# What does this PR do?
Fixes a bunch of violations.
Note: this patch touches all files but post_training.py that will be
significantly changed by #1437, hence leaving it out of the picture for
now.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
Testing with https://github.com/meta-llama/llama-stack/pull/1543
Also checked that GPU training works with the change:
```
INFO: ::1:53316 - "POST /v1/post-training/supervised-fine-tune HTTP/1.1" 200 OK
INFO: ::1:53316 - "GET /v1/post-training/job/status?job_uuid=test-jobb5ca2d84-d541-42f8-883b-762828b4c0e7 HTTP/1.1" 200 OK
INFO: ::1:53316 - "GET /v1/post-training/job/artifacts?job_uuid=test-jobb5ca2d84-d541-42f8-883b-762828b4c0e7 HTTP/1.1" 200 OK
21:24:01.161 [END] /v1/post-training/supervised-fine-tune [StatusCode.OK] (32526.75ms)
21:23:28.769 [DEBUG] Setting manual seed to local seed 3918872849. Local seed is seed + rank = 3918872849 + 0
21:23:28.996 [INFO] Identified model_type = Llama3_2. Ignoring output.weight in checkpoint in favor of the tok_embedding.weight tied weights.
21:23:29.933 [INFO] Memory stats after model init:
GPU peak memory allocation: 6.05 GiB
GPU peak memory reserved: 6.10 GiB
GPU peak memory active: 6.05 GiB
21:23:29.934 [INFO] Model is initialized with precision torch.bfloat16.
21:23:30.115 [INFO] Tokenizer is initialized.
21:23:30.118 [INFO] Optimizer is initialized.
21:23:30.119 [INFO] Loss is initialized.
21:23:30.896 [INFO] Dataset and Sampler are initialized.
21:23:30.898 [INFO] Learning rate scheduler is initialized.
21:23:31.618 [INFO] Memory stats after model init:
GPU peak memory allocation: 6.24 GiB
GPU peak memory reserved: 6.30 GiB
GPU peak memory active: 6.24 GiB
21:23:31.620 [INFO] Starting checkpoint save...
21:23:59.428 [INFO] Model checkpoint of size 6.43 GB saved to /home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/consolidated.00.pth
21:23:59.445 [INFO] Adapter checkpoint of size 0.00 GB saved to /home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/adapter/adapter.pth
```
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
Updated all instances of datetime.now() to use timezone.utc for
consistency in handling time across different systems. This ensures that
timestamps are always in Coordinated Universal Time (UTC), avoiding
issues with time zone discrepancies and promoting uniformity in
time-related data.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
It's a dict that may contain different types, as per
resolver:instantiate_provider implementation. (AFAIU it also never
contains ProviderSpecs, but *instances* of provider implementations.)
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
mypy passing if enabled checks for these modules. (See #1543)
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
Added missing shutdown handler. (Currently empty.)
Without it, when server shuts down, it posts the following warning:
```
__main__:129 server: No shutdown method for TorchtunePostTrainingImpl
```
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
(The test plan assumes shutdown logic is fixed, see #1495)
Without the patch:
```
INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit)
INFO: Shutting down
INFO: Waiting for application shutdown.
INFO 2025-03-10 20:56:43,961 __main__:140 server: Shutting down
INFO 2025-03-10 20:56:43,962 __main__:124 server: Shutting down DatasetsRoutingTable
INFO 2025-03-10 20:56:43,964 __main__:124 server: Shutting down DatasetIORouter
INFO 2025-03-10 20:56:43,965 __main__:124 server: Shutting down ScoringFunctionsRoutingTable
INFO 2025-03-10 20:56:43,966 __main__:124 server: Shutting down ScoringRouter
INFO 2025-03-10 20:56:43,967 __main__:124 server: Shutting down ModelsRoutingTable
INFO 2025-03-10 20:56:43,968 __main__:124 server: Shutting down InferenceRouter
INFO 2025-03-10 20:56:43,969 __main__:124 server: Shutting down ShieldsRoutingTable
INFO 2025-03-10 20:56:43,971 __main__:124 server: Shutting down SafetyRouter
INFO 2025-03-10 20:56:43,972 __main__:124 server: Shutting down VectorDBsRoutingTable
INFO 2025-03-10 20:56:43,973 __main__:124 server: Shutting down VectorIORouter
INFO 2025-03-10 20:56:43,974 __main__:124 server: Shutting down ToolGroupsRoutingTable
INFO 2025-03-10 20:56:43,975 __main__:124 server: Shutting down ToolRuntimeRouter
INFO 2025-03-10 20:56:43,976 __main__:124 server: Shutting down MetaReferenceAgentsImpl
INFO 2025-03-10 20:56:43,977 __main__:124 server: Shutting down TelemetryAdapter
INFO 2025-03-10 20:56:43,978 __main__:124 server: Shutting down TorchtunePostTrainingImpl
WARNING 2025-03-10 20:56:43,979 __main__:129 server: No shutdown method for TorchtunePostTrainingImpl
INFO 2025-03-10 20:56:43,979 __main__:124 server: Shutting down BenchmarksRoutingTable
INFO 2025-03-10 20:56:43,980 __main__:124 server: Shutting down EvalRouter
INFO 2025-03-10 20:56:43,981 __main__:124 server: Shutting down DistributionInspectImpl
INFO: Application shutdown complete.
INFO: Finished server process [33862]
```
Run with the patch and observe no warning:
```
$ kill -INT $(ps ax | grep llama_stack.distribution.server.server | grep -v nvim | awk -e '{print $1}' | sort | head -n 1)
```
```
INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit)
INFO: Shutting down
INFO: Waiting for application shutdown.
INFO 2025-03-11 00:32:56,863 __main__:140 server: Shutting down
INFO 2025-03-11 00:32:56,864 __main__:124 server: Shutting down DatasetsRoutingTable
INFO 2025-03-11 00:32:56,866 __main__:124 server: Shutting down DatasetIORouter
INFO 2025-03-11 00:32:56,867 __main__:124 server: Shutting down ScoringFunctionsRoutingTable
INFO 2025-03-11 00:32:56,868 __main__:124 server: Shutting down ScoringRouter
INFO 2025-03-11 00:32:56,869 __main__:124 server: Shutting down ModelsRoutingTable
INFO 2025-03-11 00:32:56,870 __main__:124 server: Shutting down InferenceRouter
INFO 2025-03-11 00:32:56,871 __main__:124 server: Shutting down ShieldsRoutingTable
INFO 2025-03-11 00:32:56,872 __main__:124 server: Shutting down SafetyRouter
INFO 2025-03-11 00:32:56,873 __main__:124 server: Shutting down VectorDBsRoutingTable
INFO 2025-03-11 00:32:56,874 __main__:124 server: Shutting down VectorIORouter
INFO 2025-03-11 00:32:56,875 __main__:124 server: Shutting down ToolGroupsRoutingTable
INFO 2025-03-11 00:32:56,876 __main__:124 server: Shutting down ToolRuntimeRouter
INFO 2025-03-11 00:32:56,877 __main__:124 server: Shutting down MetaReferenceAgentsImpl
INFO 2025-03-11 00:32:56,878 __main__:124 server: Shutting down TelemetryAdapter
INFO 2025-03-11 00:32:56,879 __main__:124 server: Shutting down TorchtunePostTrainingImpl
INFO 2025-03-11 00:32:56,880 __main__:124 server: Shutting down BenchmarksRoutingTable
INFO 2025-03-11 00:32:56,881 __main__:124 server: Shutting down EvalRouter
INFO 2025-03-11 00:32:56,882 __main__:124 server: Shutting down DistributionInspectImpl
```
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
- using `eval` is a security risk
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
- see https://github.com/meta-llama/llama-stack/pull/1327
cc @SLR722 we will need to update the corresponding dataset via
```python
def update_to_json_str():
dataset = datasets.load_dataset(...)
processed_dataset = dataset[split].map(
lambda x: {
"column": json.dumps(eval(x["column"]))
}
)
processed_dataset.push_to_hub(...)
```
[//]: # (## Documentation)
This is a follow up to:
https://github.com/meta-llama/llama-stack/pull/1140
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
[Provide a short summary of what this PR does and why. Link to relevant
issues if applicable.]
Avoid unnecessary GPU memory clean attempt when the GPU is not used for
training.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
With CPU:
```
INFO 2025-02-26 16:43:56,267 torchtune.utils._logging:121: Model checkpoint of size 6.43 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/consolidated.00.pth
INFO 2025-02-26 16:43:56,274 torchtune.utils._logging:132: Adapter checkpoint of size 0.00 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/adapter/adapter.pth
model_file_path /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0
```
With CUDA:
```
INFO 2025-02-26 21:39:24,314 torchtune.utils._logging:121: Model checkpoint of size 6.43 GB saved to /home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/consolidated.00.pth
INFO 2025-02-26 21:39:24,333 torchtune.utils._logging:132: Adapter checkpoint of size 0.00 GB saved to /home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/adapter/adapter.pth
model_file_path /home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0
```
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
## context
Now, in llama stack, we only support inference / eval a finetuned
checkpoint with meta-reference as inference provider. This is
sub-optimal since meta-reference is pretty slow.
Our vision is that developer can inference / eval a finetuned checkpoint
produced by post training apis with all the inference providers on the
stack. To achieve this, we'd like to define an unified output checkpoint
format for post training providers. So that, all the inference provider
can respect that format for customized model inference.
By spotting check how
[ollama](https://github.com/ollama/ollama/blob/main/docs/import.md) and
[fireworks](https://docs.fireworks.ai/models/uploading-custom-models) do
inference on a customized model, we defined the output checkpoint format
as /adapter/adapter_config.json and /adapter/adapter_model.safetensors
(as we only support LoRA post training now, we begin from adapter only
checkpoint)
## test
we kick off a post training job and configured checkpoint format as
'huggingface'. Output files
we did a proof of concept with ollama to see if ollama can inference our
finetuned checkpoint
1. create Modelfile like
<img width="799" alt="Screenshot 2025-01-22 at 5 04 18 PM"
src="https://github.com/user-attachments/assets/7fca9ac3-a294-44f8-aab1-83852c600609"
/>
2. create a customized model with `ollama create llama_3_2_finetuned`
and run inference successfully
This is just a proof of concept with ollama cmd line. As next step, we'd
like to wrap loading / inference customized model logic in the inference
provider implementation.
# What does this PR do?
You are now able to run a training cycle on CPU. This is useful for
debugging and testing purposes.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
On a Mac machine without CUDA devices:
```
17:00:24.417 [START] /v1/post-training/supervised-fine-tune
DEBUG 2025-02-18 12:00:24,419 torchtune.utils._logging:60: Setting manual seed to local seed 3268931494. Local seed is seed + rank = 3268931494 + 0
INFO 2025-02-18 12:00:24,463 torchtune.utils._logging:64: Identified model_type = Llama3_2. Ignoring output.weight in checkpoint in favor of the tok_embedding.weight tied weights.
INFO 2025-02-18 12:00:46,699 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:182: Model is initialized with precision torch.bfloat16.
INFO 2025-02-18 12:00:46,784 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:185: Tokenizer is initialized.
INFO 2025-02-18 12:00:46,786 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:188: Optimizer is initialized.
INFO 2025-02-18 12:00:46,786 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:192: Loss is initialized.
INFO 2025-02-18 12:00:48,997 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:209: Dataset and Sampler are initialized.
INFO 2025-02-18 12:00:48,998 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:227: Learning rate scheduler is initialized.
Writing logs to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/log_1739898049.txt
1|1|Loss: 1.7414989471435547: 100% 1/1 [03:46<00:00, 226.21s/it]INFO 2025-02-18 12:04:35,227 llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device:528: Starting checkpoint save...
INFO 2025-02-18 12:04:49,974 torchtune.utils._logging:121: Model checkpoint of size 6.43 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/consolidated.00.pth
INFO 2025-02-18 12:04:49,981 torchtune.utils._logging:132: Adapter checkpoint of size 0.00 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/adapter/adapter.pth
model_file_path /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0
1|1|Loss: 1.7414989471435547: 100% 1/1 [04:01<00:00, 241.18s/it]
INFO: ::1:64990 - "POST /v1/post-training/supervised-fine-tune HTTP/1.1" 200 OK
17:04:50.364 [END] /v1/post-training/supervised-fine-tune [StatusCode.OK] (265947.01ms)
17:00:24.419 [DEBUG] Setting manual seed to local seed 3268931494. Local seed is seed + rank = 3268931494 + 0
17:00:24.463 [INFO] Identified model_type = Llama3_2. Ignoring output.weight in checkpoint in favor of the tok_embedding.weight tied weights.
17:00:46.700 [INFO] Model is initialized with precision torch.bfloat16.
17:00:46.784 [INFO] Tokenizer is initialized.
17:00:46.786 [INFO] Optimizer is initialized.
17:00:46.786 [INFO] Loss is initialized.
17:00:48.997 [INFO] Dataset and Sampler are initialized.
17:00:48.998 [INFO] Learning rate scheduler is initialized.
17:04:35.227 [INFO] Starting checkpoint save...
17:04:49.974 [INFO] Model checkpoint of size 6.43 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/consolidated.00.pth
17:04:49.981 [INFO] Adapter checkpoint of size 0.00 GB saved to /Users/ihrachys/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0/adapter/adapter.pth
```
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
- **refactor: simplify job status extraction a bit**
- **torchtune: save job status on schedule**
- **refactor: get rid of job_list in torchtune job management code**
# What does this PR do?
A failed job is now registered in API, and one can consult its status.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
```
$ llama-stack-client post_training status --job-uuid test-jobe244b5b0-5053-4892-a4d9-d8fc8b116e73
JobStatusResponse(checkpoints=[], job_uuid='test-jobe244b5b0-5053-4892-a4d9-d8fc8b116e73', status='failed', completed_at=None, resources_allocated=None, scheduled_at=datetime.datetime(2025, 2, 18, 9, 4, 34, 3252), started_at=datetime.datetime(2025, 2, 18, 9, 4, 34, 10688))
```
[//]: # (## Documentation)
---------
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
llama-models should have extremely minimal cruft. Its sole purpose
should be didactic -- show the simplest implementation of the llama
models and document the prompt formats, etc.
This PR is the complement to
https://github.com/meta-llama/llama-models/pull/279
## Test Plan
Ensure all `llama` CLI `model` sub-commands work:
```bash
llama model list
llama model download --model-id ...
llama model prompt-format -m ...
```
Ran tests:
```bash
cd tests/client-sdk
LLAMA_STACK_CONFIG=fireworks pytest -s -v inference/
LLAMA_STACK_CONFIG=fireworks pytest -s -v vector_io/
LLAMA_STACK_CONFIG=fireworks pytest -s -v agents/
```
Create a fresh venv `uv venv && source .venv/bin/activate` and run
`llama stack build --template fireworks --image-type venv` followed by
`llama stack run together --image-type venv` <-- the server runs
Also checked that the OpenAPI generator can run and there is no change
in the generated files as a result.
```bash
cd docs/openapi_generator
sh run_openapi_generator.sh
```
# What does this PR do?
- Configured ruff linter to automatically fix import sorting issues.
- Set --exit-non-zero-on-fix to ensure non-zero exit code when fixes are
applied.
- Enabled the 'I' selection to focus on import-related linting rules.
- Ran the linter, and formatted all codebase imports accordingly.
- Removed the black dep from the "dev" group since we use ruff
Signed-off-by: Sébastien Han <seb@redhat.com>
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
[Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.*]
[//]: # (## Documentation)
[//]: # (- [ ] Added a Changelog entry if the change is significant)
Signed-off-by: Sébastien Han <seb@redhat.com>
Lint check in main branch is failing. This fixes the lint check after we
moved to ruff in https://github.com/meta-llama/llama-stack/pull/921. We
need to move to a `ruff.toml` file as well as fixing and ignoring some
additional checks.
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
# What does this PR do?
This PR changes our API to follow more idiomatic REST API approaches of
having paths being resources and methods indicating the action being
performed.
Changes made to generator:
1) removed the prefix check of "get" as its not required and is actually
needed for other method types too
2) removed _ check on path since variables can have "_"
## Test Plan
LLAMA_STACK_BASE_URL=http://localhost:5000 pytest -v
tests/client-sdk/agents/test_agents.py
## context
Currently, the GPU memory will be continuously occupied after the
training finishes. In this PR, we explicitly delete the reference and
clean up the memory after training finishes.
## test
Before the change, after training a llama 3.2 3B model, >6GB GPU memory
is still occupied
After the change, after training a llama 3.2 3B model, the GPU memory
drops to ~1GB
<img width="156" alt="Screenshot 2025-01-14 at 6 05 17 PM"
src="https://github.com/user-attachments/assets/45d212b1-a651-49f3-aad9-1c0a27fcebcf"
/>
## context
In this PR, we defined 2 llama stack dataset formats (instruct, dialog)
- For instruct dataset format, the column schema will be
[chat_completion_input, expected_answer], which is consistent with the
eval data format. This dataset format is the abstract of single turn QA
style post training data
- For dialog dataset format, the column schema will be [dialog], which
is a list of user messages and assistant messages that interleave
together. During training, the whole list will be the model input and
the loss is calculated on assistant messages only. This dataset format
is the abstract of multi turn chat style post training data
## changes
- defined the 2 llama stack dataset formats
- an adapter to convert llama stack dataset format to torchtune dataset
format
- move dataset format validation to post training level instead of
torchtune level since it's not specific to torchtune
- add localfs as datasetio provider
## test
instruct format
- use https://huggingface.co/datasets/llamastack/evals as dataset and
the training works as expected
<img width="1443" alt="Screenshot 2025-01-09 at 5 15 14 PM"
src="https://github.com/user-attachments/assets/2c37a936-c67a-4726-90e0-23fa0ba7000f"
/>
- use my generated local dataset and the training works as expected
<img width="1617" alt="Screenshot 2025-01-09 at 5 19 11 PM"
src="https://github.com/user-attachments/assets/0bdccbbf-bac2-472a-a365-15213e49bbfa"
/>
dialog format
- use my generated local dataset and the training works as expected
<img width="1588" alt="Screenshot 2025-01-09 at 5 23 16 PM"
src="https://github.com/user-attachments/assets/893915ba-41a3-4d51-948b-e872060ecede"
/>
## What does this PR do?
- Change to support llama3.1 8B instruct model other than llama3 8B
model as llama3.1 8B instruct model is a better model to finetune on top
of
- Make the copy files logic in checkpointer safer in case the file be
copied doesn't exist in source path
## test
issue a post training request from client and verify training works as
expect
<img width="1101" alt="Screenshot 2025-01-02 at 12 18 45 PM"
src="https://github.com/user-attachments/assets/47cc4df9-3edc-4afd-b5dd-abe1f039f1ed"
/>
<img width="782" alt="Screenshot 2025-01-02 at 12 18 52 PM"
src="https://github.com/user-attachments/assets/b9435274-ef1d-4570-bd8e-0880c3a4b2e9"
/>
## what does this PR do?
The current code hardcode the validation steps to run (forgot to change
it after testing). in this PR, we make it configurable by training
config
## test
On client side, issue a post training request with 20 validation steps,
server side logging shows that it runs 20 validation steps successfully
<img width="1128" alt="Screenshot 2025-01-02 at 8 21 06 PM"
src="https://github.com/user-attachments/assets/7a757516-c6ba-41d4-85c5-361a80ecf46e"
/>
### Context
In this PR, we
- Implement the post training job management and get training artifacts
apis
- get_training_jobs
- get_training_job_status
- get_training_job_artifacts
- get_training_job_logstream is deleted since the trace can be directly
accessed by UI with Jaeger
https://llama-stack.readthedocs.io/en/latest/building_applications/telemetry.html#jaeger-to-visualize-traces
- Refactor the post training and training types definition to make them
more intuitive.
- Rewrite the checkpointer to make it compatible with llama-stack file
system and can be recognized during inference
### Test
Unit test
`pytest llama_stack/providers/tests/post_training/test_post_training.py
-m "torchtune_post_training_huggingface_datasetio" -v -s --tb=short
--disable-warnings`
<img width="1506" alt="Screenshot 2024-12-10 at 4 06 17 PM"
src="https://github.com/user-attachments/assets/16225029-bdb7-48c4-9d13-e580cc769c0a">
e2e test with client side call
<img width="888" alt="Screenshot 2024-12-10 at 4 09 44 PM"
src="https://github.com/user-attachments/assets/de375e4c-ef67-4dcc-a045-4037d9489191">
### Context
This is the 1st of series PRs that integrate torchtune with llama-stack
as meta reference post-training implementation. For MVP, we will focus
on single device LoRA SFT.
Though this PR is still WIP, we want to get early feedback on the high
level design of this skeleton while still working on several details
### Scope
To limit the scope of this PR, we focus on the skeleton of the
implementation.
**What are included?**
- refine the post-training SFT apis
- skeleton of supervised_fine_tune implementation. We verified that we
can call the supervised_fine_tune API successfully from llama stack
client SDK (client side PR:
https://github.com/meta-llama/llama-stack-client-python/pull/51)
- a very basic single device LoRA training recipe based on torchtune
core components
- parity check with torchtune library and post training api unit test
**What are not includes?**
- implementation of other job management, get training artifacts apis
(separate PR)
- refactor the meta reference inference logic to support eval on
finetuned model (separate PR)
- several necessary functionality in the training recipe such as
logging, validation etc (separate PR)
- interop with telemetry for tracing and metrics logging, currently
temporarily log to local disk (separate PR)
### Testing
**e2e test**
Although we haven't added detailed testing and numerical parity check
with torchtune yet, we did a simple E2E test from client to server
1. setup server with` llama stack build --template
experimental-post-training --image-type conda` and `llama stack run
experimental-post-training `
2. On client, run `llama-stack-client --endpoint
http://devgpu018.nha2.facebook.com:5000 post_training
supervised_fine_tune`
3. Training finishes successfully. On server side, get the finetune
checkpoints under output dir. On client side, get the job uuid
server
<img width="1110" alt="Screenshot 2024-12-02 at 5 52 32 PM"
src="https://github.com/user-attachments/assets/b548eb90-7a9b-4edc-a858-ee237cc4361d">
client
<img width="807" alt="Screenshot 2024-12-02 at 5 52 37 PM"
src="https://github.com/user-attachments/assets/1138ffa8-4698-40fa-b190-3d7b99646838">
**parity check**
torchtune dataloader output and llama-stack post training dataloader
output are same
<img width="1116" alt="Screenshot 2024-12-04 at 8 18 46 PM"
src="https://github.com/user-attachments/assets/5e295cdc-4c24-4ea6-82c0-ca96ef1bd6ee">
torchtune LoRA SFT and llama-stack post training LoRA SFT on alpaca
dataset with llama3.2 3B instruct model are numerical match
<img width="860" alt="Screenshot 2024-12-04 at 8 17 01 PM"
src="https://github.com/user-attachments/assets/c05cf0a8-c674-4d2e-9f0a-c5d01b2dca99">
<img width="1049" alt="Screenshot 2024-12-04 at 8 17 06 PM"
src="https://github.com/user-attachments/assets/b911d4e2-e7b1-41a9-b62c-d75529b6d443">
**unit test **
![Uploading Screenshot 2024-12-09 at 1.35.10 PM.png…]()
