wire up squash_position_ids

* make multipack sampler patch explicit

* combining

.coderabbit.yaml

@@ -12,5 +12,6 @@ reviews:
   auto_review:
     enabled: true
     drafts: false
+    auto_incremental_review: true
 chat:
   auto_reply: true

examples/gpt-oss/README.md

@@ -41,6 +41,12 @@ model, and final model output, you may need at least 3TB of free disk space to k
 axolotl train examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
 ```
 
+To simplify fine-tuning across 2 nodes × 8x H100 (80GB) GPUs, we've partnered with [Baseten](https://baseten.co) to showcase multi-node
+training of the 120B model using Baseten Truss. You can read more about this recipe on
+[Baseten's blog](https://www.baseten.co/blog/how-to-fine-tune-gpt-oss-120b-with-baseten-and-axolotl/). The recipe can
+be found on their
+[GitHub](https://github.com/basetenlabs/ml-cookbook/tree/main/examples/oss-gpt-120b-axolotl/training).
+
 ERRATA: Transformers saves the model Architecture prefixed with `FSDP` which needs to be manually renamed in `config.json`.
 See https://github.com/huggingface/transformers/pull/40207 for the status of this issue.
 
@@ -61,9 +67,23 @@ mv ./outputs/gpt-oss-out/merged/* ./outputs/gpt-oss-out/
 
 ### Inferencing your fine-tuned model
 
+#### vLLM
+
 GPT-OSS support in vLLM does not exist in a stable release yet. See https://x.com/MaziyarPanahi/status/1955741905515323425
 for more information about using a special vllm-openai docker image for inferencing with vLLM.
 
+Optionally, vLLM can be installed from nightly:
+
+```bash
+pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
+```
+and the vLLM server can be started with the following command (modify `--tensor-parallel-size 8` to match your environment):
+```bash
+vllm serve ./outputs/gpt-oss-out/ --served-model-name axolotl/gpt-oss-20b --host 0.0.0.0 --port 8888  --tensor-parallel-size 8
+```
+
+#### SGLang
+
 SGLang has 0-day support in main, see https://github.com/sgl-project/sglang/issues/8833 for infomation on installing
 SGLang from source. Once you've installed SGLang, run the following command to launch a SGLang server:
 

examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -44,7 +44,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml

@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path: ./outputs/last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -41,7 +41,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path: ./outputs/last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml

@@ -53,7 +53,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

requirements.txt

@@ -13,8 +13,8 @@ liger-kernel==0.6.1
 packaging==23.2
 
 huggingface_hub>=0.33.0
-peft==0.17.0
+peft>=0.17.0
-transformers==4.55.2
+transformers==4.55.3
 tokenizers>=0.21.1
 accelerate==1.10.0
 datasets==4.0.0

setup.py

@@ -118,9 +118,9 @@ def get_package_version():
 
 
 extras_require = {
-    "flash-attn": ["flash-attn==2.8.2"],
+    "flash-attn": ["flash-attn==2.8.3"],
     "ring-flash-attn": [
-        "flash-attn==2.8.2",
+        "flash-attn==2.8.3",
         "ring-flash-attn>=0.1.7",
         "yunchang==0.6.0",
     ],

src/axolotl/cli/cloud/modal_.py

@@ -82,7 +82,7 @@ class ModalCloud(Cloud):
         return res
 
     def get_image(self):
-        docker_tag = "main-py3.11-cu124-2.6.0"
+        docker_tag = "main-py3.11-cu126-2.7.1"
         if self.config.docker_tag:
             docker_tag = self.config.docker_tag
         docker_image = f"axolotlai/axolotl:{docker_tag}"
@@ -200,7 +200,7 @@ class ModalCloud(Cloud):
         if family in ["a10", "a10g"]:
             return modal.gpu.A10G(count=count)
         if family == "h100":
-            return modal.gpu.H100(count=count)
+            return f"H100:{count}"
         if family == "t4":
             return modal.gpu.T4(count=count)
         if family == "l4":

src/axolotl/cli/inference.py

@@ -64,7 +64,7 @@ def do_inference(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
+        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
     elif cfg.datasets[0].type == "chat_template":
         chat_template_str = get_chat_template_from_config(
             cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer

src/axolotl/cli/preprocess.py

@@ -97,7 +97,8 @@ def do_cli(
     """
     # pylint: disable=duplicate-code
     os.environ["AXOLOTL_IS_PREPROCESS"] = "1"
-    parsed_cfg = load_cfg(config, **kwargs)
+    is_preprocess = kwargs.pop("is_preprocess", True)
+    parsed_cfg = load_cfg(config, is_preprocess=is_preprocess, **kwargs)
     parsed_cfg.is_preprocess = True
     parser = transformers.HfArgumentParser(PreprocessCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(

src/axolotl/cli/utils/sweeps.py

@@ -3,11 +3,12 @@
 import random
 from copy import deepcopy
 from itertools import product
+from typing import Any
 
 
 def generate_sweep_configs(
     base_config: dict[str, list], sweeps_config: dict[str, list]
-) -> list[dict[str, list]]:
+) -> list[dict[str, Any]]:
     """
     Recursively generates all possible configurations by applying sweeps to the base config.
 

src/axolotl/cli/utils/train.py

@@ -4,6 +4,7 @@ import os
 import subprocess  # nosec
 import sys
 import tempfile
+from pathlib import Path
 from typing import Any, Iterator, Literal
 
 import yaml
@@ -88,7 +89,12 @@ def generate_config_files(config: str, sweep: str | None) -> Iterator[tuple[str,
     # Generate all possible configurations
     permutations = generate_sweep_configs(base_config, sweep_config)
     is_group = len(permutations) > 1
-    for permutation in permutations:
+    base_output_dir = base_config.get("output_dir", "./model-out")
+    for idx, permutation in enumerate(permutations, start=1):
+        permutation_dir = Path(permutation.get("output_dir", base_output_dir))
+        permutation_id = f"sweep{idx:04d}"
+        permutation["output_dir"] = str(permutation_dir / permutation_id)
+
         # pylint: disable=consider-using-with
         temp_file = tempfile.NamedTemporaryFile(
             mode="w",

src/axolotl/common/datasets.py

@@ -6,7 +6,6 @@ from dataclasses import dataclass
 
 from datasets import Dataset
 
-import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
 from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.data import prepare_datasets, prepare_preference_datasets

src/axolotl/core/builders/causal.py

@@ -476,6 +476,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 )
             ):
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
+                if self.cfg.squash_position_ids:
+                    kwargs["squash_position_ids"] = True
             else:
                 collator = BatchSamplerDataCollatorForSeq2Seq
         else:

src/axolotl/loaders/patch_manager.py

@@ -277,6 +277,14 @@ class PatchManager:
                 has_remote_code=has_remote_code,
             )
 
+        if self.cfg.sample_packing:
+            from axolotl.monkeypatch.data.batch_dataset_fetcher import (
+                apply_multipack_dataloader_patch,
+            )
+
+            LOG.info("Applying multipack dataloader patch for sample packing...")
+            apply_multipack_dataloader_patch()
+
     def _apply_fsdp2_bnb_patches(self):
         """Apply FSDP2 BNB patches."""
         if (

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -187,7 +187,7 @@ def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
 
     # Linear4Bit will keep it's bias term in fp32. If the weight dtype is in bf16 we are not able to
     # wrap this. Therefore we must ensure the bias has the same dtype as the weight
-    if module.base_layer.bias is not None:
+    if hasattr(module.base_layer, "bias") and module.base_layer.bias is not None:
         if module.base_layer.weight.dtype != module.base_layer.bias.dtype:
             log_bias_dtype_mismatch = True
             module.base_layer.bias.data = module.base_layer.bias.data.to(

src/axolotl/monkeypatch/data/batch_dataset_fetcher.py

@@ -1,4 +1,4 @@
-"""monkey patches for the dataset fetcher to handle batches of packed indexes"""
+"""Monkey patches for the dataset fetcher to handle batches of packed indexes."""
 
 # pylint: disable=protected-access
 
@@ -6,10 +6,20 @@ import torch
 from torch.utils.data._utils.fetch import _BaseDatasetFetcher
 from torch.utils.data._utils.worker import _worker_loop
 
+_ORIGINAL_MAP_DATASET_FETCHER = None
+_ORIGINAL_WORKER_LOOP = None
+_IS_PATCHED = False
+
 
 class _MapDatasetFetcher(_BaseDatasetFetcher):
+    """
+    Custom dataset fetcher that handles nested batch structures from
+    MultipackBatchSampler.
+    """
+
     def fetch(self, possibly_batched_index):
         if isinstance(possibly_batched_index[0], list):
+            # Handle nested structure from MultipackBatchSampler
             data = [None for i in possibly_batched_index]
             for i, possibly_batched_index_ in enumerate(possibly_batched_index):
                 if self.auto_collation:
@@ -23,6 +33,7 @@ class _MapDatasetFetcher(_BaseDatasetFetcher):
                 else:
                     data[i] = self.dataset[possibly_batched_index_]
         else:
+            # Standard batch handling
             if self.auto_collation:
                 if hasattr(self.dataset, "__getitems__") and self.dataset.__getitems__:
                     data = self.dataset.__getitems__(possibly_batched_index)
@@ -34,14 +45,54 @@ class _MapDatasetFetcher(_BaseDatasetFetcher):
 
 
 def patch_fetchers():
+    """Apply patches to PyTorch's DataLoader components."""
     torch.utils.data._utils.fetch._MapDatasetFetcher = _MapDatasetFetcher
     torch.utils.data.dataloader._utils.fetch._MapDatasetFetcher = _MapDatasetFetcher
 
 
 def patched_worker_loop(*args, **kwargs):
+    """Worker loop that ensures patches are applied in worker processes."""
     patch_fetchers()
     return _worker_loop(*args, **kwargs)
 
 
-torch.utils.data._utils.worker._worker_loop = patched_worker_loop
+def apply_multipack_dataloader_patch():
-patch_fetchers()
+    """
+    This patch allows DataLoader to correctly process batches that contain multiple bins
+    of packed sequences.
+    """
+    # pylint: disable=global-statement
+    global _ORIGINAL_MAP_DATASET_FETCHER, _ORIGINAL_WORKER_LOOP, _IS_PATCHED
+
+    if _IS_PATCHED:
+        return
+
+    # Store original implementations
+    _ORIGINAL_MAP_DATASET_FETCHER = torch.utils.data._utils.fetch._MapDatasetFetcher
+    _ORIGINAL_WORKER_LOOP = torch.utils.data._utils.worker._worker_loop
+
+    # Apply patches
+    patch_fetchers()
+    torch.utils.data._utils.worker._worker_loop = patched_worker_loop
+
+    _IS_PATCHED = True
+
+
+def remove_multipack_dataloader_patch():
+    """Remove the monkeypatch and restore original PyTorch DataLoader behavior."""
+    # pylint: disable=global-statement
+    global _IS_PATCHED
+
+    if not _IS_PATCHED:
+        return
+
+    if _ORIGINAL_MAP_DATASET_FETCHER:
+        torch.utils.data._utils.fetch._MapDatasetFetcher = _ORIGINAL_MAP_DATASET_FETCHER
+        torch.utils.data.dataloader._utils.fetch._MapDatasetFetcher = (
+            _ORIGINAL_MAP_DATASET_FETCHER
+        )
+
+    if _ORIGINAL_WORKER_LOOP:
+        torch.utils.data._utils.worker._worker_loop = _ORIGINAL_WORKER_LOOP
+
+    _IS_PATCHED = False

src/axolotl/train.py

@@ -253,7 +253,9 @@ def save_trained_model(
             # final model weights have already been saved by `ReLoRACallback.on_train_end`
             return
 
-    if trainer.is_fsdp_enabled or cfg.fsdp_config:
+    if (  # pylint: disable=too-many-nested-blocks
+        trainer.is_fsdp_enabled or cfg.fsdp_config
+    ):
         if cfg.fsdp_config or cfg.fsdp:
             if cfg.fsdp_config.final_state_dict_type:
                 state_dict_type = cfg.fsdp_config.final_state_dict_type
@@ -285,6 +287,8 @@ def save_trained_model(
                 if trainer.accelerator.is_main_process:
                     # move all files in merged_path to cfg.output_dir
                     for merged_file in Path(merged_path).iterdir():
+                        if (Path(cfg.output_dir) / merged_file.name).exists():
+                            (Path(cfg.output_dir) / merged_file.name).unlink()
                         shutil.move(str(merged_file), cfg.output_dir)
                     shutil.rmtree(merged_path)  # remove what should be an empty dir
         # TODO(wing):see https://github.com/huggingface/transformers/pull/40207

src/axolotl/utils/schemas/config.py

@@ -459,6 +459,12 @@ class AxolotlInputConfig(
             "description": "The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or 'forkserver'"
         },
     )
+    squash_position_ids: bool | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Whether to squash position_ids for packing, effectively extending context length."
+        },
+    )
     eval_sample_packing: bool | None = Field(
         default=None,
         json_schema_extra={

tests/test_packed_batch_sampler.py

@@ -48,7 +48,13 @@ class TestBatchedSamplerPacking:
         max_seq_length,
         sequential,
     ):
-        import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
+        from axolotl.monkeypatch.data.batch_dataset_fetcher import (
+            apply_multipack_dataloader_patch,
+            remove_multipack_dataloader_patch,
+        )
+
+        # Apply the patch for multipack handling
+        apply_multipack_dataloader_patch()
 
         dataset = dataset_winglian_tiny_shakespeare["train"]
 
@@ -101,10 +107,14 @@ class TestBatchedSamplerPacking:
             for pack in batch:
                 batch_idxs.extend(pack)
 
-        for batch in loader:
+        try:
-            assert batch["input_ids"].numel() <= batch_size * max_seq_length
+            for batch in loader:
-            assert batch["input_ids"].shape[1] == max_seq_length
+                assert batch["input_ids"].numel() <= batch_size * max_seq_length
+                assert batch["input_ids"].shape[1] == max_seq_length
 
-        original_idxs = set(range(len(train_dataset)))
+            original_idxs = set(range(len(train_dataset)))
-        assert original_idxs == set(batch_idxs)
+            assert original_idxs == set(batch_idxs)
-        assert len(batch_idxs) == len(set(batch_idxs))
+            assert len(batch_idxs) == len(set(batch_idxs))
+        finally:
+            # Clean up: remove the patch after the test
+            remove_multipack_dataloader_patch()