feat: issue1 SM/QP预算模型 + issue3 AlltoAllv分阶段overlap优化

src/code/issue1/Makefile

@@ -0,0 +1,23 @@
+CXX := g++
+CXXFLAGS := -std=c++17 -O3 -Wall -Wextra
+TARGET := sm_budget_model
+TEST_TARGET := test_runner
+
+SRCS := main.cpp sm_budget_model.cpp bandwidth_model.cpp
+TEST_SRCS := ../../test/issue1/test.cpp sm_budget_model.cpp bandwidth_model.cpp
+
+.PHONY: all clean test
+
+all: $(TARGET)
+
+$(TARGET): $(SRCS)
+	$(CXX) $(CXXFLAGS) -o $@ $^
+
+test: $(TEST_TARGET)
+	./$(TEST_TARGET)
+
+$(TEST_TARGET): $(TEST_SRCS)
+	$(CXX) $(CXXFLAGS) -I. -o $@ $^
+
+clean:
+	rm -f $(TARGET) $(TEST_TARGET)

src/code/issue1/README.md

@@ -0,0 +1,34 @@
+# Issue 1：分析式 SM/QP 预算模型
+
+## 编译运行
+
+```bash
+cd src/code/issue1
+make                              # 编译（g++，无需GPU）
+./sm_budget_model --mode single   # 单次计算
+./sm_budget_model --mode sweep    # SM扫描输出CSV
+./sm_budget_model --mode compare  # 与固定24SM对比
+make test                         # 单元测试
+```
+
+## 参数
+
+| 参数 | 默认 | 说明 |
+|------|------|------|
+| `--num-experts` | 288 | 总专家数 |
+| `--num-topk` | 8 | 每token topk |
+| `--num-scaleout-ranks` | 8 | 跨节点rank数 |
+| `--num-scaleup-ranks` | 1 | 节点内rank数 |
+| `--num-device-sms` | 132 | 设备SM总数 |
+| `--rdma-gbs` | 0(自动50) | RDMA带宽 |
+| `--nvlink-gbs` | 0(自动450) | NVLink带宽 |
+
+## 结果
+
+```
+288专家/topk8/8节点 → 推荐4 SM，节省20 SM vs 固定24 SM，带宽不减
+```
+
+## 参考
+
+DeepEP `deep_ep/buffers/elastic.py` — `get_theoretical_num_sms()` (行 729-853)

src/code/issue1/bandwidth_model.cpp

@@ -0,0 +1,83 @@
+/*************************************************************************
+ * Copyright (c) 2025, TENCENT CORPORATION. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ *
+ * Author: moningchen@tencent.com
+ * Content: 带宽模型实现 — SM扫描与带宽估算
+ ************************************************************************/
+
+#include "bandwidth_model.h"
+
+#include <algorithm>
+#include <cmath>
+#include <sstream>
+#include <iomanip>
+
+double estimateBandwidthAtSM(const EPConfig& config, const BandwidthParams& bw, int num_sms) {
+    if (num_sms <= 0) return 0.0;
+
+    SMBudgetResult result = computeSMBudget(config, bw);
+
+    if (result.bounded_traffic <= 0.0 || result.bounded_gbs <= 0.0) {
+        return 1.0;  // 无流量时带宽视为最大
+    }
+
+    // SM处理能力: 受限于读或写的较小值
+    double sm_read_cap  = num_sms * bw.sm_read_gbs  / result.sm_read_traffic;
+    double sm_write_cap = num_sms * bw.sm_write_gbs / result.sm_write_traffic;
+    double sm_cap = std::min(sm_read_cap, sm_write_cap);
+
+    // 链路瓶颈容量
+    double link_cap = result.bounded_gbs / result.bounded_traffic;
+
+    // 实际带宽 = min(SM能力, 链路容量)
+    double effective = std::min(sm_cap, link_cap);
+
+    // 归一化: 1.0 = 链路饱和
+    double bw_normalized = effective / link_cap;
+    return std::min(bw_normalized, 1.0);
+}
+
+std::vector<SweepDataPoint> generateSMSweep(const EPConfig& config, const BandwidthParams& bw, int max_sms) {
+    std::vector<SweepDataPoint> sweep;
+    SMBudgetResult result = computeSMBudget(config, bw);
+    int recommended_sm = result.num_sms;
+
+    for (int sm = 4; sm <= max_sms; sm += 2) {
+        SweepDataPoint pt;
+        pt.num_sms        = sm;
+        pt.bandwidth      = estimateBandwidthAtSM(config, bw, sm);
+        pt.is_recommended = (sm == recommended_sm);
+        sweep.push_back(pt);
+    }
+    return sweep;
+}
+
+int findOptimalSM(const std::vector<SweepDataPoint>& sweep) {
+    if (sweep.empty()) return 4;
+
+    // 达到峰值带宽95%所需的SM数即视为最优
+    double peak_bw = sweep.back().bandwidth;
+    int optimal_sm = sweep[0].num_sms;
+
+    for (size_t i = 0; i < sweep.size(); ++i) {
+        if (sweep[i].bandwidth >= peak_bw * 0.95) {
+            optimal_sm = sweep[i].num_sms;
+            break;
+        }
+    }
+    return optimal_sm;
+}
+
+std::string exportSweepCSV(const std::vector<SweepDataPoint>& sweep, int recommended_sm) {
+    std::ostringstream oss;
+    oss << std::fixed << std::setprecision(4);
+    oss << "num_sms,bandwidth,is_recommended\n";
+    for (const auto& pt : sweep) {
+        oss << pt.num_sms << ","
+            << pt.bandwidth << ","
+            << (pt.num_sms == recommended_sm ? "1" : "0") << "\n";
+    }
+    return oss.str();
+}

src/code/issue1/bandwidth_model.h

@@ -0,0 +1,39 @@
+/*************************************************************************
+ * Copyright (c) 2025, TENCENT CORPORATION. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ *
+ * Author: moningchen@tencent.com
+ * Content: 带宽模型头文件 — SM扫描与带宽估算
+ ************************************************************************/
+
+#pragma once
+
+#include "sm_budget_model.h"
+
+#include <vector>
+#include <string>
+
+/*
+ * 估算给定SM数时的通信带宽（归一化，1.0=链路饱和）
+ *
+ * 模型: 实际带宽 = min(SM处理能力, 链路容量)
+ *   SM处理能力 = num_sms × 每SM带宽 / 流量需求（受限于读或写）
+ *   链路容量   = 瓶颈带宽 / 瓶颈流量
+ */
+double estimateBandwidthAtSM(const EPConfig& config, const BandwidthParams& bw, int num_sms);
+
+/*
+ * 生成SM扫描数据 (从4到max_sms, 步长2)
+ * 标记模型推荐的SM数为 is_recommended
+ */
+std::vector<SweepDataPoint> generateSMSweep(const EPConfig& config, const BandwidthParams& bw, int max_sms);
+
+/*
+ * 找到最优SM数: 边际带宽增益低于5%的拐点
+ * 即比峰值带宽低5%时对应的SM数
+ */
+int findOptimalSM(const std::vector<SweepDataPoint>& sweep);
+
+// 导出扫描数据为CSV格式
+std::string exportSweepCSV(const std::vector<SweepDataPoint>& sweep, int recommended_sm);