Skip to content

Navigation Menu

Appearance settings

View all features
- BY COMPANY SIZE
  Enterprises
  Small and medium teams
  Startups
  Nonprofits
- BY USE CASE
  App Modernization
  DevSecOps
  DevOps
  CI/CD
  View all use cases
- BY INDUSTRY
  Healthcare
  Financial services
  Manufacturing
  Government
  View all industries
View all solutions
- EXPLORE BY TOPIC
  AI
  Software Development
  DevOps
  Security
  View all topics
- EXPLORE BY TYPE
  Customer stories
  Events & webinars
  Ebooks & reports
  Business insights
  GitHub Skills
- SUPPORT & SERVICES
  Documentation
  Customer support
  Community forum
  Trust center
  Partners
View all resources
- COMMUNITY
  GitHub SponsorsFund open source developers
- PROGRAMS
  Security Lab
  Maintainer Community
  GitHub Stars
  Archive Program
- REPOSITORIES
  Topics
  Trending
  Collections
- ENTERPRISE SOLUTIONS
  Enterprise platformAI-powered developer platform
- AVAILABLE ADD-ONS
  GitHub Advanced SecurityEnterprise-grade security features
  Copilot for BusinessEnterprise-grade AI features
  Premium SupportEnterprise-grade 24/7 support
Pricing

Appearance settings

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

radixark / miles Public

Notifications You must be signed in to change notification settings
Fork 507
Star 3k

Code
Issues 136
Pull requests 1.1k
Actions
Projects
Security and quality
Insights

Additional navigation options

Code
Issues
Pull requests
Actions
Projects
Security and quality
Insights

[DO NOT MERGE] CI 1 - #2407

#2407

Closed

fzyzcjy wants to merge 236 commits into

tom/refactor-miles/deliver-1-baseradixark/miles:tom/refactor-miles/deliver-1-basefrom

tom/refactor-miles/deliver-1radixark/miles:tom/refactor-miles/deliver-1Copy head branch name to clipboard

Conversation Commits236 (236)Checks Files changed

Closed

[DO NOT MERGE] CI 1#2407
fzyzcjy wants to merge 236 commits into
tom/refactor-miles/deliver-1-basefrom
tom/refactor-miles/deliver-1

Commits

Commits on Sep 4, 2026

Temporarily revert PR 1313
fzyzcjy
committed
3d52698
remove duplicated field in actor group, unify dashboard hook guard, and change init order
fzyzcjy
committed
75a480d
Clear the new-engine flag from the driver
fzyzcjy
committed
a65ef5b
Extract future inference engine hooks on generation and eval
fzyzcjy
committed
6c07540
Make the CI fault injection wait asynchronously to avoid long blocking
fzyzcjy
committed
f95a6f1
Split RolloutManager into InferenceController and RolloutExecutor
fzyzcjy
committed
90a12ad
Wire up InferenceController and RolloutExecutor and make the controller API async
fzyzcjy
committed
5cd7de2
Rename RolloutExecutor.generate to get for future async and multi trainer scenarios
fzyzcjy
committed
551e8df
Support saving and loading for rollout functions and add a BaseRolloutFn base class
fzyzcjy
committed
046eb0c
Fix data source being loaded without saved
fzyzcjy
committed
4fc592f
Delete the rollout health monitor ahead of its rewrite
fzyzcjy
committed
22a6038
Drop the node_rank guards and access only node-0 engines
fzyzcjy
committed
e05149f
Extract the router registration into SGLangRouterApiClient
fzyzcjy
committed
2a50d88
Move mechanically the http-forwarding code out of SGLangEngine into SGLangApiClient
fzyzcjy
committed
1d7afc7
Make SGLangApiClient a real client and SGLangEngine a thin shell over it
fzyzcjy
committed
9e75635
Add a general httpx client provider to http_utils
fzyzcjy
committed
235b6e1
Make SGLangApiClient an async worker handle over httpx
fzyzcjy
committed
1a713ec
Make the rollout-side offload/onload fan-out async to avoid blocking sync code
fzyzcjy
committed
909ff49
Add the sync-to-async bridge for fire-then-rendezvous call sites
fzyzcjy
committed
3d1e239
Extract the engine url formatting for reusing
fzyzcjy
committed
f46b30b
Avoid infinite engine teardown blocking orchestration logic
fzyzcjy
committed
cae19ce
Avoid using SGLangEngine actor to call SGLang APIs and directly call in rollout process
fzyzcjy
committed
7ed5483
Pass concrete values into SGLangApiClient instead of reading them from args
fzyzcjy
committed
39dc26a
Pause health monitoring before snapshotting the updatable engines to fix racing condition
fzyzcjy
committed
b6cc393
Directly call SGLang engines instead of via ray actors in trainers and remove SGLangEngine shells
fzyzcjy
committed
36f2ed5
Synchronize disk-delta baseline reloads across trainer ranks
fzyzcjy
committed
d4e4a92
Register the engines with the router from the rollout process instead of ray actors
fzyzcjy
committed
b1ca809
Remove dead ServerGroup.onload_weights_from_disk
fzyzcjy
committed
4776a64
Introduce ServerCell and store the group's engines as cells
fzyzcjy
committed
e39e881
Make ServerGroup callers cell-aware and drop the all_engines property
fzyzcjy
committed
5b14ae1
Move vanilla methods onto ServerCell
fzyzcjy
committed
bee0972
Move engine teardown into ServerCell.stop
fzyzcjy
committed
bb374ce
Extract the Ray actor construction into launch_sglang_ray_actor
fzyzcjy
committed
df174aa
Drive the recover offload off the engines that actually started
fzyzcjy
committed
62e036e
Switch cell start/stop/recover from engine indices to cell indices
fzyzcjy
committed
c9ce019
Rename PortCursors to PortAllocator and key it by the node's ip
fzyzcjy
committed
d384a5c
Add PortAllocator.alloc and thread the shared allocator through allocation
fzyzcjy
committed
5c6100a
Serve HTTP from the mock engine on its allocated port
fzyzcjy
committed
69ea36f
Make the mock sglang engine reachable from other nodes
fzyzcjy
committed
a459dad
Raise for external rollout mode before launching actors
fzyzcjy
committed
5e412e7
Refactor to remove allocate_rollout_engine_addr_and_ports_normal
fzyzcjy
committed
4713b2b
Make start_engines async and await engine init inside it
fzyzcjy
committed
a8a7e8c
Restructure start_engines into per-cell blocks and assume engines in one cell have same lifecycle
fzyzcjy
committed
797a785
Move per-cell engine startup into ServerCell.start_engines
fzyzcjy
committed
e600e9f
Sink the startup config from ServerGroup into ServerCell
fzyzcjy
committed
e4a2b47
Address workers by cell index and let the cell describe itself to the router
fzyzcjy
committed
4c9097a
Move per-cell recovery into ServerCell.recover
fzyzcjy
committed
df0ee14
Move has_new_engines to RolloutServer
fzyzcjy
committed
b3fce01
Delete ServerGroup and let RolloutServer own a flat list of cells
fzyzcjy
committed
347eeb1
Extract the rollout cell state machine into cell_state.py
fzyzcjy
committed
885630c
Drive the cell state machine through whole-cell transitions
fzyzcjy
committed
4e71f18
Lift the state machine onto ServerCell and delete ServerEngine
fzyzcjy
committed
216b104
Key a server's cells by a globally unique string id instead of numerical indices
fzyzcjy
committed
b9376cd
Start the rollout engines without blocking the caller
fzyzcjy
committed
beb5f3d
Probe every node of a cell concurrently to avoid serialized initialization
fzyzcjy
committed
fb6877f
Store per-token weight version spans per generation call on Sample
fzyzcjy
committed
d6c0fba
Load rollout dumps written before per-call weight version spans
fzyzcjy
committed
430f2d6
Temporarily publish the weight version when no weight update can carry it
fzyzcjy
committed
f17555a
Add assertions to reject training data generated under the sglang default weight version
fzyzcjy
committed
62083ab
Give every launch a machine-readable run uuid
fzyzcjy
committed
deb258f
Make the structured log prefix a caller-supplied tag
fzyzcjy
committed
36ad186
Move load_function into miles.utils.function_registry
fzyzcjy
committed
aef846e
Implement a worker RPC server that submits a call and polls for its outcome
fzyzcjy
committed
378967d
Implement a typed RPC client handle
fzyzcjy
committed
cfc4de7
Support sync worker methods on per-group executor threads
fzyzcjy
committed
802b463
Retry a submit only when it provably never reached the server
fzyzcjy
committed
f602fc9
Long-poll from the client and tolerate a blip mid-poll
fzyzcjy
committed
4a39b0c
Refuse a duplicate call id with 409 instead of running it twice
fzyzcjy
committed
3105dd1
Pin a client to one server process with a boot uuid
fzyzcjy
committed
1b86716
Carry rich argument and result types and refuse ambiguous signatures
fzyzcjy
committed
ed98362
Let rpc worker handles forward positional arguments
fzyzcjy
committed
3dd7363
Support non-finite floats across the wire
fzyzcjy
committed
bdaec33
Support worker env vars before importing the server
fzyzcjy
committed
81654ce
Expire finished calls that were never retrieved
fzyzcjy
committed
18e3003
Add a FIFO ticket lock over a torch.distributed store
fzyzcjy
committed
2572e3b
Guard the broadcast weight update with a store ticket lock to remove Ray based locks
fzyzcjy
committed
44312dc
Rename control_server to api_server for general purpose
fzyzcjy
committed
b836a60
Shrink the miles router input from full args to a required-field config
fzyzcjy
committed
8a44942
Avoid session server input using full args
fzyzcjy
committed
1604c60
Add the config and argv boundaries converter
fzyzcjy
committed
940da16
Add launch_bound_subprocess with binding a child's lifetime to its parent
fzyzcjy
committed
1d92162
Launch the sgl and miles routers as command-line subprocesses
fzyzcjy
committed
7044d5f
Refactor session servers to be launched as command-line subprocesses
fzyzcjy
committed
adb7dd5
Add worker specs describing how each worker kind is launched
fzyzcjy
committed
21dd014
Add CommandActor that binds its lifetime to a launched subprocess
fzyzcjy
committed
6b6fd42
Build the dashboard engine topology on the driver instead of the actor
fzyzcjy
committed
7f9e3c3
Delete the external-engine path from SGLangEngine
fzyzcjy
committed
3a913f3
Add the sglang server argv boundary with a mandatory production roundtrip
fzyzcjy
committed
42f740e
Render the sglang server and router argv through one renderer
fzyzcjy
committed
9c21952
Dissolve RayActor and probe workers through NodeProbeMixin
fzyzcjy
committed
f960cf2
Bracket ipv6 addressing where it is built instead of in engine init
fzyzcjy
committed
4c38f72
Launch sglang engines as command subprocesses under CommandActor
fzyzcjy
committed
bfc6257
Forbid positional arguments when constructing ray actors
fzyzcjy
committed
a97504a
Move PortAllocator into miles.utils.workers
fzyzcjy
committed
c2bd753
Launch the routers and session servers through head-pinned command actors
fzyzcjy
committed
dff5d34
Add a skeleton RayWorkerManager and a wiring entry point
fzyzcjy
committed
1e1e064
Move sglang config resolution into sglang_config.py
fzyzcjy
committed
cc661ce
Cover sglang config parsing with unit tests through resolve_sglang_config
fzyzcjy
committed
2f8df47
Resolve the sglang config at construction time instead of mid-run
fzyzcjy
committed
85424c8
Parse the raw sglang config into a dedicated pydantic model
fzyzcjy
committed
80c4bd2
Convert the resolved sglang config to pydantic models
fzyzcjy
committed
209c4d5
Move needs_offload and overrides into sglang config resolution
fzyzcjy
committed
f38ab4a
Drop external router and fixed router ip mode
fzyzcjy
committed
992ee63
Add worker specs for the sglang and miles routers
fzyzcjy
committed
de86c8c
Derive session server instance ids from the run uuid and index
fzyzcjy
committed
3fad3db
Implement a worker spec for the session server
fzyzcjy
committed
a2c1f77
Use SGLANG_DG_CACHE_DIR_PER_PROCESS instead of hand-built cache dirs
fzyzcjy
committed
287ad2e
Delete the unused get_base_gpu_id function
fzyzcjy
committed
ba72c02
Stop passing a random seed and let each sglang engine pick its own
fzyzcjy
committed
2fd7604
Pass node_rank instead of rank when computing the engine launch command
fzyzcjy
committed
679c8f3
Extract compute_inference_engine_env_vars and forbid compute_server_args defaults
fzyzcjy
committed
26115cc
Derive the engine api key from the model config
fzyzcjy
committed
051c5a6
Launch inference engines through the inference engine worker spec
fzyzcjy
committed
4839e22
Rename PortAllocator internals and pass the actor positionally
fzyzcjy
committed
86e3ac3
Start dynamic worker ports above Ray's worker range
fzyzcjy
committed
b48a70e
Add a worker provider abstraction over RayWorkerManager
fzyzcjy
committed
62745c2
Stop checking actor liveness while waiting for a tcp port
fzyzcjy
committed
f245f6c
Launch the sgl and miles routers through RayWorkerManager
fzyzcjy
committed
cf545a5
Refuse a static port that something else is already serving
fzyzcjy
committed
0ca275d
Support pinning workers to the head node in worker manager
fzyzcjy
committed
5bf0fad
Drop support of configurable session server ip and port
fzyzcjy
committed
f07570a
Launch session servers through RayWorkerManager
fzyzcjy
committed
b58dc5b
Let the session servers sit on more than one host
fzyzcjy
committed
ed3f220
Support master-mode ports in RayWorkerManager
fzyzcjy
committed
8dd4ad0
Extract the placement group types handed to RayWorkerManager
fzyzcjy
committed
e6bea04
Support placement groups and gpu ids in RayWorkerManager
fzyzcjy
committed
0fa44bd
Restructure RayWorkerManager around group and cell managers
fzyzcjy
committed
61540c1
Support stopping a cell in RayWorkerManager
fzyzcjy
committed
c7ec553
Keep one addr_info per cell and read the engine topology from RayWorkerManager
fzyzcjy
committed
511df56
Remove temporarily the engine env report
fzyzcjy
committed
6e56a1e
Remove server liveness check in health waiting
fzyzcjy
committed
62bb793
Launch inference engines from RayWorkerManager and reduce cells to bookkeeping
fzyzcjy
committed
914bb7c
Remove rollout engine recovery to be rewritten later
fzyzcjy
committed
ee3f704
Remove the unnecessary debug_train_only guard when starting cells
fzyzcjy
committed
0539da3
Extract the immutable ServerCell metadata to avoid constructing ServerCell later
fzyzcjy
committed
12de814
Refactor start semantics to add semantics for cells
fzyzcjy
committed
0a127b3
Change stopping cells to dynamic removal
fzyzcjy
committed
761210e
Remove the now-unused stopped cell state
fzyzcjy
committed
1a67308
Rename the remaining cell start and stop wording to add and dispose
fzyzcjy
committed
114e19d
Fix the redundant model_path check when deciding whether a cell takes weights
fzyzcjy
committed
e0c7018
Trim ServerCellMetadata down to the fields a cell actually needs
fzyzcjy
committed
2dd15e4
Add cell watching to the worker provider
fzyzcjy
committed
9e8d8ba
Drive the rollout server cells from reconcile against the worker provider
fzyzcjy
committed
3644164
Unify the memory release and resume performed when a colocated cell starts
fzyzcjy
committed
d6f3d9d
Inline the router registration helpers and hold the router client on the cell
fzyzcjy
committed
a57e45c
Remove the always-true is_allocated flag from ServerCell
fzyzcjy
committed
6e9db1f
Restructure the rollout cell states while keeping the current behavior
fzyzcjy
committed
1099917
Rename the update weights window entry points to start and end update weights
fzyzcjy
committed
aa51984
Mark cell weights ready only once the update weights window closes
fzyzcjy
committed
753cc68
Support disposing a rollout cell from any intermediate state
fzyzcjy
committed
bb79cb3
Extract the weight update connection status into ConnStatusManager
fzyzcjy
committed
46c6374
Give the fsdp backend its own connection status manager
fzyzcjy
committed
c4a611e
Detect engine reconnection from a snapshot of the rollout cell hashes
fzyzcjy
committed
0164e62
Allocate an out-of-band gate port for every inference engine cell
fzyzcjy
committed
8e93580
Launch sglang engines gated and release them once they are running
fzyzcjy
committed
51a5e8b
Refuse to override the allocator-owned launch gate port
fzyzcjy
committed
60140ee
Add a periodic tick sweep over the rollout cells
fzyzcjy
committed
14ec98c
Rebuild the rollout cell around an explicit five-state machine
fzyzcjy
committed
f50253e
Initialize colocated engines only inside the weight update window
fzyzcjy
committed
34a7e6b
Wait for the expected fleet before the first rollout starts
fzyzcjy
committed
6355f75
Add a context lock with a with_lock decorator
fzyzcjy
committed
01d4e96
Add lock exemption and enforcement checkers to the context lock
fzyzcjy
committed
1d974a0
Add a requires_lock decorator to ensure lock is being held
fzyzcjy
committed
d9dfb8c
Add acquires_lock and releases_lock for split lock windows
fzyzcjy
committed
0bff974
Remind periodically while a context lock acquisition is blocked
fzyzcjy
committed
8e82040
Reject context lock decorators on classes that forgot enforce_lock_discipline
fzyzcjy
committed
f79029a
Support temporarily releasing the context lock
fzyzcjy
committed
28742b3
Track the context lock grant by generation instead of identity
fzyzcjy
committed
411b3d5
Guard the inference controller and its servers with ContextLock
fzyzcjy
committed
eda8868
Reuse SimpleHealthCheckerConfig for the rollout health check arguments
fzyzcjy
committed
d9c4751
Drive health checking from a pulled activeness predicate
fzyzcjy
committed
5410cf9
Give each rollout cell a health checker driven by its own state
fzyzcjy
committed
4408c2c
Drop a rollout cell whose initialization failed
fzyzcjy
committed
09b3bb4
Keep reconciling the other cells when one cell's reconcile fails
fzyzcjy
committed
6fe5ce4
Add start and stop of individual cells to RayWorkerManager
fzyzcjy
committed
cf4b2b0
Extract the health condition mapping onto CellCondition
fzyzcjy
committed
206f5f7
Let the controller report cell status
fzyzcjy
committed
ff1788c
Serve rollout cells over the api server
fzyzcjy
committed
d7c9356
Publish a worker's endpoints only once it has all of them
fzyzcjy
committed
e054751
Describe a worker that is still being given its ports
fzyzcjy
committed
0f9f065
Resolve api server cells per request instead of at startup
fzyzcjy
committed
c276c4d
Open the api server and the healing loop with --use-fault-tolerance
fzyzcjy
committed
8265523
Delete the pre-reconcile engine recovery
fzyzcjy
committed
f2f4564
Let a rollout cell be crashed on demand
fzyzcjy
committed
10ef720
Generalize the ft soak harness beyond trainer cells
fzyzcjy
committed
eb54643
Crash rollout engines from the soak harness instead of the controller
fzyzcjy
committed
8333fae
Discard stale health probe results instead of cancelling them
fzyzcjy
committed
8bd2002
Restructure the v1 train group around one keyword dispatch path
fzyzcjy
committed
ae156fe
Dispatch to trainer actors with keyword arguments only
fzyzcjy
committed
781f694
Fail fast and preserve the cause when no trainer cell can recover
fzyzcjy
committed
0b2b0af
Reject indep_dp when shared actor-critic PPO is enabled
fzyzcjy
committed
a082d16
Report lifecycle failures instead of silently continuing without cells
fzyzcjy
committed
1ad8f14
Return one train step output type from every trainer backend and scenario
fzyzcjy
committed
dfdd509
Return one train result per worker from the cell based train group
fzyzcjy
committed
bf72efe
Support external data in the cell based train group
fzyzcjy
committed
fee4116
Support multi-LoRA adapter reconciliation in trainer v2
fzyzcjy
committed
fefebee
Make the cell based train group the only trainer implementation
fzyzcjy
committed
43f2d22
Assign the trainer master address from the driver instead of rank zero
fzyzcjy
committed
1743254
Extract the shared actor creation and port allocation in the worker manager
fzyzcjy
committed
110d2f0
Give worker specs a launch context and serve mode scheduling options
fzyzcjy
committed
f742015
Launch serve mode workers from RayWorkerManager
fzyzcjy
committed
1428f3b
Add worker specs for the trainer actor and critic
fzyzcjy
committed
b406fea
Resolve GPU ids inside the worker
fzyzcjy
committed
14e1d64
Inline the fault-tolerance concurrency groups onto the trainer actor
fzyzcjy
committed
aef4506
Launch trainer actors through RayWorkerManager
fzyzcjy
committed
92f69af
Delete the trainer actor factory and the placement group plumbing
fzyzcjy
committed
1fdb9e4
Restart trainer cells through RayWorkerManager
fzyzcjy
committed
9a183d7
Drive trainer cell bookkeeping from the worker provider
fzyzcjy
committed
cac4ab2
Wait for the expected number of trainer cells at startup
fzyzcjy
committed
c108d00
Serve trainer cells over the api server like the rollout cells
fzyzcjy
committed
aa1844f
Key the trainer cells by cell id instead of cell index
fzyzcjy
committed
03f4772
Suspend trainer cells through the worker manager like the rollout cells
fzyzcjy
committed
257207d
Remove trainer cells that the provider stops reporting
fzyzcjy
committed
e8bd353
Address FT test action targets by explicit cell id
fzyzcjy
committed
58fc963
Remove the now-unused stopped trainer cell state
fzyzcjy
committed
f3b104e
Remove the pending state in trainer cells
fzyzcjy
committed
14a03fb
Merge the now duplicated actor and rollout cell handlers
fzyzcjy
committed
7292fc4
Kill failed trainer workers through their own RPC channel instead of the manager
fzyzcjy
committed
108c0ba
Add a ray worker handle with death confirmation
fzyzcjy
committed
dcdea92
Route all trainer worker calls through the worker handle instead of Ray handles
fzyzcjy
committed
a842c51
Stop the fast suite from starting its own ray clusters and leaking accelerators
fzyzcjy
committed
2a28be2
Publish the new weight version to the rollout executor
fzyzcjy
committed
b021029
Assert that the rollout executor keeps receiving weight versions
fzyzcjy
committed
75ec37a
Fail the run when the FT api server cannot start
fzyzcjy
committed
d57132b
Run the weight equality check before leaving StateInitializing
fzyzcjy
committed
9e259ad
Bound and wrap the free-port scan
fzyzcjy
committed
b4ba880
Report a rollout cell stuck initializing as unhealthy
fzyzcjy
committed
95176ea
Let workers deregister from the miles router
fzyzcjy
committed
3bb2a4b
Temporarily remove vanilla fault tolerance for the eval fleet
fzyzcjy
committed
622ba18
Make the train actor threaded only under fault tolerance
fzyzcjy
committed
c5cf37c
Serialize cell suspension and fault injection against weight updates
fzyzcjy
committed
cca4504
Rename the trainer group and cell to TrainerController and TrainerCell
fzyzcjy
committed
1d9d92b
Decide implicit healing after the config file overrides the args
fzyzcjy
committed
1e9eecf
Reacquire the context lock safely when a released section is cancelled
fzyzcjy
committed
4d0b751
Make the FT api server bind host a flag defaulting to loopback
fzyzcjy
committed
cc2682b
Await dynamic port probes instead of blocking the worker manager loop
fzyzcjy
committed
2c1590c

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Community
Docs
Contact

You can’t perform that action at this time.