Merge pull request #134 from nullnonenilNULL/master

[docs] fix: update readme & speedup

Author: nullnonenilNULL

README.md

@@ -114,7 +114,7 @@ Training speedups over mainstream open-source baselines. Each row is tagged with
 | GR00T N1.6 | VLA | LeRobot | **2.31×** | master · 2026-07 |
 | Pi0.5 | VLA | OpenPI | **2.23×** | master · 2026-07 |
 | LingBot VA | WAM | LingBot-VA | **1.80×** | master · 2026-07 |
-| X-VLA | VLA | X-VLA | **1.6×** | master · 2026-07 |
+| X-VLA | VLA | X-VLA | **1.69×** | master · 2026-07 |
 | DeepSeek-V3.2 Lite <sup>§</sup> | MoE + DSA | Megatron-LM | **5.04×** | v0.1.1 |
 | Qwen3-VL-30B-A3B | VLM | VeOmni | **1.45×** | v0.1.1 |
 | Qwen3-30B-A3B | MoE | Megatron-LM | **1.16×** | v0.1.1 |

README_zh.md

@@ -114,7 +114,7 @@
 | GR00T N1.6 | VLA | LeRobot | **2.31×** | master · 2026-07 |
 | Pi0.5 | VLA | OpenPI | **2.23×** | master · 2026-07 |
 | LingBot VA | WAM | LingBot-VA | **1.80×** | master · 2026-07 |
-| X-VLA | VLA | X-VLA | **1.6×** | master · 2026-07 |
+| X-VLA | VLA | X-VLA | **1.69×** | master · 2026-07 |
 | DeepSeek-V3.2 Lite <sup>§</sup> | MoE + DSA | Megatron-LM | **5.04×** | v0.1.1 |
 | Qwen3-VL-30B-A3B | VLM | VeOmni | **1.45×** | v0.1.1 |
 | Qwen3-30B-A3B | MoE | Megatron-LM | **1.16×** | v0.1.1 |

docs/assets/images/benchmark_image.py

@@ -27,7 +27,7 @@
     ("GR00T N1.6",         "VLA",       "LeRobot",     "GBS 96 · 224×224",               2.31, "",  "master · 2026-07"),
     ("Pi 0.5",             "VLA",       "OpenPI",      "GBS 96 · 224×224×2",             2.23, "",  "master · 2026-07"),
     ("LingBot VA",         "WAM",       "LingBot-VA",  "LIBERO",                         1.80, "",  "master · 2026-07"),
-    ("X-VLA",              "VLA",       "X-VLA",       "GBS 288",                        1.60, "",  "master · 2026-07"),
+    ("X-VLA",              "VLA",       "X-VLA",       "GBS 288",                        1.69, "",  "master · 2026-07"),
     ("DeepSeek-V3.2 Lite", "MoE + DSA", "Megatron-LM", "Reduced layers · GBS 128 · 8K",  5.04, "§", "v0.1.1"),
     ("Qwen3-VL-30B-A3B",   "VLM",       "VeOmni",      "GBS 128 · 32K",                  1.45, "",  "v0.1.1"),
     ("Qwen3-30B-A3B",      "MoE",       "Megatron-LM", "GBS 1024 · 32K",                 1.16, "",  "v0.1.1"),

docs/assets/images/benchmark_speedup.png

@@ -1,21 +1,49 @@
 # LoongForge-Embodied — Embodied Model Training Subsystem
 
-`loongforge/embodied/` is a **self-contained training subsystem for embodied foundation models**: Vision-Language-Action (VLA) policies (e.g. pi0.5) and world-action models (WAM, e.g. FastWAM). It lives inside the LoongForge monorepo alongside the Megatron-based LLM / VLM / Diffusion stacks, but is built on a torch-native DDP/FSDP engine instead.
+`loongforge/embodied/` is a **torch-native training subsystem for embodied models** — Vision-Language-Action (VLA) policies and world-action models (WAM) — combining broad open-source model support with production-grade training performance.
+
+- **Torch-native architecture** — built on vanilla PyTorch around unified `data`, `model`, and `trainer` abstractions that can be flexibly reused or extended.
+- **Extensive open-source model support** — π0.5, GR00T-N1.6 / N1.7, xVLA, Lingbot-VA, FastWAM, DreamZero, Cosmos3, and more — each supports fine-tuning, with accuracy aligned to the official baselines.
+- **High training throughput** — up to 2x+ throughput on representative models through `torch.compile`, CUDA Graph, custom kernels, and I/O optimization, plus a full range of distributed strategies: DDP, ZeRO-1, FSDP, and HSDP.
 
 ---
 
 ## Why a Separate Subsystem?
 
-Compared with typical large models, embodied models are far smaller (generally under 10B) — a typical VLA is a VLM plus an action head. Their bottleneck is data plumbing and iteration speed, not model parameter scale — so Megatron's TP/PP/EP model parallelism does little here and only adds complexity.
+Embodied models are far smaller than typical LLMs (generally under 10B) — a VLA is essentially a VLM plus an action head — so their bottleneck is not model parameter scale. Megatron's TP/PP/EP model parallelism brings little benefit at this scale.
 
-This subsystem is therefore built on **plain PyTorch DDP/FSDP**, with its own configuration, trainer, data, distributed, and evaluation layers. It shares LoongForge's repository, release, and tooling — but not the Megatron core engine. The two stacks are intentionally decoupled (no shared args/parser/core) so each evolves on its own terms:
+The subsystem is therefore built on a **torch-native DDP/FSDP** engine with its own configuration, trainer, data, distributed, and evaluation layers. It shares LoongForge's repository, release, and tooling, but stays intentionally decoupled from the Megatron core (no shared args / parser / core) so each stack evolves on its own terms. The core abstractions below all follow from that choice.
 
-| Axis | LoongForge core (LLM / VLM / Diffusion) | LoongForge-Embodied |
-|------|------------------------------------------|---------------------|
-| Compute / distributed | Megatron-LM — TP / PP / EP / CP / FSDP | torch-native DDP / FSDP |
-| Workload | large-scale, model-parallel pretrain/SFT | small-to-mid, data-parallel SFT |
+---
 
-The core abstractions below all follow from that choice.
+## Quick Start
+
+For the full framework user guide, see [User Manual](../../docs/source/embodied_tutorial/overview.md). Model-specific quick starts:
+
+- [π0.5 (pi05)](../../docs/source/embodied_tutorial/quick_start_pi05.md)
+- [GR00T-N1.6](../../docs/source/embodied_tutorial/quick_start_groot_n1_6.md)
+- [GR00T-N1.7](../../docs/source/embodied_tutorial/quick_start_groot_n1_7.md)
+- [FastWAM](../../docs/source/embodied_tutorial/quick_start_fastwam.md)
+- [DreamZero](../../docs/source/embodied_tutorial/quick_start_dreamzero.md)
+- [Cosmos3](../../docs/source/embodied_tutorial/quick_start_cosmos3.md)
+- [xVLA](../../docs/source/embodied_tutorial/quick_start_xvla.md)
+- [Lingbot-VA](../../docs/source/embodied_tutorial/quick_start_lingbot_va.md)
+
+---
+
+## Performance
+
+Training speedups over mainstream open-source baselines. Performance is still under active optimization, and these numbers will keep improving over time:
+
+| Model | Type | Baseline | Speedup |
+|---|---|---|---|
+| DreamZero (DROID Wan2.2-5B Full) | WAM | DreamZero | **2.67×** |
+| GR00T-N1.6 | VLA | LeRobot | **2.31×** |
+| π0.5 | VLA | OpenPI | **2.23×** |
+| Lingbot-VA | WAM | LingBot-VA | **1.80×** |
+| xVLA | VLA | X-VLA | **1.69×** |
+
+Numbers reflect the baseline and LoongForge versions at measurement time and may evolve as implementations change. See the [root README](../../README.md#-performance) for the full benchmark chart across all model families.
 
 ---
 
@@ -58,11 +86,11 @@ loongforge/embodied/
 
 The entry point `train.py` is self-explanatory (parse configs → build trainer → train), so we skip it. The real core is the four abstractions below.
 
-### 1. Model networking (`model/`)
+### 1. Model definition (`model/`)
 
 One directory per model, registered into a single entry via `@register_model`:
 
-- `modeling_<name>.py` — networking, forward, loss;
+- `modeling_<name>.py` — architecture, forward, loss;
 - `model_configuration_<name>.py` — model config dataclass (architecture hyperparams);
 - exposes a uniform interface upward (trainer / eval), so adding a model requires no change to the training loop.
 
@@ -102,19 +130,6 @@ Resolution flow: `--model-name` routes through `config_map.py` to the model's YA
 
 ---
 
-## Quick Start
-
-For the full framework user guide, see [User Manual](../../docs/source/embodied_tutorial/overview.md). Model-specific quick starts:
-
-- [π0.5 (pi05)](../../docs/source/embodied_tutorial/quick_start_pi05.md)
-- [GR00T-N1.6](../../docs/source/embodied_tutorial/quick_start_groot_n1_6.md)
-- [GR00T-N1.7](../../docs/source/embodied_tutorial/quick_start_groot_n1_7.md)
-- [FastWAM](../../docs/source/embodied_tutorial/quick_start_fastwam.md)
-- [DreamZero](../../docs/source/embodied_tutorial/quick_start_dreamzero.md)
-- [Cosmos3](../../docs/source/embodied_tutorial/quick_start_cosmos3.md)
-- [xVLA](../../docs/source/embodied_tutorial/quick_start_xvla.md)
-- [Lingbot-VA](../../docs/source/embodied_tutorial/quick_start_lingbot_va.md)
-
 ## Evaluation
 
 Offline benchmark evaluation (LIBERO / CALVIN / SimplerEnv / RoboTwin / ManiSkill) is a separate module. See [Eval User Guide](../../docs/source/embodied_tutorial/eval_user_guide.md).

loongforge/embodied/README.md

@@ -1,21 +1,49 @@
 # LoongForge-Embodied —— 具身模型训练子系统
 
-`loongforge/embodied/` 是 LoongForge 主仓库内一个**自成体系的具身基础模型训练子系统**，既涵盖视觉-语言-动作（VLA）策略模型（如 pi0.5），也涵盖世界-动作模型（WAM，如 FastWAM）。它与基于 Megatron 的 LLM / VLM / Diffusion 栈并列存在，但底层构建在 torch 原生的 DDP / FSDP 之上。
+`loongforge/embodied/` 是一个 **torch 原生的具身模型训练子系统**——涵盖视觉-语言-动作（VLA）策略模型与世界-动作模型（WAM）——在广泛的开源模型支持之上，兼顾生产级的训练性能。
+
+- **Torch 原生架构**：完全基于原生 PyTorch，构建了统一的 `data`、`model`、`trainer` 抽象，可灵活复用或扩展。
+- **广泛的开源模型支持**：π0.5、GR00T-N1.6 / N1.7、xVLA、Lingbot-VA、FastWAM、DreamZero、Cosmos3 等均支持微调，且精度与官方 baseline 对齐。
+- **高训练吞吐**：典型模型最高可达 2x+ 吞吐提升，集成 `torch.compile`、CUDA Graph、自定义 kernel 与 I/O 优化，并支持 DDP、ZeRO-1、FSDP、HSDP 等多种分布式策略。
 
 ---
 
 ## 为什么单独拆成一个子系统？
 
-与典型大模型相比，具身模型参数量小得多（通常在 10B 以内），典型的 VLA 多为一个 VLM 叠加一个动作头。其瓶颈在于数据管道与迭代速度，而非模型参数规模——因此 Megatron 的 TP/PP/EP 模型并行在这里作用有限，反而徒增复杂度。
+与典型大模型相比，具身模型参数量小得多（通常在 10B 以内），典型的 VLA 多为一个 VLM 叠加一个动作头，其瓶颈不在于模型参数规模。因此 Megatron 的 TP/PP/EP 模型并行在这个规模下收益有限。
+
+基于这一判断，该子系统构建在**原生 PyTorch 的 DDP/FSDP** 引擎之上，拥有独立的配置、训练器、数据、分布式与评测层。它共享 LoongForge 的代码仓库、发布流程与工具链，但与 Megatron 核心刻意保持解耦（不共享 args / parser / core），从而两套栈各自独立演进。下文的核心抽象，都是这个选择的自然结果。
+
+---
+
+## 快速开始
+
+完整的框架使用说明请参阅 [用户手册](../../docs/source_zh/embodied_tutorial/overview.md)，各模型的快速入门：
+
+- [π0.5 (pi05)](../../docs/source_zh/embodied_tutorial/quick_start_pi05.md)
+- [GR00T-N1.6](../../docs/source_zh/embodied_tutorial/quick_start_groot_n1_6.md)
+- [GR00T-N1.7](../../docs/source_zh/embodied_tutorial/quick_start_groot_n1_7.md)
+- [FastWAM](../../docs/source_zh/embodied_tutorial/quick_start_fastwam.md)
+- [DreamZero](../../docs/source_zh/embodied_tutorial/quick_start_dreamzero.md)
+- [Cosmos3](../../docs/source_zh/embodied_tutorial/quick_start_cosmos3.md)
+- [xVLA](../../docs/source_zh/embodied_tutorial/quick_start_xvla.md)
+- [Lingbot-VA](../../docs/source_zh/embodied_tutorial/quick_start_lingbot_va.md)
+
+---
+
+## 性能
 
-基于这一判断，该子系统直接构建在**原生 PyTorch 的 DDP/FSDP** 之上，拥有独立的配置、训练器、数据、分布式与评测层。它共享 LoongForge 的代码仓库、发布流程与工具链，但不共享 Megatron 核心引擎；两套栈刻意保持**解耦**（不共享 args/parser/core），从而各自独立演进：
+相较主流开源 baseline 的训练加速比（性能仍在积极优化中，这些数字后续还会持续提升）：
 
-| 维度 | LoongForge 核心（LLM / VLM / Diffusion） | LoongForge-Embodied |
-|------|------------------------------------------|---------------------|
-| 计算 / 分布式 | Megatron-LM —— TP / PP / EP / CP / FSDP | torch-native DDP / FSDP |
-| 负载 | 大规模、模型并行的预训练/SFT | 中小规模、数据并行的 SFT |
+| 模型 | 类型 | Baseline | 加速比 |
+|---|---|---|---|
+| DreamZero (DROID Wan2.2-5B Full) | WAM | DreamZero | **2.67×** |
+| GR00T-N1.6 | VLA | LeRobot | **2.31×** |
+| π0.5 | VLA | OpenPI | **2.23×** |
+| Lingbot-VA | WAM | LingBot-VA | **1.80×** |
+| xVLA | VLA | X-VLA | **1.69×** |
 
-下文的核心抽象，都是这个选择的自然结果。
+数据反映测量时刻的 baseline 与 LoongForge 版本，可能随实现演进而变化。跨所有模型族的完整基准图表见 [根 README](../../README_zh.md#-性能表现)。
 
 ---
 
@@ -56,11 +84,13 @@ loongforge/embodied/
 
 ## 核心抽象
 
-### 1. 模型组网（`model/`）
+入口 `train.py` 本身一目了然（解析配置 → 构建训练器 → 训练），此处略过，真正的核心是下面四个抽象。
+
+### 1. 模型定义（`model/`）
 
 每个模型一个目录，通过 `@register_model` 注册到统一入口：
 
-- `modeling_<name>.py` —— 组网、前向、损失计算；
+- `modeling_<name>.py` —— 架构、前向、损失计算；
 - `model_configuration_<name>.py` —— 模型配置数据类（架构超参）；
 - 对上层（Trainer / 评测）暴露统一接口，新增模型无需改训练循环。
 
@@ -100,19 +130,6 @@ loongforge/embodied/
 
 ---
 
-## 快速开始
-
-完整的框架使用说明请参阅 [用户手册](../../docs/source_zh/embodied_tutorial/overview.md)，各模型的快速入门：
-
-- [π0.5 (pi05)](../../docs/source_zh/embodied_tutorial/quick_start_pi05.md)
-- [GR00T-N1.6](../../docs/source_zh/embodied_tutorial/quick_start_groot_n1_6.md)
-- [GR00T-N1.7](../../docs/source_zh/embodied_tutorial/quick_start_groot_n1_7.md)
-- [FastWAM](../../docs/source_zh/embodied_tutorial/quick_start_fastwam.md)
-- [DreamZero](../../docs/source_zh/embodied_tutorial/quick_start_dreamzero.md)
-- [Cosmos3](../../docs/source_zh/embodied_tutorial/quick_start_cosmos3.md)
-- [xVLA](../../docs/source_zh/embodied_tutorial/quick_start_xvla.md)
-- [Lingbot-VA](../../docs/source_zh/embodied_tutorial/quick_start_lingbot_va.md)
-
 ## 评测
 
 离线 benchmark 评测（LIBERO / CALVIN / SimplerEnv / RoboTwin / ManiSkill）是一个独立模块，详见 [评测用户指南](../../docs/source_zh/embodied_tutorial/eval_user_guide.md)。