Add Nix flake with CUDA devshell for hal9000 builds

Provides a devShell with rust-overlay toolchain, CUDA packages (nvcc,
cudart, cublas, nvtx) on Linux, and proper LD_LIBRARY_PATH/CUDA_PATH
env vars for RTX 4090 (sm_89). macOS shell includes libiconv for
Metal builds. deploy.sh now always uses `nix develop` instead of
falling back to system cargo.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jamesbrink

.envrc

@@ -0,0 +1 @@
+use flake

.github/workflows/ci.yml

@@ -9,6 +9,9 @@ on:
 env:
   CARGO_TERM_COLOR: always
 
+# Note: CUDA builds are not run in CI — they happen on hal9000 (NixOS + RTX 4090)
+# via scripts/deploy.sh. CI only checks non-CUDA compilation, lints, and tests.
+
 jobs:
   check:
     name: Check

.gitignore

@@ -10,3 +10,7 @@ Cargo.lock
 *.jpeg
 *.jpg
 !docs/**/*.png
+
+# Nix / direnv
+.direnv/
+result

flake.lock

@@ -0,0 +1,72 @@
+{
+  description = "mold - like ollama, but for diffusion models";
+
+  inputs = {
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
+    flake-utils.url = "github:numtide/flake-utils";
+    rust-overlay = {
+      url = "github:oxalica/rust-overlay";
+      inputs.nixpkgs.follows = "nixpkgs";
+    };
+  };
+
+  outputs = { self, nixpkgs, flake-utils, rust-overlay }:
+    flake-utils.lib.eachDefaultSystem (system:
+      let
+        isLinux = builtins.elem system [ "x86_64-linux" "aarch64-linux" ];
+        isDarwin = builtins.elem system [ "x86_64-darwin" "aarch64-darwin" ];
+        overlays = [ rust-overlay.overlays.default ];
+        pkgs = import nixpkgs {
+          inherit system overlays;
+          config.allowUnfree = true;
+          config.cudaSupport = isLinux;
+        };
+        rustToolchain = pkgs.rust-bin.stable.latest.default.override {
+          extensions = [ "rust-src" "rustfmt" "clippy" ];
+        };
+
+        commonInputs = [
+          rustToolchain
+          pkgs.pkg-config
+          pkgs.openssl
+          pkgs.git
+        ];
+
+        linuxInputs = pkgs.lib.optionals isLinux [
+          pkgs.cudaPackages.cuda_nvcc
+          pkgs.cudaPackages.cuda_cudart
+          pkgs.cudaPackages.libcublas
+          pkgs.cudaPackages.cuda_nvtx
+        ];
+
+        darwinInputs = pkgs.lib.optionals isDarwin [
+          pkgs.libiconv
+        ];
+      in
+      {
+        devShells.default = pkgs.mkShell {
+          buildInputs = commonInputs ++ linuxInputs ++ darwinInputs;
+
+          env = pkgs.lib.optionalAttrs isLinux {
+            CUDA_PATH = "${pkgs.cudaPackages.cuda_cudart}";
+            CUDA_COMPUTE_CAP = "89"; # RTX 4090 = sm_89
+          };
+
+          LD_LIBRARY_PATH = pkgs.lib.optionalString isLinux (
+            (pkgs.lib.makeLibraryPath [
+              pkgs.cudaPackages.cuda_cudart
+              pkgs.cudaPackages.libcublas
+            ]) + ":/run/opengl-driver/lib"
+          );
+
+          shellHook = ''
+            echo "mold dev shell (${system})"
+          '' + pkgs.lib.optionalString isLinux ''
+            echo "  CUDA enabled - build with: cargo build --release -p mold-server --features cuda"
+          '' + pkgs.lib.optionalString isDarwin ''
+            echo "  macOS - build with: cargo build --release -p mold-server --features metal"
+          '';
+        };
+      }
+    );
+}

flake.nix

@@ -10,15 +10,10 @@ echo "🔨 Syncing source to hal9000..."
 rsync -av --exclude target --exclude .git --exclude "*.png" \
   ./ "$REMOTE_USER@$HAL9000:$REMOTE_DIR/"
 
-echo "🦀 Building on hal9000 (CUDA)..."
+echo "🦀 Building on hal9000 (CUDA via nix develop)..."
 ssh "$REMOTE_USER@$HAL9000" "
   cd $REMOTE_DIR
-  # Try to use cargo directly, fallback to nix develop
-  if command -v cargo &>/dev/null; then
-    cargo build --release -p mold-server --features cuda 2>&1
-  else
-    nix develop --command cargo build --release -p mold-server --features cuda 2>&1
-  fi
+  nix develop --command cargo build --release -p mold-server --features cuda 2>&1
 "
 
 echo "🛑 Stopping existing mold-server (if running)..."