Merge pull request #6 from Trainy-ai/patchfix

Patchfix

Author: asaiacai

docs/source/quickstart/finetuning.rst

@@ -14,7 +14,7 @@ To do a vicuna finetune of your first model through LLM-ATC, run the following
 .. code-block:: console
 
     # start training
-    $ llm-atc train --model_type vicuna --finetune_data ./vicuna_test.json --name myvicuna --description "This is a finetuned model that just says its name is vicuna" -c mycluster --cloud gcp --envs "MODEL_SIZE=7 WANDB_API_KEY=<my wandb key>" --accelerator A100-80G:4
+    $ llm-atc train --model_type vicuna --finetune_data ./vicuna_test.json --name myvicuna --checkpoint_bucket my-trainy-bucket --checkpoint_path ~/test_vicuna --checkpoint_store S3 --description "This is a finetuned model that just says its name is vicuna" -c mycluster --cloud gcp --envs "MODEL_SIZE=7 WANDB_API_KEY=<my wandb key>" --accelerator A100-80G:4
 
     # Once training is done, shutdown the cluster
     $ sky down

docs/source/quickstart/serving.rst

@@ -13,7 +13,7 @@ by using the :code:`llm-atc/` prefix.
 .. code-block:: console
 
     # serve an llm-atc finetuned model, requires `llm-atc/` prefix and grabs model checkpoint from object store
-    $ llm-atc serve --name llm-atc/myvicuna --accelerator A100:1 -c servecluster --cloud gcp --region asia-southeast1 --envs "HF_TOKEN=<HuggingFace_token>"
+    $ llm-atc serve --name llm-atc/myvicuna --source s3://my-bucket/my_vicuna/ --accelerator A100:1 -c servecluster --cloud gcp --region asia-southeast1 --envs "HF_TOKEN=<HuggingFace_token>"
 
     # serve a HuggingFace model, e.g. `lmsys/vicuna-13b-v1.3`
     $ llm-atc serve --name lmsys/vicuna-13b-v1.3 --accelerator A100:1 -c servecluster --cloud gcp --region asia-southeast1 --envs "HF_TOKEN=<HuggingFace_token>"

llm_atc/cli.py

@@ -58,6 +58,21 @@ def cli():
     required=True,
     help="local/cloud URI to finetuning data. (e.g ~/mychat.json, s3://my_bucket/my_chat.json)",
 )
+@click.option(
+    "--checkpoint_bucket", type=str, required=True, help="object store bucket name"
+)
+@click.option(
+    "--checkpoint_path",
+    type=str,
+    required=True,
+    help="object store path for fine tuned checkpoints, e.g. ~/datasets",
+)
+@click.option(
+    "--checkpoint_store",
+    type=str,
+    required=True,
+    help="object store type ['S3', 'GCS', 'AZURE', 'R2', 'IBM']",
+)
 @click.option("-n", "--name", type=str, help="Name of this model run.", required=True)
 @click.option(
     "--description", type=str, default="", help="description of this model run"
@@ -100,6 +115,9 @@ def cli():
 def train(
     model_type: str,
     finetune_data: str,
+    checkpoint_bucket: str,
+    checkpoint_path: str,
+    checkpoint_store: Optional[str],
     name: str,
     description: str,
     cluster: Optional[str],
@@ -118,12 +136,11 @@ def train(
             event="training launched",
             timestamp=datetime.utcnow(),
         )
-    if RunTracker.run_exists(name):
-        raise ValueError(
-            f"Task with name {name} already exists in {llm_atc.constants.LLM_ATC_PATH}. Try again with a different name"
-        )
     task = train_task(
         model_type,
+        checkpoint_bucket=checkpoint_bucket,
+        checkpoint_path=checkpoint_path,
+        checkpoint_store=checkpoint_store,
         finetune_data=finetune_data,
         name=name,
         cloud=cloud,
@@ -146,7 +163,11 @@ def train(
     "--name",
     help="name of model to serve",
     required=True,
-    multiple=True,
+)
+@click.option(
+    "--source",
+    help="object store path for llm-atc finetuned model checkpoints."
+    "e.g. s3://<bucket-name>/<path>/<to>/<checkpoints>",
 )
 @click.option(
     "-e",
@@ -189,7 +210,8 @@ def train(
     help="Don't connect to this session",
 )
 def serve(
-    name: List[str],
+    name: str,
+    source: Optional[str],
     accelerator: Optional[str],
     envs: Optional[str],
     cluster: Optional[str],
@@ -209,6 +231,7 @@ def serve(
         )
     serve_task = serve_route(
         name,
+        source=source,
         accelerator=accelerator,
         envs=envs,
         cloud=cloud,

llm_atc/config/serve/serve.yml

@@ -4,11 +4,6 @@ resources:
   ports:
     - 8000
 
-file_mounts:
-  /llm-atc:
-    name: llm-atc # Make sure it is unique or you own this bucket name
-    mode: MOUNT  # MOUNT or COPY. Defaults to MOUNT if not specified
-
 setup: |
   conda activate chatbot
   if [ $? -ne 0 ]; then
@@ -22,7 +17,7 @@ setup: |
   conda install -y -c conda-forge accelerate
   pip install sentencepiece
   pip install vllm
-  pip install git+https://github.com/lm-sys/FastChat.git
+  pip install git+https://github.com/lm-sys/FastChat.git@v0.2.28
   pip install --upgrade openai
   if [[ "$HF_TOKEN" != "" ]];
   then
@@ -49,7 +44,6 @@ run: |
   master_addr=`echo "$SKYPILOT_NODE_IPS" | head -n1`
   let x='SKYPILOT_NODE_RANK + 1'
   this_addr=`echo "$SKYPILOT_NODE_IPS" | sed -n "${x}p"`
-  MODEL_NAME=`echo "$MODELS_LIST" | sed -n "${x}p"`
 
   echo "The ip address of this machine is ${this_addr}"
   echo "The head address is ${master_addr}"
@@ -82,5 +76,5 @@ run: |
   
 
 envs:
-  MODELS_LIST: lmsys/vicuna-7b-v1.3
+  MODEL_NAME: lmsys/vicuna-7b-v1.3
   HF_TOKEN: ""

llm_atc/config/train/vicuna.yml

@@ -7,10 +7,9 @@ num_nodes: 1
 
 file_mounts:
   /artifacts:
-    name: llm-atc
+    name: ${MY_BUCKET} # Name of the bucket.
     mode: MOUNT
-
-workdir: .
+    store: ${BUCKET_TYPE} # s3, gcs, r2, ibm
 
 setup: |
   # Setup the environment
@@ -64,7 +63,7 @@ run: |
   # the training for saving checkpoints.
   mkdir -p ~/.checkpoints
   LOCAL_CKPT_PATH=~/.checkpoints
-  CKPT_PATH=/artifacts/${MODEL_NAME}
+  CKPT_PATH=/artifacts/${BUCKET_PATH}/${MODEL_NAME}
   mkdir -p $CKPT_PATH
   last_ckpt=$(ls ${CKPT_PATH} | grep -E '[0-9]+' | sort -t'-' -k1,1 -k2,2n | tail -1)
   mkdir -p ~/.checkpoints/${last_ckpt}
@@ -127,3 +126,6 @@ envs:
   WANDB_API_KEY: ""
   MODEL_NAME: "vicuna_test"
   HF_TOKEN: ""
+  MY_BUCKET: "llm-atc"
+  BUCKET_PATH: "my_vicuna" # object store path.
+  BUCKET_TYPE: "S3"

llm_atc/launch.py

@@ -2,6 +2,7 @@
 import logging
 import os
 import sky
+from sky.data.storage import Storage
 
 from omegaconf import OmegaConf
 from typing import Any, Dict, Optional
@@ -10,7 +11,7 @@
 SUPPORTED_MODELS = ("vicuna",)
 
 
-def train_task(model_type: str, **launcher_kwargs) -> sky.Task:
+def train_task(model_type: str, *args, **launcher_kwargs) -> sky.Task:
     """
     Dispatch train launch to corresponding task default config
 
@@ -39,6 +40,9 @@ class Launcher:
     def __init__(
         self,
         finetune_data: str,
+        checkpoint_bucket: str = "llm-atc",
+        checkpoint_path: str = "my_vicuna",
+        checkpoint_store: str = "S3",
         name: Optional[str] = None,
         cloud: Optional[str] = None,
         region: Optional[str] = None,
@@ -47,6 +51,9 @@ def __init__(
         envs: Optional[str] = "",
     ):
         self.finetune_data: str = finetune_data
+        self.checkpoint_bucket: str = checkpoint_bucket
+        self.checkpoint_path: str = checkpoint_path
+        self.checkpoint_store: str = checkpoint_store
         self.name: Optional[str] = name
         self.cloud: Optional[str] = cloud
         self.region: Optional[str] = region
@@ -85,8 +92,14 @@ def launch(self) -> sky.Task:
             logging.warning(
                 "No huggingface token provided. You will not be able to finetune starting from private or gated models"
             )
+        self.envs["MY_BUCKET"] = self.checkpoint_bucket
+        self.envs["BUCKET_PATH"] = self.checkpoint_path
+        self.envs["BUCKET_TYPE"] = self.checkpoint_store
         task.update_envs(self.envs)
         task.update_file_mounts({"/data/mydata.json": self.finetune_data})
+        storage = Storage(name=self.checkpoint_bucket)
+        storage.add_store(self.checkpoint_store)
+        task.update_storage_mounts({"/artifacts": storage})
         resource = list(task.get_resources())[0]
         resource._set_accelerators(self.accelerator, None)
         resource._cloud = sky.clouds.CLOUD_REGISTRY.from_str(self.cloud)

llm_atc/serve.py

@@ -8,7 +8,7 @@
 from typing import Any, Dict, List, Optional
 
 
-def serve_route(model_names: List[str], **serve_kwargs):
+def serve_route(model_name: str, source: Optional[str] = None, **serve_kwargs):
     """Routes model serve requests to the corresponding model serve config
 
     Args:
@@ -17,26 +17,30 @@ def serve_route(model_names: List[str], **serve_kwargs):
     Raises:
         ValueError: requested non-existent model from llm-atc
     """
-    model_names = list(model_names)
-    for i, name in enumerate(model_names):
-        if name.startswith("llm-atc/") and not RunTracker.run_exists(
-            name.split("/")[-1]
-        ):
-            raise ValueError(f"model = {name} does not exist within llm-atc.")
-    return Serve(model_names, **serve_kwargs).serve()
+    if model_name.startswith("llm-atc/") and source is None:
+        raise ValueError(
+            "Attempting to use a finetuned model without a corresponding object store location"
+        )
+    elif not source is None and not model_name.startswith("llm-atc/"):
+        logging.warning(
+            "Specified object store mount but model is not an llm-atc model. Skipping mounting."
+        )
+    return Serve(model_name, source, **serve_kwargs).serve()
 
 
 class Serve:
     def __init__(
         self,
-        names: List[str],
+        names: str,
+        source: Optional[str],
         accelerator: Optional[str] = None,
         cloud: Optional[str] = None,
         region: Optional[str] = None,
         zone: Optional[str] = None,
         envs: str = "",
     ):
         self.names = names
+        self.source = source
         self.num_models = len(names)
         self.accelerator = accelerator
         self.envs: Dict[Any, Any] = (
@@ -65,7 +69,7 @@ def default_serve_task(self) -> sky.Task:
     def serve(self) -> sky.Task:
         """Deploy fastchat.serve.openai_api_server with vllm_worker"""
         serve_task = self.default_serve_task
-        self.envs["MODELS_LIST"] = "\n".join(self.names)
+        self.envs["MODEL_NAME"] = self.names
         if "HF_TOKEN" not in self.envs:
             logging.warning(
                 "No huggingface token provided. You will not be able to access private or gated models"
@@ -76,5 +80,6 @@ def serve(self) -> sky.Task:
         resource._cloud = sky.clouds.CLOUD_REGISTRY.from_str(self.cloud)
         resource._set_region_zone(self.region, self.zone)
         serve_task.set_resources(resource)
-        serve_task.num_noded = self.num_models
+        if self.source and self.names.startswith("llm-atc/"):
+            serve_task.update_file_mounts({"/" + self.names: self.source})
         return serve_task

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "llm_atc"
-version = "0.1.3"
+version = "0.1.4"
 description = "Tools for fine tuning and serving LLMs"
 authors = ["Andrew Aikawa <[email protected]>"]
 readme = "README.md"

tests/test_cli.py

@@ -121,7 +121,7 @@ def test_train_vicuna():
     test = Test(
         "train_vicuna",
         [
-            f"llm-atc train --model_type vicuna --finetune_data {test_chat} --name {name} --description 'test case vicuna fine tune' -c mycluster --cloud gcp --envs 'MODEL_SIZE=7' --accelerator A100-80G:4",
+            f"llm-atc train --checkpoint_bucket llm-atc --checkpoint_path ~/test_vicuna --checkpoint_store S3 --model_type vicuna --finetune_data {test_chat} --name {name} --description 'test case vicuna fine tune' -c mycluster --cloud gcp --envs 'MODEL_SIZE=7' --accelerator A100-80G:4",
         ],
         f"sky down --purge -y {name}",
         timeout=10 * 60,

tests/test_launch.py

@@ -6,6 +6,9 @@
 def test_train():
     task = train_task(
         "vicuna",
+        checkpoint_bucket="llm-atc",
+        checkpoint_path="myvicuna",
+        checkpoint_store="S3",
         finetune_data="./vicuna_test.json",
         name="myvicuna",
         cloud="aws",