together 2.0.0a15__tar.gz → 2.0.0a17__tar.gz

together-2.0.0a17/.release-please-manifest.json

@@ -0,0 +1,3 @@
+{
+  ".": "2.0.0-alpha.17"
+}

{together-2.0.0a15 → together-2.0.0a17}/CHANGELOG.md

@@ -1,5 +1,39 @@
 # Changelog
 
+## 2.0.0-alpha.17 (2026-01-21)
+
+Full Changelog: [v2.0.0-alpha.16...v2.0.0-alpha.17](https://github.com/togethercomputer/together-py/compare/v2.0.0-alpha.16...v2.0.0-alpha.17)
+
+### Features
+
+* **cli:** add b200 and h200 GPU options for endpoint creation ([#218](https://github.com/togethercomputer/together-py/issues/218)) ([b514912](https://github.com/togethercomputer/together-py/commit/b514912a281922fefbf8a9f62b936ed1de243718))
+* Improve usage of models list cli command ([#216](https://github.com/togethercomputer/together-py/issues/216)) ([430e6c1](https://github.com/togethercomputer/together-py/commit/430e6c1e030749be474f020b677d91014ba4482c))
+
+
+### Chores
+
+* Deprecate CLI usage for endpoints create flag --no-promopt-cache ([#219](https://github.com/togethercomputer/together-py/issues/219)) ([55e9700](https://github.com/togethercomputer/together-py/commit/55e9700187b42f8baff6f567a3a657b46577ed88))
+* Mark disable_prompt_cache as deprecated for endpoint creation ([6a629b2](https://github.com/togethercomputer/together-py/commit/6a629b29e53b4374503d30ca75456184ef313b67))
+
+## 2.0.0-alpha.16 (2026-01-18)
+
+Full Changelog: [v2.0.0-alpha.15...v2.0.0-alpha.16](https://github.com/togethercomputer/together-py/compare/v2.0.0-alpha.15...v2.0.0-alpha.16)
+
+### Features
+
+* Add backwards compatible support for google colab api keys ([#211](https://github.com/togethercomputer/together-py/issues/211)) ([80dacca](https://github.com/togethercomputer/together-py/commit/80daccafb4ef69438be1d98773fe3a31ac09cace))
+* Add together clusters get-credentials CLI command ([#208](https://github.com/togethercomputer/together-py/issues/208)) ([fa54aa9](https://github.com/togethercomputer/together-py/commit/fa54aa9646e9dab2822491473856f64d5cf4688f))
+* **client:** add support for binary request streaming ([8464e12](https://github.com/togethercomputer/together-py/commit/8464e12187b42ff3a58c82ca55ec284ffa98aeea))
+
+
+### Chores
+
+* Add code samples and descriptions to instant cluster apis ([e1ad614](https://github.com/togethercomputer/together-py/commit/e1ad614f8c3189ddc991fab96013cf5f7aace1d8))
+* Improve example script ([#213](https://github.com/togethercomputer/together-py/issues/213)) ([7839058](https://github.com/togethercomputer/together-py/commit/783905873500e3228837e2cf87ebc097d9026539))
+* **internal:** update `actions/checkout` version ([dc0819c](https://github.com/togethercomputer/together-py/commit/dc0819c95429f69bf0d7ef442b7b1ea811ec40b5))
+* Port tokenize_data example ([#209](https://github.com/togethercomputer/together-py/issues/209)) ([f2714a8](https://github.com/togethercomputer/together-py/commit/f2714a84f2802989a8f519c21e5c62e9fc787038))
+* Update cluster apis to reflect their new response shape ([6be132b](https://github.com/togethercomputer/together-py/commit/6be132b210973e9c26beedd0080039e28e522096))
+
 ## 2.0.0-alpha.15 (2026-01-09)
 
 Full Changelog: [v2.0.0-alpha.14...v2.0.0-alpha.15](https://github.com/togethercomputer/together-py/compare/v2.0.0-alpha.14...v2.0.0-alpha.15)

{together-2.0.0a15 → together-2.0.0a17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 2.0.0a15
+Version: 2.0.0a17
 Summary: The official Python library for the together API
 Project-URL: Homepage, https://github.com/togethercomputer/together-py
 Project-URL: Repository, https://github.com/togethercomputer/together-py
@@ -33,12 +33,15 @@ Requires-Dist: rich>=13.7.1
 Requires-Dist: sniffio
 Requires-Dist: tabulate>=0.9.0
 Requires-Dist: tqdm>=4.67.1
+Requires-Dist: types-pyyaml>=6.0.12.20250915
 Requires-Dist: types-tabulate>=0.9.0.20240106
 Requires-Dist: types-tqdm>=4.67.0.20250516
 Requires-Dist: typing-extensions<5,>=4.10
 Provides-Extra: aiohttp
 Requires-Dist: aiohttp; extra == 'aiohttp'
 Requires-Dist: httpx-aiohttp>=0.1.9; extra == 'aiohttp'
+Provides-Extra: cli
+Requires-Dist: pyyaml; extra == 'cli'
 Provides-Extra: pyarrow
 Requires-Dist: pyarrow-stubs>=10.0.1.7; extra == 'pyarrow'
 Requires-Dist: pyarrow>=16.1.0; extra == 'pyarrow'
@@ -236,7 +239,6 @@ cluster = client.beta.clusters.create(
     billing_type="RESERVED",
     cluster_name="cluster_name",
     driver_version="CUDA_12_5_555",
-    duration_days=0,
     gpu_type="H100_SXM",
     num_gpus=0,
     region="us-central-8",

{together-2.0.0a15 → together-2.0.0a17}/README.md

@@ -190,7 +190,6 @@ cluster = client.beta.clusters.create(
     billing_type="RESERVED",
     cluster_name="cluster_name",
     driver_version="CUDA_12_5_555",
-    duration_days=0,
     gpu_type="H100_SXM",
     num_gpus=0,
     region="us-central-8",

{together-2.0.0a15 → together-2.0.0a17}/api.md

@@ -7,8 +7,6 @@ Types:
 ```python
 from together.types.beta import (
     Cluster,
-    ClusterCreateResponse,
-    ClusterUpdateResponse,
     ClusterListResponse,
     ClusterDeleteResponse,
     ClusterListRegionsResponse,
@@ -17,9 +15,9 @@ from together.types.beta import (
 
 Methods:
 
-- <code title="post /clusters">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">create</a>(\*\*<a href="src/together/types/beta/cluster_create_params.py">params</a>) -> <a href="./src/together/types/beta/cluster_create_response.py">ClusterCreateResponse</a></code>
+- <code title="post /clusters">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">create</a>(\*\*<a href="src/together/types/beta/cluster_create_params.py">params</a>) -> <a href="./src/together/types/beta/cluster.py">Cluster</a></code>
 - <code title="get /clusters/{cluster_id}">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">retrieve</a>(cluster_id) -> <a href="./src/together/types/beta/cluster.py">Cluster</a></code>
-- <code title="put /clusters/{cluster_id}">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">update</a>(cluster_id, \*\*<a href="src/together/types/beta/cluster_update_params.py">params</a>) -> <a href="./src/together/types/beta/cluster_update_response.py">ClusterUpdateResponse</a></code>
+- <code title="put /clusters/{cluster_id}">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">update</a>(cluster_id, \*\*<a href="src/together/types/beta/cluster_update_params.py">params</a>) -> <a href="./src/together/types/beta/cluster.py">Cluster</a></code>
 - <code title="get /clusters">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">list</a>() -> <a href="./src/together/types/beta/cluster_list_response.py">ClusterListResponse</a></code>
 - <code title="delete /clusters/{cluster_id}">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">delete</a>(cluster_id) -> <a href="./src/together/types/beta/cluster_delete_response.py">ClusterDeleteResponse</a></code>
 - <code title="get /clusters/regions">client.beta.clusters.<a href="./src/together/resources/beta/clusters/clusters.py">list_regions</a>() -> <a href="./src/together/types/beta/cluster_list_regions_response.py">ClusterListRegionsResponse</a></code>
@@ -29,17 +27,12 @@ Methods:
 Types:
 
 ```python
-from together.types.beta.clusters import (
-    ClusterStorage,
-    StorageCreateResponse,
-    StorageListResponse,
-    StorageDeleteResponse,
-)
+from together.types.beta.clusters import ClusterStorage, StorageListResponse, StorageDeleteResponse
 ```
 
 Methods:
 
-- <code title="post /clusters/storages">client.beta.clusters.storage.<a href="./src/together/resources/beta/clusters/storage.py">create</a>(\*\*<a href="src/together/types/beta/clusters/storage_create_params.py">params</a>) -> <a href="./src/together/types/beta/clusters/storage_create_response.py">StorageCreateResponse</a></code>
+- <code title="post /clusters/storages">client.beta.clusters.storage.<a href="./src/together/resources/beta/clusters/storage.py">create</a>(\*\*<a href="src/together/types/beta/clusters/storage_create_params.py">params</a>) -> <a href="./src/together/types/beta/clusters/cluster_storage.py">ClusterStorage</a></code>
 - <code title="get /clusters/storages/{volume_id}">client.beta.clusters.storage.<a href="./src/together/resources/beta/clusters/storage.py">retrieve</a>(volume_id) -> <a href="./src/together/types/beta/clusters/cluster_storage.py">ClusterStorage</a></code>
 - <code title="put /clusters/storages">client.beta.clusters.storage.<a href="./src/together/resources/beta/clusters/storage.py">update</a>(\*\*<a href="src/together/types/beta/clusters/storage_update_params.py">params</a>) -> <a href="./src/together/types/beta/clusters/cluster_storage.py">ClusterStorage</a></code>
 - <code title="get /clusters/storages">client.beta.clusters.storage.<a href="./src/together/resources/beta/clusters/storage.py">list</a>() -> <a href="./src/together/types/beta/clusters/storage_list_response.py">StorageListResponse</a></code>

together-2.0.0a17/examples/tokenize_data.py

@@ -0,0 +1,259 @@
+import os
+import logging
+import argparse
+from typing import Dict, List
+from functools import partial
+from multiprocessing import cpu_count
+
+from datasets import Dataset, load_dataset  # type: ignore
+from transformers import (  # type: ignore
+    AutoTokenizer,
+    BatchEncoding,
+    PreTrainedTokenizerBase,
+)
+
+# see default of ignore_index
+# for https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html#torch.nn.CrossEntropyLoss
+LOSS_IGNORE_INDEX = -100
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+
+
+def tokenize_variable_length(
+    data: Dict[str, str],
+    tokenizer: PreTrainedTokenizerBase,
+    add_special_tokens: bool = True,
+) -> BatchEncoding:
+    tokenized = tokenizer(data["text"], add_special_tokens=add_special_tokens, truncation=False)
+    return tokenized
+
+
+def tokenize_constant_length(
+    data: Dict[str, str],
+    tokenizer: PreTrainedTokenizerBase,
+    max_length: int = 2048,
+    add_special_tokens: bool = True,
+    add_labels: bool = True,
+) -> BatchEncoding:
+    # tokenized contains `input_ids` and `attention_mask`
+    tokenized: BatchEncoding = tokenizer(
+        data["text"],
+        max_length=max_length,
+        truncation=True,
+        padding="max_length",
+        add_special_tokens=add_special_tokens,
+    )
+    # add labels to mask out any padding tokens
+    if add_labels:
+        tokenized["labels"] = [
+            LOSS_IGNORE_INDEX if token_id == tokenizer.pad_token_id else token_id for token_id in tokenized["input_ids"]
+        ]
+
+    return tokenized
+
+
+def pack_sequences(
+    batch: BatchEncoding,
+    max_seq_len: int,
+    pad_token_id: int,
+    eos_token_id: int,
+    add_labels: bool,
+    cutoff_size: int = 0,
+) -> Dict[str, List[List[int]]]:
+    """
+    cutoff_size = max_seq_len means that we will drop any non-full sequences
+        (full packing without padding)
+    Example:
+        Sequence 1:
+        ['<s>', '▁usually', '▁,', '▁he', '▁would', '▁be', '▁t', 'earing']
+        Sequence 2:
+        ['▁around', '▁the', '▁living', '▁room', '▁,', '▁playing', '▁with', '▁his']
+        Sequence 3:
+        ['▁toys', '▁.', '</s>', '<s>', '▁but', '▁just', '▁one', '▁look']
+    """
+    packed_sequences = []
+    buffer = []
+
+    for input_ids in batch["input_ids"]:
+        # Add the current sequence to the buffer
+        buffer.extend(input_ids)
+        buffer.append(eos_token_id)  # Add EOS at the end of each sequence
+
+        # Check if buffer needs to be split into chunks
+        while len(buffer) > max_seq_len:
+            # Take a full chunk from the buffer and append it to packed_sequences
+            packed_sequences.append(buffer[:max_seq_len])
+            # Remove the processed chunk from the buffer
+            buffer = buffer[max_seq_len:]
+
+    # Add the last buffer if it's exactly chunk_size
+    if len(buffer) == max_seq_len:
+        packed_sequences.append(buffer)
+    elif len(buffer) > cutoff_size:
+        # if the buffer is larger than the cutoff size, pad it to the chunk_size
+        # if not, we do not include in the packed_sequences
+        buffer.extend([pad_token_id] * (max_seq_len - len(buffer)))
+        packed_sequences.append(buffer)
+
+    output = {"input_ids": packed_sequences}
+    if add_labels:
+        output["labels"] = [
+            [LOSS_IGNORE_INDEX if token_id == pad_token_id else token_id for token_id in example]
+            for example in output["input_ids"]
+        ]
+
+    # mask attention for padding tokens, a better version would also mask cross-sequence dependencies
+    output["attention_mask"] = [
+        [0 if token_id == pad_token_id else 1 for token_id in example] for example in output["input_ids"]
+    ]
+    return output
+
+
+def process_fast_packing(
+    dataset: Dataset,
+    tokenizer: PreTrainedTokenizerBase,
+    max_sequence_length: int,
+    add_labels: bool,
+    add_special_tokens: bool,
+) -> Dataset:
+    tokenized_dataset = dataset.map(
+        lambda examples: tokenize_variable_length(examples, tokenizer, add_special_tokens=add_special_tokens),
+        batched=True,
+        num_proc=cpu_count(),
+        load_from_cache_file=True,
+        remove_columns=dataset.column_names,
+    )
+    logger.info(f"tokenized dataset: {tokenized_dataset}")
+
+    packed_dataset = tokenized_dataset.map(
+        lambda batch: pack_sequences(
+            batch,
+            max_sequence_length,
+            tokenizer.pad_token_id,
+            tokenizer.eos_token_id,
+            add_labels=add_labels,
+            cutoff_size=max_sequence_length,
+        ),
+        batched=True,
+        num_proc=cpu_count() if len(tokenized_dataset) > 10000 else 1,
+        remove_columns=["attention_mask"],
+    )
+    logger.info(f"Packed dataset: {packed_dataset}")
+    return packed_dataset
+
+
+def process_data(args: argparse.Namespace) -> None:
+    if not args.out_filename.endswith(".parquet"):
+        raise ValueError("`--out_filename` should have the `.parquet` extension")
+
+    if os.path.isfile(args.dataset) and args.dataset.endswith(".jsonl"):
+        dataset = load_dataset("json", data_files=args.dataset, split="train")
+    else:
+        dataset = load_dataset(args.dataset, split="train")
+    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer)
+    tokenizer.pad_token = tokenizer.eos_token
+
+    # Handle prompt/completion format by combining into text field
+    prompt_field = args.prompt_field
+    completion_field = args.completion_field
+    separator = args.separator
+    
+    if prompt_field in dataset.column_names and completion_field in dataset.column_names:
+        if "text" not in dataset.column_names:
+            dataset = dataset.map(
+                lambda x: {"text": x[prompt_field] + separator + x[completion_field]},
+                remove_columns=[prompt_field, completion_field],
+            )
+    elif "text" not in dataset.column_names:
+        raise ValueError(
+            f"Dataset must have either 'text' field or both '{prompt_field}' and '{completion_field}' fields. "
+            f"Available columns: {dataset.column_names}"
+        )
+
+    if not args.packing:
+        tokenized_data = dataset.map(
+            partial(
+                tokenize_constant_length,
+                tokenizer=tokenizer,
+                max_length=args.max_seq_length,
+                add_special_tokens=True,
+                add_labels=args.add_labels,
+            ),
+            batched=False,
+            num_proc=cpu_count(),
+            remove_columns=dataset.column_names,
+        )
+    else:
+        tokenized_data = process_fast_packing(
+            dataset,
+            tokenizer,
+            max_sequence_length=args.max_seq_length,
+            add_labels=args.add_labels,
+            add_special_tokens=True,
+        )
+
+    assert "input_ids" in tokenized_data.column_names and "attention_mask" in tokenized_data.column_names
+
+    if args.add_labels:
+        assert "labels" in tokenized_data.column_names
+
+    logger.info("Tokenized data:")
+    print(tokenized_data)
+
+    logger.info(f"Saving data to {args.out_filename}")
+    print(len(tokenized_data[0]["input_ids"]))
+    tokenized_data.to_parquet(args.out_filename)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Pretokenize examples for finetuning via Together")
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="clam004/antihallucination_dataset",
+        help="Dataset name on the Hugging Face Hub or path to local JSONL file",
+    )
+    parser.add_argument("--max-seq-length", type=int, default=8192, help="Maximum sequence length")
+    parser.add_argument(
+        "--add-labels",
+        action="store_true",
+        help="Whether to add loss labels from padding tokens",
+    )
+    parser.add_argument(
+        "--tokenizer",
+        type=str,
+        required=True,
+        help="Tokenizer name (for example, togethercomputer/Llama-3-8b-hf)",
+    )
+    parser.add_argument(
+        "--out-filename",
+        default="processed_dataset.parquet",
+        help="Name of the Parquet file to save (should have .parquet extension)",
+    )
+    parser.add_argument(
+        "--packing",
+        action="store_true",
+        help="Whether to pack shorter sequences up to `--max-seq-length`",
+    )
+    parser.add_argument(
+        "--prompt-field",
+        type=str,
+        default="prompt",
+        help="Name of the field containing the prompt/input text (default: 'prompt')",
+    )
+    parser.add_argument(
+        "--completion-field",
+        type=str,
+        default="completion",
+        help="Name of the field containing the completion/output text (default: 'completion')",
+    )
+    parser.add_argument(
+        "--separator",
+        type=str,
+        default="",
+        help="Separator to insert between prompt and completion fields (default: empty string)",
+    )
+    args = parser.parse_args()
+
+    process_data(args)

{together-2.0.0a15 → together-2.0.0a17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "together"
-version = "2.0.0-alpha.15"
+version = "2.0.0-alpha.17"
 description = "The official Python library for the together API"
 dynamic = ["readme"]
 license = "Apache-2.0"
@@ -15,14 +15,15 @@ dependencies = [
   "anyio>=3.5.0, <5",
   "distro>=1.7.0, <2",
   "sniffio",
-    "click>=8.1.7",
-    "rich>=13.7.1",
-    "tabulate>=0.9.0",
-    "pillow>=10.4.0",
-    "types-tabulate>=0.9.0.20240106",
-    "tqdm>=4.67.1",
-    "types-tqdm>=4.67.0.20250516",
-    "filelock>=3.13.1",
+  "click>=8.1.7",
+  "rich>=13.7.1",
+  "tabulate>=0.9.0",
+  "pillow>=10.4.0",
+  "types-tabulate>=0.9.0.20240106",
+  "tqdm>=4.67.1",
+  "types-tqdm>=4.67.0.20250516",
+  "filelock>=3.13.1",
+  "types-pyyaml>=6.0.12.20250915",
 ]
 
 requires-python = ">= 3.9"
@@ -51,9 +52,10 @@ Repository = "https://github.com/togethercomputer/together-py"
 [project.optional-dependencies]
 aiohttp = ["aiohttp", "httpx_aiohttp>=0.1.9"]
 pyarrow = ["pyarrow>=16.1.0", "pyarrow-stubs>=10.0.1.7"]
+cli = ["PyYAML"]
 
 [project.scripts]
-together = "together.lib.cli.cli:main"
+together = "together.lib.cli:main"
 
 [tool.uv]
 managed = true
@@ -150,6 +152,7 @@ exclude = [
     ".venv",
     ".nox",
     ".git",
+    "examples",
 ]
 
 reportImplicitOverride = true