luminav 1.5.3.post1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Potentially problematic release.

This version of luminav might be problematic. Click here for more details.

Files changed (11) hide show
  1. luminav-1.5.3.post1/LICENSE +13 -0
  2. luminav-1.5.3.post1/PKG-INFO +115 -0
  3. luminav-1.5.3.post1/README.md +84 -0
  4. luminav-1.5.3.post1/pyproject.toml +51 -0
  5. luminav-1.5.3.post1/setup.cfg +4 -0
  6. luminav-1.5.3.post1/src/luminav/__init__.py +25 -0
  7. luminav-1.5.3.post1/src/luminav.egg-info/PKG-INFO +115 -0
  8. luminav-1.5.3.post1/src/luminav.egg-info/SOURCES.txt +9 -0
  9. luminav-1.5.3.post1/src/luminav.egg-info/dependency_links.txt +1 -0
  10. luminav-1.5.3.post1/src/luminav.egg-info/requires.txt +1 -0
  11. luminav-1.5.3.post1/src/luminav.egg-info/top_level.txt +1 -0
luminav-1.5.3.post1/LICENSE ADDED
@@ -0,0 +1,13 @@
1
+ Copyright 2026 Silver Moon (cloverxion) and Lumina Moon (cloverx-id) Contributors
2
+
3
+ Licensed under the Apache License, Version 2.0 (the "License");
4
+ you may not use this file except in compliance with the License.
5
+ You may obtain a copy of the License at
6
+
7
+ https://www.apache.org/licenses/LICENSE-2.0
8
+
9
+ Unless required by applicable law or agreed to in writing, software
10
+ distributed under the License is distributed on an "AS IS" BASIS,
11
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ See the License for the specific language governing permissions and
13
+ limitations under the License.
luminav-1.5.3.post1/PKG-INFO ADDED
@@ -0,0 +1,115 @@
1
+ Metadata-Version: 2.4
2
+ Name: luminav
3
+ Version: 1.5.3.post1
4
+ Summary: LuminaV Optimizer (Moved to 'luminamoon' - Official PROJECT LUMI Package)
5
+ Author: Lumina Moon (cloverx-id) Contributors
6
+ Author-email: Silver Moon <leafmoon.idn@gmail.com>
7
+ License-Expression: Apache-2.0
8
+ Project-URL: Homepage, https://github.com/cloverx-id/luminamoon
9
+ Project-URL: Repository, https://github.com/cloverx-id/luminamoon
10
+ Project-URL: Official PyPI (luminamoon), https://pypi.org/project/luminamoon/
11
+ Project-URL: Documentation, https://github.com/cloverx-id/luminamoon#readme
12
+ Project-URL: Changelog, https://github.com/cloverx-id/luminamoon/tree/main/changelogs
13
+ Project-URL: Bug Tracker, https://github.com/cloverx-id/luminamoon/issues
14
+ Project-URL: Organization, https://github.com/cloverx-id
15
+ Keywords: optimizer,deep-learning,pytorch,vulkan,simd,triton,luminamoon
16
+ Classifier: Development Status :: 5 - Production/Stable
17
+ Classifier: Intended Audience :: Developers
18
+ Classifier: Intended Audience :: Science/Research
19
+ Classifier: Programming Language :: Python :: 3
20
+ Classifier: Programming Language :: Python :: 3.10
21
+ Classifier: Programming Language :: Python :: 3.11
22
+ Classifier: Programming Language :: Python :: 3.12
23
+ Classifier: Programming Language :: Python :: 3.13
24
+ Classifier: Programming Language :: Python :: 3.14
25
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
26
+ Requires-Python: >=3.10
27
+ Description-Content-Type: text/markdown
28
+ License-File: LICENSE
29
+ Requires-Dist: luminamoon>=1.0.5
30
+ Dynamic: license-file
31
+
32
+ # LuminaV Optimizer — Moved to `luminamoon` (PROJECT LUMI)
33
+
34
+ > **IMPORTANT MIGRATION NOTICE:**
35
+ > **LuminaV** has officially graduated into the unified **PROJECT LUMI (`luminamoon`)** monorepo and native Rust + Python distribution!
36
+ >
37
+ > Installing `pip install luminav` (`v1.5.3`) now **automatically installs `luminamoon>=1.0.5`** and forwards `from luminamoon import LuminaV` for 100% backward compatibility.
38
+ >
39
+ > - **New Official PyPI Package**: [**`pip install luminamoon`**](https://pypi.org/project/luminamoon/)
40
+ > - **Official GitHub Repository**: [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon)
41
+ > - **Changelogs (Software & Algorithm v1.5.3)**: [**`changelogs/`**](https://github.com/cloverx-id/luminamoon/tree/main/changelogs)
42
+
43
+ ---
44
+
45
+ ## Quick Migration Guide
46
+
47
+ Replace `luminav` with `luminamoon` in your environment and imports:
48
+
49
+ ```bash
50
+ pip install --upgrade luminamoon
51
+ ```
52
+
53
+ ```python
54
+ # Recommended new import (zero deprecation warning):
55
+ from luminamoon import LuminaV
56
+
57
+ # Legacy import (still works automatically via this bridge package):
58
+ from luminav import LuminaV
59
+ ```
60
+
61
+ ---
62
+
63
+ ## What's New in Canonical LuminaV v1.5.3 (`luminamoon`)
64
+
65
+ 1. **Multi-Backend Native Rust 2024 Execution Engine**:
66
+ - **CPU SIMD (`AVX-512` & `AVX2`)**: Parallelized via Rayon with zero-copy buffer protocol support for both PyTorch `Tensor` and standalone NumPy `ndarray`.
67
+ - **Vulkan 1.3/1.4 Compute Shaders (`SPIR-V` / UMA)**: Native GPU acceleration across AMD, Intel, and NVIDIA GPUs without requiring CUDA.
68
+ - **Triton AOT/JIT GPU Kernels**: Fused single-pass (`1P`) and two-pass (`2P`) kernels for CUDA and ROCm.
69
+ - **Free-Threaded Python (`abi3t` / No-GIL)**: Full support for Python 3.10 through 3.15t.
70
+ 2. **Dual-Buffer CLM (`buffer=2`) & Single-Buffer TD-CLM (`buffer=1`)**:
71
+ - **LuminaV-2 (`buffer=2` — Dual-Buffer CLM)**:
72
+ $$m_t = \beta_1 m_{t-1} + (1 - \beta_1) g_t, \quad v_t = \beta_2 v_{t-1} + (1 - \beta_2)(g_t - m_t)^2$$
73
+ $$\tilde{m}_t = \beta_1 m_t + (1 - \beta_1) g_t, \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
74
+ - **LuminaV-1 (`buffer=1` — Single-Buffer Temporal-Defect CLM)**:
75
+ Eliminates the second-moment buffer $v_t$ (saving 50% optimizer state memory) by modulating the RMS denominator with local temporal defect $d_t = g_t - m_t$:
76
+ $$\sigma_t = \tau \cdot \text{RMS}(\tilde{m}_t) \cdot \left(1 + \lambda \frac{|g_t - m_t|}{|m_t| + \epsilon}\right), \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
77
+ 3. **Formal Mathematical Verification (`LSO-Eval` with Z3 SMT Solver)**:
78
+ - Mathematically proven division-by-zero freedom, Padé $[3/3]$ rational $\tanh$ bounds, and hardware parity.
79
+
80
+ ---
81
+
82
+ ## Standard Instantiation (`LuminaV` v1.5.3)
83
+
84
+ ```python
85
+ import torch
86
+ from luminamoon import LuminaV
87
+
88
+ model = YourModel().to(device="cuda", dtype=torch.bfloat16)
89
+
90
+ optimizer = LuminaV(
91
+ model.parameters(),
92
+ lr=8e-4, # Learning rate
93
+ betas=(0.9, 0.999), # First moment & centered variance / defect decay
94
+ eps=1e-8, # Numerical stability floor
95
+ weight_decay=0.08, # Decoupled weight decay
96
+ tau=0.8, # Soft-projection temperature threshold
97
+ lam=0.3, # Temporal defect sensitivity (for buffer=1 TD-CLM)
98
+ cautious=True, # Sign-aligned Traffic-Cop directional gate
99
+ cautious_clamp_min=0.5, # Minimum cautious scaling factor (max 2.0x boost)
100
+ buffer=2, # 2 = Dual-Buffer CLM, 1 = Single-Buffer TD-CLM (50% less state RAM)
101
+ centered_defect=False, # Mean-centered defect option for buffer=1
102
+ stochastic_rounding=True, # Stateless SplitMix64 stochastic rounding for BF16/FP16
103
+ bound=True, # Adaptive trust-region step bounding
104
+ bound_type="coordinate", # "coordinate" (zero collective sync) or "radial"
105
+ bound_ratio=0.05, # Maximum step displacement ratio relative to parameter scale
106
+ master_weights="none", # "none" (Master-Free), "semi", or "full" (FP32)
107
+ execution="auto", # "auto", "triton", "vulkan", "simd", or "scalar"
108
+ fused_single_pass=True, # Single-pass fused kernel via lagged EMA scalars
109
+ ema_decay="auto", # "auto" (0.80 for buffer=2, 0.95 for buffer=1)
110
+ fsdp_sync=False, # Distributed FSDP / ZeRO-3 scratch synchronization
111
+ seed=1337, # Deterministic PRNG seed
112
+ )
113
+ ```
114
+
115
+ For full documentation, benchmarks, and the **LuminaMoon Data Engine**, visit [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon).
luminav-1.5.3.post1/README.md ADDED
@@ -0,0 +1,84 @@
1
+ # LuminaV Optimizer — Moved to `luminamoon` (PROJECT LUMI)
2
+
3
+ > **IMPORTANT MIGRATION NOTICE:**
4
+ > **LuminaV** has officially graduated into the unified **PROJECT LUMI (`luminamoon`)** monorepo and native Rust + Python distribution!
5
+ >
6
+ > Installing `pip install luminav` (`v1.5.3`) now **automatically installs `luminamoon>=1.0.5`** and forwards `from luminamoon import LuminaV` for 100% backward compatibility.
7
+ >
8
+ > - **New Official PyPI Package**: [**`pip install luminamoon`**](https://pypi.org/project/luminamoon/)
9
+ > - **Official GitHub Repository**: [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon)
10
+ > - **Changelogs (Software & Algorithm v1.5.3)**: [**`changelogs/`**](https://github.com/cloverx-id/luminamoon/tree/main/changelogs)
11
+
12
+ ---
13
+
14
+ ## Quick Migration Guide
15
+
16
+ Replace `luminav` with `luminamoon` in your environment and imports:
17
+
18
+ ```bash
19
+ pip install --upgrade luminamoon
20
+ ```
21
+
22
+ ```python
23
+ # Recommended new import (zero deprecation warning):
24
+ from luminamoon import LuminaV
25
+
26
+ # Legacy import (still works automatically via this bridge package):
27
+ from luminav import LuminaV
28
+ ```
29
+
30
+ ---
31
+
32
+ ## What's New in Canonical LuminaV v1.5.3 (`luminamoon`)
33
+
34
+ 1. **Multi-Backend Native Rust 2024 Execution Engine**:
35
+ - **CPU SIMD (`AVX-512` & `AVX2`)**: Parallelized via Rayon with zero-copy buffer protocol support for both PyTorch `Tensor` and standalone NumPy `ndarray`.
36
+ - **Vulkan 1.3/1.4 Compute Shaders (`SPIR-V` / UMA)**: Native GPU acceleration across AMD, Intel, and NVIDIA GPUs without requiring CUDA.
37
+ - **Triton AOT/JIT GPU Kernels**: Fused single-pass (`1P`) and two-pass (`2P`) kernels for CUDA and ROCm.
38
+ - **Free-Threaded Python (`abi3t` / No-GIL)**: Full support for Python 3.10 through 3.15t.
39
+ 2. **Dual-Buffer CLM (`buffer=2`) & Single-Buffer TD-CLM (`buffer=1`)**:
40
+ - **LuminaV-2 (`buffer=2` — Dual-Buffer CLM)**:
41
+ $$m_t = \beta_1 m_{t-1} + (1 - \beta_1) g_t, \quad v_t = \beta_2 v_{t-1} + (1 - \beta_2)(g_t - m_t)^2$$
42
+ $$\tilde{m}_t = \beta_1 m_t + (1 - \beta_1) g_t, \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
43
+ - **LuminaV-1 (`buffer=1` — Single-Buffer Temporal-Defect CLM)**:
44
+ Eliminates the second-moment buffer $v_t$ (saving 50% optimizer state memory) by modulating the RMS denominator with local temporal defect $d_t = g_t - m_t$:
45
+ $$\sigma_t = \tau \cdot \text{RMS}(\tilde{m}_t) \cdot \left(1 + \lambda \frac{|g_t - m_t|}{|m_t| + \epsilon}\right), \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
46
+ 3. **Formal Mathematical Verification (`LSO-Eval` with Z3 SMT Solver)**:
47
+ - Mathematically proven division-by-zero freedom, Padé $[3/3]$ rational $\tanh$ bounds, and hardware parity.
48
+
49
+ ---
50
+
51
+ ## Standard Instantiation (`LuminaV` v1.5.3)
52
+
53
+ ```python
54
+ import torch
55
+ from luminamoon import LuminaV
56
+
57
+ model = YourModel().to(device="cuda", dtype=torch.bfloat16)
58
+
59
+ optimizer = LuminaV(
60
+ model.parameters(),
61
+ lr=8e-4, # Learning rate
62
+ betas=(0.9, 0.999), # First moment & centered variance / defect decay
63
+ eps=1e-8, # Numerical stability floor
64
+ weight_decay=0.08, # Decoupled weight decay
65
+ tau=0.8, # Soft-projection temperature threshold
66
+ lam=0.3, # Temporal defect sensitivity (for buffer=1 TD-CLM)
67
+ cautious=True, # Sign-aligned Traffic-Cop directional gate
68
+ cautious_clamp_min=0.5, # Minimum cautious scaling factor (max 2.0x boost)
69
+ buffer=2, # 2 = Dual-Buffer CLM, 1 = Single-Buffer TD-CLM (50% less state RAM)
70
+ centered_defect=False, # Mean-centered defect option for buffer=1
71
+ stochastic_rounding=True, # Stateless SplitMix64 stochastic rounding for BF16/FP16
72
+ bound=True, # Adaptive trust-region step bounding
73
+ bound_type="coordinate", # "coordinate" (zero collective sync) or "radial"
74
+ bound_ratio=0.05, # Maximum step displacement ratio relative to parameter scale
75
+ master_weights="none", # "none" (Master-Free), "semi", or "full" (FP32)
76
+ execution="auto", # "auto", "triton", "vulkan", "simd", or "scalar"
77
+ fused_single_pass=True, # Single-pass fused kernel via lagged EMA scalars
78
+ ema_decay="auto", # "auto" (0.80 for buffer=2, 0.95 for buffer=1)
79
+ fsdp_sync=False, # Distributed FSDP / ZeRO-3 scratch synchronization
80
+ seed=1337, # Deterministic PRNG seed
81
+ )
82
+ ```
83
+
84
+ For full documentation, benchmarks, and the **LuminaMoon Data Engine**, visit [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon).
luminav-1.5.3.post1/pyproject.toml ADDED
@@ -0,0 +1,51 @@
1
+ [build-system]
2
+ requires = ["setuptools>=61.0"]
3
+ build-backend = "setuptools.build_meta"
4
+
5
+ [project]
6
+ name = "luminav"
7
+ version = "1.5.3.post1"
8
+ description = "LuminaV Optimizer (Moved to 'luminamoon' - Official PROJECT LUMI Package)"
9
+ readme = "README.md"
10
+ license = "Apache-2.0"
11
+ requires-python = ">=3.10"
12
+ authors = [
13
+ { name = "Silver Moon", email = "leafmoon.idn@gmail.com" },
14
+ { name = "Lumina Moon (cloverx-id) Contributors" },
15
+ ]
16
+ keywords = [
17
+ "optimizer",
18
+ "deep-learning",
19
+ "pytorch",
20
+ "vulkan",
21
+ "simd",
22
+ "triton",
23
+ "luminamoon",
24
+ ]
25
+ classifiers = [
26
+ "Development Status :: 5 - Production/Stable",
27
+ "Intended Audience :: Developers",
28
+ "Intended Audience :: Science/Research",
29
+ "Programming Language :: Python :: 3",
30
+ "Programming Language :: Python :: 3.10",
31
+ "Programming Language :: Python :: 3.11",
32
+ "Programming Language :: Python :: 3.12",
33
+ "Programming Language :: Python :: 3.13",
34
+ "Programming Language :: Python :: 3.14",
35
+ "Topic :: Scientific/Engineering :: Artificial Intelligence",
36
+ ]
37
+ dependencies = [
38
+ "luminamoon>=1.0.5",
39
+ ]
40
+
41
+ [project.urls]
42
+ Homepage = "https://github.com/cloverx-id/luminamoon"
43
+ Repository = "https://github.com/cloverx-id/luminamoon"
44
+ "Official PyPI (luminamoon)" = "https://pypi.org/project/luminamoon/"
45
+ Documentation = "https://github.com/cloverx-id/luminamoon#readme"
46
+ Changelog = "https://github.com/cloverx-id/luminamoon/tree/main/changelogs"
47
+ "Bug Tracker" = "https://github.com/cloverx-id/luminamoon/issues"
48
+ Organization = "https://github.com/cloverx-id"
49
+
50
+ [tool.setuptools.packages.find]
51
+ where = ["src"]
luminav-1.5.3.post1/setup.cfg ADDED
@@ -0,0 +1,4 @@
1
+ [egg_info]
2
+ tag_build =
3
+ tag_date = 0
4
+
luminav-1.5.3.post1/src/luminav/__init__.py ADDED
@@ -0,0 +1,25 @@
1
+ from __future__ import annotations
2
+
3
+ import warnings
4
+ from luminamoon import LuminaV, __version__ as _luminamoon_version
5
+ from luminamoon.optimizer.luminav import (
6
+ DEFAULT_LUMINAV_VERSION as __algorithm_version__,
7
+ )
8
+
9
+ __version__ = "1.5.3"
10
+ __luminamoon_version__ = _luminamoon_version
11
+
12
+ warnings.warn(
13
+ "The 'luminav' package has moved to the unified PROJECT LUMI package 'luminamoon' "
14
+ "(now accelerated with Rust AVX-512/AVX2 SIMD, Vulkan 1.3/1.4 UMA, and Triton AOT/JIT). "
15
+ "Please switch your imports to `from luminamoon import LuminaV` and install via `pip install luminamoon`.",
16
+ FutureWarning,
17
+ stacklevel=2,
18
+ )
19
+
20
+ __all__ = [
21
+ "LuminaV",
22
+ "__algorithm_version__",
23
+ "__luminamoon_version__",
24
+ "__version__",
25
+ ]
luminav-1.5.3.post1/src/luminav.egg-info/PKG-INFO ADDED
@@ -0,0 +1,115 @@
1
+ Metadata-Version: 2.4
2
+ Name: luminav
3
+ Version: 1.5.3.post1
4
+ Summary: LuminaV Optimizer (Moved to 'luminamoon' - Official PROJECT LUMI Package)
5
+ Author: Lumina Moon (cloverx-id) Contributors
6
+ Author-email: Silver Moon <leafmoon.idn@gmail.com>
7
+ License-Expression: Apache-2.0
8
+ Project-URL: Homepage, https://github.com/cloverx-id/luminamoon
9
+ Project-URL: Repository, https://github.com/cloverx-id/luminamoon
10
+ Project-URL: Official PyPI (luminamoon), https://pypi.org/project/luminamoon/
11
+ Project-URL: Documentation, https://github.com/cloverx-id/luminamoon#readme
12
+ Project-URL: Changelog, https://github.com/cloverx-id/luminamoon/tree/main/changelogs
13
+ Project-URL: Bug Tracker, https://github.com/cloverx-id/luminamoon/issues
14
+ Project-URL: Organization, https://github.com/cloverx-id
15
+ Keywords: optimizer,deep-learning,pytorch,vulkan,simd,triton,luminamoon
16
+ Classifier: Development Status :: 5 - Production/Stable
17
+ Classifier: Intended Audience :: Developers
18
+ Classifier: Intended Audience :: Science/Research
19
+ Classifier: Programming Language :: Python :: 3
20
+ Classifier: Programming Language :: Python :: 3.10
21
+ Classifier: Programming Language :: Python :: 3.11
22
+ Classifier: Programming Language :: Python :: 3.12
23
+ Classifier: Programming Language :: Python :: 3.13
24
+ Classifier: Programming Language :: Python :: 3.14
25
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
26
+ Requires-Python: >=3.10
27
+ Description-Content-Type: text/markdown
28
+ License-File: LICENSE
29
+ Requires-Dist: luminamoon>=1.0.5
30
+ Dynamic: license-file
31
+
32
+ # LuminaV Optimizer — Moved to `luminamoon` (PROJECT LUMI)
33
+
34
+ > **IMPORTANT MIGRATION NOTICE:**
35
+ > **LuminaV** has officially graduated into the unified **PROJECT LUMI (`luminamoon`)** monorepo and native Rust + Python distribution!
36
+ >
37
+ > Installing `pip install luminav` (`v1.5.3`) now **automatically installs `luminamoon>=1.0.5`** and forwards `from luminamoon import LuminaV` for 100% backward compatibility.
38
+ >
39
+ > - **New Official PyPI Package**: [**`pip install luminamoon`**](https://pypi.org/project/luminamoon/)
40
+ > - **Official GitHub Repository**: [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon)
41
+ > - **Changelogs (Software & Algorithm v1.5.3)**: [**`changelogs/`**](https://github.com/cloverx-id/luminamoon/tree/main/changelogs)
42
+
43
+ ---
44
+
45
+ ## Quick Migration Guide
46
+
47
+ Replace `luminav` with `luminamoon` in your environment and imports:
48
+
49
+ ```bash
50
+ pip install --upgrade luminamoon
51
+ ```
52
+
53
+ ```python
54
+ # Recommended new import (zero deprecation warning):
55
+ from luminamoon import LuminaV
56
+
57
+ # Legacy import (still works automatically via this bridge package):
58
+ from luminav import LuminaV
59
+ ```
60
+
61
+ ---
62
+
63
+ ## What's New in Canonical LuminaV v1.5.3 (`luminamoon`)
64
+
65
+ 1. **Multi-Backend Native Rust 2024 Execution Engine**:
66
+ - **CPU SIMD (`AVX-512` & `AVX2`)**: Parallelized via Rayon with zero-copy buffer protocol support for both PyTorch `Tensor` and standalone NumPy `ndarray`.
67
+ - **Vulkan 1.3/1.4 Compute Shaders (`SPIR-V` / UMA)**: Native GPU acceleration across AMD, Intel, and NVIDIA GPUs without requiring CUDA.
68
+ - **Triton AOT/JIT GPU Kernels**: Fused single-pass (`1P`) and two-pass (`2P`) kernels for CUDA and ROCm.
69
+ - **Free-Threaded Python (`abi3t` / No-GIL)**: Full support for Python 3.10 through 3.15t.
70
+ 2. **Dual-Buffer CLM (`buffer=2`) & Single-Buffer TD-CLM (`buffer=1`)**:
71
+ - **LuminaV-2 (`buffer=2` — Dual-Buffer CLM)**:
72
+ $$m_t = \beta_1 m_{t-1} + (1 - \beta_1) g_t, \quad v_t = \beta_2 v_{t-1} + (1 - \beta_2)(g_t - m_t)^2$$
73
+ $$\tilde{m}_t = \beta_1 m_t + (1 - \beta_1) g_t, \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
74
+ - **LuminaV-1 (`buffer=1` — Single-Buffer Temporal-Defect CLM)**:
75
+ Eliminates the second-moment buffer $v_t$ (saving 50% optimizer state memory) by modulating the RMS denominator with local temporal defect $d_t = g_t - m_t$:
76
+ $$\sigma_t = \tau \cdot \text{RMS}(\tilde{m}_t) \cdot \left(1 + \lambda \frac{|g_t - m_t|}{|m_t| + \epsilon}\right), \quad u_t = \tanh\left(\frac{\tilde{m}_t}{\sigma_t}\right)$$
77
+ 3. **Formal Mathematical Verification (`LSO-Eval` with Z3 SMT Solver)**:
78
+ - Mathematically proven division-by-zero freedom, Padé $[3/3]$ rational $\tanh$ bounds, and hardware parity.
79
+
80
+ ---
81
+
82
+ ## Standard Instantiation (`LuminaV` v1.5.3)
83
+
84
+ ```python
85
+ import torch
86
+ from luminamoon import LuminaV
87
+
88
+ model = YourModel().to(device="cuda", dtype=torch.bfloat16)
89
+
90
+ optimizer = LuminaV(
91
+ model.parameters(),
92
+ lr=8e-4, # Learning rate
93
+ betas=(0.9, 0.999), # First moment & centered variance / defect decay
94
+ eps=1e-8, # Numerical stability floor
95
+ weight_decay=0.08, # Decoupled weight decay
96
+ tau=0.8, # Soft-projection temperature threshold
97
+ lam=0.3, # Temporal defect sensitivity (for buffer=1 TD-CLM)
98
+ cautious=True, # Sign-aligned Traffic-Cop directional gate
99
+ cautious_clamp_min=0.5, # Minimum cautious scaling factor (max 2.0x boost)
100
+ buffer=2, # 2 = Dual-Buffer CLM, 1 = Single-Buffer TD-CLM (50% less state RAM)
101
+ centered_defect=False, # Mean-centered defect option for buffer=1
102
+ stochastic_rounding=True, # Stateless SplitMix64 stochastic rounding for BF16/FP16
103
+ bound=True, # Adaptive trust-region step bounding
104
+ bound_type="coordinate", # "coordinate" (zero collective sync) or "radial"
105
+ bound_ratio=0.05, # Maximum step displacement ratio relative to parameter scale
106
+ master_weights="none", # "none" (Master-Free), "semi", or "full" (FP32)
107
+ execution="auto", # "auto", "triton", "vulkan", "simd", or "scalar"
108
+ fused_single_pass=True, # Single-pass fused kernel via lagged EMA scalars
109
+ ema_decay="auto", # "auto" (0.80 for buffer=2, 0.95 for buffer=1)
110
+ fsdp_sync=False, # Distributed FSDP / ZeRO-3 scratch synchronization
111
+ seed=1337, # Deterministic PRNG seed
112
+ )
113
+ ```
114
+
115
+ For full documentation, benchmarks, and the **LuminaMoon Data Engine**, visit [**`https://github.com/cloverx-id/luminamoon`**](https://github.com/cloverx-id/luminamoon).
luminav-1.5.3.post1/src/luminav.egg-info/SOURCES.txt ADDED
@@ -0,0 +1,9 @@
1
+ LICENSE
2
+ README.md
3
+ pyproject.toml
4
+ src/luminav/__init__.py
5
+ src/luminav.egg-info/PKG-INFO
6
+ src/luminav.egg-info/SOURCES.txt
7
+ src/luminav.egg-info/dependency_links.txt
8
+ src/luminav.egg-info/requires.txt
9
+ src/luminav.egg-info/top_level.txt
luminav-1.5.3.post1/src/luminav.egg-info/requires.txt ADDED
@@ -0,0 +1 @@
1
+ luminamoon>=1.0.5
luminav-1.5.3.post1/src/luminav.egg-info/top_level.txt ADDED
@@ -0,0 +1 @@
1
+ luminav