fix(logit_pipeline): derive temp_probs from raw logits, not penalized

build(gitignore): fix stress-test output ignore path

revert(timing): remove temporary ZeroGPU inference instrumentation

docs(lessons-learned): record ZeroGPU workshop load-capacity results

perf(zerogpu): set per-endpoint GPU durations for tighter scheduling

feat(timing): wire ZeroGPU timing into streaming endpoints

feat(timing): wire ZeroGPU timing into predict_next_token

feat(timing): wire ZeroGPU timing into the safety path

fix(timing): flush ZG_TIME output and strengthen timing tests

feat(timing): add temporary ZeroGPU inference timing helper

docs: full rewrite for Transformers + ZeroGPU backend

build: drop vLLM and nvidia-ml-py deps for Transformers backend

test(app): align test_app and test_app_token_display with sync handlers

feat(app): rewire handlers from async vLLM to sync Transformers

test: restore Transformers backend tests, delete vLLM-only tests

feat: restore Transformers backend from pre-vLLM history

docs: fix README and AGENTS to accurately describe vLLM backend

docs: move investigation docs to docs/archive/

docs: add lessons-learned.md — consolidated dead-ends reference

chore: delete merge artifact, gitignore stress-test results

docs: remove TORCH_COMPILE and PROFILE_REQUESTS documentation

refactor: remove [PROFILE] request-lifecycle instrumentation

revert(app): restore .then wiring for dataset updates

chore(app): remove dead .token-button CSS

fix(app): use visible='hidden' for distribution slots (Gradio #13494)

diag(app): log received temp/freq + distribution count per next_token click

fix(app): compute dataset updates inline to avoid stale chart_state in .then

fix(app): left-align Dataset cells via CSS (component_props text_align had no effect)

feat(app): shorten probability bar and left-align dataset cells

fix(app): pass gr.Dataset components as strings to avoid stray textboxes

docs: update Next Token rendering section for gr.Dataset

feat(app): replace @gr .render token buttons with static gr.Dataset

feat(app): add _make_dataset_select_handler factory for dataset clicks

feat(app): add _apply_distribution_updates gradio wrapper

feat(app): add _distribution_rows helper for dataset rendering

refactor(tests): consolidate vLLM fakes into tests/_fakes.py

test: add vLLM generate() API contract test

fix(generation,logit_pipeline): pass token IDs via prompt arg (V1 generate API)

fix(engines): assign correct repo to each variant (instruct/base were swapped)

fix(engines): re-budget from measured weights + make gpu fractions env-tunable

fix(engines): re-budget gpu_memory_utilization to fit a 22 GiB L4

fix(config): restore gpu_monitor startup in load_model

fix(generation): surface _merge_async_gens streaming errors instead of deadlocking

fix: async monkeypatch for handle_token_selection tests

feat(generation): parallel base+instruct streaming in compare_models

docs: update AGENTS.md for vLLM deployment (ZeroGPU dropped, L4 only)

feat(app): wire async vLLM handlers, drop ZeroGPU, eager engine load

refactor: slim config.py, drop zerogpu, update imports

feat(safety): rewrite check_safety via ShieldGemma vLLM engine

feat(generation): rewrite with async vLLM streaming generators