liger-kernel-nightly 0.4.1.dev20241115191733__tar.gz → 0.4.1.dev20241117192031__tar.gz

{liger_kernel_nightly-0.4.1.dev20241115191733/src/liger_kernel_nightly.egg-info → liger_kernel_nightly-0.4.1.dev20241117192031}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.1.dev20241115191733
+Version: 0.4.1.dev20241117192031
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.1.dev20241115191733 → liger_kernel_nightly-0.4.1.dev20241117192031}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.4.1.dev20241115191733"
+version = "0.4.1.dev20241117192031"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.4.1.dev20241115191733 → liger_kernel_nightly-0.4.1.dev20241117192031}/src/liger_kernel/transformers/model/qwen2_vl.py

@@ -1,7 +1,9 @@
 from typing import List, Optional, Tuple, 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,
@@ -80,8 +82,6 @@ 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 ..."
     ```"""
-    # FIXME: The code is outdated and not compatible with transformer >= 4.46.1
-
     output_attentions = (
         output_attentions
         if output_attentions is not None
@@ -100,27 +100,53 @@ def lce_forward(
         inputs_embeds = self.model.embed_tokens(input_ids)
         if pixel_values is not None:
             pixel_values = pixel_values.type(self.visual.get_dtype())
-            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw).to(
-                inputs_embeds.device
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_mask = (
+                (input_ids == self.config.image_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
             )
-            image_mask = input_ids == self.config.image_token_id
-            if self.training:
-                inputs_embeds = inputs_embeds.clone()
-            inputs_embeds[image_mask] = image_embeds
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
         if pixel_values_videos is not None:
             pixel_values_videos = pixel_values_videos.type(self.visual.get_dtype())
-            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw).to(
-                inputs_embeds.device
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+            video_mask = (
+                (input_ids == self.config.video_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
             )
-            video_mask = input_ids == self.config.video_token_id
-            inputs_embeds[video_mask] = video_embeds
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
+
         if attention_mask is not None:
             attention_mask = attention_mask.to(inputs_embeds.device)
-    # The code is copied from https://github.com/huggingface/transformers/pull/33487
-    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 version.parse(transformers_version) > version.parse("4.46.2"):
+        # 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
+            )
 
     outputs = self.model(
         input_ids=None,

{liger_kernel_nightly-0.4.1.dev20241115191733 → liger_kernel_nightly-0.4.1.dev20241117192031}/src/liger_kernel/transformers/monkey_patch.py

@@ -56,12 +56,15 @@ def _bind_method_to_module(module, method_name: str, new_method: Callable):
     module.__dict__[method_name] = new_method.__get__(module, module.__class__)
 
 
-def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama"):
+def _patch_rms_norm_module(
+    module, offset=0.0, eps=1e-6, casting_mode="llama", in_place=True
+):
     module.offset = offset
     module.casting_mode = casting_mode
     module.variance_epsilon = (
         getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
     )
+    module.in_place = in_place
     _bind_method_to_module(module, "forward", LigerRMSNorm.forward)
     _bind_method_to_module(module, "extra_repr", LigerRMSNorm.extra_repr)
 
@@ -510,7 +513,7 @@ def apply_liger_kernel_to_gemma2(
         LigerRMSNorm, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False
     )
     _patch_rms_norm_module_for_gemma2 = partial(
-        _patch_rms_norm_module, offset=1.0, casting_mode="gemma"
+        _patch_rms_norm_module, offset=1.0, casting_mode="gemma", in_place=False
     )
 
     if rope:

{liger_kernel_nightly-0.4.1.dev20241115191733 → liger_kernel_nightly-0.4.1.dev20241117192031/src/liger_kernel_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.1.dev20241115191733
+Version: 0.4.1.dev20241117192031
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation