kgrabko commited on
Commit
b4b1b52
·
verified ·
1 Parent(s): 1514616

Upload claims.md

Browse files
Files changed (1) hide show
  1. claims.md +64 -0
claims.md ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Intellectual Property Claims & Patent Pending Notice
2
+
3
+ **Project:** CMS Manhattan JiRack
4
+ **Inventor:** Konstantin Vladimirovich Grabko
5
+ **Contact:** grabko@cmsmanhattan.com
6
+ **Status:** [PATENT PENDING] - Formal Claims Filed/Drafted December 21, 2025.
7
+
8
+ ---
9
+
10
+ ## ⚠️ NOTICE TO DEVELOPERS AND COMMERCIAL ENTITIES
11
+
12
+ The technologies, architectures, and methods disclosed in this repository are the proprietary intellectual property of Konstantin Vladimirovich Grabko. This document serves as a formal public record of the following claims to establish **Prior Art** and notice of **Patent Pending** status.
13
+
14
+ ---
15
+
16
+ ## I. Field of Invention
17
+
18
+ This invention pertains to **machine learning optimization**, specifically the **compression and hardware-acceleration of Transformer-based models** for non-NVIDIA (ROCm/HIP) environments.
19
+
20
+ ---
21
+
22
+ ## II. Core Intellectual Property Claims
23
+
24
+ ### 1. Ternary-Quantized Optimization
25
+ A method for reducing model VRAM footprint by quantizing weights into a ternary set $\{-1, 0, +1\}$ utilizing:
26
+ - A learnable scaling factor $\gamma$.
27
+ - A straight-through estimator (STE) to maintain model perplexity.
28
+ - Achieving up to 70% memory reduction.
29
+
30
+ ### 2. Buffered Routing Embedding (BRE)
31
+ A proprietary dynamic routing architecture that utilizes **shared memory pools** on High Bandwidth Memory (HBM). This claim covers:
32
+ - The specific **per-layer buffering logic** that minimizes redundant data movement between the GPU global memory and compute units.
33
+
34
+ ### 3. SwiGLU-Attention (SWA) Fusion
35
+ A novel fused compute kernel that integrates the **SwiGLU feed-forward network (FFN)** and **Multi-Head Attention (MHA)** into a single operational pass. This claim specifically covers:
36
+ - The reduction of activation memory overhead.
37
+ - The resulting thermal optimization (maintaining $<80^\circ\text{C}$).
38
+
39
+ ### 4. Hardware-Agnostic Inference Pipeline
40
+ The specific software stack and **asynchronous memory pooling routine** optimized for ROCm/HIP runtimes, enabling **high-throughput LLM performance on non-proprietary hardware**.
41
+
42
+ ---
43
+
44
+ ## III. Legal Restrictions & Usage
45
+
46
+ - **Non-Transferable:**
47
+ Access to this code does not constitute a transfer of ownership of the underlying inventions.
48
+
49
+ - **Anti-Patent Clause:**
50
+ Any party using this code is strictly prohibited from filing patent applications based on the **BRE**, **SWA**, or **Ternary-Quantized methods** described herein.
51
+
52
+ - **Commercial Licensing:**
53
+ Any commercial use (SaaS, hardware integration, etc.) requires a **signed execution of the CMS Manhattan JiRack License V.1.2**.
54
+
55
+ ---
56
+
57
+ ## IV. Contact for IP Inquiries
58
+
59
+ For patent licensing, joint venture opportunities, or freedom-to-operate inquiries, please contact:
60
+
61
+ **Konstantin Vladimirovich Grabko**
62
+ - **Email:** grabko@cmsmanhattan.com
63
+ - **Phone:** +1 (516) 777-0945
64
+ - **Location:** New York, USA