3 Commits

Author	SHA1	Message	Date
lars	5c576fa8f3	perf: compact Stage-2 AR inference loop to active rows only ... sample_secondaries_ar ran all k_max=15 slots for every row regardless of each row's own predicted secondary count, even though the baseline checkpoint's rollout measured only 0.382 secondaries/step — so ~97% of stage-2 model calls generated tokens sec_valid then masked away. Compact the loop to the still-active row set at each slot: drop a row the moment its n_sec_pred is exhausted (or, under n_sec.mode="stop_token", the moment its own stop logit fires), so slot k's model calls cost O(active rows) instead of O(B). Exact — rows are independent given their own history — verified by comparing the compacted path against a new full_length=True escape hatch that reproduces the original uncompacted behavior bit-for-bit under deterministic noise. full_length=True is required by _assemble_stage2_ar_inputs_scheduled's scheduled-sampling self-sample, whose training contract needs a real prediction at every slot up to k_max regardless of a row's own count, so training behavior is unchanged. AttentionHistory's KV cache and MarkovHistory's O(1) state are kept aligned to the shrinking active set via a new HistoryEncoder.select_cache / Stage2Autoregressive.select_history_cache. Also fixes a latent bug the refactor surfaced: derived_n_sec (stop-token mode) could be overwritten by a later spurious re-fire of the stop logit on a row that had already stopped; now tracked via an explicit `finished` mask so only the first stop slot is recorded, matching the documented contract. No architecture or checkpoint-format change — every existing v0.3.0 Stage2Autoregressive checkpoint (flow/wgan, markov/attention, head/stop_token) picks up the speedup automatically on its next `giant rollout`/`giant predict`, no retraining needed. Measured (CPU, hidden_dim=512/6 blocks, k_max=15, batch 512, mean n_sec≈0.38 matching the baseline checkpoint's own rollout): 17.6-22.9x fewer wall-clock seconds for the AR loop alone (attention/markov history respectively). Directional only — baseline.toml's GPU inference-cost comment is updated accordingly, flagged stale pending a real rollout re-measurement via eval_cost_per_step. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HPt7bVLZYFJe5cG6V7ahqC	2026-09-03 17:56:59 +02:00
lars	50d8368415	docs: record analysis_341dfb14 baseline rollout benchmark results ... Replaces the extrapolated pre-v0.3 weak-spot claims in baseline.toml's header (which had the wrong sign on step-count error) with measured numbers from the first full rollout validation of this exact config, and adds a matching Roadmap entry in CLAUDE.md.	2026-08-28 15:01:35 +02:00
lars	d858226294	Add configs/baseline.toml as the kept reference model ... A fixed comparison point for future architecture variants, so each experimental axis (routed trunk, WGAN generators, attention history, shared conditioning) is a single edit away from one known config. flow/flow autoregressive, hidden_dim 512 / 6 blocks per stage, physical conditioning, no router, 7.70M params. Chosen by ranking the five runs in analysis_runs/ by mean Jensen-Shannon divergence against the Geant4 reference: unrouted flow wins (0.172) over routed flow (0.197/0.200) and both WGAN runs (0.218/0.234), with the lead concentrated in per-event total deposited energy and the per-PDG marginals. batch_size 36864 is sized for one L40S on deepthought2 from a measured linear fit of this config's training step (reserved MiB = 0.9736 * bs + 115), giving ~36 GiB, 78% of the card. The comments record two measured facts that are easy to get wrong: WGAN is slower to *train* than flow (n_critic plus the gradient-penalty double-backward), its advantage being inference-only; and sample_secondaries_ar loops over all k_max slots unconditionally rather than short-circuiting on n_sec, which is what makes the autoregressive decoder the dominant cost on both axes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>	2026-08-14 17:36:26 +02:00