Repository navigation

Unify LoRA target selection with HF model layouts - #3318

Merged

yueming-yuan merged 95 commits into

mainfrom

yueming/hf-lora-targets

Sep 23, 2026

Merged

Unify LoRA target selection with HF model layouts#3318
yueming-yuan merged 95 commits into
mainfrom
yueming/hf-lora-targets

Commits on Sep 16, 2026
- ci(docker): bump Megatron-Bridge to the bridge-new branch
  yushengsu-thu
  committed
  b371523
Commits on Sep 17, 2026
Commits on Sep 18, 2026
Commits on Sep 19, 2026
- ci(docker): install Megatron-Bridge from the renamed bridge branch
  yushengsu-thu
  committed
  aa28968
- tests(verifiers): pin prime-sandboxes<0.3 next to prime 0.6.19
  yushengsu-thu
  committed
  8e09d8e
- Define shared HF LoRA target groups
  yueming-yuan
  committed
  118abf2
- Expand HF LoRA target layouts and defaults
  yueming-yuan
  committed
  68c1519
- Keep HF LoRA rules with each model family
  yueming-yuan
  committed
  0af0612
- Use shared HF targets across LoRA entrypoints
  yueming-yuan
  committed
  ba52b71
- Use HF selections for LoRA export and serving
  yueming-yuan
  committed
  1b12ebb
- Trim LoRA target comments to their constraints
  yueming-yuan
  committed
  1a0c3b2
- Name LoRA selectors and gathered module names explicitly
  yueming-yuan
  committed
  95e5cd9
- Return the HF selection from Bridge target resolution
  yueming-yuan
  committed
  7b6c5cc
- Clarify LoRA selection precedence and export coverage
  yueming-yuan
  committed
  434ba32
- Place LoRA selection tests beside their source modules
  yueming-yuan
  committed
  dec888e
- docs: Regenerate the multi-LoRA example page
  yueming-yuan
  committed
  a01e63d
- style: Format the LoRA target selection changes
  yueming-yuan
  committed
  2cff19a
- style: Align LoRA imports with pinned isort
  yueming-yuan
  committed
  504078e
- Exclude DSA indexers from LoRA training groups
  yueming-yuan
  committed
  1be8802
- Align HF LoRA groups with current model definitions
  yueming-yuan
  committed
  ffb8fdb
- Check LoRA targets against native HF model structures
  yueming-yuan
  committed
  1d5406d
- style: Format HF model target test
  yueming-yuan
  committed
  d8a96d7
- Separate HF checkpoint naming from LoRA target binding
  yueming-yuan
  committed
  ee5911b
- Align LoRA routing tests with HF target selection
  yueming-yuan
  committed
  44d0a1f
- style: Sort inherited verifier requirements
  yueming-yuan
  committed
  6114b10
- Preserve exclusions and reject partial template injection
  yueming-yuan
  committed
  97db1b4
- style: Format template selection check
  yueming-yuan
  committed
  ff2ee24
- Gather missing expert names only across EP
  yueming-yuan
  committed
  cc431d6
- Accept the serialized Qwen3 MoE expert count
  yueming-yuan
  committed
  d6a5263
Commits on Sep 20, 2026
- Group HF utilities in one package
  yueming-yuan
  committed
  fe2f074
- tests(verifiers): keep requirements.txt sorted for requirements-txt-fixer
  yushengsu-thu
  committed
  291c4b9
- Keep Inkling adapter naming in native LoRA
  yueming-yuan
  committed
  43dead7
- style: Format Inkling target mapping
  yueming-yuan
  committed
  2b72f90
- Give the Inkling test a unique pytest module name
  yueming-yuan
  committed
  22cde1c
- style: Format the Inkling target assertion
  yueming-yuan
  committed
  3060f48
- ci(docker): pin Megatron-Bridge to bridge@2e09c234a
  yushengsu-thu
  committed
  93e727b
- Merge branch 'main' into bump/megatron-bridge-bridge-new
  yushengsu-thu
  committed
  1c93da0
Commits on Sep 21, 2026
Commits on Sep 22, 2026
- Normalize remote HF configs before native target inspection
  yueming-yuan
  committed
  c5cfad4
- Name LoRA target normalization and checkpoint parameters explicitly
  yueming-yuan
  committed
  36599a4
- Separate LoRA target selection from adapter binding
  yueming-yuan
  committed
  2f110c2
- style: Format LoRA target binding
  yueming-yuan
  committed
  ffe4301
- Remove exhaustive LoRA injection and export audits
  yueming-yuan
  committed
  ddf2a1d
- docs: Describe LoRA name resolution without coverage audits
  yueming-yuan
  committed
  8e25dd7
- style: Format lightweight HF name mapping
  yueming-yuan
  committed
  62b70f9
- Exclude packed expert projections from GLM attention ablation
  yueming-yuan
  committed
  8d91598
- Resolve scoped HF exclusions before retaining LoRA targets
  yueming-yuan
  committed
  ba82dee
- Restore native Inkling LoRA target discovery
  yueming-yuan
  committed
  3ce19f5
- docs: Describe native Inkling automatic adapter discovery
  yueming-yuan
  committed
  c6445c4
- style: Format Inkling target discovery
  yueming-yuan
  committed
  a39ece3
- Merge main into HF LoRA targets
  yueming-yuan
  committed
  294fe27
- Restrict Tinker targets to training groups
  yueming-yuan
  committed
  53e27ae
- Use canonical aliases for expert LoRA targets
  yueming-yuan
  committed
  80d414d
- Allow complete GDN targets with canonical LoRA
  yueming-yuan
  committed
  a7e2c05
- style: format canonical target assertion
  yueming-yuan
  committed
  c005f80
- Share LoRA target matching across backends
  yueming-yuan
  committed
  e3d70f4
- Reuse expert target detection in argument validation
  yueming-yuan
  committed
  166fe8d
- Reuse LoRA configuration during argument validation
  yueming-yuan
  committed
  52deccb
- style: Format shared LoRA target matching
  yueming-yuan
  committed
  f8c602e
Commits on Sep 23, 2026
- Select LoRA training groups through target modules
  yueming-yuan
  committed
  eed119b
- Expand paired LoRA targets for packed projections
  yueming-yuan
  committed
  8d7775c
- style: Format packed LoRA target handling
  yueming-yuan
  committed
  58cb17a
- Consolidate LoRA arguments and shared utilities
  yueming-yuan
  committed
  0e10a35
- style: Sort LoRA utility imports
  yueming-yuan
  committed
  3b52fb7
- Consolidate LoRA arguments and shared utilities
  yueming-yuan
  committed
  126f33e
- style: Sort LoRA utility imports
  yueming-yuan
  committed
  b4dd2c9
- Merge main into HF LoRA target selection
  yueming-yuan
  committed
  2a8674f
- Stack HF LoRA targets on the LoRA utility refactor
  yueming-yuan
  committed
  ddbd2da
- Group HF config utilities under hf_utils
  yueming-yuan
  committed
  277508c
- Stack HF target selection on the HF utility refactor
  yueming-yuan
  committed
  647d201
- Move HF LoRA target definitions into the LoRA package
  yueming-yuan
  committed
  874dd22
- style: Sort HF LoRA target imports
  yueming-yuan
  committed
  0b1972d
- Connect native Kimi K3 LoRA to HF targets
  yueming-yuan
  committed
  8d47c17
- style: Format Kimi K3 target definitions
  yueming-yuan
  committed
  e5b76a5
- Align Kimi K3 LoRA defaults with the training recipe
  yueming-yuan
  committed
  223a575
- style: Format Kimi K3 launcher spacing
  yueming-yuan
  committed
  faba71d
- Drop inherited verifier dependency change from LoRA PR
  yueming-yuan
  committed
  c3c0b89
- Resolve LoRA selections before assigning runtime targets
  yueming-yuan
  committed
  99563b9
- Check LoRA target resolution against Bridge and native export
  yueming-yuan
  committed
  13e6725
- style: Format LoRA resolution checks
  yueming-yuan
  committed
  2ed7329
- Merge main into LoRA utility refactor
  yueming-yuan
  committed
  1d18255
- Merge updated LoRA utility base
  yueming-yuan
  committed
  9bc0cbd
- Merge main after LoRA refactor squash
  yueming-yuan
  committed
  eff3767