[CICD] Unify dependency-coupled CI across accelerator platforms - #1285

Merged

xmhubj merged 95 commits into

flagos-ai:mainfrom

BrianPei:feat/dependency-coupling-all-platforms

Sep 15, 2026

Commits on Sep 8, 2026

feat: add Megatron-LM-FL and TE-FL dependency coupling for all platforms
AlexMa616
committed
c068f26
fix(cicd): add missing test_matrix.unit.nproc_per_node for all platforms
AlexMa616
committed
b382b50
fix(cicd): improve Megatron-LM-FL installation verification logic
AlexMa616
committed
f4a1820
fix(cicd): remove extra newlines at end of config files
AlexMa616
committed
4941a74
fix(cicd): add Python version workaround and improve verification logic
AlexMa616
committed
a5a173e
fix(cicd): improve Megatron installation verification with detailed logging
AlexMa616
committed
5f5e820
fix(cicd): verify megatron.core package installation correctly
AlexMa616
committed
30614c4
fix: integrate dependency caches into unit tests and reorganize workflows
AlexMa616
committed
09d6838
fix: align cache paths between prepare and test workflows
AlexMa616
committed
5b8c81e
fix: use TE_FL_WHEEL_DIR instead of TE_FL_INSTALL_DIR
AlexMa616
committed
a17653b
fix(ci): resolve MUSA/Kunlunxin/Enflame platform test failures
AlexMa616
committed
851d695
fix(musa): patch TE-FL flash_attn detection for torch_musa compatibility
AlexMa616
committed
53e9322
fix(metax): remove pre-installed Megatron paths to use CI-built latest version
AlexMa616
committed
a2546de
fix(musa/enflame): add NVTE env vars to disable flash_attn at build time
AlexMa616
committed
5ee5ba8
fix(all-platforms): uninstall pre-installed deps from images to prevent conflicts
AlexMa616
committed
f4b61d9
feat(all-platforms): add megatron-energon dependency for multimodal support
AlexMa616
committed
5036adb
fix(enflame): also uninstall pre-installed megatron-core from image
AlexMa616
committed
679bc97
feat(all-platforms): unify dependency management across all platforms
AlexMa616
committed
2e933ef
fix(metax): convert pip_packages to block-style YAML array to escape brackets
AlexMa616
committed
0fe8dff
fix(cuda,ascend,hygon): convert pip_packages to block-style YAML array
AlexMa616
committed
4df43b9
fix(ci): prevent dependency version conflicts across all platforms
AlexMa616
committed
fb8dd3a
style: fix ruff linting errors in compatibility_patches.py
AlexMa616
committed
d1c13ec
fix(tests): apply compatibility patches before pytest imports
AlexMa616
committed
c0895b0
style: fix import order in tests/conftest.py
AlexMa616
committed
c395bff
fix(tests): ensure megatron-lm-fl-install has absolute PYTHONPATH priority
AlexMa616
committed
a6379c7
fix(tests): apply same PYTHONPATH priority to functional tests
AlexMa616
committed
a0ac07c
fix(ci): unify prepared training runtime dependencies
AlexMa616
committed
b3971b0
fix(ci): isolate vendor startup output from env exports
AlexMa616
committed
cf6fc36
fix(cuda): disable unsupported NCCL EP build in CI
AlexMa616
committed
2cdebbb
fix(ci): prevent inspect module from receiving function as filename
AlexMa616
committed
ecd0667
fix(ci): correct FlagScale module imports and Enflame sitecustomize loading
AlexMa616
committed
630005d
fix(ci): allow cache miss fallback for dependency preparation failures
AlexMa616
committed
ea125ca
style: apply ruff formatting to test files
AlexMa616
committed
8adc10a
fix(ci): block flash_attn imports when NVTE_FLASH_ATTN=0
AlexMa616
committed
8f7134b
style: apply ruff format to sitecustomize.py
AlexMa616
committed
68e08f1
fix(ci): defer disabled flash attention stub failures
AlexMa616
committed
d63bc97
fix(ci): stabilize non-cuda training test runtime
AlexMa616
committed
de59e17
fix(ci): remove remaining legacy training imports
AlexMa616
committed
b7e10f3
fix(ci): make flash attention stubs discoverable
AlexMa616
committed
bfcd964
fix(ci): align FlagScale training imports with Megatron
AlexMa616
committed
a0d3767
fix(ci): use FlagScale-owned training imports
AlexMa616
committed
9b0836f
fix(ci): enforce prepared training dependency cache
AlexMa616
committed
b1ad210
fix(train): update checkpointing Megatron import
AlexMa616
committed
5ce226e

Commits on Sep 9, 2026

fix(train): resolve FlagScale PEFT imports
AlexMa616
committed
24a872d
fix(train): import unwrap_model from megatron core
AlexMa616
committed
f6712b1
fix(train): import unwrap_model from megatron core
AlexMa616
committed
72d94ad
fix(ci): align shared training unit tests
AlexMa616
committed
d9e71b9
fix(ci): restore prepared training dependencies reliably
AlexMa616
committed
d80f944
Merge branch 'main' into feat/dependency-coupling-all-platforms
AlexMa616
authored
1e7b447
fix(training): restore megatron overlay imports
AlexMa616
committed
4433309
fix(training): keep core utility imports in megatron
AlexMa616
committed
a991a18
fix(ci): upload prepared artifacts only on cache miss
AlexMa616
committed
5703ea7
fix(hetero): read enable_hetero from args before defaults
AlexMa616
committed
e893563
fix(training): preserve utility compatibility imports
AlexMa616
committed
4828401

Commits on Sep 10, 2026

Commits on Sep 11, 2026

fix(ci): configure prepared runtime for unit tests
AlexMa616
committed
662174b
style: format CI runtime script tests
AlexMa616
committed
5782a06
fix(ci): isolate tokenizer cache and unit runtime paths
AlexMa616
committed
cab85ca
fix(ci): clear stale prepared runtime artifacts
AlexMa616
committed
a7fd91c
fix: handle hetero memory defaults and musa attention
AlexMa616
committed
d28d589
fix(tests): update hetero baselines and musa attention
AlexMa616
committed
065d799
fix(ci): gate runtime artifacts on prepare cache misses
AlexMa616
committed
2c8c856
fix: limit cuda architecture checks to cuda
AlexMa616
committed
50b69e2
Merge branch 'main' into feat/dependency-coupling-all-platforms
AlexMa616
authored
3decbe6

Commits on Sep 12, 2026

Commits on Sep 13, 2026

ci: use two processes for unit tests
AlexMa616
committed
9c6dd43
ci: keep eight processes for cuda unit tests
AlexMa616
committed
f56b467
ci: add benchmark coverage for all platforms
AlexMa616
committed
e949132
ci: normalize platform benchmark configs
AlexMa616
committed
516203f
ci: fix platform benchmark and train names
AlexMa616
committed
59f4936
ci: avoid duplicate unit job prefix
AlexMa616
committed
0049b56
test: update platform benchmark baselines
AlexMa616
committed
2f63a6b

Commits on Sep 14, 2026

ci: tolerate non-utf8 benchmark logs
AlexMa616
committed
b3e2066
test: update MUSA benchmark baseline
AlexMa616
committed
63fd103
ci: clarify platform TE-FL preferences
AlexMa616
committed
947d0c8
ci: centralize Hugging Face module cache isolation
AlexMa616
committed
8fb31f6
fix: serialize Hugging Face dynamic module loading
AlexMa616
committed
3f902a5
fix: isolate Hugging Face module cache in CI
AlexMa616
committed
b136034
refactor: rename benchmark configs to include TE backend suffix
AlexMa616
committed
f8fd877
fix: update platform configs to match renamed benchmark cases
AlexMa616
committed
2c2e287
fix: change MetaX te_fl_prefer from flagos to reference
AlexMa616
committed
98fd10a
fix: rename MetaX config from flagos to reference backend
AlexMa616
committed
d6bbcb6
fix: rename MetaX gold baseline file to match config rename
AlexMa616
committed
92dac27
[Fix] Ascend get_device_arch_version() return None

Darryl233
authored and
FlagScale Agent
committed
a3088a5
Merge pull request #9 from Darryl233/feat/dependency-coupling-all-platforms
BrianPei
authored
494e828