sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/layers/rotary_embedding.py

@@ -7,6 +7,8 @@ from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
+import triton
+import triton.language as tl
 
 from sglang.srt.custom_op import CustomOp
 from sglang.srt.utils import (
@@ -17,6 +19,7 @@ from sglang.srt.utils import (
     is_cuda,
     is_hip,
     is_npu,
+    is_xpu,
 )
 
 _is_cuda = is_cuda()
@@ -25,6 +28,7 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 _is_npu = is_npu()
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
+_is_xpu = is_xpu()
 
 if _is_cuda:
     from sgl_kernel import FusedSetKVBufferArg, apply_rope_with_cos_sin_cache_inplace
@@ -108,9 +112,11 @@ class RotaryEmbedding(CustomOp):
         if not _is_cuda:
             cache = cache.to(dtype)
 
-        if (
-            not (_is_cuda or _is_npu) or self.head_size not in [64, 128, 256, 512]
-        ) and not (_is_cpu and _is_cpu_amx_available):
+        if dtype == torch.float32 or (
+            (not (_is_cuda or _is_npu) or self.head_size not in [64, 128, 256, 512])
+            and not (_is_cpu and _is_cpu_amx_available)
+            and not (_is_xpu)
+        ):
             from vllm._custom_ops import rotary_embedding
 
             self.vllm_rotary_embedding = rotary_embedding
@@ -248,7 +254,11 @@ class RotaryEmbedding(CustomOp):
         offsets: Optional[torch.Tensor] = None,
         fused_set_kv_buffer_arg: Optional[FusedSetKVBufferArg] = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        if _is_cuda and (self.head_size in [64, 128, 256, 512]):
+        if (
+            _is_cuda
+            and (self.head_size in [64, 128, 256, 512])
+            and self.dtype != torch.float32
+        ):
             apply_rope_with_cos_sin_cache_inplace(
                 positions=positions,
                 query=query,
@@ -284,6 +294,17 @@ class RotaryEmbedding(CustomOp):
         s += f", base={self.base}, is_neox_style={self.is_neox_style}"
         return s
 
+    def forward_xpu(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        offsets: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # TODO: make a wrapper, and XPU will implement this kernel later.
+        self.cos_sin_cache = self.cos_sin_cache.to(query.device)
+        return self.forward_native(positions, query, key, offsets)
+
 
 class LinearScalingRotaryEmbedding(RotaryEmbedding):
     """RotaryEmbedding extended with linear scaling.
@@ -1008,6 +1029,199 @@ class DynamicNTKAlphaRotaryEmbedding(RotaryEmbedding):
         return cache
 
 
+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.
+    """
+    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
+
+
+@triton.jit
+def _triton_mrope_forward(
+    q_ptr,
+    k_ptr,
+    cos,
+    sin,
+    num_tokens,
+    n_qh: tl.constexpr,
+    n_kh: tl.constexpr,
+    hd: tl.constexpr,
+    rd: tl.constexpr,
+    pad_n_qh: tl.constexpr,
+    pad_n_kh: tl.constexpr,
+    pad_hd: tl.constexpr,
+    mrope_section_t: tl.constexpr,
+    mrope_section_h: tl.constexpr,
+    mrope_section_w: tl.constexpr,
+    is_interleaved: tl.constexpr,
+):
+    # Adapted from
+    # https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/qwen2vl_mrope.py
+    # This version supports flatten input tensors from vllm
+    # and supports cos and sin cache with shape (3, num_tokens, head_dim // 2)
+    # instead of (3, bsz, seq_len, head_dim), also supports interleaved rotary
+    pid = tl.program_id(0)
+    # locate start address
+    q_ptr = q_ptr + pid * (n_qh * hd)
+    k_ptr = k_ptr + pid * (n_kh * hd)
+
+    # ####################################################################
+    # get the cos(mθ_{i...d/2}) and sin(mθ_{i...d/2}) for token position
+    # m of this program instance
+    # ####################################################################
+    # Note: cos and sin now have shape (3, num_tokens, head_dim // 2)
+
+    # Updated stride calculation for half head_dim
+    half_rd = rd // 2
+    t_cos = cos + pid * half_rd
+    h_cos = t_cos + num_tokens * half_rd
+    w_cos = h_cos + num_tokens * half_rd
+    t_sin = sin + pid * half_rd
+    h_sin = t_sin + num_tokens * half_rd
+    w_sin = h_sin + num_tokens * half_rd
+
+    # Updated offsets for half head_dim
+    cos_offsets = tl.arange(0, pad_hd // 2)
+    if is_interleaved:
+        h_mask = ((cos_offsets % 3) == 1) & (cos_offsets <= 3 * mrope_section_h)
+        w_mask = ((cos_offsets % 3) == 2) & (cos_offsets <= 3 * mrope_section_w)
+        t_mask = ~(h_mask | w_mask)
+    else:
+        t_end = mrope_section_t
+        h_end = t_end + mrope_section_h
+        t_mask = cos_offsets < mrope_section_t
+        h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end)
+        w_mask = (h_end <= cos_offsets) & (cos_offsets < half_rd)
+
+    t_cos_row = tl.load(t_cos + cos_offsets, mask=t_mask, other=0)
+    h_cos_row = tl.load(h_cos + cos_offsets, mask=h_mask, other=0)
+    w_cos_row = tl.load(w_cos + cos_offsets, mask=w_mask, other=0)
+    t_sin_row = tl.load(t_sin + cos_offsets, mask=t_mask, other=0)
+    h_sin_row = tl.load(h_sin + cos_offsets, mask=h_mask, other=0)
+    w_sin_row = tl.load(w_sin + cos_offsets, mask=w_mask, other=0)
+
+    cos_row = t_cos_row + h_cos_row + w_cos_row
+    sin_row = t_sin_row + h_sin_row + w_sin_row
+
+    # ####################################################################
+    # Load the left and right half of q and k for the current
+    # program instance (i.e. for the current token) separately
+    # ####################################################################
+    # left half of the head
+    first_half_q_offsets = (
+        tl.arange(0, pad_n_qh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    )
+    first_half_k_offsets = (
+        tl.arange(0, pad_n_kh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    )
+    first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & (
+        tl.arange(0, pad_hd // 2)[None, :] < rd // 2
+    )
+    first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & (
+        tl.arange(0, pad_hd // 2)[None, :] < rd // 2
+    )
+
+    q_tile_1 = tl.load(q_ptr + first_half_q_offsets, mask=first_q_mask, other=0).to(
+        sin_row.dtype
+    )
+    k_tile_1 = tl.load(k_ptr + first_half_k_offsets, mask=first_k_mask, other=0).to(
+        sin_row.dtype
+    )
+
+    # right half of the head
+    second_half_q_offsets = first_half_q_offsets + (rd // 2)
+    second_half_k_offsets = first_half_k_offsets + (rd // 2)
+    second_q_mask = first_q_mask
+    second_k_mask = first_k_mask
+
+    q_tile_2 = tl.load(q_ptr + second_half_q_offsets, mask=second_q_mask, other=0).to(
+        sin_row.dtype
+    )
+    k_tile_2 = tl.load(k_ptr + second_half_k_offsets, mask=second_k_mask, other=0).to(
+        sin_row.dtype
+    )
+
+    # y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]
+    # Since cos and sin are now half-size,
+    # we use the same cos_row and sin_row for both halves
+    new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
+    tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+    new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
+    tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+    new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
+    tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+    new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
+    tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+
+
+def triton_mrope(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    mrope_section: list[int],
+    head_size: int,
+    rotary_dim: int,
+    mrope_interleaved: bool,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """The mrope triton kernel.
+
+    Args:
+        q: [num_tokens, num_heads * head_size]
+        k: [num_tokens, num_kv_heads * head_size]
+        cos: [3, num_tokens, head_size //2 ]
+            (T/H/W positions with multimodal inputs)
+        sin: [3, num_tokens, head_size //2 ]
+            (T/H/W positions with multimodal inputs)
+        mrope_section: [t, h, w]
+        head_size: int
+    """
+    n_row, n_q_head_head_dim = q.shape
+    assert (
+        n_q_head_head_dim % head_size == 0
+    ), f"q shape {n_q_head_head_dim} must be divisible by head_size {head_size}"
+    n_q_head = n_q_head_head_dim // head_size
+    assert (
+        k.shape[1] % head_size == 0
+    ), f"k shape {k.shape[1]} must be divisible by head_size {head_size}"
+    n_kv_head = k.shape[1] // head_size
+    pad_hd = triton.next_power_of_2(head_size)
+    pad_n_q_head = triton.next_power_of_2(n_q_head)
+    pad_n_kv_head = triton.next_power_of_2(n_kv_head)
+
+    # ensure tensors passed into the kernel are contiguous.
+    # It will be no-op if they are already contiguous
+    q = q.contiguous()
+    k = k.contiguous()
+    cos = cos.contiguous()
+    sin = sin.contiguous()
+
+    _triton_mrope_forward[(n_row,)](
+        q,
+        k,
+        cos,
+        sin,
+        n_row,
+        n_q_head,
+        n_kv_head,
+        head_size,
+        rotary_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        mrope_section[0],
+        mrope_section[1],
+        mrope_section[2],
+        mrope_interleaved,
+    )
+    return q, k
+
+
 class MRotaryEmbedding(RotaryEmbedding):
     """Rotary Embedding with Multimodal Sections."""
 
@@ -1020,12 +1234,14 @@ class MRotaryEmbedding(RotaryEmbedding):
         is_neox_style: bool,
         dtype: torch.dtype,
         mrope_section: Optional[List[int]] = None,
+        mrope_interleaved: bool = False,
     ) -> None:
         super().__init__(
             head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype
         )
 
         self.mrope_section = mrope_section
+        self.mrope_interleaved = mrope_interleaved
         if self.mrope_section:
             expected_sum = rotary_dim // 2
             actual_sum = sum(self.mrope_section)
@@ -1059,8 +1275,17 @@ class MRotaryEmbedding(RotaryEmbedding):
                     f"Corrected mrope_section: {self.mrope_section} (sum={sum(self.mrope_section)})"
                 )
 
+    def _match_cos_sin_cache_dtype(self, query: torch.Tensor) -> None:
+        # __setattr__ in nn.Module (called by `self.cos_sin_cache = ...`)
+        # is expensive, so avoid calling it if possible
+        if (
+            self.cos_sin_cache.device != query.device
+            or self.cos_sin_cache.dtype != query.dtype
+        ):
+            self.cos_sin_cache = self.cos_sin_cache.to(query.device, dtype=query.dtype)
+
     @torch.compile(dynamic=True, backend=get_compiler_backend())
-    def forward(
+    def _forward_native(
         self,
         positions: torch.Tensor,
         query: torch.Tensor,
@@ -1086,15 +1311,18 @@ class MRotaryEmbedding(RotaryEmbedding):
         cos, sin = cos_sin.chunk(2, dim=-1)
         if positions.ndim == 2:
             assert self.mrope_section
-
-            cos = torch.cat(
-                [m[i] for i, m in enumerate(cos.split(self.mrope_section, dim=-1))],
-                dim=-1,
-            )
-            sin = torch.cat(
-                [m[i] for i, m in enumerate(sin.split(self.mrope_section, dim=-1))],
-                dim=-1,
-            )
+            if self.mrope_interleaved:
+                cos = apply_interleaved_rope(cos, self.mrope_section)
+                sin = apply_interleaved_rope(sin, self.mrope_section)
+            else:
+                cos = torch.cat(
+                    [m[i] for i, m in enumerate(cos.split(self.mrope_section, dim=-1))],
+                    dim=-1,
+                )
+                sin = torch.cat(
+                    [m[i] for i, m in enumerate(sin.split(self.mrope_section, dim=-1))],
+                    dim=-1,
+                )
 
         query_shape = query.shape
         query = query.view(num_tokens, -1, self.head_size)
@@ -1111,6 +1339,72 @@ class MRotaryEmbedding(RotaryEmbedding):
         key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
         return query, key
 
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        fused_set_kv_buffer_arg: Optional[FusedSetKVBufferArg] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass with optional Triton kernel acceleration.
+        Args:
+            positions:
+                [num_tokens,] (text only) or
+                [3, num_tokens] (T/H/W positions with multimodal inputs)
+            query: [num_tokens, num_heads * head_size]
+            key: [num_tokens, num_kv_heads * head_size]
+        """
+        assert positions.ndim == 1 or positions.ndim == 2
+
+        if positions.ndim == 2 and self.mrope_section and _is_cuda:
+            return self._forward_triton(positions, query, key)
+        else:
+            return self._forward_native(positions, query, key)
+
+    def _forward_triton(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        assert positions.ndim == 1 or positions.ndim == 2
+        assert key is not None
+
+        self._match_cos_sin_cache_dtype(query)
+        num_tokens = positions.shape[-1]
+        cos_sin = self.cos_sin_cache[positions]
+        cos, sin = cos_sin.chunk(2, dim=-1)
+        query_shape = query.shape
+        key_shape = key.shape
+        if positions.ndim == 2:
+            assert self.mrope_section
+
+            q, k = triton_mrope(
+                query,
+                key,
+                cos,
+                sin,
+                self.mrope_section,
+                self.head_size,
+                self.rotary_dim,
+                self.mrope_interleaved,
+            )
+
+            return q.reshape(query_shape), k.reshape(key_shape)
+
+        query = query.view(num_tokens, -1, self.head_size)
+        query_rot = query[..., : self.rotary_dim]
+        query_pass = query[..., self.rotary_dim :]
+        query_rot = _apply_rotary_emb(query_rot, cos, sin, self.is_neox_style)
+        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)
+
+        key = key.view(num_tokens, -1, self.head_size)
+        key_rot = key[..., : self.rotary_dim]
+        key_pass = key[..., self.rotary_dim :]
+        key_rot = _apply_rotary_emb(key_rot, cos, sin, self.is_neox_style)
+        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
+        return query, key
+
     # Copied from https://github.com/huggingface/transformers/blob/c8e0e603de9b3d49161a15fe6e8ea84badfb5d02/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1439
     @staticmethod
     def get_rope_index(
@@ -1126,6 +1420,28 @@ class MRotaryEmbedding(RotaryEmbedding):
         second_per_grid_ts: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if model_type == "qwen3_omni_moe":
+            # For qwen3-omni
+            return MRotaryEmbedding.get_rope_index_qwen3_omni(
+                spatial_merge_size,
+                image_token_id,
+                video_token_id,
+                vision_start_token_id,
+                tokens_per_second,
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                **kwargs,
+            )
+        if (
+            model_type.startswith("qwen3_vl") or model_type.startswith("qwen3_vl_moe")
+        ) and video_grid_thw is not None:
+            video_grid_thw = torch.repeat_interleave(
+                video_grid_thw, video_grid_thw[:, 0], dim=0
+            )
+            video_grid_thw[:, 0] = 1
+
         mrope_position_deltas = []
         if input_ids is not None and (
             image_grid_thw is not None or video_grid_thw is not None
@@ -1211,7 +1527,11 @@ class MRotaryEmbedding(RotaryEmbedding):
 
                         time_tensor_long = time_tensor.long()
                         t_index = time_tensor_long.flatten()
-                    elif model_type in ("qwen2_vl", "qwen3_vl", "qwen3_vl_moe"):
+                    elif model_type in (
+                        "qwen2_vl",
+                        "qwen3_vl",
+                        "qwen3_vl_moe",
+                    ):
                         t_index = (
                             torch.arange(llm_grid_t)
                             .view(-1, 1)
@@ -1219,7 +1539,7 @@ class MRotaryEmbedding(RotaryEmbedding):
                             .flatten()
                         )
                     else:
-                        raise RuntimeError("Unimplemented")
+                        raise RuntimeError(f"Unimplemented model type: {model_type}")
                     h_index = (
                         torch.arange(llm_grid_h)
                         .view(1, -1, 1)
@@ -1269,6 +1589,304 @@ class MRotaryEmbedding(RotaryEmbedding):
             mrope_position_deltas = max_position_ids + 1 - s
             return position_ids, mrope_position_deltas
 
+    @staticmethod
+    def get_rope_index_qwen3_omni(
+        spatial_merge_size: int,
+        image_token_id: int,
+        video_token_id: int,
+        vision_start_token_id: int,
+        tokens_per_second: Optional[int] = None,
+        input_ids: Optional[torch.LongTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # For qwen3-omni
+        audio_token_id = kwargs["audio_token_id"]
+        audio_start_token_id = kwargs["audio_start_token_id"]
+        position_id_per_seconds = kwargs["position_id_per_seconds"]
+        use_audio_in_video = kwargs.get("use_audio_in_video", False)
+        audio_seqlens = kwargs.get("audio_seqlens", None)
+        second_per_grids = second_per_grid_ts
+
+        mrope_position_deltas = []
+        if input_ids is not None and (
+            image_grid_thw is not None or video_grid_thw is not None
+        ):
+            total_input_ids = input_ids
+            position_ids = torch.zeros(
+                3,
+                input_ids.shape[0],
+                input_ids.shape[1],
+                dtype=torch.float,
+                device=input_ids.device,
+            )
+            image_idx, video_idx, audio_idx = 0, 0, 0
+            for i, current_input_ids in enumerate(total_input_ids):
+                image_nums, video_nums, audio_nums = 0, 0, 0
+                vision_start_indices = torch.argwhere(
+                    current_input_ids == vision_start_token_id
+                ).squeeze(1)
+                if vision_start_indices.numel() > 0:
+                    vision_tokens = current_input_ids[vision_start_indices + 1]
+                    image_nums = (vision_tokens == image_token_id).sum()
+                    video_nums = (
+                        (vision_tokens == audio_start_token_id).sum()
+                        if use_audio_in_video
+                        else (vision_tokens == video_token_id).sum()
+                    )
+                audio_nums = torch.sum(current_input_ids == audio_start_token_id)
+                input_tokens = current_input_ids.tolist()
+                llm_pos_ids_list: list = []
+                st = 0
+                remain_images, remain_videos, remain_audios = (
+                    image_nums,
+                    video_nums,
+                    audio_nums,
+                )
+                multimodal_nums = (
+                    image_nums + audio_nums
+                    if use_audio_in_video
+                    else image_nums + video_nums + audio_nums
+                )
+                for _ in range(multimodal_nums):
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+                    ed_vision_start = (
+                        input_tokens.index(vision_start_token_id, st)
+                        if (
+                            (
+                                image_token_id in input_tokens
+                                or video_token_id in input_tokens
+                            )
+                            and (remain_videos > 0 or remain_images > 0)
+                        )
+                        else len(input_tokens) + 1
+                    )
+                    ed_audio_start = (
+                        input_tokens.index(audio_start_token_id, st)
+                        if (audio_token_id in input_tokens and remain_audios > 0)
+                        else len(input_tokens) + 1
+                    )
+                    min_ed = min(ed_vision_start, ed_audio_start)
+
+                    text_len = min_ed - st
+                    if text_len != 0:
+                        llm_pos_ids_list.append(
+                            torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
+                        )
+                        st_idx += text_len
+                    # Audio in Video
+                    if (
+                        min_ed == ed_vision_start
+                        and ed_vision_start + 1 == ed_audio_start
+                    ):
+                        bos_len, eos_len = 2, 2
+                    else:
+                        bos_len, eos_len = 1, 1
+                    llm_pos_ids_list.append(
+                        torch.arange(bos_len).view(1, -1).expand(3, -1) + st_idx
+                    )
+                    st_idx += bos_len
+                    # Audio Only
+                    if min_ed == ed_audio_start:
+                        audio_len = MRotaryEmbedding._get_feat_extract_output_lengths(
+                            audio_seqlens[audio_idx]
+                        )
+                        llm_pos_ids = (
+                            torch.arange(audio_len).view(1, -1).expand(3, -1) + st_idx
+                        )
+                        llm_pos_ids_list.append(llm_pos_ids)
+
+                        st += int(text_len + bos_len + audio_len + eos_len)
+                        audio_idx += 1
+                        remain_audios -= 1
+
+                    # Image Only
+                    elif (
+                        min_ed == ed_vision_start
+                        and current_input_ids[ed_vision_start + 1] == image_token_id
+                    ):
+                        grid_t = image_grid_thw[image_idx][0]
+                        grid_hs = image_grid_thw[:, 1]
+                        grid_ws = image_grid_thw[:, 2]
+                        t_index = (
+                            torch.arange(grid_t) * 1 * position_id_per_seconds
+                        ).float()
+                        llm_pos_ids = MRotaryEmbedding._get_llm_pos_ids_for_vision(
+                            st_idx,
+                            image_idx,
+                            spatial_merge_size,
+                            t_index,
+                            grid_hs,
+                            grid_ws,
+                            input_ids.device,
+                        )
+                        image_len = image_grid_thw[image_idx].prod() // (
+                            spatial_merge_size**2
+                        )
+                        llm_pos_ids_list.append(llm_pos_ids)
+
+                        st += int(text_len + bos_len + image_len + eos_len)
+                        image_idx += 1
+                        remain_images -= 1
+
+                    # Video Only
+                    elif (
+                        min_ed == ed_vision_start
+                        and current_input_ids[ed_vision_start + 1] == video_token_id
+                    ):
+                        grid_t = video_grid_thw[video_idx][0]
+                        grid_hs = video_grid_thw[:, 1]
+                        grid_ws = video_grid_thw[:, 2]
+                        t_index = (
+                            torch.arange(grid_t)
+                            * second_per_grids[video_idx].cpu().float()
+                            * position_id_per_seconds
+                        ).float()
+                        llm_pos_ids = MRotaryEmbedding._get_llm_pos_ids_for_vision(
+                            st_idx,
+                            video_idx,
+                            spatial_merge_size,
+                            t_index,
+                            grid_hs,
+                            grid_ws,
+                            input_ids.device,
+                        )
+                        video_len = video_grid_thw[video_idx].prod() // (
+                            spatial_merge_size**2
+                        )
+                        llm_pos_ids_list.append(llm_pos_ids)
+
+                        st += int(text_len + bos_len + video_len + eos_len)
+                        video_idx += 1
+                        remain_videos -= 1
+
+                    # Audio in Video
+                    elif (
+                        min_ed == ed_vision_start
+                        and ed_vision_start + 1 == ed_audio_start
+                    ):
+                        audio_len = MRotaryEmbedding._get_feat_extract_output_lengths(
+                            audio_seqlens[audio_idx]
+                        )
+                        audio_llm_pos_ids = (
+                            torch.arange(audio_len).view(1, -1).expand(3, -1) + st_idx
+                        )
+                        grid_t = video_grid_thw[video_idx][0]
+                        grid_hs = video_grid_thw[:, 1]
+                        grid_ws = video_grid_thw[:, 2]
+
+                        t_index = (
+                            torch.arange(grid_t)
+                            * second_per_grids[video_idx].cpu().float()
+                            * position_id_per_seconds
+                        ).float()
+                        video_llm_pos_ids = (
+                            MRotaryEmbedding._get_llm_pos_ids_for_vision(
+                                st_idx,
+                                video_idx,
+                                spatial_merge_size,
+                                t_index,
+                                grid_hs,
+                                grid_ws,
+                                input_ids.device,
+                            )
+                        )
+                        video_data_index, audio_data_index = 0, 0
+                        while (
+                            video_data_index < video_llm_pos_ids.shape[-1]
+                            and audio_data_index < audio_llm_pos_ids.shape[-1]
+                        ):
+                            if (
+                                video_llm_pos_ids[0][video_data_index]
+                                <= audio_llm_pos_ids[0][audio_data_index]
+                            ):
+                                llm_pos_ids_list.append(
+                                    video_llm_pos_ids[
+                                        :, video_data_index : video_data_index + 1
+                                    ]
+                                )
+                                video_data_index += 1
+                            else:
+                                llm_pos_ids_list.append(
+                                    audio_llm_pos_ids[
+                                        :, audio_data_index : audio_data_index + 1
+                                    ]
+                                )
+                                audio_data_index += 1
+                        if video_data_index < video_llm_pos_ids.shape[-1]:
+                            llm_pos_ids_list.append(
+                                video_llm_pos_ids[
+                                    :, video_data_index : video_llm_pos_ids.shape[-1]
+                                ]
+                            )
+                        if audio_data_index < audio_llm_pos_ids.shape[-1]:
+                            llm_pos_ids_list.append(
+                                audio_llm_pos_ids[
+                                    :, audio_data_index : audio_llm_pos_ids.shape[-1]
+                                ]
+                            )
+                        video_len = video_grid_thw[video_idx].prod() // (
+                            spatial_merge_size**2
+                        )
+
+                        st += int(text_len + bos_len + audio_len + video_len + eos_len)
+
+                        audio_idx += 1
+                        video_idx += 1
+                        remain_videos -= 1
+                        remain_audios -= 1
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+                    llm_pos_ids_list.append(
+                        torch.arange(eos_len).view(1, -1).expand(3, -1) + st_idx
+                    )
+
+                if st < len(input_tokens):
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+                    text_len = len(input_tokens) - st
+                    llm_pos_ids_list.append(
+                        torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
+                    )
+
+                llm_positions = torch.cat(
+                    [item.float() for item in llm_pos_ids_list], dim=1
+                ).reshape(3, -1)
+
+                position_ids[..., i, :] = llm_positions.to(position_ids.device)
+                mrope_position_deltas.append(
+                    llm_positions.max() + 1 - len(current_input_ids)
+                )
+            mrope_position_deltas = torch.tensor(
+                mrope_position_deltas, device=input_ids.device
+            ).unsqueeze(1)
+
+            return position_ids, mrope_position_deltas
+        else:
+            s = input_ids.shape[1]
+            position_ids = torch.arange(s)
+            position_ids = (
+                position_ids.unsqueeze(0).expand(3, -1, -1).to(input_ids.device)
+            )
+            max_position_ids = position_ids.max(0, keepdim=False)[0].max(
+                -1, keepdim=True
+            )[0]
+            mrope_position_deltas = max_position_ids + 1 - s
+
+            return position_ids, mrope_position_deltas
+
     # Adapted from https://github.com/vllm-project/vllm/blob/3779eb8c81449b924a23457fc77e45a0e6171178/vllm/model_executor/layers/rotary_embedding.py#L1120
     @staticmethod
     def get_rope_index_glm4v(
@@ -1467,6 +2085,44 @@ class MRotaryEmbedding(RotaryEmbedding):
 
             return position_ids, mrope_position_deltas
 
+    # For qwen3-omni
+    @staticmethod
+    def _get_feat_extract_output_lengths(input_lengths):
+        """
+        Computes the output length of the convolutional layers and the output length of the audio encoder
+        """
+        input_lengths_leave = input_lengths % 100
+        feat_lengths = (input_lengths_leave - 1) // 2 + 1
+        output_lengths = (
+            ((feat_lengths - 1) // 2 + 1 - 1) // 2 + 1 + (input_lengths // 100) * 13
+        )
+        return output_lengths
+
+    # For qwen3-omni
+    @staticmethod
+    def _get_llm_pos_ids_for_vision(
+        st_idx, vision_idx, spatial_merge_size, t_index, grid_hs, grid_ws, device
+    ):
+        grid_h = grid_hs[vision_idx] // spatial_merge_size
+        grid_w = grid_ws[vision_idx] // spatial_merge_size
+
+        h_index = (
+            torch.arange(grid_h, device=device)
+            .view(1, -1, 1)
+            .expand(len(t_index), -1, grid_w)
+            .flatten()
+        )
+        w_index = (
+            torch.arange(grid_w, device=device)
+            .view(1, 1, -1)
+            .expand(len(t_index), grid_h, -1)
+            .flatten()
+        )
+        t_index = t_index.view(-1, 1).expand(-1, grid_h * grid_w).flatten()
+
+        llm_pos_ids = torch.stack([t_index, h_index, w_index], dim=0) + st_idx
+        return llm_pos_ids
+
 
 class DualChunkRotaryEmbedding(CustomOp):
     """Rotary positional embedding for Dual Chunk Attention."""
@@ -1768,6 +2424,7 @@ def get_rope(
                     is_neox_style,
                     dtype,
                     mrope_section=rope_scaling["mrope_section"],
+                    mrope_interleaved=rope_scaling.get("mrope_interleaved", False),
                 )
             else:
                 rotary_emb = RotaryEmbedding(