liger-kernel-nightly 0.4.0.dev20241107052928__py3-none-any.whl → 0.6.3.dev20251121010306__py3-none-any.whl

liger_kernel/transformers/model/llama4.py

@@ -0,0 +1,121 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.cache_utils import Cache
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerCausalLMOutputWithPast
+
+
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs,
+) -> Union[Tuple, LigerCausalLMOutputWithPast]:
+    r"""
+    labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+        Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+        config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+        (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Llama4ForCausalLM
+
+    >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=True,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+
+    # Compute loss
+    if self.training and (labels is not None or shift_labels is not None):
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = ((loss,) + output) if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/model/llava.py

@@ -0,0 +1,344 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.nn import CrossEntropyLoss
+from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast
+from transformers.utils import is_torchdynamo_compiling
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerLlavaCausalLMOutputWithPast
+
+
+def lce_forward_deprecated(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[int] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: torch.Tensor = None,
+    skip_logits: Optional[bool] = None,
+    **lm_kwargs,
+) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+    >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
+    >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+    >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "USER:  \nWhat's the content of the image? ASSISTANT: The image features a busy city street with a stop sign prominently displayed"
+    ```"""
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+    vision_feature_layer = (
+        vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+    )
+    vision_feature_select_strategy = (
+        vision_feature_select_strategy
+        if vision_feature_select_strategy is not None
+        else self.config.vision_feature_select_strategy
+    )
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    if pixel_values is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+        )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+
+    if pixel_values is not None:
+        image_features = self.get_image_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            image_sizes=image_sizes,
+        )
+
+        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+        if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+            n_image_tokens = (input_ids == self.config.image_token_index).sum()
+            n_image_features = image_features.shape[0] * image_features.shape[1]
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+    outputs = self.language_model.model(
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        **lm_kwargs,
+    )
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    # Overwrite skip_logits, since llava never materializes logits
+    skip_logits = labels is not None
+
+    if skip_logits:
+        # Shift so that tokens < n predict n
+        if attention_mask is not None:
+            # we use the input attention mask to shift the logits and labels, because it is 2D.
+            # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+            shift_attention_mask = attention_mask[:, -(hidden_states.shape[1] - 1) :].to(hidden_states.device)
+            shift_hidden_states = hidden_states[..., :-1, :][
+                shift_attention_mask.to(hidden_states.device) != 0
+            ].contiguous()
+            shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+        else:
+            shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+
+        lce = LigerFusedLinearCrossEntropyLoss()
+        loss = lce(
+            self.language_model.lm_head.weight,
+            shift_hidden_states.view(-1, shift_hidden_states.size(-1)),
+            shift_labels.view(-1).to(shift_hidden_states.device),
+        )
+    else:
+        logits = self.language_model.lm_head(hidden_states)
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            shift_labels = labels[..., 1:]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+
+    if not return_dict:
+        # NOTE: This part has not been tested.
+        output = outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return LlavaCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[int] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: torch.Tensor = None,
+    skip_logits: Optional[bool] = None,
+    **lm_kwargs,
+) -> Union[Tuple, LigerLlavaCausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+    >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
+    >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+    >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "USER:  \nWhat's the content of the image? ASSISTANT: The image features a busy city street with a stop sign prominently displayed"
+    ```"""
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+    vision_feature_layer = (
+        vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
+    )
+    vision_feature_select_strategy = (
+        vision_feature_select_strategy
+        if vision_feature_select_strategy is not None
+        else self.config.vision_feature_select_strategy
+    )
+
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        vision_feature_layer=vision_feature_layer,
+        vision_feature_select_strategy=vision_feature_select_strategy,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=True,
+        cache_position=cache_position,
+        image_sizes=image_sizes,
+        **lm_kwargs,
+    )
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = lm_kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    token_accuracy = None
+
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        result = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.text_config.hidden_size,
+            **lm_kwargs,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        output = (loss,) + output if loss is not None else output
+        output = output + (token_accuracy,) if token_accuracy is not None else output
+        return output
+
+    # Return custom output class with token_accuracy field
+    return LigerLlavaCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=outputs.image_hidden_states,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/model/loss_utils.py

@@ -0,0 +1,95 @@
+from typing import Optional
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+import liger_kernel.transformers.functional as F
+
+from liger_kernel.transformers.functional import CrossEntropyOutput
+
+
+def unpack_cross_entropy_result(
+    result,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+    if isinstance(result, CrossEntropyOutput):
+        return result.loss, result.z_loss, result.token_accuracy
+
+    if isinstance(result, tuple):
+        loss = result[0]
+        z_loss = result[1] if len(result) > 1 else None
+        token_accuracy = result[2] if len(result) > 2 else None
+        return loss, z_loss, token_accuracy
+
+    return result, None, None
+
+
+def fixed_fused_linear_cross_entropy(
+    hidden_states: torch.Tensor,
+    lm_head_weight: torch.Tensor,
+    target: torch.Tensor,
+    num_items_in_batch: Optional[int] = None,
+    ignore_index: int = -100,
+    final_logit_softcapping: Optional[float] = None,
+    accum_dtype: Optional[torch.dtype] = None,
+    return_token_accuracy: bool = False,
+    **kwargs,
+):
+    reduction = "sum" if num_items_in_batch is not None else "mean"
+    result = F.liger_fused_linear_cross_entropy(
+        hidden_states,
+        lm_head_weight,
+        target,
+        reduction=reduction,
+        ignore_index=ignore_index,
+        softcap=final_logit_softcapping,
+        accum_dtype=accum_dtype,
+        return_token_accuracy=return_token_accuracy,
+        **kwargs,
+    )
+
+    loss, _, token_accuracy = unpack_cross_entropy_result(result)
+
+    if reduction == "sum":
+        loss = loss / num_items_in_batch
+
+    if return_token_accuracy:
+        return CrossEntropyOutput(loss=loss, token_accuracy=token_accuracy)
+
+    return loss
+
+
+def LigerForCausalLMLoss(
+    hidden_states,
+    lm_head_weight,
+    labels,
+    hidden_size: int,
+    num_items_in_batch: Optional[int] = None,
+    ignore_index: int = -100,
+    shift_labels: Optional[torch.Tensor] = None,
+    final_logit_softcapping: Optional[float] = None,
+    return_token_accuracy: bool = False,
+    **kwargs,
+):
+    # Skip upcast since intermediate values for the loss are all fp32 in kernel
+    if shift_labels is None:
+        # Shift so that token < n predict n
+        labels = nn.functional.pad(labels, (0, 1), value=ignore_index)
+        shift_labels = labels[..., 1:].contiguous()
+
+    # Flatten the tokens
+    hidden_states = hidden_states.view(-1, hidden_size)
+    shift_labels = shift_labels.view(-1)
+    # Enable model parallelism
+    shift_labels = shift_labels.to(hidden_states.device)
+    result = fixed_fused_linear_cross_entropy(
+        hidden_states,
+        lm_head_weight,
+        shift_labels,
+        num_items_in_batch,
+        ignore_index,
+        final_logit_softcapping,
+        return_token_accuracy=return_token_accuracy,
+        **kwargs,
+    )
+    return result