Multi-turn UX: new-chat button; example click resets conversation; 1k-token sliding-window history; 4th example (sleep)

INT8 dynamic quant (~2x CPU, verified all 3 modes); add self-spec to chat mode selector

Compare/chat prompts: drop count-based "three tips" (looks truncated at 48 tok); use self-contained short-answer prompts (ocean / neural net / exercise)

Fix chat crash: gradio 6 Chatbot content is a list of parts; flatten to text before apply_chat_template

Fix gradio 6.14: gr.Chatbot dropped the type= arg (messages is default)

Custom chat layout: compact mode/max-tokens/temp controls above input; single-column examples (readable); header link to codelion/dhara-250m

Disable example caching (HF Spaces defaults cache_examples=True) so example clicks generate live, not pre-baked

Coherence: rep-penalty 1.3 in diffusion/self-spec (no more collapse); clip to last full sentence (no mid-word cutoff); exposed config panel (mode/max-tokens/temperature); greedy AR in compare; better default prompts

Fix gradio 6 build: examples as list-of-lists (additional_inputs); move theme/css to launch()

Chat mode toggle (AR/diffusion w/ live ▓▒░ denoise); streaming 3-mode compare reveal; max_new 128 + no_repeat_ngram

Fixed-height scrollable chat; pass attention_mask (silence pad==eos warning); _enc helper

Denoiser-terminal UI (mask-glyph aesthetic, amber CRT) + streaming per-mode compare with progress

gradio SDK 6.14.0 (proven to import with hub 1.16) + transformers 5.5 + messages-format app

Pin gradio==4.44.1 + use 4.x tuples ChatInterface API (model already loads)

Switch to Docker SDK; pip-resolved transformers==5.5.0 + gradio (fixes tokenizer/HfFolder)

Pin coherent stack (py3.12, gradio4.44, transformers<5, hub<1) to fix HfFolder/audioop; title ChatDhara

dhara-chat tri-mode CPU demo