Description Already Opened PRs
Miles
Fix typo environment variable and unbuffer python outputs #1895 Add a shell launch script test harness for future protection #1896 Fix various launch scripts errors about missing line concatenations or paths #1897 Derive the miles checkout location instead of hardcoding it in launch scripts #1898 Snapshot the external commands of every shell launch script #1899 Read the slurm allocation when the train config is built #1900 Snapshot the commands and generated configs of every python launch script #1901 Cover the public surface of command_utils with unit tests #1902 Rename exec_command by the resource its command needs #1903 Move the shell exec helpers next to their only consumers #1904 Remove non-reproducible file arguments by supporting inline base64 payloads #1905 Snapshot the launchers that build their own command line #1906 Fix p2p profile's rotary_base not reaching the model script it configures #1907 Snapshot test the argv of all model scripts #1908 Expand the model args in python before building the command #1909 Replace the model config shell scripts with python #1910 Quote the model args miles inlines into the launch command #1911 Run the launch script snapshot tests by hand instead of in CI #2279 [~] Add a cli.py for the miles-workbench chart #2195
[~] Compute a served worker's ctor kwargs inside its own actor #2312
remove duplicated field in actor group, unify dashboard hook guard, and change init order #1838 Clear the new-engine flag from the driver #1839 Extract future inference engine hooks on generation and eval #1840 Make the CI fault injection wait asynchronously to avoid long blocking #1841 Split RolloutManager into InferenceController and RolloutExecutor #1842 Wire up InferenceController and RolloutExecutor and make the controller API async #1843 Rename RolloutExecutor.generate to get for future async and multi trainer scenarios #1844 Support saving and loading for rollout functions and add a BaseRolloutFn base class #1845 Fix data source being loaded without saved #1846 Delete the rollout health monitor ahead of its rewrite #1847 Drop the node_rank guards and access only node-0 engines #1848 Extract the router registration into SGLangRouterApiClient #1849 Move mechanically the http-forwarding code out of SGLangEngine into SGLangApiClient #1850 Make SGLangApiClient a real client and SGLangEngine a thin shell over it #1851 Add a general httpx client provider to http_utils #1852 Make SGLangApiClient an async worker handle over httpx #1853 Make the rollout-side offload/onload fan-out async to avoid blocking sync code #1854 Add the sync-to-async bridge for fire-then-rendezvous call sites #1855 Extract the engine url formatting for reusing #1856 Avoid infinite engine teardown blocking orchestration logic #1857 Avoid using SGLangEngine actor to call SGLang APIs and directly call in rollout process #1858 Pass concrete values into SGLangApiClient instead of reading them from args #1859 Pause health monitoring before snapshotting the updatable engines to fix racing condition #1860 Directly call SGLang engines instead of via ray actors in trainers and remove SGLangEngine shells #1861 Synchronize disk-delta baseline reloads across trainer ranks #2500 Register the engines with the router from the rollout process instead of ray actors #1862 Remove dead ServerGroup.onload_weights_from_disk #1863 Introduce ServerCell and store the group's engines as cells #1864 Make ServerGroup callers cell-aware and drop the all_engines property #1865 Move vanilla methods onto ServerCell #1866 Move engine teardown into ServerCell.stop #1867 Extract the Ray actor construction into launch_sglang_ray_actor #1868 Drive the recover offload off the engines that actually started #1869 Switch cell start/stop/recover from engine indices to cell indices #1870 Rename PortCursors to PortAllocator and key it by the node's ip #1871 Add PortAllocator.alloc and thread the shared allocator through allocation #1872 Serve HTTP from the mock engine on its allocated port #1873 Raise for external rollout mode before launching actors #1874 Refactor to remove allocate_rollout_engine_addr_and_ports_normal #1875 Make start_engines async and await engine init inside it #1876 Restructure start_engines into per-cell blocks and assume engines in one cell have same lifecycle #1877 Move per-cell engine startup into ServerCell.start_engines #1878 Sink the startup config from ServerGroup into ServerCell #1879 Address workers by cell index and let the cell describe itself to the router #1880 Move per-cell recovery into ServerCell.recover #1881 Move has_new_engines to RolloutServer #1882 Delete ServerGroup and let RolloutServer own a flat list of cells #1883 Extract the rollout cell state machine into cell_state.py #1884 Drive the cell state machine through whole-cell transitions #1885 Lift the state machine onto ServerCell and delete ServerEngine #1886 Key a server's cells by a globally unique string id instead of numerical indices #1887 Start the rollout engines without blocking the caller #1888 Probe every node of a cell concurrently to avoid serialized initialization #1889 Update the mechanical-refactor-verify skill to the latest sglang version #1890 Store per-token weight version spans per generation call on Sample #1891 Load rollout dumps written before per-call weight version spans #1892 Temporarily publish the weight version when no weight update can carry it #1893 Add assertions to reject training data generated under the sglang default weight version #1894 Give every launch a machine-readable run uuid #1970 Make the structured log prefix a caller-supplied tag #1940 Move load_function into miles.utils.function_registry #1941 Implement a worker RPC server that submits a call and polls for its outcome #1942 Implement a typed RPC client handle #1943 Support sync worker methods on per-group executor threads #1944 Retry a submit only when it provably never reached the server #1945 Long-poll from the client and tolerate a blip mid-poll #1946 Refuse a duplicate call id with 409 instead of running it twice #1947 Pin a client to one server process with a boot uuid #1948 Carry rich argument and result types and refuse ambiguous signatures #1949 Support non-finite floats across the wire #1950 Support worker env vars before importing the server #1951 Expire finished calls that were never retrieved #1952 Add a FIFO ticket lock over a torch.distributed store #2143 Guard the broadcast weight update with a store ticket lock to remove Ray based locks #2144 Rename control_server to api_server for general purpose #1971 Shrink the miles router input from full args to a required-field config #1972 Avoid session server input using full args #1973 Add the config and argv boundaries converter #1974 Add launch_bound_subprocess with binding a child's lifetime to its parent #1975 Launch the sgl and miles routers as command-line subprocesses #1976 Refactor session servers to be launched as command-line subprocesses #1977 Add worker specs describing how each worker kind is launched #1978 Add CommandActor that binds its lifetime to a launched subprocess #1979 Build the dashboard engine topology on the driver instead of the actor #1980 Delete the external-engine path from SGLangEngine #1981 Add the sglang server argv boundary with a mandatory production roundtrip #1982 Render the sglang server and router argv through one renderer #2304 Dissolve RayActor and probe workers through NodeProbeMixin #1983 Bracket ipv6 addressing where it is built instead of in engine init #1984 Launch sglang engines as command subprocesses under CommandActor #1985 Forbid positional arguments when constructing ray actors #1986 Move PortAllocator into miles.utils.workers #1987 Launch the routers and session servers through head-pinned command actors #1988 Add a skeleton RayWorkerManager and a wiring entry point #1989 Move sglang config resolution into sglang_config.py #1990 Cover sglang config parsing with unit tests through resolve_sglang_config #1991 Resolve the sglang config at construction time instead of mid-run #1992 Parse the raw sglang config into a dedicated pydantic model #1993 Convert the resolved sglang config to pydantic models #1994 Move needs_offload and overrides into sglang config resolution #1995 Drop external router and fixed router ip mode #1996 Add worker specs for the sglang and miles routers #1997 Derive session server instance ids from the run uuid and index #1998 Implement a worker spec for the session server #1999 Use SGLANG_DG_CACHE_DIR_PER_PROCESS instead of hand-built cache dirs #2000 Delete the unused get_base_gpu_id function #2001 Stop passing a random seed and let each sglang engine pick its own #2002 Pass node_rank instead of rank when computing the engine launch command #2003 Extract compute_inference_engine_env_vars and forbid compute_server_args defaults #2004 Derive the engine api key from the model config #2005 Launch inference engines through the inference engine worker spec #2006 Rename PortAllocator internals and pass the actor positionally #2048 Start dynamic worker ports above Ray's worker range #2501 Add a worker provider abstraction over RayWorkerManager #2049 Stop checking actor liveness while waiting for a tcp port #2050 Launch the sgl and miles routers through RayWorkerManager #2051 Support pinning workers to the head node in worker manager #2052 Drop support of configurable session server ip and port #2053 Launch session servers through RayWorkerManager #2054 Let the session servers sit on more than one host #2305 Support master-mode ports in RayWorkerManager #2055 Extract the placement group types handed to RayWorkerManager #2056 Support placement groups and gpu ids in RayWorkerManager #2057 Restructure RayWorkerManager around group and cell managers #2058 Support stopping a cell in RayWorkerManager #2059 Keep one addr_info per cell and read the engine topology from RayWorkerManager #2060 Remove temporarily the engine env report #2061 Remove server liveness check in health waiting #2062 Launch inference engines from RayWorkerManager and reduce cells to bookkeeping #2063 Remove rollout engine recovery to be rewritten later #2064 Remove the unnecessary debug_train_only guard when starting cells #2065 Extract the immutable ServerCell metadata to avoid constructing ServerCell later #2066 Refactor start semantics to add semantics for cells #2067 Change stopping cells to dynamic removal #2068 Remove the now-unused stopped cell state #2069 Rename the remaining cell start and stop wording to add and dispose #2070 Fix the redundant model_path check when deciding whether a cell takes weights #2081 Trim ServerCellMetadata down to the fields a cell actually needs #2082 Add cell watching to the worker provider #2083 Drive the rollout server cells from reconcile against the worker provider #2084 Unify the memory release and resume performed when a colocated cell starts #2085 Inline the router registration helpers and hold the router client on the cell #2086 Remove the always-true is_allocated flag from ServerCell #2087 Restructure the rollout cell states while keeping the current behavior #2088 Rename the update weights window entry points to start and end update weights #2089 Mark cell weights ready only once the update weights window closes #2090 Support disposing a rollout cell from any intermediate state #2091 Extract the weight update connection status into ConnStatusManager #2092 Give the fsdp backend its own connection status manager #2093 Detect engine reconnection from a snapshot of the rollout cell hashes #2094 Allocate an out-of-band gate port for every inference engine cell #2095 Launch sglang engines gated and release them once they are running #2096 Add a periodic tick sweep over the rollout cells #2097 Rebuild the rollout cell around an explicit five-state machine #2098 Initialize colocated engines only inside the weight update window #2099 Wait for the expected fleet before the first rollout starts #2100 Add a context lock with a with_lock decorator #2101 Add lock exemption and enforcement checkers to the context lock #2102 Add a requires_lock decorator to ensure lock is being held #2103 Add acquires_lock and releases_lock for split lock windows #2104 Remind periodically while a context lock acquisition is blocked #2105 Reject context lock decorators on classes that forgot enforce_lock_discipline #2106 Support temporarily releasing the context lock #2107 Track the context lock grant by generation instead of identity #2306 Guard the inference controller and its servers with ContextLock #2108 Reuse SimpleHealthCheckerConfig for the rollout health check arguments #2109 Drive health checking from a pulled activeness predicate #2110 Give each rollout cell a health checker driven by its own state #2111 Add start and stop of individual cells to RayWorkerManager #2112 Extract the health condition mapping onto CellCondition #2113 Let the controller report cell status #2114 Serve rollout cells over the api server #2115 Resolve api server cells per request instead of at startup #2116 Open the api server and the healing loop with --use-fault-tolerance #2307 Delete the pre-reconcile engine recovery #2117 Let a rollout cell be crashed on demand #2118 Generalize the ft soak harness beyond trainer cells #2119 Crash rollout engines from the soak harness instead of the controller #2120 Discard stale health probe results instead of cancelling them #2308 Restructure the v1 train group around one keyword dispatch path #2146 Dispatch to trainer actors with keyword arguments only #2147 Fail fast and preserve the cause when no trainer cell can recover #2148 Report lifecycle failures instead of silently continuing without cells #2149 Return one train step output type from every trainer backend and scenario #2150 Return one train result per worker from the cell based train group #2151 Support external data in the cell based train group #2152 Support multi-LoRA adapter reconciliation in trainer v2 #2153 Make the cell based train group the only trainer implementation #2154 Assign the trainer master address from the driver instead of rank zero #2155 Extract the shared actor creation and port allocation in the worker manager #2156 Give worker specs a launch context and serve mode scheduling options #2157 Launch serve mode workers from RayWorkerManager #2158 Add worker specs for the trainer actor and critic #2159 Inline the fault-tolerance concurrency groups onto the trainer actor #2160 Launch trainer actors through RayWorkerManager #2161 Delete the trainer actor factory and the placement group plumbing #2162 Restart trainer cells through RayWorkerManager #2163 Drive trainer cell bookkeeping from the worker provider #2164 Wait for the expected number of trainer cells at startup #2165 Serve trainer cells over the api server like the rollout cells #2166 Key the trainer cells by cell id instead of cell index #2167 Suspend trainer cells through the worker manager like the rollout cells #2168 Remove trainer cells that the provider stops reporting #2169 Address FT test action targets by explicit cell id #2309 Remove the now-unused stopped trainer cell state #2170 Remove the pending state in trainer cells #2171 Merge the now duplicated actor and rollout cell handlers #2172 Kill failed trainer workers through their own RPC channel instead of the manager #2173 Add a ray worker handle with death confirmation #2174 Route all trainer worker calls through the worker handle instead of Ray handles #2175 Point the GLM recipes at the model args module #2310 Stop the fast suite from starting its own ray clusters and leaking accelerators #2311 Publish the new weight version to the rollout executor #2502 Assert that the rollout executor keeps receiving weight versions #2470 Fail the run when the FT api server cannot start #2471 Run the weight equality check before leaving StateInitializing #2472 Bound and wrap the free-port scan #2473 Report a rollout cell stuck initializing as unhealthy #2474 Let workers deregister from the miles router #2475 Temporarily remove vanilla fault tolerance for the eval fleet #2497 Rename the trainer group and cell to TrainerController and TrainerCell #2176 Add a self-provisioned Kubernetes apiserver test environment #2177 Add a kind cluster test environment for Kubernetes testing #2178 Add a level-triggered reconcile loop over a source stream #2179 Synthesize deletions when a relist replaces the store #2180 Reopen a dying source stream and gate startup on the first listing #2181 Retry a failed key with per-key exponential backoff #2182 Re-drive parents on a resync period #2183 Add a Kubernetes LIST/WATCH adapter #2184 Track pods with a reflector that manages its cursor #2185 Document how the reconcile loop maps onto the Kubernetes Go stack #2186 Add a process supervisor for running several ranks in one container #2187 Add --cluster-backend flag to choose between Ray and Kubernetes native backend #2188 Read the train parallel config from the driver to avoid cross-process handle passing #2189 Skip the prometheus collector under the k8s backend #2190 Install kubectl and helm into the training image #2191 Lint every Helm chart from pre-commit and CI #2192 Add the miles-common Helm library chart #2193 Add the miles-workbench Helm chart #2194 Stop printing from the launch command layer #2414 Turn command_utils into a package #2415 Generate every chart's values schema from one set of python types #2416 Let a checkout on shared storage override the code baked into the image #2417 Hand the workbench's own infra values to every run launched from it #2418 Give the namespace one account that may uninstall a run #2419 Give the workbench the actions that talk to a cluster #2420 Ask the cluster what this account may do before installing anything #2421 Check the objects, the API and the namespace, and let the verdict gate the install #2422 Give the workbench a command line #2423 Compute a served worker's ctor kwargs inside its own actor #2424 Let worker handles forward positional arguments #2313 Implement RpcWorkerHandle.wait_dead #2314 Annotate the controllers' and rollout executor's public methods #2315 Resolve router addresses once, before the controller is built #2316 Wait for the session servers from the driver, not the controller #2317 Make get_cell_statuses async on both controllers #2318 Make the inference controller an independent worker #2319 Ship the actor-critic values handoff through the object store #2320 Release the critic's values on every path #2321 Give the rpc layer a codec for an argparse Namespace #2322 Set up the trainer controller's args-derived state in init() #2323 Make the trainer controller an independent worker #2324 Make the rollout executor an independent worker #2325 Make the multi-LoRA async data source async #2326 Make the multi-lora controller an independent worker #2327 Name the cell operations a backend has to provide #2328 Give a worker's constructor the backend it runs on #2329 Hand every worker the providers its spec names #2330 Give each model the router provider scoped to its own pool #2331 Type the pods the reconcile loop reports #2332 Mint the independent-DP quorum store in the trainer controller #2333 Derive a cell's pod packing from the gpus a node has #2334 Answer static addresses from the chart's naming #2335 Read an observed pod into a ParsedPod #2336 Project a cell's pods into workers and addresses #2337 Implement worker provider for dynamic Kubernetes scenarios #2338 Support cell operations in Kubernetes #2339 Treat a pod under deletion as gone from its cell #2340 Assemble the kubernetes backend capability #2341 Ship an object store reference the rpc layer can encode #2342 Run a kubernetes-backed driver against platform-created workers #2343 Read a worker's own identity out of its pod #2344 Build a pod's worker from the run its own argv describes #2345 Skip the values a CLI action cannot render #2425 Make launching a training run a method on a backend object #2426 Read and rewrite train argv through one small helper #2427 Let a launch carry the commands that prepare the machines it runs on #2428 Split the typer wrapper into pieces a second caller can reuse #2429 Let a launch script read its own config out of the environment #2430 Give every launch script the same three subcommands #2431 Have every launch script talk to a backend object #2432 Move the ray-only command helpers under the ray backend #2433 Start the mooncake master when a run asks for that object store #2434 Call a trainer pool a trainer engine #2435 Add the miles-run chart and let it install a run's orchestrator #2436 Install a run's workers, one workload per kind of pool #2437 Compute a pool's chart entry from the worker spec that declares it #2438 Assemble a run's values from its specs and its launch plan #2439 Write a run's verdict where the launcher can read it #2440 Read a rendered manifest as objects rather than as a string #2441 Run the orchestration script under a wrapper that publishes its verdict #2442 Wait for a run's verdict from outside the cluster #2443 Report what the pods of a run are actually doing #2444 Build a run's environment in one place before a second backend needs it #2445 Install a training run into Kubernetes and wait for its verdict #2446 Point a kubernetes run's api server at the orchestrator's own service #2447 Run the mooncake master as part of the run that needs it #2448 Stream every container's logs while the run is going #2449 Announce the cluster events that explain a stuck run #2450 Watch a run while it runs, and say how to look at it afterwards #2451 Collect the scene when a run fails, and let the workbench collect it too #2452 Say which fields of an installed run a relaunch may change #2453 Let a relaunch change a run's size and nothing else #2454 Attach a relaunch to the verdict the installed run is already writing #2455 Remove the releases a previous ci run left behind #2456 Ship a run with the job that will uninstall it #2457 Arm the escape job once the run has reported its verdict #2458 Defuse the previous launch's escape job before installing over it #2459 Pin what a typical run renders, and the commands one launch runs #2460 Render a one-off command as a Job of the run chart #2461 Extract the placeholder substitution the ray path does inline #2462 Run a command in the cluster the same way a run is launched #2463 Compute which engine pod pairs with which trainer pod #2464 Hold a colocated engine pod until something places it #2465 Place each gated engine pod on the node its trainer landed on #2466 Skip a backend's tests when that backend is not there #2467 Let every e2e script pick its backend from the environment #2468 Document the cluster backend for the people who will use it #2469 Let every miles process name itself as the component it is #2503 Split env_report.py into a package #2504 Record the environment as a redacted audit event #2505 Re-record the environment periodically, not only at startup #2506 Hash the uncommitted changes a run actually ran with #2507 Record what the kubernetes launcher launched #2508 Record the environment every sglang engine serves with #2509 Hand the weight-update group the gpus each engine really has #2510 Give a worker provider a lifecycle hook and let it size its own fleet #2511 Support cells without launch gates #2512 Describe an external engine fleet on the command line #2513 Discover a static fleet of external sglang engines #2514
SGLang
Megatron
Reactions are currently unavailable
You can’t perform that action at this time.
Already Opened PRs
Miles
SGLang
Megatron