ollama: ship deepseek-coder-v2-16b-lite-instruct-q5_1,qwen2_5-32b-instruct-q2_K

modules/services/ollama.nix

@@ -24,6 +24,7 @@ let
       codestral-22b
       # deepseek-coder-7b  # subpar to deepseek-coder-v2 in nearly every way
       deepseek-coder-v2-16b  # GREAT balance between speed and code quality. code is superior to qwen2_5 in some ways, and inferior in others
+      deepseek-coder-v2-16b-lite-instruct-q5_1  # higher-res version of default 16b (so, better answers?)
       # falcon2-11b  # code examples are lacking
       # gemma2-9b  # fast, but not great for code
       # glm4-9b  # it generates invalid code
@@ -40,6 +41,7 @@ let
       # phi3_5-3b  # generates invalid code
       # qwen2_5-7b   # notably less quality than 32b (i.e. generates invalid code)
       qwen2_5-14b  # *almost* same quality to 32b variant, but faster
+      qwen2_5-32b-instruct-q2_K  # lower-res version of default 32b (so, faster?)
       qwen2_5-32b  # generates 3~5 words/sec, but notably more accurate than coder-7b
       # qwen2_5-coder-7b  # fast, and concise, but generates invalid code
       # solar-pro-22b  # generates invalid code

pkgs/by-name/ollamaPackages/deepseek-coder-v2-16b-lite-instruct-q5_1.nix

@@ -0,0 +1,11 @@
+# <https://ollama.com/library/deepseek-coder-v2>
+# - normal 16b is 8.9 GB; 16b-lite-instruct-q5_1 is 12 GB
+{ mkOllamaModel }: mkOllamaModel {
+  modelName = "deepseek-coder-v2";
+  variant = "16b-lite-instruct-q5_1";
+  manifestHash = "sha256-8U8LWP7e8jUMjjeooHn9tT/GX9UXIfIJrzGAB8aLRmQ=";
+  modelBlob = "7d9a741801702d521029cf790c40f9b473e9ecfb5fd65b6eb6b03eb07f8e4ff4";
+  modelBlobHash = "sha256-fZp0GAFwLVIQKc95DED5tHPp7Ptf1ltutrA+sH+OT/Q=";
+  paramsBlob = "19f2fb9e8bc65a143f47903ec07dce010fd2873f994b900ea735a4b5022e968d";
+  paramsBlobHash = "sha256-GfL7novGWhQ/R5A+wH3OAQ/Shz+ZS5AOpzWktQIulo0=";
+}

pkgs/by-name/ollamaPackages/mkOllamaModel.nix

@@ -10,11 +10,11 @@
   manifestHash ? "",
   # grab the *Blob from the manifest (trim the `sha256:` prefix).
   # the manifest can be acquired by providing just the above parameters and building this package, then viewing the output
-  modelBlob ? null,
+  modelBlob ? "",
   modelBlobHash ? "",
-  paramsBlob ? null,
+  paramsBlob ? "",
   paramsBlobHash ? "",
-  systemBlob ? null,
+  systemBlob ? "",
   systemBlobHash ? "",
 }:
 stdenv.mkDerivation {
@@ -24,17 +24,17 @@ stdenv.mkDerivation {
       url = "https://registry.ollama.ai/v2/library/${modelName}/manifests/${variant}";
       hash = manifestHash;
     })
-  ] ++ lib.optionals (modelBlob != null) [
+  ] ++ lib.optionals (modelBlob != "") [
     (fetchurl {
       url = "https://registry.ollama.ai/v2/llama/${modelName}:${variant}/blobs/sha256-${modelBlob}";
       hash = modelBlobHash;
     })
-  ] ++ lib.optionals (paramsBlob != null) [
+  ] ++ lib.optionals (paramsBlob != "") [
     (fetchurl {
       url = "https://registry.ollama.ai/v2/llama/${modelName}:${variant}/blobs/sha256-${paramsBlob}";
       hash = paramsBlobHash;
     })
-  ] ++ lib.optionals (systemBlob != null) [
+  ] ++ lib.optionals (systemBlob != "") [
     (fetchurl {
       url = "https://registry.ollama.ai/v2/llama/${modelName}:${variant}/blobs/sha256-${systemBlob}";
       hash = systemBlobHash;

pkgs/by-name/ollamaPackages/package.nix

@@ -19,6 +19,7 @@ lib.recurseIntoAttrs (lib.makeScope newScope (self: with self; {
   codestral-22b = callPackage ./codestral-22b.nix { };
   deepseek-coder-7b = callPackage ./deepseek-coder-7b.nix { };
   deepseek-coder-v2-16b = callPackage ./deepseek-coder-v2-16b.nix { };
+  deepseek-coder-v2-16b-lite-instruct-q5_1 = callPackage ./deepseek-coder-v2-16b-lite-instruct-q5_1.nix { };
   falcon2-11b = callPackage ./falcon2-11b.nix { };
   gemma2-9b = callPackage ./gemma2-9b.nix { };
   glm4-9b = callPackage ./glm4-9b.nix { };
@@ -36,6 +37,7 @@ lib.recurseIntoAttrs (lib.makeScope newScope (self: with self; {
   qwen2_5-7b = callPackage ./qwen2_5-7b.nix { };
   qwen2_5-14b = callPackage ./qwen2_5-14b.nix { };
   qwen2_5-32b = callPackage ./qwen2_5-32b.nix { };
+  qwen2_5-32b-instruct-q2_K = callPackage ./qwen2_5-32b-instruct-q2_K.nix { };
   qwen2_5-coder-7b = callPackage ./qwen2_5-coder-7b.nix { };
   solar-pro-22b = callPackage ./solar-pro-22b.nix { };
   starcoder2-15b-instruct = callPackage ./starcoder2-15b-instruct.nix { };

pkgs/by-name/ollamaPackages/qwen2_5-32b-instruct-q2_K.nix

@@ -0,0 +1,12 @@
+# <https://ollama.com/library/qwen2.5>
+# - 32b-instruct-q2_K is a variant of 32b that's 12 GB instead of 20 GB
+#   14b-instruct is 9 GB, for comparison
+{ mkOllamaModel }: mkOllamaModel {
+  modelName = "qwen2.5";
+  variant = "32b-instruct-q2_K";
+  manifestHash = "sha256-mYGFobYa4TEkAb+xBICAnmieSgYsTT+FbQNxGSWReNI=";
+  modelBlob = "ed7a653589dcc230070e02058681ad28ba202ddbd6b79e3dcbf23e436d5e8f00";
+  modelBlobHash = "sha256-7XplNYncwjAHDgIFhoGtKLogLdvWt549y/I+Q21ejwA=";
+  systemBlob = "66b9ea09bd5b7099cbb4fc820f31b575c0366fa439b08245566692c6784e281e";
+  systemBlobHash = "sha256-ZrnqCb1bcJnLtPyCDzG1dcA2b6Q5sIJFVmaSxnhOKB4=";
+}