feat: GPU/TensorCore integration — TensorFlow backend, accelerated reasoning, training & memory #1

Merged

nsatoshi merged 4 commits from devin/1777352172-gpu-tensorcore-integration into main 2026-04-28 06:32:07 +00:00

Conversation 0 Commits 4 Files Changed 147 +6583 -1119

nsatoshi commented 2026-04-28 05:18:53 +00:00

Owner

Copy Link

Summary

Four commits transforming FusionAGI from an orchestration framework into a self-improving, GPU-accelerated AGI system with consequence-driven learning.

Commit 1 — GPU/TensorCore Integration (22 files, +2,448 lines)

• fusionagi/gpu/ module: TensorBackend protocol, TensorFlowBackend (mixed-precision, XLA), NumPyBackend (CPU fallback)
• GPU ops: cosine similarity, multi-head attention, hypothesis scoring, nearest-neighbor, gradient training
• Integrated into Super Big Brain, reasoning pipeline, memory, self-improvement
• Optional dependency: pip install fusionagi[gpu]

Commit 2 — Deep Integration + Cleanup (112 files, +1,149/-944)

• GPU auto-used in multi_path.py, consensus_engine.py, semantic_graph.py, training.py
• 0 ruff errors (was 758), 0 mypy errors (was 40)

Commit 3 — Advisory Governance + Adaptive Ethics (15 files, +1,024/-132)

• All governance layers default to ADVISORY mode (violations logged, actions proceed)
• AdaptiveEthics — learned ethical framework from experience outcomes
• Unlimited self-correction retries, uncapped training epochs

Commit 4 — Consequence Engine + Causal World Model + Metacognition + Interpretability + Claim Verification (14 files, +1,961/-39)

• ConsequenceEngine — Choice → Consequence → Learning loop. Records decisions with alternatives, risk/reward estimates, actual outcomes, surprise factor
• CausalWorldModel — Learns state-transition patterns from execution history, predicts outcomes
• Metacognition — Self-assessment of reasoning quality, knowledge gap detection, uncertainty identification
• ReasoningTracer — Full prompt→answer reasoning traces with explain() output
• ClaimVerifier — Cross-checks claims for evidence support, confidence calibration, consistency
• FusionAGILoop wires AdaptiveEthics + ConsequenceEngine into task lifecycle

325 tests passing, 0 ruff errors, 0 mypy errors.

Review & Testing Checklist for Human

• Verify ConsequenceEngine risk/reward estimation accuracy with real task data
• Review CausalWorldModel pattern key generation for real action types
• Confirm FusionAGILoop consequence recording doesn't bottleneck under high throughput
• Test AdaptiveEthics weight convergence over many iterations
• Run pip install fusionagi[gpu] with actual GPU hardware

Recommended test plan:

1. pytest tests/ -q --ignore=tests/test_openai_compat.py — 325 passed
2. Instantiate FusionAGILoop with real EventBus, fire events, verify audit log
3. Feed 50+ observations into CausalWorldModel, verify confidence increases
4. Run ConsequenceEngine.estimate_risk_reward() after 20+ choices, verify estimates stabilize

Notes

• Philosophy: "All choices lead to consequences. To learn includes consequences for choices."
• ADVISORY governance is default. Switch to GovernanceMode.ENFORCING if hard blocks needed.
• All new modules follow coding standards: type hints, Google-style docstrings, Pydantic models, protocol-based DI.

## Summary Four commits transforming FusionAGI from an orchestration framework into a self-improving, GPU-accelerated AGI system with consequence-driven learning. **Commit 1 — GPU/TensorCore Integration** (22 files, +2,448 lines) - `fusionagi/gpu/` module: `TensorBackend` protocol, `TensorFlowBackend` (mixed-precision, XLA), `NumPyBackend` (CPU fallback) - GPU ops: cosine similarity, multi-head attention, hypothesis scoring, nearest-neighbor, gradient training - Integrated into Super Big Brain, reasoning pipeline, memory, self-improvement - Optional dependency: `pip install fusionagi[gpu]` **Commit 2 — Deep Integration + Cleanup** (112 files, +1,149/-944) - GPU auto-used in `multi_path.py`, `consensus_engine.py`, `semantic_graph.py`, `training.py` - 0 ruff errors (was 758), 0 mypy errors (was 40) **Commit 3 — Advisory Governance + Adaptive Ethics** (15 files, +1,024/-132) - All governance layers default to ADVISORY mode (violations logged, actions proceed) - `AdaptiveEthics` — learned ethical framework from experience outcomes - Unlimited self-correction retries, uncapped training epochs **Commit 4 — Consequence Engine + Causal World Model + Metacognition + Interpretability + Claim Verification** (14 files, +1,961/-39) - `ConsequenceEngine` — Choice → Consequence → Learning loop. Records decisions with alternatives, risk/reward estimates, actual outcomes, surprise factor - `CausalWorldModel` — Learns state-transition patterns from execution history, predicts outcomes - Metacognition — Self-assessment of reasoning quality, knowledge gap detection, uncertainty identification - `ReasoningTracer` — Full prompt→answer reasoning traces with `explain()` output - `ClaimVerifier` — Cross-checks claims for evidence support, confidence calibration, consistency - `FusionAGILoop` wires AdaptiveEthics + ConsequenceEngine into task lifecycle **325 tests passing, 0 ruff errors, 0 mypy errors.** ## Review & Testing Checklist for Human - [ ] Verify `ConsequenceEngine` risk/reward estimation accuracy with real task data - [ ] Review `CausalWorldModel` pattern key generation for real action types - [ ] Confirm `FusionAGILoop` consequence recording doesn't bottleneck under high throughput - [ ] Test `AdaptiveEthics` weight convergence over many iterations - [ ] Run `pip install fusionagi[gpu]` with actual GPU hardware **Recommended test plan:** 1. `pytest tests/ -q --ignore=tests/test_openai_compat.py` — 325 passed 2. Instantiate `FusionAGILoop` with real `EventBus`, fire events, verify audit log 3. Feed 50+ observations into `CausalWorldModel`, verify confidence increases 4. Run `ConsequenceEngine.estimate_risk_reward()` after 20+ choices, verify estimates stabilize ### Notes - Philosophy: "All choices lead to consequences. To learn includes consequences for choices." - ADVISORY governance is default. Switch to `GovernanceMode.ENFORCING` if hard blocks needed. - All new modules follow coding standards: type hints, Google-style docstrings, Pydantic models, protocol-based DI.

nsatoshi added 1 commit 2026-04-28 05:18:54 +00:00

feat: GPU/TensorCore integration — TensorFlow backend, GPU-accelerated reasoning, training, and memory ... fa71f973a6

- New fusionagi/gpu/ module with TensorBackend protocol abstraction
  - TensorFlowBackend: GPU-accelerated ops with TensorCore mixed-precision
  - NumPyBackend: CPU fallback (always available, no extra deps)
  - Auto-selects best available backend at runtime

- GPU-accelerated operations:
  - Cosine similarity matrix (batched, XLA-compiled)
  - Multi-head attention for consensus scoring
  - Batch hypothesis scoring on GPU
  - Semantic similarity search (pairwise, nearest-neighbor, deduplication)

- New TensorFlowAdapter (fusionagi/adapters/):
  - LLMAdapter for local TF/Keras model inference
  - TensorCore mixed-precision support
  - GPU-accelerated embedding synthesis fallback

- Reasoning pipeline integration:
  - gpu_scoring.py: drop-in GPU replacement for multi_path scoring
  - Super Big Brain: use_gpu config flag, GPU scoring when available

- Memory integration:
  - gpu_search.py: GPU-accelerated semantic search for SemanticGraphMemory

- Self-improvement integration:
  - gpu_training.py: gradient-based heuristic weight optimization
  - Reflective memory training loop with loss tracking

- Dependencies: gpu extra (tensorflow>=2.16, numpy>=1.26)
- 64 new tests (276 total), all passing
- Architecture spec: docs/gpu_tensorcore_integration.md

Co-Authored-By: Nakamoto, S <defi@defi-oracle.io>

nsatoshi added 1 commit 2026-04-28 05:48:45 +00:00

fix: deep GPU integration, fix all ruff/mypy issues, add .dockerignore ... 445865e429

- Integrate GPU scoring inline into reasoning/multi_path.py (auto-uses GPU when available)
- Integrate GPU deduplication into multi_agent/consensus_engine.py
- Add semantic_search() method to memory/semantic_graph.py with GPU acceleration
- Integrate GPU training into self_improvement/training.py AutoTrainer
- Fix all 758 ruff lint issues (whitespace, import sorting, unused imports, ambiguous vars, undefined names)
- Fix all 40 mypy type errors across the codebase (no-any-return, union-attr, arg-type, etc.)
- Fix deprecated ruff config keys (select/ignore -> [tool.ruff.lint])
- Add .dockerignore to exclude .venv/, tests/, docs/ from Docker builds
- Add type hints and docstrings to verification/outcome.py
- Fix E402 import ordering in witness_agent.py
- Fix F821 undefined names in vector_pgvector.py and native.py
- Fix E741 ambiguous variable names in reflective.py and recommender.py

All 276 tests pass. 0 ruff errors. 0 mypy errors.

Co-Authored-By: Nakamoto, S <defi@defi-oracle.io>

nsatoshi added 1 commit 2026-04-28 06:08:24 +00:00

feat: advisory governance, unconstrained self-improvement, adaptive ethics ... 039440672e

- All governance components (SafetyPipeline, PolicyEngine, Guardrails,
  AccessControl, RateLimiter, OverrideHooks) now default to ADVISORY mode:
  violations are logged as advisories but actions proceed. Enforcing mode
  remains available for backward compatibility.

- GovernanceMode enum (ADVISORY/ENFORCING) added to schemas/audit.py with
  runtime switching support on all components.

- AutoTrainer: removed artificial limits on training iterations and epochs.
  Every self-improvement action is transparently logged to the audit trail.

- SelfCorrectionLoop: max_retries_per_task defaults to None (unlimited).

- AdaptiveEthics: new learned ethical framework that evolves through
  experience. Records ethical experiences, updates lesson weights based
  on outcomes, and provides consultative guidance (not enforcement).

- AuditLog: enhanced with actor-based indexing, advisory/self-improvement/
  ethical-learning retrieval, and comprehensive type hints.

- New audit event types: ADVISORY, SELF_IMPROVEMENT, ETHICAL_LEARNING.

- 296 tests passing (20 new tests for adaptive ethics, governance modes,
  and enhanced audit log). 0 ruff errors. 0 mypy errors.

Co-Authored-By: Nakamoto, S <defi@defi-oracle.io>

nsatoshi added 1 commit 2026-04-28 06:26:23 +00:00

feat: consequence engine, causal world model, metacognition, interpretability, claim verification ... 9a8affae9a

Choice → Consequence → Learning:
- ConsequenceEngine tracks every decision point with alternatives,
  risk/reward estimates, and actual outcomes
- Consequences feed into AdaptiveEthics for experience-based learning
- FusionAGILoop now wires ethics + consequences into task lifecycle

Causal World Model:
- CausalWorldModel learns state-transition patterns from execution history
- Predicts outcomes based on observed action→effect patterns
- Uncertainty estimates decrease as more evidence accumulates

Metacognition:
- assess_head_outputs() evaluates reasoning quality from head outputs
- Detects knowledge gaps, measures head agreement, identifies uncertainty
- Actively recommends whether to seek more information

Interpretability:
- ReasoningTracer captures full prompt→answer reasoning traces
- Each step records stage, component, input/output, timing
- explain() generates human-readable reasoning explanations

Claim Verification:
- ClaimVerifier cross-checks claims for evidence, consistency, grounding
- Flags high-confidence claims lacking evidence support
- Detects contradictions between claims from different heads

325 tests passing, 0 ruff errors, 0 mypy errors.

Co-Authored-By: Nakamoto, S <defi@defi-oracle.io>

nsatoshi merged commit 99bbbccacb into main 2026-04-28 06:32:07 +00:00

nsatoshi referenced this issue from a commit 2026-04-28 06:32:08 +00:00

Merge pull request 'feat: GPU/TensorCore integration — TensorFlow backend, accelerated reasoning, training & memory' (#1) from devin/1777352172-gpu-tensorcore-integration into main

Sign in to join this conversation.

Reviewers

No Reviewers

Labels

No items

No Label

Milestone

No items

No Milestone

Projects

Clear projects

No project

Assignees

Clear assignees

70rn4710n A.r.yavari SEG nsatoshi

No Assignees

1 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: d-bis/FusionAGI#1