liger-kernel 0.5.2__py3-none-any.whl → 0.5.4__py3-none-any.whl

liger_kernel/transformers/model/mllama.py

@@ -1,24 +1,22 @@
-from typing import List, Optional, Tuple, Union
+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.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.mllama.modeling_mllama import MLLAMA_INPUTS_DOCSTRING
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
 
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 
 
 @add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig"
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -66,19 +64,11 @@ def lce_forward_deprecated(
     I love the idea of snowflakes gently falling, each one
     ```
     """
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
+    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
+        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(
@@ -143,9 +133,7 @@ def lce_forward_deprecated(
 
 
 @add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig"
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig")
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -198,19 +186,11 @@ def lce_forward(
     I love the idea of snowflakes gently falling, each one
     ```
     """
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
+    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
+        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(

liger_kernel/transformers/model/olmo2.py

@@ -0,0 +1,124 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.olmo2.modeling_olmo2 import _CONFIG_FOR_DOC
+from transformers.models.olmo2.modeling_olmo2 import OLMO2_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+
+
+@add_start_docstrings_to_model_forward(OLMO2_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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[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,
+    num_logits_to_keep: int = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    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]`.
+
+        num_logits_to_keep (`int`, *optional*):
+            Calculate logits for the last `num_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.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Olmo2ForCausalLM
+
+    >>> model = Olmo2ForCausalLM.from_pretrained("allenai/Olmo2-1B-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo2-1B-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 sure if you’re conscious of this, but I’m'
+    ```
+    """
+    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=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        # We do the same thing as ForCausalLMLoss but using Liger FLCE
+
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        reduction = "sum" if "num_items_in_batch" in loss_kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            loss /= loss_kwargs["num_items_in_batch"]
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

liger_kernel/transformers/model/phi3.py

@@ -1,26 +1,22 @@
-from typing import List, Optional, Tuple, Union
+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.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.phi3.modeling_phi3 import (
-    _CONFIG_FOR_DOC,
-    PHI3_INPUTS_DOCSTRING,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
+from transformers.models.phi3.modeling_phi3 import _CONFIG_FOR_DOC
+from transformers.models.phi3.modeling_phi3 import PHI3_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
 
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 
 
 @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -64,19 +60,11 @@ def lce_forward_deprecated(
     'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
     ```"""
 
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
+    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
+        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(
@@ -138,9 +126,7 @@ def lce_forward_deprecated(
 
 
 @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -202,19 +188,11 @@ def lce_forward(
             f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed."
         )
 
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
+    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
+        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(

liger_kernel/transformers/model/qwen2.py

@@ -1,26 +1,22 @@
-from typing import List, Optional, Tuple, Union
+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.modeling_outputs import CausalLMOutputWithPast
-from transformers.models.qwen2.modeling_qwen2 import (
-    _CONFIG_FOR_DOC,
-    QWEN2_INPUTS_DOCSTRING,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
+from transformers.models.qwen2.modeling_qwen2 import _CONFIG_FOR_DOC
+from transformers.models.qwen2.modeling_qwen2 import QWEN2_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
 
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 
 
 @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
@@ -63,19 +59,11 @@ def lce_forward_deprecated(
     >>> 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_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
+        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(
@@ -137,9 +125,7 @@ def lce_forward_deprecated(
 
 
 @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -187,19 +173,11 @@ def lce_forward(
     "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_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
+        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(

liger_kernel/transformers/model/qwen2_vl.py

@@ -1,28 +1,24 @@
-from typing import List, Optional, Tuple, Union
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
 
 import torch
+
 from packaging import version
 from torch.nn import CrossEntropyLoss
 from transformers import __version__ as transformers_version
-from transformers.models.qwen2_vl.modeling_qwen2_vl import (
-    _CONFIG_FOR_DOC,
-    QWEN2_VL_INPUTS_DOCSTRING,
-    Qwen2VLCausalLMOutputWithPast,
-)
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
+from transformers.models.qwen2_vl.modeling_qwen2_vl import _CONFIG_FOR_DOC
+from transformers.models.qwen2_vl.modeling_qwen2_vl import QWEN2_VL_INPUTS_DOCSTRING
+from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLCausalLMOutputWithPast
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
 
-from liger_kernel.transformers.fused_linear_cross_entropy import (
-    LigerFusedLinearCrossEntropyLoss,
-)
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
 
 
 @add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
+@replace_return_docstrings(output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,
@@ -40,6 +36,7 @@ def lce_forward(
     image_grid_thw: Optional[torch.LongTensor] = None,
     video_grid_thw: Optional[torch.LongTensor] = None,
     rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
 ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
     r"""
     Copy paste Qwen2VL's forward but replace torch cross entropy with liger fused linear cross entropy
@@ -82,19 +79,11 @@ def lce_forward(
     >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
     "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
     ```"""
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
+    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
+        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
 
     if inputs_embeds is None:
         inputs_embeds = self.model.embed_tokens(input_ids)
@@ -137,16 +126,30 @@ def lce_forward(
         if attention_mask is not None:
             attention_mask = attention_mask.to(inputs_embeds.device)
 
-    if version.parse(transformers_version) > version.parse("4.46.2"):
+    if version.parse(transformers_version) > version.parse("4.46.3"):
         # NOTE: this bug fix for qwen2-vl is not applied until transformers 4.47.0
         # https://github.com/huggingface/transformers/issues/33401
         # While correct, this breaks equivalence with past versions of Qwen2-VL from
         # transformers and leads to failed tests or users noticing differences in results.
         # TODO: remove above conditional when liger drops support for transformers<4.47.0
-        if position_ids is None and input_ids is not None:
-            position_ids, _ = self.get_rope_index(
-                input_ids, image_grid_thw, video_grid_thw, attention_mask
-            )
+        # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+        if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+            # calculate RoPE index once per generation in the pre-fill stage only
+            if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
+                position_ids, rope_deltas = self.get_rope_index(
+                    input_ids, image_grid_thw, video_grid_thw, attention_mask
+                )
+                self.rope_deltas = rope_deltas
+            # then use the prev pre-calculated rope-deltas to get the correct position ids
+            else:
+                batch_size, seq_length, _ = inputs_embeds.shape
+                delta = cache_position[0] + self.rope_deltas if cache_position is not None else 0
+                position_ids = torch.arange(seq_length, device=inputs_embeds.device)
+                position_ids = position_ids.view(1, -1).expand(batch_size, -1)
+                if cache_position is not None:  # otherwise `deltas` is an int `0`
+                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+                position_ids = position_ids.add(delta)
+                position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
 
     outputs = self.model(
         input_ids=None,
@@ -158,6 +161,7 @@ def lce_forward(
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
+        cache_position=cache_position,
     )
 
     hidden_states = outputs[0]