docs(perf): use idle-card numbers for the GQA attention result
All checks were successful
ci / test_and_build (pull_request) Successful in 41s
perf(cuda): make decode attention GQA-aware
All checks were successful
ci / test_and_build (pull_request) Successful in 41s
perf(cuda): tile long decode spans across blocks
All checks were successful
ci / test_and_build (pull_request) Successful in 1m57s
feat(engine): --power-budget-wait waits out a startup power refusal
All checks were successful
ci / test_and_build (pull_request) Successful in 2m51s
docs: energy metering at inference time
All checks were successful
ci / test_and_build (pull_request) Successful in 50s
docs(readme): thinking-budget guidance measured on four models
All checks were successful
ci / test_and_build (pull_request) Successful in 48s
feat(tokenizer): OLMo 3 Think chat template
All checks were successful
ci / test_and_build (pull_request) Successful in 48s
feat(model): OLMo 3 (Olmo3ForCausalLM) in the dense-transformer package
All checks were successful
ci / test_and_build (pull_request) Successful in 56s
feat(backend): sliding-window PagedAttention
All checks were successful
ci / test_and_build (pull_request) Successful in 51s
feat(config): v4 rope_scaling, per-layer-type rope, and layer_types
All checks were successful
ci / test_and_build (pull_request) Successful in 46s
feat: stop on generation_config's eos_token_id as well as the tokenizer's EOS
All checks were successful
ci / test_and_build (pull_request) Successful in 1m1s
refactor: reasoning markers that are text, not one token
All checks were successful
ci / test_and_build (pull_request) Successful in 51s
perf(engine): sample without sorting the vocabulary, rows side by side
All checks were successful
ci / test_and_build (pull_request) Successful in 55s
docs: knobs for the operator's goals, and measuring what they do
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
feat: soft and hard thinking-token budgets for reasoning turns
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
fix(engine): apply a response_format grammar to the answer, not the reasoning
All checks were successful
ci / test_and_build (pull_request) Successful in 34s
feat(engine): name the served model on each perfstats line
All checks were successful
ci / test_and_build (pull_request) Successful in 30s
feat(engine): response debug preview, unlimited previews in files, and closing lines for unfinished requests
All checks were successful
ci / test_and_build (pull_request) Successful in 34s
test(cuda): add cold-weight and skewed cases to the grouped GEMV benchmark
All checks were successful
ci / test_and_build (pull_request) Successful in 27s
perf(engine): read logits into a reused buffer and pick greedy tokens in parallel
All checks were successful
ci / test_and_build (pull_request) Successful in 36s