Merge pull request #27 from FederatedAI/dev-2.1.1

Dev 2.1.1

RELEASE.md

@@ -1,3 +1,8 @@
+## Release 2.1.1
+### Major Features and Improvments
+> Fate-Test: FATE Automated Testing Tool
+* Add new subcommand `llmsuite` for FATE-LLM training and evaluation
+
 ## Release 2.1.0
 ### Major Features and Improvements
 > Fate-Test: FATE Automated Testing Tool

doc/fate_test.md

@@ -9,7 +9,7 @@ A collection of useful tools to running FATE tests and PipeLine tasks.
     ```bash
     pip install -e python/fate_test
     ```
-2. edit default fate\_test\_config.yaml
+2. edit default fate\_test\_config.yaml; edit path to fate base/data base accordingly
 
    ```bash
    # edit priority config file with system default editor
@@ -88,4 +88,16 @@ shown in last step
   ```bash
   fate_test data generate -i <path contains *performance.yaml> -ng 10000 -fg 10 -fh 10 -m 1.0 --upload-data
   fate_test performance -i <path contains *performance.yaml> --skip-data
-  ```
+  ```
+
+- [llm-suite](./fate_test_command.md#llmsuite): used for running FATE-Llm testsuites, collection of FATE-Llm jobs and/or evaluations
+
+  Before running llmsuite for the first time, make sure to install FATE-Llm and allow its import in FATE-Test scripts:
+
+  ```bash
+  fate_test config include fate-llm
+  ```
+
+  ```bash
+  fate_test llmsuite -i <path contains *llmsuite.yaml>
+  ```

doc/fate_test_command.md

@@ -867,3 +867,155 @@ fate_test data --help
     data after generate and upload dataset in testsuites
     *path1*
 
+
+## Llmsuite
+
+Llmsuite is used for running a collection of FATE-Llm jobs in sequence and then evaluate them on user-specified tasks.
+It also allows users to compare the results of different llm jobs.
+
+### command options
+
+```bash
+fate_test llmsuite --help
+```
+
+1. include:
+
+   ```bash
+   fate_test llmsuite -i <path1 contains *llmsuite.yaml>
+   ```
+
+   will run llm testsuites in
+   *path1*
+
+2. exclude:
+
+   ```bash
+   fate_test llmsuite -i <path1 contains *llmsuite.yaml> -e <path2 to exclude> -e <path3 to exclude> ...
+   ```
+
+   will run llm testsuites in *path1* but not in *path2* and *path3*
+
+3. glob:
+
+   ```bash
+   fate_test llmsuite -i <path1 contains *llmsuite.yaml> -g "hetero*"
+   ```
+
+   will run llm testsuites in sub directory start with *hetero* of
+   *path1*
+
+4. algorithm-suite:
+
+   ```bash
+   fate_test llmsuite -a "pellm"
+   ```
+
+   will run built-in 'pellm' llm testsuite, which will train and evaluate a FATE-Llm model and a zero-shot model
+
+5. timeout:
+
+   ```bash
+   fate_test llmsuite -i <path1 contains *llmsuite.yaml> -m 3600
+   ```
+
+   will run llm testsuites in *path1* and timeout when job does not finish
+   within 3600s; if tasks need more time, use a larger threshold
+
+6. task-cores
+
+   ```bash
+   fate_test llmsuite -i <path1 contains *llmsuite.yaml> -p 4
+   ```
+
+   will run llm testsuites in *path1* with script config "task-cores" set to 4
+
+7. eval-config:
+
+    ```bash
+    fate_test llmsuite -i <path1 contains *llmsuite.yaml> --eval-config <path2>
+    ```
+
+   will run llm testsuites in *path1* with evaluation configuration set to *path2*
+
+8. skip-evaluate:
+
+    ```bash
+    fate_test llmsuite -i <path1 contains *llmsuite.yaml> --skip-evaluate
+    ```
+
+    will run llm testsuites in *path1* without running evaluation
+
+9. provider:
+
+    ```bash
+    fate_test llmsuite -i <path1 contains *llmsuite.yaml> --provider <provider_name>
+    ```
+
+    will run llm testsuites in *path1* with FATE provider set to *provider_name*
+
+10. yes:
+
+    ```bash
+    fate_test llmsuite -i <path1 contains *llmsuite.yaml> --yes
+    ```
+
+    will run llm testsuites in *path1* directly, skipping double check
+
+
+### FATE-Llm job configuration
+
+Configuration of jobs should be specified in a llm testsuite whose
+file name ends with "\*llmsuite.yaml". For llm testsuite example,
+please refer [here](https://github.com/FederatedAI/FATE-LLM).
+
+A FATE-Llm testsuite includes the following elements:
+
+- job group: each group includes arbitrary number of jobs with paths
+  to corresponding script and configuration
+
+    - job: name of evaluation job to be run, must be unique within each group
+      list
+
+        - script: path to [testing script](#testing-script), should be
+          relative to testsuite, optional for evaluation-only jobs;
+          note that pretrained model, if available, should be returned at the end of the script
+        - conf: path to job configuration file for script, should be
+          relative to testsuite, optional for evaluation-only jobs
+        - pretrained: path to pretrained model, should be either model name from Huggingface or relative path to
+          testsuite, optional for jobs needed to run FATE-Llm training job, where the 
+          script should return path to the pretrained model
+        - peft: path to peft file, should be relative to testsuite, 
+          optional for jobs needed to run FATE-Llm training job
+        - tasks: list of tasks to be evaluated, optional for jobs skipping evaluation
+        - include_path: should be specified if tasks are user-defined
+        - eval_conf: path to evaluation configuration file, should be
+          relative to testsuite; if not provided, will use default conf
+
+      ```yaml
+          bloom_lora:
+            pretrained: "models/bloom-560m"
+            script: "./test_bloom_lora.py"
+            conf: "./bloom_lora_config.yaml"
+            peft_path_format: "{{fate_base}}/fate_flow/model/{{job_id}}/guest/{{party_id}}/{{model_task_name}}/0/output/output_model/model_directory"
+            tasks:
+              - "dolly-15k"
+
+      ```
+
+- llm suite
+
+  ```yaml
+     hetero_nn_sshe_binary_0:
+      bloom_lora: 
+        pretrained: "bloom-560m"
+        script: "./test_bloom_lora.py"
+        conf: "./bloom_lora_config.yaml"
+        peft_path_format: "{{fate_base}}/fate_flow/model/{{job_id}}/guest/{{party_id}}/{{model_task_name}}/0/output/output_model/model_directory"
+        tasks:
+          - "dolly-15k"
+      bloom_zero_shot:
+        pretrained: "bloom-560m"
+        tasks:
+          - "dolly-15k"
+  ```

python/fate_test/_config.py

@@ -36,20 +36,20 @@
 # st_config_directory: examples/flow_test_template/hetero_lr/flow_test_config.yaml
 
 # directory stores testsuite file with min_test data sets to upload,
-# default location={FATE}/examples/data/upload_config/min_test_data_testsuite.json
-min_test_data_config: examples/data/upload_config/min_test_data_testsuite.json
+# default location={FATE}/examples/data/upload_config/min_test_data_testsuite.yaml
+min_test_data_config: examples/data/upload_config/min_test_data_testsuite.yaml
 # directory stores testsuite file with all example data sets to upload,
-# default location={FATE}/examples/data/upload_config/all_examples_data_testsuite.json
-all_examples_data_config: examples/data/upload_config/all_examples_data_testsuite.json
+# default location={FATE}/examples/data/upload_config/all_examples_data_testsuite.yaml
+all_examples_data_config: examples/data/upload_config/all_examples_data_testsuite.yaml
 
 # directory where FATE code locates, default installation location={FATE}/fate
 # python/ml -> $fate_base/python/ml
-fate_base: path(FATE)/fate
+fate_base: path(FATE)/
 
 # whether to delete data in suites after all jobs done
 clean_data: true
 
-# participating parties' id and correponding flow service ip & port information
+# participating parties' id and corresponding flow service ip & port information
 parties:
   guest: ['9999']
   host: ['10000', '9999']

python/fate_test/_flow_client.py

@@ -41,6 +41,29 @@ def __init__(self,
     def set_address(self, address):
         self.address = address
 
+    def bind_table(self, data: Data, callback=None):
+        conf = data.config
+        conf['file'] = os.path.join(str(self._data_base_dir), conf.get('file'))
+        path = Path(conf.get('file'))
+        if not path.exists():
+            raise Exception('The file is obtained from the fate flow client machine, but it does not exist, '
+                            f'please check the path: {path}')
+        response = self._client.table.bind_path(path=str(path),
+                                                namespace=data.namespace,
+                                                name=data.table_name)
+        try:
+            if callback is not None:
+                callback(response)
+                status = str(response['message']).lower()
+            else:
+                status = response["message"]
+            code = response["code"]
+            if code != 0:
+                raise RuntimeError(f"Return code {code} != 0, bind path failed")
+        except BaseException:
+            raise ValueError(f"Bind path failed, response={response}")
+        return status
+
     def transform_local_file_to_dataframe(self, data: Data, callback=None, output_path=None):
         #data_warehouse = self.upload_data(data, callback, output_path)
         #status = self.transform_to_dataframe(data.namespace, data.table_name, data_warehouse, callback)
@@ -82,44 +105,6 @@ def upload_file_and_convert_to_dataframe(self, data: Data, callback=None, output
         self._awaiting(job_id, "local", 0)
         return status
 
-    """def upload_data(self, data: Data, callback=None, output_path=None):
-        response, file_path = self._upload_data(data, output_path=output_path)
-        try:
-            if callback is not None:
-                callback(response)
-            code = response["code"]
-            if code != 0:
-                raise ValueError(f"Return code {code}!=0")
-
-            namespace = response["data"]["namespace"]
-            name = response["data"]["name"]
-            job_id = response["job_id"]
-        except BaseException:
-            raise ValueError(f"Upload data fails, response={response}")
-        # self.monitor_status(job_id, role=self.role, party_id=self.party_id)
-        self._awaiting(job_id, "local", 0)
-
-        return dict(namespace=namespace, name=name)
-
-    def transform_to_dataframe(self, namespace, table_name, data_warehouse, callback=None):
-        response = self._client.data.dataframe_transformer(namespace=namespace,
-                                                           name=table_name,
-                                                           data_warehouse=data_warehouse)
-
-        try:
-            if callback is not None:
-                callback(response)
-                status = self._awaiting(response["job_id"], "local", 0)
-                status = str(status).lower()
-            else:
-                status = response["retmsg"]
-
-        except Exception as e:
-            raise RuntimeError(f"upload data failed") from e
-        job_id = response["job_id"]
-        self._awaiting(job_id, "local", 0)
-        return status"""
-
     def delete_data(self, data: Data):
         try:
             table_name = data.config['table_name'] if data.config.get(
@@ -154,27 +139,6 @@ def _awaiting(self, job_id, role, party_id, callback=None):
                 callback(response)
             time.sleep(1)
 
-    """def _upload_data(self, data, output_path=None, verbose=0, destroy=1):
-        conf = data.config
-        # if conf.get("engine", {}) != "PATH":
-        if output_path is not None:
-            conf['file'] = os.path.join(os.path.abspath(output_path), os.path.basename(conf.get('file')))
-        else:
-            if _config.data_switch is not None:
-                conf['file'] = os.path.join(str(self._cache_directory), os.path.basename(conf.get('file')))
-            else:
-                conf['file'] = os.path.join(str(self._data_base_dir), conf.get('file'))
-        path = Path(conf.get('file'))
-        if not path.exists():
-            raise Exception('The file is obtained from the fate flow client machine, but it does not exist, '
-                            f'please check the path: {path}')
-        response = self._client.data.upload(file=str(path),
-                                            head=data.head,
-                                            meta=data.meta,
-                                            extend_sid=data.extend_sid,
-                                            partitions=data.partitions)
-        return response, conf["file"]"""
-
     def _output_data_table(self, job_id, role, party_id, task_name):
         response = self._client.output.data_table(job_id, role=role, party_id=party_id, task_name=task_name)
         if response.get("code") is not None:
@@ -223,7 +187,7 @@ def get_version(self):
     """def _add_notes(self, job_id, role, party_id, notes):
         data = dict(job_id=job_id, role=role, party_id=party_id, notes=notes)
         response = AddNotesResponse(self._post(url='job/update', json=data))
-        return response"""
+        return response
 
     def _table_bind(self, data):
         response = self._post(url='table/bind', json=data)
@@ -235,6 +199,7 @@ def _table_bind(self, data):
         except Exception as e:
             raise RuntimeError(f"table bind error: {response}") from e
         return response
+    """
 
 
 class Status(object):

python/fate_test/_io.py

@@ -32,6 +32,10 @@ def echo(cls, message, **kwargs):
         click.secho(message, **kwargs)
         click.secho(message, file=cls._file, **kwargs)
 
+    @classmethod
+    def sep_line(cls):
+        click.secho("-------------------------------------------------")
+
     @classmethod
     def file(cls, message, **kwargs):
         click.secho(message, file=cls._file, **kwargs)