-
v0.5: Add streaming HF dataset support for art/anime training, new notebook with dataset options
-
v0.5: notebook — document auto patch_size, fix total_memory, all configs L≤256"
-
v0.5 CRITICAL: Fix patch_size to scale with image size (L≤256 always), reduce d_state for speed, add config table
-
v0.4.1: Fix total_memory bug, add ETA, recommend 15-20 epochs for CIFAR, better logging
-
v0.4: notebook — add mambapy install, force re-clone
-
v0.4: Use mambapy.pscan (proven, grad-safe parallel scan) — no more torch.associative_scan
-
v0.3: Notebook — force re-clone, log step 1 immediately, show progress instantly
-
v0.3: update __init__.py
-
v0.3: PARALLEL SSM scan (torch.associative_scan), patch_size 4→8, no more Python for-loop
-
Fix deprecated torch.cuda.amp API in training cell — use torch.amp instead
-
Fix: remove duplicate forward method in LiquidSSMBlock, clean up dead code"
-
CRITICAL FIX: OOM on T4 — rewrite SSM scan to not materialize 4D tensors, add gradient checkpointing, optimize Liquid CfC memory
-
Add Colab/Kaggle training notebook with all dataset options
-
Add smoke_test.py — 25 comprehensive CPU tests
-
Add train.py — full training script with CLI
-
Add sampling.py — Euler and Heun ODE samplers
-
Add losses.py, sampling.py, train.py, smoke_test.py
-
Add model.py — core LiquidFlow architecture
-
Add liquidflow package: __init__.py
-
Initial release: LiquidFlow architecture + training code + notebook
-
initial commit
