liger-kernel-nightly 0.6.4.dev20251206103502__tar.gz → 0.6.4.dev20251209171241__tar.gz

{liger_kernel_nightly-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.4.dev20251206103502
+Version: 0.6.4.dev20251209171241
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -113,6 +113,8 @@ We've also added optimized Post-Training kernels that deliver **up to 80% memory
 
 You can view the documentation site for additional installation, usage examples, and API references:https://linkedin.github.io/Liger-Kernel/
 
+You can view the Liger Kernel Technical Report: https://openreview.net/forum?id=36SjAIT42G
+
 ## Supercharge Your Model with Liger Kernel
 
 ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -442,3 +444,4 @@ url={https://openreview.net/forum?id=36SjAIT42G}
         ↑ Back to Top ↑
     </a>
 </p>
+

{liger_kernel_nightly-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/README.md

@@ -65,6 +65,8 @@ We've also added optimized Post-Training kernels that deliver **up to 80% memory
 
 You can view the documentation site for additional installation, usage examples, and API references:https://linkedin.github.io/Liger-Kernel/
 
+You can view the Liger Kernel Technical Report: https://openreview.net/forum?id=36SjAIT42G
+
 ## Supercharge Your Model with Liger Kernel
 
 ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -394,3 +396,4 @@ url={https://openreview.net/forum?id=36SjAIT42G}
         ↑ Back to Top ↑
     </a>
 </p>
+

{liger_kernel_nightly-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.4.dev20251206103502"
+version = "0.6.4.dev20251209171241"
 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-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/src/liger_kernel/transformers/monkey_patch.py

@@ -35,8 +35,7 @@ from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_f
 from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
-from liger_kernel.transformers.rope import liger_rotary_pos_emb_with_cast
-from liger_kernel.transformers.rope import liger_rotary_pos_emb_with_cast_and_leading_batch
+from liger_kernel.transformers.rope import liger_rotary_pos_emb_vision
 from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
@@ -1754,8 +1753,8 @@ def apply_liger_kernel_to_qwen3_vl(
     from liger_kernel.transformers.model.qwen3_vl import lce_forward as qwen3_vl_lce_forward
 
     if rope:
-        modeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
-        modeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch
+        modeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb
+        modeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
 
     if rms_norm:
         modeling_qwen3_vl.Qwen3VLTextRMSNorm = LigerRMSNorm
@@ -1829,8 +1828,8 @@ def apply_liger_kernel_to_qwen3_vl_moe(
     from liger_kernel.transformers.model.qwen3_vl_moe import lce_forward as qwen3_vl_moe_lce_forward
 
     if rope:
-        modeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
-        modeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch
+        modeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
+        modeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_vision
 
     if rms_norm:
         modeling_qwen3_vl_moe.Qwen3VLMoeTextRMSNorm = LigerRMSNorm

liger_kernel_nightly-0.6.4.dev20251209171241/src/liger_kernel/transformers/rope.py

@@ -0,0 +1,64 @@
+from typing import Tuple
+
+import torch
+
+from liger_kernel.ops.rope import LigerRopeFunction
+
+
+def liger_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """
+    Applies Rotary Positional Embedding (RoPE) operation to query and key states.
+
+    Args:
+        q (torch.Tensor): The query tensor of shape (bsz, n_q_head, seq_len, head_dim).
+        k (torch.Tensor): The key tensor of shape (bsz, n_kv_head, seq_len, head_dim).
+        cos (torch.Tensor): The cosine tensor of shape (1, seq_len, head_dim) or (bsz, seq_len, head_dim).
+        sin (torch.Tensor): The sine tensor of shape (1, seq_len, head_dim) or (bsz, seq_len, head_dim).
+        position_ids (torch.Tensor, optional): The position ids tensor. Defaults to None.
+        unsqueeze_dim (int, optional): The dimension to unsqueeze. Defaults to 1.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The query and key tensors after applying the RoPE operation.
+    """
+
+    return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
+
+
+def liger_rotary_pos_emb_vision(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Modified version of liger_rotary_pos_emb for qwen3_vl's apply_rotary_pos_emb_vision function.
+    Manually tranposed the input and output to match the expected shape for liger_rotary_pos_emb.
+    Reference: https://https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L116
+
+    Args:
+        q (torch.Tensor): The query tensor of shape (seq_length, num_heads, head_dim),
+        with stride (num_heads * head_dim, head_dim, 1).
+        k (torch.Tensor): The query tensor of shape (seq_length, num_heads, head_dim),
+        with stride (num_heads * head_dim, head_dim, 1). Same as q.
+        cos (torch.Tensor): The cosine tensor of shape (seq_length, head_dim).
+        sin (torch.Tensor): The sine tensor of shape (seq_length, head_dim).
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The query and key tensors with the same shape and stride as inputs.
+    """
+    orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
+
+    # tranpose to (1, num_heads, seq_length, head_dim) and cast to float32 to match liger_rotary_pos_emb input shape
+    # also unsqueeze for batch dim
+    q32 = q.to(torch.float32).unsqueeze(0).transpose(1, 2)
+    k32 = k.to(torch.float32).unsqueeze(0).transpose(1, 2)
+    cos32 = cos.to(torch.float32)
+    sin32 = sin.to(torch.float32)
+
+    q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32)
+
+    # transpose back to (seq_length, num_heads, head_dim) and cast back to original dtype
+    # also squeeze out batch dim
+    q_out = q_out.transpose(1, 2).squeeze(0).to(orig_q_dtype)
+    k_out = k_out.transpose(1, 2).squeeze(0).to(orig_k_dtype)
+    return q_out, k_out

{liger_kernel_nightly-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.4.dev20251206103502
+Version: 0.6.4.dev20251209171241
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -113,6 +113,8 @@ We've also added optimized Post-Training kernels that deliver **up to 80% memory
 
 You can view the documentation site for additional installation, usage examples, and API references:https://linkedin.github.io/Liger-Kernel/
 
+You can view the Liger Kernel Technical Report: https://openreview.net/forum?id=36SjAIT42G
+
 ## Supercharge Your Model with Liger Kernel
 
 ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -442,3 +444,4 @@ url={https://openreview.net/forum?id=36SjAIT42G}
         ↑ Back to Top ↑
     </a>
 </p>
+

{liger_kernel_nightly-0.6.4.dev20251206103502 → liger_kernel_nightly-0.6.4.dev20251209171241}/test/transformers/test_monkey_patch.py

@@ -1019,8 +1019,8 @@ def test_apply_liger_kernel_to_instance_for_qwen3_vl_moe_text():
 def test_qwen3_vl_rope_hooks_applied():
     # Ensure any monkey patching is cleaned up for subsequent tests
     with patch("transformers.models.qwen3_vl.modeling_qwen3_vl") as modeling_mod:
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast_and_leading_batch
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_vision
 
         # Before applying, make sure attributes exist but are not the liger implementations
         setattr(modeling_mod, "apply_rotary_pos_emb", object())
@@ -1028,16 +1028,16 @@ def test_qwen3_vl_rope_hooks_applied():
 
         _apply_liger_kernel("qwen3_vl")
 
-        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb_with_cast
-        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_with_cast_and_leading_batch
+        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb
+        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_vision
 
 
 @pytest.mark.skipif(not is_qwen3_vl_moe_available(), reason="qwen3_vl_moe module not available")
 def test_qwen3_vl_moe_rope_hooks_applied():
     # Ensure any monkey patching is cleaned up for subsequent tests
     with patch("transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe") as modeling_mod:
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast_and_leading_batch
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_vision
 
         # Before applying, make sure attributes exist but are not the liger implementations
         setattr(modeling_mod, "apply_rotary_pos_emb", object())
@@ -1045,8 +1045,8 @@ def test_qwen3_vl_moe_rope_hooks_applied():
 
         _apply_liger_kernel("qwen3_vl_moe")
 
-        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb_with_cast
-        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_with_cast_and_leading_batch
+        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb
+        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_vision
 
 
 @pytest.mark.skipif(not is_falcon_h1_available(), reason="falcon_h1 module not available")

liger_kernel_nightly-0.6.4.dev20251206103502/src/liger_kernel/transformers/rope.py

@@ -1,63 +0,0 @@
-from typing import Optional
-from typing import Tuple
-
-import torch
-
-from liger_kernel.ops.rope import LigerRopeFunction
-
-
-def liger_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
-    """
-    Applies Rotary Positional Embedding (RoPE) operation to query and key states.
-
-    Args:
-        q (torch.Tensor): The query tensor of shape (bsz, n_q_head, seq_len, head_dim).
-        k (torch.Tensor): The key tensor of shape (bsz, n_kv_head, seq_len, head_dim).
-        cos (torch.Tensor): The cosine tensor of shape (1, seq_len, head_dim) or (bsz, seq_len, head_dim).
-        sin (torch.Tensor): The sine tensor of shape (1, seq_len, head_dim) or (bsz, seq_len, head_dim).
-        position_ids (torch.Tensor, optional): The position ids tensor. Defaults to None.
-        unsqueeze_dim (int, optional): The dimension to unsqueeze. Defaults to 1.
-
-    Returns:
-        Tuple[torch.Tensor, torch.Tensor]: The query and key tensors after applying the RoPE operation.
-    """
-
-    return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
-
-
-def liger_rotary_pos_emb_with_cast(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-    position_ids: Optional[torch.Tensor] = None,
-    unsqueeze_dim: int = 1,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
-
-    q32 = q.to(torch.float32)
-    k32 = k.to(torch.float32)
-    cos32 = cos.to(torch.float32)
-    sin32 = sin.to(torch.float32)
-
-    q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
-    return q_out.to(orig_q_dtype), k_out.to(orig_k_dtype)
-
-
-def liger_rotary_pos_emb_with_cast_and_leading_batch(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-    position_ids: Optional[torch.Tensor] = None,
-    unsqueeze_dim: int = 1,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
-
-    q32 = q.to(torch.float32).unsqueeze(0)
-    k32 = k.to(torch.float32).unsqueeze(0)
-    cos32 = cos.to(torch.float32).unsqueeze(0)
-    sin32 = sin.to(torch.float32).unsqueeze(0)
-
-    q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
-    return q_out.to(orig_q_dtype).squeeze(0), k_out.to(orig_k_dtype).squeeze(0)