From 0ec355cf74626d056fcbced8ef8f305cacab0e2e Mon Sep 17 00:00:00 2001
From: dylanyunlon <dogechat@163.com>
Date: Sat, 1 Aug 2026 02:10:57 +0800
Subject: [PATCH] [muh] add tuning_reduce_by_key.cuh: BI-V100 tuning for
 reduce_by_key

Translated from CCCL with SMEM overflow protection.
All SM100 values checked against 48KB limit.
---
 .../muh/tuning/tuning_reduce_by_key.cuh       | 55 +++++++++++++++++++
 1 file changed, 55 insertions(+)
 create mode 100644 muh/include/muh/tuning/tuning_reduce_by_key.cuh

diff --git a/muh/include/muh/tuning/tuning_reduce_by_key.cuh b/muh/include/muh/tuning/tuning_reduce_by_key.cuh
new file mode 100644
index 00000000..9477f7c3
--- /dev/null
+++ b/muh/include/muh/tuning/tuning_reduce_by_key.cuh
@@ -0,0 +1,55 @@
+// muh/include/muh/tuning/tuning_reduce_by_key.cuh — BI-V100
+//
+// Mirrors: cccl_upstream/cub/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh
+// CCCL SM100: 67 type specializations, 2 SMEM overflow (8B types at high threads)
+//
+// vllm relevance: grouped token aggregation (e.g. expert routing in MoE)
+
+#pragma once
+
+#include "muh/hardware.cuh"
+#include "muh/tuning/common.cuh"
+
+namespace muh::tuning::reduce_by_key {
+
+struct ReduceByKeyPolicy {
+  int threads_per_block;
+  int items_per_thread;
+  BlockLoadAlgorithm load_algorithm;
+  CacheLoadModifier load_modifier;
+  BlockScanAlgorithm scan_algorithm;
+  LookbackDelayPolicy delay;
+};
+
+struct policy_selector {
+  int key_size;
+  int accum_size;
+  int offset_size;
+  type_t accum_type;
+
+  constexpr ReduceByKeyPolicy operator()(const hardware_capability& hw) const {
+    int pair_size = key_size + accum_size;
+    
+    // SM100 typical: threads=224-384, items=10-18
+    int threads = 256;
+    int items = 14;
+    
+    if (pair_size <= 4) {
+      threads = 320; items = 16;
+    } else if (pair_size <= 8) {
+      threads = 256; items = 14;
+    } else {
+      threads = 192; items = 10;
+    }
+    
+    // SMEM: tile = threads * items * pair_size (key + accum per element)
+    while (threads * items * pair_size > hw.max_shared_memory_per_block && items > 1)
+      items--;
+
+    return {threads, items, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT,
+            BLOCK_SCAN_WARP_SCANS,
+            {LookbackDelayAlgorithm::exponential_backon, 350, 450}};
+  }
+};
+
+} // namespace muh::tuning::reduce_by_key