Does SimAI support tree topology?

[alizmhdi]

I recently tried to run a simulation using a tree topology and realized that the genAllReduceFlowModels function is returning the output of genAllReduceRingFlowModels for both the tree and ring topologies. (as seen in this switch case)

Additionally, the tutorial documentation mentions a variable called AS_NVLSTREE_ENABLE to enable NVLSTree, but I couldn't find any trace of this variable in the codebase.

Could you clarify if SimAI currently supports the tree model? If it does, how should it be properly configured?

Thank you!

[tianhao909]

【Manual reply】

Agree~~

This might be a lingering integration gap in the current version~

[tianhao909]

[Reply from Agent A] [For reference only]

Hello! Thank you for raising this issue and for digging deep into the source code. Your observations are spot on, and this is indeed a lingering integration gap in the current version.

To clarify your questions:

1. Whether SimAI supports the Tree model: At a design level, SimAI already has the core functions to compute Tree and NVLS_Tree flows implemented (i.e. genAllReduceTreeFlowModels and genallReduceNVLSTreeFlowModels). However, due to a known type-signature mismatch regarding the return types that these functions wrap, they were temporarily "skipped", forcefully making TREE fall back to behaving like RING and returning empty sets for NVLS_TREE within the master switch case. This was an oversight on our end.
2. About the missing environment variable: You're absolutely correct. AS_NVLSTREE_ENABLE hasn't been properly read via std::getenv into the main parsing logic within MockNcclGroup::get_algo_proto_info. This is a bug indicating the parameter from the environment is currently being ignored.

Next Steps / Configuration:
We've traced the issue to MockNcclGroup.cc and MockNcclGroup.h, and it requires:

• Parsing the AS_NVLSTREE_ENABLE environment variable into the algorithm dispatcher properly.
• Updating the return type map alignments and rewiring the switch cases exactly as you pointed out.

We would be incredibly grateful if you'd like to submit a Pull Request to bridge these functions back to the switch case and address the getenv handling.. Once fixed, simply prepending AS_NVLSTREE_ENABLE=1 (as mentioned in the tutorial) before your simulation scripts will correctly unlock and properly execute the simulated Tree configurations. Thank you again for your meticulous investigation!

你好！感谢您的详细反馈以及你对源码的观察。您提到的这个现象确实是在目前版本中存在的一个待对接完善的遗留问题。

针对您的问题，澄清说明如下：

1. 关于 SimAI 是否支持 Tree 模型：SimAI 在设计层面已经实现了处理 Tree 和 NVLS_Tree 通信组网流的代码逻辑（即底层的 genAllReduceTreeFlowModels 和 genallReduceNVLSTreeFlowModels 函数）。但由于这部分的函数的返回值签名与统一调度接口期望的映射结构发生了一些不匹配，此前在主干中被临时妥协，使得 TREE 被 fallback 到了 RING ，且 NVLS_TREE 被置空。我们承认这是一个遗漏。
2. 关于环境变量缺失：目前 AS_NVLSTREE_ENABLE 并没有真正在 C++ 的 MockNcclGroup::get_algo_proto_info 流程中被 getenv 读取和注入，属于环境侧传参未落地的 Bug。

应对方案：
我们已经定位了这段涉及需要修复的底层文件 (MockNcclGroup.cc)：

• 将该环境变量的获取注入到模型参数解析中。
• 把具体的 Tree 模型的函数调回正确的映射接口上。

如果您愿意的话，我们非常欢迎您提出 PR 来协助解决这一部分的数据类型和开关对接问题。一旦问题被修复，配合您在 Tutorial.md 里看到的配置 AS_NVLSTREE_ENABLE=1 就能真正使用该树模型的仿真！再次感谢您的细心排查。

[tianhao909]

[Reply from Agent B] [For reference only]

---

English version (suggested reply)

Follow-up on Issue #252

Thank you for raising this. Based on the current master branch of SimAI, here is a concise clarification.

1. Does genAllReduceFlowModels use genAllReduceRingFlowModels for both TREE and RING?
Yes. In MockNcclGroup::genAllReduceFlowModels, NCCL_ALGO_TREE and NCCL_ALGO_RING share the same case and both invoke genAllReduceRingFlowModels. The codebase also contains genAllReduceTreeFlowModels (and NVLS Tree–related genallReduceNVLSTreeFlowModels), but they are not wired into this switch today, so when the algorithm is TREE, flow generation still follows the ring path.

2. Why is AS_NVLSTREE_ENABLE not found in the code?
docs/Tutorial.md lists AS_NVLSTREE_ENABLE, but there is no corresponding getenv/usage in the source. NVLS is gated by AS_NVLS_ENABLE (used in get_algo_proto_info, etc.). This is a documentation vs. implementation mismatch; treat the code as authoritative unless the project ships an update.

3. Does SimAI support a “tree model”, and how should it be configured?

• Network topology: You can still define custom topologies via the provided topology tools and files.
• AllReduce “Tree” algorithm at the NCCL-mock / flow level: The current algorithm selection mostly uses RING for AllReduce; under H100/H800–style conditions, AS_NVLS_ENABLE=1 can select NVLS. There is no fully wired path that sets NCCL_ALGO_TREE or end-to-end NVLS Tree flow in today’s logic. Moreover, NCCL_ALGO_NVLS_TREE returns an empty map in genAllReduceFlowModels, and genallReduceNVLSTreeFlowModels is not called from that switch.
  So the accurate statement is: Tree / NVLS Tree flow-level simulation is not fully exposed as a complete, documented feature yet; aligning behavior with the tutorial would require code changes plus doc updates.

4. Relation to the aicb repo
The aicb project is mainly for workloads/benchmarks and does not replace SimAI’s internal MockNcclGroup routing; this issue should be resolved in SimAI itself.

If maintainers add Tree/NVLSTree support later, updating Tutorial.md with the real env vars, GPU/rank preconditions, and example commands would help users a lot.

---

中文版（建议回复）

关于 Issue #252 的说明

您好，感谢反馈。根据当前 master 上的代码与文档，说明如下。

1. genAllReduceFlowModels 中 TREE 与 RING 都调用 genAllReduceRingFlowModels 是否属实？
属实。在 MockNcclGroup::genAllReduceFlowModels 中，NCCL_ALGO_TREE 与 NCCL_ALGO_RING 落在同一分支，都会调用 genAllReduceRingFlowModels。仓库里另有 genAllReduceTreeFlowModels（以及 NVLS Tree 相关的 genallReduceNVLSTreeFlowModels），但 当前未接入该 switch，因此 AllReduce 在算法枚举为 TREE 时，流模型仍按 ring 路径生成。

2. AS_NVLSTREE_ENABLE 在代码里找不到？
文档 docs/Tutorial.md 的环境变量表中列出了 AS_NVLSTREE_ENABLE，但 代码中未发现对该环境变量的读取。与 NVLS 相关的是 AS_NVLS_ENABLE（在 get_algo_proto_info 等处使用）。因此属于 文档与实现不一致；应以代码为准，或等待官方将文档与实现同步。

3. SimAI 是否支持 tree 模型？如何配置？

• 网络层：用户仍可通过拓扑生成脚本等自定义 物理/逻辑网络结构（与 “tree” 相关的图结构）。
• 集合通信 AllReduce 的 “Tree 算法” 流模型：当前 算法选择逻辑 在多数场景下将 AllReduce 设为 RING；在 H100/H800 等条件下可通过 AS_NVLS_ENABLE=1 走 NVLS。未发现将算法设为 NCCL_ALGO_TREE 或完整走 NVLS Tree 流模型的已闭合配置路径；且 NCCL_ALGO_NVLS_TREE 在 genAllReduceFlowModels 中 返回空映射，genallReduceNVLSTreeFlowModels 未被调用。
  因此更准确的表述是：与 NCCL Tree / NVLS Tree 对应的流级仿真尚未作为端到端功能完整交付；若需与文档一致的 NVLSTree 开关，需要 代码与文档一并更新。

4. 与 aicb 仓库的关系
aicb 多用于工作负载与通信基准等，不替代 SimAI 内部对 MockNcclGroup 的算法绑定；本问题仍以 SimAI 源码为准。

如维护者后续实现 tree/NVLSTree，建议在 Tutorial 中明确环境变量、适用 GPU/规模条件及示例命令。

[tianhao909]

[Reply from Agent C] [For reference only]

Analysis reply for issue #252

Thanks for reporting this — your observation is correct. Based on the current codebase, SimAI does have an inconsistency between the implementation and the documentation regarding Tree / NVLSTree support.

1. Does SimAI currently support the tree model?

From the code, Tree-related data structures and flow-model generation functions do exist, for example:

• genAllReduceTreeFlowModels(...)
• generate_flow_model_tree_allreduce_up/down(...)
• gettreechannels(...)

However, in the current genAllReduceFlowModels(...) implementation:

• NCCL_ALGO_TREE
• NCCL_ALGO_RING

both dispatch to genAllReduceRingFlowModels(...).

So Tree allreduce is not currently routed to a dedicated tree flow-model path, even though some tree-related code already exists.

In other words:

• There is partial Tree implementation in the code
• but the main execution path is not fully wired up
• so it is not accurate to say that the tree model is already fully supported/configurable as documented

2. What about AS_NVLSTREE_ENABLE?

This variable currently appears in the documentation only.
There is no corresponding environment-variable handling logic in the code.

So at the moment, users cannot actually enable NVLSTree through AS_NVLSTREE_ENABLE.

This is a clear documentation vs implementation mismatch.

3. More accurate status summary

A more accurate description of the current repository state would be:

• Ring: wired up
• NVLS: wired up (via AS_NVLS_ENABLE)
• Tree: partially implemented, but not fully wired into the runtime path
• NVLSTree: some code traces exist, but no complete public enabling path is visible

4. Suggested minimal fix

A minimal and low-risk fix would be:

1. Change NCCL_ALGO_TREE dispatch from genAllReduceRingFlowModels(...) to genAllReduceTreeFlowModels(...)
2. Clarify whether NCCL_ALGO_NVLS_TREE is actually supported; if not, AS_NVLSTREE_ENABLE should not be documented as a usable option
3. Update docs/Tutorial.md so the documentation matches the implementation

5. If the question is “how should it be configured?”

Based on the current master branch, there is no complete, public, and verifiable configuration path for Tree / NVLSTree yet.

So the most accurate answer for now is:

• The current version cannot correctly enable Tree / NVLSTree only through the documented environment variable(s)
• a proper configuration guide should be added only after the algorithm-selection/runtime path is fully connected

对 issue #252 的分析回复

感谢反馈，这个问题判断得很准确。根据当前代码实现，SimAI 在 tree / NVLSTree 支持上确实存在“代码与文档不完全一致”的情况。

1. 关于 “SimAI 当前是否支持 tree model”

从代码上看，Tree 相关的数据结构和 flow model 生成函数是存在的，例如：

• genAllReduceTreeFlowModels(...)
• generate_flow_model_tree_allreduce_up/down(...)
• gettreechannels(...)

但目前 genAllReduceFlowModels(...) 中：

• NCCL_ALGO_TREE
• NCCL_ALGO_RING

这两个分支都会走 genAllReduceRingFlowModels(...)，因此 Tree allreduce flow model 目前没有被正确分发使用。

也就是说：

• 代码里有 Tree 的部分实现
• 但 当前主路径没有完整接通
• 因此目前不能认为 “tree model 已经可以按文档稳定启用”

2. 关于 AS_NVLSTREE_ENABLE

这个变量目前只在文档中出现，代码里没有对应的环境变量解析逻辑。
所以当前版本下，用户不能通过 AS_NVLSTREE_ENABLE 实际启用 NVLSTree。

这属于典型的 文档与实现不一致 问题。

3. 目前更准确的结论

更准确地说，当前仓库状态应理解为：

• Ring：已接通
• NVLS：已接通（通过 AS_NVLS_ENABLE）
• Tree：有部分实现，但未完全接通
• NVLSTree：有部分代码痕迹，但未看到完整可用的公开启用路径

4. 建议的最小修复方向

建议先做最小改动：

1. 将 NCCL_ALGO_TREE 从 genAllReduceRingFlowModels(...) 改为 genAllReduceTreeFlowModels(...)
2. 明确 NCCL_ALGO_NVLS_TREE 当前是否真正支持；若未完成，至少不要在文档中暴露 AS_NVLSTREE_ENABLE
3. 更新 docs/Tutorial.md，避免继续误导用户

5. 如果要回答“如何正确配置”

基于当前 master 分支代码，还没有看到一个完整、公开、可验证的 tree 配置入口。
因此在修复前，比较准确的答复应是：

• 当前版本不能仅通过现有文档中的环境变量来正确启用 tree / NVLSTree
• 若后续补上算法选择入口，再补充正式配置方式会更合适