mx.int32)
         b = mx.array([1, 2, 3])
-        assert _get_prefix_length(a, b) == 0
+        assert get_prefix_length(a, b) == 0
 
     def test_empty_cached(self):
         a = mx.array([1, 2, 3])
         b = mx.array([]).astype(mx.int32)
-        assert _get_prefix_length(a, b) == 0
+        assert get_prefix_length(a, b) == 0
 
     def test_both_empty(self):
         a = mx.array([]).astype(mx.int32)
         b = mx.array([]).astype(mx.int32)
-        assert _get_prefix_length(a, b) == 0
+        assert get_prefix_length(a, b) == 0
 
 
 class TestKVPrefix:
@@ -146,7 +146,7 @@ class TestKVPrefixCacheWithModel:
         prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
 
         # Cache should now hold the prompt tokens
-        assert _cache_length(cache) == len(tokens)
+        assert cache_length(cache) == len(tokens)
 
     def test_add_and_get_exact_match(self, model_and_tokenizer):
         model, tokenizer = model_and_tokenizer
@@ -166,7 +166,7 @@ class TestKVPrefixCacheWithModel:
         kv_prefix_cache.add_kv_cache(prompt, cache)
 
         assert len(kv_prefix_cache.prompts) == 1
-        stored_length = _cache_length(kv_prefix_cache.caches[0])
+        stored_length = cache_length(kv_prefix_cache.caches[0])
         assert stored_length > 0
 
         # Retrieve with same prompt: exact match
@@ -209,7 +209,7 @@ class TestKVPrefixCacheWithModel:
         long_tokens = encode_prompt(tokenizer, long_prompt)
 
         # The prompts share a prefix (chat template preamble + "Hi")
-        expected_prefix = _get_prefix_length(long_tokens, short_tokens)
+        expected_prefix = get_prefix_length(long_tokens, short_tokens)
         assert expected_prefix > 0, (
             "Prompts should share a prefix from the chat template"
         )
@@ -243,7 +243,7 @@ class TestKVPrefixCacheWithModel:
         kv_prefix_cache = KVPrefixCache(tokenizer)
         kv_prefix_cache.add_kv_cache(prompt, cache)
 
-        stored_length = _cache_length(kv_prefix_cache.caches[0])
+        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)
@@ -259,7 +259,7 @@ class TestKVPrefixCacheWithModel:
         mx.eval([c.keys for c in result_cache])
 
         # Stored cache must be unchanged
-        assert _cache_length(kv_prefix_cache.caches[0]) == stored_length
+        assert cache_length(kv_prefix_cache.caches[0]) == stored_length
 
     def test_stored_cache_survives_repeated_get_mutate_cycles(
         self, model_and_tokenizer
@@ -281,7 +281,7 @@ class TestKVPrefixCacheWithModel:
         kv_prefix_cache = KVPrefixCache(tokenizer)
         kv_prefix_cache.add_kv_cache(prompt, cache)
 
-        stored_length = _cache_length(kv_prefix_cache.caches[0])
+        stored_length = cache_length(kv_prefix_cache.caches[0])
 
         for i in range(3):
             result_cache, _, _ = kv_prefix_cache.get_kv_cache(model, prompt)
@@ -293,7 +293,7 @@ class TestKVPrefixCacheWithModel:
                 layer_cache.update_and_fetch(extra, extra)
             mx.eval([c.keys for c in result_cache])
 
-            assert _cache_length(kv_prefix_cache.caches[0]) == stored_length, (
+            assert cache_length(kv_prefix_cache.caches[0]) == stored_length, (
                 f"Failed on loop {i}"
             )
 
@@ -325,7 +325,7 @@ class TestKVPrefixCacheWithModel:
         assert len(kv_prefix_cache.caches) == 1
         # Cache should contain prompt + generated tokens
         expected_length = len(prompt_tokens) + generated_tokens
-        assert _cache_length(kv_prefix_cache.caches[0]) == expected_length
+        assert cache_length(kv_prefix_cache.caches[0]) == expected_length
 
     def test_mlx_generate_second_call_gets_prefix_hit(self, model_and_tokenizer):
         """Second mlx_generate call with same prompt should get a prefix hit from stored cache."""
@@ -400,7 +400,7 @@ class TestKVPrefixCacheWithModel:
         first_gen_time = time.perf_counter() - t0
 
         assert len(kv_prefix_cache.prompts) == 1
-        first_cache_length = _cache_length(kv_prefix_cache.caches[0])
+        first_cache_length = cache_length(kv_prefix_cache.caches[0])
 
         # Second generation: same long prompt + extra content (simulating multi-turn)
         task2 = ChatCompletionTaskParams(
@@ -416,7 +416,7 @@ class TestKVPrefixCacheWithModel:
         prompt2_tokens = encode_prompt(tokenizer, prompt2)
 
         # Verify the prompts share a long prefix
-        prefix_len = _get_prefix_length(prompt2_tokens, prompt1_tokens)
+        prefix_len = get_prefix_length(prompt2_tokens, prompt1_tokens)
         assert prefix_len > 1000, "Prompts must share > 1000 token prefix"
 
         # Second generation should reuse the cached prefix (only prefill new tokens)
@@ -440,7 +440,7 @@ class TestKVPrefixCacheWithModel:
         # With prefix_hit > 1000, should update in-place (not add a second entry)
         assert len(kv_prefix_cache.prompts) == 1
         # Updated cache should be longer (prompt2 + generated > prompt1 + generated)
-        updated_cache_length = _cache_length(kv_prefix_cache.caches[0])
+        updated_cache_length = cache_length(kv_prefix_cache.caches[0])
         assert updated_cache_length > first_cache_length
 
     def test_mlx_generate_stored_cache_not_mutated(self, model_and_tokenizer):
@@ -465,7 +465,7 @@ class TestKVPrefixCacheWithModel:
         ):
             pass
 
-        first_cache_length = _cache_length(kv_prefix_cache.caches[0])
+        firstcache_length = cache_length(kv_prefix_cache.caches[0])
 
         # Second generation gets the cache and mutates it during generation
         for _response in mlx_generate(
@@ -478,7 +478,7 @@ class TestKVPrefixCacheWithModel:
             pass
 
         # The first stored cache must not have been mutated by the second generation
-        assert _cache_length(kv_prefix_cache.caches[0]) == first_cache_length
+        assert cache_length(kv_prefix_cache.caches[0]) == firstcache_length
 
     def test_evicts_lru_entry_under_memory_pressure(self, model_and_tokenizer):
         """Under memory pressure, adding a new cache entry evicts the least recently used one."""
@@ -540,6 +540,6 @@ class TestKVPrefixCacheWithModel:
         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(
+        assert get_prefix_length(kv_prefix_cache.prompts[0], new_tokens) == len(
             new_tokens
         )

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

@@ -16,7 +16,7 @@ from exo.download.download_utils import (
     ensure_models_dir,
     fetch_file_list_with_cache,
 )
-from exo.shared.models.model_cards import ModelCard, ModelId, get_model_cards
+from exo.shared.models.model_cards import MODEL_CARDS, ModelCard, ModelId
 from exo.worker.engines.mlx.utils_mlx import (
     get_eos_token_ids_for_model,
     load_tokenizer_for_model_id,
@@ -76,7 +76,7 @@ def get_test_models() -> list[ModelCard]:
     """Get a representative sample of models to test."""
     # Pick one model from each family to test
     families: dict[str, ModelCard] = {}
-    for card in asyncio.run(get_model_cards()):
+    for card in MODEL_CARDS.values():
         # Extract family name (e.g., "llama-3.1" from "llama-3.1-8b")
         parts = card.model_id.short().split("-")
         family = "-".join(parts[:2]) if len(parts) >= 2 else parts[0]
@@ -296,7 +296,7 @@ async def test_tokenizer_special_tokens(model_card: ModelCard) -> None:
 async def test_kimi_tokenizer_specifically():
     """Test Kimi tokenizer with its specific patches and quirks."""
     kimi_models = [
-        card for card in await get_model_cards() if "kimi" in card.model_id.lower()
+        card for card in MODEL_CARDS.values() if "kimi" in card.model_id.lower()
     ]
 
     if not kimi_models:
@@ -343,7 +343,7 @@ async def test_kimi_tokenizer_specifically():
 async def test_glm_tokenizer_specifically():
     """Test GLM tokenizer with its specific EOS tokens."""
     glm_model_cards = [
-        card for card in await get_model_cards() if "glm" in card.model_id.lower()
+        card for card in MODEL_CARDS.values() if "glm" in card.model_id.lower()
     ]
 
     if not glm_model_cards:

src/exo/worker/tests/unittests/test_runner/test_event_ordering.py

@@ -109,8 +109,8 @@ def assert_events_equal(test_events: Iterable[Event], true_events: Iterable[Even
 
 @pytest.fixture
 def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
-    # initialize_mlx returns a "group" equal to 1
-    monkeypatch.setattr(mlx_runner, "initialize_mlx", make_nothin(1))
+    # initialize_mlx returns a mock group
+    monkeypatch.setattr(mlx_runner, "initialize_mlx", make_nothin(MockGroup()))
     monkeypatch.setattr(mlx_runner, "load_mlx_items", make_nothin((1, MockTokenizer)))
     monkeypatch.setattr(mlx_runner, "warmup_inference", make_nothin(1))
     monkeypatch.setattr(mlx_runner, "_check_for_debug_prompts", nothin)
@@ -120,7 +120,7 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.setattr(mlx_runner, "detect_thinking_prompt_suffix", make_nothin(False))
 
     def fake_generate(*_1: object, **_2: object):
-        yield GenerationResponse(token=0, text="hi", finish_reason="stop")
+        yield GenerationResponse(token=0, text="hi", finish_reason="stop", usage=None)
 
     monkeypatch.setattr(mlx_runner, "mlx_generate", fake_generate)
 
@@ -147,6 +147,14 @@ class MockTokenizer:
     has_tool_calling = False
 
 
+class MockGroup:
+    def rank(self) -> int:
+        return 0
+
+    def size(self) -> int:
+        return 1
+
+
 def _run(tasks: Iterable[Task]):
     bound_instance = get_bound_mlx_ring_instance(
         instance_id=INSTANCE_1_ID,
@@ -182,6 +190,8 @@ def test_events_processed_in_correct_order(patch_out_mlx: pytest.MonkeyPatch):
             text="hi",
             token_id=0,
             finish_reason="stop",
+            usage=None,
+            stats=None,
         ),
     )
 

tests/headless_runner.py

@@ -16,7 +16,7 @@ from exo.download.impl_shard_downloader import (
     exo_shard_downloader,
 )
 from exo.shared.logging import InterceptLogger, logger_setup
-from exo.shared.models.model_cards import ModelId
+from exo.shared.models.model_cards import MODEL_CARDS, ModelId
 from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
 from exo.shared.types.commands import CommandId
 from exo.shared.types.common import Host, NodeId
@@ -89,26 +89,22 @@ async def tb_detection():
 
 async def assert_downloads():
     sd = exo_shard_downloader()
+    # await sd.ensure_shard(await build_full_shard(MODEL_CARDS["qwen3-0.6b"].model_id))
     await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/Qwen3-0.6B-4bit"))
+        await build_full_shard(MODEL_CARDS["llama-3.1-8b-bf16"].model_id)
+    )
+    await sd.ensure_shard(await build_full_shard(MODEL_CARDS["qwen3-30b"].model_id))
+    await sd.ensure_shard(
+        await build_full_shard(MODEL_CARDS["gpt-oss-120b-MXFP4-Q8"].model_id)
     )
     await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"))
+        await build_full_shard(MODEL_CARDS["gpt-oss-20b-4bit"].model_id)
     )
     await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/Meta-Llama-3.1-8B-Instruct-bf16"))
+        await build_full_shard(MODEL_CARDS["glm-4.7-8bit-gs32"].model_id)
     )
     await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/gpt-oss-120b-MXFP4-Q8"))
-    )
-    await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/gpt-oss-20b-MXFP4-Q8"))
-    )
-    await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/GLM-4.7-8bit-gs32"))
-    )
-    await sd.ensure_shard(
-        await build_full_shard(ModelId("mlx-community/MiniMax-M2.1-8bit"))
+        await build_full_shard(MODEL_CARDS["minimax-m2.1-8bit"].model_id)
     )
 
 

tmp/config_examples/opencode.json

@@ -0,0 +1,18 @@
+{
+  "$schema": "https://opencode.ai/config.json",
+  "model": "exo/mlx-community/gpt-oss-120b-MXFP4-Q8",
+  "provider": {
+    "exo": {
+      "api": "http://localhost:52415/v1",
+      "models": {
+        "mlx-community/gpt-oss-120b-MXFP4-Q8": {
+          "name": "GPT OSS 120B",
+          "limit": {
+            "context": 32768,
+            "output": 8192
+          }
+        }
+      }
+    }
+  }
+}