[V1] LoRA - Add triton kernels for V1 (#13096)

Signed-off-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-03-10 17:27:53 -04:00
committed by GitHub
co-authored by Varun Sundar Rabindranath
parent 0967110e42
commit 5ff0d32580
11 changed files with 1165 additions and 191 deletions
+2 -2
View File
@@ -54,7 +54,7 @@ _LORA_A_PTR_DICT: Dict[Tuple[int, ...], Tuple[torch.tensor, ...]] = {}
_LORA_B_PTR_DICT: Dict[Tuple[int, ...], Tuple[torch.tensor, ...]] = {}
def _get_lora_a_ptr(lora_a_weights: List[torch.Tensor], device: str):
def _get_lora_a_ptr(lora_a_weights: List[torch.Tensor], device: torch.device):
"""
`_LORA_A_PTR_DICT` collects the required information during `profile_run`,
After this, it remains constant and subsequent usage is through LUT.
@@ -100,7 +100,7 @@ def _get_lora_a_ptr(lora_a_weights: List[torch.Tensor], device: str):
def _get_lora_b_ptr(lora_weights: List[torch.Tensor], offset_start: int,
device: str):
device: torch.device):
"""
`_LORA_B_PTR_DICT` collects the required information during `profile_run`,
After this, it remains constant and subsequent usage is through LUT.
+11
View File
@@ -0,0 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
from vllm.lora.ops.triton_ops.v1.v1_expand import v1_expand
from vllm.lora.ops.triton_ops.v1.v1_kernel_metadata import V1KernelMeta
from vllm.lora.ops.triton_ops.v1.v1_shrink import v1_shrink
__all__ = [
"v1_expand",
"v1_shrink",
"V1KernelMeta",
]
+282
View File
@@ -0,0 +1,282 @@
# SPDX-License-Identifier: Apache-2.0
"""
Based on:
Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023).
Punica: Multi-Tenant LoRA Serving.
https://arxiv.org/abs/2310.18547
"""
from typing import List
import torch
import triton
import triton.language as tl
from vllm.lora.ops.triton_ops.kernel_utils import do_expand_kernel
from vllm.lora.ops.triton_ops.utils import _get_lora_b_ptr
from vllm.utils import direct_register_custom_op
@triton.jit
def _v1_expand_kernel(
input_ptr,
lora_ptr,
out_ptr,
M,
N,
K,
token_indices_sorted_by_lora_ids,
num_tokens_per_lora,
lora_token_start_loc,
lora_ids,
slice_start_loc,
input_d0_stride,
input_d1_stride,
input_d2_stride, # 1
ls_d0_ptr,
ls_d1_ptr,
ls_d2_ptr, # 1
output_d0_stride,
output_d1_stride, # 1
output_hs_ptr,
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
BLOCK_K: tl.constexpr,
EVEN_K: tl.constexpr,
ADD_INPUTS: tl.constexpr,
CAST_TYPE: tl.constexpr,
SLICE_NUM: tl.constexpr,
SAME_STRIDE: tl.constexpr):
cta_n_num = tl.cdiv(N, BLOCK_N)
cta_m_num = tl.cdiv(M, BLOCK_M)
pid_mn = tl.program_id(axis=0)
pid_m = pid_mn % cta_m_num
pid_n = (pid_mn // cta_m_num) % cta_n_num
slice_id = tl.program_id(axis=1)
lora_idx = tl.program_id(axis=2)
lora_id = tl.load(lora_ids + lora_idx)
if lora_id == -1:
# Early exit for the no-lora case.
return
lora_m_size = tl.load(num_tokens_per_lora + lora_idx)
cta_m_offset = pid_m * BLOCK_M
if cta_m_offset >= lora_m_size:
# Early exit CTA.
return
# When the output dimensions of each slice are the same,cur_n=N, otherwise
# cur_n=tl.load(output_hs_ptr + slice_id), this situation exists in GQA's
# qkv linear.
curr_N = N if SAME_STRIDE else tl.load(output_hs_ptr + slice_id)
if pid_n * BLOCK_N >= curr_N:
# Early exit CTA.
return
# num rows this CTA should process.
cta_m_len = min(BLOCK_M, lora_m_size - cta_m_offset)
# Identify all rows that this CTA should process.
lora_m_indices_start = tl.load(lora_token_start_loc + lora_idx)
cta_lora_seq_indices = (token_indices_sorted_by_lora_ids +
lora_m_indices_start + cta_m_offset)
# Load all relevant row indices.
offset_m = tl.arange(0, BLOCK_M) % cta_m_len
ram = tl.load(cta_lora_seq_indices + offset_m)
do_expand_kernel(
pid_n,
lora_id,
slice_id,
input_ptr,
lora_ptr,
out_ptr,
curr_N,
K,
cta_m_len,
ram, # array identifying the rows of Input ptr to operate on
slice_start_loc,
# input ptr strides
input_d0_stride,
input_d1_stride,
input_d2_stride,
# lora ptr strides
ls_d0_ptr,
ls_d1_ptr,
ls_d2_ptr,
# out ptr strides
output_d0_stride,
output_d1_stride,
# constants
BLOCK_M,
BLOCK_N,
BLOCK_K,
SAME_STRIDE,
SLICE_NUM,
EVEN_K,
CAST_TYPE,
ADD_INPUTS)
@torch.inference_mode()
def _v1_expand(
inputs: torch.Tensor, # shape [num_slices, num_tokens, lora_rank]
lora_b_weights: List[
torch.Tensor], # shape [num_lora, hidden_size, lora_rank]
output_tensor: torch.
Tensor, # shape [num_tokens, hidden_size * num_slices]
token_lora_mapping: torch.Tensor, # shape [num_tokens]
token_indices_sorted_by_lora_ids: torch.Tensor, # shape [num_tokens]
num_tokens_per_lora: torch.Tensor, # shape [max-loras + 1]
lora_token_start_loc: torch.Tensor, # shape [max-loras + 2]
lora_ids: torch.Tensor, # shape [max-loras + 1]
offset_start: int = 0,
add_inputs: bool = False,
) -> None:
"""
Args:
inputs (torch.Tensor): input tensor
lora_b_weights (List[torch.Tensor]): lora'b weight
output_tensor (torch.Tensor): output tensor
token_lora_mapping (torch.Tensor): A tensor mapping each input token
to the lora-id related to that token. A value of -1 indicates that
LoRA doesn't apply to that token.
token_indices_sorted_by_lora_ids (torch.Tensor): Row/Token indices from
the A matrix grouped by LoRA IDs.
num_tokens_per_lora (torch.Tensor): num_tokens_per_lora[i] is the number
of tokens that are to be processed by LoRA ID lora_ids[i]
lora_token_start_loc (torch.Tensor): A cumulative sum of
num_tokens_per_lora. lora_token_start_loc[0] is always 0 so that
lora_token_start_loc[i], along with num_tokens_per_lora[i]
identifies the the region in token_indices_sorted_by_lora_ids that
LoRA lora_ids[i] should process.
lora_ids (torch.Tensor): LoRA ids to process.
offset_start (int, optional): Offset start for output_tensor.
Defaults to 0.
add_inputs (bool, optional): Whether to add the input tensor to the
output tensor. Defaults to False.
"""
assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
for weight in lora_b_weights:
assert weight.dtype in [torch.float16, torch.bfloat16]
assert inputs.size(0) == len(lora_b_weights)
assert output_tensor.is_contiguous()
# metadata sanity check.
assert token_lora_mapping.size(0) == token_indices_sorted_by_lora_ids.size(
0)
assert lora_ids.size(0) == num_tokens_per_lora.size(0)
assert lora_token_start_loc.size(0) == lora_ids.size(0) + 1
(slice_start_tensor, lora_ptr_tensor, lora_strides_d0_tensor,
lora_strides_d1_tensor, lora_strides_d2_tensor, hidden_sizes_tensor,
same_stride, MAX_N) = _get_lora_b_ptr(lora_b_weights, offset_start,
inputs.device)
K = lora_b_weights[0].shape[-1] # K= rank
M = inputs.size(1)
ADD_INPUTS = add_inputs
MAX_LORAS = lora_ids.size(0)
CAST_TYPE = False
NUM_SLICES = len(lora_b_weights)
# Triton kernel configs.
BLOCK_M = 64
BLOCK_N = 128
BLOCK_K = 16
NUM_WARPS = 4
NUM_CTAS = 1
NUM_STAGES = 2
MAX_NREG = None
EVEN_K = K % BLOCK_K == 0 # type: ignore
if inputs.dtype == torch.float32 and lora_b_weights[0].dtype in [
torch.float16,
torch.bfloat16,
]:
CAST_TYPE = True
# TODO (varun): This grid formulation maximizes parallelization at the
# cost of wasteful thread block launch when only a few input tokens require
# LoRA. This might not be the best in all cases.
grid = (
triton.cdiv(M, BLOCK_M) * triton.cdiv(MAX_N, BLOCK_N),
NUM_SLICES,
# Each LoRA receives its own set of thread blocks for output
# computation. If some LoRA doesn't have any tokens to process, its
# thread blocks simply exit.
MAX_LORAS,
)
_v1_expand_kernel[grid](
inputs,
lora_ptr_tensor,
output_tensor,
M,
MAX_N,
K,
token_indices_sorted_by_lora_ids,
num_tokens_per_lora,
lora_token_start_loc,
lora_ids,
slice_start_tensor,
inputs.stride(0),
inputs.stride(1),
inputs.stride(2),
lora_strides_d0_tensor,
lora_strides_d1_tensor,
lora_strides_d2_tensor,
output_tensor.stride(0),
output_tensor.stride(1),
hidden_sizes_tensor,
BLOCK_M,
BLOCK_N,
BLOCK_K,
EVEN_K,
ADD_INPUTS,
CAST_TYPE,
NUM_SLICES,
same_stride,
num_warps=NUM_WARPS,
num_ctas=NUM_CTAS,
num_stages=NUM_STAGES,
maxnreg=MAX_NREG,
)
return
def _v1_expand_fake(
inputs: torch.Tensor,
lora_b_weights: List[torch.Tensor],
output_tensor: torch.Tensor,
token_lora_mapping: torch.Tensor,
token_indices_sorted_by_lora_ids: torch.Tensor,
num_tokens_per_lora: torch.Tensor,
lora_token_start_loc: torch.Tensor,
lora_ids: torch.Tensor,
offset_start: int = 0,
add_inputs: bool = False,
) -> None:
return
try:
direct_register_custom_op(
op_name="v1_expand",
op_func=_v1_expand,
mutates_args=["output_tensor"],
fake_impl=_v1_expand_fake,
)
v1_expand = torch.ops.vllm.v1_expand
except AttributeError:
v1_expand = _v1_expand
@@ -0,0 +1,117 @@
# SPDX-License-Identifier: Apache-2.0
"""
V1 LoRA kernels metadata preparation utilities.
"""
from dataclasses import dataclass
from typing import Tuple, Union
import torch
@dataclass
class V1KernelMeta:
token_lora_mapping: torch.Tensor
token_indices_sorted_by_lora_ids: torch.Tensor
active_lora_ids: torch.Tensor
num_tokens_per_lora: torch.Tensor
lora_token_start_loc: torch.Tensor
@staticmethod
def make(max_loras: int, max_num_tokens: int,
device: Union[torch.device, str]) -> "V1KernelMeta":
token_lora_mapping = torch.empty(max_num_tokens,
dtype=torch.int32,
device=device)
token_indices_sorted_by_lora_ids = torch.empty(max_num_tokens,
dtype=torch.int32,
device=device)
# +1 because "no-lora" is also a possibility
# example: let max_loras be 3, active_lora_ids of [-1, 0, 2, 1]
# is a possibility.
active_lora_ids = torch.empty(max_loras + 1,
dtype=torch.int32,
device=device)
# using running example, [3, 10, 5, 2] is a possibility.
num_tokens_per_lora = torch.zeros(max_loras + 1,
dtype=torch.int32,
device=device)
# +2 for this because, the first index is always 0.
# using running example, lora_token_start_loc
# is [0, 3, 13, 18, 20].
lora_token_start_loc = torch.zeros(max_loras + 2,
dtype=torch.int32,
device=device)
return V1KernelMeta(
token_lora_mapping=token_lora_mapping,
token_indices_sorted_by_lora_ids=token_indices_sorted_by_lora_ids,
active_lora_ids=active_lora_ids,
num_tokens_per_lora=num_tokens_per_lora,
lora_token_start_loc=lora_token_start_loc)
def _reset(self):
self.active_lora_ids.fill_(-1)
self.num_tokens_per_lora.fill_(0)
self.lora_token_start_loc.fill_(0)
def prepare_tensors(self, token_lora_mapping: torch.Tensor) -> None:
"""
Prepare kernel metadata tensors for the current forward pass.
Args:
token_lora_tensor (torch.Tensor): Tensor containing lora indices
for each input token.
"""
self._reset()
num_tokens = token_lora_mapping.size(0)
# copy token lora mapping
self.token_lora_mapping[:num_tokens].copy_(token_lora_mapping,
non_blocking=True)
# token_indices_sorted_by_lora_ids
_, token_indices_sorted_by_lora_ids = torch.sort(token_lora_mapping,
stable=True)
# start gpu transfer
self.token_indices_sorted_by_lora_ids[:num_tokens].copy_(
token_indices_sorted_by_lora_ids, non_blocking=True)
# active_lora_ids, num_tokens_per_lora
lora_ids, num_tokens_per_lora = torch.unique(token_lora_mapping,
sorted=False,
return_counts=True)
self.active_lora_ids[:lora_ids.size(0)].copy_(lora_ids,
non_blocking=True)
self.num_tokens_per_lora[:num_tokens_per_lora.size(0)].copy_(
num_tokens_per_lora, non_blocking=True)
# lora_token_start_loc
lora_token_start_loc = torch.cumsum(num_tokens_per_lora, dim=0)
self.lora_token_start_loc[1:1 + lora_token_start_loc.size(0)].copy_(
lora_token_start_loc, non_blocking=True)
def meta_args(
self, token_nums: int
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor]:
"""
This function returns the kernel metadata required for the current
forward pass execution of the kernel. The function returns all the
metadata required by the kernel, in order, as a tuple, so it can be
unpacked directly during the v1_shrink/v1_expand function call.
Args:
token_nums (int): Number of input tokens in the current forward
pass.
"""
return (self.token_lora_mapping[:token_nums],
self.token_indices_sorted_by_lora_ids[:token_nums],
self.num_tokens_per_lora, self.lora_token_start_loc,
self.active_lora_ids)
+236
View File
@@ -0,0 +1,236 @@
# SPDX-License-Identifier: Apache-2.0
"""
Based on:
Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023).
Punica: Multi-Tenant LoRA Serving.
https://arxiv.org/abs/2310.18547
"""
from typing import List
import torch
import triton
import triton.language as tl
from vllm.lora.ops.triton_ops.kernel_utils import do_shrink_kernel
from vllm.lora.ops.triton_ops.utils import _get_lora_a_ptr
from vllm.utils import direct_register_custom_op
@triton.jit
def _v1_shrink_kernel(input_ptr, lora_ptr, out_ptr, M, N, K,
token_indices_sorted_by_lora_ids, num_tokens_per_lora,
lora_token_start_loc, lora_ids, scaling, input_d0_stride,
input_d1_stride, lora_d0_stride, lora_d1_stride,
lora_d2_stride, output_d0_stride, output_d1_stride,
output_d2_stride, BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr,
EVEN_K: tl.constexpr, SPLIT_K: tl.constexpr,
SLICE_NUM: tl.constexpr):
cta_n_num = tl.cdiv(N, BLOCK_N)
cta_m_num = tl.cdiv(M, BLOCK_M)
pid_sk_m_n = tl.program_id(axis=0)
pid_sk = pid_sk_m_n % SPLIT_K
pid_m = (pid_sk_m_n // SPLIT_K) % cta_m_num
pid_n = pid_sk_m_n // (SPLIT_K * cta_m_num) % cta_n_num
slice_id = tl.program_id(axis=1)
lora_idx = tl.program_id(axis=2)
lora_id = tl.load(lora_ids + lora_idx)
if lora_id == -1:
# Early exit for the no-lora case.
return
lora_m_size = tl.load(num_tokens_per_lora + lora_idx)
cta_m_offset = pid_m * BLOCK_M
if cta_m_offset >= lora_m_size:
# Early exit CTA.
return
# num rows this CTA should process.
cta_m_len = min(BLOCK_M, lora_m_size - cta_m_offset)
# Identify all rows that this CTA should process.
lora_m_indices_start = tl.load(lora_token_start_loc + lora_idx)
cta_lora_seq_indices = (token_indices_sorted_by_lora_ids +
lora_m_indices_start + cta_m_offset)
# Load all relevant row indices.
offset_m = tl.arange(0, BLOCK_M) % cta_m_len
ram = tl.load(cta_lora_seq_indices + offset_m)
do_shrink_kernel(
pid_n,
pid_sk,
slice_id,
lora_id,
input_ptr,
lora_ptr,
out_ptr,
N,
K,
cta_m_len,
ram, # array identifying the rows of Input ptr to operate on
# input strides
input_d0_stride,
input_d1_stride,
# lora strides
lora_d0_stride,
lora_d1_stride,
lora_d2_stride,
# output strides
output_d0_stride,
output_d1_stride,
output_d2_stride,
scaling,
BLOCK_M,
BLOCK_N,
BLOCK_K,
EVEN_K,
SPLIT_K,
SLICE_NUM)
@torch.inference_mode()
def _v1_shrink(
inputs: torch.Tensor, # shape [num_tokens, hidden_size]
lora_a_weights: List[
torch.Tensor], # shape [num_loras, lora_rank, hidden_size]
output_tensor: torch.Tensor, # shape [num_slices, num_tokens, lora_rank]
token_lora_mapping: torch.Tensor, # shape [num_tokens]
token_indices_sorted_by_lora_ids: torch.Tensor, # shape [num_tokens]
num_tokens_per_lora: torch.Tensor, # shape [max-loras + 1]
lora_token_start_loc: torch.Tensor, # shape [max-loras + 2]
lora_ids: torch.Tensor, # shape [max-loras + 1]
scaling: float,
) -> None:
"""
Args:
inputs (torch.Tensor): Input tensor
lora_a_weights (List[torch.Tensor]): LoRA weights
output_tensor (torch.Tensor): output tensor
token_lora_mapping (torch.Tensor): A tensor mapping each input token
to the lora-id related to that token. A value of -1 indicates that
LoRA doesn't apply to that token.
token_indices_sorted_by_lora_ids (torch.Tensor): Row/Token indices from
the A matrix grouped by LoRA IDs.
num_tokens_per_lora (torch.Tensor): num_tokens_per_lora[i] is the number
of tokens that are to be processed by LoRA ID lora_ids[i]
lora_token_start_loc (torch.Tensor): A cumulative sum of
num_tokens_per_lora. lora_token_start_loc[0] is always 0 so that
lora_token_start_loc[i], along with num_tokens_per_lora[i]
identifies the region in token_indices_sorted_by_lora_ids that
LoRA lora_ids[i] should process.
lora_ids (torch.Tensor): LoRA ids to process.
scaling (float): Scaling factor.
"""
assert inputs.dtype == lora_a_weights[0].dtype
assert inputs.dtype in [torch.float16, torch.bfloat16]
for weight in lora_a_weights:
assert weight.dtype in [torch.float16, torch.bfloat16]
assert inputs.size(1) == lora_a_weights[0].size(-1)
assert inputs.is_contiguous()
assert output_tensor.is_contiguous()
# metadata sanity check
assert token_lora_mapping.size(0) == token_indices_sorted_by_lora_ids.size(
0)
assert lora_ids.size(0) == num_tokens_per_lora.size(0)
assert lora_token_start_loc.size(0) == lora_ids.size(0) + 1
(lora_ptr_tensor, lora_strides_d0, lora_strides_d1,
lora_strides_d2) = _get_lora_a_ptr(lora_a_weights, inputs.device)
N, K = lora_a_weights[0].shape[-2:] # K=hidden_size,N=rank
M = inputs.size(0)
NUM_SLICES = len(lora_a_weights)
MAX_LORAS = lora_ids.size(0)
# Triton kernel configs
BLOCK_M = 32
BLOCK_N = 16
BLOCK_K = 256 if M < 128 else 32
SPLIT_K = 64 if M < 128 else 8
NUM_WARPS = 4
NUM_CTAS = 1
NUM_STAGES = 2
MAX_NREG = None
EVEN_K = K % (BLOCK_K * SPLIT_K) == 0 # type: ignore
# TODO (varun): This grid formulation maximizes parallelization at the
# cost of wasteful thread block launch when only few of the input tokens
# require LoRA. This might not be the best in all cases.
grid = (
SPLIT_K * triton.cdiv(M, BLOCK_M) * triton.cdiv(N, BLOCK_N),
NUM_SLICES,
# Each LoRA receives its own set of thread blocks for output
# computation. If some LoRA doesn't have any tokens to process, its
# thread blocks exit early.
MAX_LORAS,
)
_v1_shrink_kernel[grid](
inputs,
lora_ptr_tensor,
output_tensor,
M,
N,
K,
token_indices_sorted_by_lora_ids,
num_tokens_per_lora,
lora_token_start_loc,
lora_ids,
scaling,
inputs.stride(0),
inputs.stride(1),
lora_strides_d0,
lora_strides_d1,
lora_strides_d2,
output_tensor.stride(0),
output_tensor.stride(1),
output_tensor.stride(2),
BLOCK_M,
BLOCK_N,
BLOCK_K,
EVEN_K,
SPLIT_K,
NUM_SLICES,
num_warps=NUM_WARPS,
num_ctas=NUM_CTAS,
num_stages=NUM_STAGES,
maxnreg=MAX_NREG,
)
return
def _v1_shrink_fake(
inputs: torch.Tensor,
lora_a_weights: List[torch.Tensor],
output_tensor: torch.Tensor,
token_lora_mapping: torch.Tensor,
token_indices_sorted_by_lora_ids: torch.Tensor,
num_tokens_per_lora: torch.Tensor,
lora_token_start_loc: torch.Tensor,
lora_ids: torch.Tensor,
scaling: float,
) -> None:
return
try:
direct_register_custom_op(
op_name="v1_shrink",
op_func=_v1_shrink,
mutates_args=["output_tensor"],
fake_impl=_v1_shrink_fake,
)
v1_shrink = torch.ops.vllm.v1_shrink
except AttributeError:
v1_shrink = _v1_shrink