feat: Topology aware placement - #2612

Merged

terrykong merged 21 commits into

mainfrom

youngeunk/topology-aware-placement

Jun 23, 2026

Commits on Jun 22, 2026

feat: topology-aware inference placement for non-colocated vLLM

ananthsub
authored and
terrykong
committed
6a3b248
bug fix in topology-aware placement

youngeunkwon0405
authored and
terrykong
committed
fb2b112
fix: make vllm placement group init idempotent

youngeunkwon0405
authored and
terrykong
committed
fc519cd
fix: restore SBATCH singleton dependency in ray.sub
terrykong
committed
a7dafe7
fix(ray.sub): force base-10 for TOPO_RANK fallbacks to avoid invalid JSON
terrykong
committed
1518926
Revert "fix: make vllm placement group init idempotent"
terrykong
committed
ab6841f
test
terrykong
committed
4d770ca
fix: remove unnecessary try/except in _get_gpu_id_info
terrykong
committed
afa325c
feat: add segment_size to ClusterConfig and exemplar configs
terrykong
committed
20e218f
feat: add topology-aware placement to SFT
terrykong
committed
a60c2fa
feat: add topology-aware placement to DPO
terrykong
committed
5b6bb53
feat: add topology-aware placement to distillation
terrykong
committed
d9e17d7
test: add unit tests for topology-aware placement
terrykong
committed
20ed320
refactor: extract prepare_segment_topology to eliminate topology boilerplate
terrykong
committed
b89d01f
feat: add topology-aware placement to RM
terrykong
committed
295043a
fix: use generation_config[backend] to detect vLLM vs SGLang for inference topology
terrykong
committed
d7a11aa
fix: assert bundle count matches world_size after topology sort
terrykong
committed
07a533e
fix: make ClusterConfig.segment_size NotRequired to avoid breaking existing configs
terrykong
committed
c0e484a
ci: trigger DCO re-check with updated base branch
terrykong
committed
2c53db2
refactor: route bundle reordering through topology-aware get_reordered_bundle
terrykong
committed
6874bf5

Commits on Jun 23, 2026

test: temporarily skip SGLang server unit tests (CuTe/CUDA env)
terrykong
committed
4b82af4

Navigation Menu

Uh oh!

feat: Topology aware placement - #2612

feat: Topology aware placement#2612
terrykong merged 21 commits into
mainfrom
youngeunk/topology-aware-placement

Commits on Jun 22, 2026

feat: topology-aware inference placement for non-colocated vLLM

bug fix in topology-aware placement

fix: make vllm placement group init idempotent

fix: restore SBATCH singleton dependency in ray.sub

fix(ray.sub): force base-10 for TOPO_RANK fallbacks to avoid invalid JSON

Revert "fix: make vllm placement group init idempotent"

test

fix: remove unnecessary try/except in _get_gpu_id_info

feat: add segment_size to ClusterConfig and exemplar configs

feat: add topology-aware placement to SFT

feat: add topology-aware placement to DPO

feat: add topology-aware placement to distillation

test: add unit tests for topology-aware placement

refactor: extract prepare_segment_topology to eliminate topology boilerplate

feat: add topology-aware placement to RM

fix: use generation_config[backend] to detect vLLM vs SGLang for inference topology

fix: assert bundle count matches world_size after topology sort

fix: make ClusterConfig.segment_size NotRequired to avoid breaking existing configs

ci: trigger DCO re-check with updated base branch

refactor: route bundle reordering through topology-aware get_reordered_bundle

Commits on Jun 23, 2026

test: temporarily skip SGLang server unit tests (CuTe/CUDA env)