autoresearch: generate and gate non-catalog vLLM optimization candidates#56

Open

nicholaslawrence-hub wants to merge 25 commits into

mainfrom

autoresearch-clean

Commits on Jul 8, 2026

autoresearch: generate non-catalog levers
nicholaslawrence-hub
committed
autoresearch: source valid knob domains from EngineArgs CLI args

nicholaslawrence-hub
and
claude
committed
vllm-decode: add GITM_VLLM_ENFORCE_EAGER to disable CUDA graphs

nicholaslawrence-hub
and
claude
committed
autoresearch: skip multi-GPU topology knobs on a single-GPU box

nicholaslawrence-hub
and
claude
committed
autoresearch: surface the real apply-failure reason (apply_error)

nicholaslawrence-hub
and
claude
committed
report: fix mislabeled residual_invariant on AF2/edge claims
nicholaslawrence-hub
committed
autoresearch: report the real target residual instead of a hardcoded 0.0
nicholaslawrence-hub
committed
monitor: match observed kernels to predicted nodes by op identity
nicholaslawrence-hub
committed
cleaning up verbose comments from residual fix
nicholaslawrence-hub
committed
autoresearch: weight classify_bottleneck by the roofline model's bound
nicholaslawrence-hub
committed
autoresearch: replace the partial-prefill/dbo denylist with a live prerequisite check
nicholaslawrence-hub
committed
autoresearch: joint (multi-knob) candidates, and rank generated candidates for real
nicholaslawrence-hub
committed
tests: compress redundant test functions from the last two commits
nicholaslawrence-hub
committed
replay: unify predict_delta coverage onto classify_op, fix _op_present
nicholaslawrence-hub
committed
catalog: scale max_num_batched_tokens off the live engine, not a static 8192
nicholaslawrence-hub
committed
catalog: sweep every relative lever instead of one guess, not just tokens
nicholaslawrence-hub
committed
deviation: classify_op recognizes real vLLM attention/KV-cache kernels
nicholaslawrence-hub
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

autoresearch: generate and gate non-catalog vLLM optimization candidates#56

autoresearch: generate and gate non-catalog vLLM optimization candidates#56
nicholaslawrence-hub wants to merge 25 commits into
mainfrom
autoresearch-clean

Commits on Jul 8, 2026

autoresearch: generate non-catalog levers

autoresearch: source valid knob domains from EngineArgs CLI args

vllm-decode: add GITM_VLLM_ENFORCE_EAGER to disable CUDA graphs

autoresearch: skip multi-GPU topology knobs on a single-GPU box

autoresearch: surface the real apply-failure reason (apply_error)

report: fix mislabeled residual_invariant on AF2/edge claims

autoresearch: report the real target residual instead of a hardcoded 0.0

monitor: match observed kernels to predicted nodes by op identity

cleaning up verbose comments from residual fix

autoresearch: weight classify_bottleneck by the roofline model's bound

autoresearch: replace the partial-prefill/dbo denylist with a live prerequisite check

autoresearch: joint (multi-knob) candidates, and rank generated candidates for real

tests: compress redundant test functions from the last two commits

replay: unify predict_delta coverage onto classify_op, fix _op_present

catalog: scale max_num_batched_tokens off the live engine, not a static 8192

catalog: sweep every relative lever instead of one guess, not just tokens

deviation: classify_op recognizes real vLLM attention/KV-cache kernels

Commits on Jul 9, 2026

comments: compress verbose docstrings from this session's changes

catalog: add async_scheduling, cpu_offload_gb, preemption_mode; retire from autoresearch's static table

docs: fix invalid UTF-8 byte in autoresearch.md breaking CI review bots

autoresearch: joint candidates go first so max_candidates can't crowd them out

Merge main into autoresearch-clean, resolve workloads.py conflict

apply: audit revert calls now carry knobs=, matching apply

autoresearch: fix zero-default value grid and cache _searchable()

apply: LiveEngineApplicator shutdown releases IPC memory like the vllm-decode path