together 1.5.6__tar.gz → 1.5.8__tar.gz

{together-1.5.6 → together-1.5.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.6
+Version: 1.5.8
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI

{together-1.5.6 → together-1.5.8}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.5.6"
+version = "1.5.8"
 authors = ["Together AI <support@together.ai>"]
 description = "Python client for Together's Cloud Platform!"
 readme = "README.md"

{together-1.5.6 → together-1.5.8}/src/together/cli/api/finetune.py

@@ -258,10 +258,13 @@ def create(
             raise click.BadParameter(
                 f"LoRA fine-tuning is not supported for the model `{model}`"
             )
-
+        if training_method == "dpo":
+            default_batch_size = model_limits.lora_training.max_batch_size_dpo
+        else:
+            default_batch_size = model_limits.lora_training.max_batch_size
         default_values = {
             "lora_r": model_limits.lora_training.max_rank,
-            "batch_size": model_limits.lora_training.max_batch_size,
+            "batch_size": default_batch_size,
             "learning_rate": 1e-3,
         }
 
@@ -288,7 +291,12 @@ def create(
 
         batch_size_source = ctx.get_parameter_source("batch_size")  # type: ignore[attr-defined]
         if batch_size_source == ParameterSource.DEFAULT:
-            training_args["batch_size"] = model_limits.full_training.max_batch_size
+            if training_method == "dpo":
+                training_args["batch_size"] = (
+                    model_limits.full_training.max_batch_size_dpo
+                )
+            else:
+                training_args["batch_size"] = model_limits.full_training.max_batch_size
 
     if n_evals <= 0 and validation_file:
         log_warn(

{together-1.5.6 → together-1.5.8}/src/together/resources/finetune.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import re
 from pathlib import Path
-from typing import List, Literal
+from typing import Dict, List, Literal
 
 from rich import print as rprint
 
@@ -30,16 +30,8 @@ from together.types import (
     TrainingMethodSFT,
     TrainingType,
 )
-from together.types.finetune import (
-    DownloadCheckpointType,
-    FinetuneEvent,
-    FinetuneEventType,
-)
-from together.utils import (
-    get_event_step,
-    log_warn_once,
-    normalize_key,
-)
+from together.types.finetune import DownloadCheckpointType
+from together.utils import log_warn_once, normalize_key
 
 
 _FT_JOB_WITH_STEP_REGEX = r"^ft-[\dabcdef-]+:\d+$"
@@ -102,6 +94,7 @@ def create_finetune_request(
 
     training_type: TrainingType = FullTrainingType()
     max_batch_size: int = 0
+    max_batch_size_dpo: int = 0
     min_batch_size: int = 0
     if lora:
         if model_limits.lora_training is None:
@@ -119,7 +112,7 @@ def create_finetune_request(
 
         max_batch_size = model_limits.lora_training.max_batch_size
         min_batch_size = model_limits.lora_training.min_batch_size
-
+        max_batch_size_dpo = model_limits.lora_training.max_batch_size_dpo
     else:
         if model_limits.full_training is None:
             raise ValueError(
@@ -128,13 +121,24 @@ def create_finetune_request(
 
         max_batch_size = model_limits.full_training.max_batch_size
         min_batch_size = model_limits.full_training.min_batch_size
+        max_batch_size_dpo = model_limits.full_training.max_batch_size_dpo
 
-    batch_size = batch_size if batch_size != "max" else max_batch_size
+    if batch_size == "max":
+        if training_method == "dpo":
+            batch_size = max_batch_size_dpo
+        else:
+            batch_size = max_batch_size
 
-    if batch_size > max_batch_size:
-        raise ValueError(
-            f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
-        )
+    if training_method == "sft":
+        if batch_size > max_batch_size:
+            raise ValueError(
+                f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
+            )
+    elif training_method == "dpo":
+        if batch_size > max_batch_size_dpo:
+            raise ValueError(
+                f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size_dpo}."
+            )
 
     if batch_size < min_batch_size:
         raise ValueError(
@@ -210,68 +214,38 @@ def create_finetune_request(
     return finetune_request
 
 
-def _process_checkpoints_from_events(
-    events: List[FinetuneEvent], id: str
+def _parse_raw_checkpoints(
+    checkpoints: List[Dict[str, str]], id: str
 ) -> List[FinetuneCheckpoint]:
     """
-    Helper function to process events and create checkpoint list.
+    Helper function to process raw checkpoints and create checkpoint list.
 
     Args:
-        events (List[FinetuneEvent]): List of fine-tune events to process
+        checkpoints (List[Dict[str, str]]): List of raw checkpoints metadata
         id (str): Fine-tune job ID
 
     Returns:
         List[FinetuneCheckpoint]: List of available checkpoints
     """
-    checkpoints: List[FinetuneCheckpoint] = []
-
-    for event in events:
-        event_type = event.type
-
-        if event_type == FinetuneEventType.CHECKPOINT_SAVE:
-            step = get_event_step(event)
-            checkpoint_name = f"{id}:{step}" if step is not None else id
-
-            checkpoints.append(
-                FinetuneCheckpoint(
-                    type=(
-                        f"Intermediate (step {step})"
-                        if step is not None
-                        else "Intermediate"
-                    ),
-                    timestamp=event.created_at,
-                    name=checkpoint_name,
-                )
-            )
-        elif event_type == FinetuneEventType.JOB_COMPLETE:
-            if hasattr(event, "model_path"):
-                checkpoints.append(
-                    FinetuneCheckpoint(
-                        type=(
-                            "Final Merged"
-                            if hasattr(event, "adapter_path")
-                            else "Final"
-                        ),
-                        timestamp=event.created_at,
-                        name=id,
-                    )
-                )
 
-            if hasattr(event, "adapter_path"):
-                checkpoints.append(
-                    FinetuneCheckpoint(
-                        type=(
-                            "Final Adapter" if hasattr(event, "model_path") else "Final"
-                        ),
-                        timestamp=event.created_at,
-                        name=id,
-                    )
-                )
+    parsed_checkpoints = []
+    for checkpoint in checkpoints:
+        step = checkpoint["step"]
+        checkpoint_type = checkpoint["checkpoint_type"]
+        checkpoint_name = (
+            f"{id}:{step}" if "intermediate" in checkpoint_type.lower() else id
+        )
 
-    # Sort by timestamp (newest first)
-    checkpoints.sort(key=lambda x: x.timestamp, reverse=True)
+        parsed_checkpoints.append(
+            FinetuneCheckpoint(
+                type=checkpoint_type,
+                timestamp=checkpoint["created_at"],
+                name=checkpoint_name,
+            )
+        )
 
-    return checkpoints
+    parsed_checkpoints.sort(key=lambda x: x.timestamp, reverse=True)
+    return parsed_checkpoints
 
 
 class FineTuning:
@@ -549,8 +523,21 @@ class FineTuning:
         Returns:
             List[FinetuneCheckpoint]: List of available checkpoints
         """
-        events = self.list_events(id).data or []
-        return _process_checkpoints_from_events(events, id)
+        requestor = api_requestor.APIRequestor(
+            client=self._client,
+        )
+
+        response, _, _ = requestor.request(
+            options=TogetherRequest(
+                method="GET",
+                url=f"fine-tunes/{id}/checkpoints",
+            ),
+            stream=False,
+        )
+        assert isinstance(response, TogetherResponse)
+
+        raw_checkpoints = response.data["data"]
+        return _parse_raw_checkpoints(raw_checkpoints, id)
 
     def download(
         self,
@@ -924,11 +911,9 @@ class AsyncFineTuning:
             ),
             stream=False,
         )
+        assert isinstance(events_response, TogetherResponse)
 
-        # FIXME: API returns "data" field with no object type (should be "list")
-        events_list = FinetuneListEvents(object="list", **events_response.data)
-
-        return events_list
+        return FinetuneListEvents(**events_response.data)
 
     async def list_checkpoints(self, id: str) -> List[FinetuneCheckpoint]:
         """
@@ -938,11 +923,23 @@ class AsyncFineTuning:
             id (str): Unique identifier of the fine-tune job to list checkpoints for
 
         Returns:
-            List[FinetuneCheckpoint]: Object containing list of available checkpoints
+            List[FinetuneCheckpoint]: List of available checkpoints
         """
-        events_list = await self.list_events(id)
-        events = events_list.data or []
-        return _process_checkpoints_from_events(events, id)
+        requestor = api_requestor.APIRequestor(
+            client=self._client,
+        )
+
+        response, _, _ = await requestor.arequest(
+            options=TogetherRequest(
+                method="GET",
+                url=f"fine-tunes/{id}/checkpoints",
+            ),
+            stream=False,
+        )
+        assert isinstance(response, TogetherResponse)
+
+        raw_checkpoints = response.data["data"]
+        return _parse_raw_checkpoints(raw_checkpoints, id)
 
     async def download(
         self, id: str, *, output: str | None = None, checkpoint_step: int = -1

{together-1.5.6 → together-1.5.8}/src/together/types/chat_completions.py

@@ -98,6 +98,7 @@ class ToolChoice(BaseModel):
 
 class ToolChoiceEnum(str, Enum):
     Auto = "auto"
+    Required = "required"
 
 
 class ChatCompletionRequest(BaseModel):

{together-1.5.6 → together-1.5.8}/src/together/types/finetune.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 from enum import Enum
-from typing import List, Literal
+from typing import List, Literal, Any
 
 from pydantic import StrictBool, Field, field_validator
 
@@ -329,8 +329,16 @@ class FinetuneDownloadResult(BaseModel):
 
 class FinetuneFullTrainingLimits(BaseModel):
     max_batch_size: int
+    max_batch_size_dpo: int = -1
     min_batch_size: int
 
+    def __init__(self, **data: Any) -> None:
+        super().__init__(**data)
+        if self.max_batch_size_dpo == -1:
+            half_max = self.max_batch_size // 2
+            rounded_half_max = (half_max // 8) * 8
+            self.max_batch_size_dpo = max(self.min_batch_size, rounded_half_max)
+
 
 class FinetuneLoraTrainingLimits(FinetuneFullTrainingLimits):
     max_rank: int