liger-kernel-nightly 0.3.1.dev20241031001941__py3-none-any.whl → 0.3.1.dev20241031214104__py3-none-any.whl

liger_kernel/transformers/model/llama.py

@@ -22,7 +22,7 @@ from liger_kernel.transformers.fused_linear_cross_entropy import (
 @replace_return_docstrings(
     output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
 )
-def lce_forward(
+def lce_forward_deprecated(
     self,
     input_ids: torch.LongTensor = None,
     attention_mask: Optional[torch.Tensor] = None,
@@ -145,3 +145,137 @@ def lce_forward(
         hidden_states=outputs.hidden_states,
         attentions=outputs.attentions,
     )
+
+
+@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
+def lce_forward(
+    self,
+    input_ids=None,
+    attention_mask=None,
+    position_ids=None,
+    past_key_values=None,
+    inputs_embeds=None,
+    labels=None,
+    use_cache=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    return_dict=None,
+    cache_position=None,
+    num_logits_to_keep=0,
+    **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, LlamaForCausalLM
+
+    >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-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."
+    ```"""
+
+    """
+    My TODO:
+    1. Run e2e example with hf < 4.46.0, GA on/GA off
+    2. Run e2e example with hf >= 4.46.0, GA on/GA off
+    """
+
+    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,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+
+    if self.config.pretraining_tp > 1:
+        raise Exception("Liger Kernel does not support pretraining_tp!!")
+
+    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 kwargs else "mean"
+        lce = LigerFusedLinearCrossEntropyLoss(reduction=reduction)
+
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+        if reduction == "sum":
+            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,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -3,6 +3,8 @@ import logging
 from functools import partial
 from typing import Callable
 
+import transformers
+from packaging import version
 from transformers import PreTrainedModel
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
@@ -10,6 +12,9 @@ from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.layer_norm import LigerLayerNorm
 from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
+from liger_kernel.transformers.model.llama import (
+    lce_forward_deprecated as llama_lce_forward_deprecated,
+)
 from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_forward
 from liger_kernel.transformers.model.mixtral import lce_forward as mixtral_lce_forward
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
@@ -22,6 +27,8 @@ from liger_kernel.transformers.swiglu import (
     LigerSwiGLUMLP,
 )
 
+transformer_version = version.parse(transformers.__version__)
+
 logger = logging.getLogger(__name__)
 
 
@@ -88,7 +95,14 @@ def apply_liger_kernel_to_llama(
     if cross_entropy:
         modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
     if fused_linear_cross_entropy:
-        modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+        if transformer_version >= version.parse("4.46.0"):
+            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+        else:  # if version < 4.46.0
+            logger.warning(
+                "Support for transformers versions < 4.46.0 will soon be discontinued due to issues with incorrect gradient accumulation. "
+                "Please consider upgrading to avoid potential issues. See details: https://github.com/huggingface/transformers/pull/34191"
+            )
+            modeling_llama.LlamaForCausalLM.forward = llama_lce_forward_deprecated
 
     if model is not None:
         # The model instance already exists, so we need to additionally patch the

{liger_kernel_nightly-0.3.1.dev20241031001941.dist-info → liger_kernel_nightly-0.3.1.dev20241031214104.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.3.1.dev20241031001941
+Version: 0.3.1.dev20241031214104
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -31,7 +31,7 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 License-File: LICENSE-Apache-2.0
 License-File: LICENSE-MIT-AutoAWQ
-License-File: LICENSE-MIT-Efficient Cross Entropy
+License-File: LICENSE-MIT-Efficient-Cross-Entropy
 License-File: LICENSE-MIT-llmc
 License-File: LICENSE-MIT-triton
 License-File: NOTICE

{liger_kernel_nightly-0.3.1.dev20241031001941.dist-info → liger_kernel_nightly-0.3.1.dev20241031214104.dist-info}/RECORD

@@ -23,7 +23,7 @@ liger_kernel/transformers/geglu.py,sha256=QcrME_8ooIn0xa59LaC0aoOdRrBIFd11Y0bAyF
 liger_kernel/transformers/jsd.py,sha256=W-5CypO2mx4-bUWOxq1KScfCdoXlLoYbtt5xBnRzMs4,3056
 liger_kernel/transformers/kl_div.py,sha256=qVhjBg6tjRyue5iZ3NFxo8uySY4JuIFJyv0IM_50F24,431
 liger_kernel/transformers/layer_norm.py,sha256=fd6o4kSHJWolQMWxh-l1qObfgL08ruNbUoBiANKX1ow,972
-liger_kernel/transformers/monkey_patch.py,sha256=Tr2m4QlFp_qQuARA6cUalFA60SVHttIvsz1qD4aK0oI,32876
+liger_kernel/transformers/monkey_patch.py,sha256=f8Mm3LaBB2NehbLyEK3kz1rE4u98QJna9OyM2XAy6NI,33607
 liger_kernel/transformers/rms_norm.py,sha256=4XfMQI6dORF7s_5qUqVHKWv-3IUomaimU2dg-NwnpoM,1035
 liger_kernel/transformers/rope.py,sha256=m-ah8vZBYW8tfplTXCiAPMHJWlB1tdp_JPXJeWE-Boo,943
 liger_kernel/transformers/swiglu.py,sha256=0-tVJ8xEYfhxnduc16PflXFj8sZPxdx9sHUn3hfwCI4,2468
@@ -31,7 +31,7 @@ liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx
 liger_kernel/transformers/experimental/embedding.py,sha256=HpckiAMKM8-SRxKDcGTqortVxnjhwpZsfsp9lfjqfeM,895
 liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/transformers/model/gemma.py,sha256=EcdkGbSj_qroTDFl0Sc_HLyDyY0xcDhwrgkM_wkXnw8,4987
-liger_kernel/transformers/model/llama.py,sha256=cKZp1bAMIBagfbW4GFbwXBVaDKq92zBRnkGH-rk3VoA,5424
+liger_kernel/transformers/model/llama.py,sha256=B_4mt7m2_nDGpa2bZymjshbF4sAVJWTn5pho86uGTZo,10167
 liger_kernel/transformers/model/mistral.py,sha256=_MQJrDntlxBO5cJwgTjr2rk2nNd5FAXVnzcTg_PEekQ,5079
 liger_kernel/transformers/model/mixtral.py,sha256=51FghRY8aGBWat7KSgTeFDqdStDiXY3dEJepByNhEOE,5847
 liger_kernel/transformers/model/mllama.py,sha256=S00P0pJrGHOWBx170TPYZbQ0djv0__m8Dqv1FvKZUyE,5926
@@ -40,14 +40,14 @@ liger_kernel/transformers/model/qwen2.py,sha256=3inWFXGHYT7wA10OR6bq3mDUBrr10AS5
 liger_kernel/transformers/model/qwen2_vl.py,sha256=ymsm9aQpSUiSU12GY8FO608p9dSHOz4TCnNI1htX5bk,6975
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE-Apache-2.0,sha256=NRaCIsL9eblGS35gk4WKTC0usNYnR_mgRHJTKqz2_UE,11348
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE-MIT-AutoAWQ,sha256=pfiOyInrAPY3xQbvV1i-gOqNZK7QEyIepT1IbqOYYYo,1067
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE-MIT-Efficient Cross Entropy,sha256=PaC9HqyFYTy-ClS0H8Zfa2motJuTppjECXmjHwJcaOk,1063
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE-MIT-llmc,sha256=kyFLt_XUcXS88CuxQt5-PjOcLjpJP2m-T4gtqZf3GLc,1071
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/LICENSE-MIT-triton,sha256=wL6W8IwsKiyHtzXubg8TCXhRZuo8S83EPdqXffYtqWg,1131
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/METADATA,sha256=-QULBKpe1DszKDOgeTyHreRY6sZwvjhaJ9GLkAANDwU,27717
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.3.1.dev20241031001941.dist-info/RECORD,,
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE-Apache-2.0,sha256=NRaCIsL9eblGS35gk4WKTC0usNYnR_mgRHJTKqz2_UE,11348
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE-MIT-AutoAWQ,sha256=pfiOyInrAPY3xQbvV1i-gOqNZK7QEyIepT1IbqOYYYo,1067
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE-MIT-Efficient-Cross-Entropy,sha256=PaC9HqyFYTy-ClS0H8Zfa2motJuTppjECXmjHwJcaOk,1063
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE-MIT-llmc,sha256=kyFLt_XUcXS88CuxQt5-PjOcLjpJP2m-T4gtqZf3GLc,1071
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/LICENSE-MIT-triton,sha256=wL6W8IwsKiyHtzXubg8TCXhRZuo8S83EPdqXffYtqWg,1131
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/METADATA,sha256=wnIQKAkgub7YGk78qo0YfkPlYoWGAcLUWPd36sBeHY8,27717
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.3.1.dev20241031214104.dist-info/RECORD,,