Repository navigation

Add cross-platform Docker and hardware CI - #2

Closed

haoruilee wants to merge 24 commits into

mainfrom

feat/cicd

Commits on Jun 26, 2026
- ci: add cross-platform docker and hardware workflows
  haoruilee
  committed
  3d7f2ac
Commits on Jun 27, 2026
- fix(ci): address 6 review findings in cross-platform CI
  
  haoruilee
  and
  claude
  committed
  07954dd
- fix(ci): resolve all blocking and quality gaps from review
  
  haoruilee
  and
  claude
  committed
  4d13535
- fix(ci): drop cache-to from build-pr job (fork PR cache writes fail)
  
  haoruilee
  and
  claude
  committed
  2085ccd
Commits on Jun 28, 2026
- fix(ci): add strict gpu preflight safeguards
  haoruilee
  committed
  3ef3ba9
- fix(ci): handle runpod create failures
  haoruilee
  committed
  2610c62
- fix(ci): use runpod ssh key for pods
  haoruilee
  committed
  f3c5230
- fix(ci): create runpod workspace before clone
  haoruilee
  committed
  5c6f18d
- fix(ci): harden runpod pip installs
  haoruilee
  committed
  8924cd3
- fix(ci): explicitly build cuda extension on runpod
  haoruilee
  committed
  a21a928
- fix(ci): make flashinfer install optional for gpu core tests
  haoruilee
  committed
  7a03545
- fix(ci): wait for runpod ssh readiness
  haoruilee
  committed
  daf021b
- fix(cuda): correct sm90 tma barrier sequencing
  haoruilee
  committed
  337585e
- fix(ci): retry transient runpod ssh disconnects
  haoruilee
  committed
  1524d2e
- fix(cuda): pass sm90 tensor maps as raw addresses
  haoruilee
  committed
  55a2ac2
- fix(cuda): use grid-constant addresses for sm90 tensor maps
  haoruilee
  committed
  43c6b0f
- fix(cuda): store sm90 tensor maps in device memory
  haoruilee
  committed
  df29724
- fix(cuda): use cuda12 tma tile modifier
  haoruilee
  committed
  e9ceeb8
- fix(cuda): preserve shared address state for sm90 tma
  haoruilee
  committed
  53f30d6
- fix(cuda): use cuda barriers for sm90 tma
  haoruilee
  committed
  3ebf596
- fix(cuda): force global state for sm90 tensor maps
  haoruilee
  committed
  78a023a
- fix(cuda): compile sm90 tma with cluster shared state
  haoruilee
  committed
  263e116
- fix(cuda): fall back from sm90 logp for fp32 logits
  haoruilee
  committed
  734ea0f
- fix(cuda): expose generic logp api from sm90 wrapper
  haoruilee
  committed
  d270e99

Navigation Menu

Repository navigation

Add cross-platform Docker and hardware CI - #2

Add cross-platform Docker and hardware CI#2
haoruilee wants to merge 24 commits into
mainfrom
feat/cicd

Commits on Jun 26, 2026

ci: add cross-platform docker and hardware workflows

Commits on Jun 27, 2026

fix(ci): address 6 review findings in cross-platform CI

fix(ci): resolve all blocking and quality gaps from review

fix(ci): drop cache-to from build-pr job (fork PR cache writes fail)

Commits on Jun 28, 2026

fix(ci): add strict gpu preflight safeguards

fix(ci): handle runpod create failures

fix(ci): use runpod ssh key for pods

fix(ci): create runpod workspace before clone

fix(ci): harden runpod pip installs

fix(ci): explicitly build cuda extension on runpod

fix(ci): make flashinfer install optional for gpu core tests

fix(ci): wait for runpod ssh readiness

fix(cuda): correct sm90 tma barrier sequencing

fix(ci): retry transient runpod ssh disconnects

fix(cuda): pass sm90 tensor maps as raw addresses

fix(cuda): use grid-constant addresses for sm90 tensor maps

fix(cuda): store sm90 tensor maps in device memory

fix(cuda): use cuda12 tma tile modifier

fix(cuda): preserve shared address state for sm90 tma

fix(cuda): use cuda barriers for sm90 tma

fix(cuda): force global state for sm90 tensor maps

fix(cuda): compile sm90 tma with cluster shared state

fix(cuda): fall back from sm90 logp for fp32 logits

fix(cuda): expose generic logp api from sm90 wrapper