@@ -1,23 +1,12 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
from typing import Callable, Optional, Tuple, Union
|
||||
from collections.abc import Callable
|
||||
|
||||
import torch
|
||||
|
||||
from vllm_ascend.utils import is_310p
|
||||
|
||||
if is_310p():
|
||||
from vllm.lora.ops.torch_ops import (bgmv_expand, bgmv_expand_slice,
|
||||
bgmv_shrink, sgmv_expand,
|
||||
sgmv_expand_slice, sgmv_shrink)
|
||||
else:
|
||||
from vllm_ascend.lora.lora_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
|
||||
|
||||
from vllm_ascend.lora.utils import refresh_all_lora_classes
|
||||
from vllm_ascend.utils import AscendDeviceType, get_ascend_device_type
|
||||
|
||||
|
||||
# The platforms that are compatible with the PyTorch-native implementation can
|
||||
@@ -29,11 +18,36 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
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 __init__(self, max_num_batched_tokens: int, max_batches: int, device: torch.device | str, **kwargs):
|
||||
PunicaWrapperBase.__init__(self, max_num_batched_tokens, max_batches, device)
|
||||
refresh_all_lora_classes()
|
||||
self.lora_config = kwargs.get("lora_config")
|
||||
if get_ascend_device_type() == AscendDeviceType._310P or (
|
||||
self.lora_config is not None and self.lora_config.max_lora_rank >= 128
|
||||
):
|
||||
from vllm.lora.ops.torch_ops import (
|
||||
bgmv_expand,
|
||||
bgmv_expand_slice,
|
||||
bgmv_shrink,
|
||||
sgmv_expand,
|
||||
sgmv_expand_slice,
|
||||
sgmv_shrink,
|
||||
)
|
||||
else:
|
||||
from vllm_ascend.lora.lora_ops import (
|
||||
bgmv_expand,
|
||||
bgmv_expand_slice,
|
||||
bgmv_shrink,
|
||||
sgmv_expand,
|
||||
sgmv_expand_slice,
|
||||
sgmv_shrink,
|
||||
)
|
||||
self.bgmv_expand = bgmv_expand
|
||||
self.bgmv_expand_slice = bgmv_expand_slice
|
||||
self.bgmv_shrink = bgmv_shrink
|
||||
self.sgmv_expand = sgmv_expand
|
||||
self.sgmv_expand_slice = sgmv_expand_slice
|
||||
self.sgmv_shrink = sgmv_shrink
|
||||
|
||||
def _shrink_prefill(
|
||||
self,
|
||||
@@ -42,10 +56,10 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
w_t_all: torch.Tensor,
|
||||
scale: float,
|
||||
):
|
||||
#No LoRA request, so return directly
|
||||
# No LoRA request, so return directly
|
||||
if self.no_lora:
|
||||
return
|
||||
sgmv_shrink(
|
||||
self.sgmv_shrink(
|
||||
x,
|
||||
w_t_all,
|
||||
y,
|
||||
@@ -60,7 +74,7 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
w_t_all: torch.Tensor,
|
||||
scale: float,
|
||||
):
|
||||
bgmv_shrink(x, w_t_all, y, self.token_lora_indices, scale)
|
||||
self.bgmv_shrink(x, w_t_all, y, self._get_token_lora_indices(x), scale)
|
||||
|
||||
def _expand_prefill(
|
||||
self,
|
||||
@@ -69,10 +83,10 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
w_t_all: torch.Tensor,
|
||||
add_inputs: bool,
|
||||
):
|
||||
#No LoRA request, so return directly
|
||||
# No LoRA request, so return directly
|
||||
if self.no_lora:
|
||||
return
|
||||
sgmv_expand(
|
||||
self.sgmv_expand(
|
||||
x,
|
||||
w_t_all,
|
||||
y,
|
||||
@@ -87,7 +101,7 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
w_t_all: torch.Tensor,
|
||||
add_inputs: bool,
|
||||
):
|
||||
bgmv_expand(x, w_t_all, y, self.token_lora_indices, add_inputs)
|
||||
self.bgmv_expand(x, w_t_all, y, self._get_token_lora_indices(x), add_inputs)
|
||||
|
||||
def _expand_slice_prefill(
|
||||
self,
|
||||
@@ -98,10 +112,10 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
y_slice_size: int,
|
||||
add_inputs: bool,
|
||||
):
|
||||
#No LoRA request, so return directly
|
||||
# No LoRA request, so return directly
|
||||
if self.no_lora:
|
||||
return
|
||||
sgmv_expand_slice(
|
||||
self.sgmv_expand_slice(
|
||||
x,
|
||||
w_t_all,
|
||||
y,
|
||||
@@ -120,8 +134,18 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
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)
|
||||
self.bgmv_expand_slice(
|
||||
x,
|
||||
w_t_all,
|
||||
y,
|
||||
self._get_token_lora_indices(x),
|
||||
y_offset,
|
||||
y_slice_size,
|
||||
add_inputs,
|
||||
)
|
||||
|
||||
def _get_token_lora_indices(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return torch.narrow(self._token_lora_indices, 0, 0, x.size(0))
|
||||
|
||||
def _apply_expand(
|
||||
self,
|
||||
@@ -138,13 +162,10 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
GEMM of lora'b.
|
||||
"""
|
||||
|
||||
expand_slice_fun: Callable = (self._expand_slice_prefill
|
||||
if self.is_prefill else
|
||||
self._expand_slice_decode)
|
||||
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):
|
||||
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.
|
||||
@@ -155,14 +176,18 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
"""
|
||||
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: 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):
|
||||
def add_shrink(
|
||||
self,
|
||||
y: 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
|
||||
@@ -184,43 +209,38 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
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)
|
||||
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:
|
||||
def add_expand(
|
||||
self,
|
||||
y: torch.Tensor,
|
||||
x: tuple[torch.Tensor, ...] | torch.Tensor,
|
||||
lora_b_stacked: 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]
|
||||
y[:, offset:offset+slice] += x[i] @ lora_b_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
|
||||
offset_start (int): The starting position of y, defaults to 0
|
||||
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,
|
||||
@@ -233,12 +253,9 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
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:
|
||||
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.
|
||||
|
||||
@@ -253,30 +270,31 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
"""
|
||||
|
||||
# Embedding layer only need expand op
|
||||
expand_fun: Callable = (self._expand_prefill
|
||||
if self.is_prefill else self._expand_decode)
|
||||
expand_fun: Callable = self._expand_prefill if self.is_prefill else self._expand_decode
|
||||
x = x.to(torch.float32)
|
||||
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:
|
||||
def add_lora_linear(
|
||||
self,
|
||||
y: torch.Tensor,
|
||||
x: torch.Tensor,
|
||||
lora_a_stacked: tuple[torch.Tensor, ...],
|
||||
lora_b_stacked: tuple[torch.Tensor, ...],
|
||||
scale: float,
|
||||
output_slices: tuple[int, ...],
|
||||
*,
|
||||
buffer: tuple[torch.Tensor, ...] | None = 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, :, :]
|
||||
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]
|
||||
|
||||
@@ -292,37 +310,116 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
"""
|
||||
|
||||
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)))
|
||||
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)
|
||||
self.add_expand(y, buffer, lora_b_stacked, 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:
|
||||
def add_lora_fused_moe(
|
||||
self,
|
||||
y: torch.Tensor,
|
||||
x: torch.Tensor,
|
||||
lora_a_stacked: tuple[torch.Tensor, ...],
|
||||
lora_b_stacked: tuple[torch.Tensor, ...],
|
||||
*,
|
||||
topk_weights: torch.Tensor | None = None,
|
||||
sorted_token_ids: torch.Tensor | None = None,
|
||||
expert_ids: torch.Tensor,
|
||||
num_tokens_post_padded: torch.Tensor | None = None,
|
||||
max_lora_rank: int = 0,
|
||||
top_k_num: int = 1,
|
||||
shrink_config=None,
|
||||
expand_config=None,
|
||||
adapter_enabled: torch.Tensor,
|
||||
mul_routed_weight: bool = False,
|
||||
fully_sharded: bool = False,
|
||||
offset: int = 0,
|
||||
token_lora_mapping: torch.Tensor | None = None,
|
||||
) -> None:
|
||||
"""
|
||||
Ascend-native fused MoE LoRA (v2): static-shape per-row gather via the
|
||||
same bgmv_shrink/bgmv_expand AscendC kernels (csrc/kernels/bgmv_*.cpp)
|
||||
used by the dense Linear LoRA layers, instead of grouping rows by a
|
||||
data-dependent ``torch.unique`` over active LoRA ids. The previous
|
||||
``torch.unique``/``nonzero`` version produced output whose *shape*
|
||||
depended on tensor values, which ACL Graph capture cannot record
|
||||
(it failed with an `aclnnUnique2` error as soon as `enforce_eager`
|
||||
was turned off) -- every tensor below has a shape that depends only
|
||||
on input shapes, never on values, so this stays graph-capturable.
|
||||
|
||||
Rows are already one-token-per-row (top_k_num=1). Each row needs the
|
||||
LoRA slot for (lora_id, expert_id), so we fold both into a single
|
||||
gather index into a ``[max_loras * num_experts, ...]`` view of the
|
||||
existing per-(lora, expert) weight stacks:
|
||||
combined_idx[row] = lora_id[row] * num_experts + expert_id[row]
|
||||
or -1 when the row has no active adapter, mirroring the -1 sentinel
|
||||
``PunicaWrapperBase.token_lora_indices`` already uses. bgmv_shrink/
|
||||
bgmv_expand skip any row whose index is negative (leaving the
|
||||
zero-initialized shrink buffer / unmodified ``y`` in place), so
|
||||
inactive rows get a zero delta for free -- no Python-level branching
|
||||
needed.
|
||||
"""
|
||||
del sorted_token_ids, num_tokens_post_padded, max_lora_rank
|
||||
del shrink_config, expand_config, fully_sharded
|
||||
assert top_k_num == 1, "Ascend MoE LoRA v1 expects pre-expanded rows (top_k_num=1)."
|
||||
if token_lora_mapping is None:
|
||||
token_lora_mapping = self.token_lora_indices
|
||||
|
||||
x2d = x.view(-1, x.shape[-1])
|
||||
y2d = y.view(-1, y.shape[-1])
|
||||
expert_idx = expert_ids.view(-1).to(torch.long)
|
||||
num_experts = lora_a_stacked[0].shape[1]
|
||||
|
||||
lora_idx_safe = token_lora_mapping.clamp(min=0)
|
||||
enabled = (token_lora_mapping >= 0) & adapter_enabled[lora_idx_safe].bool()
|
||||
combined_idx = torch.where(
|
||||
enabled,
|
||||
lora_idx_safe * num_experts + expert_idx,
|
||||
torch.full_like(token_lora_mapping, -1),
|
||||
).contiguous()
|
||||
|
||||
# bgmv_shrink writes fp32 (its Y_T); bgmv_expand reads fp32 (its X_T),
|
||||
# so the shrink buffer is fp32.
|
||||
rank = lora_a_stacked[0].shape[-2]
|
||||
shrink_out = torch.zeros((x2d.shape[0], rank), dtype=torch.float32, device=x2d.device)
|
||||
|
||||
cur_offset = offset
|
||||
for slice_idx in range(len(lora_a_stacked)):
|
||||
# lora_a_stacked[s]/lora_b_stacked[s]: [max_loras, num_experts, rank, *].
|
||||
# Flattening the leading two dims turns "gather by (lora, expert)"
|
||||
# into "the plain per-row gather" to reuse bgmv_shrink/bgmv_expand.
|
||||
a = lora_a_stacked[slice_idx]
|
||||
b = lora_b_stacked[slice_idx]
|
||||
out_size = b.shape[-2]
|
||||
a_flat = a.view(-1, rank, a.shape[-1])
|
||||
b_flat = b.view(-1, out_size, rank)
|
||||
|
||||
self.bgmv_shrink(x2d, a_flat, shrink_out, combined_idx, 1.0)
|
||||
|
||||
delta = shrink_out
|
||||
if mul_routed_weight and topk_weights is not None:
|
||||
delta = shrink_out * topk_weights.view(-1, 1)
|
||||
|
||||
self.bgmv_expand_slice(delta, b_flat, y2d, combined_idx, cur_offset, out_size, add_inputs=True)
|
||||
cur_offset += out_size
|
||||
|
||||
def add_lora_logits(
|
||||
self,
|
||||
y: torch.Tensor,
|
||||
x: torch.Tensor,
|
||||
lora_a_stacked: torch.Tensor,
|
||||
lora_b_stacked: torch.Tensor,
|
||||
scale,
|
||||
*,
|
||||
buffer: torch.Tensor | None = None,
|
||||
**kwargs,
|
||||
) -> None:
|
||||
"""
|
||||
Applies lora specifically for LogitsProcessorWithLoRA.
|
||||
|
||||
@@ -344,13 +441,11 @@ class PunicaWrapperNPU(PunicaWrapperBase):
|
||||
r = lora_b_stacked.size(-1)
|
||||
|
||||
if buffer is None:
|
||||
buffer = torch.zeros((x.size(0), r),
|
||||
dtype=torch.float32,
|
||||
device=x.device)
|
||||
buffer = torch.zeros((x.size(0), r), dtype=torch.float32, device=x.device)
|
||||
|
||||
indices = self.sampler_indices
|
||||
indices = torch.narrow(self._sampler_indices, 0, 0, x.size(0))
|
||||
|
||||
bgmv_shrink(x, lora_a_stacked, buffer, indices, scale)
|
||||
bgmv_expand(buffer, lora_b_stacked, y, indices, add_inputs=True)
|
||||
self.bgmv_shrink(x, lora_a_stacked, buffer, indices, scale)
|
||||
self.bgmv_expand(buffer, lora_b_stacked, y, indices, add_inputs=True)
|
||||
|
||||
y = y.view_as(y_org)
|
||||
|
||||
Reference in New Issue
Block a user