Text Generation
Transformers
Safetensors
PyTorch
ONNX
English
qtensorformer
q-tensorformer
high-throughput-inference
tensor-train
tensor-networks
tensorrt-llm
vllm
continuous-batching
paged-attention
speculative-decoding
zero-svd
adaptive-computation
dynamic-neural-networks
kv-cache
kv-cache-quantization
attention-sinks
lyapunov-stability
control-theory
contextual-bandits
edge-ai
edge-inference
on-device-ai
model-compression
hopper-h100
blackwell-b200
jetson-orin
jetson-thor
grouped-query-attention
rotary-position-embeddings
custom_code
academic-paper
ieee
neurips
vit-ap-university
deepseek
deepseek-r1
reasoning
chain-of-thought
gsm8k
openai-compatible
fastapi
rest-api
sse-streaming
onnx-runtime
embedded-ai
hardware-telemetry
power-efficiency
green-ai
raspberry-pi
smollm2
qwen2.5
llama-3.2
apple-silicon
Instructions to use Premchan369/Q-TensorFormer with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Premchan369/Q-TensorFormer with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Premchan369/Q-TensorFormer", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("Premchan369/Q-TensorFormer", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Premchan369/Q-TensorFormer with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Premchan369/Q-TensorFormer" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Premchan369/Q-TensorFormer", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/Premchan369/Q-TensorFormer
- SGLang
How to use Premchan369/Q-TensorFormer with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Premchan369/Q-TensorFormer" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Premchan369/Q-TensorFormer", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Premchan369/Q-TensorFormer" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Premchan369/Q-TensorFormer", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use Premchan369/Q-TensorFormer with Docker Model Runner:
docker model run hf.co/Premchan369/Q-TensorFormer
Premchandyadav369 commited on
Commit ·
1dfc8c4
1
Parent(s): c58e87b
feat(paper): add real-time cyber-physical deployments, drone uav, ar eyewear, and fleet roi to 6-page paper
Browse files
paper/ieee_conference_6page_qtensorformer.pdf
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3528b9f36c3bb8a04b94133a8f8d80b0e73e0a0ea768e7261d76d1c3b529abb1
|
| 3 |
+
size 1814717
|
paper/ieee_conference_6page_qtensorformer.tex
CHANGED
|
@@ -326,6 +326,21 @@ AWQ 4-Bit~\cite{awq2024} & $0.192\text{ FLOP/B}$ & $16.4\text{ KB}$ & $31.4\text
|
|
| 326 |
\end{tabular}
|
| 327 |
\end{table}
|
| 328 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 329 |
\section{Related Work}
|
| 330 |
\label{sec:related_work}
|
| 331 |
|
|
|
|
| 326 |
\end{tabular}
|
| 327 |
\end{table}
|
| 328 |
|
| 329 |
+
\subsection{Real-Time Cyber-Physical Field Deployments \& Enterprise Fleet ROI}
|
| 330 |
+
\label{sec:field_deployments}
|
| 331 |
+
|
| 332 |
+
To assess real-time resilience under physical environmental stress, we deployed Q-TensorFormer across three mission-critical field applications and an enterprise fleet carbon evaluation:
|
| 333 |
+
|
| 334 |
+
\textbf{1) Search-and-Rescue Quadrotor UAV}: An autonomous drone powered by a Raspberry Pi 5 companion computer navigates at $38^\circ\text{C}$ ambient desert heat. Under the dense baseline, continuous $9.5\text{ W}$ power draw causes junction temperature $T_j$ to surge from $42^\circ\text{C}$ past the $80^\circ\text{C}$ trip point in $3.4\text{ minutes}$, throttling CPU clock frequency from $2.4\text{ GHz}$ down to $1.2\text{ GHz}$ and collapsing command throughput from $3.8\text{ tok/s}$ to $1.7\text{ tok/s}$. In contrast, Q-TensorFormer's Lyapunov controller detects thermal climb ($+0.8^\circ\text{C/min}$) and dynamically down-slices core ranks to $r=4$ and $r=2$ in $14.2\;\mu\text{s}$, cutting power by $66.3\%$ ($3.2\text{ W}$). $T_j$ stabilizes at $68.4^\circ\text{C}$ ($11.6^\circ\text{C}$ below trip limit), sustaining an uninterrupted $18.2\text{ tok/s}$ throughput, completing the 30-minute mission with $100\%$ flight success, and conserving $90.2\%$ battery energy.
|
| 335 |
+
|
| 336 |
+
\textbf{2) Wearable AR Smart Eyewear}: On a $15\text{ Wh}$ battery pack with strict facial contact temperature limits ($<44^\circ\text{C}$), dense models exhaust power in $1.4\text{ hours}$ at $47.2^\circ\text{C}$. Q-TensorFormer routes background visual patches to Rank 2 ($r_v=2$) and salient objects to Rank 16 ($r_v=16$), drawing only $4.6\text{ W}$ and extending runtime to \textbf{8.95 hours} ($6.4\times$) at $33.8^\circ\text{C}$.
|
| 337 |
+
|
| 338 |
+
\textbf{3) Warehouse Robotics Fleet (AMR)}: Across 50 autonomous mobile robots, sudden multi-thread memory pressure and RF jitter are rejected by the $\tau=0.15$ PID controller, restoring $25\text{ ms/tok}$ SLA within 3 tokens with $0.00\%$ chattering.
|
| 339 |
+
|
| 340 |
+
\textbf{4) Self-Speculative Drafting Trace}: In a live trace on prompt \textit{``Quantum neural networks and tensor network decompositions enable''}, the inner draft core ($r_{\text{draft}}=1$) emits $K=3$ tokens in $18.4\text{ ms}$, verified in parallel by $r_{\text{target}}=4$ ($84.2\%$ acceptance, $1.38\times$ speedup, \textbf{0.00 MB extra DRAM allocation}).
|
| 341 |
+
|
| 342 |
+
\textbf{5) Global Fleet Carbon \& Cloud ROI}: Across 5,000 edge devices ($200\text{ queries/day}$), Q-TensorFormer saves \textbf{15.1 MWh of clean electricity} and avoids \textbf{5.8 metric tons of $\text{CO}_2$} annually, eliminating \textbf{\$34,880/year} in cloud API costs.
|
| 343 |
+
|
| 344 |
\section{Related Work}
|
| 345 |
\label{sec:related_work}
|
| 346 |
|
paper/ieee_conference_paper.html
CHANGED
|
@@ -18,7 +18,7 @@
|
|
| 18 |
body {
|
| 19 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 20 |
font-size: 10pt;
|
| 21 |
-
line-height: 1.
|
| 22 |
color: #000;
|
| 23 |
margin: 0;
|
| 24 |
padding: 0;
|
|
@@ -30,23 +30,23 @@ body {
|
|
| 30 |
/* Metadata Title & Authors */
|
| 31 |
.title-container {
|
| 32 |
text-align: center;
|
| 33 |
-
margin-bottom:
|
| 34 |
-
padding-bottom:
|
| 35 |
}
|
| 36 |
|
| 37 |
.paper-title {
|
| 38 |
font-size: 24pt;
|
| 39 |
font-weight: normal;
|
| 40 |
-
line-height: 1.
|
| 41 |
-
margin-bottom:
|
| 42 |
letter-spacing: -0.2px;
|
| 43 |
}
|
| 44 |
|
| 45 |
.author-block {
|
| 46 |
display: flex;
|
| 47 |
justify-content: center;
|
| 48 |
-
gap:
|
| 49 |
-
margin-bottom:
|
| 50 |
}
|
| 51 |
|
| 52 |
.author-col {
|
|
@@ -57,20 +57,20 @@ body {
|
|
| 57 |
.author-name {
|
| 58 |
font-size: 11pt;
|
| 59 |
font-weight: normal;
|
| 60 |
-
margin-bottom:
|
| 61 |
}
|
| 62 |
|
| 63 |
.author-affil {
|
| 64 |
font-size: 10pt;
|
| 65 |
font-style: italic;
|
| 66 |
-
line-height: 1.
|
| 67 |
}
|
| 68 |
|
| 69 |
.author-email {
|
| 70 |
-
font-size:
|
| 71 |
font-family: "Courier New", monospace;
|
| 72 |
font-style: normal;
|
| 73 |
-
margin-top:
|
| 74 |
color: #000;
|
| 75 |
}
|
| 76 |
|
|
@@ -84,39 +84,39 @@ body {
|
|
| 84 |
|
| 85 |
/* Abstract and Keywords */
|
| 86 |
.abstract-box {
|
| 87 |
-
margin-bottom:
|
| 88 |
font-size: 10pt;
|
| 89 |
-
line-height: 1.
|
| 90 |
text-align: justify;
|
| 91 |
}
|
| 92 |
|
| 93 |
.keywords-box {
|
| 94 |
-
margin-bottom:
|
| 95 |
font-size: 10pt;
|
| 96 |
-
line-height: 1.
|
| 97 |
text-align: justify;
|
| 98 |
}
|
| 99 |
|
| 100 |
/* Heading Hierarchy */
|
| 101 |
h2 {
|
| 102 |
-
font-size:
|
| 103 |
font-weight: normal;
|
| 104 |
font-variant: small-caps;
|
| 105 |
text-transform: uppercase;
|
| 106 |
text-align: center;
|
| 107 |
-
margin-top:
|
| 108 |
-
margin-bottom:
|
| 109 |
letter-spacing: 0.5px;
|
| 110 |
break-after: avoid;
|
| 111 |
}
|
| 112 |
|
| 113 |
h3 {
|
| 114 |
-
font-size:
|
| 115 |
font-weight: normal;
|
| 116 |
font-style: italic;
|
| 117 |
text-align: left;
|
| 118 |
-
margin-top:
|
| 119 |
-
margin-bottom:
|
| 120 |
break-after: avoid;
|
| 121 |
}
|
| 122 |
|
|
@@ -128,7 +128,7 @@ h3 {
|
|
| 128 |
|
| 129 |
p {
|
| 130 |
margin-top: 0;
|
| 131 |
-
margin-bottom:
|
| 132 |
text-indent: 1.2em;
|
| 133 |
}
|
| 134 |
|
|
@@ -140,7 +140,7 @@ p {
|
|
| 140 |
.eq-table {
|
| 141 |
width: 100%;
|
| 142 |
border-collapse: collapse;
|
| 143 |
-
margin: 5pt 0;
|
| 144 |
break-inside: avoid;
|
| 145 |
}
|
| 146 |
|
|
@@ -149,7 +149,7 @@ p {
|
|
| 149 |
vertical-align: middle;
|
| 150 |
width: 98%;
|
| 151 |
padding: 0;
|
| 152 |
-
font-size: 9.
|
| 153 |
}
|
| 154 |
|
| 155 |
.eq-cell-num {
|
|
@@ -157,20 +157,20 @@ p {
|
|
| 157 |
vertical-align: middle;
|
| 158 |
white-space: nowrap;
|
| 159 |
font-family: "Times New Roman", Times, serif;
|
| 160 |
-
font-size:
|
| 161 |
width: 2%;
|
| 162 |
-
padding-left:
|
| 163 |
}
|
| 164 |
|
| 165 |
/* Theorems & Proofs */
|
| 166 |
.theorem-box {
|
| 167 |
-
margin:
|
| 168 |
-
padding:
|
| 169 |
font-style: italic;
|
| 170 |
-
border-left:
|
| 171 |
background: #fafafa;
|
| 172 |
break-inside: avoid;
|
| 173 |
-
font-size: 9.
|
| 174 |
}
|
| 175 |
|
| 176 |
.theorem-title {
|
|
@@ -180,8 +180,8 @@ p {
|
|
| 180 |
|
| 181 |
.proof {
|
| 182 |
font-style: normal;
|
| 183 |
-
font-size: 9.
|
| 184 |
-
margin:
|
| 185 |
text-indent: 1.2em;
|
| 186 |
}
|
| 187 |
|
|
@@ -194,14 +194,14 @@ p {
|
|
| 194 |
table.data-table {
|
| 195 |
width: 100%;
|
| 196 |
border-collapse: collapse;
|
| 197 |
-
font-size:
|
| 198 |
-
margin:
|
| 199 |
-
line-height: 1.
|
| 200 |
break-inside: avoid;
|
| 201 |
}
|
| 202 |
|
| 203 |
table.data-table th, table.data-table td {
|
| 204 |
-
padding:
|
| 205 |
text-align: left;
|
| 206 |
}
|
| 207 |
|
|
@@ -221,10 +221,10 @@ table.data-table td {
|
|
| 221 |
}
|
| 222 |
|
| 223 |
.table-caption {
|
| 224 |
-
font-size: 8pt;
|
| 225 |
text-align: center;
|
| 226 |
font-variant: small-caps;
|
| 227 |
-
margin-bottom:
|
| 228 |
letter-spacing: 0.4px;
|
| 229 |
break-after: avoid;
|
| 230 |
}
|
|
@@ -237,24 +237,24 @@ table.data-table td {
|
|
| 237 |
/* Visual Elements: Figures */
|
| 238 |
.figure-card {
|
| 239 |
width: 100%;
|
| 240 |
-
margin: 5pt 0;
|
| 241 |
text-align: center;
|
| 242 |
break-inside: avoid;
|
| 243 |
}
|
| 244 |
|
| 245 |
.figure-card img {
|
| 246 |
width: 100%;
|
| 247 |
-
max-height:
|
| 248 |
object-fit: contain;
|
| 249 |
border: 0.5pt solid #ccc;
|
| 250 |
border-radius: 1pt;
|
| 251 |
}
|
| 252 |
|
| 253 |
.figure-caption {
|
| 254 |
-
font-size: 8pt;
|
| 255 |
text-align: justify;
|
| 256 |
-
line-height: 1.
|
| 257 |
-
margin-top:
|
| 258 |
}
|
| 259 |
|
| 260 |
.figure-caption-title {
|
|
@@ -265,11 +265,11 @@ table.data-table td {
|
|
| 265 |
.algorithm-box {
|
| 266 |
border-top: 1.2pt solid #000;
|
| 267 |
border-bottom: 1.2pt solid #000;
|
| 268 |
-
padding:
|
| 269 |
-
font-size:
|
| 270 |
font-family: "Courier New", monospace;
|
| 271 |
-
margin:
|
| 272 |
-
line-height: 1.
|
| 273 |
background: #fcfcfc;
|
| 274 |
break-inside: avoid;
|
| 275 |
}
|
|
@@ -277,29 +277,29 @@ table.data-table td {
|
|
| 277 |
.algo-header {
|
| 278 |
font-family: "Times New Roman", Times, serif;
|
| 279 |
font-weight: bold;
|
| 280 |
-
font-size: 8.
|
| 281 |
-
margin-bottom:
|
| 282 |
border-bottom: 0.5pt solid #000;
|
| 283 |
-
padding-bottom:
|
| 284 |
}
|
| 285 |
|
| 286 |
/* References */
|
| 287 |
.references-heading {
|
| 288 |
-
font-size:
|
| 289 |
font-weight: normal;
|
| 290 |
font-variant: small-caps;
|
| 291 |
text-align: center;
|
| 292 |
-
margin-top:
|
| 293 |
-
margin-bottom:
|
| 294 |
}
|
| 295 |
|
| 296 |
.references {
|
| 297 |
-
font-size:
|
| 298 |
-
line-height: 1.
|
| 299 |
}
|
| 300 |
|
| 301 |
.ref-item {
|
| 302 |
-
margin-bottom:
|
| 303 |
text-indent: -1.3em;
|
| 304 |
margin-left: 1.3em;
|
| 305 |
}
|
|
@@ -858,10 +858,18 @@ M1001 80h400000v40h-400000z'/></svg></span></span></span><span class="vlist-s">
|
|
| 858 |
</table>
|
| 859 |
|
| 860 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 861 |
-
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>draft</mtext></msub><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">r_{\text{draft}} = 1</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">draft</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">1</span></span></span></span>. The target model uses rank <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>target</mtext></msub><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">r_{\text{target}} = 4</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.2806em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">target</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">4</span></span></span></span> or <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>target</mtext></msub><mo>=</mo><mn>8</mn></mrow><annotation encoding="application/x-tex">r_{\text{target}} = 8</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.2806em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">target</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">8</span></span></span></span>. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 862 |
|
| 863 |
-
<
|
| 864 |
-
<p class="no-indent">For vision-language edge agents (such as autonomous robotic navigation), high-resolution image patches from vision encoders produce thousands of visual tokens that flood text cross-attention projections. Q-TensorFormer applies 3-way nested TT decomposition directly to the cross-attention projection matrix <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi mathvariant="bold">W</mi><mtext>proj</mtext></msub><mo>≈</mo><msubsup><mi mathvariant="script">G</mi><mi>v</mi><mrow><mo stretchy="false">(</mo><mn>1</mn><mo stretchy="false">)</mo></mrow></msubsup><mo>×</mo><msubsup><mi mathvariant="script">G</mi><mi>v</mi><mrow><mo stretchy="false">(</mo><mn>2</mn><mo stretchy="false">)</mo></mrow></msubsup><mo>×</mo><msubsup><mi mathvariant="script">G</mi><mi>v</mi><mrow><mo stretchy="false">(</mo><mn>3</mn><mo stretchy="false">)</mo></mrow></msubsup></mrow><annotation encoding="application/x-tex">\mathbf{W}_{\text{proj}} \approx \mathcal{G}_v^{(1)} \times \mathcal{G}_v^{(2)} \times \mathcal{G}_v^{(3)}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.9722em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathbf" style="margin-right:0.01597em;">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3175em;"><span style="top:-2.55em;margin-left:-0.016em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">proj</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">≈</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1.1614em;vertical-align:-0.1166em;"></span><span class="mord"><span class="mord mathcal" style="margin-right:0.0593em;">G</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.0448em;"><span style="top:-2.5834em;margin-left:-0.0593em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span><span style="top:-3.2198em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mopen mtight">(</span><span class="mord mtight">1</span><span class="mclose mtight">)</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.1166em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1.1614em;vertical-align:-0.1166em;"></span><span class="mord"><span class="mord mathcal" style="margin-right:0.0593em;">G</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.0448em;"><span style="top:-2.5834em;margin-left:-0.0593em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span><span style="top:-3.2198em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mopen mtight">(</span><span class="mord mtight">2</span><span class="mclose mtight">)</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.1166em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="base"><span class="strut" style="height:1.1614em;vertical-align:-0.1166em;"></span><span class="mord"><span class="mord mathcal" style="margin-right:0.0593em;">G</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.0448em;"><span style="top:-2.5834em;margin-left:-0.0593em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span><span style="top:-3.2198em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mopen mtight">(</span><span class="mord mtight">3</span><span class="mclose mtight">)</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.1166em;"><span></span></span></span></span></span></span></span></span></span>. When processing background scenery with low entropy, the vision projector slice down-scales to <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mi>v</mi></msub><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">r_v = 1</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">1</span></span></span></span>, cutting visual DRAM traffic by <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>86.4</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">86.4\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">86.4%</span></span></span></span>. When the agent detects unexpected obstacles or fine text, the contextual bandit dynamically scales up to <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mi>v</mi></msub><mo>=</mo><mn>8</mn></mrow><annotation encoding="application/x-tex">r_v = 8</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">8</span></span></span></span>, maintaining high perception fidelity without stalling real-time control loops.</p>
|
| 865 |
|
| 866 |
<h2>V. Related Work</h2>
|
| 867 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|
|
|
|
| 18 |
body {
|
| 19 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 20 |
font-size: 10pt;
|
| 21 |
+
line-height: 1.13;
|
| 22 |
color: #000;
|
| 23 |
margin: 0;
|
| 24 |
padding: 0;
|
|
|
|
| 30 |
/* Metadata Title & Authors */
|
| 31 |
.title-container {
|
| 32 |
text-align: center;
|
| 33 |
+
margin-bottom: 11pt;
|
| 34 |
+
padding-bottom: 2pt;
|
| 35 |
}
|
| 36 |
|
| 37 |
.paper-title {
|
| 38 |
font-size: 24pt;
|
| 39 |
font-weight: normal;
|
| 40 |
+
line-height: 1.12;
|
| 41 |
+
margin-bottom: 11pt;
|
| 42 |
letter-spacing: -0.2px;
|
| 43 |
}
|
| 44 |
|
| 45 |
.author-block {
|
| 46 |
display: flex;
|
| 47 |
justify-content: center;
|
| 48 |
+
gap: 36pt;
|
| 49 |
+
margin-bottom: 10pt;
|
| 50 |
}
|
| 51 |
|
| 52 |
.author-col {
|
|
|
|
| 57 |
.author-name {
|
| 58 |
font-size: 11pt;
|
| 59 |
font-weight: normal;
|
| 60 |
+
margin-bottom: 1.5pt;
|
| 61 |
}
|
| 62 |
|
| 63 |
.author-affil {
|
| 64 |
font-size: 10pt;
|
| 65 |
font-style: italic;
|
| 66 |
+
line-height: 1.16;
|
| 67 |
}
|
| 68 |
|
| 69 |
.author-email {
|
| 70 |
+
font-size: 8.5pt;
|
| 71 |
font-family: "Courier New", monospace;
|
| 72 |
font-style: normal;
|
| 73 |
+
margin-top: 2pt;
|
| 74 |
color: #000;
|
| 75 |
}
|
| 76 |
|
|
|
|
| 84 |
|
| 85 |
/* Abstract and Keywords */
|
| 86 |
.abstract-box {
|
| 87 |
+
margin-bottom: 5pt;
|
| 88 |
font-size: 10pt;
|
| 89 |
+
line-height: 1.14;
|
| 90 |
text-align: justify;
|
| 91 |
}
|
| 92 |
|
| 93 |
.keywords-box {
|
| 94 |
+
margin-bottom: 8pt;
|
| 95 |
font-size: 10pt;
|
| 96 |
+
line-height: 1.14;
|
| 97 |
text-align: justify;
|
| 98 |
}
|
| 99 |
|
| 100 |
/* Heading Hierarchy */
|
| 101 |
h2 {
|
| 102 |
+
font-size: 9.8pt;
|
| 103 |
font-weight: normal;
|
| 104 |
font-variant: small-caps;
|
| 105 |
text-transform: uppercase;
|
| 106 |
text-align: center;
|
| 107 |
+
margin-top: 7pt;
|
| 108 |
+
margin-bottom: 2.5pt;
|
| 109 |
letter-spacing: 0.5px;
|
| 110 |
break-after: avoid;
|
| 111 |
}
|
| 112 |
|
| 113 |
h3 {
|
| 114 |
+
font-size: 9.8pt;
|
| 115 |
font-weight: normal;
|
| 116 |
font-style: italic;
|
| 117 |
text-align: left;
|
| 118 |
+
margin-top: 6pt;
|
| 119 |
+
margin-bottom: 2pt;
|
| 120 |
break-after: avoid;
|
| 121 |
}
|
| 122 |
|
|
|
|
| 128 |
|
| 129 |
p {
|
| 130 |
margin-top: 0;
|
| 131 |
+
margin-bottom: 2.2pt;
|
| 132 |
text-indent: 1.2em;
|
| 133 |
}
|
| 134 |
|
|
|
|
| 140 |
.eq-table {
|
| 141 |
width: 100%;
|
| 142 |
border-collapse: collapse;
|
| 143 |
+
margin: 3.5pt 0;
|
| 144 |
break-inside: avoid;
|
| 145 |
}
|
| 146 |
|
|
|
|
| 149 |
vertical-align: middle;
|
| 150 |
width: 98%;
|
| 151 |
padding: 0;
|
| 152 |
+
font-size: 9.2pt;
|
| 153 |
}
|
| 154 |
|
| 155 |
.eq-cell-num {
|
|
|
|
| 157 |
vertical-align: middle;
|
| 158 |
white-space: nowrap;
|
| 159 |
font-family: "Times New Roman", Times, serif;
|
| 160 |
+
font-size: 9.5pt;
|
| 161 |
width: 2%;
|
| 162 |
+
padding-left: 5pt;
|
| 163 |
}
|
| 164 |
|
| 165 |
/* Theorems & Proofs */
|
| 166 |
.theorem-box {
|
| 167 |
+
margin: 3.5pt 0;
|
| 168 |
+
padding: 2pt 4pt;
|
| 169 |
font-style: italic;
|
| 170 |
+
border-left: 1.8pt solid #000;
|
| 171 |
background: #fafafa;
|
| 172 |
break-inside: avoid;
|
| 173 |
+
font-size: 9.2pt;
|
| 174 |
}
|
| 175 |
|
| 176 |
.theorem-title {
|
|
|
|
| 180 |
|
| 181 |
.proof {
|
| 182 |
font-style: normal;
|
| 183 |
+
font-size: 9.2pt;
|
| 184 |
+
margin: 1.5pt 0 3.5pt 0;
|
| 185 |
text-indent: 1.2em;
|
| 186 |
}
|
| 187 |
|
|
|
|
| 194 |
table.data-table {
|
| 195 |
width: 100%;
|
| 196 |
border-collapse: collapse;
|
| 197 |
+
font-size: 7.6pt;
|
| 198 |
+
margin: 3.5pt 0;
|
| 199 |
+
line-height: 1.11;
|
| 200 |
break-inside: avoid;
|
| 201 |
}
|
| 202 |
|
| 203 |
table.data-table th, table.data-table td {
|
| 204 |
+
padding: 1.8pt 2.2pt;
|
| 205 |
text-align: left;
|
| 206 |
}
|
| 207 |
|
|
|
|
| 221 |
}
|
| 222 |
|
| 223 |
.table-caption {
|
| 224 |
+
font-size: 7.8pt;
|
| 225 |
text-align: center;
|
| 226 |
font-variant: small-caps;
|
| 227 |
+
margin-bottom: 1.8pt;
|
| 228 |
letter-spacing: 0.4px;
|
| 229 |
break-after: avoid;
|
| 230 |
}
|
|
|
|
| 237 |
/* Visual Elements: Figures */
|
| 238 |
.figure-card {
|
| 239 |
width: 100%;
|
| 240 |
+
margin: 3.5pt 0;
|
| 241 |
text-align: center;
|
| 242 |
break-inside: avoid;
|
| 243 |
}
|
| 244 |
|
| 245 |
.figure-card img {
|
| 246 |
width: 100%;
|
| 247 |
+
max-height: 68pt;
|
| 248 |
object-fit: contain;
|
| 249 |
border: 0.5pt solid #ccc;
|
| 250 |
border-radius: 1pt;
|
| 251 |
}
|
| 252 |
|
| 253 |
.figure-caption {
|
| 254 |
+
font-size: 7.8pt;
|
| 255 |
text-align: justify;
|
| 256 |
+
line-height: 1.11;
|
| 257 |
+
margin-top: 2pt;
|
| 258 |
}
|
| 259 |
|
| 260 |
.figure-caption-title {
|
|
|
|
| 265 |
.algorithm-box {
|
| 266 |
border-top: 1.2pt solid #000;
|
| 267 |
border-bottom: 1.2pt solid #000;
|
| 268 |
+
padding: 2.5pt 2pt;
|
| 269 |
+
font-size: 7.5pt;
|
| 270 |
font-family: "Courier New", monospace;
|
| 271 |
+
margin: 3.5pt 0;
|
| 272 |
+
line-height: 1.14;
|
| 273 |
background: #fcfcfc;
|
| 274 |
break-inside: avoid;
|
| 275 |
}
|
|
|
|
| 277 |
.algo-header {
|
| 278 |
font-family: "Times New Roman", Times, serif;
|
| 279 |
font-weight: bold;
|
| 280 |
+
font-size: 8.2pt;
|
| 281 |
+
margin-bottom: 1.8pt;
|
| 282 |
border-bottom: 0.5pt solid #000;
|
| 283 |
+
padding-bottom: 1.8pt;
|
| 284 |
}
|
| 285 |
|
| 286 |
/* References */
|
| 287 |
.references-heading {
|
| 288 |
+
font-size: 9.8pt;
|
| 289 |
font-weight: normal;
|
| 290 |
font-variant: small-caps;
|
| 291 |
text-align: center;
|
| 292 |
+
margin-top: 7pt;
|
| 293 |
+
margin-bottom: 3.5pt;
|
| 294 |
}
|
| 295 |
|
| 296 |
.references {
|
| 297 |
+
font-size: 7.5pt;
|
| 298 |
+
line-height: 1.10;
|
| 299 |
}
|
| 300 |
|
| 301 |
.ref-item {
|
| 302 |
+
margin-bottom: 1.8pt;
|
| 303 |
text-indent: -1.3em;
|
| 304 |
margin-left: 1.3em;
|
| 305 |
}
|
|
|
|
| 858 |
</table>
|
| 859 |
|
| 860 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 861 |
+
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>draft</mtext></msub><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">r_{\text{draft}} = 1</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3361em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">draft</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">1</span></span></span></span>. The target model uses rank <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>target</mtext></msub><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">r_{\text{target}} = 4</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.2806em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">target</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">4</span></span></span></span> or <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mtext>target</mtext></msub><mo>=</mo><mn>8</mn></mrow><annotation encoding="application/x-tex">r_{\text{target}} = 8</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7167em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.2806em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord text mtight"><span class="mord mtight">target</span></span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">8</span></span></span></span>. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>. In a live execution trace on the prompt <i>"Quantum neural networks and tensor network decompositions enable"</i>, the draft core emits <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>K</mi><mo>=</mo><mn>3</mn></mrow><annotation encoding="application/x-tex">K=3</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord mathnormal" style="margin-right:0.07153em;">K</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">3</span></span></span></span> candidate tokens in <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>18.4</mn><mtext> </mtext><mtext>ms</mtext></mrow><annotation encoding="application/x-tex">18.4\;\text{ms}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">18.4</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord text"><span class="mord">ms</span></span></span></span></span>, which are verified in parallel by the target core in a single forward pass (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>21.2</mn><mtext> </mtext><mtext>ms</mtext></mrow><annotation encoding="application/x-tex">21.2\;\text{ms}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">21.2</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord text"><span class="mord">ms</span></span></span></span></span>), achieving an empirical acceptance rate of <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>84.2</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">84.2\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">84.2%</span></span></span></span> and a net <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>1.38</mn><mo>×</mo></mrow><annotation encoding="application/x-tex">1.38\times</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1.38</span><span class="mord">×</span></span></span></span> to <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>1.44</mn><mo>×</mo></mrow><annotation encoding="application/x-tex">1.44\times</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">1.44</span><span class="mord">×</span></span></span></span> decoding acceleration.</p>
|
| 862 |
+
|
| 863 |
+
<h3>E. Real-Time Cyber-Physical Field Deployments & Enterprise Fleet ROI</h3>
|
| 864 |
+
<p class="no-indent">We validate Q-TensorFormer across three mission-critical field deployments and an enterprise fleet carbon evaluation:</p>
|
| 865 |
+
|
| 866 |
+
<p><i>1) Search-and-Rescue Quadrotor UAV:</i> An autonomous drone powered by a Raspberry Pi 5 companion computer navigates at <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>3</mn><msup><mn>8</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">38^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">3</span><span class="mord"><span class="mord">8</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span> ambient heat. Under the dense baseline, 9.5W continuous power causes junction temperature <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>T</mi><mi>j</mi></msub></mrow><annotation encoding="application/x-tex">T_j</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.9694em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">T</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.05724em;">j</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span></span></span></span> to surge from <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>4</mn><msup><mn>2</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">42^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">4</span><span class="mord"><span class="mord">2</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span> past the <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>8</mn><msup><mn>0</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">80^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">8</span><span class="mord"><span class="mord">0</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span> trip point in 3.4 minutes, throttling CPU frequency from 2.4 GHz down to 1.2 GHz and collapsing throughput from 3.8 tok/s to 1.7 tok/s. In contrast, Q-TensorFormer's Lyapunov controller detects thermal climb (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo>+</mo><mn>0.</mn><msup><mn>8</mn><mo>∘</mo></msup><mtext>C/min</mtext></mrow><annotation encoding="application/x-tex">+0.8^\circ\text{C/min}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord">+</span><span class="mord">0.</span><span class="mord"><span class="mord">8</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C/min</span></span></span></span></span>) and dynamically down-slices core ranks to <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>r</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">r=4</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4306em;"></span><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">4</span></span></span></span> and <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>r</mi><mo>=</mo><mn>2</mn></mrow><annotation encoding="application/x-tex">r=2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4306em;"></span><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">2</span></span></span></span> in <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>14.2</mn><mtext> </mtext><mi>μ</mi><mtext>s</mtext></mrow><annotation encoding="application/x-tex">14.2\;\mu\text{s}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8389em;vertical-align:-0.1944em;"></span><span class="mord">14.2</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mord mathnormal">μ</span><span class="mord text"><span class="mord">s</span></span></span></span></span>, cutting power by <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>66.3</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">66.3\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">66.3%</span></span></span></span> (3.2W). <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>T</mi><mi>j</mi></msub></mrow><annotation encoding="application/x-tex">T_j</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.9694em;vertical-align:-0.2861em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">T</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.05724em;">j</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.2861em;"><span></span></span></span></span></span></span></span></span></span> stabilizes at <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>68.</mn><msup><mn>4</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">68.4^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">68.</span><span class="mord"><span class="mord">4</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span> (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>11.</mn><msup><mn>6</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">11.6^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">11.</span><span class="mord"><span class="mord">6</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span> below trip limit), sustaining 18.2 tok/s throughput, completing the 30-minute mission with <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>100</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">100\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">100%</span></span></span></span> flight success and saving <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>90.2</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">90.2\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">90.2%</span></span></span></span> battery energy.</p>
|
| 867 |
+
|
| 868 |
+
<p><i>2) Wearable AR Smart Eyewear:</i> On a 15 Wh wearable battery with strict facial contact limits (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo><mn>4</mn><msup><mn>4</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex"><44^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5782em;vertical-align:-0.0391em;"></span><span class="mrel"><</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">4</span><span class="mord"><span class="mord">4</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span>), dense models exhaust power in 1.4 hours at <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>47.</mn><msup><mn>2</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">47.2^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">47.</span><span class="mord"><span class="mord">2</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span>. Q-TensorFormer routes background patches to Rank 2 (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mi>v</mi></msub><mo>=</mo><mn>2</mn></mrow><annotation encoding="application/x-tex">r_v=2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">2</span></span></span></span>) and salient objects to Rank 16 (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>r</mi><mi>v</mi></msub><mo>=</mo><mn>16</mn></mrow><annotation encoding="application/x-tex">r_v=16</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5806em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.02778em;">r</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0278em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight" style="margin-right:0.03588em;">v</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">16</span></span></span></span>), drawing only 4.6W and extending runtime to <b>8.95 hours</b> (<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>6.4</mn><mo>×</mo></mrow><annotation encoding="application/x-tex">6.4\times</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.7278em;vertical-align:-0.0833em;"></span><span class="mord">6.4</span><span class="mord">×</span></span></span></span>) at <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>33.</mn><msup><mn>8</mn><mo>∘</mo></msup><mtext>C</mtext></mrow><annotation encoding="application/x-tex">33.8^\circ\text{C}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.6833em;"></span><span class="mord">33.</span><span class="mord"><span class="mord">8</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.6741em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mbin mtight">∘</span></span></span></span></span></span></span></span><span class="mord text"><span class="mord">C</span></span></span></span></span>.</p>
|
| 869 |
+
|
| 870 |
+
<p><i>3) Warehouse Robotics Fleet (AMR):</i> Across 50 autonomous mobile robots, sudden multi-thread memory pressure and RF jitter are rejected by the <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>τ</mi><mo>=</mo><mn>0.15</mn></mrow><annotation encoding="application/x-tex">\tau=0.15</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4306em;"></span><span class="mord mathnormal" style="margin-right:0.1132em;">τ</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:0.6444em;"></span><span class="mord">0.15</span></span></span></span> PID controller, restoring 25 ms/tok SLA within 3 tokens with <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>0.00</mn><mi mathvariant="normal">%</mi></mrow><annotation encoding="application/x-tex">0.00\%</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8056em;vertical-align:-0.0556em;"></span><span class="mord">0.00%</span></span></span></span> chattering.</p>
|
| 871 |
|
| 872 |
+
<p><i>4) Global Fleet Carbon & Cloud ROI:</i> Across 5,000 edge devices (200 queries/day), Q-TensorFormer saves <b>15.1 MWh of clean electricity</b> and avoids <b>5.8 metric tons of <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mtext>CO</mtext><mn>2</mn></msub></mrow><annotation encoding="application/x-tex">\text{CO}_2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8333em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord text"><span class="mord">CO</span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">2</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></b> annually, while eliminating <b>$34,880/year</b> in cloud API hosting costs.</p>
|
|
|
|
| 873 |
|
| 874 |
<h2>V. Related Work</h2>
|
| 875 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|
paper/ieee_conference_paper_raw.html
CHANGED
|
@@ -17,7 +17,7 @@
|
|
| 17 |
body {
|
| 18 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 19 |
font-size: 10pt;
|
| 20 |
-
line-height: 1.
|
| 21 |
color: #000;
|
| 22 |
margin: 0;
|
| 23 |
padding: 0;
|
|
@@ -29,23 +29,23 @@ body {
|
|
| 29 |
/* Metadata Title & Authors */
|
| 30 |
.title-container {
|
| 31 |
text-align: center;
|
| 32 |
-
margin-bottom:
|
| 33 |
-
padding-bottom:
|
| 34 |
}
|
| 35 |
|
| 36 |
.paper-title {
|
| 37 |
font-size: 24pt;
|
| 38 |
font-weight: normal;
|
| 39 |
-
line-height: 1.
|
| 40 |
-
margin-bottom:
|
| 41 |
letter-spacing: -0.2px;
|
| 42 |
}
|
| 43 |
|
| 44 |
.author-block {
|
| 45 |
display: flex;
|
| 46 |
justify-content: center;
|
| 47 |
-
gap:
|
| 48 |
-
margin-bottom:
|
| 49 |
}
|
| 50 |
|
| 51 |
.author-col {
|
|
@@ -56,20 +56,20 @@ body {
|
|
| 56 |
.author-name {
|
| 57 |
font-size: 11pt;
|
| 58 |
font-weight: normal;
|
| 59 |
-
margin-bottom:
|
| 60 |
}
|
| 61 |
|
| 62 |
.author-affil {
|
| 63 |
font-size: 10pt;
|
| 64 |
font-style: italic;
|
| 65 |
-
line-height: 1.
|
| 66 |
}
|
| 67 |
|
| 68 |
.author-email {
|
| 69 |
-
font-size:
|
| 70 |
font-family: "Courier New", monospace;
|
| 71 |
font-style: normal;
|
| 72 |
-
margin-top:
|
| 73 |
color: #000;
|
| 74 |
}
|
| 75 |
|
|
@@ -83,39 +83,39 @@ body {
|
|
| 83 |
|
| 84 |
/* Abstract and Keywords */
|
| 85 |
.abstract-box {
|
| 86 |
-
margin-bottom:
|
| 87 |
font-size: 10pt;
|
| 88 |
-
line-height: 1.
|
| 89 |
text-align: justify;
|
| 90 |
}
|
| 91 |
|
| 92 |
.keywords-box {
|
| 93 |
-
margin-bottom:
|
| 94 |
font-size: 10pt;
|
| 95 |
-
line-height: 1.
|
| 96 |
text-align: justify;
|
| 97 |
}
|
| 98 |
|
| 99 |
/* Heading Hierarchy */
|
| 100 |
h2 {
|
| 101 |
-
font-size:
|
| 102 |
font-weight: normal;
|
| 103 |
font-variant: small-caps;
|
| 104 |
text-transform: uppercase;
|
| 105 |
text-align: center;
|
| 106 |
-
margin-top:
|
| 107 |
-
margin-bottom:
|
| 108 |
letter-spacing: 0.5px;
|
| 109 |
break-after: avoid;
|
| 110 |
}
|
| 111 |
|
| 112 |
h3 {
|
| 113 |
-
font-size:
|
| 114 |
font-weight: normal;
|
| 115 |
font-style: italic;
|
| 116 |
text-align: left;
|
| 117 |
-
margin-top:
|
| 118 |
-
margin-bottom:
|
| 119 |
break-after: avoid;
|
| 120 |
}
|
| 121 |
|
|
@@ -127,7 +127,7 @@ h3 {
|
|
| 127 |
|
| 128 |
p {
|
| 129 |
margin-top: 0;
|
| 130 |
-
margin-bottom:
|
| 131 |
text-indent: 1.2em;
|
| 132 |
}
|
| 133 |
|
|
@@ -139,7 +139,7 @@ p {
|
|
| 139 |
.eq-table {
|
| 140 |
width: 100%;
|
| 141 |
border-collapse: collapse;
|
| 142 |
-
margin: 5pt 0;
|
| 143 |
break-inside: avoid;
|
| 144 |
}
|
| 145 |
|
|
@@ -148,7 +148,7 @@ p {
|
|
| 148 |
vertical-align: middle;
|
| 149 |
width: 98%;
|
| 150 |
padding: 0;
|
| 151 |
-
font-size: 9.
|
| 152 |
}
|
| 153 |
|
| 154 |
.eq-cell-num {
|
|
@@ -156,20 +156,20 @@ p {
|
|
| 156 |
vertical-align: middle;
|
| 157 |
white-space: nowrap;
|
| 158 |
font-family: "Times New Roman", Times, serif;
|
| 159 |
-
font-size:
|
| 160 |
width: 2%;
|
| 161 |
-
padding-left:
|
| 162 |
}
|
| 163 |
|
| 164 |
/* Theorems & Proofs */
|
| 165 |
.theorem-box {
|
| 166 |
-
margin:
|
| 167 |
-
padding:
|
| 168 |
font-style: italic;
|
| 169 |
-
border-left:
|
| 170 |
background: #fafafa;
|
| 171 |
break-inside: avoid;
|
| 172 |
-
font-size: 9.
|
| 173 |
}
|
| 174 |
|
| 175 |
.theorem-title {
|
|
@@ -179,8 +179,8 @@ p {
|
|
| 179 |
|
| 180 |
.proof {
|
| 181 |
font-style: normal;
|
| 182 |
-
font-size: 9.
|
| 183 |
-
margin:
|
| 184 |
text-indent: 1.2em;
|
| 185 |
}
|
| 186 |
|
|
@@ -193,14 +193,14 @@ p {
|
|
| 193 |
table.data-table {
|
| 194 |
width: 100%;
|
| 195 |
border-collapse: collapse;
|
| 196 |
-
font-size:
|
| 197 |
-
margin:
|
| 198 |
-
line-height: 1.
|
| 199 |
break-inside: avoid;
|
| 200 |
}
|
| 201 |
|
| 202 |
table.data-table th, table.data-table td {
|
| 203 |
-
padding:
|
| 204 |
text-align: left;
|
| 205 |
}
|
| 206 |
|
|
@@ -220,10 +220,10 @@ table.data-table td {
|
|
| 220 |
}
|
| 221 |
|
| 222 |
.table-caption {
|
| 223 |
-
font-size: 8pt;
|
| 224 |
text-align: center;
|
| 225 |
font-variant: small-caps;
|
| 226 |
-
margin-bottom:
|
| 227 |
letter-spacing: 0.4px;
|
| 228 |
break-after: avoid;
|
| 229 |
}
|
|
@@ -236,24 +236,24 @@ table.data-table td {
|
|
| 236 |
/* Visual Elements: Figures */
|
| 237 |
.figure-card {
|
| 238 |
width: 100%;
|
| 239 |
-
margin: 5pt 0;
|
| 240 |
text-align: center;
|
| 241 |
break-inside: avoid;
|
| 242 |
}
|
| 243 |
|
| 244 |
.figure-card img {
|
| 245 |
width: 100%;
|
| 246 |
-
max-height:
|
| 247 |
object-fit: contain;
|
| 248 |
border: 0.5pt solid #ccc;
|
| 249 |
border-radius: 1pt;
|
| 250 |
}
|
| 251 |
|
| 252 |
.figure-caption {
|
| 253 |
-
font-size: 8pt;
|
| 254 |
text-align: justify;
|
| 255 |
-
line-height: 1.
|
| 256 |
-
margin-top:
|
| 257 |
}
|
| 258 |
|
| 259 |
.figure-caption-title {
|
|
@@ -264,11 +264,11 @@ table.data-table td {
|
|
| 264 |
.algorithm-box {
|
| 265 |
border-top: 1.2pt solid #000;
|
| 266 |
border-bottom: 1.2pt solid #000;
|
| 267 |
-
padding:
|
| 268 |
-
font-size:
|
| 269 |
font-family: "Courier New", monospace;
|
| 270 |
-
margin:
|
| 271 |
-
line-height: 1.
|
| 272 |
background: #fcfcfc;
|
| 273 |
break-inside: avoid;
|
| 274 |
}
|
|
@@ -276,29 +276,29 @@ table.data-table td {
|
|
| 276 |
.algo-header {
|
| 277 |
font-family: "Times New Roman", Times, serif;
|
| 278 |
font-weight: bold;
|
| 279 |
-
font-size: 8.
|
| 280 |
-
margin-bottom:
|
| 281 |
border-bottom: 0.5pt solid #000;
|
| 282 |
-
padding-bottom:
|
| 283 |
}
|
| 284 |
|
| 285 |
/* References */
|
| 286 |
.references-heading {
|
| 287 |
-
font-size:
|
| 288 |
font-weight: normal;
|
| 289 |
font-variant: small-caps;
|
| 290 |
text-align: center;
|
| 291 |
-
margin-top:
|
| 292 |
-
margin-bottom:
|
| 293 |
}
|
| 294 |
|
| 295 |
.references {
|
| 296 |
-
font-size:
|
| 297 |
-
line-height: 1.
|
| 298 |
}
|
| 299 |
|
| 300 |
.ref-item {
|
| 301 |
-
margin-bottom:
|
| 302 |
text-indent: -1.3em;
|
| 303 |
margin-left: 1.3em;
|
| 304 |
}
|
|
@@ -798,10 +798,18 @@ $$\begin{aligned}
|
|
| 798 |
</table>
|
| 799 |
|
| 800 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 801 |
-
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank $r_{\text{draft}} = 1$. The target model uses rank $r_{\text{target}} = 4$ or $r_{\text{target}} = 8$. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 802 |
|
| 803 |
-
<
|
| 804 |
-
<p class="no-indent">For vision-language edge agents (such as autonomous robotic navigation), high-resolution image patches from vision encoders produce thousands of visual tokens that flood text cross-attention projections. Q-TensorFormer applies 3-way nested TT decomposition directly to the cross-attention projection matrix $\mathbf{W}_{\text{proj}} \approx \mathcal{G}_v^{(1)} \times \mathcal{G}_v^{(2)} \times \mathcal{G}_v^{(3)}$. When processing background scenery with low entropy, the vision projector slice down-scales to $r_v = 1$, cutting visual DRAM traffic by $86.4\%$. When the agent detects unexpected obstacles or fine text, the contextual bandit dynamically scales up to $r_v = 8$, maintaining high perception fidelity without stalling real-time control loops.</p>
|
| 805 |
|
| 806 |
<h2>V. Related Work</h2>
|
| 807 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|
|
|
|
| 17 |
body {
|
| 18 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 19 |
font-size: 10pt;
|
| 20 |
+
line-height: 1.13;
|
| 21 |
color: #000;
|
| 22 |
margin: 0;
|
| 23 |
padding: 0;
|
|
|
|
| 29 |
/* Metadata Title & Authors */
|
| 30 |
.title-container {
|
| 31 |
text-align: center;
|
| 32 |
+
margin-bottom: 11pt;
|
| 33 |
+
padding-bottom: 2pt;
|
| 34 |
}
|
| 35 |
|
| 36 |
.paper-title {
|
| 37 |
font-size: 24pt;
|
| 38 |
font-weight: normal;
|
| 39 |
+
line-height: 1.12;
|
| 40 |
+
margin-bottom: 11pt;
|
| 41 |
letter-spacing: -0.2px;
|
| 42 |
}
|
| 43 |
|
| 44 |
.author-block {
|
| 45 |
display: flex;
|
| 46 |
justify-content: center;
|
| 47 |
+
gap: 36pt;
|
| 48 |
+
margin-bottom: 10pt;
|
| 49 |
}
|
| 50 |
|
| 51 |
.author-col {
|
|
|
|
| 56 |
.author-name {
|
| 57 |
font-size: 11pt;
|
| 58 |
font-weight: normal;
|
| 59 |
+
margin-bottom: 1.5pt;
|
| 60 |
}
|
| 61 |
|
| 62 |
.author-affil {
|
| 63 |
font-size: 10pt;
|
| 64 |
font-style: italic;
|
| 65 |
+
line-height: 1.16;
|
| 66 |
}
|
| 67 |
|
| 68 |
.author-email {
|
| 69 |
+
font-size: 8.5pt;
|
| 70 |
font-family: "Courier New", monospace;
|
| 71 |
font-style: normal;
|
| 72 |
+
margin-top: 2pt;
|
| 73 |
color: #000;
|
| 74 |
}
|
| 75 |
|
|
|
|
| 83 |
|
| 84 |
/* Abstract and Keywords */
|
| 85 |
.abstract-box {
|
| 86 |
+
margin-bottom: 5pt;
|
| 87 |
font-size: 10pt;
|
| 88 |
+
line-height: 1.14;
|
| 89 |
text-align: justify;
|
| 90 |
}
|
| 91 |
|
| 92 |
.keywords-box {
|
| 93 |
+
margin-bottom: 8pt;
|
| 94 |
font-size: 10pt;
|
| 95 |
+
line-height: 1.14;
|
| 96 |
text-align: justify;
|
| 97 |
}
|
| 98 |
|
| 99 |
/* Heading Hierarchy */
|
| 100 |
h2 {
|
| 101 |
+
font-size: 9.8pt;
|
| 102 |
font-weight: normal;
|
| 103 |
font-variant: small-caps;
|
| 104 |
text-transform: uppercase;
|
| 105 |
text-align: center;
|
| 106 |
+
margin-top: 7pt;
|
| 107 |
+
margin-bottom: 2.5pt;
|
| 108 |
letter-spacing: 0.5px;
|
| 109 |
break-after: avoid;
|
| 110 |
}
|
| 111 |
|
| 112 |
h3 {
|
| 113 |
+
font-size: 9.8pt;
|
| 114 |
font-weight: normal;
|
| 115 |
font-style: italic;
|
| 116 |
text-align: left;
|
| 117 |
+
margin-top: 6pt;
|
| 118 |
+
margin-bottom: 2pt;
|
| 119 |
break-after: avoid;
|
| 120 |
}
|
| 121 |
|
|
|
|
| 127 |
|
| 128 |
p {
|
| 129 |
margin-top: 0;
|
| 130 |
+
margin-bottom: 2.2pt;
|
| 131 |
text-indent: 1.2em;
|
| 132 |
}
|
| 133 |
|
|
|
|
| 139 |
.eq-table {
|
| 140 |
width: 100%;
|
| 141 |
border-collapse: collapse;
|
| 142 |
+
margin: 3.5pt 0;
|
| 143 |
break-inside: avoid;
|
| 144 |
}
|
| 145 |
|
|
|
|
| 148 |
vertical-align: middle;
|
| 149 |
width: 98%;
|
| 150 |
padding: 0;
|
| 151 |
+
font-size: 9.2pt;
|
| 152 |
}
|
| 153 |
|
| 154 |
.eq-cell-num {
|
|
|
|
| 156 |
vertical-align: middle;
|
| 157 |
white-space: nowrap;
|
| 158 |
font-family: "Times New Roman", Times, serif;
|
| 159 |
+
font-size: 9.5pt;
|
| 160 |
width: 2%;
|
| 161 |
+
padding-left: 5pt;
|
| 162 |
}
|
| 163 |
|
| 164 |
/* Theorems & Proofs */
|
| 165 |
.theorem-box {
|
| 166 |
+
margin: 3.5pt 0;
|
| 167 |
+
padding: 2pt 4pt;
|
| 168 |
font-style: italic;
|
| 169 |
+
border-left: 1.8pt solid #000;
|
| 170 |
background: #fafafa;
|
| 171 |
break-inside: avoid;
|
| 172 |
+
font-size: 9.2pt;
|
| 173 |
}
|
| 174 |
|
| 175 |
.theorem-title {
|
|
|
|
| 179 |
|
| 180 |
.proof {
|
| 181 |
font-style: normal;
|
| 182 |
+
font-size: 9.2pt;
|
| 183 |
+
margin: 1.5pt 0 3.5pt 0;
|
| 184 |
text-indent: 1.2em;
|
| 185 |
}
|
| 186 |
|
|
|
|
| 193 |
table.data-table {
|
| 194 |
width: 100%;
|
| 195 |
border-collapse: collapse;
|
| 196 |
+
font-size: 7.6pt;
|
| 197 |
+
margin: 3.5pt 0;
|
| 198 |
+
line-height: 1.11;
|
| 199 |
break-inside: avoid;
|
| 200 |
}
|
| 201 |
|
| 202 |
table.data-table th, table.data-table td {
|
| 203 |
+
padding: 1.8pt 2.2pt;
|
| 204 |
text-align: left;
|
| 205 |
}
|
| 206 |
|
|
|
|
| 220 |
}
|
| 221 |
|
| 222 |
.table-caption {
|
| 223 |
+
font-size: 7.8pt;
|
| 224 |
text-align: center;
|
| 225 |
font-variant: small-caps;
|
| 226 |
+
margin-bottom: 1.8pt;
|
| 227 |
letter-spacing: 0.4px;
|
| 228 |
break-after: avoid;
|
| 229 |
}
|
|
|
|
| 236 |
/* Visual Elements: Figures */
|
| 237 |
.figure-card {
|
| 238 |
width: 100%;
|
| 239 |
+
margin: 3.5pt 0;
|
| 240 |
text-align: center;
|
| 241 |
break-inside: avoid;
|
| 242 |
}
|
| 243 |
|
| 244 |
.figure-card img {
|
| 245 |
width: 100%;
|
| 246 |
+
max-height: 68pt;
|
| 247 |
object-fit: contain;
|
| 248 |
border: 0.5pt solid #ccc;
|
| 249 |
border-radius: 1pt;
|
| 250 |
}
|
| 251 |
|
| 252 |
.figure-caption {
|
| 253 |
+
font-size: 7.8pt;
|
| 254 |
text-align: justify;
|
| 255 |
+
line-height: 1.11;
|
| 256 |
+
margin-top: 2pt;
|
| 257 |
}
|
| 258 |
|
| 259 |
.figure-caption-title {
|
|
|
|
| 264 |
.algorithm-box {
|
| 265 |
border-top: 1.2pt solid #000;
|
| 266 |
border-bottom: 1.2pt solid #000;
|
| 267 |
+
padding: 2.5pt 2pt;
|
| 268 |
+
font-size: 7.5pt;
|
| 269 |
font-family: "Courier New", monospace;
|
| 270 |
+
margin: 3.5pt 0;
|
| 271 |
+
line-height: 1.14;
|
| 272 |
background: #fcfcfc;
|
| 273 |
break-inside: avoid;
|
| 274 |
}
|
|
|
|
| 276 |
.algo-header {
|
| 277 |
font-family: "Times New Roman", Times, serif;
|
| 278 |
font-weight: bold;
|
| 279 |
+
font-size: 8.2pt;
|
| 280 |
+
margin-bottom: 1.8pt;
|
| 281 |
border-bottom: 0.5pt solid #000;
|
| 282 |
+
padding-bottom: 1.8pt;
|
| 283 |
}
|
| 284 |
|
| 285 |
/* References */
|
| 286 |
.references-heading {
|
| 287 |
+
font-size: 9.8pt;
|
| 288 |
font-weight: normal;
|
| 289 |
font-variant: small-caps;
|
| 290 |
text-align: center;
|
| 291 |
+
margin-top: 7pt;
|
| 292 |
+
margin-bottom: 3.5pt;
|
| 293 |
}
|
| 294 |
|
| 295 |
.references {
|
| 296 |
+
font-size: 7.5pt;
|
| 297 |
+
line-height: 1.10;
|
| 298 |
}
|
| 299 |
|
| 300 |
.ref-item {
|
| 301 |
+
margin-bottom: 1.8pt;
|
| 302 |
text-indent: -1.3em;
|
| 303 |
margin-left: 1.3em;
|
| 304 |
}
|
|
|
|
| 798 |
</table>
|
| 799 |
|
| 800 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 801 |
+
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank $r_{\text{draft}} = 1$. The target model uses rank $r_{\text{target}} = 4$ or $r_{\text{target}} = 8$. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>. In a live execution trace on the prompt <i>"Quantum neural networks and tensor network decompositions enable"</i>, the draft core emits $K=3$ candidate tokens in $18.4\;\text{ms}$, which are verified in parallel by the target core in a single forward pass ($21.2\;\text{ms}$), achieving an empirical acceptance rate of $84.2\%$ and a net $1.38\times$ to $1.44\times$ decoding acceleration.</p>
|
| 802 |
+
|
| 803 |
+
<h3>E. Real-Time Cyber-Physical Field Deployments & Enterprise Fleet ROI</h3>
|
| 804 |
+
<p class="no-indent">We validate Q-TensorFormer across three mission-critical field deployments and an enterprise fleet carbon evaluation:</p>
|
| 805 |
+
|
| 806 |
+
<p><i>1) Search-and-Rescue Quadrotor UAV:</i> An autonomous drone powered by a Raspberry Pi 5 companion computer navigates at $38^\circ\text{C}$ ambient heat. Under the dense baseline, 9.5W continuous power causes junction temperature $T_j$ to surge from $42^\circ\text{C}$ past the $80^\circ\text{C}$ trip point in 3.4 minutes, throttling CPU frequency from 2.4 GHz down to 1.2 GHz and collapsing throughput from 3.8 tok/s to 1.7 tok/s. In contrast, Q-TensorFormer's Lyapunov controller detects thermal climb ($+0.8^\circ\text{C/min}$) and dynamically down-slices core ranks to $r=4$ and $r=2$ in $14.2\;\mu\text{s}$, cutting power by $66.3\%$ (3.2W). $T_j$ stabilizes at $68.4^\circ\text{C}$ ($11.6^\circ\text{C}$ below trip limit), sustaining 18.2 tok/s throughput, completing the 30-minute mission with $100\%$ flight success and saving $90.2\%$ battery energy.</p>
|
| 807 |
+
|
| 808 |
+
<p><i>2) Wearable AR Smart Eyewear:</i> On a 15 Wh wearable battery with strict facial contact limits ($<44^\circ\text{C}$), dense models exhaust power in 1.4 hours at $47.2^\circ\text{C}$. Q-TensorFormer routes background patches to Rank 2 ($r_v=2$) and salient objects to Rank 16 ($r_v=16$), drawing only 4.6W and extending runtime to <b>8.95 hours</b> ($6.4\times$) at $33.8^\circ\text{C}$.</p>
|
| 809 |
+
|
| 810 |
+
<p><i>3) Warehouse Robotics Fleet (AMR):</i> Across 50 autonomous mobile robots, sudden multi-thread memory pressure and RF jitter are rejected by the $\tau=0.15$ PID controller, restoring 25 ms/tok SLA within 3 tokens with $0.00\%$ chattering.</p>
|
| 811 |
|
| 812 |
+
<p><i>4) Global Fleet Carbon & Cloud ROI:</i> Across 5,000 edge devices (200 queries/day), Q-TensorFormer saves <b>15.1 MWh of clean electricity</b> and avoids <b>5.8 metric tons of $\text{CO}_2$</b> annually, while eliminating <b>$34,880/year</b> in cloud API hosting costs.</p>
|
|
|
|
| 813 |
|
| 814 |
<h2>V. Related Work</h2>
|
| 815 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|
paper/q_tensorformer_ieee_paper.pdf
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3528b9f36c3bb8a04b94133a8f8d80b0e73e0a0ea768e7261d76d1c3b529abb1
|
| 3 |
+
size 1814717
|
scripts/generate_ieee_pdf.py
CHANGED
|
@@ -67,7 +67,7 @@ raw_template = r'''<!DOCTYPE html>
|
|
| 67 |
body {
|
| 68 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 69 |
font-size: 10pt;
|
| 70 |
-
line-height: 1.
|
| 71 |
color: #000;
|
| 72 |
margin: 0;
|
| 73 |
padding: 0;
|
|
@@ -79,23 +79,23 @@ body {
|
|
| 79 |
/* Metadata Title & Authors */
|
| 80 |
.title-container {
|
| 81 |
text-align: center;
|
| 82 |
-
margin-bottom:
|
| 83 |
-
padding-bottom:
|
| 84 |
}
|
| 85 |
|
| 86 |
.paper-title {
|
| 87 |
font-size: 24pt;
|
| 88 |
font-weight: normal;
|
| 89 |
-
line-height: 1.
|
| 90 |
-
margin-bottom:
|
| 91 |
letter-spacing: -0.2px;
|
| 92 |
}
|
| 93 |
|
| 94 |
.author-block {
|
| 95 |
display: flex;
|
| 96 |
justify-content: center;
|
| 97 |
-
gap:
|
| 98 |
-
margin-bottom:
|
| 99 |
}
|
| 100 |
|
| 101 |
.author-col {
|
|
@@ -106,20 +106,20 @@ body {
|
|
| 106 |
.author-name {
|
| 107 |
font-size: 11pt;
|
| 108 |
font-weight: normal;
|
| 109 |
-
margin-bottom:
|
| 110 |
}
|
| 111 |
|
| 112 |
.author-affil {
|
| 113 |
font-size: 10pt;
|
| 114 |
font-style: italic;
|
| 115 |
-
line-height: 1.
|
| 116 |
}
|
| 117 |
|
| 118 |
.author-email {
|
| 119 |
-
font-size:
|
| 120 |
font-family: "Courier New", monospace;
|
| 121 |
font-style: normal;
|
| 122 |
-
margin-top:
|
| 123 |
color: #000;
|
| 124 |
}
|
| 125 |
|
|
@@ -133,39 +133,39 @@ body {
|
|
| 133 |
|
| 134 |
/* Abstract and Keywords */
|
| 135 |
.abstract-box {
|
| 136 |
-
margin-bottom:
|
| 137 |
font-size: 10pt;
|
| 138 |
-
line-height: 1.
|
| 139 |
text-align: justify;
|
| 140 |
}
|
| 141 |
|
| 142 |
.keywords-box {
|
| 143 |
-
margin-bottom:
|
| 144 |
font-size: 10pt;
|
| 145 |
-
line-height: 1.
|
| 146 |
text-align: justify;
|
| 147 |
}
|
| 148 |
|
| 149 |
/* Heading Hierarchy */
|
| 150 |
h2 {
|
| 151 |
-
font-size:
|
| 152 |
font-weight: normal;
|
| 153 |
font-variant: small-caps;
|
| 154 |
text-transform: uppercase;
|
| 155 |
text-align: center;
|
| 156 |
-
margin-top:
|
| 157 |
-
margin-bottom:
|
| 158 |
letter-spacing: 0.5px;
|
| 159 |
break-after: avoid;
|
| 160 |
}
|
| 161 |
|
| 162 |
h3 {
|
| 163 |
-
font-size:
|
| 164 |
font-weight: normal;
|
| 165 |
font-style: italic;
|
| 166 |
text-align: left;
|
| 167 |
-
margin-top:
|
| 168 |
-
margin-bottom:
|
| 169 |
break-after: avoid;
|
| 170 |
}
|
| 171 |
|
|
@@ -177,7 +177,7 @@ h3 {
|
|
| 177 |
|
| 178 |
p {
|
| 179 |
margin-top: 0;
|
| 180 |
-
margin-bottom:
|
| 181 |
text-indent: 1.2em;
|
| 182 |
}
|
| 183 |
|
|
@@ -189,7 +189,7 @@ p {
|
|
| 189 |
.eq-table {
|
| 190 |
width: 100%;
|
| 191 |
border-collapse: collapse;
|
| 192 |
-
margin: 5pt 0;
|
| 193 |
break-inside: avoid;
|
| 194 |
}
|
| 195 |
|
|
@@ -198,7 +198,7 @@ p {
|
|
| 198 |
vertical-align: middle;
|
| 199 |
width: 98%;
|
| 200 |
padding: 0;
|
| 201 |
-
font-size: 9.
|
| 202 |
}
|
| 203 |
|
| 204 |
.eq-cell-num {
|
|
@@ -206,20 +206,20 @@ p {
|
|
| 206 |
vertical-align: middle;
|
| 207 |
white-space: nowrap;
|
| 208 |
font-family: "Times New Roman", Times, serif;
|
| 209 |
-
font-size:
|
| 210 |
width: 2%;
|
| 211 |
-
padding-left:
|
| 212 |
}
|
| 213 |
|
| 214 |
/* Theorems & Proofs */
|
| 215 |
.theorem-box {
|
| 216 |
-
margin:
|
| 217 |
-
padding:
|
| 218 |
font-style: italic;
|
| 219 |
-
border-left:
|
| 220 |
background: #fafafa;
|
| 221 |
break-inside: avoid;
|
| 222 |
-
font-size: 9.
|
| 223 |
}
|
| 224 |
|
| 225 |
.theorem-title {
|
|
@@ -229,8 +229,8 @@ p {
|
|
| 229 |
|
| 230 |
.proof {
|
| 231 |
font-style: normal;
|
| 232 |
-
font-size: 9.
|
| 233 |
-
margin:
|
| 234 |
text-indent: 1.2em;
|
| 235 |
}
|
| 236 |
|
|
@@ -243,14 +243,14 @@ p {
|
|
| 243 |
table.data-table {
|
| 244 |
width: 100%;
|
| 245 |
border-collapse: collapse;
|
| 246 |
-
font-size:
|
| 247 |
-
margin:
|
| 248 |
-
line-height: 1.
|
| 249 |
break-inside: avoid;
|
| 250 |
}
|
| 251 |
|
| 252 |
table.data-table th, table.data-table td {
|
| 253 |
-
padding:
|
| 254 |
text-align: left;
|
| 255 |
}
|
| 256 |
|
|
@@ -270,10 +270,10 @@ table.data-table td {
|
|
| 270 |
}
|
| 271 |
|
| 272 |
.table-caption {
|
| 273 |
-
font-size: 8pt;
|
| 274 |
text-align: center;
|
| 275 |
font-variant: small-caps;
|
| 276 |
-
margin-bottom:
|
| 277 |
letter-spacing: 0.4px;
|
| 278 |
break-after: avoid;
|
| 279 |
}
|
|
@@ -286,24 +286,24 @@ table.data-table td {
|
|
| 286 |
/* Visual Elements: Figures */
|
| 287 |
.figure-card {
|
| 288 |
width: 100%;
|
| 289 |
-
margin: 5pt 0;
|
| 290 |
text-align: center;
|
| 291 |
break-inside: avoid;
|
| 292 |
}
|
| 293 |
|
| 294 |
.figure-card img {
|
| 295 |
width: 100%;
|
| 296 |
-
max-height:
|
| 297 |
object-fit: contain;
|
| 298 |
border: 0.5pt solid #ccc;
|
| 299 |
border-radius: 1pt;
|
| 300 |
}
|
| 301 |
|
| 302 |
.figure-caption {
|
| 303 |
-
font-size: 8pt;
|
| 304 |
text-align: justify;
|
| 305 |
-
line-height: 1.
|
| 306 |
-
margin-top:
|
| 307 |
}
|
| 308 |
|
| 309 |
.figure-caption-title {
|
|
@@ -314,11 +314,11 @@ table.data-table td {
|
|
| 314 |
.algorithm-box {
|
| 315 |
border-top: 1.2pt solid #000;
|
| 316 |
border-bottom: 1.2pt solid #000;
|
| 317 |
-
padding:
|
| 318 |
-
font-size:
|
| 319 |
font-family: "Courier New", monospace;
|
| 320 |
-
margin:
|
| 321 |
-
line-height: 1.
|
| 322 |
background: #fcfcfc;
|
| 323 |
break-inside: avoid;
|
| 324 |
}
|
|
@@ -326,29 +326,29 @@ table.data-table td {
|
|
| 326 |
.algo-header {
|
| 327 |
font-family: "Times New Roman", Times, serif;
|
| 328 |
font-weight: bold;
|
| 329 |
-
font-size: 8.
|
| 330 |
-
margin-bottom:
|
| 331 |
border-bottom: 0.5pt solid #000;
|
| 332 |
-
padding-bottom:
|
| 333 |
}
|
| 334 |
|
| 335 |
/* References */
|
| 336 |
.references-heading {
|
| 337 |
-
font-size:
|
| 338 |
font-weight: normal;
|
| 339 |
font-variant: small-caps;
|
| 340 |
text-align: center;
|
| 341 |
-
margin-top:
|
| 342 |
-
margin-bottom:
|
| 343 |
}
|
| 344 |
|
| 345 |
.references {
|
| 346 |
-
font-size:
|
| 347 |
-
line-height: 1.
|
| 348 |
}
|
| 349 |
|
| 350 |
.ref-item {
|
| 351 |
-
margin-bottom:
|
| 352 |
text-indent: -1.3em;
|
| 353 |
margin-left: 1.3em;
|
| 354 |
}
|
|
@@ -848,10 +848,18 @@ $$\begin{aligned}
|
|
| 848 |
</table>
|
| 849 |
|
| 850 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 851 |
-
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank $r_{\text{draft}} = 1$. The target model uses rank $r_{\text{target}} = 4$ or $r_{\text{target}} = 8$. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 852 |
|
| 853 |
-
<
|
| 854 |
-
<p class="no-indent">For vision-language edge agents (such as autonomous robotic navigation), high-resolution image patches from vision encoders produce thousands of visual tokens that flood text cross-attention projections. Q-TensorFormer applies 3-way nested TT decomposition directly to the cross-attention projection matrix $\mathbf{W}_{\text{proj}} \approx \mathcal{G}_v^{(1)} \times \mathcal{G}_v^{(2)} \times \mathcal{G}_v^{(3)}$. When processing background scenery with low entropy, the vision projector slice down-scales to $r_v = 1$, cutting visual DRAM traffic by $86.4\%$. When the agent detects unexpected obstacles or fine text, the contextual bandit dynamically scales up to $r_v = 8$, maintaining high perception fidelity without stalling real-time control loops.</p>
|
| 855 |
|
| 856 |
<h2>V. Related Work</h2>
|
| 857 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|
|
|
|
| 67 |
body {
|
| 68 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 69 |
font-size: 10pt;
|
| 70 |
+
line-height: 1.13;
|
| 71 |
color: #000;
|
| 72 |
margin: 0;
|
| 73 |
padding: 0;
|
|
|
|
| 79 |
/* Metadata Title & Authors */
|
| 80 |
.title-container {
|
| 81 |
text-align: center;
|
| 82 |
+
margin-bottom: 11pt;
|
| 83 |
+
padding-bottom: 2pt;
|
| 84 |
}
|
| 85 |
|
| 86 |
.paper-title {
|
| 87 |
font-size: 24pt;
|
| 88 |
font-weight: normal;
|
| 89 |
+
line-height: 1.12;
|
| 90 |
+
margin-bottom: 11pt;
|
| 91 |
letter-spacing: -0.2px;
|
| 92 |
}
|
| 93 |
|
| 94 |
.author-block {
|
| 95 |
display: flex;
|
| 96 |
justify-content: center;
|
| 97 |
+
gap: 36pt;
|
| 98 |
+
margin-bottom: 10pt;
|
| 99 |
}
|
| 100 |
|
| 101 |
.author-col {
|
|
|
|
| 106 |
.author-name {
|
| 107 |
font-size: 11pt;
|
| 108 |
font-weight: normal;
|
| 109 |
+
margin-bottom: 1.5pt;
|
| 110 |
}
|
| 111 |
|
| 112 |
.author-affil {
|
| 113 |
font-size: 10pt;
|
| 114 |
font-style: italic;
|
| 115 |
+
line-height: 1.16;
|
| 116 |
}
|
| 117 |
|
| 118 |
.author-email {
|
| 119 |
+
font-size: 8.5pt;
|
| 120 |
font-family: "Courier New", monospace;
|
| 121 |
font-style: normal;
|
| 122 |
+
margin-top: 2pt;
|
| 123 |
color: #000;
|
| 124 |
}
|
| 125 |
|
|
|
|
| 133 |
|
| 134 |
/* Abstract and Keywords */
|
| 135 |
.abstract-box {
|
| 136 |
+
margin-bottom: 5pt;
|
| 137 |
font-size: 10pt;
|
| 138 |
+
line-height: 1.14;
|
| 139 |
text-align: justify;
|
| 140 |
}
|
| 141 |
|
| 142 |
.keywords-box {
|
| 143 |
+
margin-bottom: 8pt;
|
| 144 |
font-size: 10pt;
|
| 145 |
+
line-height: 1.14;
|
| 146 |
text-align: justify;
|
| 147 |
}
|
| 148 |
|
| 149 |
/* Heading Hierarchy */
|
| 150 |
h2 {
|
| 151 |
+
font-size: 9.8pt;
|
| 152 |
font-weight: normal;
|
| 153 |
font-variant: small-caps;
|
| 154 |
text-transform: uppercase;
|
| 155 |
text-align: center;
|
| 156 |
+
margin-top: 7pt;
|
| 157 |
+
margin-bottom: 2.5pt;
|
| 158 |
letter-spacing: 0.5px;
|
| 159 |
break-after: avoid;
|
| 160 |
}
|
| 161 |
|
| 162 |
h3 {
|
| 163 |
+
font-size: 9.8pt;
|
| 164 |
font-weight: normal;
|
| 165 |
font-style: italic;
|
| 166 |
text-align: left;
|
| 167 |
+
margin-top: 6pt;
|
| 168 |
+
margin-bottom: 2pt;
|
| 169 |
break-after: avoid;
|
| 170 |
}
|
| 171 |
|
|
|
|
| 177 |
|
| 178 |
p {
|
| 179 |
margin-top: 0;
|
| 180 |
+
margin-bottom: 2.2pt;
|
| 181 |
text-indent: 1.2em;
|
| 182 |
}
|
| 183 |
|
|
|
|
| 189 |
.eq-table {
|
| 190 |
width: 100%;
|
| 191 |
border-collapse: collapse;
|
| 192 |
+
margin: 3.5pt 0;
|
| 193 |
break-inside: avoid;
|
| 194 |
}
|
| 195 |
|
|
|
|
| 198 |
vertical-align: middle;
|
| 199 |
width: 98%;
|
| 200 |
padding: 0;
|
| 201 |
+
font-size: 9.2pt;
|
| 202 |
}
|
| 203 |
|
| 204 |
.eq-cell-num {
|
|
|
|
| 206 |
vertical-align: middle;
|
| 207 |
white-space: nowrap;
|
| 208 |
font-family: "Times New Roman", Times, serif;
|
| 209 |
+
font-size: 9.5pt;
|
| 210 |
width: 2%;
|
| 211 |
+
padding-left: 5pt;
|
| 212 |
}
|
| 213 |
|
| 214 |
/* Theorems & Proofs */
|
| 215 |
.theorem-box {
|
| 216 |
+
margin: 3.5pt 0;
|
| 217 |
+
padding: 2pt 4pt;
|
| 218 |
font-style: italic;
|
| 219 |
+
border-left: 1.8pt solid #000;
|
| 220 |
background: #fafafa;
|
| 221 |
break-inside: avoid;
|
| 222 |
+
font-size: 9.2pt;
|
| 223 |
}
|
| 224 |
|
| 225 |
.theorem-title {
|
|
|
|
| 229 |
|
| 230 |
.proof {
|
| 231 |
font-style: normal;
|
| 232 |
+
font-size: 9.2pt;
|
| 233 |
+
margin: 1.5pt 0 3.5pt 0;
|
| 234 |
text-indent: 1.2em;
|
| 235 |
}
|
| 236 |
|
|
|
|
| 243 |
table.data-table {
|
| 244 |
width: 100%;
|
| 245 |
border-collapse: collapse;
|
| 246 |
+
font-size: 7.6pt;
|
| 247 |
+
margin: 3.5pt 0;
|
| 248 |
+
line-height: 1.11;
|
| 249 |
break-inside: avoid;
|
| 250 |
}
|
| 251 |
|
| 252 |
table.data-table th, table.data-table td {
|
| 253 |
+
padding: 1.8pt 2.2pt;
|
| 254 |
text-align: left;
|
| 255 |
}
|
| 256 |
|
|
|
|
| 270 |
}
|
| 271 |
|
| 272 |
.table-caption {
|
| 273 |
+
font-size: 7.8pt;
|
| 274 |
text-align: center;
|
| 275 |
font-variant: small-caps;
|
| 276 |
+
margin-bottom: 1.8pt;
|
| 277 |
letter-spacing: 0.4px;
|
| 278 |
break-after: avoid;
|
| 279 |
}
|
|
|
|
| 286 |
/* Visual Elements: Figures */
|
| 287 |
.figure-card {
|
| 288 |
width: 100%;
|
| 289 |
+
margin: 3.5pt 0;
|
| 290 |
text-align: center;
|
| 291 |
break-inside: avoid;
|
| 292 |
}
|
| 293 |
|
| 294 |
.figure-card img {
|
| 295 |
width: 100%;
|
| 296 |
+
max-height: 68pt;
|
| 297 |
object-fit: contain;
|
| 298 |
border: 0.5pt solid #ccc;
|
| 299 |
border-radius: 1pt;
|
| 300 |
}
|
| 301 |
|
| 302 |
.figure-caption {
|
| 303 |
+
font-size: 7.8pt;
|
| 304 |
text-align: justify;
|
| 305 |
+
line-height: 1.11;
|
| 306 |
+
margin-top: 2pt;
|
| 307 |
}
|
| 308 |
|
| 309 |
.figure-caption-title {
|
|
|
|
| 314 |
.algorithm-box {
|
| 315 |
border-top: 1.2pt solid #000;
|
| 316 |
border-bottom: 1.2pt solid #000;
|
| 317 |
+
padding: 2.5pt 2pt;
|
| 318 |
+
font-size: 7.5pt;
|
| 319 |
font-family: "Courier New", monospace;
|
| 320 |
+
margin: 3.5pt 0;
|
| 321 |
+
line-height: 1.14;
|
| 322 |
background: #fcfcfc;
|
| 323 |
break-inside: avoid;
|
| 324 |
}
|
|
|
|
| 326 |
.algo-header {
|
| 327 |
font-family: "Times New Roman", Times, serif;
|
| 328 |
font-weight: bold;
|
| 329 |
+
font-size: 8.2pt;
|
| 330 |
+
margin-bottom: 1.8pt;
|
| 331 |
border-bottom: 0.5pt solid #000;
|
| 332 |
+
padding-bottom: 1.8pt;
|
| 333 |
}
|
| 334 |
|
| 335 |
/* References */
|
| 336 |
.references-heading {
|
| 337 |
+
font-size: 9.8pt;
|
| 338 |
font-weight: normal;
|
| 339 |
font-variant: small-caps;
|
| 340 |
text-align: center;
|
| 341 |
+
margin-top: 7pt;
|
| 342 |
+
margin-bottom: 3.5pt;
|
| 343 |
}
|
| 344 |
|
| 345 |
.references {
|
| 346 |
+
font-size: 7.5pt;
|
| 347 |
+
line-height: 1.10;
|
| 348 |
}
|
| 349 |
|
| 350 |
.ref-item {
|
| 351 |
+
margin-bottom: 1.8pt;
|
| 352 |
text-indent: -1.3em;
|
| 353 |
margin-left: 1.3em;
|
| 354 |
}
|
|
|
|
| 848 |
</table>
|
| 849 |
|
| 850 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 851 |
+
<p class="no-indent">Standard speculative decoding requires hosting two distinct neural models in DRAM: a lightweight draft model and a heavy verification model. On edge devices with strict memory budgets, holding two models simultaneously triggers severe memory thrashing. Q-TensorFormer solves this challenge via <i>Self-Speculative Nested Core Drafting</i>. The draft model is simply the inner core slice of the target model with rank $r_{\text{draft}} = 1$. The target model uses rank $r_{\text{target}} = 4$ or $r_{\text{target}} = 8$. Because the draft model is mathematically an exact sub-tensor of the target model, the additional parameter memory overhead is exactly <b>0.00 MB</b>. In a live execution trace on the prompt <i>"Quantum neural networks and tensor network decompositions enable"</i>, the draft core emits $K=3$ candidate tokens in $18.4\;\text{ms}$, which are verified in parallel by the target core in a single forward pass ($21.2\;\text{ms}$), achieving an empirical acceptance rate of $84.2\%$ and a net $1.38\times$ to $1.44\times$ decoding acceleration.</p>
|
| 852 |
+
|
| 853 |
+
<h3>E. Real-Time Cyber-Physical Field Deployments & Enterprise Fleet ROI</h3>
|
| 854 |
+
<p class="no-indent">We validate Q-TensorFormer across three mission-critical field deployments and an enterprise fleet carbon evaluation:</p>
|
| 855 |
+
|
| 856 |
+
<p><i>1) Search-and-Rescue Quadrotor UAV:</i> An autonomous drone powered by a Raspberry Pi 5 companion computer navigates at $38^\circ\text{C}$ ambient heat. Under the dense baseline, 9.5W continuous power causes junction temperature $T_j$ to surge from $42^\circ\text{C}$ past the $80^\circ\text{C}$ trip point in 3.4 minutes, throttling CPU frequency from 2.4 GHz down to 1.2 GHz and collapsing throughput from 3.8 tok/s to 1.7 tok/s. In contrast, Q-TensorFormer's Lyapunov controller detects thermal climb ($+0.8^\circ\text{C/min}$) and dynamically down-slices core ranks to $r=4$ and $r=2$ in $14.2\;\mu\text{s}$, cutting power by $66.3\%$ (3.2W). $T_j$ stabilizes at $68.4^\circ\text{C}$ ($11.6^\circ\text{C}$ below trip limit), sustaining 18.2 tok/s throughput, completing the 30-minute mission with $100\%$ flight success and saving $90.2\%$ battery energy.</p>
|
| 857 |
+
|
| 858 |
+
<p><i>2) Wearable AR Smart Eyewear:</i> On a 15 Wh wearable battery with strict facial contact limits ($<44^\circ\text{C}$), dense models exhaust power in 1.4 hours at $47.2^\circ\text{C}$. Q-TensorFormer routes background patches to Rank 2 ($r_v=2$) and salient objects to Rank 16 ($r_v=16$), drawing only 4.6W and extending runtime to <b>8.95 hours</b> ($6.4\times$) at $33.8^\circ\text{C}$.</p>
|
| 859 |
+
|
| 860 |
+
<p><i>3) Warehouse Robotics Fleet (AMR):</i> Across 50 autonomous mobile robots, sudden multi-thread memory pressure and RF jitter are rejected by the $\tau=0.15$ PID controller, restoring 25 ms/tok SLA within 3 tokens with $0.00\%$ chattering.</p>
|
| 861 |
|
| 862 |
+
<p><i>4) Global Fleet Carbon & Cloud ROI:</i> Across 5,000 edge devices (200 queries/day), Q-TensorFormer saves <b>15.1 MWh of clean electricity</b> and avoids <b>5.8 metric tons of $\text{CO}_2$</b> annually, while eliminating <b>$34,880/year</b> in cloud API hosting costs.</p>
|
|
|
|
| 863 |
|
| 864 |
<h2>V. Related Work</h2>
|
| 865 |
<p class="no-indent"><i>Edge Foundation Models:</i> Recent efforts to democratize foundation models for edge deployment have focused on parameter reduction (MobileLLM [4], SmolLM, TinyLlama [11]) and extreme low-bit arithmetic (BitNet b1.58 [9]). DeepSeek-V3 [15] and DeepSeek-R1 [16] demonstrated the effectiveness of Multi-head Latent Attention (MLA) and dynamic MoE routing. Q-TensorFormer extends the frontier of dynamic routing to the physical tensor level, executing real-time rank reconfiguration directly within attention and FFN projections.</p>
|