[MUH] Bootstrap muh toolchain — extract/parse/gen_yaml/gen_patch + baseline.muh

Pipeline: 1. extract.py: Parses all 26 CCCL tuning_*.cuh → 26 YAML schemas in muh/schema/ 2. parse.py: .muh file parser with extends-inheritance + schema validation 3. gen_yaml.py: .muh → computility-run.yaml (verified: matches competition reference) 4. gen_patch.py: .muh → vllm kernel unified diff patches (6 algorithm mappings) 5. baseline.muh: Competition reference config, all tuning values pending BI-V100 benchmarks Schemas extracted: 26 algorithms, 8-19 params each, SM75/80/90/100 reference tunings Priority mapping: reduce→attention, topk→sampling, scan→paged_attention, transform→activations, batch_memcpy→KV_cache, for→RoPE Tested: extract→parse→validate→gen_yaml→gen_patch full pipeline passes
2026-07-30 10:39:06 +00:00
parent 70e80c5810
commit 9b21a13119
33 changed files with 4702 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -0,0 +1 @@
+muh/__pycache__/
--- a/baseline.muh
+++ b/baseline.muh
@@ -0,0 +1,122 @@
+# baseline.muh — Competition reference configuration
+# Corresponds to: dev.modelhub.org.cn EngineX-Iluvatar/enginex-vllm-bi100-qwen36
+#
+# This is the starting point. All tuning values are pending BI-V100 benchmarks.
+# Child .muh files use 'extends: baseline.muh' to override specific algorithms.
+
+# --- Hardware description ---
+hardware:
+  name: Iluvatar-BI-V100-50c-200G
+  gpu_count: 4
+  # These need to be confirmed on actual hardware:
+  warp_size: 32
+  max_threads_per_block: 1024
+  max_shared_memory_per_block: 49152
+  max_registers_per_thread: 255
+  l2_cache_size_bytes: 6291456
+  memory_bandwidth_gbps: 900
+  compute_capability: iluvatar_bi100
+
+# --- vllm launch configuration ---
+# Maps directly to computility-run.yaml command
+vllm:
+  model_path: /model
+  served_model_name: llm
+  max_model_len: 100000
+  gpu_memory_utilization: 0.9
+  tensor_parallel: 4
+  max_num_seqs: 1
+  max_num_batched_tokens: 8192
+  max_seq_len_to_capture: 32768
+  trust_remote_code: true
+  disable_log_requests: true
+  disable_frontend_multiprocessing: true
+  enable_chunked_prefill: true
+  enable_auto_tool_choice: true
+  tool_call_parser: qwen3_coder
+  reasoning_parser: qwen3
+  enable_prefix_caching: true
+
+# --- Concurrency ---
+concurrency: 1
+
+# --- Environment ---
+env:
+  VLLM_ENGINE_ITERATION_TIMEOUT_S: 3600
+
+# --- Tuning overrides (per CCCL algorithm) ---
+# Each key corresponds to a tuning_*.cuh schema in muh/schema/
+# Values are TBD until we run benchmarks on BI-V100
+#
+# Priority order (by competition score impact):
+#   1. reduce   — attention reduction (Output TPS × 16.796)
+#   2. topk     — sampling top-k/top-p (Output TPS × 16.796)
+#   3. scan     — prefix scan in paged attention
+#   4. transform — activation kernels (SiLU, GELU)
+#   5. batch_memcpy — KV cache management (Cache TPS × 0.56)
+#   6. for      — RoPE position encoding
+
+tuning:
+  reduce:
+    _priority: P0
+    _vllm_impact: attention_reduction
+    _score_weight: Output TPS × 16.796
+    # CCCL SM90 reference: threads=128, items=24, vec_size=4
+    # CCCL SM100 reference: threads varies by accum_size
+    threads_per_block: null
+    items_per_thread: null
+    vec_size: null
+
+  topk:
+    _priority: P0
+    _vllm_impact: sampling_decode
+    _score_weight: Output TPS × 16.796
+    # CCCL reference: threads=512, items=4, bits_per_pass=11
+    threads_per_block: null
+    items_per_thread: null
+    bits_per_pass: null
+
+  scan:
+    _priority: P0
+    _vllm_impact: paged_attention_prefix_scan
+    _score_weight: Input TPS × 2.799
+    # CCCL SM90 lookback: threads=128, items=24, delay=fixed(688, 1140) for float32
+    # CCCL SM100 lookback: threads=384, items=22, delay=exponential_backon(1904, 830)
+    # CCCL SM100 lookahead: warps=4, items=80-1, lookahead_items=3
+    threads_per_block: null
+    items_per_thread: null
+    load_algorithm: null
+    store_algorithm: null
+    scan_algorithm: null
+
+  transform:
+    _priority: P1
+    _vllm_impact: activation_elementwise
+    _score_weight: Output TPS × 16.796
+    threads_per_block: null
+    items_per_thread: null
+
+  batch_memcpy:
+    _priority: P1
+    _vllm_impact: kv_cache_copy
+    _score_weight: Cache TPS × 0.56
+    threads_per_block: null
+
+  for:
+    _priority: P2
+    _vllm_impact: rope_position_encoding
+    threads_per_block: null
+    items_per_thread: null
+
+  radix_sort:
+    _priority: P2
+    _vllm_impact: beam_search_token_sort
+    threads_per_block: null
+    items_per_thread: null
+    radix_bits: null
+
+  merge:
+    _priority: P2
+    _vllm_impact: sequence_merging
+    threads_per_block: null
+    items_per_thread: null
--- a/muh/extract.py
+++ b/muh/extract.py
@@ -0,0 +1,285 @@
+#!/usr/bin/env python3
+"""muh/extract.py — Extract CCCL tuning parameter spaces into muh/schema/*.yaml
+
+Reads cccl_upstream/cub/cub/device/dispatch/tuning/tuning_*.cuh,
+parses policy struct fields and SM-specific tuning values,
+outputs one YAML file per algorithm under muh/schema/.
+
+Usage:
+    python3 muh/extract.py [--cccl-root cccl_upstream] [--out-dir muh/schema]
+"""
+
+import re
+import os
+import sys
+import glob
+import argparse
+from pathlib import Path
+from collections import OrderedDict
+
+# --- Enum value sets (from CCCL headers) ---
+
+BLOCK_LOAD_ALGORITHMS = [
+    "BLOCK_LOAD_DIRECT",
+    "BLOCK_LOAD_VECTORIZE",
+    "BLOCK_LOAD_TRANSPOSE",
+    "BLOCK_LOAD_WARP_TRANSPOSE",
+    "BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED",
+    "BLOCK_LOAD_STRIPED",
+]
+
+BLOCK_STORE_ALGORITHMS = [
+    "BLOCK_STORE_DIRECT",
+    "BLOCK_STORE_WARP_TRANSPOSE",
+    "BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED",
+    "BLOCK_STORE_STRIPED",
+]
+
+BLOCK_REDUCE_ALGORITHMS = [
+    "BLOCK_REDUCE_RAKING",
+    "BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY",
+    "BLOCK_REDUCE_WARP_REDUCTIONS",
+]
+
+BLOCK_SCAN_ALGORITHMS = [
+    "BLOCK_SCAN_RAKING",
+    "BLOCK_SCAN_RAKING_MEMOIZE",
+    "BLOCK_SCAN_WARP_SCANS",
+]
+
+CACHE_LOAD_MODIFIERS = [
+    "LOAD_DEFAULT",
+    "LOAD_CA",
+    "LOAD_CG",
+    "LOAD_CS",
+    "LOAD_CV",
+    "LOAD_LDG",
+]
+
+LOOKBACK_DELAY_ALGORITHMS = [
+    "no_delay",
+    "fixed_delay",
+    "exponential_backoff",
+    "exponential_backoff_jitter",
+    "exponential_backoff_jitter_window",
+    "exponential_backon_jitter_window",
+    "exponential_backon_jitter",
+    "exponential_backon",
+]
+
+# --- Field type → range/enum mapping ---
+
+FIELD_TYPES = {
+    "threads_per_block": {"type": "int", "range": [32, 1024], "step": 32},
+    "items_per_thread": {"type": "int", "range": [1, 32], "step": 1},
+    "vec_size": {"type": "int", "range": [1, 8], "step": 1},
+    "bits_per_pass": {"type": "int", "range": [4, 11], "step": 1},
+    "radix_bits": {"type": "int", "range": [4, 8], "step": 1},
+    "load_algorithm": {"type": "enum", "values": BLOCK_LOAD_ALGORITHMS},
+    "store_algorithm": {"type": "enum", "values": BLOCK_STORE_ALGORITHMS},
+    "reduce_algorithm": {"type": "enum", "values": BLOCK_REDUCE_ALGORITHMS},
+    "scan_algorithm": {"type": "enum", "values": BLOCK_SCAN_ALGORITHMS},
+    "load_modifier": {"type": "enum", "values": CACHE_LOAD_MODIFIERS},
+    "lookback_delay.kind": {"type": "enum", "values": LOOKBACK_DELAY_ALGORITHMS},
+    "lookback_delay.delay": {"type": "int", "range": [0, 2000], "step": 50},
+    "lookback_delay.l2_write_latency": {"type": "int", "range": [0, 2000], "step": 50},
+    "reduce_and_scan_warps": {"type": "int", "range": [1, 8], "step": 1},
+    "lookahead_items_per_thread": {"type": "int", "range": [1, 16], "step": 1},
+}
+
+
+def extract_policy_fields(content, filename):
+    """Extract policy struct field names from a tuning file."""
+    fields = []
+    # Match lines like: int threads_per_block; or BlockLoadAlgorithm load_algorithm;
+    pattern = re.compile(
+        r'^\s+(?:int|BlockLoadAlgorithm|BlockStoreAlgorithm|BlockReduceAlgorithm|'
+        r'BlockScanAlgorithm|CacheLoadModifier|LookbackDelayPolicy)\s+'
+        r'(\w+)\s*[;=]',
+        re.MULTILINE
+    )
+    for m in pattern.finditer(content):
+        field = m.group(1)
+        if field not in fields:
+            fields.append(field)
+    return fields
+
+
+def extract_sm_tunings(content):
+    """Extract SM-specific tuning values from static constexpr definitions."""
+    tunings = {}
+    # Match patterns like: sm80_tuning, sm90_tuning, sm100_tuning
+    sm_pattern = re.compile(r'struct\s+sm(\d+)_tuning')
+    for m in sm_pattern.finditer(content):
+        sm = int(m.group(1))
+        if sm not in tunings:
+            tunings[sm] = []
+
+    # Extract actual parameter values from constexpr definitions
+    # Pattern: static constexpr int threads = 512;
+    blocks = re.split(r'(?=struct\s+sm\d+_tuning)', content)
+    for block in blocks:
+        sm_m = re.match(r'struct\s+sm(\d+)_tuning', block)
+        if not sm_m:
+            continue
+        sm = int(sm_m.group(1))
+        vals = {}
+        for line in block.split('\n'):
+            # int values
+            m = re.search(r'static\s+constexpr\s+int\s+(\w+)\s*=\s*(\d+)', line)
+            if m:
+                vals[m.group(1)] = int(m.group(2))
+            # enum values
+            m = re.search(r'static\s+constexpr\s+(?:BlockLoadAlgorithm|BlockStoreAlgorithm|CacheLoadModifier)\s+(\w+)\s*=\s*(\w+)', line)
+            if m:
+                vals[m.group(1)] = m.group(2)
+        if vals:
+            tunings.setdefault(sm, []).append(vals)
+
+    return tunings
+
+
+def extract_inline_tunings(content):
+    """Extract inline tuning values from make_mem_scaled_lookback_scan_policy calls and similar."""
+    inline = []
+    # Pattern: threads_per_block, items_per_thread in constructor-style calls
+    pattern = re.compile(
+        r'(?:topk_policy|ReducePassPolicy|ScanLookbackPolicy)\s*\{'
+        r'\s*(\d+)\s*,\s*(\d+)',
+        re.MULTILINE
+    )
+    for m in pattern.finditer(content):
+        inline.append({
+            "threads_per_block": int(m.group(1)),
+            "items_per_thread": int(m.group(2)),
+        })
+    return inline
+
+
+def algo_name_from_filename(filename):
+    """tuning_topk.cuh → topk"""
+    base = os.path.basename(filename)
+    return base.replace("tuning_", "").replace(".cuh", "")
+
+
+def build_schema(algo, fields, sm_tunings, inline_tunings):
+    """Build YAML-serializable schema dict for one algorithm."""
+    schema = OrderedDict()
+    schema["algorithm"] = algo
+    schema["source"] = f"cub/cub/device/dispatch/tuning/tuning_{algo}.cuh"
+
+    # Parameter space
+    params = OrderedDict()
+    for field in fields:
+        if field in FIELD_TYPES:
+            params[field] = dict(FIELD_TYPES[field])
+        elif field == "lookback_delay":
+            # Expand to sub-fields
+            for sub in ["lookback_delay.kind", "lookback_delay.delay", "lookback_delay.l2_write_latency"]:
+                params[sub] = dict(FIELD_TYPES[sub])
+        else:
+            params[field] = {"type": "int", "range": [1, 1024], "step": 1, "note": "unknown_range"}
+    schema["parameters"] = dict(params)
+
+    # Known SM tunings (for reference when tuning BI-V100)
+    if sm_tunings:
+        ref = OrderedDict()
+        for sm, vals_list in sorted(sm_tunings.items()):
+            ref[f"sm{sm}"] = vals_list
+        schema["reference_tunings"] = dict(ref)
+
+    # BI-V100 placeholder
+    schema["bi_v100"] = {
+        "status": "pending_benchmark",
+        "note": "Run muh benchmark on Iluvatar BI-V100 to fill these values",
+        "threads_per_block": "TBD",
+        "items_per_thread": "TBD",
+    }
+
+    return schema
+
+
+def yaml_dump(data, indent=0):
+    """Simple YAML serializer (no dependency on pyyaml)."""
+    lines = []
+    prefix = "  " * indent
+    if isinstance(data, dict):
+        for k, v in data.items():
+            if isinstance(v, (dict, list)):
+                lines.append(f"{prefix}{k}:")
+                lines.append(yaml_dump(v, indent + 1))
+            else:
+                lines.append(f"{prefix}{k}: {v}")
+    elif isinstance(data, list):
+        for item in data:
+            if isinstance(item, dict):
+                lines.append(f"{prefix}-")
+                lines.append(yaml_dump(item, indent + 1))
+            else:
+                lines.append(f"{prefix}- {item}")
+    else:
+        lines.append(f"{prefix}{data}")
+    return "\n".join(lines)
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Extract CCCL tuning params to muh schema")
+    parser.add_argument("--cccl-root", default="cccl_upstream",
+                        help="Path to CCCL root (default: cccl_upstream)")
+    parser.add_argument("--out-dir", default="muh/schema",
+                        help="Output directory for YAML schemas (default: muh/schema)")
+    args = parser.parse_args()
+
+    tuning_dir = os.path.join(args.cccl_root, "cub", "cub", "device", "dispatch", "tuning")
+    out_dir = args.out_dir
+    os.makedirs(out_dir, exist_ok=True)
+
+    tuning_files = sorted(glob.glob(os.path.join(tuning_dir, "tuning_*.cuh")))
+    if not tuning_files:
+        print(f"ERROR: No tuning_*.cuh files found in {tuning_dir}", file=sys.stderr)
+        sys.exit(1)
+
+    print(f"Found {len(tuning_files)} tuning files in {tuning_dir}")
+
+    all_algos = []
+    for filepath in tuning_files:
+        if os.path.basename(filepath) == "common.cuh":
+            continue
+
+        with open(filepath, "r") as f:
+            content = f.read()
+
+        algo = algo_name_from_filename(filepath)
+        fields = extract_policy_fields(content, filepath)
+        sm_tunings = extract_sm_tunings(content)
+        inline_tunings = extract_inline_tunings(content)
+
+        schema = build_schema(algo, fields, sm_tunings, inline_tunings)
+
+        out_path = os.path.join(out_dir, f"{algo}.yaml")
+        with open(out_path, "w") as f:
+            f.write(f"# muh schema for {algo}\n")
+            f.write(f"# Auto-extracted from {schema['source']}\n")
+            f.write(f"# Generated by muh/extract.py\n\n")
+            f.write(yaml_dump(dict(schema)))
+            f.write("\n")
+
+        all_algos.append(algo)
+        print(f"  {algo}: {len(fields)} params, {len(sm_tunings)} SM tunings → {out_path}")
+
+    # Write index
+    index_path = os.path.join(out_dir, "_index.yaml")
+    with open(index_path, "w") as f:
+        f.write("# muh schema index — all extracted CCCL tuning algorithms\n\n")
+        f.write("algorithms:\n")
+        for algo in all_algos:
+            f.write(f"  - {algo}\n")
+        f.write(f"\ntotal: {len(all_algos)}\n")
+        f.write(f"source: cccl_upstream/cub/cub/device/dispatch/tuning/\n")
+
+    print(f"\nDone: {len(all_algos)} schemas → {out_dir}/")
+    print(f"Index: {index_path}")
+
+
+if __name__ == "__main__":
+    main()
--- a/muh/gen_patch.py
+++ b/muh/gen_patch.py
@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+"""muh/gen_patch.py — Generate vllm kernel patches from .muh tuning configuration
+
+Given a .muh file with tuning overrides for BI-V100, generates unified diff
+patches that can be applied to the vllm source tree to inject optimized
+kernel parameters.
+
+The key insight: vllm's CUDA kernels (attention, sampling, layernorm) have
+hardcoded launch configs. This script generates patches that replace those
+hardcodes with values tuned for Iluvatar BI-V100 via CCCL benchmark data.
+
+Usage:
+    python3 muh/gen_patch.py baseline.muh [-o patches/] [--vllm-root /path/to/vllm]
+"""
+
+import os
+import sys
+import argparse
+from datetime import datetime
+
+sys.path.insert(0, os.path.dirname(__file__))
+from parse import load_muh
+
+
+# --- Kernel location mapping ---
+# Maps CCCL algorithm names to vllm source files and the specific
+# constants/defines that control kernel launch parameters.
+
+VLLM_KERNEL_MAP = {
+    "reduce": {
+        "description": "Attention score reduction in multi-head attention",
+        "files": [
+            "csrc/attention/attention_kernels.cu",
+            "csrc/attention/paged_attention_v2.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "NUM_THREADS",
+                "default": 128,
+                "locations": ["#define NUM_THREADS 128"],
+            },
+            "items_per_thread": {
+                "pattern": "NUM_ITEMS_PER_THREAD",
+                "default": 8,
+            },
+            "vec_size": {
+                "pattern": "VEC_SIZE",
+                "default": 4,
+            },
+        },
+    },
+    "topk": {
+        "description": "Top-k / top-p sampling in decode stage",
+        "files": [
+            "csrc/sampling/sampling_kernels.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "SAMPLING_BLOCK_SIZE",
+                "default": 256,
+            },
+            "bits_per_pass": {
+                "pattern": "RADIX_BITS",
+                "default": 8,
+            },
+        },
+    },
+    "scan": {
+        "description": "Prefix scan in paged attention block table lookup",
+        "files": [
+            "csrc/attention/paged_attention_v1.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "SCAN_BLOCK_SIZE",
+                "default": 128,
+            },
+        },
+    },
+    "transform": {
+        "description": "Elementwise activation kernels (SiLU, GELU, RMSNorm)",
+        "files": [
+            "csrc/activation_kernels.cu",
+            "csrc/layernorm_kernels.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "ACTIVATION_BLOCK_SIZE",
+                "default": 512,
+            },
+        },
+    },
+    "batch_memcpy": {
+        "description": "KV cache block copy between GPU memory regions",
+        "files": [
+            "csrc/cache_kernels.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "COPY_BLOCK_SIZE",
+                "default": 256,
+            },
+        },
+    },
+    "for": {
+        "description": "Elementwise for-each kernels (position embeddings, rope)",
+        "files": [
+            "csrc/pos_encoding_kernels.cu",
+        ],
+        "params": {
+            "threads_per_block": {
+                "pattern": "ROPE_BLOCK_SIZE",
+                "default": 512,
+            },
+        },
+    },
+}
+
+
+def generate_define_patch(algo, param_name, old_value, new_value, define_name, filepath):
+    """Generate a unified diff snippet for a #define change."""
+    lines = []
+    lines.append(f"--- a/{filepath}")
+    lines.append(f"+++ b/{filepath}")
+    lines.append(f"@@ -1,1 +1,1 @@")
+    lines.append(f"-#define {define_name} {old_value}")
+    lines.append(f"+#define {define_name} {new_value}  // muh: tuned for BI-V100 ({algo}.{param_name})")
+    return "\n".join(lines)
+
+
+def generate_patches(config, vllm_root=None):
+    """Generate all patches from tuning config."""
+    tuning = config.get("tuning", {})
+    patches = []
+    summary = []
+
+    for algo, algo_params in tuning.items():
+        if not isinstance(algo_params, dict):
+            continue
+
+        mapping = VLLM_KERNEL_MAP.get(algo)
+        if mapping is None:
+            summary.append(f"SKIP {algo}: no vllm kernel mapping defined")
+            continue
+
+        for param_name, new_value in algo_params.items():
+            if param_name.startswith("_"):
+                continue
+            if new_value is None:
+                continue
+
+            param_spec = mapping.get("params", {}).get(param_name)
+            if param_spec is None:
+                continue
+
+            old_value = param_spec.get("default")
+            define_name = param_spec.get("pattern", param_name.upper())
+
+            for filepath in mapping.get("files", []):
+                patch = generate_define_patch(
+                    algo, param_name, old_value, new_value, define_name, filepath
+                )
+                patches.append({
+                    "algo": algo,
+                    "param": param_name,
+                    "file": filepath,
+                    "old": old_value,
+                    "new": new_value,
+                    "diff": patch,
+                })
+                summary.append(
+                    f"PATCH {filepath}: {define_name} {old_value} → {new_value} "
+                    f"(from {algo}.{param_name})"
+                )
+
+    return patches, summary
+
+
+def write_patches(patches, out_dir):
+    """Write patches to individual .patch files."""
+    os.makedirs(out_dir, exist_ok=True)
+
+    # Combined patch
+    combined_path = os.path.join(out_dir, "muh_bi100_tuning.patch")
+    with open(combined_path, 'w') as f:
+        f.write(f"# muh kernel tuning patch for Iluvatar BI-V100\n")
+        f.write(f"# Generated: {datetime.now().isoformat()}\n")
+        f.write(f"# Algorithms patched: {len(set(p['algo'] for p in patches))}\n")
+        f.write(f"# Total changes: {len(patches)}\n\n")
+        for p in patches:
+            f.write(p["diff"])
+            f.write("\n\n")
+
+    # Per-algorithm patches
+    by_algo = {}
+    for p in patches:
+        by_algo.setdefault(p["algo"], []).append(p)
+
+    for algo, algo_patches in by_algo.items():
+        algo_path = os.path.join(out_dir, f"{algo}.patch")
+        with open(algo_path, 'w') as f:
+            f.write(f"# muh tuning patch: {algo} for BI-V100\n\n")
+            for p in algo_patches:
+                f.write(p["diff"])
+                f.write("\n\n")
+
+    return combined_path
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate vllm kernel patches from .muh")
+    parser.add_argument("muh_file", help="Path to .muh file")
+    parser.add_argument("-o", "--output-dir", default="patches",
+                        help="Output directory for patches (default: patches)")
+    parser.add_argument("--vllm-root", default=None,
+                        help="Path to vllm source tree (for verification)")
+    parser.add_argument("--dry-run", action="store_true",
+                        help="Print patches to stdout instead of writing files")
+    args = parser.parse_args()
+
+    config = load_muh(args.muh_file)
+    patches, summary = generate_patches(config, args.vllm_root)
+
+    print(f"muh gen_patch: {len(patches)} patches from {args.muh_file}\n")
+    for s in summary:
+        print(f"  {s}")
+
+    if not patches:
+        print("\nNo patches generated. Add tuning overrides to your .muh file.")
+        return
+
+    if args.dry_run:
+        print("\n--- Patches ---\n")
+        for p in patches:
+            print(p["diff"])
+            print()
+    else:
+        combined = write_patches(patches, args.output_dir)
+        print(f"\nWritten to {args.output_dir}/")
+        print(f"Combined: {combined}")
+
+
+if __name__ == "__main__":
+    main()
--- a/muh/gen_yaml.py
+++ b/muh/gen_yaml.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+"""muh/gen_yaml.py — Generate computility-run.yaml from a .muh configuration
+
+Reads a .muh file (via parse.py), extracts the 'vllm' and 'env' sections,
+and outputs a computility-run.yaml compatible with ModelHub XC platform.
+
+Usage:
+    python3 muh/gen_yaml.py baseline.muh [-o computility-run.yaml]
+"""
+
+import os
+import sys
+import argparse
+
+# Import from sibling module
+sys.path.insert(0, os.path.dirname(__file__))
+from parse import load_muh
+
+
+# Mapping from .muh vllm config keys to CLI arguments
+VLLM_ARG_MAP = {
+    "model_path": "--model",
+    "served_model_name": "--served-model-name",
+    "max_model_len": "--max-model-len",
+    "gpu_memory_utilization": "--gpu-memory-utilization",
+    "tensor_parallel": "-tp",
+    "max_num_seqs": "--max-num-seqs",
+    "max_num_batched_tokens": "--max-num-batched-tokens",
+    "max_seq_len_to_capture": "--max-seq-len-to-capture",
+    "tool_call_parser": "--tool-call-parser",
+    "reasoning_parser": "--reasoning-parser",
+}
+
+# Boolean flags (presence = enabled, no value needed)
+VLLM_FLAG_MAP = {
+    "trust_remote_code": "--trust-remote-code",
+    "disable_log_requests": "--disable-log-requests",
+    "disable_frontend_multiprocessing": "--disable-frontend-multiprocessing",
+    "enable_chunked_prefill": "--enable-chunked-prefill",
+    "enable_auto_tool_choice": "--enable-auto-tool-choice",
+    "enable_prefix_caching": "--enable-prefix-caching",
+}
+
+
+def build_command(vllm_config):
+    """Build the command list for computility-run.yaml from vllm config."""
+    cmd = [
+        "python3",
+        "-m",
+        "vllm.entrypoints.openai.api_server",
+    ]
+
+    # Model path (required)
+    model_path = vllm_config.get("model_path", "/model")
+    cmd.extend(["--model", model_path])
+
+    # Named arguments
+    for muh_key, cli_arg in VLLM_ARG_MAP.items():
+        if muh_key == "model_path":
+            continue  # already handled
+        val = vllm_config.get(muh_key)
+        if val is not None:
+            cmd.extend([cli_arg, str(val)])
+
+    # Boolean flags
+    for muh_key, cli_flag in VLLM_FLAG_MAP.items():
+        if vllm_config.get(muh_key, False):
+            cmd.append(cli_flag)
+
+    # Extra raw args (pass-through)
+    extra = vllm_config.get("extra_args", [])
+    if isinstance(extra, list):
+        cmd.extend([str(a) for a in extra])
+
+    return cmd
+
+
+def build_env(env_config):
+    """Build environment variable list."""
+    env_list = []
+    for name, value in env_config.items():
+        env_list.append({"name": name, "value": str(value)})
+    return env_list
+
+
+def yaml_serialize_computility(concurrency, command, env_list):
+    """Serialize to computility-run.yaml format (no pyyaml dependency)."""
+    lines = []
+    lines.append(f"concurrency: {concurrency}")
+    lines.append("command:")
+    for item in command:
+        lines.append(f"    - '{item}'" if ' ' in str(item) else f"    - {item}")
+    if env_list:
+        lines.append("env:")
+        for e in env_list:
+            lines.append(f"    - name: {e['name']}")
+            lines.append(f"      value: {e['value']}")
+    return "\n".join(lines) + "\n"
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate computility-run.yaml from .muh")
+    parser.add_argument("muh_file", help="Path to .muh file")
+    parser.add_argument("-o", "--output", default="computility-run.yaml",
+                        help="Output file path (default: computility-run.yaml)")
+    parser.add_argument("--dry-run", action="store_true",
+                        help="Print to stdout instead of writing file")
+    args = parser.parse_args()
+
+    config = load_muh(args.muh_file)
+
+    vllm_config = config.get("vllm", {})
+    env_config = config.get("env", {})
+    concurrency = config.get("concurrency", 1)
+
+    command = build_command(vllm_config)
+    env_list = build_env(env_config)
+
+    output = yaml_serialize_computility(concurrency, command, env_list)
+
+    if args.dry_run:
+        print(output)
+    else:
+        with open(args.output, 'w') as f:
+            f.write(output)
+        print(f"Generated {args.output} ({len(command)} command args, {len(env_list)} env vars)")
+
+
+if __name__ == "__main__":
+    main()
--- a/muh/parse.py
+++ b/muh/parse.py
@@ -0,0 +1,229 @@
+#!/usr/bin/env python3
+"""muh/parse.py — Parse .muh configuration files
+
+A .muh file is YAML with these semantics:
+  - 'extends': inherit from another .muh file (deep merge, child overrides parent)
+  - 'hardware': target hardware description (warp_size, smem, registers, etc.)
+  - 'tuning': per-algorithm parameter overrides
+  - 'vllm': vllm-specific launch config (maps to computility-run.yaml)
+  - 'env': environment variable overrides
+
+Schema validation against muh/schema/*.yaml ensures parameter names and
+value ranges are legal.
+
+Usage:
+    python3 muh/parse.py baseline.muh [--schema-dir muh/schema] [--validate]
+"""
+
+import os
+import sys
+import copy
+import argparse
+from pathlib import Path
+
+
+def yaml_load_simple(text):
+    """Minimal YAML parser — handles flat dicts, nested dicts, lists, strings, numbers.
+    No dependency on PyYAML. Sufficient for .muh files."""
+    result = {}
+    stack = [(result, -1)]  # (dict, indent_level)
+    current_list_key = None
+
+    for line in text.split('\n'):
+        stripped = line.strip()
+        if not stripped or stripped.startswith('#'):
+            continue
+
+        indent = len(line) - len(line.lstrip())
+
+        # Pop stack to find parent at correct indent
+        while len(stack) > 1 and stack[-1][1] >= indent:
+            stack.pop()
+
+        parent, _ = stack[-1]
+
+        # List item
+        if stripped.startswith('- '):
+            val = stripped[2:].strip()
+            if current_list_key and current_list_key in parent:
+                if not isinstance(parent[current_list_key], list):
+                    parent[current_list_key] = []
+                parent[current_list_key].append(_parse_value(val))
+            continue
+
+        if ':' in stripped:
+            key, _, val = stripped.partition(':')
+            key = key.strip()
+            val = val.strip()
+
+            if val == '' or val == '|':
+                # Nested dict or upcoming list
+                parent[key] = {}
+                stack.append((parent[key], indent))
+                current_list_key = key
+            elif val.startswith('[') and val.endswith(']'):
+                # Inline list
+                items = [_parse_value(v.strip()) for v in val[1:-1].split(',') if v.strip()]
+                parent[key] = items
+                current_list_key = None
+            else:
+                parent[key] = _parse_value(val)
+                current_list_key = key if val == '' else None
+
+    return result
+
+
+def _parse_value(v):
+    """Parse a YAML scalar value."""
+    if v in ('true', 'True', 'yes'):
+        return True
+    if v in ('false', 'False', 'no'):
+        return False
+    if v in ('null', 'None', '~', 'TBD'):
+        return None
+    # Strip quotes
+    if (v.startswith('"') and v.endswith('"')) or (v.startswith("'") and v.endswith("'")):
+        return v[1:-1]
+    # Try int
+    try:
+        return int(v)
+    except ValueError:
+        pass
+    # Try float
+    try:
+        return float(v)
+    except ValueError:
+        pass
+    return v
+
+
+def deep_merge(base, override):
+    """Deep merge two dicts; override wins on conflicts."""
+    result = copy.deepcopy(base)
+    for k, v in override.items():
+        if k in result and isinstance(result[k], dict) and isinstance(v, dict):
+            result[k] = deep_merge(result[k], v)
+        else:
+            result[k] = copy.deepcopy(v)
+    return result
+
+
+def load_muh(filepath, search_dirs=None):
+    """Load a .muh file, resolving 'extends' chain."""
+    if search_dirs is None:
+        search_dirs = [os.path.dirname(filepath), '.']
+
+    with open(filepath, 'r') as f:
+        data = yaml_load_simple(f.read())
+
+    # Resolve extends
+    if 'extends' in data:
+        parent_name = data.pop('extends')
+        parent_path = None
+        for d in search_dirs:
+            candidate = os.path.join(d, parent_name)
+            if os.path.exists(candidate):
+                parent_path = candidate
+                break
+        if parent_path is None:
+            raise FileNotFoundError(f"Cannot find parent .muh file: {parent_name} (searched {search_dirs})")
+        parent_data = load_muh(parent_path, search_dirs)
+        data = deep_merge(parent_data, data)
+
+    return data
+
+
+def load_schema(schema_dir, algo):
+    """Load a schema YAML for validation."""
+    path = os.path.join(schema_dir, f"{algo}.yaml")
+    if not os.path.exists(path):
+        return None
+    with open(path, 'r') as f:
+        return yaml_load_simple(f.read())
+
+
+def validate_tuning(config, schema_dir):
+    """Validate tuning parameters against extracted schemas."""
+    errors = []
+    tuning = config.get('tuning', {})
+
+    for algo, params in tuning.items():
+        if not isinstance(params, dict):
+            continue
+        schema = load_schema(schema_dir, algo)
+        if schema is None:
+            errors.append(f"WARNING: No schema for algorithm '{algo}' — skipping validation")
+            continue
+
+        schema_params = schema.get('parameters', {})
+        for param_name, param_value in params.items():
+            if param_name.startswith('_'):  # metadata keys
+                continue
+            if param_name not in schema_params:
+                errors.append(f"{algo}.{param_name}: unknown parameter (not in schema)")
+                continue
+
+            spec = schema_params[param_name]
+            if spec.get('type') == 'int' and isinstance(param_value, (int, float)):
+                rng = spec.get('range', [0, 99999])
+                if not (rng[0] <= param_value <= rng[1]):
+                    errors.append(
+                        f"{algo}.{param_name}: value {param_value} outside range {rng}"
+                    )
+            elif spec.get('type') == 'enum' and isinstance(param_value, str):
+                valid = spec.get('values', [])
+                if param_value not in valid:
+                    errors.append(
+                        f"{algo}.{param_name}: '{param_value}' not in {valid}"
+                    )
+
+    return errors
+
+
+def print_config(config, indent=0):
+    """Pretty-print a parsed .muh config."""
+    prefix = "  " * indent
+    for k, v in config.items():
+        if isinstance(v, dict):
+            print(f"{prefix}{k}:")
+            print_config(v, indent + 1)
+        elif isinstance(v, list):
+            print(f"{prefix}{k}:")
+            for item in v:
+                print(f"{prefix}  - {item}")
+        else:
+            print(f"{prefix}{k}: {v}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Parse and validate .muh files")
+    parser.add_argument("muh_file", help="Path to .muh file")
+    parser.add_argument("--schema-dir", default="muh/schema",
+                        help="Schema directory (default: muh/schema)")
+    parser.add_argument("--validate", action="store_true",
+                        help="Validate tuning params against schemas")
+    parser.add_argument("--json", action="store_true",
+                        help="Output as JSON instead of pretty-print")
+    args = parser.parse_args()
+
+    config = load_muh(args.muh_file)
+
+    if args.validate:
+        errors = validate_tuning(config, args.schema_dir)
+        if errors:
+            print("Validation errors:", file=sys.stderr)
+            for e in errors:
+                print(f"  ✗ {e}", file=sys.stderr)
+            sys.exit(1)
+        else:
+            print("Validation passed ✓", file=sys.stderr)
+
+    if args.json:
+        import json
+        print(json.dumps(config, indent=2, ensure_ascii=False))
+    else:
+        print_config(config)
+
+
+if __name__ == "__main__":
+    main()
--- a/muh/schema/_index.yaml
+++ b/muh/schema/_index.yaml
@@ -0,0 +1,32 @@
+# muh schema index — all extracted CCCL tuning algorithms
+
+algorithms:
+  - adjacent_difference
+  - batch_memcpy
+  - batched_topk
+  - find
+  - find_bound_sorted_values
+  - for
+  - histogram
+  - merge
+  - merge_sort
+  - radix_sort
+  - reduce
+  - reduce_by_key
+  - rle_encode
+  - rle_non_trivial_runs
+  - scan
+  - scan_by_key
+  - segmented_radix_sort
+  - segmented_reduce
+  - segmented_scan
+  - segmented_sort
+  - select_if
+  - three_way_partition
+  - topk
+  - transform
+  - transform_tile
+  - unique_by_key
+
+total: 26
+source: cccl_upstream/cub/cub/device/dispatch/tuning/
--- a/muh/schema/adjacent_difference.yaml
+++ b/muh/schema/adjacent_difference.yaml
@@ -0,0 +1,56 @@
+# muh schema for adjacent_difference
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh
+# Generated by muh/extract.py
+
+algorithm: adjacent_difference
+source: cub/cub/device/dispatch/tuning/tuning_adjacent_difference.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  value_type_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/batch_memcpy.yaml
+++ b/muh/schema/batch_memcpy.yaml
@@ -0,0 +1,67 @@
+# muh schema for batch_memcpy
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_batch_memcpy.cuh
+# Generated by muh/extract.py
+
+algorithm: batch_memcpy
+source: cub/cub/device/dispatch/tuning/tuning_batch_memcpy.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  buffers_per_thread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  bytes_per_thread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  block_level_tile_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  warp_level_threshold:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  block_level_threshold:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  buffer_lookback_delay:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  block_lookback_delay:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/batched_topk.yaml
+++ b/muh/schema/batched_topk.yaml
@@ -0,0 +1,46 @@
+# muh schema for batched_topk
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_batched_topk.cuh
+# Generated by muh/extract.py
+
+algorithm: batched_topk
+source: cub/cub/device/dispatch/tuning/tuning_batched_topk.cuh
+parameters:
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/find.yaml
+++ b/muh/schema/find.yaml
@@ -0,0 +1,46 @@
+# muh schema for find
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_find.cuh
+# Generated by muh/extract.py
+
+algorithm: find
+source: cub/cub/device/dispatch/tuning/tuning_find.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  vec_size:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  input_type_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/find_bound_sorted_values.yaml
+++ b/muh/schema/find_bound_sorted_values.yaml
@@ -0,0 +1,47 @@
+# muh schema for find_bound_sorted_values
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_find_bound_sorted_values.cuh
+# Generated by muh/extract.py
+
+algorithm: find_bound_sorted_values
+source: cub/cub/device/dispatch/tuning/tuning_find_bound_sorted_values.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  range_type_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  values_type_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/for.yaml
+++ b/muh/schema/for.yaml
@@ -0,0 +1,24 @@
+# muh schema for for
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_for.cuh
+# Generated by muh/extract.py
+
+algorithm: for
+source: cub/cub/device/dispatch/tuning/tuning_for.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/histogram.yaml
+++ b/muh/schema/histogram.yaml
@@ -0,0 +1,117 @@
+# muh schema for histogram
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_histogram.cuh
+# Generated by muh/extract.py
+
+algorithm: histogram
+source: cub/cub/device/dispatch/tuning/tuning_histogram.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  pixels_per_thread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  vec_size:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  init_kernel_pdl_trigger_max_bins:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  sample_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  counter_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  sample_size_bytes:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  num_channels:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  num_active_channels:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm90:
+    -
+      threads: 768
+      items: 12
+      load_modifier: LOAD_LDG
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 960
+      items: 10
+      load_modifier: LOAD_DEFAULT
+      load_algorithm: BLOCK_LOAD_DIRECT
+  sm100:
+    -
+      items: 12
+      threads: 928
+      load_modifier: LOAD_CA
+      load_algorithm: BLOCK_LOAD_DIRECT
+      vec_size: 1
+    -
+      items: 12
+      threads: 448
+      load_modifier: LOAD_LDG
+      load_algorithm: BLOCK_LOAD_DIRECT
+      vec_size: 1
+      init_kernel_pdl_trigger_max_bins: 2048
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/merge.yaml
+++ b/muh/schema/merge.yaml
@@ -0,0 +1,89 @@
+# muh schema for merge
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_merge.cuh
+# Generated by muh/extract.py
+
+algorithm: merge
+source: cub/cub/device/dispatch/tuning/tuning_merge.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_align:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_is_trivially_relocatable:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_align:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_is_trivially_relocatable:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/merge_sort.yaml
+++ b/muh/schema/merge_sort.yaml
@@ -0,0 +1,63 @@
+# muh schema for merge_sort
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_merge_sort.cuh
+# Generated by muh/extract.py
+
+algorithm: merge_sort
+source: cub/cub/device/dispatch/tuning/tuning_merge_sort.cuh
+parameters:
+  ItemsPerThread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/radix_sort.yaml
+++ b/muh/schema/radix_sort.yaml
@@ -0,0 +1,103 @@
+# muh schema for radix_sort
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_radix_sort.cuh
+# Generated by muh/extract.py
+
+algorithm: radix_sort
+source: cub/cub/device/dispatch/tuning/tuning_radix_sort.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  private_partitions:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  radix_bits:
+    type: int
+    range:
+      - 4
+      - 8
+    step: 1
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  rank_private_partitions:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  threads:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  items:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/reduce.yaml
+++ b/muh/schema/reduce.yaml
@@ -0,0 +1,102 @@
+# muh schema for reduce
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_reduce.cuh
+# Generated by muh/extract.py
+
+algorithm: reduce
+source: cub/cub/device/dispatch/tuning/tuning_reduce.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  vec_size:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  reduce_algorithm:
+    type: enum
+    values:
+      - BLOCK_REDUCE_RAKING
+      - BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY
+      - BLOCK_REDUCE_WARP_REDUCTIONS
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  items:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  threads:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  items_per_vec_load:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm100:
+    -
+      items: 15
+      threads: 512
+      items_per_vec_load: 2
+    -
+      items: 15
+      threads: 512
+      items_per_vec_load: 1
+    -
+      items: 16
+      threads: 512
+      items_per_vec_load: 2
+    -
+      items: 16
+      threads: 640
+      items_per_vec_load: 1
+    -
+      threads_per_block: 256
+      items_per_thread: 16
+      items_per_vec_load: 4
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/reduce_by_key.yaml
+++ b/muh/schema/reduce_by_key.yaml
@@ -0,0 +1,390 @@
+# muh schema for reduce_by_key
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh
+# Generated by muh/extract.py
+
+algorithm: reduce_by_key
+source: cub/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 160
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 288
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 160
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 320
+      items: 23
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 19
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 9
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      nominal_4B_items_per_thread: 6
+  sm100:
+    -
+      items: 13
+      threads: 576
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      items: 10
+      threads: 224
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 128
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 19
+      threads: 128
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 14
+      threads: 128
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 256
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 11
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 10
+      threads: 160
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 10
+      threads: 224
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 11
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 10
+      threads: 256
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 9
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 11
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 9
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 9
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/rle_encode.yaml
+++ b/muh/schema/rle_encode.yaml
@@ -0,0 +1,155 @@
+# muh schema for rle_encode
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_rle_encode.cuh
+# Generated by muh/extract.py
+
+algorithm: rle_encode
+source: cub/cub/device/dispatch/tuning/tuning_rle_encode.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  length_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 19
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm100:
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 224
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      nominal_4B_items_per_thread: 6
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/rle_non_trivial_runs.yaml
+++ b/muh/schema/rle_non_trivial_runs.yaml
@@ -0,0 +1,155 @@
+# muh schema for rle_non_trivial_runs
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_rle_non_trivial_runs.cuh
+# Generated by muh/extract.py
+
+algorithm: rle_non_trivial_runs
+source: cub/cub/device/dispatch/tuning/tuning_rle_non_trivial_runs.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  length_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 192
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 288
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm100:
+    -
+      threads: 224
+      items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      threads: 224
+      items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 224
+      items: 13
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 288
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      nominal_4B_items_per_thread: 15
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/scan.yaml
+++ b/muh/schema/scan.yaml
@@ -0,0 +1,255 @@
+# muh schema for scan
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_scan.cuh
+# Generated by muh/extract.py
+
+algorithm: scan
+source: cub/cub/device/dispatch/tuning/tuning_scan.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  reduce_and_scan_warps:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  lookahead_items_per_thread:
+    type: int
+    range:
+      - 1
+      - 16
+    step: 1
+  lookahead_stages:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  block_idx_stages:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  delay_constructor:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  input_value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  input_value_alignment:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  output_value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  output_value_alignment:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_alignment:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm75:
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+  sm80:
+    -
+      threads: 320
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 352
+      items: 16
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 320
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 288
+      items: 22
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 288
+      items: 8
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 384
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 640
+      items: 24
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+  sm90:
+
+  sm100:
+    -
+      items: 18
+      threads: 512
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 384
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 13
+      threads: 512
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 13
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 22
+      threads: 384
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 19
+      threads: 416
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 23
+      threads: 416
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 22
+      threads: 320
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/scan_by_key.yaml
+++ b/muh/schema/scan_by_key.yaml
@@ -0,0 +1,450 @@
+# muh schema for scan_by_key
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh
+# Generated by muh/extract.py
+
+algorithm: scan_by_key
+source: cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 288
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 19
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 8
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 320
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 160
+      items: 17
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 160
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 17
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 13
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 320
+      items: 8
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 256
+      items: 16
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 128
+      items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 22
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 288
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      store_algorithm: BLOCK_STORE_DIRECT
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+  sm100:
+    -
+      items: 13
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 13
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 19
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 18
+      threads: 192
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 12
+      threads: 384
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 160
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 14
+      threads: 160
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 13
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 20
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 13
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 20
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 14
+      threads: 224
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 12
+      threads: 160
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 15
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      items: 22
+      threads: 160
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      items: 23
+      threads: 256
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      store_algorithm: BLOCK_STORE_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+      nominal_4b_items_per_thread: 9
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/segmented_radix_sort.yaml
+++ b/muh/schema/segmented_radix_sort.yaml
@@ -0,0 +1,26 @@
+# muh schema for segmented_radix_sort
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_segmented_radix_sort.cuh
+# Generated by muh/extract.py
+
+algorithm: segmented_radix_sort
+source: cub/cub/device/dispatch/tuning/tuning_segmented_radix_sort.cuh
+parameters:
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/segmented_reduce.yaml
+++ b/muh/schema/segmented_reduce.yaml
@@ -0,0 +1,60 @@
+# muh schema for segmented_reduce
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_segmented_reduce.cuh
+# Generated by muh/extract.py
+
+algorithm: segmented_reduce
+source: cub/cub/device/dispatch/tuning/tuning_segmented_reduce.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  threads_per_warp:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  vec_size:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/segmented_scan.yaml
+++ b/muh/schema/segmented_scan.yaml
@@ -0,0 +1,76 @@
+# muh schema for segmented_scan
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_segmented_scan.cuh
+# Generated by muh/extract.py
+
+algorithm: segmented_scan
+source: cub/cub/device/dispatch/tuning/tuning_segmented_scan.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  store_algorithm:
+    type: enum
+    values:
+      - BLOCK_STORE_DIRECT
+      - BLOCK_STORE_WARP_TRANSPOSE
+      - BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_STORE_STRIPED
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  max_segments:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  accum_align:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/segmented_sort.yaml
+++ b/muh/schema/segmented_sort.yaml
@@ -0,0 +1,82 @@
+# muh schema for segmented_sort
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_segmented_sort.cuh
+# Generated by muh/extract.py
+
+algorithm: segmented_sort
+source: cub/cub/device/dispatch/tuning/tuning_segmented_sort.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  radix_bits:
+    type: int
+    range:
+      - 4
+      - 8
+    step: 1
+  threads_per_warp:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  partitioning_threshold:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/select_if.yaml
+++ b/muh/schema/select_if.yaml
@@ -0,0 +1,508 @@
+# muh schema for select_if
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_select_if.cuh
+# Generated by muh/extract.py
+
+algorithm: select_if
+source: cub/cub/device/dispatch/tuning/tuning_select_if.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  input_size_bytes:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  flag_size_bytes:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size_bytes:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 992
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 576
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 384
+      items: 4
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 320
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 6
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 5
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 512
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 15
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 5
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 512
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 192
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 5
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 22
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 384
+      items: 17
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 384
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 512
+      items: 5
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 448
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 448
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 15
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 512
+      items: 3
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 320
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 128
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 192
+      items: 5
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 512
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 224
+      items: 6
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 160
+      items: 5
+      load_algorithm: BLOCK_LOAD_DIRECT
+  sm100:
+    -
+      threads: 384
+      nominal_4b_items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 448
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      nominal_4b_items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      nominal_4b_items: 19
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      nominal_4b_items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      nominal_4b_items: 5
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 512
+      nominal_4b_items: 5
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 896
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 1024
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      nominal_4b_items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 448
+      nominal_4b_items: 20
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      nominal_4b_items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      nominal_4b_items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 320
+      nominal_4b_items: 22
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 384
+      nominal_4b_items: 21
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_CA
+    -
+      threads: 512
+      nominal_4b_items: 3
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 512
+      nominal_4b_items: 3
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      nominal_4b_items: 15
+      threads: 608
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 22
+      threads: 320
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 19
+      threads: 320
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 416
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 22
+      threads: 576
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 608
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 18
+      threads: 608
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 14
+      threads: 512
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 22
+      threads: 224
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 22
+      threads: 320
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 19
+      threads: 608
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 23
+      threads: 416
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 608
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 22
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 19
+      threads: 608
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 23
+      threads: 416
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 448
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 18
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 19
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 21
+      threads: 384
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 448
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      nominal_4b_items: 14
+      threads: 320
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 14
+      threads: 640
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 19
+      threads: 384
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 24
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 18
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 11
+      threads: 448
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 20
+      threads: 384
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 12
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 12
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 14
+      threads: 352
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      nominal_4b_items: 11
+      threads: 512
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+      nominal_4B_items_per_thread: 10
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/three_way_partition.yaml
+++ b/muh/schema/three_way_partition.yaml
@@ -0,0 +1,185 @@
+# muh schema for three_way_partition
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh
+# Generated by muh/extract.py
+
+algorithm: three_way_partition
+source: cub/cub/device/dispatch/tuning/tuning_three_way_partition.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  input_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  offset_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 224
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm90:
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 320
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 384
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 24
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      threads: 640
+      items: 24
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 18
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 256
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+  sm100:
+    -
+      items: 12
+      threads: 256
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      items: 14
+      threads: 288
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 11
+      threads: 512
+      load_algorithm: BLOCK_LOAD_DIRECT
+    -
+      items: 10
+      threads: 256
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 20
+      threads: 768
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 20
+      threads: 768
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 15
+      threads: 768
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+    -
+      items: 14
+      threads: 320
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/topk.yaml
+++ b/muh/schema/topk.yaml
@@ -0,0 +1,52 @@
+# muh schema for topk
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_topk.cuh
+# Generated by muh/extract.py
+
+algorithm: topk
+source: cub/cub/device/dispatch/tuning/tuning_topk.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  bits_per_pass:
+    type: int
+    range:
+      - 4
+      - 11
+    step: 1
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/transform.yaml
+++ b/muh/schema/transform.yaml
@@ -0,0 +1,107 @@
+# muh schema for transform
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_transform.cuh
+# Generated by muh/extract.py
+
+algorithm: transform
+source: cub/cub/device/dispatch/tuning/tuning_transform.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread_no_input:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  min_items_per_thread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  max_items_per_thread:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  prefetch_byte_stride:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  unroll_factor:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  vec_size:
+    type: int
+    range:
+      - 1
+      - 8
+    step: 1
+  store_vec_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  min_bytes_in_flight:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  copy_alignment:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  smem_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  tile_padding:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  max_alignment:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/transform_tile.yaml
+++ b/muh/schema/transform_tile.yaml
@@ -0,0 +1,19 @@
+# muh schema for transform_tile
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_transform_tile.cuh
+# Generated by muh/extract.py
+
+algorithm: transform_tile
+source: cub/cub/device/dispatch/tuning/tuning_transform_tile.cuh
+parameters:
+  items:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD
--- a/muh/schema/unique_by_key.yaml
+++ b/muh/schema/unique_by_key.yaml
@@ -0,0 +1,379 @@
+# muh schema for unique_by_key
+# Auto-extracted from cub/cub/device/dispatch/tuning/tuning_unique_by_key.cuh
+# Generated by muh/extract.py
+
+algorithm: unique_by_key
+source: cub/cub/device/dispatch/tuning/tuning_unique_by_key.cuh
+parameters:
+  threads_per_block:
+    type: int
+    range:
+      - 32
+      - 1024
+    step: 32
+  items_per_thread:
+    type: int
+    range:
+      - 1
+      - 32
+    step: 1
+  load_algorithm:
+    type: enum
+    values:
+      - BLOCK_LOAD_DIRECT
+      - BLOCK_LOAD_VECTORIZE
+      - BLOCK_LOAD_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE
+      - BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED
+      - BLOCK_LOAD_STRIPED
+  load_modifier:
+    type: enum
+    values:
+      - LOAD_DEFAULT
+      - LOAD_CA
+      - LOAD_CG
+      - LOAD_CS
+      - LOAD_CV
+      - LOAD_LDG
+  scan_algorithm:
+    type: enum
+    values:
+      - BLOCK_SCAN_RAKING
+      - BLOCK_SCAN_RAKING_MEMOIZE
+      - BLOCK_SCAN_WARP_SCANS
+  lookback_delay.kind:
+    type: enum
+    values:
+      - no_delay
+      - fixed_delay
+      - exponential_backoff
+      - exponential_backoff_jitter
+      - exponential_backoff_jitter_window
+      - exponential_backon_jitter_window
+      - exponential_backon_jitter
+      - exponential_backon
+  lookback_delay.delay:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  lookback_delay.l2_write_latency:
+    type: int
+    range:
+      - 0
+      - 2000
+    step: 50
+  key_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+  value_size:
+    type: int
+    range:
+      - 1
+      - 1024
+    step: 1
+    note: unknown_range
+reference_tunings:
+  sm80:
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 224
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 15
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 320
+      items: 20
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 192
+      items: 22
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 10
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 192
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 128
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+  sm90:
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 448
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 288
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 288
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 23
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 448
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 9
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 9
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 9
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 640
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 448
+      items: 11
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+  sm100:
+    -
+      threads: 512
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 288
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 12
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_CA
+    -
+      threads: 384
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 224
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 11
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 512
+      items: 14
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 7
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 9
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 384
+      items: 10
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 576
+      items: 7
+      load_algorithm: BLOCK_LOAD_DIRECT
+      load_modifier: LOAD_DEFAULT
+    -
+      threads: 256
+      items: 9
+      load_algorithm: BLOCK_LOAD_WARP_TRANSPOSE
+      load_modifier: LOAD_DEFAULT
+bi_v100:
+  status: pending_benchmark
+  note: Run muh benchmark on Iluvatar BI-V100 to fill these values
+  threads_per_block: TBD
+  items_per_thread: TBD