Roadmap

Current stack

Layer	Complete
Core	Mamba, flow head, CPU Transformer decoder baseline, CPU kernels, BF16 weights, C/C++/Python APIs, cached-VLM SmolVLA action-expert boundary
State	Resumable flow solving, Mamba snapshots, generic job capsules
Relay service	Action and generic-job daemons, EDF, cancellation, administration
Action delivery	Multi-rate chunks, timed replacement, freshness and safety gate
Generic jobs	Contracts, IPC, bounded routing/admission, production Mamba streaming
Observability	Portable action/job traces; fixed-memory action/job metrics

Active development sequence

FlowEdge keeps Diffusion Policy and flow matching as the product path. The Transformer decoder is a fixture for later VLA work. CUDA is the next accelerator, not Tenstorrent-first. Jetson/ARM (#70) still needs a board JSON.

Order	Deliverable	Acceptance evidence
1	Transformer issue #10	Kernel gtests, tiny-gpt2 conversion JSON, optional latency JSON
2	CUDA linking TU	FLOWEDGE_BACKEND=cuda configures only with nvcc; ISA kernels link (#160)
3	Device-resident flow head	Persistent device buffers; no per-op host round-trip (#161)
4	Diffusion Policy on CUDA	Device-resident DDIM; no cudaMalloc after load; vs CPU same checkpoint (#162)
5	Matched GPU replay	GTX 1650 JSON vs PyTorch CUDA (#163); LeRobot CUDA period log (#164)
5b	Device-resident Mamba	Persistent device weights/scratch; no cudaMalloc after load (#172)
5c	CUDA DP conv occupancy	PushT-shape flowedge_cuda_dp_mix before/after; native DDIM median (#174)
5d	CUDA DP split-K conv	L=4 warp split-K; conv 2048 and native DDIM medians (#176)
5e	CUDA DDIM launch census	One native sample shape counts from flowedge_cuda_dp_mix; block-per-group GroupNorm (#179)
5f	CUDA L=8 split-K	L=8 conv and k-major upsample split-K; 256-thread L=4; mix isolate and native DDIM (#182)
6	Jetson / ARM replay	Device JSON for #70; x86 JSON is not ARM evidence
7	Tenstorrent / π0 / ACT	After CUDA evidence; see deferred table

The first DeadlineFlow selector and CPU DDIM bridge are experimental. Full SmolVLA and Tenstorrent execution remain unimplemented. See policy evaluation, DeadlineFlow, CUDA, and the Tenstorrent program.

Deferred expansion

These tracks require a concrete deployment need before taking priority over CUDA flow and Diffusion Policy.

Track	Planned
Heads	ACT, VQ-BeT, π0; Diffusion Policy diffusion_pusht is done on CPU
Backbone	Transformer policy adapter and source SmolVLA encoder
Weight traffic	NUMA replication experiments, INT8
Backends	Tenstorrent TT-Metal
Perception	External observation-encoder integration
Additional backbones	Mamba-3 after CUDA policy evidence
Transport	ROS 2, Zenoh, and distributed scheduling after deployment evidence