liger-kernel-nightly 0.5.10.dev20250526154218__py3-none-any.whl → 0.5.10.dev20250528223223__py3-none-any.whl

liger_kernel/transformers/model/llava.py

@@ -5,6 +5,7 @@ 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 transformers.utils.deprecation import deprecate_kwarg
@@ -189,7 +190,20 @@ def lce_forward_deprecated(
 
         lce = LigerFusedLinearCrossEntropyLoss()
         loss = lce(self.language_model.lm_head.weight, shift_hidden_states, shift_labels)
-
+    else:
+        logits = self.language_model.lm_head(hidden_states)
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                shift_attention_mask = attention_mask[..., 1:]
+                shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device))
     if not return_dict:
         # NOTE: This part has not been tested.
         output = outputs[1:]
@@ -349,6 +363,28 @@ def lce_forward(
             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.

liger_kernel/transformers/monkey_patch.py

@@ -776,7 +776,7 @@ def apply_liger_kernel_to_gemma3_text(
 
     from transformers.models.gemma3 import modeling_gemma3
     from transformers.models.gemma3.modeling_gemma3 import Gemma3DecoderLayer
-    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM, Gemma3TextModel
 
     from liger_kernel.transformers.gema3_rms import LigerRMSNormForGemma3
     from liger_kernel.transformers.model.gemma3 import causal_forward
@@ -807,9 +807,9 @@ def apply_liger_kernel_to_gemma3_text(
         # The model instance already exists, so we need to additionally patch the
         # instance variables that reference already-instantiated modules
 
-        if isinstance(model, Gemma3ForCausalLM):
+        if isinstance(model, Gemma3ForCausalLM) or isinstance(model, Gemma3TextModel):
             # get the base model from the model instance
-            base_model = model.model
+            base_model = model.model if isinstance(model, Gemma3ForCausalLM) else model
 
             if rms_norm:
                 _patch_rms_norm_module_for_gemma3(base_model.norm)
@@ -1625,7 +1625,6 @@ def _apply_liger_kernel_to_instance(model: PreTrainedModel, **kwargs) -> None:
         return
 
     apply_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[model_type]
-
     apply_fn_signature = inspect.signature(apply_fn)
 
     # Filter out the keyword arguments that are not supported by the apply function

{liger_kernel_nightly-0.5.10.dev20250526154218.dist-info → liger_kernel_nightly-0.5.10.dev20250528223223.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250526154218
+Version: 0.5.10.dev20250528223223
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.10.dev20250526154218.dist-info → liger_kernel_nightly-0.5.10.dev20250528223223.dist-info}/RECORD

@@ -52,7 +52,7 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=DKv5-4KyXLiVhAJ9WVFv1I1i1DzjaudTrhqx6EVYViU,74505
+liger_kernel/transformers/monkey_patch.py,sha256=a0CXSC8BwZg3vok-ns0udZLUOBkegGQgPDod3H8ilP4,74610
 liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=GqCEJuGt0YdqqlMcToE0Wp4A8YFquDa4UUSyH2uFW2A,1191
@@ -69,7 +69,7 @@ liger_kernel/transformers/model/gemma2.py,sha256=JxPTXVkuFtiaZYkaBM8bZF-ObyatHmA
 liger_kernel/transformers/model/gemma3.py,sha256=JI4jj9K660HeRsofB6cpkCHBQ0OsazElArRtKUehUmw,15945
 liger_kernel/transformers/model/glm4.py,sha256=3YJiGdZ0nNSdZidPFlXdUad8mlFwyfq44yd11OcdNns,5259
 liger_kernel/transformers/model/llama.py,sha256=cAWTCY0bk67lFXNtAVEXIWl9WNgn4JyU25Q7nhpKjE0,12505
-liger_kernel/transformers/model/llava.py,sha256=Ju2vGwpL3fMyb_wXwJtvPVddZKOUlMpZaI7aGMCDjXI,17126
+liger_kernel/transformers/model/llava.py,sha256=ONdpx96AVbbL8QDQvHSm08jMJPz3tzkbeO92IRbAb1A,19270
 liger_kernel/transformers/model/loss_utils.py,sha256=WWAMdiONPaXpIvxyOim_0igLrYh0yyOok5Q9_L9xvZw,1787
 liger_kernel/transformers/model/mistral.py,sha256=vFFZD5VAwpx6Bs4gXoXDRmyU9-7Dp50w3jIcj0q0sIo,5567
 liger_kernel/transformers/model/mixtral.py,sha256=vSmgBc91WMu9_iWkAHUJPzo0-WDkTJK5SEVYNaDRT_Y,11398
@@ -86,9 +86,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/METADATA,sha256=Htje8soAZU2663YooXwqv7WjHOGITbkEP-o5qwGXnxg,24113
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.10.dev20250526154218.dist-info/RECORD,,
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/METADATA,sha256=peyDncCLhsNKI0sXe4Fg-cjTiGK_5NFaM7vdiRwjaZY,24113
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.10.dev20250528223223.dist-info/RECORD,,