[backport 26.05] python3Packages.flash-linear-attention: init at 0.5.0; python3Packages.exllamav3: 0.0.38 -> 0.0.42 (#531303)

pkgs/development/python-modules/exllamav3/default.nix

@@ -8,6 +8,7 @@
   setuptools,
 
   flash-attn,
+  flash-linear-attention,
   formatron,
   kbnf,
   marisa-trie,
@@ -28,14 +29,14 @@ let
 in
 buildPythonPackage.override { inherit (torch) stdenv; } (finalAttrs: {
   pname = "exllamav3";
-  version = "0.0.38";
+  version = "0.0.42";
   pyproject = true;
 
   src = fetchFromGitHub {
     owner = "turboderp-org";
     repo = "exllamav3";
     tag = "v${finalAttrs.version}";
-    hash = "sha256-WlHIbnQX1Jd7y5yQzlqXVgBLQ92rnDSWy4z9bEm3WLA=";
+    hash = "sha256-kdI2BT7T2+mrdgWE7aXTeqC49WP6qEus+LfQGk0ozhA=";
   };
 
   pythonRelaxDeps = [
@@ -60,6 +61,7 @@ buildPythonPackage.override { inherit (torch) stdenv; } (finalAttrs: {
 
   dependencies = [
     flash-attn
+    flash-linear-attention
     formatron
     kbnf
     marisa-trie

pkgs/development/python-modules/flash-linear-attention/default.nix

@@ -0,0 +1,66 @@
+{
+  lib,
+  buildPythonPackage,
+  fetchFromGitHub,
+  pythonOlder,
+
+  # build-system
+  setuptools,
+
+  # dependencies
+  einops,
+  torch,
+  transformers,
+
+  # optional-dependencies
+  causal-conv1d,
+  matplotlib,
+  datasets,
+  pytest,
+}:
+
+buildPythonPackage (finalAttrs: {
+  pname = "flash-linear-attention";
+  version = "0.5.0";
+  pyproject = true;
+
+  disabled = pythonOlder "3.10";
+
+  src = fetchFromGitHub {
+    owner = "fla-org";
+    repo = "flash-linear-attention";
+    tag = "v${finalAttrs.version}";
+    hash = "sha256-g66yGHaBwEyjb+of76tKTtV/7as/2xQuqcjbGs4E3rU=";
+  };
+
+  build-system = [ setuptools ];
+
+  dependencies = [
+    einops
+    torch
+    transformers
+  ];
+
+  optional-dependencies = {
+    # tilelang = [ tilelang ];
+    conv1d = [ causal-conv1d ];
+    benchmark = [
+      matplotlib
+      datasets
+    ];
+    test = [ pytest ];
+  };
+
+  # Tests require a GPU
+  doCheck = false;
+
+  pythonImportsCheck = [ "fla" ];
+
+  meta = {
+    description = "Triton-based implementations of causal linear attention";
+    homepage = "https://github.com/fla-org/flash-linear-attention";
+    changelog = "https://github.com/fla-org/flash-linear-attention/releases/tag/v${finalAttrs.version}";
+    license = lib.licenses.mit;
+    maintainers = with lib.maintainers; [ BatteredBunny ];
+  };
+})

pkgs/top-level/python-packages.nix

@@ -5695,6 +5695,8 @@ self: super: with self; {
 
   flash-attn = callPackage ../development/python-modules/flash-attn { };
 
+  flash-linear-attention = callPackage ../development/python-modules/flash-linear-attention { };
+
   flash-mla = callPackage ../development/python-modules/flash-mla { };
 
   flashinfer = callPackage ../development/python-modules/flashinfer { };