3 files changed, 88 insertions, 0 deletions
diff --git a/nixpkgs/pkgs/tools/misc/ollama/default.nix b/nixpkgs/pkgs/tools/misc/ollama/default.nix
new file mode 100644
index 000000000000..2176582e1fe9
--- /dev/null
+++ b/nixpkgs/pkgs/tools/misc/ollama/default.nix
@@ -0,0 +1,50 @@
+{ lib
+, buildGoModule
+, fetchFromGitHub
+, llama-cpp
+}:
+
+buildGoModule rec {
+  pname = "ollama";
+  version = "0.1.17";
+
+  src = fetchFromGitHub {
+    owner = "jmorganca";
+    repo = "ollama";
+    rev = "v${version}";
+    hash = "sha256-eXukNn9Lu1hF19GEi7S7a96qktsjnmXCUp38gw+3MzY=";
+  };
+
+  patches = [
+    # disable passing the deprecated gqa flag to llama-cpp-server
+    # see https://github.com/ggerganov/llama.cpp/issues/2975
+    ./disable-gqa.patch
+
+    # replace the call to the bundled llama-cpp-server with the one in the llama-cpp package
+    ./set-llamacpp-path.patch
+  ];
+
+  postPatch = ''
+    substituteInPlace llm/llama.go \
+      --subst-var-by llamaCppServer "${llama-cpp}/bin/llama-cpp-server"
+    substituteInPlace server/routes_test.go --replace "0.0.0" "${version}"
+  '';
+
+  vendorHash = "sha256-yGdCsTJtvdwHw21v0Ot6I8gxtccAvNzZyRu1T0vaius=";
+
+  ldflags = [
+    "-s"
+    "-w"
+    "-X=github.com/jmorganca/ollama/version.Version=${version}"
+    "-X=github.com/jmorganca/ollama/server.mode=release"
+  ];
+
+  meta = with lib; {
+    description = "Get up and running with large language models locally";
+    homepage = "https://github.com/jmorganca/ollama";
+    license = licenses.mit;
+    mainProgram = "ollama";
+    maintainers = with maintainers; [ dit7ya elohmeier ];
+    platforms = platforms.unix;
+  };
+}
diff --git a/nixpkgs/pkgs/tools/misc/ollama/disable-gqa.patch b/nixpkgs/pkgs/tools/misc/ollama/disable-gqa.patch
new file mode 100644
index 000000000000..b54440cd3d53
--- /dev/null
+++ b/nixpkgs/pkgs/tools/misc/ollama/disable-gqa.patch
@@ -0,0 +1,15 @@
+diff --git a/llm/llama.go b/llm/llama.go
+index 0b460e9..b79e04a 100644
+--- a/llm/llama.go
++++ b/llm/llama.go
+@@ -299,10 +299,6 @@ func newLlama(model string, adapters []string, runners []ModelRunner, numLayers
+ 		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", numGPU))
+ 	}
+ 
+-	if opts.NumGQA > 0 {
+-		params = append(params, "--gqa", fmt.Sprintf("%d", opts.NumGQA))
+-	}
+-
+ 	if len(adapters) > 0 {
+ 		// TODO: applying multiple adapters is not supported by the llama.cpp server yet
+ 		params = append(params, "--lora", adapters[0])
diff --git a/nixpkgs/pkgs/tools/misc/ollama/set-llamacpp-path.patch b/nixpkgs/pkgs/tools/misc/ollama/set-llamacpp-path.patch
new file mode 100644
index 000000000000..e90e552bab45
--- /dev/null
+++ b/nixpkgs/pkgs/tools/misc/ollama/set-llamacpp-path.patch
@@ -0,0 +1,23 @@
+diff --git a/llm/llama.go b/llm/llama.go
+index f23d5d8..6563550 100644
+--- a/llm/llama.go
++++ b/llm/llama.go
+@@ -25,7 +25,6 @@ import (
+ 	"github.com/jmorganca/ollama/api"
+ )
+ 
+-//go:embed llama.cpp/*/build/*/bin/*
+ var llamaCppEmbed embed.FS
+ 
+ type ModelRunner struct {
+@@ -33,6 +32,10 @@ type ModelRunner struct {
+ }
+ 
+ func chooseRunners(workDir, runnerType string) []ModelRunner {
++	return []ModelRunner{
++		{Path: "@llamaCppServer@"},
++	}
++
+ 	buildPath := path.Join("llama.cpp", runnerType, "build")
+ 	var runners []string
+