resolve conflicts.

Author: sayakpaul

build2cmake/src/templates/utils.cmake

@@ -217,15 +217,15 @@ function(extract_unique_cuda_archs_ascending OUT_ARCHES CUDA_ARCH_FLAGS)
 endfunction()
 
 #
-# For a specific file set the `-gencode` flag in compile options conditionally
-# for the CUDA language.
+# For a specific file set the `-gencode` flag in compile options conditionally 
+# for the CUDA language. 
 #
 # Example:
 #   set_gencode_flag_for_srcs(
 #     SRCS "foo.cu"
 #     ARCH "compute_75"
 #     CODE "sm_75")
-#   adds: "-gencode arch=compute_75,code=sm_75" to the compile options for
+#   adds: "-gencode arch=compute_75,code=sm_75" to the compile options for 
 #    `foo.cu` (only for the CUDA language).
 #
 macro(set_gencode_flag_for_srcs)
@@ -344,13 +344,13 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   list(REMOVE_DUPLICATES _PTX_ARCHS)
   list(REMOVE_DUPLICATES _SRC_CUDA_ARCHS)
 
-  # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
-  # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
+  # If x.0a or x.0f is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
+  # remove x.0a or x.0f from SRC_CUDA_ARCHS and add x.0a or x.0f to _CUDA_ARCHS
   set(_CUDA_ARCHS)
   foreach(_arch ${_SRC_CUDA_ARCHS})
-    if(_arch MATCHES "\\a$")
+    if(_arch MATCHES "[af]$")
       list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
-      string(REPLACE "a" "" _base "${_arch}")
+      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
       if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
         list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
         list(APPEND _CUDA_ARCHS "${_arch}")

docs/source/cli.md

@@ -37,3 +37,25 @@ your kernel builds to the Hub. To know the supported arguments run: `kernels upl
 Use `kernels create-and-upload-card <kernel_source_dir> --card-path README.md` to generate a basic homepage
 for the kernel. Find an example [here](https://hf.co/kernels-community/kernel-card-template). You can
 optionally push it to the Hub by specifying a `--repo-id`.
+
+### kernels skills add
+
+Use `kernels skills add` to install the skills for AI coding assistants like Claude, Codex, and OpenCode. For now, only the `cuda-kernels` skill is supported. Skill files are downloaded from the `huggingface/kernels` directory in this [repository](https://github.com/huggingface/kernels/tree/main/skills).
+
+Skills instruct agents how to deal with hardware-specific optimizations, integrate with libraries like diffusers and transformers, and benchmark kernel performance in consistent ways.
+
+Examples:
+
+```bash
+# install for Claude in the current project
+kernels skills add --claude
+
+# install globally for Codex
+kernels skills add --codex --global
+
+# install for multiple assistants
+kernels skills add --claude --codex --opencode
+
+# install to a custom destination and overwrite if already present
+kernels skills add --dest ~/my-skills --force
+```

flake.nix

@@ -1,6 +1,13 @@
 {
   description = "Kernels";
 
+  nixConfig = {
+    extra-substituters = [ "https://huggingface.cachix.org" ];
+    extra-trusted-public-keys = [
+      "huggingface.cachix.org-1:ynTPbLS0W8ofXd9fDjk1KvoFky9K2jhxe6r4nXAkc/o="
+    ];
+  };
+
   inputs = {
     flake-utils.url = "github:numtide/flake-utils";
     nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable-small";
@@ -121,7 +128,8 @@
 
         buildSets = defaultBuildSetsPerSystem.${system};
         buildSetsByBackend = (partitionBuildSetsBySystemBackend defaultBuildSets).${system};
-        buildSet = builtins.head buildSetsByBackend.cuda;
+        defaultBackend = if system == "aarch64-darwin" then "metal" else "cuda";
+        buildSet = builtins.head buildSetsByBackend.${defaultBackend};
 
         # Dev shells per framework.
         devShellByBackend = lib.mapAttrs (

kernels/pyproject.toml

@@ -36,6 +36,7 @@ dev = [
 [project.optional-dependencies]
 abi-check = ["kernel-abi-check>=0.6.2,<0.7.0"]
 benchmark = [
+  "matplotlib>=3.7.0",
   "numpy>=2.0.2",
   "tabulate>=0.9.0",
   "torch",

kernels/src/kernels/cli/__init__.py

@@ -13,6 +13,7 @@
     KNOWN_BACKENDS,
 )
 from kernels.cli.init import run_init, parse_kernel_name
+from kernels.cli.skills import add_skill
 from kernels.cli.versions import print_kernel_versions
 from kernels.cli.doc import generate_readme_for_kernel
 from kernels.kernel_card_utils import (
@@ -98,6 +99,53 @@ def main():
     )
     upload_parser.set_defaults(func=upload_kernels)
 
+    skills_parser = subparsers.add_parser(
+        "skills",
+        help="Install kernels specific skills for agents like Claude, Codex, and OpenCode",
+    )
+    skills_subparsers = skills_parser.add_subparsers(required=True)
+    skills_add_parser = skills_subparsers.add_parser(
+        "add",
+        help="Install the cuda-kernels skill for an AI assistant",
+    )
+    skills_add_parser.add_argument(
+        "--claude",
+        action="store_true",
+        help="Install for Claude.",
+    )
+    skills_add_parser.add_argument(
+        "--codex",
+        action="store_true",
+        help="Install for Codex.",
+    )
+    skills_add_parser.add_argument(
+        "--opencode",
+        action="store_true",
+        help="Install for OpenCode.",
+    )
+    skills_add_parser.add_argument(
+        "--global",
+        "-g",
+        dest="global_",
+        action="store_true",
+        help=(
+            "Install globally (user-level) instead of in the current project "
+            "directory."
+        ),
+    )
+    skills_add_parser.add_argument(
+        "--dest",
+        type=Path,
+        default=None,
+        help="Install into a custom destination (path to skills directory).",
+    )
+    skills_add_parser.add_argument(
+        "--force",
+        action="store_true",
+        help="Overwrite existing skills in the destination.",
+    )
+    skills_add_parser.set_defaults(func=add_skill)
+
     lock_parser = subparsers.add_parser("lock", help="Lock kernel revisions")
     lock_parser.add_argument(
         "project_dir",
@@ -156,6 +204,17 @@ def main():
     )
     benchmark_parser.add_argument("--iterations", type=int, default=100)
     benchmark_parser.add_argument("--warmup", type=int, default=10)
+    benchmark_parser.add_argument(
+        "--visual",
+        type=str,
+        default=None,
+        help="Save visual outputs using this base path (e.g., --visual bench creates bench_light.svg and bench_dark.svg variants)",
+    )
+    benchmark_parser.add_argument(
+        "--rasterized",
+        action="store_true",
+        help="Output PNG and GIF formats instead of SVG",
+    )
     benchmark_parser.set_defaults(func=run_benchmark)
 
     init_parser = subparsers.add_parser(
@@ -356,4 +415,6 @@ def run_benchmark(args):
         warmup=args.warmup,
         output=args.output,
         print_json=args.json,
+        visual=args.visual,
+        rasterized=args.rasterized,
     )

kernels/src/kernels/cli/benchmark.py

@@ -682,6 +682,8 @@ def run_benchmark(
     upload: bool = False,
     output: str | None = None,
     print_json: bool = False,
+    visual: str | None = None,
+    rasterized: bool = False,
 ) -> BenchmarkResult:
     if MISSING_DEPS:
         print(
@@ -794,6 +796,56 @@ def run_benchmark(
     if print_json:
         print(json.dumps(result.to_payload(), indent=2))
 
+    if visual:
+        from kernels.cli.benchmark_graphics import (
+            save_speedup_animation,
+            save_speedup_image,
+        )
+
+        media_dir = Path("media")
+        media_dir.mkdir(exist_ok=True)
+
+        for theme in ("light", "dark"):
+            dark = theme == "dark"
+            base_path = media_dir / f"{visual}_{theme}"
+
+            # Always produce SVGs
+            save_speedup_image(
+                timing_results,
+                f"{base_path}.svg",
+                machine_info.backend,
+                repo_id,
+                machine_info.pytorch_version,
+                dark=dark,
+            )
+            save_speedup_animation(
+                timing_results,
+                f"{base_path}_animation.svg",
+                machine_info.backend,
+                repo_id,
+                machine_info.pytorch_version,
+                dark=dark,
+            )
+
+            # Additionally produce PNGs and GIFs when --rasterized is used
+            if rasterized:
+                save_speedup_image(
+                    timing_results,
+                    f"{base_path}.png",
+                    machine_info.backend,
+                    repo_id,
+                    machine_info.pytorch_version,
+                    dark=dark,
+                )
+                save_speedup_animation(
+                    timing_results,
+                    f"{base_path}_animation.gif",
+                    machine_info.backend,
+                    repo_id,
+                    machine_info.pytorch_version,
+                    dark=dark,
+                )
+
     if upload:
         submit_benchmark(repo_id=repo_id, result=result)
         print("Benchmark submitted successfully!", file=sys.stderr)