CMSManhattan
/

JiRackTernary_70b

Model card Files Files and versions

kgrabko commited on Dec 22, 2025

Commit

3af61d4

·

verified ·

1 Parent(s): 7ac032e

Update LICENSE

Files changed (1) hide show

LICENSE +6 -6

LICENSE CHANGED Viewed

@@ -31,12 +31,12 @@ Licensor grants a **non-exclusive, worldwide, royalty-bearing license** to imple
 ### 3. Key Technical Components
-| **Component**              | **Description**                                                                 |
-|-----------------------------|---------------------------------------------------------------------------------|
-| **Ternary Quantization**    | Weights restricted to $W \in \{-1, 0, +1\}$ with learnable scaling ($\gamma$).  |
-| **Buffered Routing Embedding (BRE)** | Dynamic routing logic maximizing HBM-specific memory pooling.                      |
-| **SwiGLU-Attention (SWA) Fusion** | Unified compute kernel merging FFN and MHA for efficiency.                            |
-| **Hardware Optimization**   | Optimization stack specifically for ROCm/HIP and non-NVIDIA hardware.           |
 ---

 ### 3. Key Technical Components
+| **Component**                        | **Description**                                                                 |
+|--------------------------------------|---------------------------------------------------------------------------------|
+| **Ternary Quantization**             | Weights restricted to $W \in \{-1, 0, +1\}$ with learnable scaling ($\gamma$).  |
+| **Buffered Routing Embedding (BRE)** | Dynamic routing logic maximizing HBM-specific memory pooling.                   |
+| **SwiGLU-Attention (SWA) Fusion**    | Unified compute kernel merging FFN and MHA for efficiency.                      |
+| **Hardware Optimization**            | Optimization stack specifically for ROCm/HIP and non-NVIDIA hardware.           |
 ---