Implement namespace-owned stream rendezvous with incarnation isolation, cancellation, recovery, bounded local byte rings, and host/child actor protocols.
Add a replaceable stream transport boundary plus an Iroh ALPN adapter with framed records, backpressure, reconnect handling, and terminal propagation.
Wire stream sources and sinks through Myelin job deployment and Python bindings, with guarantee coverage across blob, namespace, job, byte-ring, and transport paths.
Sanitize nested Cargo build context in xtask so lint, Nextest, doctest, and maturin invocations retain stable fingerprints.
Deny workspace warnings and lint suppressions, consolidate Rust and Python coverage under cargo xtask test with 60-second per-test limits, and remove stale flaky, stateful, Docker, and orphaned test artifacts.
Refine fleet and provisioning interactions, expand control-plane projections, and add CPU, memory, pressure, and storage telemetry for the local dashboard surfaces.
Add durable namespace control, Iroh-backed source transfer, session-independent publication lifetime, and wire managed Myelin jobs and bindings through the reusable data-plane API.
Replace the eventfd/ring job bootstrap with one inherited arena descriptor, actor-owned sessions, sealed blob leases, awaitable inbox wakeups, and zero-copy Python mappings. Route VastAI mock provisioning through image-backed local Docker workers and preserve pinned child and controller routes across directory updates.
Add TOML job-file submission to the Fleet UI with generic started, running, and completed feedback, reusable remote job controller routing, cancellation and kill invariants, Tinygrad fixture and image support, and comprehensive Rust, Python, CUDA, and lifecycle-ordering coverage.
Define the spawn-time handoff a job process consumes before any
application code runs: two inherited descriptors named by
SWACTOR_ARENA_FD and SWACTOR_WAKE_FD, and a fixed 48-byte little-endian
header at arena offset 0 naming one control-ring region.
- Canonical header layout, shared parser, and a host writer that leases
the header region and control ring disjointly under the arena's own
placement law, zeroes the ring, stamps its generation, and returns the
handoff (env map, inheritable arena and wake descriptors, host wake
eventfd) only after every write completes.
- Python binding gains swactor.run(main): map the arena read-only, take
ground truth from fstat, validate through the shared parser, arm
FD_CLOEXEC on the wake descriptor, close the arena descriptor after
mapping, and drive main on asyncio with Context.data carrying the
resolved state. Fail-fast BootstrapError before main on any defect.
- Rewrite jobs/tiny_linear_inference.py against the approved path API and
record the slice, invariants, and remaining bridge work in the handoff
doc.
Verified with 12 new data-plane bootstrap guarantee tests (parser defect
table, fuzzed pages, writer round-trip), 29 binding tests across
in-process and exec boundaries, and full data-plane and iroh-driver
suites with no regressions.
Architecture enforcement:
- Install a repository-owned rustc wrapper for ordinary cargo check,
build, and test commands. Resolve compiler item identities so renamed
imports and helper wrappers cannot hide spawning, timing, blocking,
polling, thread, or runtime-driving capabilities.
- Define the execution-owner crates and reject dependencies from those
substrates back into Myelin policy. Add compile-pass and compile-fail
contracts for actor helpers, execution owners, test waits, forbidden
capabilities, suppression attempts, and owner dependency inversions.
Execution ownership:
- Add engine-owned actor timers with cancellation and generation identity,
then migrate lifecycle deadlines and protocol ticks off application
tasks. Keep networking, process output, telemetry, and blocking provider
calls in their approved I/O substrates.
- Move process spawn, wait, signal, Unix listener, and output-following
mechanics into swactor-process. Isolate Vast.ai blocking HTTP mechanics
behind its adapter while actors retain retry, recovery, and provisioning
decisions.
Myelin control flow:
- Rework manual control, worker lifecycle, provisioning, provider recovery,
job deployment, distribution, edge orchestration, and shutdown as actor
state transitions and typed effects. Preserve durable provider adoption
and command outcomes across graceful and abrupt restarts.
- Replace controller loops and timer-forwarding tasks with actor messages;
leave substrate tasks as cancellable observation streams with no durable
policy state.
Properties and resource ownership:
- Add deterministic engine and component properties, a stateful mock-VastAI
lifecycle model, persisted regression cases, controlled fault injection,
and a bounded nightly workflow covering restart and teardown behavior.
- Terminate reply observers, cancel telemetry collectors, bound dashboard
projections, and release child observers, file descriptors, process
records, and inode-verified Unix sockets on every terminal path.
Verified with the compiler-policy contracts, 105 Myelin library tests, 32
swactor-process tests, telemetry cancellation contracts, randomized
stateful restart cases, cargo check, and formatting checks.
Add actor-backed manual node provisioning, control-plane endpoints, and fleet UI assets with durable provider lifecycle handling.
Simplify Myelin orchestration, node runtime, staging, and telemetry paths while removing obsolete engine-builder, dashboard-view, and local-mock implementations.
Align runtime delivery, data-plane, distribution, job-runner, process, telemetry, dashboard, Vast.ai integrations, and their tests with the revised actor and transport contracts.
Run Docker workers detached from the launching CLI and persist enough
bootstrap intent to distinguish a live node from stale provider state.
Treat graceful and abrupt orchestrator exits alike: surviving containers
remain authoritative instead of being torn down with the old runtime.
Discover labeled containers on startup, rebuild their local process
records, restore runtime facts, and drive the control-plane rejoin so
existing nodes recover cluster connectivity without duplicate workers.
Keep explicit kill and remove commands authoritative, clean stale
resources on terminal failure paths, and cover adoption, restart, and
provider-exit policy in lifecycle tests.
Subscribe to each worker telemetry pull endpoint during runtime readiness, retry interrupted streams, and remove the obsolete telemetry actor-address path.
Make shutdown state sticky, clean local Docker resources on Ctrl+C, validate fleet controls, and keep dashboard node state honest under failure. Wait for bidirectional iroh routes before job submission and render the orchestrator role as a card header.
Replace the single-purpose chat runtime with a persistent fleet daemon that provisions, adopts, and controls nodes through the dashboard.
Add distributed job-runner actors and provider-backed deployment so jobs can materialize workspaces, execute remotely, and return outputs over iroh.
Duty mixing between iroh-driver and data-plane is resolved: the transport
crate now owns only byte pumping, and the data-plane owns every edge
semantic.
data-plane:
- ids.rs: single EdgeId/RingId/StreamId/NodeId/RunId/LeaseRequestId/
ActorAddress definitions; arena, edge_lifecycle, and ring re-export them
(previously duplicated per module)
- edge_wire.rs: the whole transport contract — WireEvent, EdgeWriter, and
the EdgeTransport port (associated Writer/PeerAddr types)
- edge_runtime.rs: EdgeRuntime composition engine absorbing iroh-driver's
driver_pumps bookkeeping, the EdgeEstablisher lifecycle drive, arena
leasing, ingress stream buffering with object-record parsing, and ring
writes; effects go through a WorkerPort trait; progress surfaces as
structured Observations the application maps to telemetry/agent messages
- delete superseded test-only layers: actor.rs (DataPlaneNodeActor),
edge_actor.rs, ingress.rs, egress.rs and their guarantee tests
- fold ObjectIdAllocator into object_record (now edge-free, starts at 1)
iroh-driver:
- edge_transport speaks pure data_plane::edge_wire vocabulary; EdgeSendHandle
implements EdgeWriter; IrohDriver implements EdgeTransport (PeerAddr =
EndpointAddr) — the entire edge surface is open_writer + drain_events
- delete driver_pumps.rs; new dependency on data-plane (no cycle)
- IROH_DRIVER_SPEC §5 updated for the new module set and edge boundary
myelin:
- WorkerEdgeRuntime shrinks from ~830 lines of hand glue to an EdgeRuntime
holder plus a tinygrad WorkerPort impl and observation reporting; the
driver-event/edge-event translation layers and newtype re-wrapping are
gone
- orchestration/app.rs and job edge drains consume WireEvent
Tests: data-plane 31 (5 new EdgeRuntime contract tests), iroh-driver 13,
myelin 65 — all green.
Replace fleet/actor-overview/actor-dossier/workers views and the root link
list with one ControlPlaneView serving / and /view/fleet: node cards fuse
machine stats with actor rollup, node click opens machine detail + roster,
actor click opens an in-page dossier (identity, message diet, sparkline,
sampled message history via /api/view/fleet/detail).
- Message history folds view-side from messages_processed deltas: 16-receipt
ring, 250ms spacing, sampled-out counters — noisy actors cannot flood the
page and producers stay untouched.
- Stale streams render in a separate collapsed pool; superseded life
generations are evicted immediately; stale pool hard-capped at 50.
- Unified navbar injected server-side from the view registry; pages opt in
with a <!--swactor:nav--> placeholder so app-registered views appear
automatically. Hardcoded per-page navs stripped.
- Rust type names are the actor display name; address is the unique key.
Stale doc comment claiming types are not on the frame fixed.
The crate is the per-node metrics/logging pipe with a universal
subscriber endpoint, but "datastream" kept getting misread as a general
messaging plane. Rename crate, module paths, and public API
(`DatastreamEndpoint` → `TelemetryEndpoint`, etc.) so misuse is visible
on sight.
Renamed contracts (all in-repo producers/consumers migrated):
- env vars `MYELIN_DATASTREAM_*` → `MYELIN_TELEMETRY_*`
- artifact `datastream.ndjson` → `telemetry.ndjson`
- actor names `telemetry-publisher` / `telemetry-sink`
- wire ALPN `swactor/telemetry/0`
- `DATASTREAM_SPEC.md` → `TELEMETRY_SPEC.md`
Also fixes two latent test breaks: `process` and `iroh-driver` tests
imported `DatastreamEvent` from the crate root, which was never
re-exported; they now use the observer path `telemetry::frame::`.
The datastream is metrics/logging only; control decisions must never branch
on a frame. This was a recurring cultural problem with no structural
enforcement. This change makes it a compile-time and CI-enforced fact.
datastream crate (lib.rs):
- Stop re-exporting Frame, DatastreamEvent, FrameDelivery at crate root.
is now a compile error (E0425). These types live
only in datastream::frame::* and are documented as the observer surface.
- Safe identity types (ChannelId, StreamId, Position, Record, etc.) remain
re-exported at root for producer-side callers.
orchestration/app.rs:
- Extracted all frame-touching code (CollectedDatastreamFrame,
drain_datastream_connections, update_load_progress_from_frame,
drain_frames, archive_collected_frame, pump, OrchDatastream,
DashboardSupport) into two new observability modules:
frame_collector.rs and orch_datastream.rs.
- The orchestrator now interacts through a FrameCollector whose
drain/drain_with_progress methods take closures; it never names Frame,
DatastreamEvent, or CollectedDatastreamFrame.
- StageLoadProgress (the one control-relevant signal previously scraped
from frame payloads) is extracted inside FrameCollector and handed to
the control loop as plain data.
xtask:
- New check-telemetry-isolation command scans control-plane modules
(orchestration/, distribution/, data-plane/, provisioning/) for
forbidden frame-type references and fails the build if any are found.
Verified: workspace builds (myelin + dashboard feature), datastream 29
tests pass, myelin 64 lib tests pass, check-telemetry-isolation passes
clean.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Introduce a pure, level-triggered reconciler in `crates/provisioning`
that drives a declared cluster shape toward convergence over the
existing node lifecycle, replacing the edge-triggered imperative node
orchestration in `apps/myelin`.
- `reconcile`/`reconcile_node`/`observe`: pure decider and observation
folder with stable logical-node identity, per-attempt operation
identity, and deterministic retry backoff; `ClusterDriver` is the sole
writer of observed state, coalescing triggers, recording operations as
pending before dispatch, and scheduling timed requeues.
- `IdempotentEffectExecutor`: deduplicates submissions by
`(run_id, logical_node_id, attempt)` and runs provider work on the
engine-hosted blocking substrate, never blocking a reconcile pass.
- Myelin integration: `MyelinEffectBackend` bridges `ProvisionPlugin` to
the executor contract; `LocalProcessPlugin`/`LocalDockerPlugin`
provider adapters; `ProvisionedClusterGuard` pumps triggers,
observations, and due operations.
- Retire the imperative acquire/bootstrap/teardown sequencing across
`apps/myelin` orchestration, staging, observability, and provider
adapters in favor of the declarative driver.
- Move the reconciler specification to `docs/specs/archive`.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Introduce the swactor engine: a swactor-owned composite that retains a
selected execution substrate, drives the core runtime, and hosts the
async/blocking/timer work that backs actors. Integrations receive one
cloneable EngineHandle and never construct or borrow a raw Tokio
runtime/handle.
Engine crate (crates/engine):
- The contract: spawn / spawn_blocking / timer / interval / now, a
per-implementation capability model with construction-time binding
(require()), and engine-owned time. The engine owns all progression;
actor handlers stay synchronous and never .await.
- TokioBackend owns the Tokio runtime and schedules core ticks and
supporting futures on it; SteppingBackend is a single-threaded
deterministic scheduler with virtual time (the non-Tokio portability
proof). Core is driven through its existing tick() surface; a
self-rescheduling CoreDriver is installed at construction and is the
sole place permitted to call try_tick.
iroh-driver:
- Receives an EngineHandle instead of a raw Tokio Handle. Accepts,
reads, dials, writes, endpoint construction, and teardown schedule
through it; required capabilities (tasks/timers/io) are validated
before the endpoint binds. Engine-hosted interval pumps drive
actor-bridge, datastream, and edge ingress.
myelin:
- One node/orchestrator engine owns core, protocol tick injection, and
transport progression; the application loop only drains
integration-owned queues. Stage-shard process readers, delayed actor
messages, helper stdout/stderr, prompt RPC, and CPU sampling all
schedule through the engine (spawn_blocking / engine tasks / timers).
- Removed the split-engine APIs: install_actor_bridge_pump(period) and
spawn_protocol_ticker(period) use each component's stored engine;
deleted the no-op pump_network callback and its plumbing; deleted the
dashboard raw-Tokio/standalone-runtime conveniences.
Enforcement:
- A clippy disallowed-methods boundary forbids direct runtime/scheduling/
time/core-driving bypasses, denied in swactor-engine, iroh-driver, and
myelin. Retained excluded uses (VastAI provider, provider process
supervision/log capture, OS-signal/stdin/process-control sequencing)
carry narrow allowances with reasons.
Verification:
- Engine contract + unit tests (incl. the SteppingBackend portability
proof), iroh integration tests (capability rejection before binding,
multi-node actor behavior), and a production execution-composition
smoke test that observes engine-driven actor progress with no ambient
Tokio runtime and no manual tick/pump. Workspace all-target/all-feature
clippy and tests are green.
Specs co-located with their crates: ENGINE_SPEC.md in crates/engine,
IROH_DRIVER_SPEC.md in crates/iroh-driver. VastAI remains explicitly out
of scope pending its separate redesign.
Specs gain a monotonic Id and relocate by their true status: drafts
(WIP/aspirational) to docs/specs/drafts/, and accurate code-behavior
references stay in their crate dirs. docs/specs/archive/ is reserved for
superseded docs (currently empty).
Dispositions were cross-referenced against code, not the specs' own headers.
IROH_DRIVER claimed "current-state" but ~30% is unbuilt redesign, so it moves
to drafts. DATA_PLANE_ACTOR's central integration claim is unrealized (myelin
bypasses its node actor), and ACTOR_PANEL is not a reference; both are dropped
rather than reviewed or archived. DATASTREAM and MANAGED_PROCESS stay as
references; MYELIN stays in place (stale, flagged for review).
From here, commit titles reference a spec by [N] when one applies. This
bootstrap commit does not carry one.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Core no longer owns or drives OS threads. The runtime is now a single
tick-driven worker whose loop an external engine hosts and advances.
This is the cutover required before the engine seam is introduced.
Removed from core:
- Runtime::run() and its owned thread pool (spawn, park/unpark, join)
- notify_worker() and worker_threads: Vec<OnceLock<Thread>> plumbing
- Placement load-aware selector and its WorkerStats-driven next_worker()
- WorkerId newtype and the address->worker routing map; AddressMap is
now a plain AddrSet membership set
- num_threads from RuntimeConfig
Rewired for the single-worker tick API:
- python/wasm bindings, dashboard dummy node (deleted), myelin vastai
adapter, and the runtime/test suites
Cleanup folded in during review:
- prune three never-written WorkerStats counters (cross_sends,
messages_dropped, restarts)
- collapse the redundant tick_all params onto the WorkerContext handle
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Add project licensing and a full README, plus a minimal WebAssembly ping-pong example that exercises the actor runtime.
- `LICENSE`: add the full GNU AGPL-3.0 text and set `license = "AGPL-3.0-only"` on every package (`swactor`, each crate, `apps/myelin`, `tools/vastai`, `xtask`)
- `README.md`: rewrite from a stub into a full project overview, covering features (actor_id routing, WASM, iroh QUIC/SWIM, OTP-style std, process manager, zero-copy objects, datastream metrics, dashboard), architecture, examples, developing, status, and license
- `examples/ping-pong`: new standalone workspace (`pingpong` cdylib) where two actors volley on a single-threaded `wasm` runtime driven by a Node host via `tick()`, demonstrating spawning, message passing, and death monitoring (`watching`)
- `examples/ping-pong`: add a host/run harness -- `run.sh` (wasm-pack build + `run.mjs`), `serve.sh` (dashboard + static demo), `index.html`, and a pinned `Cargo.lock`
- `.gitignore`: stop ignoring `.loop/`, `.deployment-notes/`, and `.omp/`
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Add read-only actor overview and per-actor dossier pages to the dashboard, fed by enriched per-actor runtime snapshots.
- `dashboard/swactor/actor_view`: new `ActorPanelView`, a tolerant frame consumer over `runtime.actors`/`runtime.stats` that folds per-actor snapshots and serves `/view/swactor/actor-overview` (roster) and `/view/swactor/actor-dossier` (per-actor detail), each backed by an embedded HTML template (`actor_overview.html`, `actor_dossier.html`)
- `dashboard`: register both views in `DashboardHandle` and export `actor_overview_view()`/`actor_dossier_view()` from the swactor module
- `swactor` core: enrich `ActorSnapshot` with `actor_type` and `message_type` (populated from `slot.actor.metadata()` in `ActorPool`) and add `ActorAddress::to_full_hex()` for untruncated display
- `myelin/orchestration`: publish actor stats to the dashboard via a `runtime.actors` channel producer (`stats_hook_on`) threaded through the distribution stack, and carry the orchestrator actor address into readiness signaling
- workspace `Cargo.toml`: add `default-members` for native iteration and a centralized `[workspace.dependencies] tokio` so members share one feature set; `dashboard/Cargo.toml` switches to `tokio.workspace = true`
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Promote the `mvp-system` workspace library crate to a standalone application at `apps/myelin`, rebranding the MVP system along with its binaries, node image, and spec.
- workspace `Cargo.toml`: swap member `crates/mvp-system` -> `apps/myelin` and drop `apps` from `exclude` so the app joins the workspace
- `apps/myelin/Cargo.toml`: declare package `myelin` with `autobins = false` and explicit `[[bin]]` targets `myelin-worker`, `myelin-orchestrator`, `myelin-chat`
- `apps/myelin/src`: move the whole `mvp-system` source tree and rebrand module surfaces (`chat/mod.rs`, `prompt/mod.rs`); add `bin/chat.rs` (`myelin::run_chat_from_args`) and delete the old `mvp_chat.rs`
- `apps/myelin/node-image`: relocate the worker image assets from `apps/mvp-node/` (Dockerfile, Dockerfile.base, tinygrad_worker.py, entrypoint, e2e script) and rename `MVP_SYSTEM_SPEC.md` -> `MYELIN_SPEC.md`
- `xtask`: rewrite build/reference paths for the rename (~1000-line churn); add `crates/dashboard/ACTOR_PANEL_SPEC.md`
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Land end-to-end canonical benchmark observability and a synthetic datastream-connectivity preflight across the orchestrator, chat, worker-node, and Python tinygrad worker, plus an xtask validator, so a 4-stage 7B pipeline run is fully diagnosable.
- benchmark_observability: expand stamp() with canonical producer fields (producer_component/instance_id/process_id/sequence, wall_clock_unix_ms, monotonic_ms, clock_source) and schema_version so every event shares one envelope shape
- orchestrator_app + bin/{mvp_chat,worker_node}: stamp OrchBootstrap/OrchPromptEvent/ChatProgress/NodeEvent/SamplerHealth with the canonical fields plus span_id/parent_span_id, and add a 4-phase synthetic datastream preflight (ProducerConfigured/Connected/SyntheticEventSent/Observed) plus an endpoint_config_snapshot event on each process
- apps/mvp-node/tinygrad_worker: add apply_canonical_envelope()/datastream_endpoint_snapshot() and emit_python_datastream_preflight() mirroring the Rust preflight, and enrich benchmark_stamp() with the same producer fields
- bin/worker_node: pass MVP_DATASTREAM_ENDPOINT_ID/MVP_BENCHMARK_PRODUCER_INSTANCE/MVP_IROH_ENDPOINT_ADDR_MASK/MVP_IROH_RELAY_MODE env to the spawned tinygrad worker so its stamps identify the stage
- bin/mvp_chat: add --pipeline-parallel as an alias for --pipeline-stages (with a duplicate-guard) and bump recursion_limit
- xtask: add a benchmark-observability validator (ValidatorFinding/BenchmarkValidation, validate_benchmark_observability, canonical-stamp and stage/edge checks, evidence + gap-report builders) with tests for missing python datastream connectivity, wrong run_id, and missing span_id
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Complete an 8-stage pipeline-parallel run over VastAI by provisioning stages high-to-low, adding per-stage/per-step metrics, host anti-colocation, and provider state-timeout guardrails.
- orchestrator_app: select the next weight-load stage by max index (provision stages high-to-low for parallel spread), add a throttled "loaded N of M; waiting on stage X" stage_provision_wait headline, and surface min_compute_cap/state_timeout_secs in the config dump.
- orchestrator_app: enrich pipeline_token_in/out and tokenizer_decode events with token_count/token_ids/generated_index.
- worker_node: add timing metrics across the data path (helper_execute_ms, egress_ring_read_ms, send_ms, ingress_ring_write_ms, object_load_ms), refactor take_complete_ingress_record into IngressRecordBytes (object_id/sequence/extent/flags), and emit a new object_loaded event.
- vastai_provisioning: track leased host_ids and blacklist already-leased hosts in later ProvisionRequests so stages don't co-locate, and tag SSH-bootstrap retry logs with the attempt number.
- tools/vastai: add min_compute_cap (PP_MIN_COMPUTE_CAP) filter/search query and a LifecyclePolicy state_timeout (PP_STATE_TIMEOUT_SECS) that fails instances stuck in a non-running status instead of polling forever.
- xtask: raise the check timeout to 1800s/30s grace, drop --skip-rebuild for VastAI, aggregate per-stage StepExecuted metrics, add a vastai summary section, and write failure artifacts on abort.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Turn mvp-chat-check into a benchmarking harness with per-stage latency capture, a run envelope, summary artifacts, and a mvp-chat-compare command for delta analysis.
- xtask: add mvp-chat-compare <baseline> <candidate> that checks comparability (schema/scenario/workload/model/provider/pipeline_stages) and prints deltas for total, prepare, standup-to-RPC, and per-prompt roundtrip/first-token/decode/text-decode ms.
- xtask: build_benchmark_summary now writes stdout/stderr/prompts/redacted-config/summary artifacts with per-artifact byte counts and a vastai summary section; rename the dump log to datastream.ndjson.
- xtask: add write_failure_artifacts so failed checks still emit a failure summary with the prompt-corpus blake3 and artifact sizes.
- mvp_chat: add the mvp.chat.benchmark channel and emit_benchmark_envelope (BenchmarkRunEnvelope with model/runtime/provider/workload detail), and tag prompt events with prompt_index and a blake3 prompt_hash.
- tinygrad_worker: add per-phase latency metrics (encode/decode/text-decode/first-token elapsed_ms; stage_execution_ms/record_write_ms on execute_step; ring_readable/encode_prompt/decode_tokens elapsed_ms plus payload sizes).
- vastai_provisioning: emit VastAiLeaseReady and VastAiSshEndpointReady provider lines (contract/offer/host/gpu/dph, ssh host/port/user) for observability.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Land the first working two-stage pipeline-parallel run over VastAI, wiring a real inter-stage data path with observability, a max-price offer cap, and remote-image reuse.
- orchestrator_app: raise the VastAI pipeline-stage cap from 1 to 2 and let VastAI pipeline planning resolve the HuggingFace GGUF from the default local cached-model metadata path instead of requiring host mounts; add --vastai-max-dph-total (CLI/env/TOML) config.
- vastai_provisioning: make complete_bootstrap a no-op so the SSH bootstrap log tail stays alive past runtime-ready until node stop, preserving post-ready worker logs; add a test asserting the tail is only stopped on NodeStop.
- worker_node: emit data-path NodeEvents across the pipeline (iroh_edge_stream_arrived/bytes_read/bytes_sent, egress_ring_read, ingress_ring_write) with edge/byte metadata.
- tools/vastai: add max_dph_total (PP_MAX_DPH_TOTAL) to SelectionPolicy, the reachable-offer filter, and the search query, and improve the empty-pool error message.
- xtask: pass --skip-rebuild for the VastAI scenario and gate it on a new require_vastai_data_path_facts plus GPU facts (ring install, activation object load/step, interstage handoff, iroh edge read/sent).
- mvp_chat: add ChatModelConfig (model id/gguf/tokenizer/max-context) forwarded to the orchestrator; for VastAI + skip-rebuild, emit skip events and reuse the remote node image without a local build.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Get the multinode pipeline-parallel prompt loop actually running locally by supporting both tinygrad LLM backends and switching the Docker scenario to direct addressing.
- tinygrad_worker.py (load_pipeline_stage_model): try the modern tinygrad.llm gguf/model modules first and fall back to the legacy tinygrad.apps.llm TransformerBlock on ModuleNotFoundError, with PipelineStageTinygradModel constructing blocks positionally when no TransformerConfig exists
- tinygrad_worker.py (load_weights): drop the Transformer.from_gguf whole-model branch and its TinygradAppsLlmPartialStageUnsupported fatal, so partial pipeline stages build through load_pipeline_stage_model on either backend
- xtask (MultinodeDocker): stop forcing --relay-mode default --endpoint-addr-mask relay-only, so the scenario runs over the Docker network with full/direct addresses
- xtask (dump-log facts): rename relay_masked_* facts to docker_*, assert multiple workers join the coordinator via direct addresses (direct_addr_count > 0), and relax the benchmark report to skip the ensure_worker_binary span for the Docker scenario
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
- Pull ~7.4k lines out of the orchestrator bin into a new orchestrator_app library
module.
- Wire a local single-node GPU prompt loop into the mvp_chat bin; touch
gpu_worker_ingress_parser.
- Grow xtask and the mvp-node tinygrad worker.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Add end-to-end timing instrumentation and an xtask benchmark report for mvp-chat runs.
- benchmark_observability: add a shared stamping module — stamp(component) emitting schema/pid/monotonic+wall ms from a process-global start and sequence counter, plus unix_ms_now() — stamped onto every mvp-chat/orchestrator/worker-node event and frame-archive record
- mvp-chat: thread a run_id (new --run-id, defaults to 1) through config and the orchestrator CLI, add per-phase started/ready/failed emits for ensure_orch_binary/ensure_worker_binary/prepare_node_image, and a prompt_complete record carrying tokens_generated/elapsed_ms/final_text bytes
- orchestrator/worker-node: stamp bootstrap and prompt events, add arrival_unix_ms to archived frames, propagate MVP_RUN_ID/MVP_LOGICAL_NODE_ID/MVP_STAGE_INDEX into the tinygrad worker, default the device to CPU for the process provider, emit a prompt_rpc started span, and drop the MVP_TINYGRAD_TEST_MODE passthrough
- tinygrad_worker.py: stamp every control() event and tag it with run/node/stage env, add a CPU:X86 fallback when clang is absent, and remove the test_mode() short-circuits
- xtask: replace the flat dump-log fact assertions with a benchmark report builder (build_benchmark_report) that requires named spans (prepare_runtime, ensure_*_binary, weights_loaded, prompt_rpc) and emits per-prompt first-token/decode/tokens-per-second latency; wrap the cargo run in XtaskBenchmark synthetic frames and pass a unix-ms --run-id
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Replace stdout-parsed runtime-ready detection with an explicit, plugin-driven bootstrap-completion step and actorize SSH bootstrap teardown.
- provisioning: drop the `PluginObservation::RuntimeReady` variant and add `ProvisionPlugin::complete_bootstrap`, an explicit per-node completion hook (no-op for `LocalDockerPlugin`)
- bootstrap_datastream: remove `parse_runtime_ready`/`RuntimeReadyLine` so bootstrap no longer infers readiness from a parsed stdout JSON line
- vastai_provisioning: drop the `ReadyTrackingSink` ready-flag wrapper; the SSH retry loop now runs purely `while !stopping`, and `complete_bootstrap` stops the node's bootstrap with `BootstrapStopReason::RuntimeReady`
- vastai_provisioning: actorize teardown as `SshBootstrapActor` on the swactor `Runtime` (handle holds an `ActorAddress`), with `stop_bootstrap(handle, reason)` delivering a `Stop` message; add `BootstrapStopReason::{RuntimeReady,NodeStop}`
- actors/provisioner: replace the `RuntimeReady` observation arm with a `ProvisionerMsg::RuntimeReady` handler that calls `complete_bootstrap` then `mark_live`/emits NodeLive (or NodeFailed on error)
- callers/tests: wire the new explicit ready flow through node_agent, the orchestrator/worker_node binaries, and `mvp_one_node_chat`; add the `ssh_bootstrap_actor_stop_kills_child` test
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Stand up an interactive end-to-end chat over a CUDA GPU, provisioning a Dockerized node that loads a GGUF model and serves prompts over TCP.
- prompt_rpc: add the newline-JSON prompt protocol (`SubmitPrompt` + `PromptEvent::{TextDelta,Done,Fault}`) carried over TCP
- mvp_chat: add an interactive REPL client connecting to the prompt RPC port (default 127.0.0.1:19777)
- mvp_orch_one_node / mvp_one_node_chat: add the single-node orchestrator that provisions a `LocalDockerPlugin` node, loads `bartowski/Llama-3.2-1B-Instruct-GGUF` (Q4_K_M), and exposes the prompt RPC listener with boot/route/weight timeouts
- mvp_node: add the GPU worker binary that spawns `tinygrad_worker.py` (default device CUDA) and ships runtime telemetry via a `ClusterFrameSink`
- vastai_provisioning / bootstrap_datastream: add the vast.ai provider adapter (`VastAiProvisioningConfig`, `VastAiLeaseClient`) wrapping `swactor_vastai`, plus a bridge that folds provision stdout onto a per-node datastream
- apps/mvp-node: add CUDA base/runtime Dockerfiles (nvidia/cuda 12.6.3, tinygrad 0.12.0, sshd), `mvp_entrypoint.sh` (sshd + mvp-node, held for postmortem), `local_docker_e2e.sh`, the GGUF tinygrad worker, and one-node-chat/bootstrap/vastai guarantee tests
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Fold the fragmented distribution swim/routing/gossip tests into swim_core, routing,
and swim_actor. Prune the dashboard tui and command surfaces. Add mvp-system actors
(node_agent, orchestrator, stage_controller), the local_e2e harness, and gpu worker
e2e.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Move iroh_driver and the relay binary out of distribution into a dedicated
crates/iroh-driver (lib re-exports IrohDriver; relay bin renamed). Remove the node crate
and the single-gpu-inference example; drop the docker/datastream demo. Slim
pipeline-parallel vastai.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Remove the datastore crate, the top-level design/orchestration/ring specs, the
benches, and the ci config. Add the dashboard host telemetry sampler
(cpu/disk/net/gpu/mem). Localize the pipeline-parallel e2e stub/mock paths.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
SwimTelemetry observer records probe RTTs (p50), recent targets, and membership
transitions for the node's telemetry tick. Expand registry/registry_actor, datastream
catalog/emit/source, and the pipeline-parallel cluster/fleet. Wire the node main loop
to emit host/runtime/transport/membership frames.
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
Promote pipeline-parallel-inference to a first-class app and consolidate observability on the datastream wire, decoupling the dashboard crate from `distribution`.
- apps/pipeline-parallel-inference: move the example out of `examples/` into `apps/` as its own workspace, rename binaries to `pp-worker`/`pp-orchestrator`, and strip release binaries
- cluster: add `ClusterNode`, a synchronous facade over the actorized distribution protocol (IrohDriver + per-node Runtime hosting Swim/Registry/Metadata/Directory actors with a `MembershipFanout`), replacing ad-hoc `driver.node()`/`tick()` call sites
- fleet: add per-node fleet telemetry that ships identity/resource records as `DatastreamFrame`s over the cluster transport to the orchestrator's `DatastreamSink`, folded into a `FleetView` on a 3s tick
- provision: add best-effort, opt-in SSH boot-phase telemetry (`PP_DEPLOY_KEY`) that streams rented-node boot logs onto the orchestrator's datastream as `proc.boot.<stage>.*`
- dashboard: rewire the crate dependency from `distribution` to `datastream`, drop the standalone `swactor-datastream-dashboard` binary, and rewrite `datastream_source.rs` to demux per-node frames into Overview/Distribution/Fleet views with live-node TTL filtering
- distribution: refresh dist/netmap plugin copy and README from "Kademlia routing" to gossip-directory terminology
Signed-off-by: Zachery Aaron Shores-Chmielewski <zacheryasc@gmail.com>
