amv-forge 2.0.0__py3-none-any.whl

amv_forge-2.0.0.dist-info/METADATA

@@ -0,0 +1,260 @@
+Metadata-Version: 2.4
+Name: amv-forge
+Version: 2.0.0
+Summary: High-performance WebGPU-accelerated tensor computation engine with reverse-mode autograd for browser and edge deep learning.
+Author-email: AMEVA Foundation <contact@uno-km.com>
+Maintainer: AMEVA Engineering Team
+License: MIT
+Project-URL: Homepage, https://uno-km.github.io/amv-forge/
+Project-URL: Documentation, https://uno-km.github.io/amv-forge/
+Project-URL: Repository, https://github.com/uno-km/amv-forge
+Project-URL: Issues, https://github.com/uno-km/amv-forge/issues
+Project-URL: Changelog, https://github.com/uno-km/amv-forge/blob/main/RELEASE_NOTES.md
+Keywords: deep-learning,tensor,autograd,webgpu,pytorch,transformers,pyodide,wasm,machine-learning
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: numpy>=1.20.0
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0; extra == "dev"
+
+# amv-forge (`amv-forge`)
+
+[![PyPI Version](https://img.shields.io/badge/pypi-v1.0.1-blue.svg)](https://pypi.org/project/amv-forge/)
+[![Python Version](https://img.shields.io/badge/python-3.9%20%7C%203.10%20%7C%203.11%20%7C%203.12%20%7C%203.13-blue)](https://pypi.org/project/amv-forge/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/uno-km/amv-forge/blob/main/LICENSE)
+[![Test Suite](https://img.shields.io/badge/tests-279%2F279%20passed%20(100%25)-brightgreen.svg)](https://github.com/uno-km/amv-forge)
+[![WebGPU Acceleration](https://img.shields.io/badge/acceleration-WebGPU%20WGSL%20%2B%20CPU-orange.svg)](https://uno-km.github.io/amv-forge/)
+
+**High-Performance Client-Side Tensor Computation Engine & Reverse-Mode Autograd Framework Powered by WebGPU.**
+
+Developed and maintained by the **AMEVA Foundation** (?„ë©”ë°??¬ë‹¨), amv-forge is an industrial-grade, zero-server-cost deep learning library engineered to execute high-throughput tensor operations, automated differentiation, and end-to-end neural model training natively within client runtimes (WebGPU, WASM/Pyodide, and native Python environments).
+
+---
+
+## Architectural Pillars
+
+```
++-----------------------------------------------------------------------------------+
+|                            amv-forge User Space                                 |
+|   forge.nn  |  forge.optim  |  forge.linalg  |  forge.fft  |  forge.distributions |
++-----------------------------------------------------------------------------------+
+|                      Reverse-Mode Autograd DAG Engine                             |
+|       Vector-Jacobian Products (VJP)  *  In-Place Mutation Version Locks          |
++-----------------------------------------------------------------------------------+
+|                         Hardware Abstraction Layer                                |
+|   CPU Backend (Vectorized C/NumPy)  <--->  WebGPU Backend (Async WGSL Kernels)    |
+|   Staging Buffer Recycling Pool     <--->  Zero-Leak Allocation Token Ring        |
++-----------------------------------------------------------------------------------+
+```
+
+1. **Deterministic Autograd & Topological Execution**  
+   Strict reverse-mode automatic differentiation graph with cycle detection, multi-output tuple bindings, in-place version invalidation, and scalar-tensor memory optimization.
+2. **WebGPU Hardware Acceleration**  
+   Direct-to-silicon WGSL compute shaders featuring 8-dimensional non-contiguous stride dispatching, 2D workgroup partitioning ($65,535 \times 65,535$), and explicit buffer lifecycle tracking.
+3. **PyTorch 1:1 API Parity**  
+   Seamless drop-in compatibility across neural layers (`nn.Module`, `nn.MultiheadAttention`, `nn.Conv2d`), mathematical primitives (`linalg`, `fft`, `special`), and probabilistic graphical models (`distributions`).
+4. **Zero-Server Infrastructure (Edge & Browser)**  
+   Execute full model fine-tuning and inference directly inside the browser using Pyodide and WebGPU with zero cloud compute cost and total data privacy.
+
+---
+
+## Installation
+
+Install the official package from PyPI:
+
+```bash
+pip install amv-forge
+```
+
+Or install from source with development dependencies:
+
+```bash
+git clone https://github.com/uno-km/amv-forge.git
+cd amv-forge/packages/forge-py
+pip install -e .
+```
+
+---
+
+## Quick Start
+
+### 1. Basic Tensor & Automated Differentiation
+
+```python
+import forge as fg
+
+# Initialize tensors with gradient tracking
+x = fg.tensor([[1.0, 2.0], [3.0, 4.0]], requires_grad=True)
+w = fg.tensor([[0.5, -0.5], [1.0, 2.0]], requires_grad=True)
+b = fg.tensor([0.1, -0.1], requires_grad=True)
+
+# Forward pass: Linear projection + GELU activation
+y = fg.matmul(x, w) + b
+loss = fg.sum(fg.nn.functional.gelu(y))
+
+# Compute Vector-Jacobian Products (Autograd backward)
+loss.backward()
+
+print("Loss Value :", loss.numpy())
+print("Gradient dL/dw :\n", w.grad.numpy())
+```
+
+---
+
+### 2. Character-Level Transformer (NanoGPT)
+
+Train a complete causal autoregressive transformer directly on your local device:
+
+```python
+import forge as fg
+import forge.nn as nn
+from forge.models.nanogpt import GPT, GPTConfig
+
+# Define model configuration
+config = GPTConfig(
+    block_size=32,
+    vocab_size=64,
+    n_layer=4,
+    n_head=4,
+    n_embd=64,
+    bias=False
+)
+model = GPT(config)
+optimizer = fg.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-2)
+criterion = nn.CrossEntropyLoss()
+
+# Training step (Batch Size: 8, Sequence Length: 32)
+input_tokens = fg.tensor([[1, 5, 12, 3]], dtype="int32")
+target_tokens = fg.tensor([[5, 12, 3, 18]], dtype="int32")
+
+optimizer.zero_grad()
+logits = model(input_tokens)
+loss = criterion(logits, target_tokens)
+loss.backward()
+optimizer.step()
+```
+
+---
+
+### 3. Speech-to-Text & Acoustic Signal Processing (`forge.fft` + `forge.nn`)
+
+Compute real Fourier Mel-spectrograms from raw acoustic waveforms:
+
+```python
+import forge as fg
+import forge.nn as nn
+
+# 16kHz PCM audio waveform (Batch: 4, Samples: 8000)
+raw_audio = fg.tensor(audio_data, dtype="float32")
+
+# Fast Fourier Transform (Complex Spectrum)
+fft_complex = fg.fft.rfft(raw_audio, n=1024, dim=-1)
+
+# Power Spectrogram Energy
+power_spec = (fft_complex.real.pow(2.0) + fft_complex.imag.pow(2.0) + 1e-6).log()
+
+# 1D Convolutional Audio Feature Extractor
+conv = nn.Conv1d(in_channels=513, out_channels=64, kernel_size=3, padding=1)
+audio_features = conv(power_spec)
+```
+
+---
+
+## Comprehensive Module Directory
+
+| Module | Core Functionality | Key Operators / Classes |
+| :--- | :--- | :--- |
+| **`forge`** | Core Tensor Engine & Factories | `tensor`, `zeros`, `ones`, `randn`, `matmul`, `einsum`, `reshape`, `permute`, `where` |
+| **`forge.nn`** | Deep Learning Layers & Containers | `Linear`, `Conv1d`, `Conv2d`, `MultiheadAttention`, `LayerNorm`, `RMSNorm`, `BatchNorm2d`, `Embedding`, `CrossEntropyLoss`, `MSELoss` |
+| **`forge.optim`** | Optimizers & Rate Schedulers | `SGD`, `Adam`, `AdamW`, `RMSprop`, `CosineAnnealingLR`, `StepLR` |
+| **`forge.linalg`** | Linear Algebra Decomposition | `norm`, `svd`, `qr`, `cholesky`, `inv`, `pinv`, `det`, `matrix_rank`, `solve`, `eigh` |
+| **`forge.fft`** | Discrete Fourier Transforms | `rfft`, `irfft`, `fft`, `ifft`, `fft2`, `ifft2`, `rfft2`, `irfft2`, `fftfreq`, `fftshift` |
+| **`forge.special`** | Transcendental & Error Functions | `erf`, `erfc`, `erfinv`, `gammaln`, `digamma`, `expm1`, `log1p`, `expit`, `logit`, `sinc`, `i0`, `xlogy` |
+| **`forge.distributions`**| Probability Distributions & KL | `Normal(rsample)`, `Uniform`, `Bernoulli`, `Categorical`, `kl_divergence` |
+| **`forge.models`** | Pre-architected Reference Models | `GPT`, `GPTConfig`, `LLaMA` |
+
+---
+
+## In-Browser Zero-Install Execution (WebGPU + Pyodide)
+
+amv-forge packages a single bundled JavaScript distribution (`forge-py-bundle.js`) that mounts into browser-native Pyodide runtimes:
+
+```html
+<script src="https://cdn.jsdelivr.net/pyodide/v0.26.2/full/pyodide.js"></script>
+<script src="https://uno-km.github.io/amv-forge/dist/forge-py-bundle.js"></script>
+<script>
+  async function runClientDeepLearning() {
+    let pyodide = await loadPyodide();
+    await window.loadAmevaForgeBundle(pyodide);
+    
+    await pyodide.runPythonAsync(`
+      import forge as fg
+      x = fg.randn((1024, 1024), device="gpu")
+      y = fg.matmul(x, x)
+      print("Computed 1024x1024 on WebGPU Hardware:", y.shape)
+    `);
+  }
+  runClientDeepLearning();
+</script>
+```
+
+---
+
+## ?Œ The AMEVA Foundation (?„ë©”ë°??¬ë‹¨)
+
+**amv-forge** is an open-source deep learning initiative created and governed by the **AMEVA Foundation** (?„ë©”ë°??¬ë‹¨).
+
+### ?Ž¯ Our Mission & Zero-Server Paradigm
+The AMEVA Foundation is dedicated to the **democratisation of client-side artificial intelligence and edge computing**. We envision a decentralized web where deep learning training, fine-tuning, and neural inference occur directly on user devices (browsers, laptops, tablets, and smartphones)?”eliminating massive cloud server costs, safeguarding user data sovereignty with zero network leakage, and providing zero-latency neural intelligence everywhere.
+
+### ?ŒŸ Core Foundation Initiatives
+- ?Œ **amv-forge Tensor Engine**: High-performance WebGPU tensor computation & reverse-mode autograd engine.
+- ?Ž™ï¸?**On-Device Whisper STT**: Real-time acoustic FFT signal processing & speech recognition running locally in browser.
+- ?¤– **Decentralized LLM Transformers**: Zero-install character and subword transformer models with zero cloud dependency.
+- ?”’ **Privacy-First AI Sovereignty**: Sovereign on-device machine learning where private data never leaves client silicon.
+
+* **Foundation Portal**: [https://uno-km.github.io/amv-forge/](https://uno-km.github.io/amv-forge/)
+* **Official Repository**: [https://github.com/uno-km/amv-forge](https://github.com/uno-km/amv-forge)
+* **Community & Governance**: [https://github.com/uno-km/amv-forge/issues](https://github.com/uno-km/amv-forge/issues)
+* **Sponsorship & Research Partnerships**: [https://github.com/uno-km](https://github.com/uno-km)
+
+---
+
+## Quality Assurance & Verification
+
+Every release of amv-forge undergoes rigorous multi-tier verification:
+* **292 Unit & Stress Tests**: 100% automated pass rate across CPU, GPU fallback, mathematical accuracy, and memory quota managers.
+* **Finite-Difference Gradcheck**: Numerical gradient validation against analytical Vector-Jacobian backward formulations.
+* **Memory Lifecycle Audit**: Zero-leak allocation token reclamation and buffer recycling verification across 10,000+ continuous execution cycles.
+
+---
+
+## License & Citation
+
+amv-forge is licensed under the [MIT License](https://github.com/uno-km/amv-forge/blob/main/LICENSE).
+
+```bibtex
+@software{ameva_forge_2026,
+  author = {AMEVA Foundation},
+  title = {amv-forge: High-Performance WebGPU-Accelerated Tensor Computation Engine},
+  year = {2026},
+  publisher = {GitHub},
+  url = {https://github.com/uno-km/amv-forge}
+}
+```
+

amv_forge-2.0.0.dist-info/RECORD

@@ -0,0 +1,34 @@
+forge/__init__.py,sha256=dK2zXcm5fy41T2XXy99jdwS-AgpD-PlGNRNaaROlPU4,8264
+forge/__init__.pyi,sha256=ENmGiFdP3rqUMLTIZn63BVFqP4oK2braS4oNT0pHmLM,421
+forge/autograd.py,sha256=MMrwebcHruSUU_xfOzTYWCK3hzSbVtK_2307rkvyCCo,13825
+forge/bridge.py,sha256=CpnoSyKUB24tn4KZ7Z5sYLAE_5huSZPQyFN9FMcxCZU,10834
+forge/data.py,sha256=05bymKhhCCdV4l0jY2ulPY-I0Z1H9U1WfwSnW0CNeEo,7183
+forge/device.py,sha256=1fZV030cV08qtnFvoA4NJxrhvlVzJFvc-fBhOFRw_dE,6495
+forge/distributions.py,sha256=rU6esE_WmALj-y0Q1RjKW2ZhPpwcAHHkoAcNYU-QJgk,8855
+forge/errors.py,sha256=uMgyXnNRGmz_QlOifMjx0tOR1iipC-OUywNQMIoLZJE,2305
+forge/fft.py,sha256=9kTt1C8W4iMYWd-rkGa_6rmqK6VIQc-3WSAqdYsDXlU,6477
+forge/functional.py,sha256=c9cAK6UvGRaw2EEHESk5pLcbp0VzsMut_1BUh7g_RmA,43413
+forge/graph.py,sha256=7HkHuetOF9NEAR1U8_rwyW2hGvX1IVgXjvhIdooKpXg,10271
+forge/linalg.py,sha256=ZDc3ZeOuzOGpXBkuuErY3fUgEUf8wYI9LMwhcHedOKo,5708
+forge/nn.py,sha256=r4Vm2n9qXA8lhqkt69Pw35rrpnIG1KM86wqsz5awYzQ,88706
+forge/nn.pyi,sha256=p1epAZGOokCZo7EQ8N0KIiJWBHyF1M5vgJZBRiq88-A,1159
+forge/ops.py,sha256=tKWTeFTOAfZF8jyXOkZNDhdrIhyEP54KeRg2MXYaz88,210408
+forge/optim.py,sha256=Y-qnWygPPKuzbz4WCd-4oOk2tWXCKKTuISkXxU7EXD4,43765
+forge/optim.pyi,sha256=0DDDNQXpdn_giivT8UzTJguvyJ534M1pnD3_G_UNEWs,487
+forge/pipeline.py,sha256=ZZbvxfJedyGMU297mV3aCv1ZASrB4xOlEv-gLXPjePw,6322
+forge/py.typed,sha256=ZSegze1ahS5tlaVlEVAF0HG-jTkkScOwTjUnglYQOS8,22
+forge/serialization.py,sha256=XgHEixELxAbcMNHqv1BY93at5t8LzvSZsQw5eDm59eo,5114
+forge/special.py,sha256=_IjKwmVU05f8Dy0oQLWiCdF35CXCFYwzHw9x-W13vuM,10178
+forge/tensor.py,sha256=JJDugRK8U85lhIQ3WsZNXojBn_5tPO5pTVGkuAnNSCA,79457
+forge/tensor.pyi,sha256=a73RhYYShb5HjlsIJvnGwdI_d3prMWJtreus0kto8vc,2205
+forge/generated/__init__.py,sha256=aWydFp3Yrw_rjO3C1ZSbDLf4Hp0yJrlfVm8XipDxjhw,168
+forge/generated/op_schema.py,sha256=UJCvW2QHNcsaRw00hhXfN-Y-Eb3OoHr99RfvrmS4oE8,2077
+forge/models/__init__.py,sha256=IXGE6k1aq-19XcTGcup6445I9Ahic6YbwgwfP4IMF30,798
+forge/models/llama.py,sha256=XcxZFwtS31adqgg2CxojMGqZLWPwdMSq9VETk8KjIlY,10437
+forge/models/nanogpt.py,sha256=iUU4ZVb8XP6L47dEstA-rbBHOAxoVIkf_ljRhqZvyn4,6677
+forge/rl/__init__.py,sha256=DA-t-2B4AEYz9DRjxkeJaBtJuuqXimOEhSjRj1KsCAY,652
+forge/rl/cartpole.py,sha256=8Ana2x4eIeAFbp4okGOhaj33y0ypLNmJkasKRB7MnTs,5513
+amv_forge-2.0.0.dist-info/METADATA,sha256=MLXStBm9N65hocQbIlYTcS_blbpsVTi4wqpLFNkbnsA,11902
+amv_forge-2.0.0.dist-info/WHEEL,sha256=mOdXR9C3wMVov7cSY8PY-XCzo3SmmWCHvGQr5csJoao,91
+amv_forge-2.0.0.dist-info/top_level.txt,sha256=sDbe4KjRUBYyB4IABQOjHzoomNKH_4KwOv5fPPrv4ME,6
+amv_forge-2.0.0.dist-info/RECORD,,

amv_forge-2.0.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (84.2.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

amv_forge-2.0.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+forge

forge/__init__.py

@@ -0,0 +1,229 @@
+# Copyright (c) 2026 Uno Kim / AMEVA Foundation. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0 OR Commercial-License
+"""
+__init__.py ??forge ?�키지 공개 API
+
+M-07 Fix: wildcard import ?�거 ??명시??__all__ ?�의.
+NL-05 Fix: ones_like�?공개 API??추�?.
+NL-07 Fix: Quota ?��?.
+
+================================================================================
+?�일 ?�력 (Historical Metadata)
+Created: 2026-08-12 12:14:52 +0900 (�?커밋 기�?)
+Modified:
+  - 2026-08-12 12:59:35 +0900: Feat: Introduce v3.0 features (CNN, Pooling, Dropout, Serialization)
+  - 2026-08-12 12:23:09 +0900: Docs: Build Apache-style docs and unify tests
+  - 2026-08-12 12:14:52 +0900: Refactor: Rename AMEVA-Tensor to amv-forge and reorganize directories
+================================================================================
+WHAT: amv-forge ?�키지??주요 컴포?�트?�을 ??곳에 모아 ?��????�출?�는 진입???�일?�니??
+WHY: ?�용?��? ?�키지 ?��? 구조�????�요 ?�이 최상??모듈?�서 ?�리?�게 ?�요??기능?�을 가?�다 ?????�도�??�기 ?�함?�니??
+HOW: ?��???�??�브모듈(ops, tensor, nn, optim ???�서 ?�요???�래?�나 ?�수�?명시?�으�??�포?�하??__all__ 리스?��? ?�해 ?�출?�니??
+"""
+# ?�드?�어 ?�바?�스 초기??�??�태 관�?관??모듈 ?�포??
+from .device import init, is_available, current_device
+# 기본 ?�서 ?�산 �??�학/배열 조작 ?�수?�을 ops 모듈?�서 ?�포??
+from .ops import (
+    tensor, random, randn, matmul, relu, add, sub, mul, div, neg, transpose,
+    ones_like, zeros_like, zeros, ones, full,  # NL-05: 공개 API??추�?
+    to_numpy, to_numpy_async, dispose,
+    unsqueeze, squeeze, flatten, permute, max_op, max_axis, var, std, sqrt,
+    cat, where, pad, gather, scatter, dropout, conv2d, max_pool2d, avg_pool2d, bmm,
+    maximum, minimum, clamp,
+    arange, eye, linspace, triu, tril,
+    adaptive_avg_pool2d, adaptive_max_pool2d,
+    binary_cross_entropy_with_logits, smooth_l1_loss, kl_div, l1_loss,
+    topk, sort, argsort, TopKResult, SortResult,
+    isnan, isinf, isfinite, isclose, allclose, any_op, all_op,
+    conv1d, pixel_shuffle, pixel_unshuffle, interpolate,
+    stack, chunk, split, unbind,
+    multinomial, randint, bernoulli, normal,
+    roll, repeat_interleave, meshgrid, diag, diagonal, trace, outer,
+    masked_fill, index_select, masked_select, nonzero, take_along_dim,
+    cumprod, unflatten, cdist, norm, atleast_1d, atleast_2d, atleast_3d,
+    einsum, kron, tensordot, nan_to_num,
+    conv_transpose2d, affine_grid, grid_sample
+)
+any = any_op
+all = all_op
+# ?�서 객체 ?�체�??�의?�는 ?�래??�?GC ?�수 ?�포??
+from .tensor import Tensor, Parameter, flush_gc
+# ?�동 미분 기능 관??모듈 ?�포??
+from .autograd import no_grad, enable_grad, is_grad_enabled, set_grad_enabled, set_max_graph_nodes
+# ?�버�?모드�??�정?�고 가?�오???�틸리티 ?�수 ?�포??
+from .ops import set_debug_mode, get_debug_mode
+# ?�키지 ?�에??발생?????�는 ?�러 커스?� ?�러 ?�래??모음 ?�포??
+from .errors import (
+    AMEVAForgeError,
+    AMEVAForgeShapeError,
+    AMEVAForgeDTypeError,
+    AMEVAForgeDeviceError,
+    AMEVAForgeDisposedError,
+    AMEVAForgeWebGPUUnavailableError,
+    AMEVAForgeQuotaExceededError,   # NL-07: TypeScript?� ?��?
+    AMEVAForgeSecurityError,         # NL-07: TypeScript?� ?��?
+    AMEVAForgeValidationError,
+    AMEVAForgeOutOfMemoryError,
+    AMEVAForgeInternalGPUError,
+    AMEVAForgeDeviceLostError,
+    AMEVAForgeStaleHandleError,
+    AMEVAForgeUnsupportedOperationError,
+)
+
+# WHAT: ?�키지 버전???�의?�는 변?�입?�다.
+# WHY: ?�키지 ?�용?�나 ?�른 ?�구가 ?�재 ?�치???�키지??버전???�인?????�게 ?�니??
+# HOW: 문자???�태�?버전???�?�하???��??�서 ?�근 가?�하�??�니??
+__version__ = "1.0.1"
+
+# 기본 ?�계 �??�학 ?�산 ?�수?�을 ?�의???�른 ?�름?�로 매핑?�여 ?�포??
+from .ops import sum_op as sum, mean_op as mean, exp_op as exp, log_op as log
+from .ops import sigmoid, tanh_op as tanh, reshape, sum_axis, gelu, silu, leaky_relu, elu
+# ?�경�?구성 ?�소가 ?�는 nn ?�브모듈�?관???�이???�포??
+from . import nn
+from .nn import (
+    BatchNorm2d, Dropout, LayerNorm, MultiheadAttention, TransformerEncoderLayer,
+    PositionalEncoding, RNNCell, LSTMCell, RNN, LSTM, RMSNorm, RotaryEmbedding, SwiGLU,
+    Sequential, ModuleList, ModuleDict, ParameterList, ParameterDict, Parameter,
+    GELU, SiLU, LeakyReLU, ELU, CosineSimilarity,
+    Identity, ConstantPad2d, ZeroPad2d, ReflectionPad2d, ReplicationPad2d,
+    AdaptiveAvgPool2d, AdaptiveMaxPool2d,
+    CrossEntropyLoss, MSELoss, L1Loss, BCEWithLogitsLoss, SmoothL1Loss, KLDivLoss,
+    TripletMarginLoss, CosineEmbeddingLoss, MarginRankingLoss,
+    Conv1d, ConvTranspose2d, PixelShuffle, PixelUnshuffle, Upsample
+)
+# 최적???�고리즘???�는 optim ?�브모듈�??�습�??��?줄러 ?�포??
+from . import optim  
+from .optim import clip_grad_norm, clip_grad_norm_, clip_grad_value, clip_grad_value_, StepLR, CosineAnnealingLR, ReduceLROnPlateau
+# 모델 ?�??�?로드 ?�틸리티 ?�포??
+from .serialization import save_model, load_model
+# ?�수??API 묶음 ?�포??
+from . import functional as F
+from .functional import batch_norm2d, scaled_dot_product_attention, rms_norm, swiglu, rope
+# ?�이??로드 ?�틸리티 ?�포??
+from .data import DataLoader
+# 고수준 AI 모델, 강화?�습, ?�이?�라??모듈 ?�포??
+from . import models
+from . import rl
+from . import linalg
+from . import fft
+from . import special
+from .special import erf, erfc, erfinv, expm1, log1p, sinc
+from . import distributions
+from .pipeline import pipeline
+
+# WHAT: ?��??�서 `from forge import *`�??�출?????�출???�름?�의 리스?�입?�다.
+# WHY: ?�키지 ?��??�서�??�이???�겨�?private) 모듈?�나 변?��? ?�수�??�출?��? ?�도�??�어?�기 ?�해 ?�용?�니??
+# HOW: ?��???공개?�야 ???�수, ?�래?? ?�브모듈?�의 ?�름??문자??리스?�로 모아???�의?�니??
+__all__ = [
+    "linalg",
+    "fft",
+    "special",
+    "distributions",
+    "erf",
+    "erfc",
+    "erfinv",
+    "expm1",
+    "log1p",
+    "sinc",
+    # 초기??
+    "init",
+    "is_available",
+    "current_device",
+    # ?�서 ?�성
+    "tensor",
+    "random",
+    "randn",
+    "ones_like",
+    "zeros_like",
+    "zeros",
+    "sum",
+    "mean",
+    "exp",
+    "log",
+    "sigmoid",
+    "tanh",
+    "reshape",
+    "sum_axis",
+    "nn",
+    "optim",
+    "models",
+    "rl",
+    "pipeline",
+    "F",
+    "DataLoader",
+    "ones",
+    "full",
+    "Tensor",
+    "flush_gc",
+    "no_grad",
+    "BatchNorm2d",
+    "Dropout",
+    "LayerNorm",
+    "RMSNorm",
+    "RotaryEmbedding",
+    "SwiGLU",
+    "MultiheadAttention",
+    "TransformerEncoderLayer",
+    "PositionalEncoding",
+    "RNNCell",
+    "LSTMCell",
+    "RNN",
+    "LSTM",
+    "batch_norm2d",
+    "scaled_dot_product_attention",
+    "rms_norm",
+    "swiglu",
+    "rope",
+    "dropout",
+    "bmm",
+    # ?�산
+    "matmul",
+    "relu",
+    "add",
+    "sub",
+    "mul",
+    "div",
+    "neg",
+    "transpose",
+    "cat",
+    "where",
+    "pad",
+    "gather",
+    "scatter",
+    "conv2d",
+    "max_pool2d",
+    "avg_pool2d",
+    # ?�틸
+    "to_numpy",
+    "to_numpy_async",
+    "dispose",
+    # ?�러
+    "AMEVAForgeError",
+    "AMEVAForgeShapeError",
+    "AMEVAForgeDTypeError",
+    "AMEVAForgeDeviceError",
+    "AMEVAForgeDisposedError",
+    "AMEVAForgeWebGPUUnavailableError",
+    "AMEVAForgeQuotaExceededError",
+    "AMEVAForgeSecurityError",
+    "AMEVAForgeValidationError",
+    "AMEVAForgeOutOfMemoryError",
+    "AMEVAForgeInternalGPUError",
+    "AMEVAForgeDeviceLostError",
+    "AMEVAForgeStaleHandleError",
+    "AMEVAForgeUnsupportedOperationError",
+    # ?�버�??�정/Autograd
+    "no_grad",
+    "enable_grad",
+    "is_grad_enabled",
+    "set_grad_enabled",
+    "set_debug_mode",
+    "get_debug_mode",
+    "set_max_graph_nodes",
+    # 메�?
+    "__version__",
+    "unsqueeze", "squeeze", "flatten", "permute", 
+    "max_op", "max_axis", "var", "std", "sqrt", 
+    "clip_grad_norm", "clip_grad_value",
+    "StepLR", "CosineAnnealingLR", "ReduceLROnPlateau",
+    "save_model", "load_model"
+]

forge/__init__.pyi

@@ -0,0 +1,20 @@
+from typing import Any, Sequence, Union, Optional
+import numpy as np
+from .tensor import Tensor, tensor
+from . import nn as nn
+from . import optim as optim
+from . import functional as functional
+from .device import is_available, init_webgpu
+
+__version__: str
+
+__all__ = [
+    "Tensor",
+    "tensor",
+    "nn",
+    "optim",
+    "functional",
+    "is_available",
+    "init_webgpu",
+    "__version__",
+]