大改

vllm_npu/torchair/ops/sequence_parallel.py

@@ -0,0 +1,120 @@
+import torch
+from torch.nn import functional as F
+from vllm.distributed import (get_tensor_model_parallel_world_size,
+                              get_tp_group, tensor_model_parallel_all_gather,
+                              tensor_model_parallel_reduce_scatter)
+from vllm.forward_context import get_forward_context
+
+from vllm_npu.platform import NPUPlatform
+
+
+class MetadataForPadding:
+
+    def __init__(self,
+                 padding_flag=False,
+                 lengths_sum_padding=0,
+                 lengths_sum_unpadding=0,
+                 pad_size=0,
+                 not_dummy_and_is_prefill=False):
+        self.padding_flag = padding_flag
+        self.not_dummy_and_is_prefill = not_dummy_and_is_prefill
+
+        self.lengths_sum_padding = lengths_sum_padding
+        self.lengths_sum_unpadding = lengths_sum_unpadding
+        self.pad_size = pad_size
+
+        self.tp_size = get_tp_group().world_size
+        self.tp_rank_in_group = get_tp_group().rank_in_group
+
+        assert self.lengths_sum_padding % self.tp_size == 0
+        self.slice_size = self.lengths_sum_padding // self.tp_size
+
+        self.mc2_mask = torch.zeros(
+            self.lengths_sum_padding,
+            dtype=torch.bool,
+            device=NPUPlatform.device_type,
+        )
+        self.mc2_mask[:lengths_sum_unpadding] = True
+
+    def padding_aligned_reduce_scatter(self,
+                                       data: torch.Tensor) -> torch.Tensor:
+        if self.padding_flag:
+            pad_size = self.pad_size
+            padded_data = F.pad(data, (0, 0, 0, pad_size))
+        else:
+            padded_data = data
+        padded_data_reduce_scatter = tensor_model_parallel_reduce_scatter(
+            padded_data, 0)
+
+        return padded_data_reduce_scatter
+
+    def allgather_unpadding_aligned(self,
+                                    padded_data: torch.Tensor) -> torch.Tensor:
+        padded_data_allgather = tensor_model_parallel_all_gather(
+            padded_data, 0)
+        if self.padding_flag:
+            lengths_sum_unpadding = self.lengths_sum_unpadding
+            unpadding_data = padded_data_allgather[:lengths_sum_unpadding]
+        else:
+            unpadding_data = padded_data_allgather
+        return unpadding_data
+
+    def padding_slice(self, data: torch.Tensor) -> torch.Tensor:
+
+        padded_data = F.pad(data, (0, 0, 0, self.pad_size))
+        start = self.tp_rank_in_group * self.slice_size
+        end = start + self.slice_size
+        slice_data = padded_data[start:end]
+
+        return slice_data
+
+    def padding_aligned_scatter(self, data: torch.Tensor) -> torch.Tensor:
+        if self.padding_flag:
+            pad_size = self.pad_size
+            padded_data = F.pad(data, (0, 0, 0, pad_size))
+        else:
+            padded_data = data
+        # padded_data = data
+        padded_data = torch.tensor_split(padded_data, self.tp_size, dim=0)
+
+        padded_data_reduce_scatter = padded_data[self.tp_rank_in_group]
+
+        return padded_data_reduce_scatter
+
+
+def init_metadata_for_sp(input_ids, enable_sequence_parallelism):
+    if not enable_sequence_parallelism:
+        return MetadataForPadding(padding_flag=False,
+                                  not_dummy_and_is_prefill=False)
+
+    is_perifll = 0
+    attn_metadata = get_forward_context().attn_metadata
+    tp_size = get_tensor_model_parallel_world_size()
+    if attn_metadata is not None:
+        if hasattr(attn_metadata,
+                   'is_only_prefill') and attn_metadata.is_only_prefill:
+            is_perifll = 1
+        if hasattr(attn_metadata,
+                   'num_prefills') and attn_metadata.num_prefills > 0:
+            is_perifll = 1
+
+        if is_perifll:
+            lengths_sum_unpadding = input_ids.shape[0]
+            lengths_sum_padding = (
+                (lengths_sum_unpadding + tp_size - 1) // tp_size) * tp_size
+            if lengths_sum_unpadding == lengths_sum_padding:
+                padding_flag = False
+            else:
+                padding_flag = True
+            pad_size = lengths_sum_padding - lengths_sum_unpadding
+            _metadata_for_padding = MetadataForPadding(
+                lengths_sum_unpadding=lengths_sum_unpadding,
+                lengths_sum_padding=lengths_sum_padding,
+                padding_flag=padding_flag,
+                pad_size=pad_size,
+                not_dummy_and_is_prefill=True)
+
+            return _metadata_for_padding
+
+    return MetadataForPadding(padding_flag=False,
+                              not_dummy_and_is_prefill=False)

vllm_npu/torchair/ops/shared_weight_layer.py

@@ -0,0 +1,245 @@
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+import torch
+import torch.distributed as dist
+from vllm.distributed.parallel_state import GroupCoordinator
+from vllm.model_executor.layers.linear import LinearBase
+
+
+def dispose_tensor(x: torch.Tensor):
+    x.set_(torch.empty([], device=x.device, dtype=x.dtype))
+
+
+@dataclass
+class LayerMetadata:
+    """Metadata for a layer.
+    """
+    layer: Optional[LinearBase]  # The layer object.
+    post_method: Callable[[
+        torch.nn.Module
+    ], None]  # The `process_weights_after_loading` method from the quant method.
+    weight: torch.Tensor  # The weight tensor.
+    window_idx: int  # The index of the window.
+
+
+@dataclass
+class SharedWindowMetadata:
+    """Metadata for a shared window.
+    """
+    weight: torch.Tensor  # The weight tensor to be shared by layers.
+    data_layer_idx: int  # The index of the layer this window's weight is equal to.
+    work: Optional[torch.distributed.Work]  # The asynchronous broadcast work.
+
+
+@dataclass
+class SeriesMetadata:
+    """Metadata for a weight shared series.
+    """
+    group: GroupCoordinator
+    start_layer: int
+    end_layer: int
+    num_layers: int
+    prefetch_step: int
+    dummy_weight: torch.Tensor  # Dummy weight to replace the loaded weight matrix. All the layers in the series share the same dummy weight tensor.
+    layers: list[LayerMetadata]
+    shared_windows: list[
+        SharedWindowMetadata]  # Shared windows for prefetching. The window size is (`prefetch_step` + 1), as only the weights for the next (`prefetch_step` + 1) layers need to be stored.
+    window_offset: int  # The index of the window for the next coming layer.
+
+    def is_source(self, layer_idx) -> bool:
+        return layer_idx % self.group.world_size == self.group.rank_in_group
+
+    def post_process_after_loading(self):
+        # This method only needs to be called once per series.
+        if self.shared_windows:
+            return
+        for layer_idx in range(self.start_layer, self.end_layer):
+            layer = self.layers[layer_idx - self.start_layer]
+            is_source = self.is_source(layer_idx)
+            # If the weight uses dummy weight, make a copy temporary such that the post method call won't affect other layers which also uses dummy weight.
+            if not is_source:
+                layer.weight.set_(torch.empty_like(self.dummy_weight))
+            # Broadcast to get the true weight.
+            dist.broadcast(layer.weight,
+                           src=self.group.ranks[layer_idx %
+                                                self.group.world_size],
+                           group=self.group.device_group)
+            assert layer.layer is not None
+            # Call `process_weights_after_loading` from the quant method.
+            layer.post_method(layer.layer)
+            step = layer_idx - self.start_layer
+            if step < self.prefetch_step:
+                # Build the windows for the first `prefetch_step` layers. The weights can be used for the first `prefetch_step` layers in `forward()`, so also clone the weights.
+                self.shared_windows.append(
+                    SharedWindowMetadata(
+                        weight=layer.weight.clone().detach(),
+                        data_layer_idx=layer_idx,
+                        work=None,
+                    ))
+                layer.window_idx = step
+                # When the layer not intended to be stored in this device, link to the corresponding window's tensor.
+                if not is_source:
+                    layer.weight.set_(self.shared_windows[-1].weight)
+            else:
+                # Build one more window for prefetch. The weight is useless, so just keep the shape.
+                if step == self.prefetch_step:
+                    self.shared_windows.append(
+                        SharedWindowMetadata(
+                            weight=torch.empty_like(layer.weight),
+                            data_layer_idx=-1,
+                            work=None,
+                        ))
+                # When the layer not intended to be stored in this device, dispose the tensor.
+                if not is_source:
+                    dispose_tensor(layer.weight)
+
+        dispose_tensor(self.dummy_weight)
+
+    def reach_layer(self, layer_idx: int):
+        # The index of the layer to be prefetched.
+        next_layer_idx = (layer_idx + self.prefetch_step
+                          ) % self.num_layers + self.start_layer
+        next_layer = self.layers[next_layer_idx - self.start_layer]
+        # The index of the window to store the weight for the coming layer.
+        next_layer.window_idx = self.window_offset
+        window = self.shared_windows[next_layer.window_idx]
+        # When the layer not intended to be stored in this device, link to the corresponding window's tensor.
+        if not self.is_source(next_layer_idx):
+            next_layer.weight.set_(window.weight)
+        # Update `window_offset` by rolling one step.
+        self.window_offset = (self.window_offset + 1) % (self.prefetch_step +
+                                                         1)
+        assert window.data_layer_idx != next_layer_idx
+        window.data_layer_idx = next_layer_idx
+        # Start asynchronous broadcast work.
+        window.work = dist.broadcast(
+            next_layer.weight,
+            src=self.group.ranks[next_layer_idx % self.group.world_size],
+            group=self.group.device_group,
+            async_op=True)
+
+    def wait_weight(self, layer_idx: int):
+        # Find the asynchronous broadcast work and wait for it.
+        assert self.shared_windows
+        window = self.shared_windows[self.layers[layer_idx -
+                                                 self.start_layer].window_idx]
+        # Make sure the data in the corresponding shared window is for the current layer.
+        assert window.data_layer_idx == layer_idx
+        if window.work is not None:
+            window.work.wait()
+            window.work = None
+
+
+@dataclass
+class LayerExternalMetadata:
+    """External metadata for a layer.
+    """
+    series: SeriesMetadata
+    layer_idx: int
+
+
+_series_dict: dict[str, SeriesMetadata] = {}
+
+_layer_external_dict: dict[int, LayerExternalMetadata] = {}
+
+
+def _create_forward_wrapper(forward: Callable, series: SeriesMetadata,
+                            layer_idx: int) -> Callable:
+
+    def wrapped_forward(*args, **kwargs):
+        # Wait for the weight.
+        series.wait_weight(layer_idx)
+        return forward(*args, **kwargs)
+
+    return wrapped_forward
+
+
+"""
+Register linear layers into a shared storage series.
+
+In a parallel group, each device stores a distinct, non-overlapping subset of layers from the series. All layers in a series must have the same structure (are isomorphic). The weight matrix for the i-th layer is stored on device (i % n), where n is the number of devices.
+
+After loading the model, you must call `post_process_after_loading_for_shared_weight_series(layer)` on any layer of this series to complete the initialization.
+
+During execution, each time a new layer is reached, you must call `reach_layer_for_shared_weight_series(layer)` for that layer to prefetch the weights. The argument `prefetch_step` is a non-negative integer k that manages asynchronous weight prefetching. Each call to `reach_layer_for_shared_weight_series(current_layer)` method will trigger an asynchronous prefetch for the weights of the k-th subsequent layer after `current_layer` within the series.
+
+Note: The layers are managed as a circular buffer. The index of the layer to prefetch is determined by the formula:
+- total_layers = end_layer - start_layer
+- prefetch_layer_idx = (layer_idx + prefetch_step) % total_layers + start_layer
+
+To hold the weights for the current layer and the k prefetched layers, a pool of (k + 1) shared tensor buffers will be created for this series.
+
+Arguments:
+    series_name: This name identifies which series this layer belongs to.
+    group: The group coordinator for handling asynchronous communications. It is recommended to create a new group coordinator for each new series.
+    start_layer: The index of the first layer in the series (inclusive).
+    end_layer: The index of the last layer in the series (exclusive). Thus, the series includes all layers with indices in the range [start_layer, end_layer).
+    layer_idx: The index of the current layer.
+    layer: The linear layer object to register.
+    prefetch_step: An integer that manages asynchronous weight prefetching. Setting it to 0 or 1 can cover most cases.
+"""
+
+
+def register_layer_to_shared_weight_series(
+    series_name: str,
+    group: GroupCoordinator,
+    start_layer: int,
+    end_layer: int,
+    layer_idx: int,
+    layer: LinearBase,
+    prefetch_step: int = 1,
+):
+    global _series_dict
+    if series_name not in _series_dict:
+        num_layers = end_layer - start_layer
+        assert num_layers > 0
+        assert prefetch_step >= 0 and prefetch_step <= num_layers - 2
+        _series_dict[series_name] = SeriesMetadata(
+            group=group,
+            start_layer=start_layer,
+            end_layer=end_layer,
+            num_layers=num_layers,
+            prefetch_step=prefetch_step,
+            dummy_weight=torch.empty_like(layer.weight),
+            layers=[
+                LayerMetadata(
+                    layer=None,
+                    post_method=lambda layer: None,
+                    weight=torch.empty([]),
+                    window_idx=-1,
+                ) for _ in range(num_layers)
+            ],
+            shared_windows=[],
+            window_offset=prefetch_step,
+        )
+    series = _series_dict[series_name]
+    assert layer.quant_method is not None
+    series.layers[layer_idx - start_layer] = LayerMetadata(
+        layer=layer,
+        post_method=layer.quant_method.process_weights_after_loading,
+        weight=layer.weight,
+        window_idx=-1,
+    )
+    # Discard the original `process_weights_after_loading` method such that it won't be called by others.
+    layer.quant_method.process_weights_after_loading = lambda layer: None
+    # When the layer not intended to be stored in this device, dispose the tensor and skip weight loading.
+    if not series.is_source(layer_idx):
+        dispose_tensor(layer.weight)
+        layer.weight.weight_loader = lambda *args, **kwargs: None
+    layer.forward = _create_forward_wrapper(layer.forward, series, layer_idx)
+    global _layer_external_dict
+    _layer_external_dict[id(layer)] = LayerExternalMetadata(
+        series=series,
+        layer_idx=layer_idx,
+    )
+
+
+def post_process_after_loading_for_shared_weight_series(layer: LinearBase):
+    ext = _layer_external_dict[id(layer)]
+    ext.series.post_process_after_loading()
+
+
+def reach_layer_for_shared_weight_series(layer: LinearBase):
+    ext = _layer_external_dict[id(layer)]
+    ext.series.reach_layer(ext.layer_idx)

vllm_npu/torchair/ops/torchair_activation.py

@@ -0,0 +1,37 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+import torch
+
+
+def torchair_silu_and_mul_forward_oot(self, x: torch.Tensor) -> torch.Tensor:
+    """AscendSiluAndMul forward in torchair mode.
+
+    The key difference from the original implementation is the removal of operators
+    from the torch.ops.vllm class, as these operators only function in non-torchair
+    modes. Adding them back would cause the graph compilation to fail.
+    """
+
+    import torch_npu
+
+    from vllm_npu.utils import is_310p
+
+    if is_310p():
+        out = torch_npu.npu_swiglu(x.to(torch.float32)).to(torch.float16)
+    else:
+        out = torch_npu.npu_swiglu(x)
+    return out

vllm_npu/torchair/ops/torchair_layernorm.py

@@ -0,0 +1,78 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+from typing import Optional, Tuple, Union
+
+import torch
+from vllm.config import get_current_vllm_config
+from vllm.model_executor.layers.layernorm import RMSNorm
+
+_original_re_init = RMSNorm.__init__
+
+
+def torchair_rmsnorm_init_(
+    self,
+    hidden_size: int,
+    eps: float = 1e-6,
+    var_hidden_size: Optional[int] = None,
+    has_weight: bool = True,
+    dtype: Optional[torch.dtype] = None,
+) -> None:
+    _original_re_init(self, hidden_size, eps, var_hidden_size, has_weight,
+                      dtype)
+    vllm_config = get_current_vllm_config()
+    self.bias = None
+    # quantization with anti_method m4 will generate none-zero norm bias
+    if vllm_config.quant_config is not None and \
+            any("norm.bias" in name for name in vllm_config.quant_config.quant_description.keys()):
+        self.bias = torch.nn.Parameter(torch.zeros(hidden_size),
+                                       requires_grad=False)
+
+
+def torchair_rmsnorm_forward_oot(
+    self,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    """AscendRMSNorm forward in torchair mode.
+
+    The key difference from the original implementation is the removal of operators
+    from the torch.ops.vllm class, as these operators only function in non-torchair
+    modes. Adding them back would cause the graph compilation to fail.
+    """
+
+    import torch_npu
+
+    from vllm_npu.utils import is_310p
+    if residual is not None:
+        if is_310p():
+            orig_dtype = residual.dtype
+            x = x + residual.to(x.dtype)
+            residual = x.to(orig_dtype)
+            x, _ = torch_npu.npu_rms_norm(x, self.weight,
+                                          self.variance_epsilon)
+        else:
+            x, _, residual = torch_npu.npu_add_rms_norm(
+                x, residual, self.weight, self.variance_epsilon)
+        if self.bias is not None:
+            x.add_(self.bias)
+        return x, residual
+
+    x, residual = torch_npu.npu_rms_norm(x, self.weight, self.variance_epsilon)
+    if self.bias is not None:
+        x.add_(self.bias)
+    return x

vllm_npu/torchair/ops/torchair_rotary_embedding.py

@@ -0,0 +1,365 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import torch_npu
+from vllm.model_executor.layers.rotary_embedding import (
+    DeepseekScalingRotaryEmbedding, RotaryEmbedding)
+
+from vllm_npu.ascend_config import get_ascend_config
+from vllm_npu.utils import enable_custom_op, is_310p
+
+
+def custom_rotary_embedding_enabled(query, neox_style, head_size):
+    return query.dtype == torch.float16 and neox_style and head_size % 32 == 0 and enable_custom_op(
+    )
+
+
+def rope_forward_oot(
+    self,
+    positions: torch.Tensor,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    offsets: Optional[torch.Tensor] = None,
+    is_neox_style_override: Optional[bool] = None,
+    is_qwen_torchair: Optional[bool] = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if get_ascend_config(
+    ).torchair_graph_config.enabled and not is_qwen_torchair:
+        return self.forward_native(
+            positions,
+            query,
+            key,
+            offsets,
+        )
+
+    query_shape, key_shape = query.shape, key.shape
+    if self.cos_sin_cache.device != query.device:
+        self.cos_sin_cache = self.cos_sin_cache.to(query.device)
+    if self.cos_sin_cache.dtype != query.dtype:
+        self.cos_sin_cache = self.cos_sin_cache.to(query.dtype)
+    neox_style = self.is_neox_style
+    if is_neox_style_override is not None:
+        neox_style = is_neox_style_override
+    # adopt custom kernel path for rotary_embedding
+    if custom_rotary_embedding_enabled(query, neox_style,
+                                       self.head_size) and not is_310p():
+        query, key = torch.ops._C_ascend.rotary_embedding(
+            positions,
+            query,
+            key,
+            self.head_size,
+            self.cos_sin_cache,
+            neox_style,
+        )
+        return query.view(query_shape), key.view(key_shape)
+    if offsets is not None:
+        raise NotImplementedError(
+            "Batched rotary embedding is currently not supported on NPU.")
+    else:
+        # TODO: Remove the contiguous in the future.
+        query = query.contiguous().view(query.shape[0], -1)
+        key = key.contiguous().view(key.shape[0], -1)
+        torch_npu._npu_rotary_embedding(
+            positions,
+            query,
+            key,
+            self.head_size,
+            self.cos_sin_cache,
+            neox_style,
+        )
+    return query.view(query_shape), key.view(key_shape)
+
+
+def native_rope_deepseek_forward(self,
+                                 positions: torch.Tensor,
+                                 query: torch.Tensor,
+                                 key: torch.Tensor,
+                                 offsets: Optional[torch.Tensor] = None):
+    if len(key.shape) == 2:
+        key = key[:, None, :]
+    # Note: we implement the non neox_style method with shuffle the last dim and neox style
+    # calculation method which is also more compute friendly to the ascend machine
+    # https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/blob/main/modeling_deepseek.py
+    neox_style = True
+    if self.is_neox_style is False:
+        b, h_q, d = query.shape
+        query = query.view(b, h_q, d // 2, 2).transpose(3,
+                                                        2).reshape(b, h_q, d)
+        b, h_k, d = key.shape
+        key = key.view(b, h_k, d // 2, 2).transpose(3, 2).reshape(b, h_k, d)
+    q_pe, k_pe = rope_forward_oot(self, positions, query, key, offsets,
+                                  neox_style)
+    return q_pe, k_pe
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., :x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Inverse dim formula to find dim based on number of rotations
+def yarn_find_correction_dim(num_rotations,
+                             dim,
+                             base=10000,
+                             max_position_embeddings=2048):
+    # Note: use torch instead of math to solve MTP compilation error.
+    return (dim * torch.log(
+        torch.tensor(max_position_embeddings) /
+        (num_rotations * 2 * torch.pi))) / (2 * torch.log(torch.tensor(base)))
+
+
+def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+# Find dim range bounds based on rotations
+def yarn_find_correction_range(low_rot,
+                               high_rot,
+                               dim,
+                               base=10000,
+                               max_position_embeddings=2048):
+    # Note: use torch instead of math to solve MTP compilation error.
+    low = torch.floor(
+        yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings))
+    high = torch.ceil(
+        yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings))
+    # Note: use torch instead of max/min to solve MTP compilation error.
+    return torch.clamp(low, min=0), torch.clamp(high, max=dim - 1)
+
+
+def yarn_linear_ramp_mask(min_value, max_value, dim):
+    # Note: The if conditional branch is not used here
+    # to solve MTP compilation error.
+    max_value += (min_value == max_value).float() * 0.001
+    linear_func = (torch.arange(dim, dtype=torch.float32) -
+                   min_value) / (max_value - min_value)
+    ramp_func = torch.clamp(linear_func, 0, 1)
+    return ramp_func
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos[position_ids]
+    sin = sin[position_ids]
+    cos = cos[:, None, None, :]
+    sin = sin[:, None, None, :]
+
+    if len(q.shape) == 3:
+        q = q[:, :, None, :]
+    if len(k.shape) == 2:
+        k = k[:, None, None, :]
+    elif len(k.shape) == 3:
+        k = k[:, :, None, :]
+
+    b, h_q, s, d = q.shape
+    q = q.view(b, h_q, s, d // 2, 2).transpose(4, 3).reshape(b, h_q, s, d)
+
+    b, h_k, s, d = k.shape
+    k = k.view(b, h_k, s, d // 2, 2).transpose(4, 3).reshape(b, h_k, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+
+    q_embed = q_embed.view(b, h_q, d)
+    k_embed = k_embed.view(b, h_k, d)
+
+    return q_embed, k_embed
+
+
+def _set_cos_sin_cache(self, max_seq_len, device, dtype):
+    dim = self.rotary_dim
+
+    freq_extra = 1.0 / (self.base**(
+        torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
+    freq_inter = 1.0 / (self.scaling_factor * self.base**(
+        torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
+
+    low, high = yarn_find_correction_range(
+        self.beta_fast,
+        self.beta_slow,
+        dim,
+        self.base,
+        self.max_position_embeddings,
+    )
+    inv_freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2).to(
+        device=device, dtype=torch.float32)
+    inv_freq = freq_inter * (1 - inv_freq_mask) + freq_extra * inv_freq_mask
+    self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    t = torch.arange(max_seq_len, device=device, dtype=torch.float32)
+
+    freqs = torch.outer(t, inv_freq)
+    cos_cached = torch.cat([freqs, freqs], dim=-1).cos() * self.mscale
+    sin_cached = torch.cat([freqs, freqs], dim=-1).sin() * self.mscale
+    cos_cached = cos_cached.to(dtype)
+    sin_cached = sin_cached.to(dtype)
+    cache = torch.cat([freqs.cos() * self.mscale,
+                       freqs.sin() * self.mscale],
+                      dim=-1).to(dtype)
+    self.register_buffer("cos_sin_cache", cache, persistent=False)
+    self.register_buffer("cos_cached", cos_cached, persistent=False)
+    self.register_buffer("sin_cached", sin_cached, persistent=False)
+
+
+def __set_cos_sin_cache(self, seq_len, device, dtype):
+    inv_freq = 1.0 / (self.base**(torch.arange(
+        0, self.rotary_dim, 2, device=device, dtype=torch.float32) *
+                                  (1 / self.rotary_dim)))
+    self.register_buffer("inv_freq", inv_freq)
+
+    t = torch.arange(self.max_position_embeddings,
+                     device=self.inv_freq.device,
+                     dtype=torch.float32)
+    freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+
+    emb = torch.cat((freqs, freqs), dim=-1)
+    self.register_buffer("cos", emb.cos().to(dtype=dtype), persistent=False)
+    self.register_buffer("sin", emb.sin().to(dtype=dtype), persistent=False)
+    self.embed = F.embedding
+
+
+_original_re_init = RotaryEmbedding.__init__
+
+
+def qwen_rope_init_func(
+    self,
+    head_size: int,
+    rotary_dim: int,
+    max_position_embeddings: int,
+    base: float,
+    is_neox_style: bool,
+    dtype: torch.dtype,
+) -> None:
+    _original_re_init(self, head_size, rotary_dim, max_position_embeddings,
+                      base, is_neox_style, dtype)
+    if get_ascend_config().torchair_graph_config.enabled:
+        __set_cos_sin_cache(self,
+                            seq_len=max_position_embeddings,
+                            device="npu",
+                            dtype=dtype)
+
+
+def rope_forward(
+    self,
+    positions: torch.Tensor,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    offsets: Optional[torch.Tensor] = None,
+    is_neox_style_override: Optional[bool] = None,
+    max_seq_len: Optional[int] = None,
+    is_prefill: Optional[bool] = True,
+    is_qwen_torchair: Optional[bool] = False,
+):
+    if get_ascend_config().torchair_graph_config.enabled \
+            and is_qwen_torchair and not is_prefill:
+        if max_seq_len is not None and torch.gt(max_seq_len,
+                                                self.max_position_embeddings):
+            __set_cos_sin_cache(self,
+                                seq_len=max_seq_len,
+                                device=query.device,
+                                dtype=torch.float32)
+
+        # bsnd/bnsd
+        if positions is not None:
+            cos = self.embed(positions, self.cos)
+            sin = self.embed(positions, self.sin)
+            self.cos_embed = cos
+            self.sin_embed = sin
+        else:
+            cos = self.cos_embed
+            sin = self.sin_embed
+
+        query = query.view(*query.shape[:-1], -1, self.head_size).contiguous()
+        key = key.view(*key.shape[:-1], -1, self.head_size).contiguous()
+
+        cos = cos.unsqueeze(-2).unsqueeze(-2)
+        sin = sin.unsqueeze(-2).unsqueeze(-2)
+
+        query = query.unsqueeze(1)
+        key = key.unsqueeze(1)
+
+        q_embed, k_embed = torch_npu.npu_apply_rotary_pos_emb(
+            query, key, cos, sin)
+        return q_embed.flatten(-2), k_embed.flatten(-2)
+    else:
+        return rope_forward_oot(self, positions, query, key, offsets,
+                                is_neox_style_override,
+                                is_qwen_torchair)  # type: ignore
+
+
+def deepseek_rope_init_func(
+    self,
+    head_size: int,
+    rotary_dim: int,
+    max_position_embeddings: int,
+    base: int,
+    is_neox_style: bool,
+    scaling_factor: float,
+    dtype: torch.dtype,
+    *,
+    extrapolation_factor: float = 1,
+    attn_factor: float = 1,
+    beta_fast: int = 32,
+    beta_slow: int = 1,
+    mscale: float = 1,
+    mscale_all_dim: float = 0,
+) -> None:
+    self.scaling_factor = scaling_factor
+    self.extrapolation_factor = extrapolation_factor
+    self.attn_factor = attn_factor
+    self.beta_fast = beta_fast
+    self.beta_slow = beta_slow
+    # Get n-d magnitude scaling corrected for interpolation.
+    self.mscale = float(
+        yarn_get_mscale(self.scaling_factor, float(mscale)) /
+        yarn_get_mscale(self.scaling_factor, float(mscale_all_dim)) *
+        attn_factor)
+    super(DeepseekScalingRotaryEmbedding,
+          self).__init__(head_size, rotary_dim, max_position_embeddings, base,
+                         is_neox_style, dtype)
+
+    # NOTE: For ascend friendly computing, reorder sin and cos cache
+    self.max_seq_len = math.ceil(max_position_embeddings * scaling_factor)
+    _set_cos_sin_cache(self, self.max_seq_len, dtype=dtype, device="npu")

vllm_npu/torchair/ops/torchair_vocab_parallel_embedding.py

@@ -0,0 +1,38 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+from vllm.distributed import tensor_model_parallel_all_reduce
+
+
+def vocab_embedding_forward(self, input_):
+    if self.tp_size > 1:
+        # Build the mask.
+        masked_input, input_mask = self._get_masked_input_and_mask(
+            input_, self.shard_indices.org_vocab_start_index,
+            self.shard_indices.org_vocab_end_index,
+            self.shard_indices.num_org_vocab_padding,
+            self.shard_indices.added_vocab_start_index,
+            self.shard_indices.added_vocab_end_index)
+    else:
+        masked_input = input_
+    # Get the embeddings.
+    output_parallel = self.quant_method.embedding(self, masked_input.long())
+    # Mask the output embedding.
+    if self.tp_size > 1:
+        output_parallel.masked_fill_(input_mask.unsqueeze(-1), 0)
+    # Reduce across all the model parallel GPUs.
+    output = tensor_model_parallel_all_reduce(output_parallel)
+    return output