neurocuda 0.2.0__tar.gz

neurocuda-0.2.0/CLAUDE.md

@@ -0,0 +1,74 @@
+# CLAUDE.md — NeuroCUDA
+
+## PRIME DIRECTIVE — HONESTY RULES (read before every task)
+
+1. **A failed run is a bug, never a "finding."** If a published method produces bad results, your implementation is broken. Investigate. Do not claim you discovered the method doesn't work.
+2. **Full test set only.** CIFAR-10 = 10,000 images. Never report 500-image subsets as results.
+3. **≥3 seeds.** Every number is mean ± std. Single runs are not results.
+4. **Label hardware precisely.** "Loihi 2 simulator validated against Lava" — never "Loihi 3" or "silicon" unless physically run on it.
+5. **Gate failure = STOP.** Do not proceed. Do not relabel the target.
+6. **Report failures first.** "Gate 2 FAILED. Cause: X. Options: Y."
+7. **No marketing language.** No "world-class," "nobody has done this," "🔥." Just measurements.
+
+## Project Identity
+
+**NeuroCUDA is a systems/tooling contribution.** A pip-installable compiler that takes PyTorch models and deploys them across GPU, CPU, Loihi 2 simulator, and FPGA through one API call.
+
+**Goal:** Build an honest, working seed. NOT beat SOTA accuracy. NOT claim novel science.
+
+## Current State (June 19, 2026)
+
+### Real Numbers (full test set, need multi-seed verification)
+- MLP/MNIST: ANN 97.8%, SNN 97.4%, gap 0.4% (1 seed)
+- StrongCNN/CIFAR-10: ANN 80.4%, SNN 74.3%, gap 6.0% (1 seed, convert + FT)
+- SEW-ResNet/CIFAR-10: 67.7% at T=8 (direct SNN, 50 epochs, 1 seed)
+- ResNet/CIFAR-10: ANN 92.1%, SNN 70.1%, gap 22.0% (convert + FT, 1 seed)
+
+### What Works
+- Multi-backend deployment: GPU/CPU/Loihi ≤1.2% deviation
+- Loihi 2 bit-accurate validation: 0/256K spike deviations
+- BN folding: lossless
+- Post-conversion fine-tuning: +7-52% gain
+
+### Known Bugs
+- QCFS: λ frozen at ~1.0 for layers 1-2 (gradient/learning-rate bug)
+- Weight normalization: destroys ANN accuracy (implementation error)
+- "Spike-density trap" and "QCFS non-generalization" are bugs, not discoveries
+
+## The Gates
+
+### GATE 1 — Ground Truth
+Re-measure everything on full test set, 3 seeds. Produce honest baseline table.
+
+### GATE 2 — Fix Base ANN
+Train ResNet-18 on CIFAR-10 to ≥93% (solved problem, standard recipe).
+
+### GATE 3 — Fix Converter (QCFS)
+Initialize from pretrained ANN. Separate higher LR on λ. Gap ≤5%.
+
+### GATE 4 — Re-test Methods
+Re-run per-channel, SPIKE-NORM, weight-norm on fixed pipeline.
+
+### GATE 5 — NeuroBench
+Standard-format, multi-seed, multi-backend reporting.
+
+### GATE 6 — Ship
+Clean README, reproducible benchmarks, honest paper.
+
+## NON-Goals
+- Do NOT chase SOTA accuracy
+- Do NOT claim physical silicon without physical silicon
+- Do NOT describe bugs as discoveries
+- Do NOT add scope until Gates 1-6 pass
+
+## Reporting Format
+After each gate:
+```
+GATE N — PASS/FAIL
+Headline: <metric> = <mean> ± <std> (3 seeds, full test set)
+Target: <target>
+Changes: <1-3 bullets>
+Surprises: <honest>
+Verified: YES/NO
+Next: YES/NO — if NO, why
+```

neurocuda-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 NeuroCUDA
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

neurocuda-0.2.0/MANIFEST.in

@@ -0,0 +1,15 @@
+include README.md
+include LICENSE
+include requirements.txt
+include CLAUDE.md
+include reproduce.py
+recursive-include examples *.py
+recursive-include tests *.py
+recursive-include results *.md *.csv *.json
+prune .venv
+prune __pycache__
+prune .git
+prune .claude
+prune checkpoints
+prune data
+prune cache