Digital Drosophila Archive
4.3 KiB
Conditioning acceleration v1 — independent engineering benchmark
Authority: user requested V100 acceleration and algorithm optimization on 2026-10-01. This work DOES NOT resume or change Phase 4AR, and does not authorize a new 30,000-epoch scientific experiment.
Frozen scientific semantics
Preserve original Phase 4J/AM eligibility, compartment mapping, the complete candidate mask, anatomical base weights, float32 update semantics, all time steps and original event ordering. No threshold/gain/eta/tau/sign change, reduced trial count claim, surrogate gradients, skip-time power approximation, new connections or model learning claim. GPU outputs are diagnostic replay products only and are never installed into a brain.
The oracle is the exact SHA-pinned original Torch CPU update running in the original science environment. Candidate CuPy CUDA code runs in a separate environment with NumPy 2.2.6 and CuPy 13.6.0, using the installed CUDA 12.6 toolkit. Do not modify the original environment, ComfyUI or frozen source files.
Fixed validation and measurement plan
Select the first and last fully ledger-committed training task for each observed arm by deterministic task-ID order, without inspecting learning scores. This is coverage of available saved inputs only, not all registered conditions. Export fixed original inputs and recomputed CPU reference outputs into a new directory; compare them with the saved scientific outputs before use.
Compare all sampled traces, final traces and local delta arrays byte-for-byte against the original CPU oracle on both physical V100s and the 4090. Add synthetic edge cases: no spikes, non-KC spikes, no US, US at zero/final tick, duplicate/invalid events rejection, multiple US snapshots for trace-only API, saturation, D=0, frozen identity, all compartments, and seeded sparse/random traces. Do not select only passing cases. Any mismatch blocks promotion.
Use 3 warmups and 20 timed repeats for every real fixture on each GPU, with the same fixture order. Record first-use compilation/context cost separately, CPU preparation, transfer-inclusive host wall time, kernel event time, GPU identity, memory and exact software/source hashes. Repeat within a resident process to expose avoidable per-process overhead. Never present kernel-only speedup as full training speedup.
Budget: at most 15 minutes of candidate benchmark execution after dependencies/compilation, one process per GPU initially; do not occupy a GPU with a nonempty ComfyUI queue or evict services. No automatic scientific retry. Implementation failures may be diagnosed and corrected in new benchmark attempts, retaining earlier reports. No modification of any frozen experiment. Independent CPU verification remains required before any future production integration.
Gate and limits
Passing this benchmark means only ELIGIBILITY-UPDATE-REPLAY-EQUIVALENCE for the tested inputs. It does not establish full-brain event equivalence, end-to-end acceleration, learning success, resumability, or completion of Phase 4AR. Integration, persistent brain workers and independent-seed multi-GPU scheduling require separately validated execution/audit changes. Stop before installing any update or starting learning.
Keep source, configuration, environment installation report, fixed input manifest, raw timing samples, comparisons, failures and SHA-256 manifest. Check Phase 4AR registration/source and the prior saved-file inventory before/after; report the exact scope (not a new Phase 1–4AQ numerical regression).
Recorded implementation correction (before successful GPU execution)
The first candidate installation attempted pip's --report with the new venv's inherited pip 22.0.2, which does not support it. Only the candidate venv was upgraded to pip 25.2; both logs remain preserved.
GPU attempt v100a-01 failed before running kernels because CuPy 13.6's RawKernel cache compiler appends -ftz=true, conflicting with the required --ftz=false. Do not remove the precision requirement. The corrected candidate compiles the same CUDA recurrence with installed nvcc 12.6, explicit separate rounding and --ftz=false, then loads the precompiled module through CuPy. Compiler commands, sources, output and binary hashes are preserved in candidate-only caches. The original scientific recurrence and all predeclared fixtures remain unchanged; failed attempt records are not replaced.