wxsys commited on
Commit
b4d74bc
·
verified ·
1 Parent(s): c27551f

docs: add available model variants (Slim 1.3GB & Baseline 1.6GB)

Browse files
Files changed (1) hide show
  1. README.md +10 -1
README.md CHANGED
@@ -119,12 +119,21 @@ The generative projection head (`lm_head`, 311M parameters) has been removed. Ra
119
  | **Output Gating** | Headwise Sigmoid Projection (`g_proj`) |
120
  | **Context Window** | Up to 1,048,576 tokens via circular SWA buffer |
121
  | **Decision Mechanism** | Dual-Softmax Differential Cortex Head (`DiffCAMQPCortexHead`) |
122
- | **Format & Quantization** | Outlier-Preserved AWQ INT4 / W8A8 (1.6 GB binary on disk) |
123
  | **Resident Memory** | 4.08 MB VmRSS (C++ daemon via UNIX domain socket) |
124
  | **Inference Latency** | 0.044 ms per diff (CPU Intel AVX2 SIMD) |
125
 
126
  ---
127
 
 
 
 
 
 
 
 
 
 
128
  ## Core Capabilities
129
 
130
  ### 1. Single-Pass 3-Way Verdict Gate
 
119
  | **Output Gating** | Headwise Sigmoid Projection (`g_proj`) |
120
  | **Context Window** | Up to 1,048,576 tokens via circular SWA buffer |
121
  | **Decision Mechanism** | Dual-Softmax Differential Cortex Head (`DiffCAMQPCortexHead`) |
122
+ | **Format & Quantization** | Outlier-Preserved AWQ INT4 / W8A8 (1.3 GB Slim & 1.6 GB Baseline) |
123
  | **Resident Memory** | 4.08 MB VmRSS (C++ daemon via UNIX domain socket) |
124
  | **Inference Latency** | 0.044 ms per diff (CPU Intel AVX2 SIMD) |
125
 
126
  ---
127
 
128
+ ## Available Model Variants
129
+
130
+ | Variant | Path in Repository | Size on Disk | Embedding Precision | Verdict Parity vs Baseline | Description |
131
+ |---|---|---|---|---|---|
132
+ | **Spark-Servitor Slim (Recommended)** | `slim/spark-servitor-slim.servitor` | **1.3 GB** (~1,312 MB) | INT8 Quantized (Cosine: 0.999966) | **100% Identical** ($\Delta r^* = 0.0000$) | Recommended for production. Saves ~300 MB on disk while maintaining full 151k vocabulary and zero quality loss. |
133
+ | **Spark-Servitor Baseline** | `int4/spark-servitor-awq_int4.servitor` | **1.6 GB** (~1,608 MB) | FP16 High-Precision | Baseline Standard | Golden uncompressed-embedding checkpoint preserving full FP16 token lookup representations. |
134
+
135
+ ---
136
+
137
  ## Core Capabilities
138
 
139
  ### 1. Single-Pass 3-Way Verdict Gate