[peft][ckpt] feat: add HF PEFT adapter export for LoRA/DoRA checkpoints#2574
Merged
[peft][ckpt] feat: add HF PEFT adapter export for LoRA/DoRA checkpoints#2574
Conversation
The infer_train_mode function only checked for 'finetune' in recipe names. Recipes named with 'sft' or 'peft' were not recognized as finetune mode, causing a ValueError. Add these keywords to the has_finetune check. Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Add `save_hf_adapter()` and `export_adapter_ckpt()` to AutoBridge for exporting Megatron-Bridge LoRA/DoRA adapters to HuggingFace PEFT format (adapter_config.json + adapter_model.safetensors). Key changes: - peft_bridge: perform LoRA merge in float32 to avoid bf16 precision loss - peft_bridge: add helpers to infer target_modules and build adapter config - auto_bridge: add save_hf_adapter() for direct model export - auto_bridge: add export_adapter_ckpt() for checkpoint-based export - Move adapter examples to examples/conversion/adapter/ with export, stream, and verification scripts Signed-off-by: Yu Yao <yaoyu.094@gmail.com> Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
- Add unit tests for peft_bridge helpers (infer_target_modules, build_adapter_config_dict) and save_hf_adapter in auto_bridge - Add functional test for Qwen3 LoRA adapter export end-to-end: creates toy model, attaches LoRA, exports via AutoBridge, verifies output files, config, weight shapes, and PEFT library loading - Add README for examples/conversion/adapter/ with usage docs - Update parent examples/conversion/README.md with adapter section Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
📝 WalkthroughWalkthroughThis PR introduces adapter export and verification functionality for Megatron-Bridge LoRA/DoRA adapters, adding new public methods to AutoBridge and helper utilities, three example scripts demonstrating the workflow, and comprehensive tests validating the implementation. Changes
Sequence Diagram(s)sequenceDiagram
participant User
participant export_adapter.py
participant AutoBridge
participant MegatronModel
participant HFAdapter as HuggingFace PEFT
participant FileSystem
User->>export_adapter.py: Run export script
export_adapter.py->>AutoBridge: Load from pretrained HF model
AutoBridge->>HFAdapter: Initialize with HF model
export_adapter.py->>AutoBridge: export_adapter_ckpt(megatron_ckpt)
AutoBridge->>MegatronModel: Load LoRA config from run_config.yaml
MegatronModel->>MegatronModel: Materialize base + LoRA weights
AutoBridge->>MegatronModel: Load adapter weights from checkpoint
AutoBridge->>AutoBridge: save_hf_adapter()
AutoBridge->>HFAdapter: Extract adapter weights
AutoBridge->>FileSystem: Write adapter_config.json
AutoBridge->>FileSystem: Write adapter_model.safetensors
FileSystem-->>User: Adapter artifacts saved
sequenceDiagram
participant User
participant verify_adapter.py
participant HFModel as HF Base Model
participant PEFTAdapter as PEFT Adapter
participant MegatronModel
participant Comparison
User->>verify_adapter.py: Run verification script
verify_adapter.py->>HFModel: Load base model
verify_adapter.py->>HFModel: Compute logits for prompt
HFModel-->>verify_adapter.py: Base logits
verify_adapter.py->>PEFTAdapter: Load via PEFT library
verify_adapter.py->>PEFTAdapter: Compute logits
PEFTAdapter-->>verify_adapter.py: PEFT logits
verify_adapter.py->>Comparison: Compare top-k logits
Comparison-->>verify_adapter.py: PEFT verification result
alt Megatron verification requested
verify_adapter.py->>MegatronModel: Load from checkpoint
MegatronModel->>MegatronModel: Construct LoRA model
MegatronModel->>MegatronModel: Load merged weights
verify_adapter.py->>MegatronModel: Compute logits
MegatronModel-->>verify_adapter.py: Megatron logits
verify_adapter.py->>Comparison: Compare PEFT vs Megatron
Comparison-->>User: Final PASSED/FAILED
else
verify_adapter.py-->>User: PEFT verification result
end
Estimated code review effort🎯 4 (Complex) | ⏱️ ~50 minutes Possibly related PRs
Suggested labels
Suggested reviewers
🚥 Pre-merge checks | ✅ 3 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (3 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 7
🧹 Nitpick comments (3)
tests/unit_tests/models/test_adapter_export.py (1)
39-297: Add a module-level test category marker and reuse fixtures for common setup.The file repeats the same distributed patch context and does not label tests with a pytest category.
As per coding guidelines, `tests/**/*.py: Use pytest fixtures for common setup in unit tests` and `Use 'pytest.mark' to categorize tests (unit, integration, system)`.Suggested refactor
import pytest import torch @@ from megatron.bridge.models.conversion.peft_bridge import ( @@ ) +pytestmark = pytest.mark.unit + +@pytest.fixture +def no_distributed(): + with ( + patch("torch.distributed.is_available", return_value=False), + patch("torch.distributed.is_initialized", return_value=False), + ): + yield + @@ - def test_save_creates_files(self, tmp_path): + def test_save_creates_files(self, tmp_path, no_distributed): @@ - with ( - patch("torch.distributed.is_available", return_value=False), - patch("torch.distributed.is_initialized", return_value=False), - ): - from megatron.bridge.models.conversion.auto_bridge import AutoBridge - - AutoBridge.save_hf_adapter( - mock_bridge, - model=[MagicMock()], - path=output_dir, - peft_config=lora, - base_model_name_or_path="test/model", - ) + from megatron.bridge.models.conversion.auto_bridge import AutoBridge + AutoBridge.save_hf_adapter( + mock_bridge, + model=[MagicMock()], + path=output_dir, + peft_config=lora, + base_model_name_or_path="test/model", + )🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@tests/unit_tests/models/test_adapter_export.py` around lines 39 - 297, Add a module-level pytest marker (e.g., pytestmark = pytest.mark.unit) at top of the file and refactor repeated setup into fixtures: create a fixture (e.g., mock_distributed) that yields the patched torch.distributed.is_available/is_initialized context and another fixture (e.g., mock_bridge_with_weights) that returns a MagicMock configured similarly to how mock_bridge is created in TestSaveHfAdapter tests; update tests like TestSaveHfAdapter.test_save_creates_files, test_save_raises_on_empty_adapter, and test_save_infers_base_model_path to accept and use these fixtures instead of repeating patch/context creation and mock_bridge construction, and keep references to AutoBridge.save_hf_adapter and the MagicMock export_adapter_weights behavior unchanged.tests/functional_tests/models/qwen/test_qwen3_peft_export.py (2)
131-133: Add pytest markers to categorize the test class.Per coding guidelines, tests should use
pytest.markfor categorization. Consider adding markers to indicate this is a functional test and document any hardware/environment requirements.♻️ Proposed fix
+@pytest.mark.functional +@pytest.mark.slow # if applicable class TestQwen3PeftExport: """Functional tests for Qwen3 LoRA adapter export to HuggingFace PEFT format."""🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@tests/functional_tests/models/qwen/test_qwen3_peft_export.py` around lines 131 - 133, Add pytest markers to the TestQwen3PeftExport class by decorating it with appropriate pytest.mark annotations (e.g., pytest.mark.functional and any required hardware/environment markers such as pytest.mark.gpu or pytest.mark.requires_internet) so the test suite can categorize and selectively run it; update the class definition (TestQwen3PeftExport) to include these markers and document specific environment needs in the marker names or a short docstring comment immediately above the class.
158-179: Consider moving repeated import to module level.The
from safetensors.torch import load_fileimport appears in three test methods (lines 160, 183, 195). Moving it to the module-level imports would be cleaner.♻️ Proposed refactor
Add to imports section:
from megatron.bridge.training.model_load_save import temporary_distributed_context +from safetensors.torch import load_fileThen remove the local imports from test methods.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@tests/functional_tests/models/qwen/test_qwen3_peft_export.py` around lines 158 - 179, The repeated local import "from safetensors.torch import load_file" used in tests like test_safetensors_weight_pairs should be moved to the module-level imports: add a single "from safetensors.torch import load_file" at the top of the test file and remove the in-function imports from each test method (e.g., test_safetensors_weight_pairs and the other safetensors-related test functions) so all tests reuse the same module-level symbol.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@examples/conversion/adapter/README.md`:
- Around line 51-55: The fenced output block showing the adapter file tree is
missing a language label (triggers MD040); update the opening fence from ``` to
```text for the block containing "my_adapter/ ├── adapter_config.json └──
adapter_model.safetensors" (in the README's fenced code block) so the snippet is
marked as plain text—apply the same change to any other similar unlabeled
file-tree fences in the README.
- Around line 35-78: Update the example command invocations in README.md to use
the project's recommended runner by prefixing each "python" call with "uv run"
for both export_adapter.py and verify_adapter.py examples (e.g., change the
export command invoking export_adapter.py with --hf-model-id,
--megatron-peft-checkpoint, --output-hf-path and the verify commands invoking
verify_adapter.py with --hf-model-id, --hf-adapter-path,
--megatron-peft-checkpoint to use "uv run python ..."). Ensure every occurrence
of "python examples/conversion/adapter/export_adapter.py" and "python
examples/conversion/adapter/verify_adapter.py" in the README is replaced so the
shown CLI examples match the rest of the conversion docs.
In `@examples/conversion/adapter/verify_adapter.py`:
- Around line 163-177: The block that enters a distributed context with
temporary_distributed_context(ctx.__enter__()) can leak the context on
exceptions and currently only loads weights into model[0]; wrap the code after
ctx.__enter__() in try/finally (or try/except/finally) to ensure ctx.__exit__()
is always called, and when loading the checkpoint (dist_checkpointing.load)
determine whether keys are "model" or "model0"/"model1"/... then iterate over
the list returned by provide_distributed_model and call load_state_dict on each
chunk (e.g., for i, m in enumerate(model): select loaded_sd[f"model{i}"] if
per-chunk keys exist, otherwise use loaded_sd["model"]) so all model chunks
receive their corresponding weights; reference functions:
temporary_distributed_context, provide_distributed_model,
_generate_model_state_dict, apply_peft_adapter_filter_to_state_dict,
dist_checkpointing.load, and model[0].load_state_dict.
In `@scripts/training/run_recipe.py`:
- Line 212: Update the inference failure message to list the current accepted
finetune keywords to match the detection logic: where the code computes
has_finetune (variable name has_finetune) from lowered including "finetune",
"sft", or "peft", change the error/help text emitted on inference failure to
mention "finetune", "sft", and "peft" (and still include "pretrain" if relevant)
so the guidance matches the logic in run_recipe.py around has_finetune.
In `@src/megatron/bridge/models/conversion/auto_bridge.py`:
- Around line 1051-1056: The code only calls model[0].load_state_dict(...),
leaving other chunks uninitialized; change this to iterate over all model chunks
and load each chunk's state dict: after computing sharded_state_dict,
apply_peft_adapter_filter_to_state_dict, and loaded_sd/model_key, loop over
enumerate(model) and for each index i determine a chunk-specific key by
preferring f"{model_key}.{i}" (or search for the first loaded_sd key that
startswith f"{model_key}.{i}") and call
model[i].load_state_dict(loaded_sd[chunk_key], strict=False); if no per-index
keys exist, fall back to loading the same loaded_sd[model_key] into every
model[i].load_state_dict(...) so all shards are populated.
In `@src/megatron/bridge/models/conversion/peft_bridge.py`:
- Around line 839-851: Add explicit shape validation before performing the LoRA
merge: compute or infer the expected delta tensor shape (based on
base_weight.shape and dim) and validate that the transformed tensors used by
LoRAMerge (e.g., linear_out_weight, linear_in_weight after casting/moving:
linear_out_on_base, linear_in_on_base) will produce a delta whose shape matches
base_weight.shape; if the shapes mismatch, raise a clear exception. Update the
merge call site in the function that uses orig_dtype, base_weight, LoRAMerge,
linear_out_weight, linear_in_weight, alpha, and dim to perform this check right
before calling merger.merge and only proceed to return merged.to(orig_dtype)
when the validation passes.
In `@tests/functional_tests/models/qwen/test_qwen3_peft_export.py`:
- Around line 73-89: The fixture qwen3_toy_model_dir currently pulls a tokenizer
from the network via AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B"); replace
this network dependency by creating and saving a minimal local tokenizer into
model_dir (e.g., construct a simple tokenizer instance with minimal vocab/config
and call tokenizer.save_pretrained(model_dir) instead of
AutoTokenizer.from_pretrained) so tests don’t require HF Hub access, or
alternatively annotate the fixture with `@pytest.mark.network` if you intend to
allow external network in CI; update the code references around the
qwen3_toy_model_dir fixture and remove the AutoTokenizer.from_pretrained call.
---
Nitpick comments:
In `@tests/functional_tests/models/qwen/test_qwen3_peft_export.py`:
- Around line 131-133: Add pytest markers to the TestQwen3PeftExport class by
decorating it with appropriate pytest.mark annotations (e.g.,
pytest.mark.functional and any required hardware/environment markers such as
pytest.mark.gpu or pytest.mark.requires_internet) so the test suite can
categorize and selectively run it; update the class definition
(TestQwen3PeftExport) to include these markers and document specific environment
needs in the marker names or a short docstring comment immediately above the
class.
- Around line 158-179: The repeated local import "from safetensors.torch import
load_file" used in tests like test_safetensors_weight_pairs should be moved to
the module-level imports: add a single "from safetensors.torch import load_file"
at the top of the test file and remove the in-function imports from each test
method (e.g., test_safetensors_weight_pairs and the other safetensors-related
test functions) so all tests reuse the same module-level symbol.
In `@tests/unit_tests/models/test_adapter_export.py`:
- Around line 39-297: Add a module-level pytest marker (e.g., pytestmark =
pytest.mark.unit) at top of the file and refactor repeated setup into fixtures:
create a fixture (e.g., mock_distributed) that yields the patched
torch.distributed.is_available/is_initialized context and another fixture (e.g.,
mock_bridge_with_weights) that returns a MagicMock configured similarly to how
mock_bridge is created in TestSaveHfAdapter tests; update tests like
TestSaveHfAdapter.test_save_creates_files, test_save_raises_on_empty_adapter,
and test_save_infers_base_model_path to accept and use these fixtures instead of
repeating patch/context creation and mock_bridge construction, and keep
references to AutoBridge.save_hf_adapter and the MagicMock
export_adapter_weights behavior unchanged.
ℹ️ Review info
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (11)
3rdparty/Megatron-LMexamples/conversion/README.mdexamples/conversion/adapter/README.mdexamples/conversion/adapter/export_adapter.pyexamples/conversion/adapter/stream_adapter_weights.pyexamples/conversion/adapter/verify_adapter.pyscripts/training/run_recipe.pysrc/megatron/bridge/models/conversion/auto_bridge.pysrc/megatron/bridge/models/conversion/peft_bridge.pytests/functional_tests/models/qwen/test_qwen3_peft_export.pytests/unit_tests/models/test_adapter_export.py
Comment on lines
+35
to
+78
| ```bash | ||
| python examples/conversion/adapter/export_adapter.py \ | ||
| --hf-model-id meta-llama/Llama-3.2-1B \ | ||
| --megatron-peft-checkpoint /path/to/finetune_ckpt \ | ||
| --output-hf-path ./my_adapter | ||
| ``` | ||
|
|
||
| | Argument | Description | | ||
| |---|---| | ||
| | `--hf-model-id` | HuggingFace model name or local path (architecture + base weights) | | ||
| | `--megatron-peft-checkpoint` | Path to the Megatron-Bridge distributed checkpoint containing LoRA adapter weights | | ||
| | `--output-hf-path` | Output directory (default: `./my_adapter`) | | ||
| | `--trust-remote-code` | Allow custom code from the HuggingFace repository | | ||
|
|
||
| **Output structure:** | ||
|
|
||
| ``` | ||
| my_adapter/ | ||
| ├── adapter_config.json | ||
| └── adapter_model.safetensors | ||
| ``` | ||
|
|
||
| ### 2. `verify_adapter.py` — Export Verification | ||
|
|
||
| Loads the exported adapter with the `peft` library and runs verification | ||
| checks: | ||
|
|
||
| - The PEFT model logits must differ from the base model (adapter has effect). | ||
| - When `--megatron-peft-checkpoint` is provided, the top-k predicted tokens | ||
| from the PEFT model must match those from the Megatron model with merged | ||
| weights. | ||
|
|
||
| ```bash | ||
| # Quick check (PEFT-only, no Megatron comparison) | ||
| python examples/conversion/adapter/verify_adapter.py \ | ||
| --hf-model-id meta-llama/Llama-3.2-1B \ | ||
| --hf-adapter-path ./my_adapter | ||
|
|
||
| # Full verification (compares against Megatron checkpoint) | ||
| python examples/conversion/adapter/verify_adapter.py \ | ||
| --hf-model-id meta-llama/Llama-3.2-1B \ | ||
| --hf-adapter-path ./my_adapter \ | ||
| --megatron-peft-checkpoint /path/to/finetune_ckpt/iter_0000020 | ||
| ``` |
Contributor
There was a problem hiding this comment.
Align command examples with uv run usage.
Use uv run python ... in this README to match the rest of the conversion examples and avoid environment mismatch.
Suggested doc fix
-python examples/conversion/adapter/export_adapter.py \
+uv run python examples/conversion/adapter/export_adapter.py \
--hf-model-id meta-llama/Llama-3.2-1B \
--megatron-peft-checkpoint /path/to/finetune_ckpt \
--output-hf-path ./my_adapter
@@
-python examples/conversion/adapter/verify_adapter.py \
+uv run python examples/conversion/adapter/verify_adapter.py \
--hf-model-id meta-llama/Llama-3.2-1B \
--hf-adapter-path ./my_adapter
@@
-python examples/conversion/adapter/verify_adapter.py \
+uv run python examples/conversion/adapter/verify_adapter.py \
--hf-model-id meta-llama/Llama-3.2-1B \
--hf-adapter-path ./my_adapter \
--megatron-peft-checkpoint /path/to/finetune_ckpt/iter_0000020🧰 Tools
🪛 markdownlint-cli2 (0.21.0)
[warning] 51-51: Fenced code blocks should have a language specified
(MD040, fenced-code-language)
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@examples/conversion/adapter/README.md` around lines 35 - 78, Update the
example command invocations in README.md to use the project's recommended runner
by prefixing each "python" call with "uv run" for both export_adapter.py and
verify_adapter.py examples (e.g., change the export command invoking
export_adapter.py with --hf-model-id, --megatron-peft-checkpoint,
--output-hf-path and the verify commands invoking verify_adapter.py with
--hf-model-id, --hf-adapter-path, --megatron-peft-checkpoint to use "uv run
python ..."). Ensure every occurrence of "python
examples/conversion/adapter/export_adapter.py" and "python
examples/conversion/adapter/verify_adapter.py" in the README is replaced so the
shown CLI examples match the rest of the conversion docs.
Comment on lines
+51
to
+55
| ``` | ||
| my_adapter/ | ||
| ├── adapter_config.json | ||
| └── adapter_model.safetensors | ||
| ``` |
Contributor
There was a problem hiding this comment.
Add a language to the fenced output block.
The output structure block is missing a fence language and triggers MD040.
Suggested markdownlint fix
-```
+```text
my_adapter/
├── adapter_config.json
└── adapter_model.safetensors</details>
<!-- suggestion_start -->
<details>
<summary>📝 Committable suggestion</summary>
> ‼️ **IMPORTANT**
> Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
```suggestion
🧰 Tools
🪛 markdownlint-cli2 (0.21.0)
[warning] 51-51: Fenced code blocks should have a language specified
(MD040, fenced-code-language)
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@examples/conversion/adapter/README.md` around lines 51 - 55, The fenced
output block showing the adapter file tree is missing a language label (triggers
MD040); update the opening fence from ``` to ```text for the block containing
"my_adapter/ ├── adapter_config.json └── adapter_model.safetensors" (in the
README's fenced code block) so the snippet is marked as plain text—apply the
same change to any other similar unlabeled file-tree fences in the README.
| lowered = recipe_name.lower() | ||
| has_pretrain = "pretrain" in lowered | ||
| has_finetune = "finetune" in lowered | ||
| has_finetune = "finetune" in lowered or "sft" in lowered or "peft" in lowered |
Contributor
There was a problem hiding this comment.
Update the inference error text to reflect new accepted finetune keywords.
Line 212 now recognizes "sft" and "peft" for finetune inference, but the failure message still instructs only "pretrain"/"finetune". Please align the message so failed inference guidance is accurate.
Suggested patch
ERR_INFER_MODE_FAILED = (
"Unable to infer training mode from recipe name. "
- "Please include 'pretrain' or 'finetune' in the recipe name or pass --mode explicitly."
+ "Please include 'pretrain', 'finetune', 'sft', or 'peft' in the recipe name, "
+ "or pass --mode explicitly."
)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@scripts/training/run_recipe.py` at line 212, Update the inference failure
message to list the current accepted finetune keywords to match the detection
logic: where the code computes has_finetune (variable name has_finetune) from
lowered including "finetune", "sft", or "peft", change the error/help text
emitted on inference failure to mention "finetune", "sft", and "peft" (and still
include "pretrain" if relevant) so the guidance matches the logic in
run_recipe.py around has_finetune.
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test 53a1d64 |
Patch get_tensor_model_parallel_world_size to return 1 in TestMergeSingleAdapterWeight so LoRAMerge.merge() takes the tp_size==1 path and avoids requiring an initialized tensor model parallel group. Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test 578403d |
1 similar comment
Contributor
Author
|
/ok to test 578403d |
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Contributor
Author
|
/ok to test d0c76b2 |
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>
Contributor
Author
|
/ok to test cd79d06 |
3 similar comments
Contributor
Author
|
/ok to test cd79d06 |
Contributor
Author
|
/ok to test cd79d06 |
Contributor
Author
|
/ok to test cd79d06 |
…xport Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor # Conflicts: # pyproject.toml # uv.lock
Contributor
Author
|
/ok to test 3372000 |
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test 6eb53b9 |
Add TestExportAdapterCkpt with 10 tests covering the orchestrator method that was previously uncovered by Codecov. Tests mock heavy infrastructure (dist_checkpointing, distributed context, model materialisation) and exercise config parsing, VLMLoRA selection, error paths, float32 dtype enforcement, and base_model_name fallback logic. Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test eabcce3 |
cuichenx
previously approved these changes
Mar 11, 2026
8 tasks
…xport Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test 8fa62fa |
…xport Made-with: Cursor # Conflicts: # uv.lock
Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com> Made-with: Cursor
Contributor
Author
|
/ok to test f2104df |
cuichenx
previously approved these changes
Mar 14, 2026
Contributor
Author
|
/ok to test 4e1fdd2 |
Signed-off-by: oliver könig <okoenig@nvidia.com>
Contributor
|
/ok to test 3ba210d |
ko3n1g
approved these changes
Mar 16, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
save_hf_adapter()andexport_adapter_ckpt()toAutoBridgefor exporting Megatron-Bridge LoRA/DoRA adapters to HuggingFace PEFT format (adapter_config.json+adapter_model.safetensors)infer_target_modules_from_adapter_weights()andbuild_adapter_config_dict()inpeft_bridge.pyexamples/conversion/adapter/for export, verification, and streamingsave_hf_adapterTest plan
pytest tests/unit_tests/models/test_adapter_export.py(19 passed)pytest tests/functional_tests/models/qwen/test_qwen3_peft_export.py(7 passed)pytest tests/unit_tests/models/test_model_bridge_lora.pyMade with Cursor
Summary by CodeRabbit
New Features
Documentation
Tests