Update README.md
Browse files
README.md
CHANGED
|
@@ -16,24 +16,26 @@ tags:
|
|
| 16 |
|
| 17 |
NICoLE is a compact LLM-based controller for congestion-aware RTP/WebRTC adaptive video streaming.
|
| 18 |
|
| 19 |
-
It predicts:
|
| 20 |
- ECN
|
| 21 |
- Current Profile (CP)
|
| 22 |
- Next Profile (NP)
|
| 23 |
|
| 24 |
from RTP packetization and queue telemetry using compact symbolic prompting.
|
| 25 |
|
| 26 |
-
Optimized for:
|
| 27 |
- low-latency inference
|
| 28 |
- edge deployment
|
| 29 |
- GGUF quantization
|
| 30 |
- deterministic structured outputs
|
| 31 |
|
| 32 |
-
Applications:
|
|
|
|
| 33 |
- WebRTC adaptive streaming
|
| 34 |
- congestion-aware real-time video encoding adaptation
|
| 35 |
- in-Network QoE Optimization
|
| 36 |
- edge AI networking
|
|
|
|
| 37 |
---
|
| 38 |
|
| 39 |
# Profiles
|
|
@@ -50,21 +52,21 @@ The dataset was generated using real-time WebRTC streaming under a 40 Mbps bottl
|
|
| 50 |
|
| 51 |
# Prompt Format
|
| 52 |
|
| 53 |
-
Input order:
|
| 54 |
|
| 55 |
```text
|
| 56 |
|
| 57 |
PS FS IFGS IFGR CQ LQ E
|
| 58 |
```
|
| 59 |
|
| 60 |
-
Output order:
|
| 61 |
|
| 62 |
```text
|
| 63 |
|
| 64 |
E C N
|
| 65 |
```
|
| 66 |
|
| 67 |
-
Example:
|
| 68 |
|
| 69 |
```text
|
| 70 |
I:PS FS IFGS IFGR CQ LQ E
|
|
@@ -81,7 +83,7 @@ Expected output:
|
|
| 81 |
```
|
| 82 |
---
|
| 83 |
|
| 84 |
-
# Hugging Face Usage
|
| 85 |
|
| 86 |
```python
|
| 87 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
@@ -150,12 +152,12 @@ A:" \
|
|
| 150 |
| 16 | 1043 | 0.96 | 132.46 |
|
| 151 |
| 32 | 1432 | 0.70 | 104.29 |
|
| 152 |
|
| 153 |
-
Best CPU deployment:
|
| 154 |
- 4 threads
|
| 155 |
- 343 ms response time
|
| 156 |
- 2.91 decisions/sec
|
| 157 |
|
| 158 |
-
Compact symbolic prompting significantly reduces:
|
| 159 |
- prompt tokens
|
| 160 |
- KV-cache usage
|
| 161 |
- inference latency
|
|
@@ -169,10 +171,10 @@ compared to verbose natural-language prompting.
|
|
| 169 |
Available quantization:
|
| 170 |
- Q4_K_M (recommended)
|
| 171 |
|
| 172 |
-
Runtime:
|
| 173 |
- llama.cpp
|
| 174 |
|
| 175 |
-
Designed for:
|
| 176 |
- edge deployment
|
| 177 |
- CPU inference
|
| 178 |
- bounded symbolic control inference
|
|
@@ -192,5 +194,5 @@ Designed for:
|
|
| 192 |
|
| 193 |
If you use this model, please cite the NICoLE paper and repository.
|
| 194 |
|
| 195 |
-
- Alireza Shirmarz, Fabio Luciano Verdi, Gyanesh Patra, Gergely Pongracz,*"NICoLE: Are In-Network LLM-Based Agents Cost-Feasible for RTP Video Streaming?"*,
|
| 196 |
-
IEEE/IFIP Networking, Switzerland 2026.
|
|
|
|
| 16 |
|
| 17 |
NICoLE is a compact LLM-based controller for congestion-aware RTP/WebRTC adaptive video streaming.
|
| 18 |
|
| 19 |
+
It **predicts**:
|
| 20 |
- ECN
|
| 21 |
- Current Profile (CP)
|
| 22 |
- Next Profile (NP)
|
| 23 |
|
| 24 |
from RTP packetization and queue telemetry using compact symbolic prompting.
|
| 25 |
|
| 26 |
+
**Optimized** for:
|
| 27 |
- low-latency inference
|
| 28 |
- edge deployment
|
| 29 |
- GGUF quantization
|
| 30 |
- deterministic structured outputs
|
| 31 |
|
| 32 |
+
**Applications**:
|
| 33 |
+
|
| 34 |
- WebRTC adaptive streaming
|
| 35 |
- congestion-aware real-time video encoding adaptation
|
| 36 |
- in-Network QoE Optimization
|
| 37 |
- edge AI networking
|
| 38 |
+
|
| 39 |
---
|
| 40 |
|
| 41 |
# Profiles
|
|
|
|
| 52 |
|
| 53 |
# Prompt Format
|
| 54 |
|
| 55 |
+
**Input order**:
|
| 56 |
|
| 57 |
```text
|
| 58 |
|
| 59 |
PS FS IFGS IFGR CQ LQ E
|
| 60 |
```
|
| 61 |
|
| 62 |
+
**Output order**:
|
| 63 |
|
| 64 |
```text
|
| 65 |
|
| 66 |
E C N
|
| 67 |
```
|
| 68 |
|
| 69 |
+
**Example**:
|
| 70 |
|
| 71 |
```text
|
| 72 |
I:PS FS IFGS IFGR CQ LQ E
|
|
|
|
| 83 |
```
|
| 84 |
---
|
| 85 |
|
| 86 |
+
# Hugging Face Usage (Python code)
|
| 87 |
|
| 88 |
```python
|
| 89 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
| 152 |
| 16 | 1043 | 0.96 | 132.46 |
|
| 153 |
| 32 | 1432 | 0.70 | 104.29 |
|
| 154 |
|
| 155 |
+
**Best CPU deployment:**
|
| 156 |
- 4 threads
|
| 157 |
- 343 ms response time
|
| 158 |
- 2.91 decisions/sec
|
| 159 |
|
| 160 |
+
## Compact symbolic prompting significantly reduces:
|
| 161 |
- prompt tokens
|
| 162 |
- KV-cache usage
|
| 163 |
- inference latency
|
|
|
|
| 171 |
Available quantization:
|
| 172 |
- Q4_K_M (recommended)
|
| 173 |
|
| 174 |
+
**Runtime**:
|
| 175 |
- llama.cpp
|
| 176 |
|
| 177 |
+
**Designed** for:
|
| 178 |
- edge deployment
|
| 179 |
- CPU inference
|
| 180 |
- bounded symbolic control inference
|
|
|
|
| 194 |
|
| 195 |
If you use this model, please cite the NICoLE paper and repository.
|
| 196 |
|
| 197 |
+
- **Alireza Shirmarz, Fabio Luciano Verdi, Gyanesh Patra, Gergely Pongracz,*"NICoLE: Are In-Network LLM-Based Agents Cost-Feasible for RTP Video Streaming?"*,
|
| 198 |
+
IEEE/IFIP Networking, Switzerland 2026.**
|