skip to content
[readonly]

51 packages match "node"

  • total 51
  • drwxr-xr-x-../
  • drwxr-xr-x-Oct 27 02:57@ajax/mesh-worker-node/# 0.1.0 mesh-rt's node loop on CPU threads: claim published tickets, run init or task, complete them with results, marks, node records and traces
  • drwxr-xr-x-Oct 27 01:45@ajax/mesh-abi-node/# 0.1.0 Node records (N_*): a node's SM, heartbeat, tasks and busy time
  • drwxr-xr-x-Oct 27 01:45@ajax/mesh-abi-node-load/# 0.1.0 Node load records (NL_*): how much of the GPU the nodes got, measured from inside
  • drwxr-xr-x-Oct 27 02:57@ajax/mesh-occupancy/# 0.1.0 Co-residency of resident CUDA kernels: whether every program's nodes fit on the GPU at once, and how many more do
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobs-node/# 0.1.0 A meshd's job node, sans IO: its replica of the fleet's job table, the leader's dispatcher (leases, rates, stealing, retries, deadlines, exactly-once results, payload copies) and the member side
  • drwxr-xr-x-Oct 27 04:12@ajax/mesh-top-ui-mesh/# 0.1.0 mesh-top's Mesh tab: the cluster overview (a row per system, even one known only through its peers) and a panel per machine with its GPU, nodes, peers, program chips, sparkline, network, load and node grid
  • drwxr-xr-x-Oct 27 04:27@ajax/mesh-lease/# 0.1.0 The lease: mesh-rt's node loop played by a host thread against the mailbox, for executors without resident kernels (wasm lanes)
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-load-share/# 0.1.0 The GPU share maths of the load sampler: node load record deltas, share, slices, gaps, and nearest-rank quantiles
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-loop/# 0.1.0 The node loop of every mesh program's kernel: tickets from the slot's ring, completion, heartbeat, load records, shutdown
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-linalg/# 0.1.0 Exact-mode GEMM tile and GEMV at mesh node scope: sequential sums, unfused, bit-identical on every GPU
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-load-prog/# 0.1.0 One program's load sample: GPU share from its node load records, task latencies from its trace, busy time, throughput, its CPU executor's cost
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-blocks/# 0.1.0 blocks, the mesh program that exercises every block collective of a multi-warp node, as PTX from C
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-block/# 0.1.0 Block collectives of a multi-warp mesh node in a fixed, bit-reproducible order (reduce, scan, broadcast), and typed shared memory
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobs-out/# 0.1.0 The job node's outputs (sends, blob pushes and fetches, persists, runs, cuts, answers, logs) and the rule that every persist is synced before the next send
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-sim/# 0.1.0 mesh-rt's device primitives on the CPU, and a launcher that runs a kernel's nodes as threads, for testing GPU programs without a GPU
  • drwxr-xr-x-Oct 27 02:57@ajax/mesh-worker/# 0.1.0 The SDK of native worker programs: mesh_worker_run(init, task) maps meshd's mailbox, says ready, and runs a slot's nodes at each run until they halt
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobd-text/# 0.1.0 The jobs control answers: `jobs`, `job status <id>` and `job results <id> [from] [n]`, from a member's job node
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-sha256d/# 0.1.0 sha256d, Bitcoin-style double SHA-256 nonce grinding as a mesh GPU job program (multi-warp nodes), as PTX from C
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt/# 0.1.0 The mesh device runtime: MESH_PROGRAM(init, task) makes a resident GPU kernel (mesh_main) on the node loop, or a CPU-simulated one
  • drwxr-xr-x-Oct 27 01:59@ajax/mesh-ref-block/# 0.1.0 The block collectives of a multi-warp node on the CPU: warp butterflies and scans, block reduce and scan, step for step as the GPU
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobd/# 0.1.0 meshd's jobs thread and blob thread: the job node over the peer links, the driver and the job log; the blob store's thread and its writers; the jobs view and the host block's job fields
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-matmul/# 0.1.0 matmul, the mesh's exact-mode GPU GEMM and GEMV (bit-identical on every GPU), as PTX from C
  • drwxr-xr-x-Oct 27 02:57@ajax/mesh-inproc/# 0.1.0 A GPU program driven in-process: compile, load, install a program-table slot in a mailbox of its own, launch, publish, stop and relaunch, through meshd's backend
  • drwxr-xr-x-Oct 27 04:09@ajax/mesh-sched-score/# 0.1.0 The fleet scheduler's feasibility filter and score: why a program cannot run on a member, or the weighted terms of running it there
  • drwxr-xr-x-Oct 27 01:59@ajax/mesh-ref-gpt2-norm/# 0.1.0 GPT-2's layer normalization of a row: 256-thread strided sums, then the block tree, in Z
  • drwxr-xr-x-Oct 27 04:27@ajax/mesh-exec-wasm-testkit/# 0.1.0 Runs wasm guest programs through meshd's wasm executor on a test mailbox: load, launch, publish tickets, read views, unload (for tests)
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-pipesrc-n/# 0.1.0 pipesrc-n, the pipe demo's source as a native mesh worker: msgs into pipehash or digests straight to a sink, with backpressure
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-hello/# 0.1.0 hello, the smallest mesh GPU program (task = arg * 3 + 7), as PTX from C
  • drwxr-xr-x-Oct 27 02:57@ajax/mesh-mps/# 0.1.0 meshd's CUDA backend: one context that blinks, or --isolation mps with a private MPS daemon and a pool of worker processes; nvJitLink as its compiler
  • drwxr-xr-x-Oct 27 05:23@ajax/mesh-snapshot/# 0.1.0 Snapshot v3: the mailbox header, one section per loaded program, and the host block, in one frame for viewers
  • drwxr-xr-x-Oct 27 02:56@ajax/nvjitlink/# 0.1.0 nvJitLink loaded at runtime: PTX to a cubin for the GPU's arch, off the blink, with no driver JIT fallback
  • drwxr-xr-x-Oct 27 01:59@ajax/mesh-ref-gpt2-shape/# 0.1.0 GPT-2 124M's shape: layers, heads, widths, vocabulary, context, and the flat weight buffer's layout in checkpoint order
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-pipesink/# 0.1.0 pipesink, the pipe demo's checking sink as a native mesh worker: drains digests, verifies each sha256d, histograms latency
  • drwxr-xr-x-Oct 27 04:27@ajax/mesh-exec-wasm/# 0.1.0 meshd's wasm executor (BACKEND_WASM): CPU programs as WebAssembly modules on lease lanes, behind the backend interface
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-pathtrace/# 0.1.0 pathtrace, a deterministic path tracer as a mesh GPU job program (bit-identical images on every GPU and CPU), as PTX from C
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-prog-gemm/# 0.1.0 gemm, f32 matrix multiply throughput as a native mesh worker: a cache-blocked multi-threaded GEMM, timed tasks and job items
  • drwxr-xr-x-Oct 27 03:27@ajax/mesh-worker-ctx/# 0.1.1 What a native worker's init and task see: mesh-rt's Ctx on a CPU thread (views with fault recording, peer views, the input queue, pipes, the request window, counters)
  • drwxr-xr-x-Oct 27 02:56@ajax/mesh-cuda/# 0.1.0 meshd's CUDA: a primary or green context, eager module loading, the pinned mailbox, cubin modules, streams and resident-kernel launches
  • drwxr-xr-x-Oct 27 04:08@ajax/mesh-sched-needs/# 0.1.0 What a program needs and prefers from the fleet scheduler (backends, arch, terms, replicas, groups, reservations) and its one-token wire form
  • drwxr-xr-x-Oct 27 02:56@ajax/cuda-driver-fake/# 0.1.0 A simulated CUDA device behind the cuda-driver table: modules, globals, streams and launches whose kernels run as C functions on CPU threads, for testing without a GPU
  • drwxr-xr-x-Oct 27 01:59@ajax/mesh-ref-gpt2-attend/# 0.1.0 GPT-2's attention for one head and query: scaled dots, block-tree max and sum, softmax, the 4-group value sum, in Z
  • drwxr-xr-x-Oct 27 05:23@ajax/mesh-driver/# 0.1.0 meshd's driver thread: the only ring publisher and program-table writer; lifecycle, blinks, publishing, pipes and job items
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-ptx/# 0.1.0 mesh-rt's device primitives on NVIDIA GPUs as inline PTX: special registers, sys-scope ordering, atomics, warp shuffles, the block barrier
  • drwxr-xr-x-Oct 27 04:27@ajax/mesh-rt-wasm/# 0.1.0 The guest runtime of wasm CPU mesh programs: Ctx, View, Exchange and MESH_PROGRAM over the mesh::* imports (wasm32)
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobs-rate/# 0.1.0 A member's measured rate for a job program (an EWMA over windows) with slow start for its lease size
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-ctx/# 0.1.0 What a mesh program sees (Ctx): views, peer views, input queue, pipes, request window, exchange page, counters, ids
  • drwxr-xr-x-Oct 27 03:26@ajax/mesh-rt-sim-mailbox/# 0.1.0 A host mailbox with one mesh program slot and the host side of its ticket ring, request window and input queue, for testing programs on the CPU simulator
  • drwxr-xr-x-Oct 27 05:24@ajax/mesh-jobs-sched/# 0.1.0 The jobs' lease selection over the fleet scheduler's model: members, each program's needs, measured performance, weights (F4's filter and score)
  • drwxr-xr-x-Oct 27 01:45@ajax/mesh-abi-base/# 0.1.0 Mailbox-wide constants of the mesh contract: magic, version, runtime ABI, warp and shared-memory sizes, slots, systems, names
  • drwxr-xr-x-Oct 27 04:09@ajax/mesh-sched-place/# 0.1.0 The fleet scheduler's placement pass: every program placed where it fits and scores best, in a fixed order, with an explain per program

-- more --