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 ·
3d24ee3
1
Parent(s): 114f6bd
docs(paper): exact IEEE layout standards, 6 pages, table-aligned non-overlapping equations & 7 figures
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:b978b9c212c2ab6c7b09e696e4c5e5968d50b66fb40b1f1de454931413b7c31c
|
| 3 |
+
size 1805737
|
paper/ieee_conference_paper.html
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
paper/ieee_conference_paper_raw.html
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
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:b978b9c212c2ab6c7b09e696e4c5e5968d50b66fb40b1f1de454931413b7c31c
|
| 3 |
+
size 1805737
|
scripts/generate_ieee_pdf.py
CHANGED
|
@@ -1,15 +1,21 @@
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
"""
|
| 3 |
scripts/generate_ieee_pdf.py
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"""
|
| 14 |
|
| 15 |
import base64
|
|
@@ -35,13 +41,14 @@ def get_base64_image(filename):
|
|
| 35 |
return ""
|
| 36 |
|
| 37 |
fig1_b64 = get_base64_image("figure1_architecture.png")
|
| 38 |
-
fig2_b64 = get_base64_image("
|
| 39 |
-
fig3_b64 = get_base64_image("
|
| 40 |
-
fig4_b64 = get_base64_image("
|
| 41 |
-
fig5_b64 = get_base64_image("
|
| 42 |
-
fig6_b64 = get_base64_image("
|
| 43 |
-
|
| 44 |
-
|
|
|
|
| 45 |
<html lang="en">
|
| 46 |
<head>
|
| 47 |
<meta charset="utf-8">
|
|
@@ -51,13 +58,16 @@ raw_template = '''<!DOCTYPE html>
|
|
| 51 |
|
| 52 |
@page {
|
| 53 |
size: letter;
|
| 54 |
-
margin: 0.
|
|
|
|
|
|
|
|
|
|
| 55 |
}
|
| 56 |
|
| 57 |
body {
|
| 58 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 59 |
font-size: 10pt;
|
| 60 |
-
line-height: 1.
|
| 61 |
color: #000;
|
| 62 |
margin: 0;
|
| 63 |
padding: 0;
|
|
@@ -66,105 +76,108 @@ body {
|
|
| 66 |
print-color-adjust: exact;
|
| 67 |
}
|
| 68 |
|
|
|
|
| 69 |
.title-container {
|
| 70 |
text-align: center;
|
| 71 |
-
margin-bottom:
|
| 72 |
-
padding-bottom:
|
| 73 |
-
border-bottom: 0.5pt solid #ddd;
|
| 74 |
}
|
| 75 |
|
| 76 |
.paper-title {
|
| 77 |
-
font-size:
|
| 78 |
-
font-weight:
|
| 79 |
-
line-height: 1.
|
| 80 |
-
margin-bottom:
|
| 81 |
letter-spacing: -0.2px;
|
| 82 |
}
|
| 83 |
|
| 84 |
.author-block {
|
| 85 |
display: flex;
|
| 86 |
justify-content: center;
|
| 87 |
-
gap:
|
| 88 |
-
margin-bottom:
|
| 89 |
}
|
| 90 |
|
| 91 |
.author-col {
|
| 92 |
text-align: center;
|
| 93 |
-
max-width:
|
| 94 |
}
|
| 95 |
|
| 96 |
.author-name {
|
| 97 |
-
font-size:
|
| 98 |
-
font-weight:
|
| 99 |
margin-bottom: 2pt;
|
| 100 |
}
|
| 101 |
|
| 102 |
.author-affil {
|
| 103 |
-
font-size:
|
| 104 |
font-style: italic;
|
| 105 |
-
line-height: 1.
|
| 106 |
}
|
| 107 |
|
| 108 |
.author-email {
|
| 109 |
-
font-size:
|
| 110 |
font-family: "Courier New", monospace;
|
|
|
|
| 111 |
margin-top: 3pt;
|
| 112 |
-
color: #
|
| 113 |
}
|
| 114 |
|
|
|
|
| 115 |
.two-col {
|
| 116 |
column-count: 2;
|
| 117 |
-
column-gap: 18pt;
|
| 118 |
text-align: justify;
|
|
|
|
| 119 |
}
|
| 120 |
|
|
|
|
| 121 |
.abstract-box {
|
| 122 |
-
margin-bottom:
|
| 123 |
-
font-size:
|
| 124 |
-
line-height: 1.
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
.abstract-title {
|
| 128 |
-
font-weight: bold;
|
| 129 |
-
font-style: italic;
|
| 130 |
}
|
| 131 |
|
| 132 |
.keywords-box {
|
| 133 |
-
margin-bottom:
|
| 134 |
-
font-size:
|
| 135 |
-
line-height: 1.
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
.keywords-title {
|
| 139 |
-
font-weight: bold;
|
| 140 |
-
font-style: italic;
|
| 141 |
}
|
| 142 |
|
|
|
|
| 143 |
h2 {
|
| 144 |
-
font-size:
|
| 145 |
-
font-weight:
|
|
|
|
| 146 |
text-transform: uppercase;
|
| 147 |
text-align: center;
|
| 148 |
-
margin-top:
|
| 149 |
-
margin-bottom:
|
| 150 |
-
letter-spacing: 0.
|
| 151 |
break-after: avoid;
|
| 152 |
-
border-bottom: 0.5pt solid #ccc;
|
| 153 |
-
padding-bottom: 2pt;
|
| 154 |
}
|
| 155 |
|
| 156 |
h3 {
|
| 157 |
font-size: 10pt;
|
| 158 |
-
font-weight:
|
| 159 |
font-style: italic;
|
|
|
|
| 160 |
margin-top: 9pt;
|
| 161 |
margin-bottom: 3pt;
|
| 162 |
break-after: avoid;
|
| 163 |
}
|
| 164 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
p {
|
| 166 |
margin-top: 0;
|
| 167 |
-
margin-bottom: 5pt;
|
| 168 |
text-indent: 1.2em;
|
| 169 |
}
|
| 170 |
|
|
@@ -172,13 +185,41 @@ p {
|
|
| 172 |
text-indent: 0;
|
| 173 |
}
|
| 174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
.theorem-box {
|
| 176 |
-
margin:
|
| 177 |
-
padding:
|
| 178 |
font-style: italic;
|
| 179 |
-
border-left:
|
| 180 |
-
background: #
|
| 181 |
break-inside: avoid;
|
|
|
|
| 182 |
}
|
| 183 |
|
| 184 |
.theorem-title {
|
|
@@ -188,8 +229,8 @@ p {
|
|
| 188 |
|
| 189 |
.proof {
|
| 190 |
font-style: normal;
|
| 191 |
-
font-size: 9.
|
| 192 |
-
margin:
|
| 193 |
text-indent: 1.2em;
|
| 194 |
}
|
| 195 |
|
|
@@ -198,135 +239,128 @@ p {
|
|
| 198 |
font-weight: bold;
|
| 199 |
}
|
| 200 |
|
| 201 |
-
|
|
|
|
| 202 |
width: 100%;
|
| 203 |
border-collapse: collapse;
|
| 204 |
-
font-size:
|
| 205 |
-
margin:
|
| 206 |
-
line-height: 1.
|
| 207 |
break-inside: avoid;
|
| 208 |
}
|
| 209 |
|
| 210 |
-
th, td {
|
| 211 |
-
padding:
|
| 212 |
text-align: left;
|
| 213 |
}
|
| 214 |
|
| 215 |
-
th {
|
| 216 |
font-weight: bold;
|
| 217 |
-
border-top: 1.
|
| 218 |
-
border-bottom: 0.
|
| 219 |
-
background: #
|
| 220 |
}
|
| 221 |
|
| 222 |
-
td {
|
| 223 |
border-bottom: 0.4pt solid #e0e0e0;
|
| 224 |
}
|
| 225 |
|
| 226 |
.table-bottom {
|
| 227 |
-
border-bottom: 1.
|
| 228 |
}
|
| 229 |
|
| 230 |
.table-caption {
|
| 231 |
-
font-size:
|
| 232 |
text-align: center;
|
| 233 |
-
font-
|
| 234 |
-
margin-bottom:
|
| 235 |
-
|
| 236 |
-
|
| 237 |
}
|
| 238 |
|
| 239 |
.highlight-row {
|
| 240 |
-
background-color: #
|
| 241 |
font-weight: bold;
|
| 242 |
}
|
| 243 |
|
|
|
|
| 244 |
.figure-card {
|
| 245 |
width: 100%;
|
| 246 |
-
margin:
|
| 247 |
text-align: center;
|
| 248 |
break-inside: avoid;
|
| 249 |
}
|
| 250 |
|
| 251 |
.figure-card img {
|
| 252 |
width: 100%;
|
| 253 |
-
max-height:
|
| 254 |
object-fit: contain;
|
| 255 |
-
border: 0.5pt solid #
|
| 256 |
-
border-radius:
|
| 257 |
}
|
| 258 |
|
| 259 |
.figure-caption {
|
| 260 |
-
font-size:
|
| 261 |
text-align: justify;
|
| 262 |
-
line-height: 1.
|
| 263 |
-
margin-top:
|
| 264 |
}
|
| 265 |
|
| 266 |
.figure-caption-title {
|
| 267 |
-
font-weight:
|
| 268 |
}
|
| 269 |
|
|
|
|
| 270 |
.algorithm-box {
|
| 271 |
-
border-top: 1.
|
| 272 |
-
border-bottom: 1.
|
| 273 |
-
padding:
|
| 274 |
-
font-size:
|
| 275 |
font-family: "Courier New", monospace;
|
| 276 |
-
margin:
|
| 277 |
-
line-height: 1.
|
| 278 |
-
background: #
|
| 279 |
break-inside: avoid;
|
| 280 |
}
|
| 281 |
|
| 282 |
.algo-header {
|
| 283 |
font-family: "Times New Roman", Times, serif;
|
| 284 |
font-weight: bold;
|
| 285 |
-
font-size:
|
| 286 |
-
margin-bottom:
|
| 287 |
-
border-bottom: 0.
|
| 288 |
padding-bottom: 2pt;
|
| 289 |
}
|
| 290 |
|
| 291 |
-
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
align-items: center;
|
| 295 |
-
margin: 5pt 0;
|
| 296 |
-
break-inside: avoid;
|
| 297 |
-
}
|
| 298 |
-
|
| 299 |
-
.eq-math {
|
| 300 |
-
flex-grow: 1;
|
| 301 |
-
text-align: center;
|
| 302 |
-
}
|
| 303 |
-
|
| 304 |
-
.eq-num {
|
| 305 |
-
font-size: 9.5pt;
|
| 306 |
font-weight: normal;
|
| 307 |
-
|
| 308 |
-
|
|
|
|
|
|
|
| 309 |
}
|
| 310 |
|
| 311 |
.references {
|
| 312 |
-
font-size:
|
| 313 |
-
line-height: 1.
|
| 314 |
}
|
| 315 |
|
| 316 |
.ref-item {
|
| 317 |
-
margin-bottom:
|
| 318 |
-
text-indent: -1.
|
| 319 |
-
margin-left: 1.
|
| 320 |
}
|
| 321 |
|
| 322 |
.perf-badge {
|
| 323 |
display: inline-block;
|
| 324 |
-
padding:
|
| 325 |
font-size: 7.5pt;
|
| 326 |
font-weight: bold;
|
| 327 |
color: #0b6623;
|
| 328 |
background: #eaf8ec;
|
| 329 |
-
border-radius:
|
| 330 |
}
|
| 331 |
</style>
|
| 332 |
</head>
|
|
@@ -337,12 +371,12 @@ td {
|
|
| 337 |
<div class="author-block">
|
| 338 |
<div class="author-col">
|
| 339 |
<div class="author-name">Dr. Madala Guru Brahmam</div>
|
| 340 |
-
<div class="author-affil">
|
| 341 |
<div class="author-email">madala.gurubrahmam@vitap.ac.in</div>
|
| 342 |
</div>
|
| 343 |
<div class="author-col">
|
| 344 |
<div class="author-name">V C Premchand Yadav</div>
|
| 345 |
-
<div class="author-affil">
|
| 346 |
<div class="author-email">premchandyadav.vc21bce9124@vitap.ac.in</div>
|
| 347 |
</div>
|
| 348 |
</div>
|
|
@@ -351,61 +385,67 @@ td {
|
|
| 351 |
<div class="two-col">
|
| 352 |
|
| 353 |
<div class="abstract-box">
|
| 354 |
-
<
|
| 355 |
</div>
|
| 356 |
|
| 357 |
<div class="keywords-box">
|
| 358 |
-
<
|
| 359 |
</div>
|
| 360 |
|
| 361 |
<h2>I. Introduction</h2>
|
| 362 |
-
<p class="no-indent">Decentralized on-device artificial intelligence requires deploying multi-billion-parameter Transformer foundation models directly on constrained edge
|
| 363 |
|
| 364 |
-
<p><
|
| 365 |
|
| 366 |
-
<p><
|
| 367 |
|
| 368 |
<p>To overcome these foundational limitations, we introduce <b>Q-TensorFormer</b>. Rather than relying on heuristic early-exit shortcuts [12] or uncalibrated token skipping [13], Q-TensorFormer formulates edge inference as a closed-loop cyber-physical control problem. It harmonizes algebraic Tensor-Train core nesting with discrete-time Lyapunov stability and contextual bandit routing.</p>
|
| 369 |
|
| 370 |
<p><b>Primary Contributions of this Work:</b></p>
|
| 371 |
-
<p>• <
|
| 372 |
-
<p>• <
|
| 373 |
-
<p>• <
|
| 374 |
-
<p>• <
|
| 375 |
-
<p>• <
|
| 376 |
|
| 377 |
<div class="figure-card">
|
| 378 |
<img src="__FIG1__" alt="Q-TensorFormer Architecture">
|
| 379 |
-
<div class="figure-caption">
|
| 380 |
</div>
|
| 381 |
|
| 382 |
-
<h2>II.
|
| 383 |
|
| 384 |
<h3>A. Problem Formulation: Constrained Markov Decision Process</h3>
|
| 385 |
-
<p class="no-indent">We formulate edge foundation inference as a Constrained Markov Decision Process (CMDP) over discrete decoding steps $t \
|
| 386 |
|
| 387 |
<p>The objective is to minimize cumulative task loss subject to physical hardware constraints:</p>
|
| 388 |
|
| 389 |
-
$$\
|
|
|
|
|
|
|
|
|
|
| 390 |
|
| 391 |
<p class="no-indent">where $C_k$ represents physical resource costs (DRAM memory traffic, thermal dissipation, latency deadline) and $B_k$ denotes the specified budget ceiling.</p>
|
| 392 |
|
| 393 |
<h3>B. Zero-SVD Nested Tensor-Train Parameterization</h3>
|
| 394 |
-
<p class="no-indent">Let $\
|
| 395 |
|
| 396 |
-
$$\
|
|
|
|
|
|
|
|
|
|
| 397 |
|
| 398 |
-
<p class="no-indent">where $\
|
| 399 |
|
| 400 |
-
<p>In Q-TensorFormer, we parameterize cores such that lower-rank sub-tensors are nested directly inside the maximum rank allocation $r_{\
|
| 401 |
|
| 402 |
-
$$\
|
| 403 |
|
| 404 |
<p class="no-indent">Dynamic rank adjustment requires zero memory allocations, zero tensor copies, and zero singular value decompositions. The runtime execution kernel merely modifies loop trip-count bounds in processor registers, switching active rank in <b>14.2 μs</b>.</p>
|
| 405 |
|
| 406 |
<div class="figure-card">
|
| 407 |
-
<img src="__FIG2__" alt="
|
| 408 |
-
<div class="figure-caption">
|
| 409 |
</div>
|
| 410 |
|
| 411 |
<div class="algorithm-box">
|
|
@@ -423,21 +463,21 @@ $$\\mathcal{G}^{(k)}_{\\text{active}} = \\mathcal{G}^{(k)}[0:r_{\\text{curr}}, \
|
|
| 423 |
<h3>C. Discrete-Time Lyapunov-Stable Controller</h3>
|
| 424 |
<p class="no-indent">Dynamic rank adjustment without feedback stabilization can induce limit-cycle oscillations (chattering), where the model continuously flips between high and low ranks, destabilizing memory buses and degrading output coherence. Q-TensorFormer stabilizes rank transitions using a dual-subgradient controller with analytical hysteresis:</p>
|
| 425 |
|
| 426 |
-
$$e_t = L_t - L_{\
|
| 427 |
|
| 428 |
-
$$u_t = e_t + \
|
| 429 |
|
| 430 |
-
$$\
|
| 431 |
|
| 432 |
-
<p class="no-indent">where $\
|
| 433 |
|
| 434 |
-
$$\
|
| 435 |
|
| 436 |
-
<p class="no-indent">The active rank is mapped via $r_t = g(\
|
| 437 |
|
| 438 |
<div class="figure-card">
|
| 439 |
-
<img src="__FIG3__" alt="
|
| 440 |
-
<div class="figure-caption">
|
| 441 |
</div>
|
| 442 |
|
| 443 |
<div class="algorithm-box">
|
|
@@ -455,37 +495,54 @@ $$\\phi_\\tau(u) = \\begin{cases} u - \\tau, & u > \\tau \\\\ 0, & |u| \\le \\ta
|
|
| 455 |
9: <b>return</b> $r_{t+1}$ (provably stable: $\Delta V_t \le 0$)
|
| 456 |
</div>
|
| 457 |
|
| 458 |
-
<
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 459 |
|
| 460 |
<div class="theorem-box">
|
| 461 |
-
<span class="theorem-title">Theorem 1 (Nested TT-Core Truncation Error Bound):</span> Let $\
|
| 462 |
-
$$\
|
| 463 |
-
|
|
|
|
|
|
|
|
|
|
| 464 |
</div>
|
| 465 |
|
| 466 |
-
<p class="proof"><i>Proof Sketch:</i> By telescoping difference over TT factor contractions $\
|
| 467 |
|
| 468 |
<div class="theorem-box">
|
| 469 |
-
<span class="theorem-title">Theorem 2 (Discrete-Time Lyapunov Stability & Anti-Chattering):</span> Consider the Lyapunov candidate $V_t = \
|
| 470 |
-
$$\
|
| 471 |
-
where $\
|
| 472 |
</div>
|
| 473 |
|
| 474 |
-
<p class="proof"><i>Proof Sketch:</i> Evaluating
|
| 475 |
|
| 476 |
<div class="figure-card">
|
| 477 |
-
<img src="
|
| 478 |
-
<div class="figure-caption">
|
| 479 |
</div>
|
| 480 |
|
| 481 |
<div class="theorem-box">
|
| 482 |
-
<span class="theorem-title">Theorem 3 (Entropy-Gated KV Information Retention):</span> Let query-key attention distribution be $p_i = \
|
| 483 |
-
$$\
|
| 484 |
</div>
|
| 485 |
|
| 486 |
<div class="theorem-box">
|
| 487 |
-
<span class="theorem-title">Theorem 4 (LinUCB Contextual Bandit Regret Bound):</span> Under feature dimension $d$ and confidence parameter $\
|
| 488 |
-
$$\
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 489 |
</div>
|
| 490 |
|
| 491 |
<h2>IV. Quantitative Experimental Evaluation</h2>
|
|
@@ -498,14 +555,14 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 498 |
<p>• <b>Apple M3 Edge SoC:</b> 8-core CPU / 10-core GPU with 16-core Neural Engine, unified 100 GB/s memory architecture.</p>
|
| 499 |
<p>• <b>Intel Core Ultra 7 (Meteor Lake):</b> Intel 4 process, integrated Arc GPU + Intel AI Boost NPU, 32 GB LPDDR5X-7467.</p>
|
| 500 |
|
| 501 |
-
<div class="table-caption">
|
| 502 |
-
<table>
|
| 503 |
<thead>
|
| 504 |
<tr>
|
| 505 |
<th>Platform & Model Architecture</th>
|
| 506 |
-
<th>Params
|
| 507 |
-
<th>Peak RAM
|
| 508 |
-
<th>Throughput
|
| 509 |
<th>Speedup</th>
|
| 510 |
<th>Adaptation Latency</th>
|
| 511 |
<th>Thermal Drop (15 min)</th>
|
|
@@ -513,11 +570,11 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 513 |
</thead>
|
| 514 |
<tbody>
|
| 515 |
<tr>
|
| 516 |
-
<td colspan="7" style="background:#
|
| 517 |
</tr>
|
| 518 |
<tr>
|
| 519 |
<td>Dense Baseline (FP16 TinyLlama) [11]</td>
|
| 520 |
-
<td>1,
|
| 521 |
<td>2,180 MB</td>
|
| 522 |
<td>3.8 tok/s</td>
|
| 523 |
<td>1.0x</td>
|
|
@@ -526,7 +583,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 526 |
</tr>
|
| 527 |
<tr>
|
| 528 |
<td>AWQ 4-Bit Baseline [7]</td>
|
| 529 |
-
<td>1,
|
| 530 |
<td>840 MB</td>
|
| 531 |
<td>7.2 tok/s</td>
|
| 532 |
<td>1.9x</td>
|
|
@@ -535,7 +592,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 535 |
</tr>
|
| 536 |
<tr>
|
| 537 |
<td>MobileLLM-1.5B (ICLR 2024) [4]</td>
|
| 538 |
-
<td>1,
|
| 539 |
<td>1,420 MB</td>
|
| 540 |
<td>4.6 tok/s</td>
|
| 541 |
<td>1.2x</td>
|
|
@@ -544,7 +601,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 544 |
</tr>
|
| 545 |
<tr>
|
| 546 |
<td>Llama-3.2-1B Edge (Meta 2024) [1]</td>
|
| 547 |
-
<td>1,
|
| 548 |
<td>1,280 MB</td>
|
| 549 |
<td>5.4 tok/s</td>
|
| 550 |
<td>1.4x</td>
|
|
@@ -553,7 +610,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 553 |
</tr>
|
| 554 |
<tr>
|
| 555 |
<td>TensorGPT-TT Baseline (AAAI 2023) [14]</td>
|
| 556 |
-
<td>
|
| 557 |
<td>720 MB</td>
|
| 558 |
<td>8.9 tok/s</td>
|
| 559 |
<td>2.3x</td>
|
|
@@ -562,7 +619,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 562 |
</tr>
|
| 563 |
<tr class="highlight-row">
|
| 564 |
<td><b>Q-TensorFormer (Static Balanced)</b></td>
|
| 565 |
-
<td><b>
|
| 566 |
<td><b>640 MB</b></td>
|
| 567 |
<td><b>14.6 tok/s</b></td>
|
| 568 |
<td><b>3.8x</b></td>
|
|
@@ -571,7 +628,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 571 |
</tr>
|
| 572 |
<tr class="highlight-row">
|
| 573 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 574 |
-
<td><b>140–
|
| 575 |
<td><b>593 MB</b></td>
|
| 576 |
<td><b>18.2 tok/s</b></td>
|
| 577 |
<td><span class="perf-badge">+380% (4.8x)</span></td>
|
|
@@ -579,11 +636,11 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 579 |
<td><b>-2.1% (Stable)</b></td>
|
| 580 |
</tr>
|
| 581 |
<tr>
|
| 582 |
-
<td colspan="7" style="background:#
|
| 583 |
</tr>
|
| 584 |
<tr>
|
| 585 |
<td>Dense Baseline (FP16)</td>
|
| 586 |
-
<td>1,
|
| 587 |
<td>2,180 MB</td>
|
| 588 |
<td>24.1 tok/s</td>
|
| 589 |
<td>1.0x</td>
|
|
@@ -592,7 +649,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 592 |
</tr>
|
| 593 |
<tr>
|
| 594 |
<td>AWQ 4-Bit Baseline [7]</td>
|
| 595 |
-
<td>1,
|
| 596 |
<td>840 MB</td>
|
| 597 |
<td>48.6 tok/s</td>
|
| 598 |
<td>2.0x</td>
|
|
@@ -601,7 +658,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 601 |
</tr>
|
| 602 |
<tr>
|
| 603 |
<td>MobileLLM-1.5B [4]</td>
|
| 604 |
-
<td>1,
|
| 605 |
<td>1,420 MB</td>
|
| 606 |
<td>29.4 tok/s</td>
|
| 607 |
<td>1.2x</td>
|
|
@@ -610,7 +667,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 610 |
</tr>
|
| 611 |
<tr class="highlight-row">
|
| 612 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 613 |
-
<td><b>140–
|
| 614 |
<td><b>593 MB</b></td>
|
| 615 |
<td><b>94.7 tok/s</b></td>
|
| 616 |
<td><span class="perf-badge">+293% (3.9x)</span></td>
|
|
@@ -618,11 +675,11 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 618 |
<td><b>-1.8% (Stable)</b></td>
|
| 619 |
</tr>
|
| 620 |
<tr>
|
| 621 |
-
<td colspan="7" style="background:#
|
| 622 |
</tr>
|
| 623 |
<tr>
|
| 624 |
<td>Dense Baseline (FP16)</td>
|
| 625 |
-
<td>1,
|
| 626 |
<td>2,180 MB</td>
|
| 627 |
<td>46.2 tok/s</td>
|
| 628 |
<td>1.0x</td>
|
|
@@ -631,7 +688,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 631 |
</tr>
|
| 632 |
<tr>
|
| 633 |
<td>Llama-3.2-1B Edge (Meta 2024) [1]</td>
|
| 634 |
-
<td>1,
|
| 635 |
<td>1,280 MB</td>
|
| 636 |
<td>62.8 tok/s</td>
|
| 637 |
<td>1.4x</td>
|
|
@@ -640,7 +697,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 640 |
</tr>
|
| 641 |
<tr class="highlight-row">
|
| 642 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 643 |
-
<td><b>140–
|
| 644 |
<td><b>593 MB</b></td>
|
| 645 |
<td><b>164.5 tok/s</b></td>
|
| 646 |
<td><span class="perf-badge">+256% (3.6x)</span></td>
|
|
@@ -651,18 +708,18 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 651 |
</table>
|
| 652 |
|
| 653 |
<p class="no-indent">As shown in Table I, Q-TensorFormer achieves unmatched inference efficiency:</p>
|
| 654 |
-
<p>• <
|
| 655 |
-
<p>• <
|
| 656 |
-
<p>• <
|
| 657 |
-
<p>• <
|
| 658 |
|
| 659 |
<div class="figure-card">
|
| 660 |
-
<img src="
|
| 661 |
-
<div class="figure-caption">
|
| 662 |
</div>
|
| 663 |
|
| 664 |
-
<div class="table-caption">
|
| 665 |
-
<table>
|
| 666 |
<thead>
|
| 667 |
<tr>
|
| 668 |
<th>Model Architecture</th>
|
|
@@ -734,90 +791,85 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 734 |
|
| 735 |
<p class="no-indent">Table II demonstrates that Q-TensorFormer retains <b>98.6% of full dense model reasoning accuracy</b> across challenging benchmarks (GSM8K, ARC-Challenge, HellaSwag). Naive SVD truncation collapses accuracy to 68.2% because un-nested singular vectors do not share structural alignment. In contrast, Q-TensorFormer's joint nested training aligns sub-core principal components, guaranteeing near-lossless compression.</p>
|
| 736 |
|
| 737 |
-
<div class="table-caption">
|
| 738 |
-
<table>
|
| 739 |
<thead>
|
| 740 |
<tr>
|
| 741 |
<th>Architecture Setup</th>
|
| 742 |
-
<th>Arithmetic Intensity
|
| 743 |
-
<th>Intermediate DRAM Traffic
|
| 744 |
-
<th>Energy / Token
|
| 745 |
<th>Battery Life Impact</th>
|
| 746 |
-
<th>
|
| 747 |
</tr>
|
| 748 |
</thead>
|
| 749 |
<tbody>
|
| 750 |
<tr>
|
| 751 |
<td>Standard Dense Transformer</td>
|
| 752 |
-
<td>0.048</td>
|
| 753 |
-
<td>65.6 KB</td>
|
| 754 |
<td>84.2 mJ</td>
|
| 755 |
<td>Baseline (1.0x)</td>
|
| 756 |
<td>None</td>
|
| 757 |
</tr>
|
| 758 |
<tr>
|
| 759 |
<td>INT8 Weight-Only PTQ</td>
|
| 760 |
-
<td>0.096</td>
|
| 761 |
-
<td>32.8 KB</td>
|
| 762 |
<td>48.6 mJ</td>
|
| 763 |
<td>1.73x Extension</td>
|
| 764 |
<td>None</td>
|
| 765 |
</tr>
|
| 766 |
<tr>
|
| 767 |
<td>AWQ 4-Bit (MLSys 2024) [7]</td>
|
| 768 |
-
<td>0.192</td>
|
| 769 |
-
<td>16.4 KB</td>
|
| 770 |
<td>31.4 mJ</td>
|
| 771 |
<td>2.68x Extension</td>
|
| 772 |
<td>None</td>
|
| 773 |
</tr>
|
| 774 |
<tr class="highlight-row">
|
| 775 |
<td><b>Q-TensorFormer (Fused TT)</b></td>
|
| 776 |
-
<td><b>1.549</b></td>
|
| 777 |
-
<td><b>2.03 KB (-96.9%)</b></td>
|
| 778 |
<td><b>12.8 mJ</b></td>
|
| 779 |
<td><b>6.58x Extension</b></td>
|
| 780 |
<td><b>1.38x Self-Drafting</b></td>
|
| 781 |
</tr>
|
| 782 |
<tr class="highlight-row">
|
| 783 |
<td><b>Q-TensorFormer (Full Adaptive)</b></td>
|
| 784 |
-
<td><b>1.549</b></td>
|
| 785 |
-
<td><b>1.48 KB (-97.7%)</b></td>
|
| 786 |
<td><b>9.4 mJ</b></td>
|
| 787 |
<td><span class="perf-badge">8.95x Battery Extension</span></td>
|
| 788 |
-
<td><b>1.44x
|
| 789 |
</tr>
|
| 790 |
</tbody>
|
| 791 |
</table>
|
| 792 |
|
| 793 |
-
<div class="figure-card">
|
| 794 |
-
<img src="__FIG6__" alt="Hardware Roofline Model">
|
| 795 |
-
<div class="figure-caption"><span class="figure-caption-title">Fig. 6.</span> Hardware roofline model across edge microarchitectures. Illustrates how in-register fused TT contraction moves the operational point from severe memory-bandwidth saturation ($0.048$ FLOP/Byte) toward peak processor arithmetic intensity ($1.549$ FLOP/Byte).</div>
|
| 796 |
-
</div>
|
| 797 |
-
|
| 798 |
<h3>D. Self-Speculative Drafting with Zero Parameter Overhead</h3>
|
| 799 |
-
<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 <
|
| 800 |
|
| 801 |
<h3>E. Multi-Modal Cross-Attention Projector Slicing</h3>
|
| 802 |
-
<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 $\
|
| 803 |
|
| 804 |
<h2>V. Related Work</h2>
|
| 805 |
-
<p class="no-indent"><
|
| 806 |
|
| 807 |
-
<p><
|
| 808 |
|
| 809 |
-
<p><
|
| 810 |
|
| 811 |
-
<p><
|
| 812 |
|
| 813 |
<h2>VI. Conclusion and Open-Source Artifacts</h2>
|
| 814 |
<p class="no-indent">In this paper, we introduced <b>Q-TensorFormer</b>, a closed-loop cyber-physical architecture for real-time, high-throughput foundation model inference on edge hardware. By harmonizing nested Tensor-Train decompositions, discrete-time Lyapunov-stable feedback control, and contextual bandit token routing, Q-TensorFormer eliminates runtime SVD re-factorization, achieves up to a <b>4.8x throughput speedup</b> on edge SoCs, slashes peak RAM by <b>72.8%</b>, and preserves over <b>98.6% of full dense reasoning accuracy</b>.</p>
|
| 815 |
|
| 816 |
-
<p>The entire production codebase, verification test suites (107 unit tests passed), pre-trained checkpoints, and interactive deployment spaces are open-sourced under the Apache-2.0 license at:</p>
|
| 817 |
<p>• <b>Model Hub Repository:</b> <a href="https://huggingface.co/Premchan369/Q-TensorFormer">https://huggingface.co/Premchan369/Q-TensorFormer</a></p>
|
| 818 |
<p>• <b>Interactive Web Space:</b> <a href="https://huggingface.co/spaces/Premchan369/Q-TensorFormer">https://huggingface.co/spaces/Premchan369/Q-TensorFormer</a></p>
|
| 819 |
|
| 820 |
-
<
|
| 821 |
<div class="references">
|
| 822 |
<div class="ref-item">[1] Llama 3 Team (Meta AI), "The Llama 3 Herd of Models," <i>arXiv preprint arXiv:2407.21783</i>, Jul. 2024.</div>
|
| 823 |
<div class="ref-item">[2] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, "LoRA: Low-Rank Adaptation of Large Language Models," in <i>Proc. Int. Conf. Learn. Representations (ICLR)</i>, Apr. 2022.</div>
|
|
@@ -826,7 +878,7 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 826 |
<div class="ref-item">[5] G. Xiao, J. Lin, M. Han, S. Tung, and S. Han, "Efficient Streaming Language Models with Attention Sinks," in <i>Proc. Int. Conf. Learn. Representations (ICLR)</i>, May 2024.</div>
|
| 827 |
<div class="ref-item">[6] J. Shah, G. Dao, S. Gu, K. Li, and C. Ré, "FlashAttention-3: Fast and Memory-Efficient Language Model Attention with Asynchrony and Low-Precision," <i>arXiv preprint arXiv:2407.08608</i>, Jul. 2024.</div>
|
| 828 |
<div class="ref-item">[7] J. Lin, J. Tang, H. Tang, C. Yang, X. Dang, C. Gan, and S. Han, "AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration," in <i>Proc. Conf. Mach. Learn. Syst. (MLSys)</i>, May 2024.</div>
|
| 829 |
-
<div class="ref-item">[8] G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, "SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models," in <i>Proc. Int. Conf.
|
| 830 |
<div class="ref-item">[9] H. Wang, S. Ma, L. Dong, S. Huang, H. Wang, Z. Ling, F. Wei et al., "The Era of 1-bit LLMs: All Large Language Models in 1.58 Bits (BitNet b1.58)," <i>arXiv preprint arXiv:2402.17764</i>, Feb. 2024.</div>
|
| 831 |
<div class="ref-item">[10] T. Schuster, A. Fisch, J. Gupta, M. Dehghani, D. Bahri, V. Tran, Y. Tay, and D. Metzler, "Confident Adaptive Language Modeling," in <i>Proc. Conf. Neural Inf. Process. Syst. (NeurIPS)</i>, Dec. 2022.</div>
|
| 832 |
<div class="ref-item">[11] P. Zhang, G. Zeng, T. Wang, and W. Lu, "TinyLlama: An Open-Source Small Language Model," <i>arXiv preprint arXiv:2401.02385</i>, Jan. 2024.</div>
|
|
@@ -846,20 +898,14 @@ $$\\mathcal{R}(T) \\le 2 \\sqrt{2 d T \\ln(1 + T / \\lambda)} \\left( \\sqrt{\\l
|
|
| 846 |
</html>
|
| 847 |
'''
|
| 848 |
|
| 849 |
-
# Replace placeholders
|
| 850 |
html_content = raw_template.replace("__FIG1__", fig1_b64)
|
| 851 |
html_content = html_content.replace("__FIG2__", fig2_b64)
|
| 852 |
html_content = html_content.replace("__FIG3__", fig3_b64)
|
| 853 |
html_content = html_content.replace("__FIG4__", fig4_b64)
|
| 854 |
html_content = html_content.replace("__FIG5__", fig5_b64)
|
| 855 |
html_content = html_content.replace("__FIG6__", fig6_b64)
|
| 856 |
-
|
| 857 |
-
# Verify zero em dashes
|
| 858 |
-
em_dash_present = any(char in html_content for char in ["\u2014", "\u2013\u2013", "---"])
|
| 859 |
-
print(f"[Verification] Em dash present in source: {em_dash_present}")
|
| 860 |
-
if em_dash_present:
|
| 861 |
-
print("[Error] Em dashes detected! Removing...")
|
| 862 |
-
html_content = html_content.replace("\u2014", ": ").replace("---", ": ")
|
| 863 |
|
| 864 |
# Write raw HTML
|
| 865 |
HTML_RAW_PATH.write_text(html_content, encoding="utf-8")
|
|
|
|
| 1 |
#!/usr/bin/env python3
|
| 2 |
"""
|
| 3 |
scripts/generate_ieee_pdf.py
|
| 4 |
+
Compiles camera-ready IEEE Conference Paper PDF for Q-TensorFormer.
|
| 5 |
+
Strict adherence to official IEEE Conference Publishing standards:
|
| 6 |
+
- US Letter (8.5 x 11 in)
|
| 7 |
+
- Margins: Top 0.75 in (19 mm), Bottom 1.0 in (25.4 mm), Left/Right 0.625 in (15.9 mm)
|
| 8 |
+
- Column Gutter: 0.25 in (18 pt)
|
| 9 |
+
- Times New Roman body font (10 pt, justified, single spaced)
|
| 10 |
+
- Centered 24 pt regular Title Case paper title
|
| 11 |
+
- Author block: Dr. Madala Guru Brahmam (1st) & V C Premchand Yadav (2nd)
|
| 12 |
+
- Abstract (single paragraph, 150-250 words) with Abstract- and Index Terms- labels
|
| 13 |
+
- Primary headings: Roman numerals, small caps, centered, 10 pt
|
| 14 |
+
- Subheadings: Capital letters, italic, left-aligned, 10 pt
|
| 15 |
+
- Multi-line broken equations to completely eliminate equation overlap
|
| 16 |
+
- Equation table layout: centered math with flush-right parenthetical numbers (1)
|
| 17 |
+
- 7 comprehensive scientific figures embedded as base64 images
|
| 18 |
+
- No printed dates, page numbers, or URLs in headers/footers (--no-pdf-header-footer)
|
| 19 |
"""
|
| 20 |
|
| 21 |
import base64
|
|
|
|
| 41 |
return ""
|
| 42 |
|
| 43 |
fig1_b64 = get_base64_image("figure1_architecture.png")
|
| 44 |
+
fig2_b64 = get_base64_image("figure16_baseline_improvement_radar.png")
|
| 45 |
+
fig3_b64 = get_base64_image("figure12_hardware_roofline.png")
|
| 46 |
+
fig4_b64 = get_base64_image("figure11_subsystem_latency_breakdown.png")
|
| 47 |
+
fig5_b64 = get_base64_image("figure4_nested_tt_suboptimality.png")
|
| 48 |
+
fig6_b64 = get_base64_image("figure8_controller_convergence.png")
|
| 49 |
+
fig7_b64 = get_base64_image("figure7_kv_rate_distortion.png")
|
| 50 |
+
|
| 51 |
+
raw_template = r'''<!DOCTYPE html>
|
| 52 |
<html lang="en">
|
| 53 |
<head>
|
| 54 |
<meta charset="utf-8">
|
|
|
|
| 58 |
|
| 59 |
@page {
|
| 60 |
size: letter;
|
| 61 |
+
margin-top: 0.75in;
|
| 62 |
+
margin-bottom: 1.00in;
|
| 63 |
+
margin-left: 0.625in;
|
| 64 |
+
margin-right: 0.625in;
|
| 65 |
}
|
| 66 |
|
| 67 |
body {
|
| 68 |
font-family: "Times New Roman", Times, "Nimbus Roman No9 L", serif;
|
| 69 |
font-size: 10pt;
|
| 70 |
+
line-height: 1.16;
|
| 71 |
color: #000;
|
| 72 |
margin: 0;
|
| 73 |
padding: 0;
|
|
|
|
| 76 |
print-color-adjust: exact;
|
| 77 |
}
|
| 78 |
|
| 79 |
+
/* Metadata Title & Authors */
|
| 80 |
.title-container {
|
| 81 |
text-align: center;
|
| 82 |
+
margin-bottom: 14pt;
|
| 83 |
+
padding-bottom: 4pt;
|
|
|
|
| 84 |
}
|
| 85 |
|
| 86 |
.paper-title {
|
| 87 |
+
font-size: 24pt;
|
| 88 |
+
font-weight: normal;
|
| 89 |
+
line-height: 1.15;
|
| 90 |
+
margin-bottom: 14pt;
|
| 91 |
letter-spacing: -0.2px;
|
| 92 |
}
|
| 93 |
|
| 94 |
.author-block {
|
| 95 |
display: flex;
|
| 96 |
justify-content: center;
|
| 97 |
+
gap: 40pt;
|
| 98 |
+
margin-bottom: 12pt;
|
| 99 |
}
|
| 100 |
|
| 101 |
.author-col {
|
| 102 |
text-align: center;
|
| 103 |
+
max-width: 250pt;
|
| 104 |
}
|
| 105 |
|
| 106 |
.author-name {
|
| 107 |
+
font-size: 11pt;
|
| 108 |
+
font-weight: normal;
|
| 109 |
margin-bottom: 2pt;
|
| 110 |
}
|
| 111 |
|
| 112 |
.author-affil {
|
| 113 |
+
font-size: 10pt;
|
| 114 |
font-style: italic;
|
| 115 |
+
line-height: 1.20;
|
| 116 |
}
|
| 117 |
|
| 118 |
.author-email {
|
| 119 |
+
font-size: 9pt;
|
| 120 |
font-family: "Courier New", monospace;
|
| 121 |
+
font-style: normal;
|
| 122 |
margin-top: 3pt;
|
| 123 |
+
color: #000;
|
| 124 |
}
|
| 125 |
|
| 126 |
+
/* Strict Two-Column Layout */
|
| 127 |
.two-col {
|
| 128 |
column-count: 2;
|
| 129 |
+
column-gap: 18pt; /* 0.25 in */
|
| 130 |
text-align: justify;
|
| 131 |
+
column-fill: balance;
|
| 132 |
}
|
| 133 |
|
| 134 |
+
/* Abstract and Keywords */
|
| 135 |
.abstract-box {
|
| 136 |
+
margin-bottom: 6pt;
|
| 137 |
+
font-size: 10pt;
|
| 138 |
+
line-height: 1.16;
|
| 139 |
+
text-align: justify;
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
}
|
| 141 |
|
| 142 |
.keywords-box {
|
| 143 |
+
margin-bottom: 10pt;
|
| 144 |
+
font-size: 10pt;
|
| 145 |
+
line-height: 1.16;
|
| 146 |
+
text-align: justify;
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
}
|
| 148 |
|
| 149 |
+
/* Heading Hierarchy */
|
| 150 |
h2 {
|
| 151 |
+
font-size: 10pt;
|
| 152 |
+
font-weight: normal;
|
| 153 |
+
font-variant: small-caps;
|
| 154 |
text-transform: uppercase;
|
| 155 |
text-align: center;
|
| 156 |
+
margin-top: 10pt;
|
| 157 |
+
margin-bottom: 4pt;
|
| 158 |
+
letter-spacing: 0.5px;
|
| 159 |
break-after: avoid;
|
|
|
|
|
|
|
| 160 |
}
|
| 161 |
|
| 162 |
h3 {
|
| 163 |
font-size: 10pt;
|
| 164 |
+
font-weight: normal;
|
| 165 |
font-style: italic;
|
| 166 |
+
text-align: left;
|
| 167 |
margin-top: 9pt;
|
| 168 |
margin-bottom: 3pt;
|
| 169 |
break-after: avoid;
|
| 170 |
}
|
| 171 |
|
| 172 |
+
.sub-subhead {
|
| 173 |
+
font-style: italic;
|
| 174 |
+
text-indent: 1.2em;
|
| 175 |
+
display: inline;
|
| 176 |
+
}
|
| 177 |
+
|
| 178 |
p {
|
| 179 |
margin-top: 0;
|
| 180 |
+
margin-bottom: 3.5pt;
|
| 181 |
text-indent: 1.2em;
|
| 182 |
}
|
| 183 |
|
|
|
|
| 185 |
text-indent: 0;
|
| 186 |
}
|
| 187 |
|
| 188 |
+
/* Mathematics & Equation Formatting */
|
| 189 |
+
.eq-table {
|
| 190 |
+
width: 100%;
|
| 191 |
+
border-collapse: collapse;
|
| 192 |
+
margin: 5pt 0;
|
| 193 |
+
break-inside: avoid;
|
| 194 |
+
}
|
| 195 |
+
|
| 196 |
+
.eq-cell-math {
|
| 197 |
+
text-align: center;
|
| 198 |
+
vertical-align: middle;
|
| 199 |
+
width: 98%;
|
| 200 |
+
padding: 0;
|
| 201 |
+
font-size: 9.5pt;
|
| 202 |
+
}
|
| 203 |
+
|
| 204 |
+
.eq-cell-num {
|
| 205 |
+
text-align: right;
|
| 206 |
+
vertical-align: middle;
|
| 207 |
+
white-space: nowrap;
|
| 208 |
+
font-family: "Times New Roman", Times, serif;
|
| 209 |
+
font-size: 10pt;
|
| 210 |
+
width: 2%;
|
| 211 |
+
padding-left: 6pt;
|
| 212 |
+
}
|
| 213 |
+
|
| 214 |
+
/* Theorems & Proofs */
|
| 215 |
.theorem-box {
|
| 216 |
+
margin: 6pt 0;
|
| 217 |
+
padding: 3pt 5pt;
|
| 218 |
font-style: italic;
|
| 219 |
+
border-left: 2pt solid #000;
|
| 220 |
+
background: #fafafa;
|
| 221 |
break-inside: avoid;
|
| 222 |
+
font-size: 9.5pt;
|
| 223 |
}
|
| 224 |
|
| 225 |
.theorem-title {
|
|
|
|
| 229 |
|
| 230 |
.proof {
|
| 231 |
font-style: normal;
|
| 232 |
+
font-size: 9.5pt;
|
| 233 |
+
margin: 2pt 0 5pt 0;
|
| 234 |
text-indent: 1.2em;
|
| 235 |
}
|
| 236 |
|
|
|
|
| 239 |
font-weight: bold;
|
| 240 |
}
|
| 241 |
|
| 242 |
+
/* Visual Elements: Tables */
|
| 243 |
+
table.data-table {
|
| 244 |
width: 100%;
|
| 245 |
border-collapse: collapse;
|
| 246 |
+
font-size: 8pt;
|
| 247 |
+
margin: 6pt 0;
|
| 248 |
+
line-height: 1.15;
|
| 249 |
break-inside: avoid;
|
| 250 |
}
|
| 251 |
|
| 252 |
+
table.data-table th, table.data-table td {
|
| 253 |
+
padding: 2.5pt 3pt;
|
| 254 |
text-align: left;
|
| 255 |
}
|
| 256 |
|
| 257 |
+
table.data-table th {
|
| 258 |
font-weight: bold;
|
| 259 |
+
border-top: 1.2pt solid #000;
|
| 260 |
+
border-bottom: 0.6pt solid #000;
|
| 261 |
+
background: #f8f8f8;
|
| 262 |
}
|
| 263 |
|
| 264 |
+
table.data-table td {
|
| 265 |
border-bottom: 0.4pt solid #e0e0e0;
|
| 266 |
}
|
| 267 |
|
| 268 |
.table-bottom {
|
| 269 |
+
border-bottom: 1.2pt solid #000;
|
| 270 |
}
|
| 271 |
|
| 272 |
.table-caption {
|
| 273 |
+
font-size: 8pt;
|
| 274 |
text-align: center;
|
| 275 |
+
font-variant: small-caps;
|
| 276 |
+
margin-bottom: 2pt;
|
| 277 |
+
letter-spacing: 0.4px;
|
| 278 |
+
break-after: avoid;
|
| 279 |
}
|
| 280 |
|
| 281 |
.highlight-row {
|
| 282 |
+
background-color: #f2f7fc;
|
| 283 |
font-weight: bold;
|
| 284 |
}
|
| 285 |
|
| 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: 85pt;
|
| 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: 1.15;
|
| 306 |
+
margin-top: 3pt;
|
| 307 |
}
|
| 308 |
|
| 309 |
.figure-caption-title {
|
| 310 |
+
font-weight: normal;
|
| 311 |
}
|
| 312 |
|
| 313 |
+
/* Algorithms */
|
| 314 |
.algorithm-box {
|
| 315 |
+
border-top: 1.2pt solid #000;
|
| 316 |
+
border-bottom: 1.2pt solid #000;
|
| 317 |
+
padding: 4pt 2pt;
|
| 318 |
+
font-size: 8pt;
|
| 319 |
font-family: "Courier New", monospace;
|
| 320 |
+
margin: 6pt 0;
|
| 321 |
+
line-height: 1.20;
|
| 322 |
+
background: #fcfcfc;
|
| 323 |
break-inside: avoid;
|
| 324 |
}
|
| 325 |
|
| 326 |
.algo-header {
|
| 327 |
font-family: "Times New Roman", Times, serif;
|
| 328 |
font-weight: bold;
|
| 329 |
+
font-size: 8.5pt;
|
| 330 |
+
margin-bottom: 2pt;
|
| 331 |
+
border-bottom: 0.5pt solid #000;
|
| 332 |
padding-bottom: 2pt;
|
| 333 |
}
|
| 334 |
|
| 335 |
+
/* References */
|
| 336 |
+
.references-heading {
|
| 337 |
+
font-size: 10pt;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 338 |
font-weight: normal;
|
| 339 |
+
font-variant: small-caps;
|
| 340 |
+
text-align: center;
|
| 341 |
+
margin-top: 10pt;
|
| 342 |
+
margin-bottom: 6pt;
|
| 343 |
}
|
| 344 |
|
| 345 |
.references {
|
| 346 |
+
font-size: 8pt;
|
| 347 |
+
line-height: 1.15;
|
| 348 |
}
|
| 349 |
|
| 350 |
.ref-item {
|
| 351 |
+
margin-bottom: 3pt;
|
| 352 |
+
text-indent: -1.3em;
|
| 353 |
+
margin-left: 1.3em;
|
| 354 |
}
|
| 355 |
|
| 356 |
.perf-badge {
|
| 357 |
display: inline-block;
|
| 358 |
+
padding: 0.5pt 2.5pt;
|
| 359 |
font-size: 7.5pt;
|
| 360 |
font-weight: bold;
|
| 361 |
color: #0b6623;
|
| 362 |
background: #eaf8ec;
|
| 363 |
+
border-radius: 1.5pt;
|
| 364 |
}
|
| 365 |
</style>
|
| 366 |
</head>
|
|
|
|
| 371 |
<div class="author-block">
|
| 372 |
<div class="author-col">
|
| 373 |
<div class="author-name">Dr. Madala Guru Brahmam</div>
|
| 374 |
+
<div class="author-affil">Department of Computer Science and Engineering (AIML)<br>School of Computer Science and Engineering (SCOPE)<br>VIT-AP University, Amaravati, Andhra Pradesh, India</div>
|
| 375 |
<div class="author-email">madala.gurubrahmam@vitap.ac.in</div>
|
| 376 |
</div>
|
| 377 |
<div class="author-col">
|
| 378 |
<div class="author-name">V C Premchand Yadav</div>
|
| 379 |
+
<div class="author-affil">Department of Computer Science and Engineering (AIML)<br>School of Computer Science and Engineering (SCOPE)<br>VIT-AP University, Amaravati, Andhra Pradesh, India</div>
|
| 380 |
<div class="author-email">premchandyadav.vc21bce9124@vitap.ac.in</div>
|
| 381 |
</div>
|
| 382 |
</div>
|
|
|
|
| 385 |
<div class="two-col">
|
| 386 |
|
| 387 |
<div class="abstract-box">
|
| 388 |
+
<b><i>Abstract</i>—Deploying autoregressive vision-language foundation models on resource-constrained edge microarchitectures (mobile SoCs, robotics platforms, embedded NPUs) is severely impeded by memory bus saturation, thermal throttling, and unpredictable dynamic execution budgets. Existing compression paradigms, including post-training quantization (AWQ, SmoothQuant), structural pruning, and static low-rank adaptation, permanently fix parameter geometry offline. They lack the plasticity to scale throughput dynamically during real-time hardware stress or burst reasoning demands. In this paper, we present Q-TensorFormer, a closed-loop cyber-physical architecture delivering real-time, high-throughput edge LLM inference. Q-TensorFormer establishes five unified innovations: (1) nested Tensor-Train parameterization enabling instant sub-15-microsecond active rank slicing ($r \in \{1, 2, 4, 8, 16\}$) with zero runtime singular value decomposition (Zero-SVD) and zero memory reallocation; (2) a discrete-time Lyapunov-stable dual-subgradient controller with analytical hysteresis deadband ($\tau = 0.15$), proven to eliminate limit-cycle chattering ($\Delta V_t \le 0$); (3) an entropy-gated semantic anchor KV-cache slashing memory footprint by up to $93.36\%$ with guaranteed $100\%$ precision retention on critical needle tokens; (4) an in-register fused contraction kernel eliminating $96.90\%$ of intermediate DRAM traffic; and (5) a LinUCB contextual bandit token scheduler with proven sub-linear regret $\mathcal{O}(d \sqrt{T \ln(T/\delta)})$. Across 50 empirical trials on edge hardware, Q-TensorFormer delivers a 4.8x throughput speedup on Raspberry Pi 5 (18.2 vs 3.8 tok/s) and 3.9x on NVIDIA Jetson Orin Nano (94.7 vs 24.1 tok/s), cuts peak RAM by 72.8%, and switches rank in 14.2 μs (9,042x faster than 128 ms SVD re-factorization) with zero accuracy collapse ($p < 10^{-15}$, Cohen's $d = 1.354$).</b>
|
| 389 |
</div>
|
| 390 |
|
| 391 |
<div class="keywords-box">
|
| 392 |
+
<b><i>Index Terms</i>—Closed-loop control, Contextual bandits, Edge artificial intelligence, Foundation models, Lyapunov stability, Tensor-Train decomposition.</b>
|
| 393 |
</div>
|
| 394 |
|
| 395 |
<h2>I. Introduction</h2>
|
| 396 |
+
<p class="no-indent">Decentralized on-device artificial intelligence requires deploying multi-billion-parameter Transformer foundation models directly on constrained edge devices, including mobile systems-on-chip, autonomous robotic units, and battery-powered embedded devices [1], [2], [3]. Edge execution eliminates round-trip cloud latency, preserves data privacy, and functions reliably in disconnected environments. However, deploying autoregressive foundation models on edge microarchitectures exposes two critical physical walls:</p>
|
| 397 |
|
| 398 |
+
<p><i>1) The Edge Memory Bandwidth Wall:</i> Autoregressive decoding is inherently memory-bound. Emitting each single token demands transferring gigabytes of weight tensors and accumulating Key-Value (KV) cache states from external Dynamic RAM (DRAM) into on-chip cache lines. Narrow edge buses (16-bit to 64-bit wide LPDDR4X/LPDDR5 channels) quickly saturate, throttling arithmetic intensity below $0.05$ FLOP/Byte [4], [5], [6].</p>
|
| 399 |
|
| 400 |
+
<p><i>2) Rigid Static Model Geometries:</i> Mainstream compression techniques, including 4-bit weight quantization (AWQ [7], SmoothQuant [8]), static low-rank adaptation (LoRA [2]), and 1-bit representations (BitNet b1.58 [9]), permanently freeze parameter shapes at compilation. When real-time operating conditions deteriorate, such as processor junction temperature spiking above safe thermal limits ($T_j > 95^\circ\text{C}$), battery state-of-charge decaying, or concurrent threads starving DRAM bandwidth, static models suffer severe thermal throttling, causing frame drops and system lockups [10], [11].</p>
|
| 401 |
|
| 402 |
<p>To overcome these foundational limitations, we introduce <b>Q-TensorFormer</b>. Rather than relying on heuristic early-exit shortcuts [12] or uncalibrated token skipping [13], Q-TensorFormer formulates edge inference as a closed-loop cyber-physical control problem. It harmonizes algebraic Tensor-Train core nesting with discrete-time Lyapunov stability and contextual bandit routing.</p>
|
| 403 |
|
| 404 |
<p><b>Primary Contributions of this Work:</b></p>
|
| 405 |
+
<p>• <i>Real-Time Zero-SVD Rank Slicing:</i> We introduce nested Tensor-Train cores where lower-rank cores reside continuously inside higher-rank cores. Modulating capacity requires simply shifting slice pointer offsets ($14.2\;\mu\text{s}$), completely eliminating expensive runtime singular value decompositions ($128.4\;\text{ms}$).</p>
|
| 406 |
+
<p>• <i>Provable Lyapunov-Stable Controller:</i> We formulate a dual-subgradient PID controller with an analytical hysteresis deadband ($\tau = 0.15$), proving discrete-time Lyapunov stability ($\Delta V_t \le 0$) and mathematically preventing high-frequency chattering under thermal and latency variations.</p>
|
| 407 |
+
<p>• <i>Rigorous Theoretical Foundations:</i> We establish a comprehensive quad-theorem suite bounding TT truncation error, dynamical stability, semantic KV retention, and online bandit regret.</p>
|
| 408 |
+
<p>• <i>Hardware-Fused In-Register Contraction:</i> We design an in-register block-tiled contraction kernel cutting intermediate DRAM traffic by $96.90\%$ and boosting operational arithmetic intensity by $32.27\times$.</p>
|
| 409 |
+
<p>• <i>Extensive Edge Hardware Validation:</i> We benchmark Q-TensorFormer across 50 rigorous trials on Raspberry Pi 5, NVIDIA Jetson Orin Nano, Apple M-series, and Intel Core Ultra edge SoCs, demonstrating up to <b>4.8x higher throughput</b>, <b>72.8% RAM reduction</b>, and <b>9,042x faster dynamic adaptation</b>.</p>
|
| 410 |
|
| 411 |
<div class="figure-card">
|
| 412 |
<img src="__FIG1__" alt="Q-TensorFormer Architecture">
|
| 413 |
+
<div class="figure-caption">Fig. 1. System architecture of Q-TensorFormer. Shows the closed-loop cyber-physical control pipeline: in-register fused Tensor-Train core contraction, LinUCB contextual bandit token scheduling, semantic anchor KV-cache compression, and discrete-time Lyapunov rank adaptation.</div>
|
| 414 |
</div>
|
| 415 |
|
| 416 |
+
<h2>II. System Architecture</h2>
|
| 417 |
|
| 418 |
<h3>A. Problem Formulation: Constrained Markov Decision Process</h3>
|
| 419 |
+
<p class="no-indent">We formulate edge foundation inference as a Constrained Markov Decision Process (CMDP) over discrete decoding steps $t \in \{1, 2, \dots, T\}$. Let $\mathbf{s}_t \in \mathcal{S}$ denote the edge hardware state vector, encompassing instantaneous memory bandwidth consumption, junction temperature $T_j(t)$, cache hit ratios, and prompt context complexity. The system policy $\pi(\mathbf{a}_t | \mathbf{s}_t)$ selects an action tuple $\mathbf{a}_t = (r_t, m_t, k_t)$, where $r_t \in \mathcal{R} = \{1, 2, 4, 8, 16\}$ is the active Tensor-Train rank, $m_t \in \{0, 1\}$ is the compute grid gating mask, and $k_t$ is the KV-cache anchor compression ratio.</p>
|
| 420 |
|
| 421 |
<p>The objective is to minimize cumulative task loss subject to physical hardware constraints:</p>
|
| 422 |
|
| 423 |
+
$$\begin{aligned}
|
| 424 |
+
\min_{\pi} \;& \mathbb{E}_{\pi} \left[ \sum_{t=1}^T \mathcal{L}_{\text{task}}(\mathbf{y}_t, \hat{\mathbf{y}}_t) \right] \\
|
| 425 |
+
\text{s.t.} \;& \mathbb{E}[C_k(\mathbf{s}_t, \mathbf{a}_t)] \le B_k, \quad \forall k
|
| 426 |
+
\end{aligned}$$
|
| 427 |
|
| 428 |
<p class="no-indent">where $C_k$ represents physical resource costs (DRAM memory traffic, thermal dissipation, latency deadline) and $B_k$ denotes the specified budget ceiling.</p>
|
| 429 |
|
| 430 |
<h3>B. Zero-SVD Nested Tensor-Train Parameterization</h3>
|
| 431 |
+
<p class="no-indent">Let $\mathbf{W} \in \mathbb{R}^{M \times N}$ denote a dense linear projection weight (e.g., in multi-head self-attention or feed-forward network). We reshape $\mathbf{W}$ into high-dimensional tensors factorized via Tensor-Train (TT) decomposition [3], [14]:</p>
|
| 432 |
|
| 433 |
+
$$\begin{aligned}
|
| 434 |
+
\mathbf{W}_{(i_1 \dots i_d), (j_1 \dots j_d)} &= \mathcal{G}^{(1)}[:, i_1, j_1, :] \cdot \mathcal{G}^{(2)}[:, i_2, j_2, :] \\
|
| 435 |
+
&\quad \cdots \mathcal{G}^{(d)}[:, i_d, j_d, :]
|
| 436 |
+
\end{aligned}$$
|
| 437 |
|
| 438 |
+
<p class="no-indent">where $\mathcal{G}^{(k)} \in \mathbb{R}^{r_{k-1} \times m_k \times n_k \times r_k}$ are 4-way TT-cores. Under standard low-rank methods, altering rank $r$ requires dynamic matrix re-factorization or re-allocating weight memory buffers, introducing catastrophic $100\text{ to }150\;\text{ms}$ pauses per token.</p>
|
| 439 |
|
| 440 |
+
<p>In Q-TensorFormer, we parameterize cores such that lower-rank sub-tensors are nested directly inside the maximum rank allocation $r_{\max}$:</p>
|
| 441 |
|
| 442 |
+
$$\mathcal{G}^{(k)}_{\text{active}} = \mathcal{G}^{(k)}[0:r_{\text{curr}}, \; :, \; :, \; 0:r_{\text{curr}}]$$
|
| 443 |
|
| 444 |
<p class="no-indent">Dynamic rank adjustment requires zero memory allocations, zero tensor copies, and zero singular value decompositions. The runtime execution kernel merely modifies loop trip-count bounds in processor registers, switching active rank in <b>14.2 μs</b>.</p>
|
| 445 |
|
| 446 |
<div class="figure-card">
|
| 447 |
+
<img src="__FIG2__" alt="Improvement Radar">
|
| 448 |
+
<div class="figure-caption">Fig. 2. Multi-objective improvement radar chart comparing Q-TensorFormer against Dense FP16, AWQ 4-Bit, and MobileLLM across throughput, memory compression, dynamic rank adaptation speed, thermal tolerance, and reasoning retention.</div>
|
| 449 |
</div>
|
| 450 |
|
| 451 |
<div class="algorithm-box">
|
|
|
|
| 463 |
<h3>C. Discrete-Time Lyapunov-Stable Controller</h3>
|
| 464 |
<p class="no-indent">Dynamic rank adjustment without feedback stabilization can induce limit-cycle oscillations (chattering), where the model continuously flips between high and low ranks, destabilizing memory buses and degrading output coherence. Q-TensorFormer stabilizes rank transitions using a dual-subgradient controller with analytical hysteresis:</p>
|
| 465 |
|
| 466 |
+
$$e_t = L_t - L_{\text{target}}$$
|
| 467 |
|
| 468 |
+
$$u_t = e_t + \kappa_D (e_t - e_{t-1})$$
|
| 469 |
|
| 470 |
+
$$\lambda_{t+1} = \Pi_{[\lambda_{\min}, \lambda_{\max}]} \left[ \lambda_t + \eta \cdot \phi_\tau(u_t) \right]$$
|
| 471 |
|
| 472 |
+
<p class="no-indent">where $\phi_\tau(u)$ is a deadband hysteresis operator defined as:</p>
|
| 473 |
|
| 474 |
+
$$\phi_\tau(u) = \begin{cases} u - \tau, & u > \tau \\ 0, & |u| \le \tau \\ u + \tau, & u < -\tau \end{cases}$$
|
| 475 |
|
| 476 |
+
<p class="no-indent">The active rank is mapped via $r_t = g(\lambda_t)$. The deadband threshold $\tau = 0.15$ ensures that minor stochastic latency fluctuations do not trigger parameter re-slicing.</p>
|
| 477 |
|
| 478 |
<div class="figure-card">
|
| 479 |
+
<img src="__FIG3__" alt="Roofline Model">
|
| 480 |
+
<div class="figure-caption">Fig. 3. Hardware roofline model across edge microarchitectures. Illustrates how in-register fused TT contraction moves the operational point from severe memory-bandwidth saturation ($0.048$ FLOP/Byte) toward peak processor arithmetic intensity ($1.549$ FLOP/Byte).</div>
|
| 481 |
</div>
|
| 482 |
|
| 483 |
<div class="algorithm-box">
|
|
|
|
| 495 |
9: <b>return</b> $r_{t+1}$ (provably stable: $\Delta V_t \le 0$)
|
| 496 |
</div>
|
| 497 |
|
| 498 |
+
<div class="figure-card">
|
| 499 |
+
<img src="__FIG4__" alt="Subsystem Latency Breakdown">
|
| 500 |
+
<div class="figure-caption">Fig. 4. Subsystem latency breakdown across generation steps. Fused register contractions require only $14.2\;\mu\text{s}$, while eliminating DRAM transfer stalls.</div>
|
| 501 |
+
</div>
|
| 502 |
+
|
| 503 |
+
<h2>III. Theoretical Analysis</h2>
|
| 504 |
|
| 505 |
<div class="theorem-box">
|
| 506 |
+
<span class="theorem-title">Theorem 1 (Nested TT-Core Truncation Error Bound):</span> Let $\mathbf{W}$ be a tensor represented by TT-cores $\{\mathcal{G}^{(k)}\}_{k=1}^d$ with maximal ranks $R_k$. Let $\mathbf{W}(r)$ denote the nested slice with uniform active rank $r \le \min_k R_k$. The Frobenius norm truncation error satisfies:
|
| 507 |
+
$$\begin{aligned}
|
| 508 |
+
\|\mathbf{W} - \mathbf{W}(r)\|_F &\le \sum_{k=1}^{d-1} \|\mathcal{G}^{(k)}_{>r}\|_F \\
|
| 509 |
+
&\quad \times \prod_{j \ne k} \|\mathcal{G}^{(j)}\|_F
|
| 510 |
+
\end{aligned}$$
|
| 511 |
+
where $\mathcal{G}^{(k)}_{>r}$ denotes the complementary sub-tensor slice outside rank $r$.
|
| 512 |
</div>
|
| 513 |
|
| 514 |
+
<p class="proof"><i>Proof Sketch:</i> By telescoping difference over TT factor contractions $\mathbf{W} - \mathbf{W}(r) = \sum_{k=1}^{d-1} (\mathcal{T}_k - \mathcal{T}_{k-1})$, where $\mathcal{T}_k$ contracts sliced cores up to index $k$ and full cores thereafter. Applying sub-multiplicativity of Frobenius norms across core tensor contractions yields the bounded summation. <span class="proof-qed">■</span></p>
|
| 515 |
|
| 516 |
<div class="theorem-box">
|
| 517 |
+
<span class="theorem-title">Theorem 2 (Discrete-Time Lyapunov Stability & Anti-Chattering):</span> Consider the Lyapunov candidate $V_t = \frac{1}{2} (\lambda_t - \lambda^*)^2$. For step-size $\eta < \frac{2}{L_f}$ and hysteresis threshold $\tau > \frac{\sigma_{\text{noise}}}{1 - \rho}$, the sequence $\{V_t\}$ satisfies:
|
| 518 |
+
$$\Delta V_t = V_{t+1} - V_t \le -\alpha \|e_t\|^2, \quad \forall |u_t| > \tau$$
|
| 519 |
+
where $\alpha > 0$. The closed-loop system is asymptotically stable and eliminates limit-cycle chattering.
|
| 520 |
</div>
|
| 521 |
|
| 522 |
+
<p class="proof"><i>Proof Sketch:</i> Evaluating forward difference $\Delta V_t = (\lambda_{t+1} - \lambda_t)(\lambda_t - \lambda^*) + \frac{1}{2}(\lambda_{t+1} - \lambda_t)^2$. Substituting the deadband gradient update and applying strict monotonicity of the cost objective confirms negative definiteness outside the deadband $|u_t| > \tau$. Inside the deadband, $\phi_\tau(u_t) = 0$, producing $\Delta V_t = 0$ with zero parameter churn. <span class="proof-qed">■</span></p>
|
| 523 |
|
| 524 |
<div class="figure-card">
|
| 525 |
+
<img src="__FIG5__" alt="Nested TT Suboptimality">
|
| 526 |
+
<div class="figure-caption">Fig. 5. Nested TT truncation suboptimality gap across ranks $r \in \{1, 2, 4, 8, 16\}$. Demonstrates that joint multi-rank gradient descent preserves over $98.6\%$ of full-rank singular value energy within lower-rank sub-cores.</div>
|
| 527 |
</div>
|
| 528 |
|
| 529 |
<div class="theorem-box">
|
| 530 |
+
<span class="theorem-title">Theorem 3 (Entropy-Gated KV Information Retention):</span> Let query-key attention distribution be $p_i = \text{softmax}(q k_i^T / \sqrt{d_k})$. If keys with attention entropy surprise $\mathcal{H}(k_i) \ge \gamma$ are retained as exact uncompressed anchors while remaining keys are compressed to 2-bit representations, the total variational distance between true attention output $\mathbf{o}$ and compressed output $\hat{\mathbf{o}}$ is bounded by:
|
| 531 |
+
$$\|\mathbf{o} - \hat{\mathbf{o}}\|_1 \le 2 \sum_{i \notin \mathcal{A}} p_i \le 2 \exp(-\gamma)$$
|
| 532 |
</div>
|
| 533 |
|
| 534 |
<div class="theorem-box">
|
| 535 |
+
<span class="theorem-title">Theorem 4 (LinUCB Contextual Bandit Regret Bound):</span> Under feature dimension $d$ and confidence parameter $\delta \in (0, 1)$, the cumulative regret of the Q-TensorFormer token routing bandit over $T$ generation steps satisfies with probability at least $1 - \delta$:
|
| 536 |
+
$$\begin{aligned}
|
| 537 |
+
\mathcal{R}(T) &\le 2 \sqrt{2 d T \ln(1 + T / \lambda)} \\
|
| 538 |
+
&\quad \times \left( \sqrt{\lambda} S + \sigma \sqrt{2 \ln(1/\delta) + d \ln\left(1 + \frac{T}{d \lambda}\right)} \right) \\
|
| 539 |
+
&= \mathcal{O}\left(d \sqrt{T \ln(T / \delta)}\right)
|
| 540 |
+
\end{aligned}$$
|
| 541 |
+
</div>
|
| 542 |
+
|
| 543 |
+
<div class="figure-card">
|
| 544 |
+
<img src="__FIG6__" alt="Lyapunov Convergence">
|
| 545 |
+
<div class="figure-caption">Fig. 6. Lyapunov stability phase portrait and controller error convergence under non-stationary thermal shock. Demonstrates negative drift $\Delta V_t \le 0$ and rapid convergence to equilibrium without limit-cycle chattering.</div>
|
| 546 |
</div>
|
| 547 |
|
| 548 |
<h2>IV. Quantitative Experimental Evaluation</h2>
|
|
|
|
| 555 |
<p>• <b>Apple M3 Edge SoC:</b> 8-core CPU / 10-core GPU with 16-core Neural Engine, unified 100 GB/s memory architecture.</p>
|
| 556 |
<p>• <b>Intel Core Ultra 7 (Meteor Lake):</b> Intel 4 process, integrated Arc GPU + Intel AI Boost NPU, 32 GB LPDDR5X-7467.</p>
|
| 557 |
|
| 558 |
+
<div class="table-caption">TABLE I: REAL-TIME EDGE INFERENCE THROUGHPUT, MEMORY FOOTPRINT, AND LATENCY ACROSS PLATFORMS</div>
|
| 559 |
+
<table class="data-table">
|
| 560 |
<thead>
|
| 561 |
<tr>
|
| 562 |
<th>Platform & Model Architecture</th>
|
| 563 |
+
<th>Params</th>
|
| 564 |
+
<th>Peak RAM</th>
|
| 565 |
+
<th>Throughput</th>
|
| 566 |
<th>Speedup</th>
|
| 567 |
<th>Adaptation Latency</th>
|
| 568 |
<th>Thermal Drop (15 min)</th>
|
|
|
|
| 570 |
</thead>
|
| 571 |
<tbody>
|
| 572 |
<tr>
|
| 573 |
+
<td colspan="7" style="background:#f0f0f0; font-weight:bold; font-size:7.5pt;">Raspberry Pi 5 (ARM Cortex-A76, 17.0 GB/s DRAM Bandwidth)</td>
|
| 574 |
</tr>
|
| 575 |
<tr>
|
| 576 |
<td>Dense Baseline (FP16 TinyLlama) [11]</td>
|
| 577 |
+
<td>1,100M</td>
|
| 578 |
<td>2,180 MB</td>
|
| 579 |
<td>3.8 tok/s</td>
|
| 580 |
<td>1.0x</td>
|
|
|
|
| 583 |
</tr>
|
| 584 |
<tr>
|
| 585 |
<td>AWQ 4-Bit Baseline [7]</td>
|
| 586 |
+
<td>1,100M</td>
|
| 587 |
<td>840 MB</td>
|
| 588 |
<td>7.2 tok/s</td>
|
| 589 |
<td>1.9x</td>
|
|
|
|
| 592 |
</tr>
|
| 593 |
<tr>
|
| 594 |
<td>MobileLLM-1.5B (ICLR 2024) [4]</td>
|
| 595 |
+
<td>1,500M</td>
|
| 596 |
<td>1,420 MB</td>
|
| 597 |
<td>4.6 tok/s</td>
|
| 598 |
<td>1.2x</td>
|
|
|
|
| 601 |
</tr>
|
| 602 |
<tr>
|
| 603 |
<td>Llama-3.2-1B Edge (Meta 2024) [1]</td>
|
| 604 |
+
<td>1,230M</td>
|
| 605 |
<td>1,280 MB</td>
|
| 606 |
<td>5.4 tok/s</td>
|
| 607 |
<td>1.4x</td>
|
|
|
|
| 610 |
</tr>
|
| 611 |
<tr>
|
| 612 |
<td>TensorGPT-TT Baseline (AAAI 2023) [14]</td>
|
| 613 |
+
<td>210M</td>
|
| 614 |
<td>720 MB</td>
|
| 615 |
<td>8.9 tok/s</td>
|
| 616 |
<td>2.3x</td>
|
|
|
|
| 619 |
</tr>
|
| 620 |
<tr class="highlight-row">
|
| 621 |
<td><b>Q-TensorFormer (Static Balanced)</b></td>
|
| 622 |
+
<td><b>200M</b></td>
|
| 623 |
<td><b>640 MB</b></td>
|
| 624 |
<td><b>14.6 tok/s</b></td>
|
| 625 |
<td><b>3.8x</b></td>
|
|
|
|
| 628 |
</tr>
|
| 629 |
<tr class="highlight-row">
|
| 630 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 631 |
+
<td><b>140–200M</b></td>
|
| 632 |
<td><b>593 MB</b></td>
|
| 633 |
<td><b>18.2 tok/s</b></td>
|
| 634 |
<td><span class="perf-badge">+380% (4.8x)</span></td>
|
|
|
|
| 636 |
<td><b>-2.1% (Stable)</b></td>
|
| 637 |
</tr>
|
| 638 |
<tr>
|
| 639 |
+
<td colspan="7" style="background:#f0f0f0; font-weight:bold; font-size:7.5pt;">NVIDIA Jetson Orin Nano (1024-Core Ampere GPU, 68.0 GB/s Bandwidth)</td>
|
| 640 |
</tr>
|
| 641 |
<tr>
|
| 642 |
<td>Dense Baseline (FP16)</td>
|
| 643 |
+
<td>1,100M</td>
|
| 644 |
<td>2,180 MB</td>
|
| 645 |
<td>24.1 tok/s</td>
|
| 646 |
<td>1.0x</td>
|
|
|
|
| 649 |
</tr>
|
| 650 |
<tr>
|
| 651 |
<td>AWQ 4-Bit Baseline [7]</td>
|
| 652 |
+
<td>1,100M</td>
|
| 653 |
<td>840 MB</td>
|
| 654 |
<td>48.6 tok/s</td>
|
| 655 |
<td>2.0x</td>
|
|
|
|
| 658 |
</tr>
|
| 659 |
<tr>
|
| 660 |
<td>MobileLLM-1.5B [4]</td>
|
| 661 |
+
<td>1,500M</td>
|
| 662 |
<td>1,420 MB</td>
|
| 663 |
<td>29.4 tok/s</td>
|
| 664 |
<td>1.2x</td>
|
|
|
|
| 667 |
</tr>
|
| 668 |
<tr class="highlight-row">
|
| 669 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 670 |
+
<td><b>140–200M</b></td>
|
| 671 |
<td><b>593 MB</b></td>
|
| 672 |
<td><b>94.7 tok/s</b></td>
|
| 673 |
<td><span class="perf-badge">+293% (3.9x)</span></td>
|
|
|
|
| 675 |
<td><b>-1.8% (Stable)</b></td>
|
| 676 |
</tr>
|
| 677 |
<tr>
|
| 678 |
+
<td colspan="7" style="background:#f0f0f0; font-weight:bold; font-size:7.5pt;">Apple M3 Edge SoC (Unified 100 GB/s Memory Architecture)</td>
|
| 679 |
</tr>
|
| 680 |
<tr>
|
| 681 |
<td>Dense Baseline (FP16)</td>
|
| 682 |
+
<td>1,100M</td>
|
| 683 |
<td>2,180 MB</td>
|
| 684 |
<td>46.2 tok/s</td>
|
| 685 |
<td>1.0x</td>
|
|
|
|
| 688 |
</tr>
|
| 689 |
<tr>
|
| 690 |
<td>Llama-3.2-1B Edge (Meta 2024) [1]</td>
|
| 691 |
+
<td>1,230M</td>
|
| 692 |
<td>1,280 MB</td>
|
| 693 |
<td>62.8 tok/s</td>
|
| 694 |
<td>1.4x</td>
|
|
|
|
| 697 |
</tr>
|
| 698 |
<tr class="highlight-row">
|
| 699 |
<td><b>Q-TensorFormer (Lyapunov Slicing)</b></td>
|
| 700 |
+
<td><b>140–200M</b></td>
|
| 701 |
<td><b>593 MB</b></td>
|
| 702 |
<td><b>164.5 tok/s</b></td>
|
| 703 |
<td><span class="perf-badge">+256% (3.6x)</span></td>
|
|
|
|
| 708 |
</table>
|
| 709 |
|
| 710 |
<p class="no-indent">As shown in Table I, Q-TensorFormer achieves unmatched inference efficiency:</p>
|
| 711 |
+
<p>• <i>Throughput Superiority:</i> On the bandwidth-constrained Raspberry Pi 5, Q-TensorFormer delivers <b>18.2 tok/s</b> compared to 3.8 tok/s for the dense model and 7.2 tok/s for AWQ 4-bit, representing a <b>4.8x acceleration (+380% gain)</b>.</p>
|
| 712 |
+
<p>• <i>Extreme Memory Footprint Reduction:</i> Peak memory drops from 2,180 MB to <b>593 MB</b>, an overall <b>72.8% reduction</b>, allowing comfortable execution alongside operating system tasks on 1 GB embedded RAM partitions.</p>
|
| 713 |
+
<p>• <i>Instant Sub-15 μs Dynamic Adaptation:</i> Whereas classical low-rank methods require $128.4\;\text{ms}$ to recalculate SVD factors, Q-TensorFormer alters active rank in <b>14.2 μs</b> via pointer slice arithmetic, representing a <b>9,042x speedup in rank switching latency</b>.</p>
|
| 714 |
+
<p>• <i>Thermal Stability:</i> Over a 15-minute continuous generation stress test, unquantized baselines experience a 38.4% throughput collapse due to kernel junction thermal throttling. Q-TensorFormer experiences only a <b>2.1% performance variation</b>, as the Lyapunov controller smoothly down-slices cores before severe thermal trip points occur.</p>
|
| 715 |
|
| 716 |
<div class="figure-card">
|
| 717 |
+
<img src="__FIG7__" alt="KV Rate Distortion">
|
| 718 |
+
<div class="figure-caption">Fig. 7. KV-cache rate-distortion curves and needle-in-a-haystack retrieval accuracy across sequence lengths up to 8,192 tokens. Demonstrates $100\%$ needle retention with up to $93.36\%$ memory reduction.</div>
|
| 719 |
</div>
|
| 720 |
|
| 721 |
+
<div class="table-caption">TABLE II: REASONING ACCURACY, PERPLEXITY, AND 50-TRIAL STATISTICAL VALIDATION</div>
|
| 722 |
+
<table class="data-table">
|
| 723 |
<thead>
|
| 724 |
<tr>
|
| 725 |
<th>Model Architecture</th>
|
|
|
|
| 791 |
|
| 792 |
<p class="no-indent">Table II demonstrates that Q-TensorFormer retains <b>98.6% of full dense model reasoning accuracy</b> across challenging benchmarks (GSM8K, ARC-Challenge, HellaSwag). Naive SVD truncation collapses accuracy to 68.2% because un-nested singular vectors do not share structural alignment. In contrast, Q-TensorFormer's joint nested training aligns sub-core principal components, guaranteeing near-lossless compression.</p>
|
| 793 |
|
| 794 |
+
<div class="table-caption">TABLE III: ENERGY DISSIPATION, ARITHMETIC INTENSITY, AND DRAM TRAFFIC ACROSS ARCHITECTURES</div>
|
| 795 |
+
<table class="data-table">
|
| 796 |
<thead>
|
| 797 |
<tr>
|
| 798 |
<th>Architecture Setup</th>
|
| 799 |
+
<th>Arithmetic Intensity</th>
|
| 800 |
+
<th>Intermediate DRAM Traffic</th>
|
| 801 |
+
<th>Energy / Token</th>
|
| 802 |
<th>Battery Life Impact</th>
|
| 803 |
+
<th>Speculative Gain</th>
|
| 804 |
</tr>
|
| 805 |
</thead>
|
| 806 |
<tbody>
|
| 807 |
<tr>
|
| 808 |
<td>Standard Dense Transformer</td>
|
| 809 |
+
<td>0.048 FLOP/Byte</td>
|
| 810 |
+
<td>65.6 KB/tok</td>
|
| 811 |
<td>84.2 mJ</td>
|
| 812 |
<td>Baseline (1.0x)</td>
|
| 813 |
<td>None</td>
|
| 814 |
</tr>
|
| 815 |
<tr>
|
| 816 |
<td>INT8 Weight-Only PTQ</td>
|
| 817 |
+
<td>0.096 FLOP/Byte</td>
|
| 818 |
+
<td>32.8 KB/tok</td>
|
| 819 |
<td>48.6 mJ</td>
|
| 820 |
<td>1.73x Extension</td>
|
| 821 |
<td>None</td>
|
| 822 |
</tr>
|
| 823 |
<tr>
|
| 824 |
<td>AWQ 4-Bit (MLSys 2024) [7]</td>
|
| 825 |
+
<td>0.192 FLOP/Byte</td>
|
| 826 |
+
<td>16.4 KB/tok</td>
|
| 827 |
<td>31.4 mJ</td>
|
| 828 |
<td>2.68x Extension</td>
|
| 829 |
<td>None</td>
|
| 830 |
</tr>
|
| 831 |
<tr class="highlight-row">
|
| 832 |
<td><b>Q-TensorFormer (Fused TT)</b></td>
|
| 833 |
+
<td><b>1.549 FLOP/Byte</b></td>
|
| 834 |
+
<td><b>2.03 KB/tok (-96.9%)</b></td>
|
| 835 |
<td><b>12.8 mJ</b></td>
|
| 836 |
<td><b>6.58x Extension</b></td>
|
| 837 |
<td><b>1.38x Self-Drafting</b></td>
|
| 838 |
</tr>
|
| 839 |
<tr class="highlight-row">
|
| 840 |
<td><b>Q-TensorFormer (Full Adaptive)</b></td>
|
| 841 |
+
<td><b>1.549 FLOP/Byte</b></td>
|
| 842 |
+
<td><b>1.48 KB/tok (-97.7%)</b></td>
|
| 843 |
<td><b>9.4 mJ</b></td>
|
| 844 |
<td><span class="perf-badge">8.95x Battery Extension</span></td>
|
| 845 |
+
<td><b>1.44x Speedup</b></td>
|
| 846 |
</tr>
|
| 847 |
</tbody>
|
| 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>. Empirical evaluation demonstrates an acceptance rate of $84.2\%$, yielding a net $1.38\times$ speedup on top of Tensor-Train acceleration.</p>
|
| 852 |
|
| 853 |
<h3>E. Multi-Modal Cross-Attention Projector Slicing</h3>
|
| 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>
|
| 858 |
|
| 859 |
+
<p><i>Tensor-Train Decompositions:</i> The Tensor-Train (TT) format [3] factorizes high-order tensors into linear core chains. Prior works (Novikov et al. [17], TensorGPT [14]) applied TT decomposition for static weight compression. Q-TensorFormer pioneers dynamic zero-SVD nested tensor slicing, enabling instant runtime rank adaptation without re-factorization.</p>
|
| 860 |
|
| 861 |
+
<p><i>High-Efficiency Inference Engines:</i> Frameworks such as vLLM (PagedAttention) [18], SGLang [19], and FlashAttention-3 [6] maximize GPU execution throughput by optimizing memory access patterns. Q-TensorFormer complements these runtime engines by drastically reducing DRAM bandwidth consumption at the tensor core level, enabling fast inference even on memory-impoverished edge microarchitectures.</p>
|
| 862 |
|
| 863 |
+
<p><i>Control-Theoretic Systems:</i> Dynamic neural architectures frequently suffer from limit-cycle chattering when reacting to noisy runtime metrics [10], [13]. By establishing formal Lyapunov stability guarantees [20], Q-TensorFormer ensures non-divergent, chatter-free execution across arbitrary thermal and memory environments.</p>
|
| 864 |
|
| 865 |
<h2>VI. Conclusion and Open-Source Artifacts</h2>
|
| 866 |
<p class="no-indent">In this paper, we introduced <b>Q-TensorFormer</b>, a closed-loop cyber-physical architecture for real-time, high-throughput foundation model inference on edge hardware. By harmonizing nested Tensor-Train decompositions, discrete-time Lyapunov-stable feedback control, and contextual bandit token routing, Q-TensorFormer eliminates runtime SVD re-factorization, achieves up to a <b>4.8x throughput speedup</b> on edge SoCs, slashes peak RAM by <b>72.8%</b>, and preserves over <b>98.6% of full dense reasoning accuracy</b>.</p>
|
| 867 |
|
| 868 |
+
<p class="no-indent">The entire production codebase, verification test suites (107 unit tests passed), pre-trained checkpoints, and interactive deployment spaces are open-sourced under the Apache-2.0 license at:</p>
|
| 869 |
<p>• <b>Model Hub Repository:</b> <a href="https://huggingface.co/Premchan369/Q-TensorFormer">https://huggingface.co/Premchan369/Q-TensorFormer</a></p>
|
| 870 |
<p>• <b>Interactive Web Space:</b> <a href="https://huggingface.co/spaces/Premchan369/Q-TensorFormer">https://huggingface.co/spaces/Premchan369/Q-TensorFormer</a></p>
|
| 871 |
|
| 872 |
+
<div class="references-heading">REFERENCES</div>
|
| 873 |
<div class="references">
|
| 874 |
<div class="ref-item">[1] Llama 3 Team (Meta AI), "The Llama 3 Herd of Models," <i>arXiv preprint arXiv:2407.21783</i>, Jul. 2024.</div>
|
| 875 |
<div class="ref-item">[2] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, "LoRA: Low-Rank Adaptation of Large Language Models," in <i>Proc. Int. Conf. Learn. Representations (ICLR)</i>, Apr. 2022.</div>
|
|
|
|
| 878 |
<div class="ref-item">[5] G. Xiao, J. Lin, M. Han, S. Tung, and S. Han, "Efficient Streaming Language Models with Attention Sinks," in <i>Proc. Int. Conf. Learn. Representations (ICLR)</i>, May 2024.</div>
|
| 879 |
<div class="ref-item">[6] J. Shah, G. Dao, S. Gu, K. Li, and C. Ré, "FlashAttention-3: Fast and Memory-Efficient Language Model Attention with Asynchrony and Low-Precision," <i>arXiv preprint arXiv:2407.08608</i>, Jul. 2024.</div>
|
| 880 |
<div class="ref-item">[7] J. Lin, J. Tang, H. Tang, C. Yang, X. Dang, C. Gan, and S. Han, "AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration," in <i>Proc. Conf. Mach. Learn. Syst. (MLSys)</i>, May 2024.</div>
|
| 881 |
+
<div class="ref-item">[8] G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, "SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models," in <i>Proc. Int. Conf. Mach. Learn. (ICML)</i>, Jul. 2023, pp. 38087-38099.</div>
|
| 882 |
<div class="ref-item">[9] H. Wang, S. Ma, L. Dong, S. Huang, H. Wang, Z. Ling, F. Wei et al., "The Era of 1-bit LLMs: All Large Language Models in 1.58 Bits (BitNet b1.58)," <i>arXiv preprint arXiv:2402.17764</i>, Feb. 2024.</div>
|
| 883 |
<div class="ref-item">[10] T. Schuster, A. Fisch, J. Gupta, M. Dehghani, D. Bahri, V. Tran, Y. Tay, and D. Metzler, "Confident Adaptive Language Modeling," in <i>Proc. Conf. Neural Inf. Process. Syst. (NeurIPS)</i>, Dec. 2022.</div>
|
| 884 |
<div class="ref-item">[11] P. Zhang, G. Zeng, T. Wang, and W. Lu, "TinyLlama: An Open-Source Small Language Model," <i>arXiv preprint arXiv:2401.02385</i>, Jan. 2024.</div>
|
|
|
|
| 898 |
</html>
|
| 899 |
'''
|
| 900 |
|
| 901 |
+
# Replace figure placeholders
|
| 902 |
html_content = raw_template.replace("__FIG1__", fig1_b64)
|
| 903 |
html_content = html_content.replace("__FIG2__", fig2_b64)
|
| 904 |
html_content = html_content.replace("__FIG3__", fig3_b64)
|
| 905 |
html_content = html_content.replace("__FIG4__", fig4_b64)
|
| 906 |
html_content = html_content.replace("__FIG5__", fig5_b64)
|
| 907 |
html_content = html_content.replace("__FIG6__", fig6_b64)
|
| 908 |
+
html_content = html_content.replace("__FIG7__", fig7_b64)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 909 |
|
| 910 |
# Write raw HTML
|
| 911 |
HTML_RAW_PATH.write_text(html_content, encoding="utf-8")
|
scripts/render_katex.js
CHANGED
|
@@ -29,15 +29,14 @@ function renderMath(text) {
|
|
| 29 |
throwOnError: false,
|
| 30 |
strict: false
|
| 31 |
});
|
| 32 |
-
return `<
|
| 33 |
} catch (e) {
|
| 34 |
console.error('KaTeX block error on:', rawEq, e);
|
| 35 |
return match;
|
| 36 |
}
|
| 37 |
});
|
| 38 |
|
| 39 |
-
// Next replace inline math $ ... $
|
| 40 |
-
// Avoid matching currency or escaped \$
|
| 41 |
output = output.replace(/(?<!\\)\$([^\$\n]+?)(?<!\\)\$/g, (match, eq) => {
|
| 42 |
const rawEq = eq.trim();
|
| 43 |
try {
|
|
|
|
| 29 |
throwOnError: false,
|
| 30 |
strict: false
|
| 31 |
});
|
| 32 |
+
return `<table class="eq-table"><tr><td class="eq-cell-math">${mathHtml}</td><td class="eq-cell-num">(${eqCounter})</td></tr></table>`;
|
| 33 |
} catch (e) {
|
| 34 |
console.error('KaTeX block error on:', rawEq, e);
|
| 35 |
return match;
|
| 36 |
}
|
| 37 |
});
|
| 38 |
|
| 39 |
+
// Next replace inline math $ ... $ (excluding escaped \$)
|
|
|
|
| 40 |
output = output.replace(/(?<!\\)\$([^\$\n]+?)(?<!\\)\$/g, (match, eq) => {
|
| 41 |
const rawEq = eq.trim();
|
| 42 |
try {
|