From 71b0da7c4ecca10a9f11f535fc370cf988ce12d7 Mon Sep 17 00:00:00 2001
From: bastefaniak <bstefaniak@nvidia.com>
Date: Fri, 14 Aug 2026 00:08:40 +0200
Subject: [PATCH] [Bugfix] Fix .../mrope.py::apply_interleaved_rope() when
 torch.compile is used in torch==2.13 (#52005)

Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com>
---
 tests/kernels/core/test_mrope.py              | 45 +++++++++++++++++++
 .../layers/rotary_embedding/mrope.py          | 12 ++---
 2 files changed, 52 insertions(+), 5 deletions(-)

diff --git a/tests/kernels/core/test_mrope.py b/tests/kernels/core/test_mrope.py
index 6f64fabfe9..81e0a916a3 100644
--- a/tests/kernels/core/test_mrope.py
+++ b/tests/kernels/core/test_mrope.py
@@ -6,6 +6,7 @@ import pytest
 import torch
 
 from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.rotary_embedding.mrope import apply_interleaved_rope
 from vllm.platforms import current_platform
 from vllm.transformers_utils.config import get_config
 from vllm.utils.torch_utils import set_random_seed
@@ -60,6 +61,50 @@ MODELS_TO_TEST = [
 num_tokens_list = [11, 8192]
 
 
+def test_apply_interleaved_rope():
+    mrope_section = [3, 1, 1]
+    x = torch.tensor(
+        [
+            [[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]],
+            [[10, 11, 12, 13, 14], [15, 16, 17, 18, 19]],
+            [[20, 21, 22, 23, 24], [25, 26, 27, 28, 29]],
+        ]
+    )
+
+    result = apply_interleaved_rope(x, mrope_section)
+
+    expected = torch.tensor([[0, 11, 22, 3, 4], [5, 16, 27, 8, 9]])
+    torch.testing.assert_close(result, expected, rtol=0, atol=0)
+
+
+@pytest.mark.skipif(
+    not current_platform.is_cuda_alike(), reason="Skipping CUDA/ROCm only test."
+)
+def test_apply_interleaved_rope_torch_compile():
+    mrope_section = [24, 20, 20]
+    num_tokens = 8192
+    rotary_dim = sum(mrope_section) * 2
+    cache = torch.randn(
+        3,
+        num_tokens,
+        rotary_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    x = cache[..., : rotary_dim // 2]
+
+    expected = apply_interleaved_rope(x, mrope_section)
+    compiled_fn = torch.compile(
+        apply_interleaved_rope,
+        backend="inductor",
+        fullgraph=True,
+    )
+
+    result = compiled_fn(x, mrope_section)
+
+    torch.testing.assert_close(result, expected, rtol=0, atol=0)
+
+
 @pytest.mark.skipif(
     not current_platform.is_cuda_alike(), reason="Skipping CUDA/ROCm only tests."
 )
diff --git a/vllm/model_executor/layers/rotary_embedding/mrope.py b/vllm/model_executor/layers/rotary_embedding/mrope.py
index 29ce9e5000..4da8350171 100644
--- a/vllm/model_executor/layers/rotary_embedding/mrope.py
+++ b/vllm/model_executor/layers/rotary_embedding/mrope.py
@@ -226,12 +226,14 @@ def triton_mrope(
 def apply_interleaved_rope(x: torch.Tensor, mrope_section: list[int]) -> torch.Tensor:
     """Apply interleaved MRoPE to 3D rotary embeddings.
     Reorganizes frequency layout from chunked [TTT...HHH...WWW] to
-    interleaved [THTHWHTHW...TT], preserving frequency continuity.
+    interleaved [THWTHWTHW...TT], preserving frequency continuity.
     """
-    x_t = x[0].clone()
-    x_t[..., 1 : mrope_section[1] * 3 : 3] = x[1, ..., 1 : mrope_section[1] * 3 : 3]
-    x_t[..., 2 : mrope_section[2] * 3 : 3] = x[2, ..., 2 : mrope_section[2] * 3 : 3]
-    return x_t
+    channels = torch.arange(x.shape[-1], device=x.device)
+    is_height = (channels % 3 == 1) & (channels < mrope_section[1] * 3)
+    is_width = (channels % 3 == 2) & (channels < mrope_section[2] * 3)
+
+    result = torch.where(is_height, x[1], x[0])
+    return torch.where(is_width, x[2], result)
 
 
 class MRotaryEmbedding(RotaryEmbeddingBase):