rocmPackages: 7.1.1 -> 7.2.0 (#484792)

2026-03-05 13:38:58 +00:00
parent 9c0a49094d 3bc0e6a4c5
commit 5fab99fdcf
54 changed files with 442 additions and 399 deletions
@@ -21,12 +21,12 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "amdsmi";
-  version = "7.1.1";
+  version = "7.2.0";
  src = fetchFromGitHub {
    owner = "rocm";
    repo = "amdsmi";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-BGe3+8YFwu41ZVAF+VtN5Cn9pfzGxmCg/Rpq8qWOEoM=";
+    hash = "sha256-3V3B9+B3cpg0ebvsmX6wwvKoTIHbXtvbpIQHs4tkeWU=";
  };

  postPatch = ''
@@ -8,12 +8,12 @@

 stdenv.mkDerivation {
  pname = "aqlprofile";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "aqlprofile";
-    tag = "rocm-7.1.1";
+    tag = "rocm-7.2.0";
    hash = "sha256-MAZUHo52gb0aZSFnKugMlXxcDkmMwhy1AFF1RDRgRVk=";
  };

@@ -70,7 +70,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "clr";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -84,7 +84,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "clr";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-ofsq1uqMixtum5C6cp/UgTDpgGPfj+rAd6PoDx5iLLw=";
+    hash = "sha256-zz2O4Qsl1zXMC25L714azsFR2PROAvdpjgKhRolmt1w=";
  };

  nativeBuildInputs = [
@@ -148,8 +148,8 @@ stdenv.mkDerivation (finalAttrs: {
    ./cmake-find-x11-libgl.patch
    (fetchpatch {
      # [PATCH] rocclr: Extend HIP ISA compatibility checks
-      sha256 = "sha256-InUSIFI1MgkfocBEoZjO2BCgXNyfF10ehh9jkTtAPXs=";
-      url = "https://github.com/GZGavinZhao/rocm-systems/commit/937dcfdd316b589509c061809186fe5451d22431.patch";
+      hash = "sha256-3MsDL+OQg24wH1RDhbao74RuIbzEAmduwla9KOPzQ/M=";
+      url = "https://github.com/GZGavinZhao/rocm-systems/commit/039cb23b24d739adb8c0f9de8b550d9f557de031.patch";
      relative = "projects/clr";
    })
  ];
@@ -42,7 +42,7 @@ stdenv.mkDerivation (finalAttrs: {
  '';

  pname = "composable_kernel_base";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -58,7 +58,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "composable_kernel";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-exdkyTIK03dzlCXHm3j8ehEb9NxLOxPX9QyfMsiCgSs=";
+    hash = "sha256-ABL0MSmWtqAeY5uyw8Ib64npB2v82baUnzLpmrEgDn4=";
  };

  nativeBuildInputs = [
@@ -54,6 +54,7 @@ let
      targets = [
        "device_grouped_conv1d_fwd_instance"
        "device_grouped_conv2d_fwd_instance"
+        "device_grouped_conv2d_fwd_bias_bnorm_clamp_instance"
        "device_grouped_conv2d_fwd_bias_clamp_instance"
        "device_grouped_conv2d_fwd_clamp_instance"
        "device_grouped_conv2d_fwd_dynamic_op_instance"
@@ -77,6 +78,7 @@ let
      targets = [
        "device_grouped_conv3d_fwd_instance"
        "device_grouped_conv3d_fwd_clamp_instance"
+        "device_grouped_conv3d_fwd_bias_bnorm_clamp_instance"
        "device_grouped_conv3d_fwd_bias_clamp_instance"
        "device_grouped_conv3d_fwd_bilinear_instance"
        "device_grouped_conv3d_fwd_convinvscale_instance"
@@ -9,7 +9,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "half";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
@@ -7,13 +7,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "hip-common";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "HIP";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-aI41HdyzyE5HtCbzFmTQYGMiFyQELM/MTVOAbQsg8So=";
+    hash = "sha256-eQ+jHc6MlZePPIwJQMB8NKiEcE26i83+U21vKgfGFFM=";
  };

  dontConfigure = true;
@@ -8,13 +8,13 @@
 }:
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipblas-common";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "hipBLAS-common";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-sswqYIv3c5ArT6dQ3MRru7ZZVA0VxoKW/a0pYTBpjgE=";
+    hash = "sha256-oOelDrk7TLe9/17fzaw6CDPL4MGAoITJL90ahHzhaAE=";
  };

  nativeBuildInputs = [
@@ -24,7 +24,7 @@
 # Can also use cuBLAS
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipblas";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -43,7 +43,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipBLAS";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-2wDnYZapJ4fU/JZpy6qOf5+DYuatWmCw/hA8WO5x2T0=";
+    hash = "sha256-w7myy4FO2vKm91OOfSGWO3bxR8jAvjU68oqI7aJX82c=";
  };

  postPatch = ''
@@ -1,53 +1,3 @@
-diff --git a/tensilelite/Tensile/SolutionStructs/Naming.py b/tensilelite/Tensile/SolutionStructs/Naming.py
-index 4f220960db1d..99535e246650 100644
--- a/tensilelite/Tensile/SolutionStructs/Naming.py
-+++ b/tensilelite/Tensile/SolutionStructs/Naming.py
-@@ -105,7 +105,6 @@ def _getName(state, requiredParameters: frozenset, splitGSU: bool, ignoreInterna
-     if splitGSU:
-       state["GlobalSplitU"] = "M" if (state["GlobalSplitU"] > 1 or state["GlobalSplitU"] == -1) else state["GlobalSplitU"]
- 
-
-   requiredParametersTemp = set(requiredParameters.union(["GlobalSplitU"]))
- 
-   if ignoreInternalArgs:
-diff --git a/tensilelite/Tensile/CustomYamlLoader.py b/tensilelite/Tensile/CustomYamlLoader.py
-index bab8c687509..e03f456fbec 100644
--- a/tensilelite/Tensile/CustomYamlLoader.py
-+++ b/tensilelite/Tensile/CustomYamlLoader.py
-@@ -1,3 +1,6 @@
-+# Copyright © Advanced Micro Devices, Inc., or its affiliates.
-+# SPDX-License-Identifier: MIT
-+
- import yaml
- from pathlib import Path
- 
-Author: Luna Nova <git@lunnova.dev>
-Date:   Sun Oct 12 11:52:10 2025 -0700
-
-    [hipblaslt] intern strings to reduce duplicate memory for solution keys
-
-diff --git a/tensilelite/Tensile/CustomYamlLoader.py b/tensilelite/Tensile/CustomYamlLoader.py
-index 685e69220c..9fdf38d8e5 100644
--- a/tensilelite/Tensile/CustomYamlLoader.py
-+++ b/tensilelite/Tensile/CustomYamlLoader.py
-@@ -1,6 +1,7 @@
- # Copyright © Advanced Micro Devices, Inc., or its affiliates.
- # SPDX-License-Identifier: MIT
- 
-+import sys
- import yaml
- from pathlib import Path
- 
-@@ -85,7 +86,7 @@ def parse_scalar(loader: yaml.Loader):
-         if not evt.style:
-             return None
- 
-    return value
-+    return sys.intern(value)
- 
- def load_yaml_stream(yaml_path: Path, loader_type: yaml.Loader):
-     with open(yaml_path, 'r') as f:
-
 diff --git a/tensilelite/Tensile/Common/Parallel.py b/tensilelite/Tensile/Common/Parallel.py
 index 1a2bf9e119..f46100c7b8 100644
 --- a/tensilelite/Tensile/Common/Parallel.py
@@ -55,7 +5,7 @@ index 1a2bf9e119..f46100c7b8 100644
@@ -22,43 +22,58 @@
 #
 ################################################################################
- 
+
 -import concurrent.futures
 -import itertools
 +import multiprocessing
@@ -67,10 +17,10 @@ index 1a2bf9e119..f46100c7b8 100644
 -from joblib import Parallel, delayed
 +from functools import partial
 +from typing import Any, Callable
- 
+
 from .Utilities import tqdm
- 
- 
+
+
 -def joblibParallelSupportsGenerator():
 -    import joblib
 -    from packaging.version import Version
@@ -81,7 +31,7 @@ index 1a2bf9e119..f46100c7b8 100644
 +            return int(os.environ['CMAKE_BUILD_PARALLEL_LEVEL'])
 +        except ValueError:
 +            pass
- 
+
 -    joblibVer = joblib.__version__
 -    return Version(joblibVer) >= Version("1.4.0")
 +    # 2. Parse MAKEFLAGS for -jN
@@ -89,12 +39,12 @@ index 1a2bf9e119..f46100c7b8 100644
 +    match = re.search(r'-j\s*(\d+)', makeflags)
 +    if match:
 +        return int(match.group(1))
- 
+
 +    return -1
- 
+
 -def CPUThreadCount(enable=True):
 -    from .GlobalParameters import globalParameters
- 
+
 +def CPUThreadCount(enable=True):
     if not enable:
         return 1
@@ -121,7 +71,7 @@ index 1a2bf9e119..f46100c7b8 100644
 -        if cpuThreads == -1:
 -            return cpu_count
 +            cpuThreads = len(os.sched_getaffinity(0))
- 
+
 -        return min(cpu_count, cpuThreads)
 +    if os.name == "nt":
 +        # Windows supports at most 61 workers because the scheduler uses
@@ -129,13 +79,13 @@ index 1a2bf9e119..f46100c7b8 100644
 +        # is actually 64, but some handles are needed for accounting).
 +        cpuThreads = min(cpuThreads, 61)
 +    return max(1, cpuThreads)
- 
- 
+
+
 def pcallWithGlobalParamsMultiArg(f, args, newGlobalParameters):
@@ -71,19 +86,22 @@ def pcallWithGlobalParamsSingleArg(f, arg, newGlobalParameters):
     return f(arg)
- 
- 
+
+
 -def apply_print_exception(item, *args):
 -    # print(item, args)
 +def OverwriteGlobalParameters(newGlobalParameters):
@@ -166,8 +116,8 @@ index 1a2bf9e119..f46100c7b8 100644
     finally:
@@ -98,154 +116,121 @@ def OverwriteGlobalParameters(newGlobalParameters):
     GlobalParameters.globalParameters.update(newGlobalParameters)
- 
- 
+
+
 -def ProcessingPool(enable=True, maxTasksPerChild=None):
 -    import multiprocessing
 -    import multiprocessing.dummy
@@ -191,10 +141,10 @@ index 1a2bf9e119..f46100c7b8 100644
 +def progress_logger(iterable, total, message, min_log_interval=5.0):
 +    """
 +    Generator that wraps an iterable and logs progress with time-based throttling.
- 
+
 +    Only logs progress if at least min_log_interval seconds have passed since last log.
 +    Only prints completion message if task took >= min_log_interval seconds.
- 
+
 -def ParallelMap(function, objects, message="", enable=True, method=None, maxTasksPerChild=None):
 +    Yields (index, item) tuples.
     """
@@ -212,7 +162,7 @@ index 1a2bf9e119..f46100c7b8 100644
 +    start_time = time.time()
 +    last_log_time = start_time
 +    log_interval = 1 + (total // 100)
- 
+
 -    threadCount = CPUThreadCount(enable)
 -    pool = ProcessingPool(enable, maxTasksPerChild)
 -
@@ -251,12 +201,12 @@ index 1a2bf9e119..f46100c7b8 100644
 +                print(f"{message}\t{idx+1: 5d}/{total: 5d}")
 +                last_log_time = current_time
 +        yield idx, item
- 
+
 -    if message != "":
 -        message += ": "
 +    elapsed = time.time() - start_time
 +    final_idx = idx + 1 if 'idx' in locals() else 0
- 
+
 -    print("{0}Launching {1} threads{2}...".format(message, threadCount, countMessage))
 -    sys.stdout.flush()
 -    currentTime = time.time()
@@ -268,8 +218,8 @@ index 1a2bf9e119..f46100c7b8 100644
 -    return rv
 +    if elapsed >= min_log_interval or last_log_time > start_time:
 +        print(f"{message} done in {elapsed:.1f}s!\t{final_idx: 5d}/{total: 5d}")
- 
- 
+
+
 -def ParallelMapReturnAsGenerator(function, objects, message="", enable=True, multiArg=True):
 -    from .GlobalParameters import globalParameters
 +def imap_with_progress(pool, func, iterable, total, message, chunksize):
@@ -277,10 +227,10 @@ index 1a2bf9e119..f46100c7b8 100644
 +    for _, result in progress_logger(pool.imap(func, iterable, chunksize=chunksize), total, message):
 +        results.append(result)
 +    return results
- 
+
 -    threadCount = CPUThreadCount(enable)
 -    print("{0}Launching {1} threads...".format(message, threadCount))
- 
+
 -    if threadCount <= 1 and globalParameters["ShowProgressBar"]:
 -        # Provide a progress bar for single-threaded operation.
 -        callFunc = lambda args: function(*args) if multiArg else lambda args: function(args)
@@ -288,7 +238,7 @@ index 1a2bf9e119..f46100c7b8 100644
 +def _ParallelMap_generator(worker, objects, objLen, message, chunksize, threadCount, globalParameters, maxtasksperchild):
 +    # separate fn because yield makes the entire fn a generator even if unreachable
 +    ctx = multiprocessing.get_context('forkserver' if os.name != 'nt' else 'spawn')
- 
+
 -    with concurrent.futures.ProcessPoolExecutor(max_workers=threadCount) as executor:
 -        resultFutures = (executor.submit(function, *arg if multiArg else arg) for arg in objects)
 -        for result in concurrent.futures.as_completed(resultFutures):
@@ -297,8 +247,8 @@ index 1a2bf9e119..f46100c7b8 100644
 +                  initializer=OverwriteGlobalParameters, initargs=(globalParameters,)) as pool:
 +        for _, result in progress_logger(pool.imap_unordered(worker, objects, chunksize=chunksize), objLen, message):
 +            yield result
- 
- 
+
+
 def ParallelMap2(
 -    function, objects, message="", enable=True, multiArg=True, return_as="list", procs=None
 +    function: Callable,
@@ -333,7 +283,7 @@ index 1a2bf9e119..f46100c7b8 100644
     """
 -    Generally equivalent to list(map(function, objects)), possibly executing in parallel.
 +    from .GlobalParameters import globalParameters
- 
+
 -      message: A message describing the operation to be performed.
 -      enable: May be set to false to disable parallelism.
 -      multiArg: True if objects represent multiple arguments
@@ -342,23 +292,23 @@ index 1a2bf9e119..f46100c7b8 100644
 -    if return_as in ("generator", "generator_unordered") and not joblibParallelSupportsGenerator():
 -        return ParallelMapReturnAsGenerator(function, objects, message, enable, multiArg)
 +    threadCount = CPUThreadCount(enable)
- 
+
 -    from .GlobalParameters import globalParameters
 +    if not hasattr(objects, "__len__"):
 +        objects = list(objects)
- 
+
 -    threadCount = procs if procs else CPUThreadCount(enable)
 +    objLen = len(objects)
 +    if objLen == 0:
 +        return [] if return_as == "list" else iter([])
- 
+
 -    threadCount = CPUThreadCount(enable)
 +    f = (lambda x: function(*x)) if multiArg else function
 +    if objLen == 1:
 +        print(f"{message}: (1 task)")
 +        result = [f(x) for x in objects]
 +        return result if return_as == "list" else iter(result)
- 
+
 -    if threadCount <= 1 and globalParameters["ShowProgressBar"]:
 -        # Provide a progress bar for single-threaded operation.
 -        return [function(*args) if multiArg else function(args) for args in tqdm(objects, message)]
@@ -367,7 +317,7 @@ index 1a2bf9e119..f46100c7b8 100644
 +        if objLen
 +        else ""
 +    )
- 
+
 -    countMessage = ""
 -    try:
 -        countMessage = " for {} tasks".format(len(objects))
@@ -419,19 +369,19 @@ index ffceb636f5..127b3386a1 100644
 --- a/tensilelite/Tensile/CustomKernels.py
 +++ b/tensilelite/Tensile/CustomKernels.py
@@ -24,7 +24,9 @@
- 
+
 from . import CUSTOM_KERNEL_PATH
 from Tensile.Common.ValidParameters import checkParametersAreValid, validParameters, newMIValidParameters
 +from Tensile.CustomYamlLoader import DEFAULT_YAML_LOADER
- 
+
 +from functools import lru_cache
 import yaml
- 
+
 import os
@@ -58,10 +60,13 @@ def getCustomKernelConfigAndAssembly(name, directory=CUSTOM_KERNEL_PATH):
- 
+
     return (config, assembly)
- 
+
 +# getCustomKernelConfig will get called repeatedly on the same file
 +# 20x logic loading speedup for aquavanjaram_Cijk_Ailk_Bljk_F8NH_HHS_BH_Bias_HAS_SAB_SAV_freesize_custom_GSUs
 +@lru_cache
@@ -442,13 +392,34 @@ index ffceb636f5..127b3386a1 100644
 +        return yaml.load(rawConfig, Loader=DEFAULT_YAML_LOADER)["custom.config"]
     except yaml.scanner.ScannerError as e:
         raise RuntimeError("Failed to read configuration for custom kernel: {0}\nDetails:\n{1}".format(name, e))
- 
+
+diff --git a/tensilelite/Tensile/CustomYamlLoader.py b/tensilelite/Tensile/CustomYamlLoader.py
+index e03f456fbe..ed7510f2ce 100644
+--- a/tensilelite/Tensile/CustomYamlLoader.py
+++ b/tensilelite/Tensile/CustomYamlLoader.py
+@@ -1,6 +1,7 @@
+ # Copyright © Advanced Micro Devices, Inc., or its affiliates.
+ # SPDX-License-Identifier: MIT
+
+import sys
+ import yaml
+ from pathlib import Path
+
+@@ -70,7 +71,7 @@ def parse_scalar(loader: yaml.Loader):
+     elif is_float(value_lower):
+         return float(value_lower)
+
+-    return value
+    return sys.intern(value)
+
+ def load_yaml_stream(yaml_path: Path, loader_type: yaml.Loader):
+     with open(yaml_path, 'r') as f:
 diff --git a/tensilelite/Tensile/TensileCreateLibrary/Run.py b/tensilelite/Tensile/TensileCreateLibrary/Run.py
-index 835ed9c019..024c6c49c1 100644
+index 22d19851a3..348068b3bf 100644
 --- a/tensilelite/Tensile/TensileCreateLibrary/Run.py
 +++ b/tensilelite/Tensile/TensileCreateLibrary/Run.py
@@ -26,8 +26,10 @@ import rocisa
- 
+
 import functools
 import glob
 +import gc
@@ -460,8 +431,8 @@ index 835ed9c019..024c6c49c1 100644
 from timeit import default_timer as timer
@@ -78,6 +80,25 @@ from Tensile.Utilities.Decorators.Timing import timing
 from .ParseArguments import parseArguments
- 
- 
+
+
 +def getMemoryUsage():
 +    """Get peak and current memory usage in MB."""
 +    rusage = resource.getrusage(resource.RUSAGE_SELF)
@@ -484,11 +455,11 @@ index 835ed9c019..024c6c49c1 100644
 class KernelCodeGenResult(NamedTuple):
     err: int
     src: str
-@@ -115,6 +136,29 @@ def processKernelSource(kernelWriterAssembly, data, splitGSU, kernel) -> KernelC
+@@ -115,6 +136,29 @@ def processKernelSource(kernelWriterAssembly, data, outOptions, splitGSU, kernel
     )
- 
- 
-+def processAndAssembleKernelTCL(kernelWriterAssembly, rocisa_data, splitGSU, kernel, assemblyTmpPath, assembler):
+
+
+def processAndAssembleKernelTCL(kernelWriterAssembly, rocisa_data, outOptions, splitGSU, kernel, assemblyTmpPath, assembler):
 +    """
 +    Pipeline function for TCL mode that:
 +    1. Generates kernel source
@@ -496,7 +467,7 @@ index 835ed9c019..024c6c49c1 100644
 +    3. Assembles to .o file
 +    4. Deletes .s file
 +    """
-+    result = processKernelSource(kernelWriterAssembly, rocisa_data, splitGSU, kernel)
+    result = processKernelSource(kernelWriterAssembly, rocisa_data, outOptions, splitGSU, kernel)
 +    return writeAndAssembleKernel(result, assemblyTmpPath, assembler)
 +
 +
@@ -516,8 +487,8 @@ index 835ed9c019..024c6c49c1 100644
     removeKernelNames = []
@@ -189,6 +233,24 @@ def writeAssembly(asmPath: Union[Path, str], result: KernelCodeGenResult):
     return path, isa, wfsize, minResult
- 
- 
+
+
 +def writeAndAssembleKernel(result: KernelCodeGenResult, asmPath: Union[Path, str], assembler):
 +    """Write assembly file and immediately assemble it to .o file"""
 +    if result.err:
@@ -539,13 +510,14 @@ index 835ed9c019..024c6c49c1 100644
 def writeHelpers(
     outputPath, kernelHelperObjs, KERNEL_HELPER_FILENAME_CPP, KERNEL_HELPER_FILENAME_H
 ):
-@@ -268,13 +330,14 @@ def writeSolutionsAndKernels(
+@@ -272,14 +334,15 @@ def writeSolutionsAndKernels(
     numAsmKernels = len(asmKernels)
     numKernels = len(asmKernels)
     assert numKernels == numAsmKernels, "Only assembly kernels are supported in TensileLite"
 -    asmIter = zip(
 -        itertools.repeat(kernelWriterAssembly),
 -        itertools.repeat(rocisa.rocIsa.getInstance().getData()),
+-        itertools.repeat(outOptions),
 -        itertools.repeat(splitGSU),
 -        asmKernels
 +
@@ -553,6 +525,7 @@ index 835ed9c019..024c6c49c1 100644
 +        processKernelSource,
 +        kernelWriterAssembly=kernelWriterAssembly,
 +        data=rocisa.rocIsa.getInstance().getData(),
+        outOptions=outOptions,
 +        splitGSU=splitGSU
     )
 -    asmResults = ParallelMap2(processKernelSource, asmIter, "Generating assembly kernels", return_as="list")
@@ -560,12 +533,12 @@ index 835ed9c019..024c6c49c1 100644
     removeInvalidSolutionsAndKernels(
         asmResults, asmKernels, solutions, errorTolerant, getVerbosity(), splitGSU
     )
-@@ -282,19 +345,21 @@ def writeSolutionsAndKernels(
+@@ -287,19 +350,21 @@ def writeSolutionsAndKernels(
         asmResults, asmKernels, solutions, splitGSU
     )
- 
+
 -    def assemble(ret):
-        p, isa, wavefrontsize, result = ret
+-        p, isa, wavefrontsize, _ = ret
 -        asmToolchain.assembler(isaToGfx(isa), wavefrontsize, str(p), str(p.with_suffix(".o")))
 -
 -    unaryWriteAssembly = functools.partial(writeAssembly, assemblyTmpPath)
@@ -586,33 +559,41 @@ index 835ed9c019..024c6c49c1 100644
     )
 +    del asmResults
 +    gc.collect()
- 
+
     writeHelpers(outputPath, kernelHelperObjs, KERNEL_HELPER_FILENAME_CPP, KERNEL_HELPER_FILENAME_H)
     srcKernelFile = Path(outputPath) / "Kernels.cpp"
-@@ -369,32 +434,31 @@ def writeSolutionsAndKernelsTCL(
- 
+@@ -376,40 +441,35 @@ def writeSolutionsAndKernelsTCL(
+
     uniqueAsmKernels = [k for k in asmKernels if not k.duplicate]
- 
-    def assemble(ret):
-        p, isa, wavefrontsize, result = ret
-        asmToolchain.assembler(isaToGfx(isa), wavefrontsize, str(p), str(p.with_suffix(".o")))
+
+-    def assemble(ret, removeTemporaries: bool):
+-        asmPath, isa, wavefrontsize, result = ret
+-        asmToolchain.assembler(isaToGfx(isa), wavefrontsize, str(asmPath), str(asmPath.with_suffix(".o")))
+-        if removeTemporaries:
+-            asmPath.unlink()
 -        return result
 -
+-    unaryAssemble = functools.partial(assemble, removeTemporaries=removeTemporaries)
+-
+     outOptions = rocisa.rocIsa.getInstance().getOutputOptions()
+     outOptions.outputNoComment = not disableAsmComments
+
 -    unaryProcessKernelSource = functools.partial(
 -        processKernelSource,
 +    processKernelFn = functools.partial(
 +        processAndAssembleKernelTCL,
         kernelWriterAssembly,
         rocisa.rocIsa.getInstance().getData(),
+         outOptions,
         splitGSU,
 +        assemblyTmpPath=assemblyTmpPath,
 +        assembler=asmToolchain.assembler
     )
- 
+
 -    unaryWriteAssembly = functools.partial(writeAssembly, assemblyTmpPath)
 -    compose = lambda *F: functools.reduce(lambda f, g: lambda x: f(g(x)), F)
 -    ret = ParallelMap2(
-        compose(assemble, unaryWriteAssembly, unaryProcessKernelSource),
+-        compose(unaryAssemble, unaryWriteAssembly, unaryProcessKernelSource),
 +    results = ParallelMap2(
 +        processKernelFn,
         uniqueAsmKernels,
@@ -635,10 +616,10 @@ index 835ed9c019..024c6c49c1 100644
     buildAssemblyCodeObjectFiles(
         asmToolchain.linker,
         asmToolchain.bundler,
-@@ -493,6 +557,15 @@ def generateKernelHelperObjects(solutions: List[Solution], cxxCompiler: str, isa
+@@ -508,6 +568,15 @@ def generateKernelHelperObjects(solutions: List[Solution], cxxCompiler: str, isa
     return sorted(khos, key=sortByEnum, reverse=True) # Ensure that we write Enum kernel helpers are first in list
- 
- 
+
+
 +def libraryIter(lib: MasterSolutionLibrary):
 +    if len(lib.solutions):
 +        for i, s in enumerate(lib.solutions.items()):
@@ -650,11 +631,11 @@ index 835ed9c019..024c6c49c1 100644
 +
 @timing
 def generateLogicDataAndSolutions(logicFiles, args, assembler: Assembler, isaInfoMap):
- 
-@@ -508,26 +581,23 @@ def generateLogicDataAndSolutions(logicFiles, args, assembler: Assembler, isaInf
+
+@@ -523,26 +592,23 @@ def generateLogicDataAndSolutions(logicFiles, args, assembler: Assembler, isaInf
     printSolutionRejectionReason = True
     printIndexAssignmentInfo = False
- 
+
 -    fIter = zip(
 -        logicFiles,
 -        itertools.repeat(assembler),
@@ -672,7 +653,7 @@ index 835ed9c019..024c6c49c1 100644
 +        isaInfoMap=isaInfoMap,
 +        lazyLibraryLoading=args["LazyLibraryLoading"]
     )
- 
+
 -    def libraryIter(lib: MasterSolutionLibrary):
 -        if len(lib.solutions):
 -            for i, s in enumerate(lib.solutions.items()):
@@ -691,31 +672,31 @@ index 835ed9c019..024c6c49c1 100644
 +        multiArg=False,
     ):
         _, architectureName, _, _, _, newLibrary = library
- 
-@@ -539,6 +609,9 @@ def generateLogicDataAndSolutions(logicFiles, args, assembler: Assembler, isaInf
+
+@@ -554,6 +620,9 @@ def generateLogicDataAndSolutions(logicFiles, args, assembler: Assembler, isaInf
         else:
             masterLibraries[architectureName] = newLibrary
             masterLibraries[architectureName].version = args["CodeObjectVersion"]
 +        del library, newLibrary
 +
 +    gc.collect()
- 
+
     # Sort masterLibraries to make global soln index values deterministic
     solnReIndex = 0
-@@ -734,6 +807,9 @@ def run():
+@@ -751,6 +820,9 @@ def run():
     )
     stop_wsk = timer()
     print(f"Time to generate kernels (s): {(stop_wsk-start_wsk):3.2f}")
 +    numKernelHelperObjs = len(kernelHelperObjs)
 +    del kernelWriterAssembly, kernelHelperObjs
 +    gc.collect()
- 
+
     archs = [ # is this really different than the other archs above?
         isaToGfx(arch)
-@@ -751,13 +827,10 @@ def run():
+@@ -768,13 +840,10 @@ def run():
             if kName not in solDict:
                 solDict["%s"%kName] = kernel
- 
+
 -    def writeMsl(name, lib):
 -        filename = os.path.join(newLibraryDir, name)
 -        lib.applyNaming(splitGSU)
@@ -725,13 +706,13 @@ index 835ed9c019..024c6c49c1 100644
     LibraryIO.write(filename, libraryMapping, "msgpack")
 +    del libraryMapping
 +    gc.collect()
- 
+
     start_msl = timer()
     for archName, newMasterLibrary in masterLibraries.items():
-@@ -774,12 +847,22 @@ def run():
+@@ -791,12 +860,22 @@ def run():
                     kName = getKeyNoInternalArgs(s.originalSolution, splitGSU)
                     s.sizeMapping.CUOccupancy = solDict["%s"%kName]["CUOccupancy"]
- 
+
 -            ParallelMap2(writeMsl,
 +            writeFn = functools.partial(
 +                writeMasterSolutionLibrary,
@@ -749,15 +730,15 @@ index 835ed9c019..024c6c49c1 100644
     print(f"Time to write master solution libraries (s): {(stop_msl-start_msl):3.2f}")
 +    del masterLibraries, solutions, kernels, solDict
 +    gc.collect()
- 
+
     if not arguments["KeepBuildTmp"]:
         buildTmp = Path(arguments["OutputPath"]).parent / "library" / "build_tmp"
-@@ -796,8 +879,11 @@ def run():
+@@ -813,8 +892,11 @@ def run():
     print("")
- 
+
     stop = timer()
 +    peak_memory_mb, current_memory_mb = getMemoryUsage()
- 
+
     print(f"Total time (s): {(stop-start):3.2f}")
     print(f"Total kernels processed: {numKernels}")
     print(f"Kernels processed per second: {(numKernels/(stop-start)):3.2f}")
@@ -772,7 +753,7 @@ index e33c617b6f..ba163e9918 100644
@@ -303,8 +303,7 @@ def avoidRegressions(originalDir, incrementalDir, outputPath, forceMerge, noEff=
         logicsFiles[origFile] = origFile
         logicsFiles[incFile] = incFile
- 
+
 -    iters = zip(logicsFiles.keys())
 -    logicsList = ParallelMap2(loadData, iters, "Loading Logics...", return_as="list")
 +    logicsList = ParallelMap2(loadData, logicsFiles.keys(), "Loading Logics...", return_as="list", multiArg=False)
@@ -785,11 +766,11 @@ index 5ff265d0ed..c1803a6349 100644
 +++ b/tensilelite/Tensile/TensileUpdateLibrary.py
@@ -26,7 +26,7 @@ from . import LibraryIO
 from .Tensile import addCommonArguments, argUpdatedGlobalParameters
- 
+
 from .Common import assignGlobalParameters, print1, restoreDefaultGlobalParameters, HR, \
 -                    globalParameters, architectureMap, ensurePath, ParallelMap, __version__
 +                    globalParameters, architectureMap, ensurePath, ParallelMap2, __version__
- 
+
 import argparse
 import copy
@@ -149,7 +149,7 @@ def TensileUpdateLibrary(userArgs):
@@ -798,8 +779,8 @@ index 5ff265d0ed..c1803a6349 100644
     fIter = zip(logicFiles, itertools.repeat(args.logic_path), itertools.repeat(outputPath))
 -    libraries = ParallelMap(UpdateLogic, fIter, "Updating logic files", method=lambda x: x.starmap)
 +    libraries = ParallelMap2(UpdateLogic, fIter, "Updating logic files", multiArg=True, return_as="list")
- 
- 
+
+
 def main():
 diff --git a/tensilelite/Tensile/Toolchain/Assembly.py b/tensilelite/Tensile/Toolchain/Assembly.py
 index a8b91e8d62..265e1d532c 100644
@@ -808,16 +789,16 @@ index a8b91e8d62..265e1d532c 100644
@@ -30,7 +30,7 @@ import subprocess
 from pathlib import Path
 from typing import List, Union, NamedTuple
- 
+
 -from Tensile.Common import print2
 +from Tensile.Common import print1, print2
 from Tensile.Common.Architectures import isaToGfx
 from ..SolutionStructs import Solution
- 
+
@@ -92,8 +92,26 @@ def buildAssemblyCodeObjectFiles(
         if coName:
           coFileMap[asmDir / (coName + extCoRaw)].add(str(asmDir / (kernel["BaseName"] + extObj)))
- 
+
 +      # Build reference count map for .o files to handle shared object files
 +      # (.o files from kernels marked .duplicate in TensileCreateLibrary)
 +      objFileRefCount = collections.Counter()
@@ -70,14 +70,17 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipblaslt${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocm-libraries";
-    rev = "a676499add42941ff6af1e8d3f0504416dac7429";
-    hash = "sha256-zIYdHFbHyP2V6dkx6Ueb6NBqWu8tJji2hSWF9zWEJa4=";
-    sparseCheckout = [ "projects/hipblaslt" ];
+    rev = "rocm-${finalAttrs.version}";
+    hash = "sha256-I2dGn4Ld5lZeML8GePcLPssplBZ+4weNR6uBEqFdZVg=";
+    sparseCheckout = [
+      "projects/hipblaslt"
+      "shared"
+    ];
  };
  sourceRoot = "${finalAttrs.src.name}/projects/hipblaslt";
  env.CXX = compiler;
@@ -203,7 +206,7 @@ stdenv.mkDerivation (finalAttrs: {
    # Move binaries to appropriate outputs and delete leftover /bin
    + ''
      mkdir -p $benchmark/bin
-      mv $out/bin/hipblaslt-{api-overhead,sequence,bench*} $out/bin/*.yaml $out/bin/*.py $benchmark/bin
+      mv $out/bin/hipblaslt-{api-overhead,bench*} $out/bin/*.yaml $out/bin/*.py $benchmark/bin
      ${lib.optionalString buildTests ''
        mkdir -p $test/bin
        mv $out/bin/hipblas-test $test/bin
@@ -17,7 +17,7 @@
 # CUB can also be used as a backend instead of rocPRIM.
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipcub";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -33,7 +33,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipCUB";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-3/499pX7hFKJwbhSEH3/bMhNdyabK36jLv8s5TC2fuw=";
+    hash = "sha256-hIIpU1E/X6bPIhFfauNLPnFNBYSwtoxZV335Huw9fCA=";
  };

  nativeBuildInputs = [
@@ -21,7 +21,7 @@
 # Can also use cuFFT
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipfft";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -40,7 +40,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipFFT";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-6FyI9s6H/lHFhm8aUqB9+vvJ0CRIVWCFLNoJprvsI6o=";
+    hash = "sha256-5zeL3o/Dfh0iyZOj5CHyfEQdFNehD2BLmllnZRdpsQI=";
    fetchSubmodules = true;
  };

@@ -1,7 +1,6 @@
 {
  lib,
  stdenv,
-  fetchpatch,
  fetchFromGitHub,
  rocmUpdateScript,
  cmake,
@@ -11,24 +10,15 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "hipfort";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "hipfort";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-IVLhp8rYtKhkt8K8Mc0qyrp670oKoK0QeclnJjO36pY=";
+    hash = "sha256-XaB4jauCN41tgD1YHHA2td/yckwfMBemBe/iL0SCxQo=";
  };

-  patches = [
-    (fetchpatch {
-      name = "hipfort-fix-cmake-4.patch";
-      url = "https://github.com/ROCm/hipfort/commit/75552c7ec48e3bd6a914c57c9475ec573ccb37d9.patch";
-      hash = "sha256-S9r1V6cUo9QbKbu/NK4wIvXMV6BFv7+/n9mjCScVk40=";
-      includes = [ "bin/*" ];
-    })
-  ];
-
  nativeBuildInputs = [
    cmake
    rocm-cmake
@@ -12,13 +12,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "hipify";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "HIPIFY";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-NbFFHAAvMGpvIryhEbktN5w03Cpay9lEqelqkUT9dpQ=";
+    hash = "sha256-LC0lnYetV7RPVw92zew6za6bDH4zmnERXUM4MVaRVtc=";
  };

  strictDeps = true;
@@ -14,7 +14,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "hiprand";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -27,7 +27,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipRAND";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-yypKwJ5p2aSAqapEHre9eOjjaFAI8Q1dDYWhwfrrdhg=";
+    hash = "sha256-mhhwYswewcRYKGoVAB/ZTIY8ubFSlTDXKZFoJpoDf2o=";
  };

  nativeBuildInputs = [
@@ -21,7 +21,7 @@
 # Can also use cuSOLVER
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipsolver";
-  version = "7.1.0";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -40,7 +40,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipSOLVER";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-n0+Xy92mBhI2vjB7pFHGt2UgPIHo9rO2ZxGxkz30Otw=";
+    hash = "sha256-xrzRU3X0zlmK6t8Bz4p5ihnJnadrqiW1BW2HLRDpu1M=";
  };

  nativeBuildInputs = [
@@ -19,7 +19,7 @@
 # This can also use cuSPARSE as a backend instead of rocSPARSE
 stdenv.mkDerivation (finalAttrs: {
  pname = "hipsparse";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -35,7 +35,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "hipSPARSE";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-zaMAsLfUYkMUiA2vnGiK0//+vZwugqt5T6xJU4FBrRo=";
+    hash = "sha256-DmanHHlDny/SweYrSes4xQeWoF6TV+AoThyDHQpz+hQ=";
  };

  nativeBuildInputs = [
@@ -2,7 +2,7 @@
  lib,
  stdenv,
  # LLVM version closest to ROCm fork to override
-  llvmPackages_20,
+  llvmPackages_22,
  overrideCC,
  lndir,
  rocm-device-libs,
@@ -22,23 +22,27 @@
  # leaving compressed line tables (-g1 -gz) unstripped
  # TODO: Should also apply to downstream packages which use rocmClangStdenv?
  profilableStdenv ? false,
+  # FIXME: proper two-stage bootstrap & PGO/BOLT/LTO
+  # LTO currently disabled due to llvm 22 vs 22.1 bitcode mismatch between
+  # llvmPackages_22 (22.1) and ROCm LLVM (22.0). Uses new bitcode attr
+  # 105 nocreateundeforpoison fails to link hipify.
  # Whether to use LTO when building the ROCm toolchain
  # Slows down this toolchain's build, for typical ROCm usecase
  # time saved building composable_kernel and other heavy packages
  # will outweight that. ~3-4% speedup multiplied by thousands
  # of corehours.
-  withLto ? true,
+  withLto ? false,
  # whether rocm stdenv uses libcxx (clang c++ stdlib) instead of gcc stdlibc++
  withLibcxx ? false,
 }:

 let
-  version = "7.1.1";
+  version = "7.2.0";
  # major version of this should be the clang version ROCm forked from
-  rocmLlvmVersion = "20.0.0-rocm";
+  rocmLlvmVersion = "22.0.0-rocm";
  # llvmPackages_base version should match rocmLlvmVersion
  # so libllvm's bitcode is compatible with the built toolchain
-  llvmPackages_base = llvmPackages_20;
+  llvmPackages_base = llvmPackages_22;
  llvmPackagesNoBintools = llvmPackages_base.override {
    bootBintools = null;
    bootBintoolsNoLibc = null;
@@ -115,7 +119,7 @@ let
    owner = "ROCm";
    repo = "llvm-project";
    rev = "rocm-${version}";
-    hash = "sha256-CfknIRVeR1bCKh1xzXKl3ehVp0kWT0uGrI9C1HTSKVo=";
+    hash = "sha256-I/Bxq1JjU9N4h3vXj/tbD3xPYY4+N7QzYA8UTIq1EL0=";
  };
  llvmMajorVersion = lib.versions.major rocmLlvmVersion;
  # An llvmPackages (pkgs/development/compilers/llvm/) built from ROCm LLVM's source tree
@@ -197,10 +201,23 @@ let
        rm -rf $out/lib/cmake $out/lib/lib*.a
        mkdir -p $out/lib/clang/${llvmMajorVersion}/lib/linux/
        ln -s $out/lib/linux/libclang_rt.* $out/lib/clang/${llvmMajorVersion}/lib/linux/
+        mkdir -p $out/lib/clang/${llvmMajorVersion}/lib/${stdenv.hostPlatform.config}/
+        for f in $out/lib/linux/*-${stdenv.hostPlatform.parsed.cpu.name}.*; do
+          [ -e "$f" ] || continue
+          base="$(basename "$f" | sed 's/-${stdenv.hostPlatform.parsed.cpu.name}\././')"
+          ln -sf "$f" "$out/lib/clang/${llvmMajorVersion}/lib/${stdenv.hostPlatform.config}/$base"
+        done

        find $out -type f -exec sed -i "s|${cc.out}|$out|g" {} +
        find $out -type f -exec sed -i "s|${cc.dev}|$out|g" {} +

+        # Clang config file: redirect resource dir to the sysroot (where compiler-rt
+        # lives) and set GCC install prefix for header/library search
+        cat > $out/bin/${stdenv.hostPlatform.config}.cfg <<CFG
+        -resource-dir=$out/lib/clang/${llvmMajorVersion}
+        ${lib.optionalString (!withLibcxx) "--gcc-toolchain=${gcc-prefix}"}
+        CFG
+
        ${lib.getExe rdfind} -makesymlinks true ${
          builtins.concatStringsSep " " (map (x: "${x}/lib") paths)
        } $out/ # create links *within* the sysroot to save space
@@ -233,6 +250,9 @@ let
    (lib.cmakeFeature "LLVM_ENABLE_ZSTD" "FORCE_ON")
    # required for threaded ThinLTO to work
    (lib.cmakeBool "LLVM_ENABLE_THREADS" true)
+    # third-party/benchmark is broken in rocm-7.2.0
+    # error: '__COUNTER__' is a C2y extension [-Werror,-Wc2y-extensions]
+    (lib.cmakeBool "LLVM_INCLUDE_BENCHMARKS" false)
    # LLVM tries to call git to embed VCS info if FORCE_VC_ aren't set
    (lib.cmakeFeature "LLVM_FORCE_VC_REVISION" "rocm-${version}")
    (lib.cmakeFeature "LLVM_FORCE_VC_REPOSITORY" "https://github.com/ROCm/llvm-project")
@@ -276,11 +296,14 @@ overrideLlvmPackagesRocm (s: {
  libllvm = (s.prev.libllvm.override { }).overrideAttrs (old: {
    patches = old.patches ++ [
      (fetchpatch {
-        # fix compile error in tools/gold/gold-plugin.cpp
-        name = "gold-plugin-fix.patch";
-        url = "https://github.com/llvm/llvm-project/commit/b0baa1d8bd68a2ce2f7c5f2b62333e410e9122a1.patch";
-        hash = "sha256-yly93PvGIXOnFeDGZ2W+W6SyhdWFM6iwA+qOeaptrh0=";
-        relative = "llvm";
+        # Revert of a patch that cause perf regression by pessimising loop unrolling decisions
+        # See PR and issue discussion
+        # https://github.com/ROCm/rocm-systems/issues/2865 https://github.com/ROCm/llvm-project/pull/1349
+        name = "rocm-llvm-revert-unrolling-regression.patch";
+        url = "https://github.com/ROCm/llvm-project/commit/f58b06dce1f9c15707c5f808fd002e18c2accf7e.patch";
+        hash = "sha256-pH+3C7PSDqNfOF014sA5Rvm+sc2IJMQJfysS2bvj/o0=";
+        # stripLen instead of relative to avoid filterdiff mangling /dev/null on the deleted test file
+        stripLen = 1;
      })
      ./perf-increase-namestring-size.patch
      # TODO: consider reapplying "Don't include aliases in RegisterClassInfo::IgnoreCSRForAllocOrder"
@@ -342,7 +365,17 @@ overrideLlvmPackagesRocm (s: {
        passthru = old.passthru // {
          inherit gcc-prefix;
        };
-        patches = old.patches ++ [
+        patches = [
+          (fetchpatch {
+            # [clang][cmake] Add option to control scan-build-py installation (#172727)
+            name = "clang-scan-build-py-configurable.patch";
+            url = "https://github.com/llvm/llvm-project/commit/f5759eeb63a3a5ce7d555c13c3126cea84e0c7b1.patch";
+            relative = "clang";
+            hash = "sha256-73IDPGZWKX4vny3x5FJ3/NQw8XRad9UNwfYkvQdMB4s=";
+          })
+        ]
+        ++ old.patches
+        ++ [
          # Never add FHS include paths
          ./clang-bodge-ignore-systemwide-incls.diff
          # Prevents builds timing out if a single compiler invocation is very slow but
@@ -352,11 +385,23 @@ overrideLlvmPackagesRocm (s: {
          ./perf-shorten-gcclib-include-paths.patch
          (fetchpatch {
            # [ClangOffloadBundler]: Add GetBundleIDsInFile to OffloadBundler
-            sha256 = "sha256-G/mzUdFfrJ2bLJgo4+mBcR6Ox7xGhWu5X+XxT4kH2c8=";
-            url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/6d296f879b0fed830c54b2a9d26240da86c8bb3a.patch";
+            hash = "sha256-OsarDZXuJ5vAXTP4i0NBUeK/r6tQPumaqmMWkf29UtM=";
+            url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/c7de294b0d1d25f277f9d1cbb2c9e09c7600e210.patch";
            relative = "clang";
          })
        ];
+        # ROCm 7.2 commits 4dda51261a6 "Replace hostexec with upstream rpc"
+        # and 2ca1509d6d2 "Put the RTL, Back!", added CGEmitEmissaryExec.cpp which
+        # includes ../../openmp/device/include/EmissaryIds.h, breaking
+        # standalone clang builds. The upstream PR llvm/llvm-project#175265
+        # ("[OpenMP] support for Emissary APIs") moves EmissaryIds.h into
+        # clang/lib/Headers/ so this should not be needed once that lands
+        # and ROCm rebases onto it.
+        postUnpack = ''
+          ${old.postUnpack or ""}
+          mkdir -p "''${sourceRoot}/openmp/device/include"
+          ln -s "${llvmSrc}/openmp/device/include/EmissaryIds.h" "''${sourceRoot}/openmp/device/include/"
+        '';
        hardeningDisable = [ "all" ];
        nativeBuildInputs = old.nativeBuildInputs ++ [
          removeReferencesTo
@@ -375,14 +420,7 @@ overrideLlvmPackagesRocm (s: {
        # https://github.com/llvm/llvm-project/blob/6976deebafa8e7de993ce159aa6b82c0e7089313/clang/cmake/caches/DistributionExample-stage2.cmake#L9-L11
        cmakeFlags =
          # TODO: Remove in followup, tblgen now works correctly but would rebuild
-          (builtins.filter tablegenUsage old.cmakeFlags)
-          ++ commonCmakeFlags
-          ++ lib.optionals (!withLibcxx) [
-            # FIXME: Config file in rocm-toolchain instead of GCC_INSTALL_PREFIX?
-            # Expected to be fully removed eventually
-            "-DUSE_DEPRECATED_GCC_INSTALL_PREFIX=ON"
-            "-DGCC_INSTALL_PREFIX=${gcc-prefix}"
-          ];
+          (builtins.filter tablegenUsage old.cmakeFlags) ++ commonCmakeFlags;
        preFixup = ''
          ${toString old.preFixup or ""}
          moveToOutput "lib/lib*.a" "$dev"
@@ -437,14 +475,6 @@ overrideLlvmPackagesRocm (s: {
      isGNU = false;
    };
  compiler-rt-libc = s.prev.compiler-rt-libc.overrideAttrs (old: {
-    patches = old.patches ++ [
-      # fix build with glibc >= 2.42
-      (fetchpatch {
-        url = "https://github.com/llvm/llvm-project/commit/59978b21ad9c65276ee8e14f26759691b8a65763.patch";
-        hash = "sha256-ys5SMLfO3Ay9nCX9GV5yRCQ6pLsseFu/ZY6Xd6OL4p0=";
-        relative = "compiler-rt";
-      })
-    ];
    meta = old.meta // llvmMeta;
  });
  compiler-rt = s.final.compiler-rt-libc;
@@ -17,6 +17,6 @@ index 33f12fb2d075d..ba35f65210c7a 100644
   if (TCArgs.hasFlag(options::OPT_offload_compress,
 -                     options::OPT_no_offload_compress, false))
 +                     options::OPT_no_offload_compress, true))
-     CmdArgs.push_back("-compress");
+     CmdArgs.push_back("--compress");
   if (TCArgs.hasArg(options::OPT_v))
     CmdArgs.push_back("-verbose");
@@ -1,4 +1,4 @@
-From ef6c5b353861be727c98f1319c81d0c6b609d644 Mon Sep 17 00:00:00 2001
+From 8c64deff2a82fcca5b521362193a1fd8b1a0dca7 Mon Sep 17 00:00:00 2001
 From: Luna Nova <git@lunnova.dev>
 Date: Tue, 17 Dec 2024 04:29:11 -0800
 Subject: [PATCH] HACK: Get canonical GCC include path so doesn't have
@@ -6,21 +6,20 @@ Subject: [PATCH] HACK: Get canonical GCC include path so doesn't have

 This allows more of the strings used in compilation to fit inside
 fixed size stack allocated buffers instead of spilling into the heap
-
 ---
- clang/lib/Driver/ToolChains/Gnu.cpp | 16 ++++++++++------
+ lib/Driver/ToolChains/Gnu.cpp | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

-diff --git a/clang/lib/Driver/ToolChains/Gnu.cpp b/clang/lib/Driver/ToolChains/Gnu.cpp
-index af9fd46f0ce7b..a63a7a93f6a78 100644
+diff --git a/lib/Driver/ToolChains/Gnu.cpp b/lib/Driver/ToolChains/Gnu.cpp
+index 07df9864636..b1201b999f0 100644
 --- a/lib/Driver/ToolChains/Gnu.cpp
 +++ b/lib/Driver/ToolChains/Gnu.cpp
-@@ -3394,29 +3394,33 @@ bool Generic_GCC::addLibStdCXXIncludePaths(Twine IncludeDir, StringRef Triple,
-   if (!getVFS().exists(IncludeDir))
+@@ -3326,29 +3326,33 @@ static bool addLibStdCXXIncludePaths(llvm::vfs::FileSystem &vfs,
+   if (!vfs.exists(IncludeDir))
     return false;
 
 +  SmallString<260> CanonicalIncludeDir;
-+  if (getVFS().getRealPath(IncludeDir, CanonicalIncludeDir))
+  if (vfs.getRealPath(IncludeDir, CanonicalIncludeDir))
 +    return false;
 +
   // Debian native gcc uses g++-multiarch-incdir.diff which uses
@@ -34,24 +33,27 @@ index af9fd46f0ce7b..a63a7a93f6a78 100644
 -      (Include + "/" + Triple + Dir.substr(Include.size()) + IncludeSuffix)
 +      (Include + "/" + Triple + CanonicalIncludeDir.substr(Include.size()) + IncludeSuffix)
           .str();
-   if (DetectDebian && !getVFS().exists(Path))
+   if (DetectDebian && !vfs.exists(Path))
     return false;
 
   // GPLUSPLUS_INCLUDE_DIR
-  addSystemInclude(DriverArgs, CC1Args, IncludeDir);
-+  addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir);
+-  ToolChain::addSystemInclude(DriverArgs, CC1Args, IncludeDir);
+  ToolChain::addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir);
   // GPLUSPLUS_TOOL_INCLUDE_DIR. If Triple is not empty, add a target-dependent
   // include directory.
   if (DetectDebian)
-     addSystemInclude(DriverArgs, CC1Args, Path);
+     ToolChain::addSystemInclude(DriverArgs, CC1Args, Path);
   else if (!Triple.empty())
-     addSystemInclude(DriverArgs, CC1Args,
-                     IncludeDir + "/" + Triple + IncludeSuffix);
-+                     CanonicalIncludeDir + "/" + Triple + IncludeSuffix);
+     ToolChain::addSystemInclude(DriverArgs, CC1Args,
+-                                IncludeDir + "/" + Triple + IncludeSuffix);
+                                CanonicalIncludeDir + "/" + Triple + IncludeSuffix);
   // GPLUSPLUS_BACKWARD_INCLUDE_DIR
-  addSystemInclude(DriverArgs, CC1Args, IncludeDir + "/backward");
-+  if (getVFS().exists(CanonicalIncludeDir + "/backward"))
-+    addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir + "/backward");
+-  ToolChain::addSystemInclude(DriverArgs, CC1Args, IncludeDir + "/backward");
+  if (vfs.exists(CanonicalIncludeDir + "/backward"))
+    ToolChain::addSystemInclude(DriverArgs, CC1Args, CanonicalIncludeDir + "/backward");
   return true;
 }
 
+-- 
+2.52.0
+
@@ -58,7 +58,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "migraphx";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -74,7 +74,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "AMDMIGraphX";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-s6w4bF7koK4wnf6leVKBzwIB4X2ROHa3EgX6XuJIAew=";
+    hash = "sha256-FBAVsk4x3ATLDYtfzcPUPnUTwe36maIAY1/FwqG6jD0=";
  };

  nativeBuildInputs = [
@@ -4,7 +4,7 @@
  callPackage,
  fetchFromGitHub,
  fetchpatch,
-  rocmUpdateScript,
+  fetchurl,
  runCommand,
  pkg-config,
  cmake,
@@ -43,7 +43,7 @@
 let
  # FIXME: cmake files need patched to include this properly
  cFlags = "-Wno-documentation-pedantic --offload-compress -I${hipblas-common}/include -I${hipblas}/include -I${roctracer}/include -I${nlohmann_json}/include -I${sqlite.dev}/include -I${rocrand}/include";
-  version = "7.1.1";
+  version = "7.2.0";

  # Targets outside this list will get
  # error: use of undeclared identifier 'CK_BUFFER_RESOURCE_3RD_DWORD'
@@ -65,42 +65,12 @@ let
    "gfx1201"
  ] gpuTargets;

-  kdbTargets = lib.intersectLists [
-    "gfx900"
-    "gfx906"
-    "gfx908"
-    "gfx90a"
-    "gfx942"
-    "gfx1030"
-  ] gpuTargets;
-
  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "MIOpen";
    rev = "rocm-${version}";
-    hash = "sha256-g0AEpuUiwKCu/doiRh9mW34W04m4ynHoarSyl6tR/aE=";
-    fetchLFS = true;
+    hash = "sha256-OwBFzUruHHeJD7n3zLs/NsU5cNEjwkwFgim4m2/1hR0=";
    fetchSubmodules = true;
-    # WORKAROUND: .lfsconfig is incorrectly set to exclude everything upstream
-    leaveDotGit = true;
-    # FIXME: if someone can reduce the level of awful here that would be really nice
-    postFetch = ''
-      export HOME=$(mktemp -d)
-      cd $out
-      git remote add origin $url
-      git fetch origin +refs/tags/rocm-${version}:refs/tags/rocm-${version}
-      git clean -fdx
-      git switch -c rocm-${version} refs/tags/rocm-${version}
-      git config lfs.fetchexclude "none"
-      rm .lfsconfig
-      git lfs install
-      git lfs track "*.kdb.bz2"
-      git lfs fetch --include="src/kernels/**"
-      git lfs pull --include="src/kernels/**"
-      git lfs checkout
-
-      rm -rf .git
-    '';
  };

  latex = lib.optionalAttrs buildDocs (
@@ -120,19 +90,27 @@ let
    )
  );

-  kernelDatabases = lib.genAttrs kdbTargets (
-    target:
-    runCommand "miopen-${target}.kdb" { preferLocalBuild = true; } ''
-      ${lbzip2}/bin/lbzip2 -ckd ${src}/src/kernels/${target}.kdb.bz2 > $out
-    ''
-  );
+  # Kernel databases moved from Git LFS to DVC (anonymous s3 bucket s3://therock-dvc/rocm-libraries)
+  fetchKdb =
+    name:
+    { url, hash }:
+    runCommand "miopen-${name}.kdb" { preferLocalBuild = true; } ''
+      ${lbzip2}/bin/lbzip2 -ckd ${
+        fetchurl {
+          inherit url hash;
+          name = "${name}.kdb.bz2";
+        }
+      } > $out
+    '';
+
+  kdbs = lib.mapAttrs fetchKdb (import ./kdbs.nix);

  linkKDBsTo =
    targetPath:
    lib.concatStringsSep "" (
-      map (target: ''
-        ln -sf ${kernelDatabases.${target}} ${targetPath}/${target}.kdb
-      '') kdbTargets
+      lib.mapAttrsToList (name: kdb: ''
+        ln -sf ${kdb} ${targetPath}/${name}.kdb
+      '') kdbs
    );
 in
 stdenv.mkDerivation (finalAttrs: {
@@ -153,14 +131,13 @@ stdenv.mkDerivation (finalAttrs: {
      url = "https://github.com/ROCm/MIOpen/commit/3413d2daaeb44b7d6eadcc03033a5954a118491e.patch";
      hash = "sha256-ST4snUcTmmSI1Ogx815KEX9GdMnmubsavDzXCGJkiKs=";
    })
-    # FIXME: We need to rebase or drop this arch compat patch
-    # https://github.com/ROCm/MIOpen/issues/3540 suggests that
-    # arch compat patching doesn't work correctly for gfx1031
-    # (fetchpatch {
-    #   name = "Extend-MIOpen-ISA-compatibility.patch";
-    #   url = "https://github.com/GZGavinZhao/MIOpen/commit/416088b534618bd669a765afce59cfc7197064c1.patch";
-    #   hash = "sha256-OwONCA68y8s2GqtQj+OtotXwUXQ5jM8tpeM92iaD4MU=";
-    # })
+    (fetchpatch {
+      # [miopen] Extend HIP ISA compatibility
+      name = "Extend-MIOpen-ISA-compatibility.patch";
+      url = "https://github.com/GZGavinZhao/rocm-libraries/commit/02f0fedffdc197f146dd45f41e10990a00cde3ee.patch";
+      hash = "sha256-My32iZw75rvB4fyvUJJ2kw2bU9/39awGteFGjzijixw=";
+      relative = "projects/miopen";
+    })
  ];

  outputs = [
@@ -308,11 +285,7 @@ stdenv.mkDerivation (finalAttrs: {
      inherit frugally-deep nlohmann_json;
    };
  };
-  passthru.updateScript = rocmUpdateScript {
-    name = finalAttrs.pname;
-    inherit (finalAttrs.src) owner;
-    inherit (finalAttrs.src) repo;
-  };
+  passthru.updateScript = ./update.sh;

  meta = {
    description = "Machine intelligence library for ROCm";
@@ -0,0 +1,27 @@
+# Generated by update.sh
+{
+  gfx1030 = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/22/c4bd38cc34d3906add061ee57b099b";
+    hash = "sha256-LESExp7sFl7osCM7HdVMpzOE0dU7953e2yzgeHSGgfI=";
+  };
+  gfx900 = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/cf/9c12dafe25b15ea1e93088b7b061af";
+    hash = "sha256-lFFB5GQFTjUnx3JDx816yH+96OGRt9XvBzfLFFHtUrE=";
+  };
+  gfx906 = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/04/b3be6a15c700ff382e0a5e136010e1";
+    hash = "sha256-cn1ODLcUQi+/QJUlWbELukMUj4+b1vjjRhhTseL9tdM=";
+  };
+  gfx908 = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/35/b9082b72e661dc5e37f14de1d9d4ed";
+    hash = "sha256-pI30zpZL3OEP3aeLdVpM00nHeTyD5K+XDH47g0W2A+M=";
+  };
+  gfx90a = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/a1/1615323e06c5774fe19cf34c51baef";
+    hash = "sha256-TkzSgY3V/Xs5YXfr69TrL8JjqDr1MZMgeBeQ/e2kEjE=";
+  };
+  gfx942 = {
+    url = "https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5/6a/fe97d7bd441fe15280b28b6e3b1748";
+    hash = "sha256-fyhnR3+xPFkHLN36dI9UYIhqoVstt5RpareGn2S9lMM=";
+  };
+}
@@ -0,0 +1,48 @@
+#!/usr/bin/env nix-shell
+#!nix-shell -i bash -p curl jq nix common-updater-scripts
+
+set -euo pipefail
+
+gh_curl() { curl ${GITHUB_TOKEN:+-u ":$GITHUB_TOKEN"} -sf "$@"; }
+
+version="${1:-$(gh_curl "https://api.github.com/repos/ROCm/MIOpen/releases?per_page=4" \
+  | jq -re 'map(.tag_name // .name)
+    | map(select(test("^rocm-[0-9]+\\.[0-9]+(\\.[0-9]+)?$")))
+    | first | ltrimstr("rocm-")')}"
+
+# main source version
+echo "Updating MIOpen to $version" >&2
+update-source-version rocmPackages.miopen "$version" --ignore-same-hash
+
+# kdbs.nix from s3 bucket
+cd "$(dirname "${BASH_SOURCE[0]}")"
+bucket="https://therock-dvc.s3.amazonaws.com/rocm-libraries/files/md5"
+
+# Fetch directory listing; extract name + download_url for .dvc files
+dvc_entries=$(gh_curl "https://api.github.com/repos/ROCm/MIOpen/contents/src/kernels?ref=rocm-${version}" \
+  | jq -r '.[] | select(.name | endswith(".kdb.bz2.dvc")) | "\(.name)\t\(.download_url)"')
+
+[[ -n "$dvc_entries" ]] || { echo "No .kdb.bz2.dvc files found, upstream likely changed packaging again" >&2; exit 1; }
+
+{
+  echo "# Generated by update.sh"
+  echo "{"
+
+  while IFS=$'\t' read -r dvc_file raw_url; do
+    arch="${dvc_file%.kdb.bz2.dvc}"
+    md5=$(curl -sf "$raw_url" | awk '/- md5:/{print $3}')
+    url="${bucket}/${md5:0:2}/${md5:2}"
+    sri=$(nix --extra-experimental-features nix-command store prefetch-file --json "$url" --name "${arch}.kdb.bz2" | jq -r .hash)
+    echo "  $arch: $sri" >&2
+    cat <<EOF
+  $arch = {
+    url = "$url";
+    hash = "$sri";
+  };
+EOF
+  done <<< "$dvc_entries"
+
+  echo "}"
+} > kdbs.nix
+
+echo "Wrote kdbs.nix" >&2
@@ -40,13 +40,13 @@ stdenv.mkDerivation (finalAttrs: {
        "cpu"
    );

-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "MIVisionX";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-aC6GUPK6ZSOAx+PHHB4gLPNAG5U/kapqX7YWilusDw8=";
+    hash = "sha256-rdYQt001kFi9FjVL5R8XVzT+Y+rZVeMsbjXCJEB887A=";
  };

  patches = [
@@ -40,7 +40,7 @@ in
 # infiniband ib_peer_mem support isn't in the mainline kernel but is carried by some distros
 stdenv.mkDerivation (finalAttrs: {
  pname = "rccl${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -58,7 +58,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rccl";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-3u7D3Gre1n+4Lf+cK+RMfCUM9c46pXZjdhGOrwIKM0w=";
+    hash = "sha256-Kpobn0fR7/mXGVWYjNQ35sToRWLfOQTHRbjuSYj1of4=";
  };

  requiredSystemFeatures = [ "big-parallel" ]; # Very resource intensive LTO
@@ -49,7 +49,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rdc";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -65,7 +65,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rdc";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-oJS0fBeISpgjZkaMfFl/Yq7mPOsdaZDqiRjSPv7kT1Q=";
+    hash = "sha256-gF17mcbsqsL2K9pwTAVG9E6nVY4pi2lC88WyPa0xMes=";
  };

  patches = [
@@ -22,7 +22,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocalution";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -41,7 +41,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocALUTION";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-UkXbcbhaTulCqW1FZHwTyUZLFSnj7WmLYqlPqu5m6YM=";
+    hash = "sha256-7ZlY0GVIsUJE8OOAkQwQ4pXNHKzgCQomQOET9+TKQN4=";
  };

  nativeBuildInputs = [
@@ -40,13 +40,13 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rocblas${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocBLAS";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-obKypbYmnSeOtOr7g0pxOz02YfzZ0bGugTtznkeHz14=";
+    hash = "sha256-oY6yELFnpnDDksu63Go88TGLAgC64tUaPIbaRgD2qus=";
  };

  outputs = [ "out" ] ++ lib.optional buildBenchmarks "benchmark" ++ lib.optional buildTests "test";
@@ -137,9 +137,9 @@ stdenv.mkDerivation (finalAttrs: {
  patches = [
    (fetchpatch {
      name = "Extend-rocBLAS-HIP-ISA-compatibility.patch";
-      url = "https://github.com/GZGavinZhao/rocm-libraries/commit/2850f22f80f90c9e498f520608a82989a4932ec3.patch";
-      hash = "sha256-SPsdEwGe+r8bQudkChRzBDAgu3tPQWFweZCgzh+4nOE=";
-      stripLen = 2;
+      url = "https://github.com/GZGavinZhao/rocm-libraries/commit/49f21f3cfe7eb4b8ac724eef81fa2cae97a3c22e.patch";
+      hash = "sha256-vAVVpTwt49lGHu2YopR1X68v5LwFOjUYuSC4ucBpFGg=";
+      relative = "projects/rocblas";
    })
  ];

@@ -47,7 +47,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rocdbgapi";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -62,7 +62,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "ROCdbgapi";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-RwYZJPwGhNtSSvmSgy0AsNTc98cav0/u9jH5f93sB9M=";
+    hash = "sha256-KqvhwfIv8pbr8WbnfAKl71fg5yxbwYcpzZcGU9Htdkc=";
  };

  nativeBuildInputs = [
@@ -20,13 +20,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocfft${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocFFT";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-1Ho3b5NmnzfLhDnvH6FECigs6OgpbJrxw4EnqKyaHA0=";
+    hash = "sha256-AFBV5tlhNYH9+d0KG7mCjSifKfRt+8c1ie5wASTiQXM=";
  };

  nativeBuildInputs = [
@@ -23,13 +23,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocgdb";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "ROCgdb";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-LDVGO++voqwVMM6RcRtgbXTUvFLTyg/TFdSZanv5Xdc=";
+    hash = "sha256-oml/HLExnnjh7+axeWPZRWecpwK2BnzVOaGvXYhrxKs=";
  };

  nativeBuildInputs = [
@@ -9,13 +9,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocm-cmake";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocm-cmake";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-im6UO0crO0Jc27zkTsdvJYPHit8IGlw/vDPGrmP1XqY=";
+    hash = "sha256-gY6jzIIN1pSXGbCMN6y35Q/VJgbIqWDRjD8aI/fc1L0=";
  };

  nativeBuildInputs = [ cmake ];
@@ -30,20 +30,20 @@ stdenv.mkDerivation (finalAttrs: {
  patches = [
    # [Comgr] Extend ISA compatibility
    (fetchpatch {
-      sha256 = "sha256-dgow0kwSWM1TnkqWOZDRQrh5nuF8p5jbYyOLCpQsH4k=";
-      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/a439e4f37ce71de48d4a979594276e3be0e6278f.patch";
+      hash = "sha256-X2VPGigK582J+a/u2Kg74w25/+CTpVWU9D3Eqgnb2PU=";
+      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/7002dc04863d38c57cfd2e6fc60a1cf5a613fd8e.patch";
      relative = "amd/comgr";
    })
-    #[Comgr] Extend ISA compatibility for CCOB
+    # [Comgr] Extend ISA compatibility for CCOB
    (fetchpatch {
-      sha256 = "sha256-PCi0QHLiEQCTIYRtSSbhOjXANJ3zC3VLdMED1BEfQeg=";
-      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/fa80abb77d5ae6f8d89ab956e7ebda9c802a804f.patch";
+      hash = "sha256-/50I+PqxL3oaQMqg5vR7+ibUcXO1SvfXBdw/sybRt1o=";
+      url = "https://github.com/GZGavinZhao/rocm-llvm-project/commit/2c1e44fc3eacadcafdd4ada3e3184a092b6f26c5.patch";
      relative = "amd/comgr";
    })
  ];

  postPatch = ''
-    substituteInPlace cmake/opencl_pch.cmake \
+    substituteInPlace cmake/opencl_header.cmake \
      --replace-fail "\''${CLANG_CMAKE_DIR}/../../../" "${llvm.clang-unwrapped.lib}"
  '';

@@ -17,13 +17,13 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rocm-core";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocm-core";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-3WQPR7v7LRoo4ioCXzkUEqnJdXA3uxSgoZZQvEG5Gqg=";
+    hash = "sha256-AqD6MByFAtY6IzAFXdCrCANPXqROaGHybdKixv3NbXE=";
  };

  patches = [
@@ -41,6 +41,7 @@ stdenv.mkDerivation (finalAttrs: {
    "-DROCM_VERSION=${finalAttrs.version}"
    "-DBUILD_ID=${finalAttrs.env.BUILD_ID}"
    "-DCMAKE_INSTALL_LIBDIR=lib"
+    "-DCMAKE_INSTALL_LIBEXECDIR=libexec"
    "-DCMAKE_INSTALL_INCLUDEDIR=include"
    "-DCMAKE_INSTALL_BINDIR=bin"
  ];
@@ -37,12 +37,6 @@ stdenv.mkDerivation (finalAttrs: {

  patches = [
    ./cmake.patch
-    (fetchpatch {
-      name = "cmake-4-compat-dont-set-cmp0053.patch";
-      url = "https://github.com/ROCm/llvm-project/commit/a18cc4c7cb51f94182b6018c7c73acde1b8ebddb.patch";
-      hash = "sha256-kp/Ld0IhjWgRbRR9R/CKdkI9ELvPkQSAMqPsAPFxzhM=";
-      relative = "amd/device-libs";
-    })
  ];

  nativeBuildInputs = [
@@ -16,13 +16,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocm-runtime";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "ROCR-Runtime";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-gqe1trGc/Cu1XFA4aYjOzFurUgebLbdTHEJi4iw2+kk=";
+    hash = "sha256-6xELKQ/uqAoorsCR/H7d8iNK7LsVNsW2DRRZo5cU7UM=";
  };

  cmakeBuildType = "RelWithDebInfo";
@@ -58,8 +58,8 @@ stdenv.mkDerivation (finalAttrs: {
    ./queue-failure.patch
    (fetchpatch {
      # [PATCH] rocr: Extend HIP ISA compatibility check
-      sha256 = "sha256-8r2Lb5lBfFaZC3knCxfXGcnkzNv6JxOKyJn2rD5gus4=";
-      url = "https://github.com/GZGavinZhao/rocm-systems/commit/c13cd118fcc8e0bc9ae8de62897542dca7352b71.patch";
+      hash = "sha256-8r2Lb5lBfFaZC3knCxfXGcnkzNv6JxOKyJn2rD5gus4=";
+      url = "https://github.com/GZGavinZhao/rocm-systems/commit/dcef23cf896f4dcbc7ed81abeaa4ec2208dcdd8c.patch";
      relative = "projects/rocr-runtime";
    })
    (fetchpatch {
@@ -67,17 +67,6 @@ stdenv.mkDerivation (finalAttrs: {
      url = "https://github.com/ROCm/ROCR-Runtime/commit/41bfc66aef437a5b349f71105fa4b907cc7e17d5.patch";
      hash = "sha256-A7VhPR3eSsmjq2cTBSjBIz9i//WiNjoXm0EsRKtF+ns=";
    })
-    (fetchpatch {
-      # [PATCH] hsakmt: bump vgpr count for gfx1151 (#1807) (#1986)
-      # We apply only the change that adds a GFX_VERSION_GFX1151 define but
-      # *not* the default vgpr change which causes crashes on old kernels
-      # This was later partially reverted in # [PATCH] Revert "hsakmt: bump vgpr count for gfx1151 (#1807) (#1986)"
-      name = "rocr-runtime-gfx1151-vgpr.patch";
-      url = "https://github.com/ROCm/rocm-systems/commit/09ba45b3f43ec333a84a0ca178fcd1e3ea9400a9.patch";
-      relative = "projects/rocr-runtime";
-      includes = [ "libhsakmt/src/libhsakmt.h" ];
-      hash = "sha256-/V5i+sr88n7fK4yNjR/FpY0ZpiEG5xAD6Oq+9ZOikd4=";
-    })
    (fetchpatch {
      # [PATCH] hsakmt: Expose and use CWSR and Control stack sizes (#2200)
      # Fixes potential mismatches between stack sizes configured in amdgpu kernel module
@@ -11,13 +11,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocm-smi";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocm_smi_lib";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-/3to1+45n+D8snwrFVhApSMWp/rIXmTFK+BiA0mcKZo=";
+    hash = "sha256-IeOutA/Mpt/75VpdcfPnZrWvx6GG/JLaQNfBrNKb+Mw=";
  };

  propagatedBuildInputs = [
@@ -12,14 +12,14 @@
 }:

 stdenv.mkDerivation (finalAttrs: {
-  version = "7.1.1";
+  version = "7.2.0";
  pname = "rocminfo";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocminfo";
    rev = "rocm-${finalAttrs.version}";
-    sha256 = "sha256-SuW34m2ep69+dj1rb0vqfQcK83FBtjlMLqSMCUZltU4=";
+    hash = "sha256-Md91iH2YWU7ziNE//xxKuk2RrXdujO9mY6SZY5R9TqA=";
  };

  strictDeps = true;
@@ -42,7 +42,7 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rocmlir${suffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -55,7 +55,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocMLIR";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-A9vUvsEZrZlNEW4cscF66L48rJQ1zJYmIzwXQ2QzJ3s=";
+    hash = "sha256-gg37odxZETpH7HqY+44KQta6f0vxfcxhzMAqxU6vicY=";
  };

  nativeBuildInputs = [
@@ -15,7 +15,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocprim";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocPRIM";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-/O6aEF23IaUSVzEDRqIlpYpxNCSTa/CwVDgXOhvTrQc=";
+    hash = "sha256-/6x4S8W4MbWY+M2b46Tmebj2echBIs31OnKq+50M21M=";
  };

  nativeBuildInputs = [
@@ -18,13 +18,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocprofiler-register";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocprofiler-register";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-XlBz5rDMvnPxyCiqpEpIUWzbHJ1BbMzeCJTrP5kKf+w=";
+    hash = "sha256-6jr6dfmCjeP5PCNfLDMmh8IO9Hxz6RvLlacNwIWQduQ=";
    fetchSubmodules = true;
  };

@@ -44,13 +44,13 @@ let
 in
 stdenv.mkDerivation (finalAttrs: {
  pname = "rocprofiler";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocprofiler";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-V+E+6D/mBW/NVdHSBVUVMG5Z/QOF73vfZ5rryW2VBtI=";
+    hash = "sha256-W32VAdtZdFmkHeUwyHIIHk8yRnABq6gUaHLZ9Bj8kYI=";
    fetchSubmodules = true;
  };

@@ -12,13 +12,13 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocr-debug-agent";
-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rocr_debug_agent";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-zW92HyK4nGcDaJxmBaHeURAivs6dU059WxaywL6dAk4=";
+    hash = "sha256-w2Zg4kpuKy68DkVGsBTUsjRZoV/Y/Z3Q8s0oSIR3Ask=";
  };

  nativeBuildInputs = [
@@ -15,7 +15,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocrand${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocRAND";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-QhHxpjr3SqeSSMS4QmnDq8zIxEeUtrgwb9JG84jfJgY=";
+    hash = "sha256-BX1A74ZBXE0fqYqA7WijezCezUOpEciUm6k2vhFQlA4=";
  };

  nativeBuildInputs = [
@@ -37,7 +37,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocsolver${clr.gpuArchSuffix}";
-  version = "7.1.0";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -53,7 +53,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocSOLVER";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-U5RrV90UZiIUgLYleLJ6rFHWRFMQNyh3zezRusj7T0M=";
+    hash = "sha256-aCmNTF8P0Y7piPkRnu/+DWFhTd7X8mDcaQpotfw/4hM=";
  };

  nativeBuildInputs = [
@@ -7,6 +7,7 @@
  cmake,
  rocm-cmake,
  rocprim,
+  rocblas,
  clr,
  gfortran,
  gtest,
@@ -19,7 +20,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocsparse${clr.gpuArchSuffix}";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -35,7 +36,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocSPARSE";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-ZNeFslIQZlFCNqWHmrpcxDuOvVu1cjiEd8b6ZNXcZzo=";
+    hash = "sha256-jmrX1a09DlQTkuXwkrInt9pKX5pk3oCO88S0pKXXhO0=";
  };

  nativeBuildInputs = [
@@ -48,6 +49,7 @@ stdenv.mkDerivation (finalAttrs: {

  buildInputs = [
    rocprim
+    rocblas
  ]
  ++ lib.optionals (buildTests || buildBenchmarks) [
    gtest
@@ -6,7 +6,7 @@

 {
  matrix-01 = fetchzip {
-    sha256 = "sha256-AHur5ZIDZTFRrO2GV0ieXrffq4KUiGWiZ59pv0fUtEQ=";
+    hash = "sha256-AHur5ZIDZTFRrO2GV0ieXrffq4KUiGWiZ59pv0fUtEQ=";

    urls = [
      "${mirror1}/SNAP/amazon0312.tar.gz"
@@ -15,7 +15,7 @@
  };

  matrix-02 = fetchzip {
-    sha256 = "sha256-0rSxaN4lQcdaCLsvlgicG70FXUxXeERPiEmQ4MzbRdE=";
+    hash = "sha256-0rSxaN4lQcdaCLsvlgicG70FXUxXeERPiEmQ4MzbRdE=";

    urls = [
      "${mirror1}/Muite/Chebyshev4.tar.gz"
@@ -24,7 +24,7 @@
  };

  matrix-03 = fetchzip {
-    sha256 = "sha256-hDzDWDUnHEyFedX/tMNq83ZH8uWyM4xtZYUUAD3rizo=";
+    hash = "sha256-hDzDWDUnHEyFedX/tMNq83ZH8uWyM4xtZYUUAD3rizo=";

    urls = [
      "${mirror1}/FEMLAB/sme3Dc.tar.gz"
@@ -33,7 +33,7 @@
  };

  matrix-04 = fetchzip {
-    sha256 = "sha256-GmN2yOt/MoX01rKe05aTyB3ypUP4YbQGOITZ0BqPmC0=";
+    hash = "sha256-GmN2yOt/MoX01rKe05aTyB3ypUP4YbQGOITZ0BqPmC0=";

    urls = [
      "${mirror1}/Williams/webbase-1M.tar.gz"
@@ -42,7 +42,7 @@
  };

  matrix-05 = fetchzip {
-    sha256 = "sha256-gQNjfVyWzNM9RwImJGhkhahRmZz74LzDs1oijL7mI7k=";
+    hash = "sha256-gQNjfVyWzNM9RwImJGhkhahRmZz74LzDs1oijL7mI7k=";

    urls = [
      "${mirror1}/Williams/mac_econ_fwd500.tar.gz"
@@ -51,7 +51,7 @@
  };

  matrix-06 = fetchzip {
-    sha256 = "sha256-87cdZjntNcTuz5BtO59irhcuRbPllWSbhCEX3Td02qc=";
+    hash = "sha256-87cdZjntNcTuz5BtO59irhcuRbPllWSbhCEX3Td02qc=";

    urls = [
      "${mirror1}/Williams/mc2depi.tar.gz"
@@ -60,7 +60,7 @@
  };

  matrix-07 = fetchzip {
-    sha256 = "sha256-WRamuJX3D8Tm+k0q67RjUDG3DeNAxhKiaPkk5afY5eU=";
+    hash = "sha256-WRamuJX3D8Tm+k0q67RjUDG3DeNAxhKiaPkk5afY5eU=";

    urls = [
      "${mirror1}/Bova/rma10.tar.gz"
@@ -69,7 +69,7 @@
  };

  matrix-08 = fetchzip {
-    sha256 = "sha256-5dhkm293Mc3lzakKxHy5W5XIn4Rw+gihVh7gyrjEHXo=";
+    hash = "sha256-5dhkm293Mc3lzakKxHy5W5XIn4Rw+gihVh7gyrjEHXo=";

    urls = [
      "${mirror1}/JGD_BIBD/bibd_22_8.tar.gz"
@@ -78,7 +78,7 @@
  };

  matrix-09 = fetchzip {
-    sha256 = "sha256-czjLWCjXAjZCk5TGYHaEkwSAzQu3TQ3QyB6eNKR4G88=";
+    hash = "sha256-czjLWCjXAjZCk5TGYHaEkwSAzQu3TQ3QyB6eNKR4G88=";

    urls = [
      "${mirror1}/Hamm/scircuit.tar.gz"
@@ -87,7 +87,7 @@
  };

  matrix-10 = fetchzip {
-    sha256 = "sha256-bYuLnJViAIcIejAkh69/bsNAVIDU4wfTLtD+nmHd6FM=";
+    hash = "sha256-bYuLnJViAIcIejAkh69/bsNAVIDU4wfTLtD+nmHd6FM=";

    urls = [
      "${mirror1}/Sandia/ASIC_320k.tar.gz"
@@ -96,7 +96,7 @@
  };

  matrix-11 = fetchzip {
-    sha256 = "sha256-aDwn8P1khYjo2Agbq5m9ZBInJUxf/knJNvyptt0fak0=";
+    hash = "sha256-aDwn8P1khYjo2Agbq5m9ZBInJUxf/knJNvyptt0fak0=";

    urls = [
      "${mirror1}/GHS_psdef/bmwcra_1.tar.gz"
@@ -105,7 +105,7 @@
  };

  matrix-12 = fetchzip {
-    sha256 = "sha256-8OJqA/byhlAZd869TPUzZFdsOiwOoRGfKyhM+RMjXoY=";
+    hash = "sha256-8OJqA/byhlAZd869TPUzZFdsOiwOoRGfKyhM+RMjXoY=";

    urls = [
      "${mirror1}/HB/nos1.tar.gz"
@@ -114,7 +114,7 @@
  };

  matrix-13 = fetchzip {
-    sha256 = "sha256-FS0rKqmg+uHwsM/yGfQLBdd7LH/rUrdutkNGBD/Mh1I=";
+    hash = "sha256-FS0rKqmg+uHwsM/yGfQLBdd7LH/rUrdutkNGBD/Mh1I=";

    urls = [
      "${mirror1}/HB/nos2.tar.gz"
@@ -123,7 +123,7 @@
  };

  matrix-14 = fetchzip {
-    sha256 = "sha256-DANnlrNJikrI7Pst9vRedtbuxepyHmCIu2yhltc4Qcs=";
+    hash = "sha256-DANnlrNJikrI7Pst9vRedtbuxepyHmCIu2yhltc4Qcs=";

    urls = [
      "${mirror1}/HB/nos3.tar.gz"
@@ -132,7 +132,7 @@
  };

  matrix-15 = fetchzip {
-    sha256 = "sha256-21mUgqjWGUfYgiWwSrKh9vH8Vdt3xzcefmqYNYRpxiY=";
+    hash = "sha256-21mUgqjWGUfYgiWwSrKh9vH8Vdt3xzcefmqYNYRpxiY=";

    urls = [
      "${mirror1}/HB/nos4.tar.gz"
@@ -141,7 +141,7 @@
  };

  matrix-16 = fetchzip {
-    sha256 = "sha256-FOuXvGqBBFNkVS6cexmkluret54hCfCOdK+DOZllE4c=";
+    hash = "sha256-FOuXvGqBBFNkVS6cexmkluret54hCfCOdK+DOZllE4c=";

    urls = [
      "${mirror1}/HB/nos5.tar.gz"
@@ -150,7 +150,7 @@
  };

  matrix-17 = fetchzip {
-    sha256 = "sha256-+7NI1rA/qQxYPpjXKHvAaCZ+LSaAJ4xuJvMRMBEUYxg=";
+    hash = "sha256-+7NI1rA/qQxYPpjXKHvAaCZ+LSaAJ4xuJvMRMBEUYxg=";

    urls = [
      "${mirror1}/HB/nos6.tar.gz"
@@ -159,7 +159,7 @@
  };

  matrix-18 = fetchzip {
-    sha256 = "sha256-q3NxJjbwGGcFiQ9nhWfUKgZmdVwCfPmgQoqy0AqOsNc=";
+    hash = "sha256-q3NxJjbwGGcFiQ9nhWfUKgZmdVwCfPmgQoqy0AqOsNc=";

    urls = [
      "${mirror1}/HB/nos7.tar.gz"
@@ -168,7 +168,7 @@
  };

  matrix-19 = fetchzip {
-    sha256 = "sha256-0GAN6qmVfD+tprIigzuUUUwm5KVhkN9X65wMEvFltDY=";
+    hash = "sha256-0GAN6qmVfD+tprIigzuUUUwm5KVhkN9X65wMEvFltDY=";

    urls = [
      "${mirror1}/DNVS/shipsec1.tar.gz"
@@ -177,7 +177,7 @@
  };

  matrix-20 = fetchzip {
-    sha256 = "sha256-f28Du/Urxsiq5NkRmRO10Zz9vvGRjEchquzHzbZpZ7U=";
+    hash = "sha256-f28Du/Urxsiq5NkRmRO10Zz9vvGRjEchquzHzbZpZ7U=";

    urls = [
      "${mirror1}/Cote/mplate.tar.gz"
@@ -186,7 +186,7 @@
  };

  matrix-21 = fetchzip {
-    sha256 = "sha256-O+Wy0NfCU1hVUOfNR1dJpvDHLBwwa301IRJDrQJnhak=";
+    hash = "sha256-O+Wy0NfCU1hVUOfNR1dJpvDHLBwwa301IRJDrQJnhak=";

    urls = [
      "${mirror1}/Bai/qc2534.tar.gz"
@@ -195,7 +195,7 @@
  };

  matrix-22 = fetchzip {
-    sha256 = "sha256-oxMnt8U5Cf1ILWcBdU6W9jdSMMm+U6bIVl8nm3n3+OA=";
+    hash = "sha256-oxMnt8U5Cf1ILWcBdU6W9jdSMMm+U6bIVl8nm3n3+OA=";

    urls = [
      "${mirror1}/Chevron/Chevron2.tar.gz"
@@ -204,7 +204,7 @@
  };

  matrix-23 = fetchzip {
-    sha256 = "sha256-MFD9BxFI/3IS7yatW121BAI04fbqrXpgYDT5UKjeKcU=";
+    hash = "sha256-MFD9BxFI/3IS7yatW121BAI04fbqrXpgYDT5UKjeKcU=";

    urls = [
      "${mirror1}/Chevron/Chevron3.tar.gz"
@@ -213,7 +213,7 @@
  };

  matrix-24 = fetchzip {
-    sha256 = "sha256-ikS8O51pe1nt3BNyhvfvqCbVL0+bg/da9bqGqeBDkTg=";
+    hash = "sha256-ikS8O51pe1nt3BNyhvfvqCbVL0+bg/da9bqGqeBDkTg=";

    urls = [
      "${mirror1}/Chevron/Chevron4.tar.gz"
@@ -15,7 +15,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocthrust";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -31,7 +31,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocThrust";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-Gncy0wvN8M4JSmnjbxaED+M9rLo4A/7LCCq/6temiZU=";
+    hash = "sha256-BDxFeCR2FfZI+TGMi2jBDHORCoAnFwk8WGGKMckCkmU=";
  };

  nativeBuildInputs = [
@@ -19,7 +19,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "roctracer";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -19,7 +19,7 @@

 stdenv.mkDerivation (finalAttrs: {
  pname = "rocwmma";
-  version = "7.1.1";
+  version = "7.2.0";

  outputs = [
    "out"
@@ -38,7 +38,7 @@ stdenv.mkDerivation (finalAttrs: {
    owner = "ROCm";
    repo = "rocWMMA";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-6hvdq3J8UhldcyNqXxi9UEF+P0RgG8o5YdIZLzvpgXw=";
+    hash = "sha256-O+P+pobPnlKnmJQylFp/qUQLE1kJKbxjMWyGS6lc3Mo=";
  };

  patches = lib.optionals (buildTests || buildBenchmarks) [
@@ -29,13 +29,13 @@ stdenv.mkDerivation (finalAttrs: {
        "cpu"
    );

-  version = "7.1.1";
+  version = "7.2.0";

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "rpp";
    rev = "rocm-${finalAttrs.version}";
-    hash = "sha256-ZlJRV57ybPsczvoWwd5vr2n7EKje0vj2GJYsvVY+qas=";
+    hash = "sha256-mET9nZReO3BrNZkcNeB7hOwIQaVZ8OD7lPD3b1Jot2k=";
  };

  nativeBuildInputs = [
@@ -18,14 +18,14 @@

 buildPythonPackage rec {
  pname = "tensile";
-  version = "7.1.1";
+  version = "7.2.0";
  pyproject = true;

  src = fetchFromGitHub {
    owner = "ROCm";
    repo = "Tensile";
    rev = "rocm-${version}";
-    hash = "sha256-uYLlbtXrc8m8/RpMbP7WaS7VYGIUm2LHYzv00DZAzvE=";
+    hash = "sha256-pTuEeu+0E+RlpdeFho0mJmx/V/tou/ZRLYX9KoWEPpA=";
  };

  # TODO: It should be possible to run asm caps test ONCE for all supported arches
@@ -58,8 +58,9 @@ buildPythonPackage rec {
    ./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";
+      hash = "sha256-ZHXNYSeLkhhNDaIfdqQm68Pxmh1shUL7mAVmh8/I6Xk=";
+      url = "https://github.com/GZGavinZhao/rocm-libraries/commit/1f7135dfc0cdb175c8f0e5eb71b2d24699942873.patch";
+      relative = "shared/tensile";
    })
  ];