feat: add ROCm ATOM backend and Infera routing - #3

Open

cquil11 wants to merge 48 commits into

mainfrom

agent/atom-atomesh

Commits on Aug 10, 2026

feat: add native static router frontends
cquil11
committed
b8856c1
fix: run setup scripts inside router containers
cquil11
committed
a888bea
fix: advertise vLLM NIXL bootstrap ports
cquil11
committed
fedc844
validate SGLang tensor and data parallelism
cquil11
committed
c328cbc
fix: label static router health checks accurately
cquil11
committed
8605639
fix: align vLLM Router startup timeout
cquil11
committed
1630a9a
fix: skip local SGLang P/D warmup behind router
cquil11
committed
8affae4
docs: cover stable vllm device binding
cquil11
committed
0afc6c2
make router frontend names semantic
cquil11
committed
5d13899
Revert "make router frontend names semantic"
cquil11
committed
d573ba6
preserve native frontend compatibility
cquil11
committed
be02143
test router log capture path
cquil11
committed
ab98030
support router-native vLLM data parallelism
cquil11
committed
cebe610

Commits on Aug 11, 2026

feat(runtime): add AMD accelerator device masking
cquil11
committed
f8b8e36
feat(slurm): support configurable GPU directives
cquil11
committed
c99c0a4
feat(vllm): add vendor-neutral device binding
cquil11
committed
0d61ad7
fix(ci): format GPU directive fallback
cquil11
committed
1a63cb9
Support node-local runtime config transport
cquil11
committed
750dc88
fix: translate node-local log mounts in containers
cquil11
committed
4e2355a
fix: support immutable worker containers
cquil11
committed
96f10cb
feat: separate submit and runtime source paths
cquil11
committed
98a7448
fix: support immutable AMD control plane
cquil11
committed
29b9ddd
fix: prefer private fabric addresses
cquil11
committed
f2e95e4
feat: orchestrate MoRI-IO vLLM workers
cquil11
committed
1d1ce61
fix: embed cluster profile for node-local jobs
cquil11
committed
10d2bc1
fix(router): probe dynamic MoRI readiness end to end
cquil11
committed
6e3e9b9
fix(runtime): preserve default and mock compatibility
cquil11
committed
b0c60c3
style: format transport fallback
cquil11
committed
cfa04c3
fix(setup): skip unused infra for native routers
cquil11
committed
24de885
build: add compute-only setup target
cquil11
committed
0e18f98
fix(identity): detect native router and AMD MoRI
cquil11
committed
9d44a0e
fix(hf): propagate cache environment during prefetch
cquil11
committed
e1ac409
fix(hf): resolve the worker hub cache exactly
cquil11
committed
3b6b13c
fix(router): propagate runtime srun options
cquil11
committed
ab0ba14
test(router): cover runtime srun options
cquil11
committed
f94026b
fix(router): skip Dynamo infra for SGLang Router
cquil11
committed
dd0109d
fix(router): use the native SGLang frontend key
cquil11
committed
31e72da
fix(logs): tolerate non-UTF-8 backend output
cquil11
committed
8bd8aef
Gate benchmarks on advertised vLLM workers
cquil11
committed
b4bf6eb
Merge srt-slurm PR #2 for ATOM integration
cquil11
committed
aca0e48
Merge srt-slurm PR #1 for ATOM integration
cquil11
committed
633e71f
feat: add ATOM backend and Infera frontend
cquil11
committed
9bded70
fix(atom): pin Mooncake to advertised worker IP
cquil11
committed
d93b481
fix(infera): honor worker registry health endpoint
cquil11
committed
297da66
fix(atom): pool Mooncake TCP fallback connections
cquil11
committed
f71cbb1
feat(atom): make Mooncake transport explicit
cquil11
committed
83ef50b
fix(fingerprint): detect ATOM and Infera versions
cquil11
committed
141f035
fix: force Mooncake TCP transport for ATOM
cquil11
committed
5ecfb13