Add FlyDSL fused RoPE + KV Cache backend

.githooks/pre-commit

@@ -55,17 +55,22 @@ for file in $files; do
 done
 
 # if clang-format exists, run it on C/C++ files
-if command -v clang-format >/dev/null; then
+CLANG_FORMAT=""
+if command -v clang-format-18 >/dev/null; then
+    CLANG_FORMAT="clang-format-18"
+fi
+
+if [[ -n "$CLANG_FORMAT" ]]; then
     for file in $files; do
     [[ -L $file ]] && continue
        if [[ -e $file ]] && echo $file | grep -Eq '\.c$|\.h$|\.cuh$|\.hpp$|\.cpp$|\.cl$|\.h\.in$|\.hpp\.in$|\.cpp\.in$'; then
-            echo "clang-format $file"
-            clang-format -i -style=file "$file"
+            echo "$CLANG_FORMAT $file"
+            $CLANG_FORMAT -i -style=file "$file"
             git add -u "$file"
         fi
     done
 else
-    echo "clang-format command not found, skipping file formatting."
+    echo "clang-format-18 not found, skipping file formatting."
 fi
 
 if command -v black >/dev/null; then

.github/scripts/build_aiter_triton.sh

@@ -8,12 +8,13 @@ dpkg -l | grep rocm || echo "No ROCm packages found."
 
 echo
 echo "==== Install dependencies and aiter ===="
+git config --global --add safe.directory /workspace
 pip install --upgrade pandas zmq einops numpy==1.26.2
 pip uninstall -y aiter || true
 pip install --upgrade "pybind11>=3.0.1"
 pip install --upgrade "ninja>=1.11.1"
 pip install tabulate
-python3 setup.py develop
+pip install -e .
 
 # Read BUILD_TRITON env var, default to 1. If 1, install Triton; if 0, skip installation.
 BUILD_TRITON=${BUILD_TRITON:-1}

.github/scripts/list_jobs.py

@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+import argparse
+import json
+from pathlib import Path
+
+import yaml
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Parse workflow files and output job matrix for GitHub Actions."
+    )
+    parser.add_argument(
+        "--workflow-dir",
+        default=".github/workflows",
+        help="Directory containing workflow yaml files.",
+    )
+    parser.add_argument(
+        "--workflows",
+        default="",
+        help="Comma-separated workflow file names. Empty means auto-discover.",
+    )
+    parser.add_argument(
+        "--exclude-workflows",
+        default="",
+        help="Comma-separated workflow file names to skip.",
+    )
+    parser.add_argument(
+        "--exclude-jobs",
+        default="",
+        help="Comma-separated job names to skip.",
+    )
+    parser.add_argument(
+        "--out-matrix", required=True, help="Output path for matrix JSON."
+    )
+    parser.add_argument(
+        "--out-workflow-map",
+        required=True,
+        help="Output path for workflow -> jobs mapping JSON.",
+    )
+    return parser.parse_args()
+
+
+def parse_csv(raw: str):
+    return [item.strip() for item in raw.split(",") if item.strip()]
+
+
+def discover_workflows(workflow_dir: Path):
+    return sorted(
+        [path.name for path in workflow_dir.glob("*.yml") if path.is_file()]
+        + [path.name for path in workflow_dir.glob("*.yaml") if path.is_file()]
+    )
+
+
+def main():
+    args = parse_args()
+    workflow_dir = Path(args.workflow_dir)
+    excluded_workflows = set(parse_csv(args.exclude_workflows))
+    excluded_jobs = set(parse_csv(args.exclude_jobs))
+
+    if args.workflows.strip():
+        workflow_files = parse_csv(args.workflows)
+    else:
+        workflow_files = discover_workflows(workflow_dir)
+
+    workflow_files = [wf for wf in workflow_files if wf not in excluded_workflows]
+
+    matrix = []
+    workflow_map = {}
+
+    for workflow_file in workflow_files:
+        workflow_path = workflow_dir / workflow_file
+        if not workflow_path.exists():
+            continue
+
+        with workflow_path.open("r", encoding="utf-8") as file_obj:
+            content = yaml.safe_load(file_obj) or {}
+
+        jobs = list((content.get("jobs") or {}).keys())
+        jobs = [job for job in jobs if job not in excluded_jobs]
+
+        workflow_map[workflow_file] = jobs
+        for job in jobs:
+            matrix.append({"workflow": workflow_file, "job_name": job})
+
+    Path(args.out_matrix).write_text(
+        json.dumps(matrix, ensure_ascii=False), encoding="utf-8"
+    )
+    Path(args.out_workflow_map).write_text(
+        json.dumps(workflow_map, ensure_ascii=False), encoding="utf-8"
+    )
+
+    print(f"Discovered workflows: {len(workflow_map)}")
+    print(f"Total jobs in matrix: {len(matrix)}")
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/op_tune.sh

@@ -20,7 +20,7 @@ declare -a tune_jobs=(
   "ck_batched_gemm_bf16:csrc/ck_batched_gemm_bf16:op_tests/test_batched_gemm_bf16.py:python3 csrc/ck_batched_gemm_bf16/batched_gemm_bf16_tune.py -i aiter/configs/bf16_untuned_batched_gemm.csv -o aiter/configs/bf16_tuned_batched_gemm.csv"
   "ck_gemm_a8w8:csrc/ck_gemm_a8w8:op_tests/test_gemm_a8w8.py:python3 csrc/ck_gemm_a8w8/gemm_a8w8_tune.py -i aiter/configs/a8w8_untuned_gemm.csv -o aiter/configs/a8w8_tuned_gemm.csv"
   "ck_gemm_a8w8_blockscale:csrc/ck_gemm_a8w8_blockscale:op_tests/test_gemm_a8w8_blockscale.py:python3 csrc/ck_gemm_a8w8_blockscale/gemm_a8w8_blockscale_tune.py -i aiter/configs/a8w8_blockscale_untuned_gemm.csv -o aiter/configs/a8w8_blockscale_tuned_gemm.csv"
-  "ck_gemm_a8w8_blockscale_bpreshuffle:csrc/ck_gemm_a8w8_blockscale_bpreshuffle:op_tests/test_gemm_a8w8_blockscale.py:python3 csrc/ck_gemm_a8w8_blockscale_bpreshuffle/gemm_a8w8_blockscale_bpreshuffle_tune.py -i aiter/configs/a8w8_blockscale_bpreshuffle_untuned_gemm.csv -o aiter/configs/a8w8_blockscale_bpreshuffle_tuned_gemm.csv"
+  "ck_gemm_a8w8_blockscale_bpreshuffle:csrc/ck_gemm_a8w8_blockscale_bpreshuffle:op_tests/test_gemm_a8w8_blockscale.py:python3 csrc/ck_gemm_a8w8_blockscale/gemm_a8w8_blockscale_tune.py --preshuffle -i aiter/configs/a8w8_blockscale_bpreshuffle_untuned_gemm.csv -o aiter/configs/a8w8_blockscale_bpreshuffle_tuned_gemm.csv"
   "ck_gemm_a8w8_bpreshuffle:csrc/ck_gemm_a8w8_bpreshuffle:op_tests/test_gemm_a8w8.py:python3 csrc/ck_gemm_a8w8_bpreshuffle/gemm_a8w8_bpreshuffle_tune.py -i aiter/configs/a8w8_bpreshuffle_untuned_gemm.csv -o aiter/configs/a8w8_bpreshuffle_tuned_gemm.csv"
   "ck_gemm_moe_2stages_codegen:csrc/ck_gemm_moe_2stages_codegen:op_tests/test_moe.py:python3 csrc/ck_gemm_moe_2stages_codegen/gemm_moe_tune.py -i aiter/configs/untuned_fmoe.csv -o aiter/configs/tuned_fmoe.csv"
   #"ck_gemm_a4w4_blockscale:csrc/ck_gemm_a4w4_blockscale:op_tests/test_gemm_a4w4_blockscale.py:python3 csrc/ck_gemm_a4w4_blockscale/gemm_a4w4_blockscale_tune.py -i aiter/configs/a4w4_blockscale_untuned_gemm.csv -o aiter/configs/a4w4_blockscale_tuned_gemm.csv"