Use generate_compiled (CUDA graph) for ~2.75x faster inference

Fix Haiku model ID — claude-haiku-4-5-20251001

Filter threshold 90%, wrap annotation description text

Filter annotations to >=95% percmatch

Parallel annotation, scoring tweaks, UI polish

UI overhaul + speed improvements

Fix BaseStreamer import — use duck-typed class instead

Fix GPU detection and bokeh plasmid map rendering

Streaming progress bar, rename to Top N, fix scoring

Token-aware scoring, KV cache, torch.compile

Pin transformers<5: model built for 4.x, 5.3.0 breaks sampling

Add startup diagnostic: test forward pass + greedy gen

Match model card loading exactly: no dtype, no device_map

Fix: load without device_map to avoid accelerate dispatch hooks

Fix: use dtype= and device_map= instead of torch_dtype + .to()

Add detailed logging to generate_dna for debugging

Simplify: remove ZeroGPU, use dedicated GPU directly

Use KV cache in manual generation + stop cleanly on NaN

Fix: force eos/pad token IDs from vocab when tokenizer doesn't set them

Fix: handle None pad_token_id in manual generate loop

Replace model.generate() with manual loop + Gumbel-max sampling

Fix numpy.bool8 compat: shim for bokeh 2.4.3 + numpy 2.x

Fix ZeroGPU NaN: re-init non-persistent RoPE buffers on GPU

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Delete Dockerfile with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload requirements.txt with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Delete packages.txt with huggingface_hub

Upload folder using huggingface_hub

initial commit