foxhunt/docs/superpowers/plans/2026-03-18-gpu-dedup-consolidation.md

GPU Implementation Deduplication Plan

Post-candle-elimination cleanup. Consolidate duplicate GPU tensor, linear, activation, and optimizer implementations.

Goal: Single canonical implementation per GPU primitive in ml-core, with ml-ppo CudaLinear kept for its distinct ownership model.

---

Phase 1: Delete ml-supervised GpuTensor/GpuLinear (CRITICAL)

• DELETE crates/ml-supervised/src/gpu_tensor.rs GpuTensor + GpuLinear definitions
• Re-export ml_core::cuda_autograd::{GpuTensor, GpuLinear} in ml-supervised
• Update all ml-supervised imports
• ~200 LOC reduction

Phase 2: Consolidate activations into ml-core

• MERGE crates/ml-ppo/src/cuda_nn/activations.rs CUDA kernels into ml-core
• DELETE crates/ml-ppo/src/cuda_nn/activations.rs
• Add thin wrapper in ml-ppo returning CudaVec for PPO compat
• ~400 LOC reduction

Phase 3: Merge AdamW kernels

• MERGE crates/ml-ppo/src/cuda_nn/adam.rs kernel source into ml-core GpuAdamW
• Keep wrapper in ml-ppo for flat param group interface
• ~200 LOC reduction

Phase 4: GPU-ify host-side workarounds

Priority order:

1. gpu_clone() — cudaMemcpyDtoDAsync (trivial, HIGH frequency)
2. add/sub/mul() — 3 element-wise CUDA kernels (~10 lines each)
3. broadcast_mul/div() — strided CUDA kernel
4. softmax — standard CUDA softmax kernel for attention
5. narrow(dim>0) — 2D copy kernel
6. argmax/mean_all — CUB reductions (LOW frequency)

Phase 5: Consolidate pointer helpers

• Define raw_ptr()/raw_ptr_mut() once in ml-core, re-export
• Delete 3 duplicate definitions