together 1.5.6__tar.gz → 1.5.7__tar.gz

{together-1.5.6 → together-1.5.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.6
+Version: 1.5.7
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI

{together-1.5.6 → together-1.5.7}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.5.6"
+version = "1.5.7"
 authors = ["Together AI <support@together.ai>"]
 description = "Python client for Together's Cloud Platform!"
 readme = "README.md"

{together-1.5.6 → together-1.5.7}/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-1.5.7}/src/together/resources/finetune.py

@@ -102,6 +102,7 @@ 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 +120,7 @@ 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 +129,24 @@ 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(

{together-1.5.6 → together-1.5.7}/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