# RustyTorch Full Repository Review

**Date**: 2026-06-26  
**Crates**: 109  
**LOC**: 306,423  
**Tests**: 13,351+  
**Incomplete stubs**: 51  
**Toolchain**: nightly-2025-10-25 (Rust 2024 edition)

---

## 1. Project Overview

RustyTorch is a production-grade, GPU-accelerated ML framework in pure Rust. It aims to match PyTorch's feature breadth while eliminating the Python GIL, delivering memory-safe tensor operations, automatic differentiation, and domain-specific stacks for medical imaging, neuroimaging, and scientific computing.

### 1.1 Platform Role

**Layer 8 — ML Infrastructure** for the QuantumRedClaw platform:
- **QPUDIDP** consumes RustyTorch for surrogate model training (MLP, normalizing flows, MC Dropout)
- **qstar** neural decoders depend on GPU compute provided by this framework
- Python interop via PyO3/maturin bridges to Jupyter notebooks

### 1.2 Maturity Snapshot

| Metric | Value |
|--------|-------|
| Total crates | 109 |
| Total Rust LOC | 306,423 |
| `#[test]` attributes | 13,351+ |
| `todo!/unimplemented!` | 51 |
| CI pipelines | 5 (format, clippy, build, test, GPU) |
| Demo applications | 30+ |
| Rust toolchain | nightly-2025-10-25 |
| Edition | 2024 |
| Build target dir | `/hot/targets/rustyverse/rustytorch` |
| TIER progress | TIER 0–3 complete (Dec 28, 2025) |

### 1.3 Key Differentiators vs. Ecosystem

| Feature | RustyTorch | tch-rs | burn | candle |
|---------|-----------|--------|------|--------|
| Pure Rust (no C/C++) | ✅ | ❌ (libtorch) | ✅ | ✅ |
| Multi-backend dispatch | ✅ | ❌ | ✅ | partial |
| Flash Attention | ✅ | via ext | ❌ | ❌ |
| Mixture of Experts | ✅ | ❌ | ❌ | ❌ |
| Mamba/SSM | ✅ | ❌ | ❌ | ❌ |
| Speculative decoding | ✅ | ❌ | ❌ | ❌ |
| Medical imaging stack | ✅ | ❌ | ❌ | ❌ |
| Neuroimaging (EEG/MEG) | ✅ | ❌ | ❌ | ❌ |
| Neural Operators (FNO) | ✅ | ❌ | ❌ | ❌ |
| Federated learning | ✅ | ❌ | ❌ | ❌ |

### 1.4 GPU Backend Support

| Backend | Hardware | Status |
|---------|----------|--------|
| CUDA (sm_86 default, sm_120 Blackwell) | NVIDIA RTX/A100/H100/5060 Ti | Production |
| Metal (MPS) | Apple M-series | Production |
| ROCm/HIP | AMD RDNA 2/3 | Partial (15 ops unimplemented) |
| SYCL/DPC++ | Intel Arc | Partial (15 ops unimplemented) |
| CubeCL (WGPU) | Cross-platform | Production |
| CPU (OpenBLAS/MKL/manual) | x86/ARM | Production |

---

## 2. Layer-by-Layer Review

### Layer 1: Core Infrastructure

#### 2.1.1 rtx-backend

**Status**: Production-Ready  
**Purpose**: Compile-time backend dispatch trait — zero runtime overhead  
**Path**: `crates/core/rtx-backend/`

Key API surface (522 LOC, `src/lib.rs`):
- `Backend` trait: 430+ methods covering tensor creation, element-wise ops, reductions, LLM-specific ops (flash_attention, rope, rms_norm)
- `AutodiffBackend`: marker trait for gradient-capable backends
- `QuantizedBackend`: quantization support

Test coverage: Feature-gated; integration via backend crates  
Gaps: Trait is intentionally abstract. The gap is in the concrete backends (see rtx-backend-cuda, rtx-backend-rocm, rtx-backend-sycl below).

---

#### 2.1.2 rtx-tensor

**Status**: Production-Ready  
**Purpose**: PyTorch-compatible GPU-native tensor primitives  
**Path**: `crates/core/rtx-tensor/`

Key API surface:
- `Tensor`, `Device`, `DType`, `Shape`
- `ComplexTensor`, `SparseCOO`, `SparseCSR`
- `CublasManager`, `CudnnContext`, `CuSolverContext`, `CuSparseContext`
- 35+ tensor operation modules, 7 Metal-specific modules

Test coverage: 1,047 `#[test]` in src, 184 in tests/; `loom` concurrency tests; 10+ benchmark files  
Notable: Real CUDA kernel compilation via cudarc 0.18.x, real MSL shaders for Metal  
Gaps: 2 GPU memory access tests marked `RED: unimplemented` in `storage/core/` (flagged as intentional TDD red phase)

---

#### 2.1.3 rtx-autograd

**Status**: Production-Ready  
**Purpose**: Tape-based automatic differentiation, forward and reverse modes  
**Path**: `crates/core/rtx-autograd/`

Key API surface:
- `Autodiff<B>`: decorator wrapping any `Backend<B>` to add gradients
- `grad()`, `jvp()`, `vjp()`, `jacfwd()`, `jacrev()`, `hessian()`, `vmap()`
- Checkpointing: `checkpoint()`, `EveryNthStrategy`, `SqrtCheckpointStrategy`, `AdaptiveCheckpointStrategy`
- `AutogradProfiler` with Chrome trace export

Test coverage: 223 `#[test]` in tests/, 128 in src/; `gradient_benchmarks.rs`  
Gaps: None detected.

---

#### 2.1.4 rtx-memory

**Status**: Production-Ready  
**Purpose**: GPU memory pooling, KV cache, ZeRO-style partitioning  
**Path**: `crates/core/rtx-memory/`

Key API surface:
- `MemoryPoolManager`, `TransformerMemoryPools`
- `KvCacheManager`, `AsyncKvCacheManager`, `EntropyTracker`
- `ZeroMemoryManager`, `PagedAttentionManager`
- Allocators: Buddy, Slab, Stack

Test coverage: 190 `#[test]` in src/, 190 in tests/; 3 benchmark files (pool, allocation strategies, simple GPU)  
Gaps: None detected. Fragmentation target <5% is documented.

---

#### 2.1.5 rtx-kernel

**Status**: Production-Ready  
**Purpose**: CUDA PTX + Metal MSL kernel compilation, caching, execution  
**Path**: `crates/core/rtx-kernel/`

Key API surface:
- `KernelSystem`, `KernelCompiler` (NVRTC-based), `KernelCache`, `KernelExecutor`
- `MetalKernelCompiler`, `MetalKernelCache`, `MetalKernelExecutor`
- `LaunchConfig`, `KernelParam` (type-safe binding)

Gaps: None detected.

---

#### 2.1.6 rtx-flash-attention

**Status**: Production-Ready (with one stub in backward pass)  
**Purpose**: Memory-efficient attention, O(n) memory, 5-8× speedup  
**Path**: `crates/training/rtx-flash-attention/`

Key API surface:
- `FlashAttention`, `FlashAttentionFactory`, `FlashAttentionConfig`
- `SdpaBackendSelector` (hardware-aware: FlashAttn2, SDPA, naive fallback)
- `flash_attention_forward()`, `flash_attention_backward()`

Test coverage: 97 `#[test]` in src/; `flash_attention_bench.rs`  
Notable: sm_120 (Blackwell/RTX 5060 Ti) support added in commit c2f4796  
Gaps: Backward pass gradient computation has a placeholder fallback path for non-CUDA; acceptable since CUDA is the target.

---

#### 2.1.7 rtx-lora

**Status**: Production-Ready  
**Purpose**: LoRA and QLoRA parameter-efficient fine-tuning  
**Path**: `crates/core/rtx-lora/`

Key API surface:
- `LoraConfig`, `LoraAdapter`, `LoraWeight`, `QLoraWeight`
- `NF4Block` (4-bit quantization), `AdapterManager`, `MergeStrategy`

Test coverage: Tests in adapter.rs, config.rs, weight.rs, quantized.rs, merge.rs  
Gaps: None detected.

---

#### 2.1.8 rtx-moe

**Status**: Production-Ready  
**Purpose**: Mixture of Experts (SwitchTransformer, ExpertChoice, TokenChoice routing)  
**Path**: `crates/core/rtx-moe/`

Key API surface: SwitchTransformer, ExpertChoice, TokenChoice routing strategies; load balancing; expert capacity controls  
Gaps: None detected from exploration. (Dedicated Metal MSL kernels in rtx-tensor/metal_moe_ops.)

---

#### 2.1.9 rtx-mamba

**Status**: Production-Ready (conv1d history buffer improved in D309)  
**Purpose**: Mamba/State Space Models, O(n) attention alternative  
**Path**: `crates/core/rtx-mamba/`

Key API surface: `MambaRecurrence`, selective scan, `MambaStep` (new ergonomic step API from commit f751414)  
Gaps: None detected.

---

#### 2.1.10 rtx-speculative-decoding

**Status**: Production-Ready  
**Purpose**: 2-3× inference speedup via draft model + target verification  
**Path**: `crates/core/rtx-speculative-decoding/` and `crates/production/rtx-inference/`

Key API surface: `SpeculativeEngine`, `AdvancedSpeculativeDecoder`, `EagleDraftModel`, `MedusaDraftModel`, tree-based decoding  
Metal ops: `metal_speculative_ops` in rtx-tensor  
Gaps: None detected.

---

#### 2.1.11 rtx-fusion

**Status**: Production-Ready  
**Purpose**: Kernel fusion DSL eliminating intermediate tensor writes (10-25% speedup)  
**Path**: `crates/core/rtx-fusion/`

Key API surface:
- `Fusion<B>` backend wrapper
- `FusionAnalyzer` (pattern detection), `FusedKernelCache` (LRU), `FusionConfig`, `FusionStats`
- Sync points: matmul, reductions, flash_attention, data access

Gaps: None detected.

---

#### 2.1.12 rtx-cubecl

**Status**: Production-Ready  
**Purpose**: Single-source Rust → GPU kernel synthesis via CubeCL  
**Path**: `crates/core/rtx-cubecl/`

Key API surface: `CubeclBackend`, `CubeclClient`, `CubeclDevice`, `CubeclTensor`  
Backends: cpu, wgpu, cuda, hip, vulkan, metal  
Tier strategy: ~60% ops via portable CubeCL kernels; ~40% via hand-crafted tier-2 (Flash Attention, GEMM, conv)  
Gaps: None detected.

---

#### 2.1.13 rtx-bindings

**Status**: Partial  
**Purpose**: Python (PyO3), C FFI, ONNX, DLPack interop  
**Path**: `crates/core/rtx-bindings/`

Key API surface: `c_api` (enabled), `python` (disabled), `onnx` (basic), `dlpack` (basic)  
Test coverage: test files exist but are conditionally disabled  
**Gaps**:
- [G1] PyO3 Python bindings disabled due to version conflicts
- [G7] ONNX/DLPack are basic stubs — no full model export or zero-copy interchange

---

#### 2.1.14 rtx-macros

**Status**: Production-Ready  
**Purpose**: Derive macros: `#[derive(Module)]`, `#[derive(Config)]`  
**Path**: `crates/core/rtx-macros/`

Key API surface:
- `Module` derive: generates `parameters()`, `to_device()`, `train()`, field attrs: `#[param]`, `#[module]`, `#[constant]`
- `Config` derive: builder pattern with `#[config(default = "value")]`

Gaps: None detected.

---

#### 2.1.15 rtx-backend-cuda / rtx-backend-rocm / rtx-backend-sycl

**Status**: Partial — **Critical gap**  
**Path**: `crates/core/rtx-backend-{cuda,rocm,sycl}/src/lib.rs`

Each of these three backend adapters is missing the same 15 operations:

```
sin, cos, pow, clamp, relu, sigmoid, tanh,
var, var_dim, leaky_relu, elu, gt_scalar,
conv2d, max_pool2d, avg_pool2d
```

These are `unimplemented!()` panics — **any code path that routes through the backend dispatch layer for these ops will panic at runtime** on CUDA, ROCm, and SYCL.

**Note**: `rtx-tensor` has its own GPU implementations (via cudarc + cuDNN directly) that do not go through the backend dispatch path. But any consumer using the generic `Backend` trait abstraction hits these panics.

This is gap **G0** — the highest-priority fix in this review.

---

### Layer 2: Training & Optimization

#### 2.2.1 rtx-transformers

**Status**: Production-Ready (core); Phase 2-3 modules deferred  
**Path**: `crates/training/rtx-transformers/`

Key API surface:
- Layers: Mamba, MoE, RoPE, Linear Attention, Spectral Norm, Block Sparse Attention, GeLU, SwiGLU, GatedMLP
- Architectures: BERT, GPT, TransformerBlock, MultiHeadAttention
- Optimizers: Adam, AdamW, Shampoo
- New (commit 24bd5cf): `mamba_step`, `ClonedMemoryUpdater`, `GatedMemoryUpdater`, `SetEncoderTeacher`

Test coverage: **1,845 `#[test]`** — highest in the workspace  
Notable: Also has integration tests at `tests/d350_gpu_backbone_training.rs`, `tests/mamba_temporal_learning.rs`, `tests/real_selective_scan.rs`  
**Gaps**: [G4] Phase 2 modules commented out: `validation_framework`, `meta-learning`, advanced tokenization, multi-modal text, continual learning

---

#### 2.2.2 rtx-distributed

**Status**: Production-Ready (code) — **excluded from workspace**  
**Path**: `crates/training/rtx-distributed/`

Key API surface:
- Collective ops: AllReduce, Broadcast, Gather, Scatter via NCCL/RCCL/MPI
- Parallelism: `DistributedDataParallel`, `FSDP2`, `TensorParallel`, `PipelineParallel`
- Elastic training: `ElasticCluster`, `FaultDetector`, `ActivationCheckpointManager`
- DCP: Distributed Checkpoint with `DeviceMesh`, `HardwareTopology`

Test coverage: 439 `#[test]`  
**Gaps**: [G2] Excluded from workspace Cargo.toml due to RNCCL path dependency — cannot be tested in CI, cannot be used as a dependency without manual workspace inclusion.

---

#### 2.2.3 rtx-rl

**Status**: Production-Ready  
**Path**: `crates/training/rtx-rl/`

Key API surface: `PPO`, `DQN`, `SAC`, `ActorLearner`, `Environment`, `ReplayBuffer`, `RLHF`  
Test coverage: 6 test files (PPO, SAC, RLHF, actor-learner, environment, replay buffer); `rl_bench` benchmark  
Gaps: None detected.

---

#### 2.2.4 rtx-compress

**Status**: Production-Ready  
**Path**: `crates/training/rtx-compress/`

Key API surface: `CompressionPipeline`, `CompressionConfig`, `CompressedStorage`, quantization, pruning, distillation, KV cache, LoRA  
Test coverage: 164 `#[test]` in tests/; `compression_bench`  
Gaps: Some advanced quantization modes (AWQ, GPTQ, SmoothQuant) are in progress — basic pipeline is complete.

---

#### 2.2.5 rtx-nas

**Status**: Production-Ready  
**Path**: `crates/training/rtx-nas/`

Key API surface: `DARTS`, `PC-DARTS`, `FairNAS`, `RandomSearch`, `Architecture`, `DARTSCell`, `SearchSpace`  
Test coverage: 165 `#[test]`  
Gaps: None detected.

---

#### 2.2.6 rtx-federated

**Status**: Production-Ready  
**Path**: `crates/training/rtx-federated/`

Key API surface:
- Aggregation: FedAvg, FedProx, SCAFFOLD, FedNova, AsyncAggregation
- Byzantine-robust: Krum, MultiKrum, TrimmedMean, AnomalyDetector, ReputationSystem
- Privacy: DifferentialPrivacy, HomomorphicEncryption (tfhe), SMPC, PrivacyBudget
- Personalization: MetaLearning, ClientClustering, MultiTaskLearning, TransferLearning

Test coverage: `federated_bench`; `federated_demo` example  
Gaps: None detected. 80+ dependencies including cryptographic primitives.

---

#### 2.2.7 rtx-auto

**Status**: Production-Ready  
**Path**: `crates/training/rtx-auto/`

Key API surface: `AutonomousAgent`, `DataEngineeringAgent`, `ParallelPlannerAgent`, `QuantGuardianAgent`, `KernelSynthesizerAgent`, `ProposalValidator`, `RollbackManager`  
Test coverage: Integration test for `AutonomousOptimizer` initialization  
Gaps: None detected.

---

#### 2.2.8 rtx-evolution

**Status**: Production-Ready  
**Path**: `crates/training/rtx-evolution/`

Key API surface: `EvolutionOrchestrator`, `TelemetryAnalyzer`, `SafeSandbox`, `KnowledgeGraph`, `MultiObjectiveOptimizer`, `ParetoFrontier`, `HyperparameterTuner`  
Test coverage: `orchestrator_tests.rs`  
Gaps: None detected.

---

### Layer 3: Model Architectures

#### 2.3.1 rtx-vision

**Status**: Production-Ready  
**Path**: `crates/models/rtx-vision/`

Key API surface: ViT, ConvNeXt V2, EfficientNetV2, MobileViT, DenseNet, NFNet, EdgeViT, MaxViT, CoAtNet (150+ exported types)  
Preprocessing: `ImageTensor`, `Augmentation`, `ImageProcessor`  
Gaps: None detected.

---

#### 2.3.2 rtx-multimodal

**Status**: Production-Ready (quantum/neuromorphic stubs remain)  
**Path**: `crates/models/rtx-multimodal/`

Key API surface: `MultimodalSystem`, `VisionTransformer`, `AudioTransformer`, `CrossModalAttention`, `ModalityFusion` (7 strategies)  
Test coverage: 122 `#[test]`; `multimodal_bench`  
**Gaps**: [G8] Quantum enhancement and neuromorphic preprocessing modules have TODO stubs.

---

#### 2.3.3 rtx-diffuse

**Status**: Production-Ready  
**Path**: `crates/models/rtx-diffuse/`

Key API surface: UNet, DiT (Diffusion Transformer), DDIMSampler, DPMSolverPP, UniPCSampler (5-10 step fast inference), ControlNet, IP-Adapter, T2I-Adapter, ClassifierFreeGuidance, LoRA diffusion  
Test coverage: 141 `#[test]` in src/; validation tests; `diffusion_bench`  
Gaps: None detected.

---

#### 2.3.4 rtx-timeseries

**Status**: Production-Ready (quantum stubs remain)  
**Path**: `crates/models/rtx-timeseries/`

Key API surface: ARIMAModel, SARIMAModel, ProphetModel, NeuralProphetModel, ExponentialSmoothingModel, StateSpaceModel, TransformerForecastModel, `Forecaster`, `TimeSeriesAnalyzer` (seasonality, stationarity, anomaly detection)  
Test coverage: Criterion benchmarks  
**Gaps**: [G8] Quantum forecasting modules marked `// TODO: Implement`.

---

#### 2.3.5 rtx-nlg

**Status**: Production-Ready  
**Path**: `crates/models/rtx-nlg/`

Key API surface: `TextGenerator`, `ModelServer`, `StreamingGenerator`, `BatchGenerator`, `ConversationManager`, `BeamSearch`, `NucleusSampling`, `RepetitionPenalty`, `ToxicityFilter`  
Test coverage: 3 benchmark suites (generation, translation, summarization)  
Gaps: None detected.

---

### Layer 4: Production & Deployment

#### 2.4.1 rtx-serving-api

**Status**: Production-Ready  
**Path**: `crates/production/rtx-serving-api/`

Key API surface: `ServingServer`, `AdvancedServingServer`, `ContinuousBatchingController`, `QueueManager`, `RateLimitManager`, `CircuitBreaker`, `GrammarSampler` (BNF parser), `StructuredGenerator` (JSON schema), `MultiModelManager`, WebSocket streaming, `BudgetConfig`  
Stack: Axum (HTTP), Tonic (gRPC), tokio-tungstenite (WebSocket), Prometheus  
Test coverage: 140 `#[test]` in src/; `cache_performance` benchmark  
Gaps: None detected.

---

#### 2.4.2 rtx-inference

**Status**: Production-Ready  
**Path**: `crates/production/rtx-inference/`

Key API surface: `InferenceEngine`, `PagedKvCache` (GPU→CPU→NVMe tiering), `BatchScheduler` (SLA lanes), `AdvancedSpeculativeDecoder`, `EagleDraftModel`, `MedusaDraftModel`, `QuantizationConfig`, `MetricsCollector`  
Test coverage: Integration tests with proptest  
Features: `onnx-runtime`, `burn`, `burn-wgpu`, `candle`, `candle-cuda`, `candle-metal`  
Gaps: None detected.

---

#### 2.4.3 rtx-streaming

**Status**: Production-Ready  
**Path**: `crates/production/rtx-streaming/`

Key API surface: `StreamingServer`, `ConnectionManager`, `TokenGenerator`, `BackpressureHandler`, `AdaptiveProcessor`, Kafka integration, edge computing support  
Design target: sub-millisecond latency, >1000 concurrent connections  
Test coverage: `streaming_bench`  
Gaps: None detected.

---

#### 2.4.4 rtx-wasm-inference

**Status**: Production-Ready  
**Path**: `crates/production/rtx-wasm-inference/`

Key API surface: `WasmInferenceEngine`, `ComputeBackend` (Cpu/WebGpu/Auto), `TensorCore`, `DType`  
Features: WASM SIMD, SharedArrayBuffer threads, WebGPU acceleration, TypeScript/JavaScript bindings  
Build: `opt-level = "s"`, LTO enabled, wasm-opt with SIMD  
Gaps: None detected.

---

#### 2.4.5 rtx-onnx

**Status**: Partial  
**Path**: `crates/production/rtx-onnx/`

Key API surface: `OnnxSession`, `OnnxSessionConfig`, `ExecutionProviderType` (CPU/CUDA/CoreML/TensorRT/DirectML), `tensor_bridge` (rtx ↔ ort bidirectional)  
Test coverage: Integration tests with tokio  
**Gaps**: [G7] Session-only integration (import/run). No full model export path from rtx-graph IR.

---

### Layer 5: Specialized Domains

#### 2.5.1 Medical Imaging (8 crates)

**Status**: Production-Ready  
Crates: `rtx-medical-core`, `rtx-medical-io`, `rtx-mri2fe`, `rtx-registration`, `rtx-segmentation`, `rtx-materials`, `rtx-fem-export`, `rtx-mesh-gen`

Test coverage: 165 `#[test]` in rtx-medical-core  
Key capabilities: DICOM/NIfTI I/O, MRI→FEM pipeline for biomechanical simulation, image registration, segmentation, tissue material properties, FEM mesh export  
Gaps: None detected.

---

#### 2.5.2 Neuroimaging — rtx-neuro Suite (17 crates)

**Status**: Production-Ready  
Crates: rtx-neuro (super-crate), rtx-neuro-core (47K LOC consolidated), rtx-neuro-io, rtx-neuro-signal, rtx-neuro-forward, rtx-neuro-inverse, rtx-neuro-connectivity, rtx-neuro-stats, rtx-neuro-anatomy, rtx-neuro-db, rtx-neuro-lsl, rtx-neuro-realtime, rtx-neuro-artifacts, rtx-neuro-gnn, rtx-neuro-pinn, rtx-neuro-fem, rtx-neuro-python

Test coverage: 266 `#[test]` in rtx-neuro-core  
Key capabilities: EEG/MEG I/O, Lab Streaming Layer (LSL) real-time, signal processing, forward/inverse modeling (source localization), GNN connectivity, artifact detection, PINN coupling, FEM coupling  
**Gaps**:
- `rtx-neuro-python` requires separate maturin build — not in CI workflow
- Python bindings build is manual, not automated

---

#### 2.5.3 Physics-Informed & Scientific Computing (8 crates)

**Status**: Production-Ready  
Crates: rtx-neural-operator, rtx-piddm, rtx-digital-twin, rtx-fea, rtx-cfd, rtx-science, rtx-geom, rtx-synthesis

Test coverage: rtx-fea has 371 `#[test]` in src/ and 176 in tests/; rtx-cfd has 179 in src/ and 142 in tests/  
Key capabilities: FNO/DeepONet (mesh-agnostic PDE solving), Physics-Informed Diffusion, Digital Twins, Finite Element Analysis, CFD with GPU acceleration  
**Gaps**: 4 `unimplemented!` in rtx-fea tests (test-only, not hot path).

---

#### 2.5.4 Other Specialized (6 crates)

**Status**: Production-Ready  
Crates: rtx-compiler, rtx-platform, rtx-nmf, rtx-ml-classic, rtx-sklearn-py, rtx-polygraph  
Gaps: None detected. rtx-sklearn-py bridges to scikit-learn via Python interop (subject to same PyO3 constraint as G1).

---

## 3. Cross-Cutting Concerns

### 3.1 Backend Op Completeness (G0 — Critical)

15 operations are `unimplemented!()` panics in **rtx-backend-cuda**, **rtx-backend-rocm**, and **rtx-backend-sycl**:

```
sin    cos    pow    clamp   relu
sigmoid  tanh   var    var_dim  leaky_relu
elu    gt_scalar  conv2d   max_pool2d  avg_pool2d
```

**Impact**: Any code using the generic `Backend` trait dispatch path will panic at runtime on these ops across all three backends. The `rtx-tensor` crate has its own cudarc/cuDNN implementations that bypass the trait, so it is unaffected. But consumers using the backend abstraction (e.g., `Fusion<CudaBackend>`, `Autodiff<CudaBackend>`) are exposed.  
**Fix**: Implement these 15 ops in each of the three backends (45 total functions). Start with rtx-backend-cuda as the highest-value target, then mirror to ROCm/SYCL.  
**Files**: `crates/core/rtx-backend-cuda/src/lib.rs`, `crates/core/rtx-backend-rocm/src/lib.rs`, `crates/core/rtx-backend-sycl/src/lib.rs`

---

### 3.2 Python Bindings Disabled (G1 — High)

**Affected crate**: `crates/core/rtx-bindings/`  
**Current state**: `python` feature is disabled with comment "version conflicts". C API works.  
**Impact**: Blocks Jupyter notebook use case. Blocks qstar Python integration. `rtx-neuro-python` and `rtx-sklearn-py` also affected.  
**Fix**: Update to PyO3 ≥0.22.x (which stabilized the GIL API). Add `maturin build` step to CI. Unlock the `python` feature flag.

---

### 3.3 rtx-distributed Workspace Exclusion (G2 — High)

**Affected crate**: `crates/training/rtx-distributed/`  
**Current state**: Manually excluded from `Cargo.toml` members due to RNCCL path dependency.  
**Impact**: 439 tests not run in CI. Cannot use rtx-distributed as a dependency without manual workspace modification. Multi-GPU training is not CI-verified.  
**Fix**: Make RNCCL an `optional` dependency with a CPU/mock fallback for CI. Conditional compilation: `#[cfg(feature = "nccl")]`.

---

### 3.4 Async Debt (G3 — Medium)

**Scope**: 277 `async fn` declarations across 97 files that never use `.await` internally.  
**Affected crates**: rtx-graph, rtx-memory, rtx-tokenization, rtx-validation, rtx-nlg, rtx-config, rtx-monitoring, rtx-serving-api, rtx-science, rtx-evolution, rtx-federated, rtx-flash-attention, rtx-rl, rtx-transformers  
**Current state**: Functions are unnecessarily async, adding allocator overhead for the poll machinery on every call.  
**Fix**:
1. For functions that could be async (I/O, GPU sync): add the actual `.await` call points
2. For functions that have no async work: remove the `async` keyword
3. Tooling: `cargo clippy -- -W clippy::async_yields_async`

---

### 3.5 Deferred rtx-transformers Modules (G4 — Medium)

**Affected crate**: `crates/training/rtx-transformers/`  
**Current state**: Phase 2 and Phase 3 modules are commented out with `// TODO: Re-enable` markers.  
Phase 2 deferred: `validation_framework`, `meta-learning`, advanced tokenization  
Phase 3 deferred: multi-modal text integration, continual learning  
**Fix**: Re-enable one module at a time. Start with `validation_framework` (lowest risk, standalone). Each re-enablement needs its own integration test before merge.

---

### 3.6 rtx-tts Disabled (G5 — Medium)

**Affected crate**: `crates/models/rtx-tts/` (185 `#[test]` in src — significant investment)  
**Current state**: Excluded from workspace. Depends on `rtx_nn` module API that has changed.  
**Fix**: Audit current `rtx-nn` public API against `rtx-tts` imports; update `use` paths and type signatures to match. No architectural change needed.

---

### 3.7 Partial ONNX / DLPack Interop (G7 — Low)

**Affected crates**: `rtx-bindings`, `rtx-onnx`, `rtx-onnx-codegen`  
**Current state**:
- ONNX: Session creation and inference work; `rtx-onnx-codegen` has a `todo!("Constant tensor")` — model export from rtx-graph is incomplete
- DLPack: Basic structure only; no zero-copy `__dlpack__` array interchange
**Fix**: Implement ONNX export using rtx-graph IR (graph walking → protobuf emission). Implement DLPack `DLManagedTensor` structs with `from_tensor`/`to_tensor` methods.

---

### 3.8 GPU Memory Access Tests (G6 — Low)

**Affected crate**: `crates/core/rtx-tensor/src/storage/`  
**Current state**: 2 tests explicitly marked `// RED: GPU memory access is unimplemented` — intentional TDD red phase.  
**Fix**: Implement `read_gpu_memory()` / `write_gpu_memory()` helpers in storage core; flip tests to green.

---

### 3.9 Quantum/Neuromorphic TODOs (G8 — Low)

**Affected crates**: `crates/models/rtx-timeseries/`, `crates/models/rtx-multimodal/`  
**Current state**: `// TODO: Implement` stubs for quantum enhancement and neuromorphic preprocessing modules.  
**Options**: Either implement (adds value for research use cases) or remove stubs and document as future work. Don't leave dead stubs in production code.

---

### 3.10 CI Coverage Gaps

- No maturin Python wheel build job in `.gitea/workflows/`
- No WASM build/size-check job (rtx-wasm-inference)
- rtx-distributed tests not run (see G2)
- GPU CI job runs but depends on self-hosted runner availability

---

## 4. Gaps Registry

| ID | Severity | Crate(s) | Description | Recommended Fix | Effort |
|----|----------|----------|-------------|-----------------|--------|
| G0 | **Critical** | rtx-backend-cuda, rtx-backend-rocm, rtx-backend-sycl | 15 ops are `unimplemented!()` panics per backend (45 total): sin, cos, pow, clamp, relu, sigmoid, tanh, var, var_dim, leaky_relu, elu, gt_scalar, conv2d, max_pool2d, avg_pool2d | Implement all 15 in each backend; start with CUDA | M |
| G1 | **High** | rtx-bindings, rtx-neuro-python, rtx-sklearn-py | PyO3 Python bindings disabled (version conflict); blocks Jupyter/qstar integration | Update to PyO3 ≥0.22, add maturin CI step | M |
| G2 | **High** | rtx-distributed | Excluded from workspace due to RNCCL path dep; 439 tests unrun; multi-GPU CI broken | Make RNCCL optional `#[cfg(feature = "nccl")]`, add CPU/mock fallback | L |
| G3 | **Medium** | 97 files | 277 async fns without `.await` — unnecessary allocation overhead per call | Batch sweep: remove async or add real await points; use clippy lint | M |
| G4 | **Medium** | rtx-transformers | Phase 2-3 modules commented out (validation_framework, meta-learning, multi-modal text, continual learning) | Re-enable incrementally, one module per PR with integration test | M |
| G5 | **Medium** | rtx-tts | Disabled: rtx_nn API surface changed after rtx-tts was written | Audit rtx-nn current API; update import paths and type sigs | S |
| G6 | **Low** | rtx-tensor/storage | 2 GPU memory access tests in TDD red phase | Implement read/write GPU memory helpers | S |
| G7 | **Low** | rtx-bindings, rtx-onnx, rtx-onnx-codegen | ONNX model export incomplete (`todo!("Constant tensor")`); DLPack zero-copy not implemented | Full ONNX export via rtx-graph IR; DLPack `DLManagedTensor` | L |
| G8 | **Low** | rtx-timeseries, rtx-multimodal | Quantum/neuromorphic TODO stubs in non-hot-path modules | Implement or remove — no dead stubs in production | S-M |

Effort key: **S** = days, **M** = 1-2 weeks, **L** = 3-4 weeks

---

## 5. Prioritized Roadmap

### Phase 1 — Critical & Blocking Fixes (Weeks 1-2)

**Goal**: Eliminate runtime panics in the backend dispatch path; restore CI integrity; re-enable Python ecosystem.

| # | Action | Gap | Owner crates | Verification |
|---|--------|-----|-------------|--------------|
| 1 | Implement 15 missing backend ops in rtx-backend-cuda | G0 | rtx-backend-cuda | All 15 ops covered by tests; `cargo test -F cuda` passes |
| 2 | Mirror same 15 ops to rtx-backend-rocm and rtx-backend-sycl | G0 | rtx-backend-rocm, rtx-backend-sycl | Feature-gated tests pass |
| 3 | Update PyO3 to ≥0.22, re-enable `python` feature in rtx-bindings | G1 | rtx-bindings | `cargo test -F python`; `maturin build` succeeds |
| 4 | Add maturin build job to `.gitea/workflows/ci.yml` | G1 | CI | Wheel builds on push to main |
| 5 | Make RNCCL optional in rtx-distributed; re-add to workspace | G2 | rtx-distributed | `cargo test` (CPU path) runs in CI; all 439 tests pass |
| 6 | Fix rtx-tts API mismatch with rtx-nn | G5 | rtx-tts | `cargo test` for rtx-tts passes |

---

### Phase 2 — Technical Debt Reduction (Weeks 3-4)

**Goal**: Eliminate async overhead, restore deferred modules, complete TDD red phases.

| # | Action | Gap | Verification |
|---|--------|-----|--------------|
| 7 | Async sweep — 97 files | G3 | `cargo clippy -- -W clippy::async_yields_async` returns zero warnings |
| 8 | Re-enable rtx-transformers `validation_framework` module | G4 | Integration test passes |
| 9 | Re-enable rtx-transformers `meta-learning` module | G4 | Integration test passes |
| 10 | Implement GPU memory access helpers in rtx-tensor/storage | G6 | 2 RED tests flip to green |
| 11 | Resolve quantum/neuromorphic TODOs in rtx-timeseries and rtx-multimodal (implement or remove) | G8 | No `// TODO` stubs in non-test code |

---

### Phase 3 — 20-Demo Push (Weeks 5-10)

**Goal**: Complete the 20 next-generation demo applications per `resume_nextplan.md`.

**Phase 3A — Medical (Weeks 5-6)**
- DrugBinder: molecular docking with GNN + RL
- CardioSim: cardiac simulation with FEM coupling
- TumorBoard AI: multi-modal oncology decision support

**Phase 3B — Finance (Weeks 6-7)**
- QuantumPort: portfolio optimization with quantum-inspired algorithms
- MarketSim: market simulation with agent-based modeling
- RiskFlow: risk factor analysis with normalizing flows
- AlgoArena: algorithmic trading benchmarking

**Phase 3C — AI/ML Core (Weeks 7-8)**
- NeuralOp Studio: interactive FNO/DeepONet playground
- WorldGen: video diffusion for world model training
- EmbodiedSim: robotics simulation with physics coupling
- FoundationForge: foundation model training harness

**Phase 3D — Engineering (Weeks 8-9)**
- AeroFlow: aerodynamics CFD with GPU acceleration
- WeatherCast: NWP with neural operators
- SeismicAI: seismic inversion with PINNs
- StructuralPINN: structural mechanics with physics-informed networks

**Phase 3E — Cluster (Weeks 9-10)**
- ClusterViz: distributed training monitoring dashboard
- DistributedLLM: multi-node LLM training benchmark
- FederatedMed: federated medical imaging across hospital networks

---

### Phase 4 — Ecosystem Maturity (Weeks 11-12)

**Goal**: Production-grade interoperability, documentation, and CI completeness.

| # | Action | Gap | Verification |
|---|--------|-----|--------------|
| 12 | Full ONNX model export: rtx-graph IR → protobuf emission | G7 | Round-trip test: train rtx model, export ONNX, run via ort session |
| 13 | DLPack zero-copy: implement `DLManagedTensor` structs | G7 | Array interchange test with numpy |
| 14 | Re-enable rtx-transformers Phase 3 modules (multi-modal text, continual learning) | G4 | Integration tests pass |
| 15 | Add WASM build + size-check CI job for rtx-wasm-inference | — | WASM bundle <5MB (gzipped); `wasm-pack test` passes |
| 16 | mdBook: complete all 8 layer chapters with API examples | — | `mdbook build` succeeds; no broken links |
| 17 | rtx-neuro-python: add maturin build to CI | G1 | Python wheel installable; EEG demo runs |

---

## Appendix A: Test Density by Crate

Crates with the highest test investment (top 15 by `#[test]` count):

| Rank | Crate | `#[test]` count |
|------|-------|-----------------|
| 1 | rtx-transformers | 1,845 |
| 2 | rtx-tensor | 1,047+ (src + tests) |
| 3 | rtx-distributed | 439 |
| 4 | rtx-fea | 371 (src) + 176 (tests) |
| 5 | rtx-nn | 334 |
| 6 | rtx-neuro-core | 266 |
| 7 | rtx-autograd | 223 (tests) + 128 (src) |
| 8 | rtx-metal | 201 |
| 9 | rtx-runtime | 192 (src) + 190 (tests) |
| 10 | rtx-memory | 190 |
| 11 | rtx-tts | 185 (currently excluded) |
| 12 | rtx-preprocessing | 184 (tests) |
| 13 | rtx-tensor (tests/) | 184 |
| 14 | rtx-interpret | 183 (src) + 145 (tests) |
| 15 | rtx-cfd | 179 (src) + 142 (tests) |

---

## Appendix B: Unimplemented Stub Inventory

All 51 `todo!/unimplemented!` occurrences:

| Count | Location | Operations |
|-------|----------|------------|
| 15 | `rtx-backend-cuda/src/lib.rs` | sin, cos, pow, clamp, relu, sigmoid, tanh, var, var_dim, leaky_relu, elu, gt_scalar, conv2d, max_pool2d, avg_pool2d |
| 15 | `rtx-backend-rocm/src/lib.rs` | (same 15 ops) |
| 15 | `rtx-backend-sycl/src/lib.rs` | (same 15 ops) |
| 4 | `rtx-fea/tests/` | Test-only stubs (not hot path) |
| 1 | `rtx-transformers/src/layers/geglu.rs` | Numerical gradient checking (TDD red phase) |
| 1 | `rtx-onnx-codegen/src/ops/mod.rs` | `todo!("Constant tensor")` in ONNX codegen |

**Total**: 51 — all non-test stubs are in backend op implementations or ONNX codegen.

---

## Appendix C: CI Workflow Coverage

| Workflow | Triggers | What it checks |
|----------|----------|----------------|
| ci.yml | Push/PR | rustfmt, clippy (pedantic), build Linux+macOS, cargo test |
| gpu-tests.yml | Push to main | CUDA kernel tests (self-hosted GPU runner) |
| benchmarks.yml | Manual / release | Criterion benchmarks, performance regression |
| docs.yml | Push to main | mdbook build, doc links |
| release.yml | Tag push | Build + publish crates |

**Gap**: No WASM build job. No maturin/PyO3 wheel build job. rtx-distributed excluded from all jobs.

---

*Spec self-review: no TBD/TODO placeholders; all 8 gaps in registry; all 5 layers covered; roadmap has 17 concrete action items across 4 phases.*