Add LoRA & Multi-LoRA support for V0.7.3 dev by Cherry Pick (#700)

### What this PR does / why we need it?
According to this [RFC](
#396 ) and
[this](#448), we pull
request relavant code to support (1) Multi-LoRA and (2) Multi-LoRA
Dynamic Serving.
LoRA reference is here: [LoRA
reference](https://docs.vllm.ai/en/latest/features/lora.html)

### Does this PR introduce _any_ user-facing change?
Following openai HTTP apis will be supported:
/v1/load_lora_adapter
/v1/unload_lora_adapter

### How was this patch tested?
git clone
[https://github.com/vllm-project/vllm.git](https://github.com/vllm-project/vllm.git)
cd vllm/examples/offline_inference/ && python3 multilora_inference.py

> [[Release]: vLLM Ascend v0.7.3 release checklist
](#644 (comment))

---------

Signed-off-by: paulyu <paulyu0307@gmail.com>
Co-authored-by: paulyu12 <507435917@qq.com>
Co-authored-by: paulyu <paulyu0307@gmail.com>

Author: 3 people

vllm_ascend/lora/__init__.py

@@ -0,0 +1,346 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+from vllm.lora.ops.torch_ops import (bgmv_expand, bgmv_expand_slice,
+                                     bgmv_shrink, sgmv_expand,
+                                     sgmv_expand_slice, sgmv_shrink)
+from vllm.lora.punica_wrapper.punica_base import PunicaWrapperBase
+
+
+# The platforms that are compatible with the PyTorch-native implementation can
+# inherit this class
+class PunicaWrapperNPU(PunicaWrapperBase):
+    """
+    PunicaWrapperNPU is designed to manage and provide metadata for the punica 
+    kernel. The main function is to maintain the state information for 
+    Multi-LoRA, and to provide the interface for the pytorch punica ops.
+    """
+
+    def __init__(self, max_num_batched_tokens: int, max_batches: int,
+                 device: Union[torch.device, str], **kwargs):
+        PunicaWrapperBase.__init__(self, max_num_batched_tokens, max_batches,
+                                   device)
+
+    def _shrink_prefill(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        scale: float,
+    ):
+        #No LoRA request, so return directly
+        if self.no_lora:
+            return
+        sgmv_shrink(
+            x,
+            w_t_all,
+            y,
+            *self.prefill_metadata,
+            scale,
+        )
+
+    def _shrink_decode(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        scale: float,
+    ):
+        bgmv_shrink(x, w_t_all, y, self.token_lora_indices, scale)
+
+    def _expand_prefill(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        add_inputs: bool,
+    ):
+        #No LoRA request, so return directly
+        if self.no_lora:
+            return
+        sgmv_expand(
+            x,
+            w_t_all,
+            y,
+            *self.prefill_metadata,
+            add_inputs,
+        )
+
+    def _expand_decode(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        add_inputs: bool,
+    ):
+        bgmv_expand(x, w_t_all, y, self.token_lora_indices, add_inputs)
+
+    def _expand_slice_prefill(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        y_offset: int,
+        y_slice_size: int,
+        add_inputs: bool,
+    ):
+        #No LoRA request, so return directly
+        if self.no_lora:
+            return
+        sgmv_expand_slice(
+            x,
+            w_t_all,
+            y,
+            *self.prefill_metadata,
+            y_offset,
+            y_slice_size,
+            add_inputs,
+        )
+
+    def _expand_slice_decode(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        y_offset: int,
+        y_slice_size: int,
+        add_inputs: bool,
+    ):
+        bgmv_expand_slice(x, w_t_all, y, self.token_lora_indices, y_offset,
+                          y_slice_size, add_inputs)
+
+    def _apply_expand(
+        self,
+        y: torch.Tensor,
+        x: torch.Tensor,
+        w_t_all: torch.Tensor,
+        y_offset: int,
+        y_slice_size: int,
+        add_inputs: bool = True,
+    ):
+        """
+        Perform the ` y[:,y_offset:y_offset+y_slice_size]+=x@w_t_all` 
+        computation, which is suitable for the
+        GEMM of lora'b.
+        """
+
+        expand_slice_fun: Callable = (self._expand_slice_prefill
+                                      if self.is_prefill else
+                                      self._expand_slice_decode)
+        expand_slice_fun(y, x, w_t_all, y_offset, y_slice_size, add_inputs)
+
+    def _apply_shrink(self, y: torch.Tensor, x: torch.Tensor,
+                      w_t_all: torch.Tensor, scale: float):
+        """
+        Perform the ` y+=x@w_t_all` computation, which is suitable for the
+        GEMM of lora'a.
+        When `is_prefill is` true, it indicates that it is currently the
+        prefill stage, and the `_shrink_prefill` function should be called.
+        Otherwise, it is the decode stage, and the _shrink_decode function
+        should be called.
+        """
+        y_org = y
+        y = y.view(-1, y.shape[-1])
+        shrink_fun: Callable = (self._shrink_prefill
+                                if self.is_prefill else self._shrink_decode)
+        shrink_fun(y, x, w_t_all, scale)
+        y = y.view_as(y_org)
+
+    def add_shrink(self, y: Union[Tuple[torch.Tensor, ...], torch.Tensor],
+                   x: torch.Tensor, lora_a_stacked: Tuple[torch.Tensor, ...],
+                   scale: float, **kwargs):
+        """
+        Performs GEMM  for multiple slices of lora_a.
+        When `is_prefill is` true, it indicates that it is currently the
+        prefill stage, and the `_shrink_prefill` function should be called.
+        Otherwise, it is the decode stage, and the _shrink_decode function
+        should be called.
+            
+        Semantics:
+        for i in range(len(lora_a_stacked)):
+            y[i] += (x @ lora_a_stacked[i]) * scale
+        
+        Args:
+            y (Union[Tuple[torch.Tensor, ...], torch.Tensor]): Output tensors
+            x (torch.Tensor): Input tensor
+            lora_a_stacked (Tuple[torch.Tensor, ...]): lora_a's weights
+            scale (float): Scaling factor for the operation
+        """
+
+        x = x.view(-1, x.shape[-1])
+        # TODO fuse these kernels
+        for slice_idx in range(len(lora_a_stacked)):
+            self._apply_shrink(y[slice_idx], x, lora_a_stacked[slice_idx],
+                               scale)
+
+    def add_expand(self,
+                   y: torch.Tensor,
+                   x: Union[Tuple[torch.Tensor, ...], torch.Tensor],
+                   lora_b_stacked: Tuple[torch.Tensor, ...],
+                   lora_bias_stacked: Optional[Tuple[torch.Tensor, ...]],
+                   output_slices: Tuple[int, ...],
+                   offset_start: int = 0,
+                   add_inputs=True,
+                   **kwargs) -> None:
+        """
+        Performs GEMM and bias addition for multiple slices of lora_b.
+      
+        Semantics:
+            for i in range(len(lora_b_stacked)):
+                slice = output_slices[i]
+                y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i] + 
+                    lora_bias_stacked[i] 
+                offset += slice
+            
+        Args:
+            y (torch.Tensor): Output tensor.
+            x (Union[Tuple[torch.Tensor, ...], torch.Tensor]): Input tensors
+            lora_b_stacked (Tuple[torch.Tensor, ...]): lora_b's weight
+            lora_bias_stacked (Optional[Tuple[torch.Tensor, ...]]): 
+                bias's weight
+            output_slices (Tuple[int, ...]): Every slice's size
+            add_inputs (bool):  Defaults to True.
+        """
+        y_org = y
+        y = y.view(-1, y.shape[-1])
+        offset_left = offset_start
+        if lora_bias_stacked is not None:
+            self._apply_bias(self.token_lora_indices, y, output_slices,
+                             lora_bias_stacked)
+        for slice_idx in range(len(lora_b_stacked)):
+            self._apply_expand(
+                y,
+                x[slice_idx],
+                lora_b_stacked[slice_idx],
+                offset_left,
+                output_slices[slice_idx],
+                add_inputs=add_inputs,
+            )
+            offset_left += output_slices[slice_idx]
+        y = y.view_as(y_org)
+
+    def add_lora_embedding(self,
+                           y: torch.Tensor,
+                           x: torch.Tensor,
+                           lora_b_stacked: torch.Tensor,
+                           add_inputs: bool = True,
+                           **kwargs) -> None:
+        """
+        Applies lora  specifically for VocabParallelEmbeddingWithLoRA.
+
+        Semantics:
+            y += x @ lora_b_stacked
+
+        Args:
+            y (torch.Tensor): Output tensor.
+            x (torch.Tensor): Input tensor.
+            lora_b_stacked (torch.Tensor): lora_b's weights.
+            add_inputs (bool): Default to True.
+        """
+
+        # Embedding layer only need expand op
+        expand_fun: Callable = (self._expand_prefill
+                                if self.is_prefill else self._expand_decode)
+        expand_fun(y, x, lora_b_stacked, add_inputs)
+
+    def add_lora_linear(self,
+                        y: torch.Tensor,
+                        x: torch.Tensor,
+                        lora_a_stacked: Tuple[torch.Tensor, ...],
+                        lora_b_stacked: Tuple[torch.Tensor, ...],
+                        lora_bias_stacked: Optional[Tuple[torch.Tensor, ...]],
+                        scale: float,
+                        output_slices: Tuple[int, ...],
+                        *,
+                        buffer: Optional[Tuple[torch.Tensor, ...]] = None,
+                        **kwargs) -> None:
+        """
+        Applicable to linear-related lora. 
+
+        Semantics:
+            for i in range(len(lora_a_stacked)):
+                y[i] += (
+                    x[i].unsqueeze(0)
+                    @ lora_a_stacked[indices[i], layer_idx, :, :]
+                    @ lora_b_stacked[indices[i], layer_idx, :, :]
+                    * scale
+                    ).squeeze(0)+lora_bias_stacked[i]
+
+        Args:
+            y (torch.Tensor): Output tensor. Will be changed in-place.
+            x (torch.Tensor): Input tensor
+            lora_a_stacked (Tuple[torch.Tensor, ...]): lora_a's weight.
+            lora_b_stacked (Tuple[torch.Tensor, ...]): lora_b's weight.
+            lora_bias_stacked (Optional[Tuple[torch.Tensor, ...]]): lora's bias.
+            scale (float): Scaling factor.
+            output_slices (Tuple[int, ...]): Every slice's size.
+            buffer (Optional[Tuple[torch.Tensor, ...]]): Defaults to None.
+        """
+
+        assert len(lora_a_stacked) == len(lora_b_stacked) == len(output_slices)
+        if lora_bias_stacked is not None:
+            assert len(lora_bias_stacked) == len(output_slices)
+            y = self._apply_bias(self.token_lora_indices, y, output_slices,
+                                 lora_bias_stacked)
+
+        if buffer is None:
+            r = lora_b_stacked[0].size(-1)
+            # We set the buffer to be float32 by default, consistent with the
+            # triton op
+            buffer = tuple(
+                torch.zeros(
+                    (x.size(0), r), dtype=torch.float32, device=x.device)
+                for _ in range(len(output_slices)))
+        self.add_shrink(buffer, x, lora_a_stacked, scale, **kwargs)
+        self.add_expand(y,
+                        buffer,
+                        lora_b_stacked,
+                        None,
+                        output_slices,
+                        add_inputs=True,
+                        **kwargs)
+
+    def add_lora_logits(self,
+                        y: torch.Tensor,
+                        x: torch.Tensor,
+                        lora_a_stacked: torch.Tensor,
+                        lora_b_stacked: torch.Tensor,
+                        scale,
+                        *,
+                        buffer: Optional[torch.Tensor] = None,
+                        **kwargs) -> None:
+        """
+        Applies lora  specifically for LogitsProcessorWithLoRA.
+        
+        Semantics:
+            buffer = (x @ lora_a_stacked) * scale
+            y += buffer @ lora_b_stacked
+
+        Args:
+            y (torch.Tensor): Output tensor.
+            x (torch.Tensor): Input tensor.
+            lora_a_stacked (torch.Tensor): lora_a's weights.
+            lora_b_stacked (torch.Tensor):lora_b's weights.
+            scale (float): Scaling factor.
+            buffer (Optional[torch.Tensor]):Default to None.
+        """
+        y_org = y
+        y = y.view(-1, y.shape[-1])
+        x = x.view(-1, x.shape[-1])
+        r = lora_b_stacked.size(-1)
+        if buffer is None:
+            # We set the buffer to be float32 by default, consistent with the
+            # triton op
+            buffer = torch.zeros((x.size(0), r),
+                                 dtype=torch.float32,
+                                 device=x.device)
+        # LogitsProcessorWithLoRA always using bgmv.
+        bgmv_shrink(x, lora_a_stacked, buffer, self.sampler_indices, scale)
+        bgmv_expand(buffer,
+                    lora_b_stacked,
+                    y,
+                    self.sampler_indices,
+                    add_inputs=True)
+        y = y.view_as(y_org)

vllm_ascend/lora/punica_wrapper/__init__.py

@@ -172,6 +172,10 @@ def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
             return "vllm_ascend.attention.attention.AscendMLAAttentionBackend"
         return "vllm_ascend.attention.attention.AscendAttentionBackend"
 
+    @classmethod
+    def get_punica_wrapper(cls) -> str:
+        return "vllm_ascend.lora.punica_wrapper.punica_npu.PunicaWrapperNPU"
+
     @classmethod
     def get_current_memory_usage(cls,
                                  device: Optional[torch.types.Device] = None