push sheeet

pkgs/development/rocm-modules/6/tensile/default.nix

@@ -0,0 +1,92 @@
+{
+  lib,
+  fetchFromGitHub,
+  fetchpatch,
+  rocmUpdateScript,
+  buildPythonPackage,
+  pytestCheckHook,
+  setuptools,
+  distro,
+  pyyaml,
+  msgpack,
+  pandas,
+  joblib,
+  filelock,
+  clr,
+  rich,
+}:
+
+buildPythonPackage rec {
+  pname = "tensile";
+  # Using a specific commit which has compression support from after the 6.4 release
+  # Without compression packages are too large for hydra
+  version = "6.4-unstable-2025-06-12";
+  format = "pyproject";
+
+  src = fetchFromGitHub {
+    owner = "ROCm";
+    repo = "Tensile";
+    rev = "1ce87a9fe73610ffb962082f0a882360cd39b103";
+    hash = "sha256-qIuoIbmridy1HQVV10qPTzbccuxNJPsOvePaQQnClZc=";
+  };
+
+  # TODO: It should be possible to run asm caps test ONCE for all supported arches
+  # We currently disable the test because it's slow and runs each time tensile launches
+  postPatch = ''
+    substituteInPlace Tensile/Common.py \
+      --replace-fail 'if globalParameters["AssemblerPath"] is not None:' "if False:"
+    # Add an assert that the fallback 9,0,0 is supported before setting the kernel to it
+    # If it's not detected as supported we have an issue with compiler paths or the compiler is broken
+    # and it's better to stop immediately
+    substituteInPlace Tensile/KernelWriter.py \
+      --replace-fail '= (9,0,0)' '= (9,0,0);assert(globalParameters["AsmCaps"][(9,0,0)]["SupportedISA"])'
+    find . -type f -iname "*.sh" -exec chmod +x {} \;
+    patchShebangs Tensile
+  '';
+
+  buildInputs = [ setuptools ];
+
+  propagatedBuildInputs = [
+    pyyaml
+    msgpack
+    pandas
+    joblib
+    distro
+    rich
+  ];
+
+  patches = [
+    ./tensile-solutionstructs-perf-fix.diff
+    ./tensile-create-library-dont-copy-twice.diff
+    (fetchpatch {
+      # [PATCH] Extend Tensile HIP ISA compatibility
+      sha256 = "sha256-d+fVf/vz+sxGqJ96vuxe0jRMgbC5K6j5FQ5SJ1e3Sl8=";
+      url = "https://github.com/GZGavinZhao/Tensile/commit/855cb15839849addb0816a6dde45772034a3e41f.patch";
+    })
+  ];
+
+  doCheck = false; # Too many errors, not sure how to set this up properly
+
+  nativeCheckInputs = [
+    pytestCheckHook
+    filelock
+    clr
+  ];
+
+  env.ROCM_PATH = "${clr}";
+
+  pythonImportsCheck = [ "Tensile" ];
+
+  passthru.updateScript = rocmUpdateScript {
+    name = pname;
+    inherit (src) owner repo;
+  };
+
+  meta = with lib; {
+    description = "GEMMs and tensor contractions";
+    homepage = "https://github.com/ROCm/Tensile";
+    license = with licenses; [ mit ];
+    teams = [ teams.rocm ];
+    platforms = platforms.linux;
+  };
+}

pkgs/development/rocm-modules/6/tensile/tensile-create-library-dont-copy-twice.diff

@@ -0,0 +1,20 @@
+diff --git a/Tensile/TensileCreateLibrary.py b/Tensile/TensileCreateLibrary.py
+index a1644606..c6ca2882 100644
+--- a/Tensile/TensileCreateLibrary.py
++++ b/Tensile/TensileCreateLibrary.py
+@@ -852,9 +852,14 @@ def copyStaticFiles(outputPath=None):
+         "KernelHeader.h",
+     ]
+ 
++    import filecmp
+     for fileName in libraryStaticFiles:
+         # copy file
+-        shutil.copy(os.path.join(globalParameters["SourcePath"], fileName), outputPath)
++        # no need to copy twice if it has already been copied
++        src = os.path.join(globalParameters["SourcePath"], fileName)
++        dst = os.path.join(outputPath, os.path.basename(src))
++        if not os.path.isfile(dst) or not filecmp.cmp(src, dst):
++            shutil.copyfile(src, dst)
+ 
+     return libraryStaticFiles
+ 

pkgs/development/rocm-modules/6/tensile/tensile-solutionstructs-perf-fix.diff

@@ -0,0 +1,48 @@
+diff --git a/Tensile/SolutionStructs.py b/Tensile/SolutionStructs.py
+index f663c6f1..17bcf897 100644
+--- a/Tensile/SolutionStructs.py
++++ b/Tensile/SolutionStructs.py
+@@ -4828,24 +4828,26 @@ class Solution(collections.abc.Mapping):
+   # create a dictionary of lists of parameter values
+   @staticmethod
+   def getSerialNaming(objs):
++    valid_params = sorted(validParameters.keys())
+     data = {}
+-    for objIdx in range(0, len(objs)):
+-      obj = objs[objIdx]
+-      for paramName in sorted(obj.keys()):
+-        if paramName in list(validParameters.keys()):
+-          paramValue = obj[paramName]
+-          if paramName in data:
+-            if paramValue not in data[paramName]:
+-              data[paramName].append(paramValue)
+-          else:
+-            data[paramName] = [ paramValue ]
+-    maxObjs = 1
+-    for paramName in data:
+-      if not isinstance(data[paramName][0],dict):
+-        data[paramName] = sorted(data[paramName])
+-      maxObjs *= len(data[paramName])
+-    numDigits = len(str(maxObjs))
+-    return [ data, numDigits ]
++
++    objs = [getattr(obj, "_state", obj) for obj in objs]
++
++    for param in valid_params:
++      d = []
++      for obj in objs:
++        if param in obj:
++          v = obj[param]
++          if v not in d:
++            d.append(v)
++      if len(d):
++        if not isinstance(d[0], dict): d.sort()
++        data[param] = d
++
++    # Calculate max objects using prod() from math module
++    max_objs = math.prod(len(values) for values in data.values())
++    num_digits = len(str(max_objs))
++    return data, num_digits
+ 
+   ########################################
+   # Get Name Serial