feat: stable-diffussion & updates

2026-01-02 09:50:44 -05:00
parent bb3305adbf
commit f3ceb57e5e
6 changed files with 190 additions and 30 deletions
--- a/homes/x86_64-linux/evanreichard@lin-va-thinkpad/default.nix
+++ b/homes/x86_64-linux/evanreichard@lin-va-thinkpad/default.nix
@@ -51,6 +51,7 @@ in
        git = enabled;
        k9s = enabled;
        nvim = enabled;
+        opencode = enabled;
      };
    };
  };
--- a/modules/home/programs/terminal/opencode/default.nix
+++ b/modules/home/programs/terminal/opencode/default.nix
@@ -28,15 +28,6 @@ in
          doom_loop = "ask";
          external_directory = "ask";
        };
-        lsp = {
-          nil = {
-            command = [
-              "${pkgs.nil}/bin/nil"
-              "--stdio"
-            ];
-            extensions = [ ".nix" ];
-          };
-        };
        provider = {
          "llama-swap" = {
            npm = "@ai-sdk/openai-compatible";
@@ -44,6 +35,9 @@ in
              baseURL = "https://llm-api.va.reichard.io/v1";
            };
            models = {
+              nemotron-3-nano-30b-thinking = {
+                name = "Nemotron 3 Nano (30B) - Thinking";
+              };
              gpt-oss-20b-thinking = {
                name = "GPT OSS (20B)";
              };
@@ -62,6 +56,9 @@ in
              qwen3-30b-2507-instruct = {
                name = "Qwen3 2507 (30B) Instruct";
              };
+              qwen3-4b-2507-instruct = {
+                name = "Qwen3 2507 (4B) - Instruct";
+              };
            };
          };
        };
--- a/packages/llama-swap/default.nix
+++ b/packages/llama-swap/default.nix
@@ -13,13 +13,13 @@ let
 in
 buildGoModule (finalAttrs: {
  pname = "llama-swap";
-  version = "179";
+  version = "180";

  src = fetchFromGitHub {
    owner = "mostlygeek";
    repo = "llama-swap";
    tag = "v${finalAttrs.version}";
-    hash = "sha256-7iftF1KMz+2DUifRG/ESHcWXYsOJ3NiEF7oHuJKxmUE=";
+    hash = "sha256-WPDmENGH1uwNrobcIPA2vuNEsb9sP7Wl7T0wtUv1H/s=";
    # populate values that require us to use git. By doing this in postFetch we
    # can delete .git afterwards and maintain better reproducibility of the src.
    leaveDotGit = true;
--- a/packages/opencode/default.nix
+++ b/packages/opencode/default.nix
@@ -13,12 +13,12 @@
 }:
 let
  pname = "opencode";
-  version = "1.0.170";
+  version = "1.0.223";
  src = fetchFromGitHub {
    owner = "sst";
    repo = "opencode";
    tag = "v${version}";
-    hash = "sha256-Y0thIZ20p0FSBAH0mJfFn8e+OEUvlZyTuk+/yEt8Sy8=";
+    hash = "sha256-CzbWv48UySgXfNgtWdIdFBcqx8GHT4rSJNDdpn39b0c="; # "sha256-Y0thIZ20p0FSBAH0mJfFn8e+OEUvlZyTuk+/yEt8Sy8=";
  };

  node_modules = stdenvNoCC.mkDerivation {
@@ -75,7 +75,7 @@ let
    # NOTE: Required else we get errors that our fixed-output derivation references store paths
    dontFixup = true;

-    outputHash = "sha256-Aq774bgU12HkrF2oAtfu9kyQFlxUeDbmwlS9lz4Z4ZI=";
+    outputHash = "sha256-+HEd3I11VqejTi7cikbTL5+DmNGyvUC4Cm4ysfujwes=";
    outputHashAlgo = "sha256";
    outputHashMode = "recursive";
  };
--- a/packages/stable-diffusion-cpp/default.nix
+++ b/packages/stable-diffusion-cpp/default.nix
@@ -0,0 +1,118 @@
+{ lib
+, stdenv
+, fetchFromGitHub
+, cmake
+, ninja
+, pkg-config
+, autoAddDriverRunpath
+, config ? { }
+, cudaSupport ? (config.cudaSupport or false)
+, cudaPackages ? { }
+, rocmSupport ? (config.rocmSupport or false)
+, rocmPackages ? { }
+, rocmGpuTargets ? (rocmPackages.clr.localGpuTargets or rocmPackages.clr.gpuTargets or [ ])
+, openclSupport ? false
+, clblast
+, vulkanSupport ? false
+, shaderc
+, vulkan-headers
+, vulkan-loader
+, spirv-tools
+, metalSupport ? (stdenv.hostPlatform.isDarwin && stdenv.hostPlatform.isAarch64)
+, apple-sdk
+,
+}:
+
+let
+  inherit (lib)
+    cmakeBool
+    cmakeFeature
+    optionals
+    ;
+
+  effectiveStdenv = if cudaSupport then cudaPackages.backendStdenv else stdenv;
+in
+effectiveStdenv.mkDerivation (finalAttrs: {
+  pname = "stable-diffusion-cpp";
+  version = "master-453-4ff2c8c";
+
+  src = fetchFromGitHub {
+    owner = "leejet";
+    repo = "stable-diffusion.cpp";
+    rev = "master-453-4ff2c8c";
+    hash = "sha256-8cN6dYOQAKnJpuQdtayp6+o71s64lG+FcTn8GsIM4jI=";
+    fetchSubmodules = true;
+  };
+
+  nativeBuildInputs = [
+    cmake
+    ninja
+    pkg-config
+  ]
+  ++ optionals cudaSupport [
+    (cudaPackages.cuda_nvcc)
+    autoAddDriverRunpath
+  ];
+
+  buildInputs =
+    (optionals cudaSupport (
+      with cudaPackages;
+      [
+        cuda_cccl
+        cuda_cudart
+        libcublas
+      ]
+    ))
+    ++ (optionals rocmSupport (
+      with rocmPackages;
+      [
+        clr
+        hipblas
+        rocblas
+      ]
+    ))
+    ++ (optionals vulkanSupport [
+      shaderc
+      vulkan-headers
+      vulkan-loader
+      spirv-tools
+    ])
+    ++ (optionals openclSupport [
+      clblast
+    ])
+    ++ (optionals metalSupport [
+      apple-sdk
+    ]);
+
+  cmakeFlags = [
+    (cmakeBool "SD_BUILD_EXAMPLES" true)
+    (cmakeBool "SD_BUILD_SHARED_LIBS" true)
+    (cmakeBool "SD_USE_SYSTEM_GGML" false)
+    (cmakeBool "SD_CUDA" cudaSupport)
+    (cmakeBool "SD_HIPBLAS" rocmSupport)
+    (cmakeBool "SD_VULKAN" vulkanSupport)
+    (cmakeBool "SD_OPENCL" openclSupport)
+    (cmakeBool "SD_METAL" metalSupport)
+    (cmakeBool "SD_FAST_SOFTMAX" false)
+  ]
+  ++ optionals cudaSupport [
+    (cmakeFeature "CMAKE_CUDA_ARCHITECTURES" cudaPackages.flags.cmakeCudaArchitecturesString)
+  ]
+  ++ optionals rocmSupport [
+    (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmGpuTargets))
+  ];
+
+  meta = with lib; {
+    description = "Stable Diffusion inference in pure C/C++";
+    homepage = "https://github.com/leejet/stable-diffusion.cpp";
+    license = licenses.mit;
+    mainProgram = "sd";
+    maintainers = with lib.maintainers; [
+      dit7ya
+      adriangl
+    ];
+    platforms = platforms.unix;
+    badPlatforms = optionals (cudaSupport || openclSupport) platforms.darwin;
+    broken = metalSupport && !stdenv.hostPlatform.isDarwin;
+  };
+})
--- a/systems/x86_64-linux/lin-va-desktop/default.nix
+++ b/systems/x86_64-linux/lin-va-desktop/default.nix
@@ -5,6 +5,12 @@
 }:
 let
  inherit (lib.${namespace}) enabled;
+
+  llama-swap = pkgs.reichard.llama-swap;
+  llama-cpp = pkgs.reichard.llama-cpp;
+  stable-diffusion-cpp = pkgs.reichard.stable-diffusion-cpp.override {
+    cudaSupport = true;
+  };
 in
 {
  system.stateVersion = "25.11";
@@ -88,14 +94,14 @@ in
  services.llama-swap = {
    enable = true;
    openFirewall = true;
-    package = pkgs.reichard.llama-swap;
+    package = llama-swap;
    settings = {
      models = {
        # https://huggingface.co/unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF/tree/main
        "devstral-small-2-instruct" = {
          name = "Devstral Small 2 (24B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Devstral/Devstral-Small-2-24B-Instruct-2512-UD-Q4_K_XL.gguf \
              --chat-template-file /mnt/ssd/Models/Devstral/Devstral-Small-2-24B-Instruct-2512-UD-Q4_K_XL_template.jinja \
@@ -113,7 +119,7 @@ in
        "gpt-oss-20b-thinking" = {
          name = "GPT OSS (20B) - Thinking";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/GPT-OSS/gpt-oss-20b-F16.gguf \
              -c 131072 \
@@ -128,7 +134,7 @@ in
        "gpt-oss-csec-20b-thinking" = {
          name = "GPT OSS CSEC (20B) - Thinking";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/GPT-OSS/GPT-OSS-Cybersecurity-20B-Merged.i1-MXFP4_MOE.gguf \
              -c 131072 \
@@ -143,7 +149,7 @@ in
        "qwen3-next-80b-instruct" = {
          name = "Qwen3 Next (80B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Qwen3/Qwen3-Next-80B-A3B-Instruct-UD-Q2_K_XL.gguf \
              -c 262144 \
@@ -162,7 +168,7 @@ in
        "qwen3-30b-2507-instruct" = {
          name = "Qwen3 2507 (30B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Qwen3/Qwen3-30B-A3B-Instruct-2507-Q4_K_M.gguf \
              -c 262144 \
@@ -181,10 +187,10 @@ in
        "qwen3-coder-30b-instruct" = {
          name = "Qwen3 Coder (30B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
-              -m /mnt/ssd/Models/Qwen3/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf \
-              -c 262144 \
+              -m /mnt/ssd/Models/Qwen3/Qwen3-Coder-30B-A3B-Instruct-UD-Q6_K_XL.gguf \
+              -c 131072 \
              --temp 0.7 \
              --min-p 0.0 \
              --top-p 0.8 \
@@ -200,7 +206,7 @@ in
        "qwen3-30b-2507-thinking" = {
          name = "Qwen3 2507 (30B) - Thinking";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Qwen3/Qwen3-30B-A3B-Thinking-2507-UD-Q4_K_XL.gguf \
              -c 262144 \
@@ -219,7 +225,7 @@ in
        "nemotron-3-nano-30b-thinking" = {
          name = "Nemotron 3 Nano (30B) - Thinking";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Nemotron/Nemotron-3-Nano-30B-A3B-UD-Q4_K_XL.gguf \
              -c 1048576 \
@@ -233,7 +239,7 @@ in
        "qwen3-8b-vision" = {
          name = "Qwen3 Vision (8B) - Thinking";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Qwen3/Qwen3-VL-8B-Instruct-UD-Q4_K_XL.gguf \
              --mmproj /mnt/ssd/Models/Qwen3/Qwen3-VL-8B-Instruct-UD-Q4_K_XL_mmproj-F16.gguf \
@@ -253,11 +259,12 @@ in
        "qwen2.5-coder-7b-instruct" = {
          name = "Qwen2.5 Coder (7B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              -m /mnt/ssd/Models/Qwen2.5/Qwen2.5-Coder-7B-Instruct-Q8_0.gguf \
              --fim-qwen-7b-default \
              -c 131072 \
              --port ''${PORT} \
+              -fit off \
              -dev CUDA1
          '';
        };
@@ -266,7 +273,7 @@ in
        "qwen2.5-coder-3b-instruct" = {
          name = "Qwen2.5 Coder (3B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              -m /mnt/ssd/Models/Qwen2.5/Qwen2.5-Coder-3B-Instruct-Q8_0.gguf \
              --fim-qwen-3b-default \
              --port ''${PORT} \
@@ -279,7 +286,7 @@ in
        "qwen3-4b-2507-instruct" = {
          name = "Qwen3 2507 (4B) - Instruct";
          cmd = ''
-            ${pkgs.reichard.llama-cpp}/bin/llama-server \
+            ${llama-cpp}/bin/llama-server \
              --port ''${PORT} \
              -m /mnt/ssd/Models/Qwen3/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
              -c 98304 \
@@ -289,6 +296,40 @@ in
              -dev CUDA1
          '';
        };
+
+        "z-image-turbo" = {
+          name = "Z-Image-Turbo";
+          checkEndpoint = "/";
+          cmd = ''
+            ${stable-diffusion-cpp}/bin/sd-server \
+              --listen-port ''${PORT} \
+              --diffusion-fa \
+              --diffusion-model /mnt/ssd/StableDiffusion/ZImageTurbo/z-image-turbo-Q8_0.gguf \
+              --vae /mnt/ssd/StableDiffusion/ZImageTurbo/ae.safetensors \
+              --llm /mnt/ssd/Models/Qwen3/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
+              --cfg-scale 1.0 \
+              --steps 9 \
+              --rng cuda
+          '';
+        };
+
+        "qwen-image-edit" = {
+          name = "Qwen Image Edit";
+          checkEndpoint = "/";
+          cmd = ''
+            ${stable-diffusion-cpp}/bin/sd-server \
+              --listen-port ''${PORT} \
+              --diffusion-fa \
+              --diffusion-model /mnt/ssd/StableDiffusion/QwenImageEdit/Qwen-Rapid-v18_Q5_K.gguf \
+              --vae /mnt/ssd/StableDiffusion/QwenImageEdit/qwen_image_vae.safetensors \
+              --llm /mnt/ssd/Models/Qwen2.5/Qwen2.5-VL-7B-Instruct.Q4_K_M.gguf \
+              --cfg-scale 2.5 \
+              --sampling-method euler \
+              --flow-shift 3 \
+              --steps 9 \
+              --rng cuda
+          '';
+        };
      };

      groups = {
@@ -335,6 +376,9 @@ in
    git
    tmux
    vim
-    reichard.llama-cpp
+
+    # Local Packages
+    llama-cpp
+    stable-diffusion-cpp
  ];
 }