PR #2077

open

ANS Compression and Hyperloop

by tranphat180603View on GitHub
val_bpb
1.0877
Architecture
Transformer
Optimizer
Artifact Size
15,875,599 bytes

Training Techniques

Quantization
GPTQ
bits: null
scope: model weights
Compression
custom
level: null
Evaluation
sliding window eval
parameters: null
Test-Time Training
full TTT
parameters: {"learning_rate":0.005,"epochs":3}
Sequence Length
sequence_length
train_length: 8192
eval_length: null
Architecture
depth recurrence
3-layer recurrence / recurrent middle block with parallel residuals
parameters: {"layers":3}
weight tying
tied embeddings / tied embedding baseline implied by canonical method list
parameters: null

Novel Contributions

  • Replaced the quantized artifact compression path with custom ANS-coded compression
  • Reported improved quantized artifact size versus Brotli/LZMA after quantization
  • Explored a Hyperloop-lite loop mechanism, but found it slower and not beneficial under the time budget