add colab callback to fix inference post train

fix error with lambda and pickling

make sure things are in float instead of np.float

src/axolotl/cli/art.py

@@ -16,8 +16,15 @@ AXOLOTL_LOGO = """
     @@@@  @@@@@@@@@@@@@@@@
 """
 
+HAS_PRINTED_LOGO = False
+
 
 def print_axolotl_text_art():
     """Prints axolotl ASCII art."""
+
+    global HAS_PRINTED_LOGO  # pylint: disable=global-statement
+    if HAS_PRINTED_LOGO:
+        return
     if is_main_process():
+        HAS_PRINTED_LOGO = True
         print(AXOLOTL_LOGO)

src/axolotl/common/datasets.py

@@ -48,6 +48,7 @@ def load_datasets(
     *,
     cfg: DictDefault,
     cli_args: PreprocessCliArgs | TrainerCliArgs | None = None,
+    debug: bool = False,
 ) -> TrainDatasetMeta:
     """
     Loads one or more training or evaluation datasets, calling
@@ -56,6 +57,7 @@ def load_datasets(
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: Command-specific CLI arguments.
+        debug: Whether to print out tokenization of sample
 
     Returns:
         Dataclass with fields for training and evaluation datasets and the computed
@@ -77,20 +79,25 @@ def load_datasets(
         preprocess_iterable=preprocess_iterable,
     )
 
-    if cli_args and (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
+    if (  # pylint: disable=too-many-boolean-expressions
+        cli_args
+        and (
+            cli_args.debug
+            or cfg.debug
+            or cli_args.debug_text_only
+            or int(cli_args.debug_num_examples) > 0
+        )
+    ) or debug:
         LOG.info("check_dataset_labels...")
 
-        train_samples = sample_dataset(train_dataset, cli_args.debug_num_examples)
+        num_examples = cli_args.debug_num_examples if cli_args else 1
+        text_only = cli_args.debug_text_only if cli_args else False
+        train_samples = sample_dataset(train_dataset, num_examples)
         check_dataset_labels(
             train_samples,
             tokenizer,
-            num_examples=cli_args.debug_num_examples,
-            text_only=cli_args.debug_text_only,
+            num_examples=num_examples,
+            text_only=text_only,
         )
 
         LOG.info("printing prompters...")

src/axolotl/core/trainer_builder.py

@@ -21,6 +21,7 @@ import importlib.util
 import inspect
 import logging
 import math
+import os
 import sys
 from abc import abstractmethod
 from pathlib import Path
@@ -72,6 +73,7 @@ from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
@@ -293,6 +295,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
 
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
         callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 

src/axolotl/core/trainers/base.py

@@ -114,6 +114,8 @@ class AxolotlTrainer(
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            group_size=self.args.sample_packing_group_size,
+            bin_size=self.args.sample_packing_bin_size,
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -72,7 +72,7 @@ class CutCrossEntropyPlugin(BasePlugin):
         if cfg.cut_cross_entropy:
             self._check_requirements()
 
-            from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
+            from .monkeypatch.patch import (
                 cce_patch,
             )
 

src/axolotl/monkeypatch/attention/__init__.py

@@ -0,0 +1,19 @@
+"""
+attention module for attention monkeypatches
+"""
+
+from transformers.integrations.flash_attention import flash_attention_forward
+
+
+def patch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    from .xformers import xformers_attention_forward
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = xformers_attention_forward
+
+
+def unpatch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward()

src/axolotl/monkeypatch/attention/xformers.py

@@ -0,0 +1,160 @@
+"""
+xformers attention implementation for packing
+"""
+
+from typing import Optional
+
+import torch
+import xformers
+import xformers.ops.fmha
+from transformers.modeling_flash_attention_utils import (
+    _upad_input,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+xformers_attention = xformers.ops.fmha.memory_efficient_attention
+
+
+def xformers_attention_forward(
+    module: torch.nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    dropout: float = 0.0,  # pylint: disable=unused-argument
+    scaling: Optional[float] = None,  # pylint: disable=unused-argument
+    sliding_window: Optional[int] = None,  # pylint: disable=unused-argument
+    softcap: Optional[float] = None,  # pylint: disable=unused-argument
+    cu_seq_lens_q: Optional[torch.LongTensor] = None,
+    cu_seq_lens_k: Optional[torch.LongTensor] = None,
+    max_length_q: Optional[int] = None,
+    max_length_k: Optional[int] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
+):
+    # Get dimensions
+    # query: [batch, heads, seq_len, hidden_dim]
+    batch_size = query.size(0)
+    query_length = query.shape[2]
+    key_length = key.shape[2]
+
+    # Default causal mask
+    attn_bias = xformers.ops.LowerTriangularMask()
+
+    # Check if we have sliding window attention
+    has_sliding_window = sliding_window is not None and sliding_window < query_length
+
+    # Transpose dimensions for xformers (Q: [b, h, s, d] -> [b, s, h, d])
+    query = query.transpose(1, 2)
+    key = key.transpose(1, 2)
+    value = value.transpose(1, 2)
+
+    # Get GQA parameters
+    num_attention_heads = module.config.num_attention_heads
+    num_key_value_heads = module.config.num_key_value_heads
+    head_dim = query.size(-1)
+    is_gqa = num_attention_heads != num_key_value_heads
+    n_groups = num_attention_heads // num_key_value_heads if is_gqa else 1
+
+    # If position_ids is provided and check all examples do not contain only 1 sequence, If tensor in increasing
+    # then we probably have one sequence, otherwise it is packed. Additionally check we are in pre-fill/training stage.
+    # Use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach
+    if position_ids is not None and (
+        max_length_q is not None
+        or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all())
+    ):
+        if cu_seq_lens_q is None or cu_seq_lens_k is None:
+            cu_seq_lens_q = get_cu_seqlens_from_pos_ids(position_ids)[0]
+            cu_seq_lens_q = cu_seq_lens_q.squeeze()
+            seq_lengths = cu_seq_lens_q[1:] - cu_seq_lens_q[:-1]
+            attn_bias = (
+                xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+                    q_seqlen=seq_lengths.tolist(),
+                )
+            )
+        else:
+            query = query.reshape(-1, query.size(-2), query.size(-1))
+            key = key.reshape(-1, key.size(-2), key.size(-1))
+            value = value.reshape(-1, value.size(-2), value.size(-1))
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+
+    elif attention_mask is not None:
+        query, key, value, _, cu_seq_lens, _ = _upad_input(
+            query, key, value, attention_mask, query_length
+        )
+        cu_seq_lens_q, cu_seq_lens_k = cu_seq_lens
+        seq_lengths = []
+        for i in range(len(cu_seq_lens_q) - 1):
+            seq_lengths.append(cu_seq_lens_q[i + 1] - cu_seq_lens_q[i])
+        attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+            q_seqlen=seq_lengths,
+            kv_seqlen=seq_lengths,
+        )
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+    else:
+        # Handle Group Query Attention (GQA) using view/expand approach from reference
+        key = key.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        value = value.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        key = key.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+        value = value.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+
+        if module.training:
+            key = key.reshape(batch_size, key_length, num_attention_heads, head_dim)
+            value = value.reshape(batch_size, key_length, num_attention_heads, head_dim)
+
+            if has_sliding_window:
+                query = query.view(
+                    1, batch_size * query_length, num_attention_heads, head_dim
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+        else:
+            query = query.view(
+                batch_size, query_length, num_key_value_heads, n_groups, head_dim
+            )
+
+            # If we need a sliding window attention
+            if has_sliding_window:
+                query = query.view(
+                    1,
+                    batch_size * query_length,
+                    num_key_value_heads,
+                    n_groups,
+                    head_dim,
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+
+    # Run the xformers attention
+    attn_output = xformers_attention(
+        query,
+        key,
+        value,
+        attn_bias=attn_bias,
+    )
+
+    attn_output = attn_output.view(
+        batch_size, -1, attn_output.size(-2), attn_output.size(-1)
+    )
+    return attn_output, None

src/axolotl/monkeypatch/loss/chunked.py

@@ -0,0 +1,134 @@
+"""
+chunked ce loss
+"""
+
+from typing import List, Optional
+
+import torch
+import torch.nn.functional as F
+
+
+# copied and modified from torchtune.modules.loss.CEWithChunkedOutputLoss
+class CEWithChunkedOutputLoss(torch.nn.Module):
+    """
+    Cross-entropy with chunked outputs that saves memory by only upcasting one chunk at a time.
+
+    For more details, please refer to: https://github.com/pytorch/torchtune/pull/1390
+    """
+
+    def __init__(self, num_output_chunks: int = 8, ignore_index: int = -100):
+        super().__init__()
+        self.num_output_chunks = num_output_chunks
+        self.ignore_index = ignore_index
+
+    def compute_cross_entropy(
+        self,
+        logits: torch.Tensor,
+        labels: torch.Tensor,
+        normalize: bool = True,  # pylint: disable=unused-argument
+    ) -> torch.Tensor:
+        """
+        Upcast logits to fp32 and compute cross entropy loss.
+        """
+        return F.cross_entropy(
+            logits.float(), labels, ignore_index=self.ignore_index, reduction="sum"
+        )
+
+    def forward(
+        self, logits: List[torch.Tensor], labels: torch.Tensor, reduction="sum"
+    ) -> torch.Tensor:
+        """
+        Args:
+            logits (List[torch.Tensor]): List of chunked logits of length
+                ``self.num_output_chunks``, where each chunk has shape
+                ``(batch_size, num_tokens / num_output_chunks, vocab_size)``.
+            labels (torch.Tensor): Ground truth labels of shape ``(batch_size, num_tokens)``.
+            reduction (str): The reduction to apply to the output.
+
+        Returns:
+            torch.Tensor: Cross entropy loss of shape (1,).
+        """
+
+        total_elements = (labels != self.ignore_index).sum()
+
+        # chunk and reshape labels (bsz, num_tokens, vocab) -> [(bsz*num_tokens/num_chunks, vocab)]
+        labels = [
+            target_chunk.reshape(-1)
+            for target_chunk in labels.chunk(self.num_output_chunks, dim=1)
+        ]
+        # reshape logits [(bsz, num_tokens/num_chunks, vocab)] -> [(bsz*num_tokens/num_chunks, vocab)]
+        logits = [
+            logit_chunk.reshape(-1, logit_chunk.size(-1)) for logit_chunk in logits
+        ]
+
+        # compute one chunk at a time
+        total_loss = 0.0
+        for logits_chunk, labels_chunk in zip(logits, labels):
+            total_loss += self.compute_cross_entropy(logits_chunk, labels_chunk)
+
+        if reduction == "sum":
+            return total_loss
+        return total_loss / total_elements
+
+
+def _build_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
+    loss_fn_ce = CEWithChunkedOutputLoss(num_output_chunks, ignore_index)
+    loss_fn_ce.compute_cross_entropy = torch.compile(
+        loss_fn_ce.compute_cross_entropy, backend="inductor"
+    )
+    return loss_fn_ce
+
+
+def get_causal_lm_loss(num_output_chunks: int = 8, ignore_index: int = -100):
+    loss_fn_ce = _build_chunked_ce_loss_fn(num_output_chunks, ignore_index)
+
+    def chunked_fix_cross_entropy(
+        source,
+        target,
+        num_items_in_batch: int = None,
+        ignore_index: int = -100,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        reduction = "sum" if num_items_in_batch is not None else "mean"
+        logit_chunks = [  # pylint: disable=unnecessary-comprehension
+            chunk for chunk in source.chunk(loss_fn_ce.num_output_chunks, dim=1)
+        ]
+        loss = loss_fn_ce(logit_chunks, target, reduction=reduction)
+        if reduction == "sum":
+            loss = loss / num_items_in_batch
+        return loss
+
+    def for_causal_lm_chunked_loss(
+        logits,
+        labels,
+        vocab_size: int = None,  # pylint: disable=unused-argument
+        num_items_in_batch: Optional[int] = None,
+        ignore_index: int = -100,
+        shift_labels: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        # skip the upcast to float since we handle that in the chunking loss
+        if shift_labels is None:
+            # Shift so that tokens < n predict n
+            labels = F.pad(labels, (0, 1), value=ignore_index)
+            shift_labels = labels[..., 1:].contiguous()
+
+        # Skip Flattening the tokens
+        # Enable model parallelism
+        shift_labels = shift_labels.to(logits.device)
+        loss = chunked_fix_cross_entropy(
+            logits, shift_labels, num_items_in_batch, ignore_index, **kwargs
+        )
+        return loss
+
+    return for_causal_lm_chunked_loss
+
+
+def patch_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
+    import transformers.loss.loss_utils
+
+    for_causal_lm_chunked_loss = get_causal_lm_loss(num_output_chunks, ignore_index)
+    transformers.loss.loss_utils.ForCausalLMLoss = for_causal_lm_chunked_loss
+    transformers.loss.loss_utils.LOSS_MAPPING["ForCausalLM"] = (
+        for_causal_lm_chunked_loss
+    )

src/axolotl/monkeypatch/peft/utils.py

@@ -0,0 +1,78 @@
+"""
+Patch prepare_model_for_kbit_training to not upcast everything
+"""
+
+import inspect
+import logging
+
+import peft
+
+import axolotl
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_PREPARE_CODE = """
+        for param in model.parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit":
+                param.data = param.data.to(torch.float32)
+"""
+
+PATCHED_PREPARE_CODE = """
+        for name, param in model.named_parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit" and "norm" in name:
+                param.data = param.data.to(torch.float32)
+"""
+
+
+def get_peft_prep_code() -> str:
+    prepare = inspect.getsource(peft.utils.other.prepare_model_for_kbit_training)
+    return prepare
+
+
+def check_peft_prep_code_is_patchable() -> bool:
+    prep_code = get_peft_prep_code()
+    prep_code, _ = detab_code(prep_code)
+    return ORIGINAL_PREPARE_CODE in prep_code
+
+
+def patch_peft_prep_code():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        prep_code = get_peft_prep_code()
+    except OSError:
+        return
+    peft.utils.other._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        prep_code
+    )
+    prep_code, _ = detab_code(prep_code)
+    if ORIGINAL_PREPARE_CODE not in prep_code:
+        return
+
+    prep_code = prep_code.replace(ORIGINAL_PREPARE_CODE, PATCHED_PREPARE_CODE)
+    prep_code = prep_code.replace(
+        "def prepare_model_for_kbit_training(",
+        "def fixed_prepare_model_for_kbit_training(",
+        1,
+    )
+
+    items_to_import = []
+    for item in dir(peft.utils.other):
+        if item in prep_code:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from peft.utils.other import (" + ", ".join(x for x in items_to_import) + ")",
+        globals(),
+    )
+    exec(prep_code, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching prepare_model_for_kbit_training to allow for overrides")
+    peft.utils.other.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821
+    axolotl.utils.models.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/train.py

@@ -21,6 +21,7 @@ from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from transformers.trainer import Trainer
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
@@ -516,6 +517,8 @@ def train(
     Returns:
         Tuple of (model, tokenizer) after training
     """
+    print_axolotl_text_art()
+
     # Setup model, tokenizer, (causal or RLHF) trainer, etc.
     (
         trainer,

src/axolotl/utils/callbacks/__init__.py

@@ -868,3 +868,29 @@ class GCCallback(TrainerCallback):
     ):
         torch.cuda.empty_cache()
         gc.collect()
+
+
+def colab_inference_post_train_callback(trainer: Trainer):
+    class ColabCallback(TrainerCallback):
+        """Callback to prep model for inference on Google Colab"""
+
+        def __init__(self, cfg):
+            self.gpu_name = torch.cuda.get_device_name(0)
+            self.cfg = cfg
+
+        def on_train_end(
+            self, args, state, control, **kwargs
+        ):  # pylint: disable=unused-argument
+            """
+            handle T4 gpu, we need to convert attention to eager for inference
+            """
+            if "Tesla T4" in self.gpu_name and self.cfg.xformers_attention:
+                trainer.model.eval()
+            trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
+                "eager"
+            )
+            trainer.model.gradient_checkpointing_disable()
+            trainer.model.config.use_cache = True
+            trainer.model.eval()
+
+    return ColabCallback

src/axolotl/utils/config/__init__.py

@@ -70,6 +70,9 @@ def resolve_dtype(cfg):
             if cfg.fp16 is None and not cfg.float16:
                 cfg.fp16 = True
 
+    if cfg.fp16 and cfg.bf16 == "auto":
+        cfg.bf16 = False
+
     if cfg.device == "mps":
         cfg.load_in_8bit = False
         cfg.tf32 = False

src/axolotl/utils/models.py

@@ -556,11 +556,30 @@ class ModelLoader:
         self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
 
     def apply_patches(self) -> None:
+        if self.cfg.xformers_attention and self.cfg.sample_packing:
+            from axolotl.monkeypatch.attention import patch_xformers_attn_over_fa2
+
+            patch_xformers_attn_over_fa2()
+            self.cfg.flash_attention = True
+
+        if self.cfg.chunked_cross_entropy:
+            from axolotl.monkeypatch.loss.chunked import patch_chunked_ce_loss_fn
+
+            if self.cfg.chunked_cross_entropy_num_chunks:
+                patch_chunked_ce_loss_fn(self.cfg.chunked_cross_entropy_num_chunks)
+            else:
+                patch_chunked_ce_loss_fn()
+
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 
             patch_accelerate_fsdp_utils()
 
+        if self.cfg.adapter:
+            from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
+
+            patch_peft_prep_code()
+
         if self.cfg.flex_attention:
             from axolotl.monkeypatch.attention.flex_attn import (
                 patch_flex_make_mask,
@@ -1180,7 +1199,7 @@ class ModelLoader:
                 ],
             )
 
-    def prepare_model(self, qlora_fsdp) -> None:
+    def prepare_model(self, qlora_fsdp: bool) -> None:
         skip_prepare_model_for_kbit_training = False
         if self.cfg.model_config_type == "qwen" and self.cfg.adapter == "lora":
             # Qwen doesn't play nicely with LoRA if this is enabled
@@ -1309,7 +1328,7 @@ class ModelLoader:
 
         # make sure these are fp32 per Ramesh et al. (2021)
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
-        if not self.cfg.fsdp:
+        if self.cfg.fsdp:
             # FSDP doesn't like mixed Float and BFloat16
             self.convert_embedding_modules_dtype(
                 embedding_modules,

src/axolotl/utils/samplers/multipack.py

@@ -1,10 +1,13 @@
-# pylint: skip-file
 """
-Multipack Batch Sampler
+Multipack Batch Sampler - An efficient batch sampler for packing variable-length sequences
+into fixed-capacity batches to optimize memory usage and training throughput.
 """
+
 import logging
 import math
-from typing import Any, Iterable, List, Union
+from concurrent.futures import ProcessPoolExecutor
+from multiprocessing import cpu_count
+from typing import Iterable, List, Union
 
 import numba
 import numpy as np
@@ -13,26 +16,39 @@ from torch.utils.data import BatchSampler, Sampler, SequentialSampler
 from axolotl.utils.distributed import reduce_and_broadcast
 
 LOG = logging.getLogger(__name__)
-
 LOG.setLevel(logging.INFO)
 
 
 @numba.njit
-def ffd_check(a: np.ndarray, c: int, n: int):
-    # First-fit-decreasing bin packing
-    # Check if a[] could fit in n bins with capacity c
-    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
+def ffd_check(sequence_lengths: np.ndarray, bin_capacity: int, num_bins: int):
+    """
+    First-fit-decreasing bin packing algorithm check
 
-    a = np.sort(a)[::-1]
-    bins = np.full((n,), c, dtype=a.dtype)
-    for size in a:
+    Checks if sequences with the given lengths could fit in the specified number of bins
+
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin
+        num_bins: Number of bins available
+
+    Returns:
+        True if all sequences can be packed, False otherwise
+    """
+    # Sort sequence lengths in descending order for optimal packing
+    sequence_lengths = np.sort(sequence_lengths)[::-1]
+    # Initialize all bins with full capacity
+    bins = np.full((num_bins,), bin_capacity, dtype=sequence_lengths.dtype)
+
+    # Try to place each sequence in the first bin it fits
+    for size in sequence_lengths:
         not_found = True
-        for idx in range(n):
+        for idx in range(num_bins):
             if bins[idx] >= size:
                 bins[idx] -= size
                 not_found = False
                 break
 
+        # If no bin could fit this sequence, packing failed
         if not_found:
             return False
 
@@ -40,240 +56,380 @@ def ffd_check(a: np.ndarray, c: int, n: int):
 
 
 @numba.njit
-def ffd_with_result(a: np.ndarray, c: int, start_index: int):
-    # First-fit-decreasing bin packing (with result return)
+def pack_group(
+    sequence_lengths: np.ndarray,
+    group_offset: int,
+    bin_capacity: int,
+    max_bins: int,
+    bin_size: int,
+    safe_mode: bool = True,
+):
+    """
+    Pack a group of sequences into bins using First-Fit Decreasing algorithm
 
-    indices = np.argsort(a)[::-1]
-    a = a[indices]
+    Args:
+        sequence_lengths: Array of sequence lengths
+        group_offset: Offset to apply to indices when returning results
+        bin_capacity: Maximum capacity of each bin
+        max_bins: Maximum number of bins to use
+        bin_size: Maximum number of sequences per bin
+        safe_mode: If True, use a more conservative packing approach
 
-    bins: List[Any] = []
-    bins_result: List[Any] = []
-    for a_id, size in enumerate(a):
-        add_new = True
-        for idx in range(len(bins)):
-            if bins[idx] >= size:
-                bins[idx] -= size
-                bins_result[idx].append(indices[a_id] + start_index)
-                add_new = False
+    Returns:
+        List of bins, where each bin contains indices of sequences assigned to it
+    """
+    # Get sorting indices and sort lengths in descending order
+    indices = np.argsort(sequence_lengths)[::-1]
+    sorted_lengths = sequence_lengths[indices]
+
+    bins_remaining_space: list = []  # Tracks remaining capacity in each bin
+    bins_assigned_sequences: list = []  # Tracks sequence indices assigned to each bin
+
+    for seq_id, size in enumerate(sorted_lengths):
+        global_idx = indices[seq_id] + group_offset
+
+        # Try to place sequence in existing bins
+        add_new_bin = True
+        for bin_idx, _ in enumerate(bins_remaining_space):
+            if (
+                bins_remaining_space[bin_idx] >= size
+                and len(bins_assigned_sequences[bin_idx]) < bin_size
+            ):
+                bins_remaining_space[bin_idx] -= size
+                bins_assigned_sequences[bin_idx].append(global_idx)
+                add_new_bin = False
                 break
 
-        if add_new:
-            bins.append(c - size)
-            bins_result.append([indices[a_id] + start_index])
+        # Create a new bin if needed and if we haven't reached the limit
+        if add_new_bin:
+            if len(bins_remaining_space) >= max_bins and safe_mode:
+                # In safe mode, skip items that would exceed max_bins
+                continue
+            bins_remaining_space.append(bin_capacity - size)
+            bins_assigned_sequences.append([global_idx])
 
-    return bins_result
+            # Safety check to avoid infinite bins
+            if len(bins_remaining_space) > len(sequence_lengths):
+                break
+
+    return bins_assigned_sequences
 
 
-@numba.njit
-def allocate(
-    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
+# Define a standalone function for multiprocessing
+def _process_group(args):
+    group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode = args
+    return pack_group(
+        group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode
+    )
+
+
+def pack_parallel(
+    sequence_lengths: np.ndarray,
+    bin_capacity: int,
+    group_size: int,
+    bin_size: int,
+    num_processes: int | None = None,
+    safe_mode: bool = True,
 ):
-    # Dynamic batch allocator, similar to Multifit
-    # https://en.wikipedia.org/wiki/Multifit_algorithm
-    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
+    """
+    Pack sequences into bins using parallel processing
 
-    s = 0
-    start_index = 0
-    result = []
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin as total number of tokens
+        group_size: Number of sequences to process in each group
+        bin_size: Maximum number of bins to use
+        num_processes: Number of parallel processes to use
+        safe_mode: If True, use a more conservative packing approach
 
-    while True:
-        # binary search [l, r)
-        left = 1
-        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
+    Returns:
+        List of bins, where each bin contains indices of sequences assigned to it
+    """
+    num_items = len(sequence_lengths)
+    if num_processes is None:
+        num_processes = max(1, min(num_items // group_size, cpu_count()))
 
-        while right - left > 1:
-            mid = (left + right) // 2
-            if ffd_check(lengths[start_index : start_index + mid], c, n):
-                left = mid
-            else:
-                right = mid
+    # Create tasks for parallel processing
+    tasks = []
+    for i in range(0, num_items, group_size):
+        group_lengths = sequence_lengths[i : i + group_size]
+        max_bins = len(group_lengths)  # Allow as many bins as items in the group
+        tasks.append((group_lengths, i, bin_capacity, max_bins, bin_size, safe_mode))
 
-        # use length l
-        batch = ffd_with_result(
-            lengths[start_index : start_index + left], c, start_index
-        )
-        assert len(batch) <= n
-        if len(batch) < n:
-            break
+    # Process groups in parallel
+    all_bins = []
+    with ProcessPoolExecutor(max_workers=num_processes) as executor:
+        for group_bins in executor.map(_process_group, tasks):
+            all_bins.extend(group_bins)
 
-        start_index += left
-        s = lengths_cumsum[start_index - 1]
-
-        # add local rank
-        result.append(batch[rank])
-
-    return result, s, len(result) * c * n
+    return all_bins
 
 
 @numba.njit
-def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
+def allocate_sequentially(
+    sequence_lengths: np.ndarray, rank: int, bin_capacity: int, num_ranks: int
+):
     """
     Sequential allocator that preserves example order
 
     Parameters:
-    - lengths: The lengths of all examples
-    - rank: The current rank (for distributed training)
-    - c: The capacity of each bin (maximum sequence length)
-    - n: Number of ranks
+        sequence_lengths: The lengths of all examples
+        rank: The current rank (for distributed training)
+        bin_capacity: The capacity of each bin (maximum sequence length)
+        num_ranks: Number of ranks (processes/GPUs)
 
     Returns:
-    - result: List of batches for the current rank
-    - total_used: Number of actual example tokens
-    - total_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
+        rank_batches: List of batches for the current rank
+        total_tokens_used: Number of actual example tokens
+        total_token_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
     """
-    result = []
-    total_used = 0
+    rank_batches = []
+    total_tokens_used = 0
 
     # First, do sequential packing into bins
     all_bins = []
-    current_bin = [0 for i in range(0)]  # numba hint
-    remaining_capacity = c
+    current_bin = []
+    remaining_capacity = bin_capacity
 
-    for idx, size in enumerate(lengths):
+    # Process each sequence in order
+    for idx, size in enumerate(sequence_lengths):
         if size <= remaining_capacity:
             # Example fits in current bin
             current_bin.append(idx)
             remaining_capacity -= size
-            total_used += size
+            total_tokens_used += size
         else:
             # Example doesn't fit, start a new bin
             if current_bin:  # Add non-empty bin to all_bins
                 all_bins.append(current_bin)
             current_bin = [idx]
-            remaining_capacity = c - size
-            total_used += size
+            remaining_capacity = bin_capacity - size
+            total_tokens_used += size
 
     # Add the last bin if not empty
     if current_bin:
         all_bins.append(current_bin)
 
-    # Assign bins to ranks - each rank gets every n-th bin
-    for bin_idx in range(rank, len(all_bins), n):
-        result.append(all_bins[bin_idx])
+    # Assign bins to ranks - each rank gets every num_ranks-th bin
+    for bin_idx in range(rank, len(all_bins), num_ranks):
+        rank_batches.append(all_bins[bin_idx])
 
-    return result, total_used, len(all_bins) * c
+    return rank_batches, total_tokens_used, len(all_bins) * bin_capacity
 
 
 class MultipackBatchSampler(BatchSampler):
-    """Batch sampler class for multipack"""
+    """
+    Batch sampler class for efficient packing of variable-length sequences
+
+    This sampler packs sequences into fixed-capacity bins (batches) to maximize
+    GPU memory utilization and training throughput by reducing padding.
+
+    It supports both parallel packing (using FFD algorithm) and
+    sequential packing (preserving original sequence order).
+    """
 
     def __init__(
         self,
         sampler: Union[Sampler[int], Iterable[int]],
-        batch_size: int,
-        batch_max_len: int,
-        lengths: np.ndarray,
-        packing_efficiency_estimate: float = 1.0,
-        drop_last: bool = False,
-        num_count_samples: int = 16,
-        sequential: bool = False,
-        **kwargs,
+        batch_size: int,  # Number of bins per batch
+        batch_max_len: int,  # Maximum sequence length (bin capacity)
+        lengths: np.ndarray,  # Sequence lengths
+        packing_efficiency_estimate: float = 1.0,  # Initial efficiency estimate
+        drop_last: bool = False,  # Whether to drop incomplete batches
+        num_count_samples: int = 16,  # Number of samples to estimate batch count
+        sequential: bool = False,  # Whether to use sequential packing
+        group_size: int = 100_000,  # Size of groups for parallel packing
+        bin_size: int = 200,  # The max number of samples that can be packed in a single bin
+        num_processes: int | None = None,  # Number of processes for parallel packing
+        safe_mode: bool = True,  # Conservative packing to prevent training instability
+        **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
         self.batch_size = batch_size
         self.batch_max_len = batch_max_len
-        self.lengths: np.ndarray = lengths
+        self.lengths = np.array(lengths, dtype=np.int32)
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
         self.sequential = sequential
+        self.group_size = group_size
+        self.bin_size = bin_size
+        self.num_processes = num_processes
+        self.safe_mode = safe_mode
 
         assert isinstance(self.lengths, np.ndarray)
 
         self.epoch = 0
 
-        # statistics
-        self.eff_total_used = 0
-        self.eff_total_slots = 0
+        # Efficiency statistics tracking
+        self.total_tokens_used = 0
+        self.total_token_slots = 0
 
-        # The number of times to calculate the batches to determine the minimum packed dataset length for the local rank
+        # The number of times to calculate batches to determine minimum packed dataset length
         self.num_count_samples = num_count_samples
-        # the minimum packed dataset length across all ranks determined by a gather/broadcast
+        # Minimum packed dataset length across all ranks (determined by gather/broadcast)
         self.len_across_ranks = None
 
+        # Cache for batches
+        self._batches = None
+
         if self.sequential and not isinstance(sampler, SequentialSampler):
-            LOG.warn(
+            LOG.warning(
                 "using sequential sample packing with non-sequential sampler, did you want to also enable curriculum_sampling?"
             )
 
     def set_epoch(self, epoch: int):
+        """Set the epoch number, used for reproducible shuffling across epochs"""
         self.epoch = epoch
+        self._batches = None  # Invalidate batch cache
 
     def generate_batches(self, set_stats=False):
-        indices = [idx for idx in self.sampler]
+        """
+        Generate packed batches for training
 
-        lengths = self.lengths[indices]
-        lengths_cumsum = np.cumsum(lengths)
+        Args:
+            set_stats: Whether to update efficiency statistics
 
-        if self.sequential:
-            batches, total_used, total_slots = allocate_sequentially(
-                lengths=lengths,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
-        else:
-            batches, total_used, total_slots = allocate(
-                lengths=lengths,
-                lengths_cumsum=lengths_cumsum,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
+        Returns:
+            List of batches, where each batch contains multiple bins,
+            and each bin contains multiple sequence indices
+        """
+        if self._batches is not None:
+            return self._batches
 
-        batches = [
-            [
-                [indices[b_idx] for b_idx in batch]
-                for batch in batches[i : i + self.batch_size]
-            ]
-            for i in range(0, len(batches), self.batch_size)
+        # Get indices from the sampler
+        indices = [  # pylint: disable=unnecessary-comprehension
+            idx for idx in self.sampler
         ]
 
-        # statistics
-        if set_stats:
-            self.eff_total_used += total_used
-            self.eff_total_slots += total_slots
+        # Get lengths of the selected sequences
+        lengths = self.lengths[indices]
 
+        # Pack sequences into bins using either sequential or parallel packing
+        if self.sequential:
+            bins, total_used, total_slots = allocate_sequentially(
+                lengths,
+                rank=0,
+                bin_capacity=self.batch_max_len,
+                num_ranks=1,
+            )
+        else:
+            # Use parallel packing
+            all_bins = pack_parallel(
+                lengths,
+                bin_capacity=self.batch_max_len,
+                group_size=self.group_size,
+                bin_size=self.bin_size,
+                num_processes=self.num_processes,
+                safe_mode=self.safe_mode,
+            )
+
+            # Map bin indices back to original indices
+            bins = [
+                [indices[b_idx] for b_idx in bin_indices] for bin_indices in all_bins
+            ]
+
+            # Calculate efficiency statistics
+            total_used = lengths.sum()
+            total_slots = len(all_bins) * self.batch_max_len
+
+        # Group bins into batches (each batch contains batch_size bins)
+        batches = [
+            bins[i : i + self.batch_size] for i in range(0, len(bins), self.batch_size)
+        ]
+
+        # Drop last batch if requested and it's incomplete
+        if self.drop_last and len(batches[-1]) < self.batch_size:
+            batches = batches[:-1]
+            # Adjust total_slots if we dropped a batch
+            if not self.sequential:
+                total_slots -= (self.batch_size - len(batches[-1])) * self.batch_max_len
+
+        # Update statistics if requested
+        if set_stats:
+            self.total_tokens_used += total_used
+            self.total_token_slots += total_slots
+
+        self._batches = batches
         return batches
 
     def __iter__(self):
+        """
+        Return an iterator over batches
+
+        The batches are truncated to match the minimum number of batches across all ranks
+        to ensure distributed training balance
+        """
         batches = self.generate_batches(set_stats=True)
         if self.len_across_ranks:
-            # make sure the batches we iterate over is truncated to the same min length across all ranks
+            # Truncate batches to ensure all ranks have the same number of batches
             batches = batches[: self.len_across_ranks]
         return iter(batches)
 
-    def num_batches(self):
-        batches = self.generate_batches(set_stats=True)
-        return len(batches)
-
     def efficiency(self):
-        return self.eff_total_used / self.eff_total_slots
+        """
+        Calculate the packing efficiency (ratio of tokens used to total token slots)
+        Higher is better - 1.0 would mean perfect packing with no wasted space
+        """
+        if self.total_token_slots == 0:
+            self.generate_batches(set_stats=True)
+        if self.total_token_slots == 0:
+            return 0.0
+        # Return a Python float instead of potentially a numpy float
+        return float(self.total_tokens_used / self.total_token_slots)
 
     def gather_efficiency(self):
+        """
+        Gather and synchronize packing efficiency estimates across all distributed ranks
+        Returns a conservative efficiency estimate based on the measurements
+        """
+
         def calc_sample_packing_eff_est(estimates: List[float]):
             LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
-            return math.floor(0.997 * max(estimates))
+            # Use 99.7% of max observed efficiency as a safe estimate
+            max_eff = max(float(eff) for eff in estimates)
+            return math.floor(0.997 * max_eff)
 
+        # Gather efficiency from all ranks and apply the calculation function
         sample_packing_actual_eff_all = reduce_and_broadcast(
-            lambda: self.efficiency(),  # pylint: disable=unnecessary-lambda
+            lambda: float(self.efficiency()),  # pylint: disable=unnecessary-lambda
             calc_sample_packing_eff_est,
         )
+
+        # Quantize to 0.5% intervals for stability
         sample_packing_eff_est = (
             math.ceil(sample_packing_actual_eff_all * 200.0) / 200.0
         )
         return sample_packing_eff_est
 
     def gather_len_batches(self, num):
+        """
+        Gather and synchronize batch counts across all distributed ranks
+        Returns the minimum number of batches available on any rank
+        """
+
         def calc_min_len(estimates: list[(int, float)]):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(min(estimates))
 
+        # Find minimum batch count across ranks to ensure balance
         min_len_batches = reduce_and_broadcast(lambda: num, calc_min_len)
         return min_len_batches
 
     def __len__(self):
-        if not self.len_across_ranks:
-            len_batches = min(
-                [self.num_batches() for _ in range(self.num_count_samples)]
+        """
+        Return the total number of batches that will be yielded by this sampler
+
+        This is calculated as the minimum number of batches available on any rank
+        to ensure balanced distributed training
+        """
+        if self._batches is None:
+            self._batches = self.generate_batches(set_stats=True)
+
+        if self.len_across_ranks is None:
+            # Sample multiple times to get stable estimate
+            len_batches = min(  # pylint: disable=consider-using-generator
+                [len(self._batches) for _ in range(self.num_count_samples)]
             )
+            # Gather minimum across all ranks
             self.len_across_ranks = self.gather_len_batches(len_batches)
+
         return self.len_across_ranks

src/axolotl/utils/schemas/config.py

@@ -242,6 +242,9 @@ class AxolotlInputConfig(
     unsloth_rms_norm: bool | None = None
     unsloth_rope: bool | None = None
 
+    chunked_cross_entropy: bool | None = None
+    chunked_cross_entropy_num_chunks: int | None = None
+
     lora_mlp_kernel: bool | None = None
     lora_qkv_kernel: bool | None = None
     lora_o_kernel: bool | None = None
@@ -435,16 +438,6 @@ class AxolotlInputConfig(
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_sample_packing_w_xformers(cls, data):
-        if data.get("sample_packing") and data.get("xformers_attention"):
-            raise ValueError(
-                "sample_packing not compatible with xformers_attention. Use flash_attention"
-            )
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     # pylint: disable=duplicate-code

tests/test_chunked_xentropy.py

@@ -0,0 +1,40 @@
+"""
+test suite for chunked cross entropy
+"""
+
+import pytest
+import torch
+from torch import nn
+
+from axolotl.monkeypatch.loss.chunked import get_causal_lm_loss
+
+
+@pytest.fixture
+def chunked_fixtures():
+    model_dim = 512
+    vocab_size = 1024 * 256
+    seq_len = 2048
+    batch_size = 1
+
+    lm_head = nn.Linear(model_dim, vocab_size)
+    hidden_state = torch.randn(batch_size, seq_len, model_dim)
+    labels = torch.randint(low=0, high=vocab_size, size=(batch_size, seq_len))
+    return lm_head, hidden_state, labels, vocab_size
+
+
+def test_chunked_forward(chunked_fixtures):  # pylint: disable=redefined-outer-name
+    lm_head, hidden_state, labels, vocab_size = chunked_fixtures
+    lm_loss = get_causal_lm_loss()
+
+    logits = lm_head(hidden_state)
+
+    chunked_lm_loss = lm_loss(logits, labels)
+
+    logits_flattened = logits.view(-1, vocab_size)
+    labels_flattened = labels.view(-1)
+
+    loss = nn.functional.cross_entropy(
+        logits_flattened.float(), labels_flattened, reduction="mean"
+    )
+
+    assert torch.allclose(chunked_lm_loss, loss, atol=1e-2, rtol=1e-2)