[Kernel] LoRA triton kernels support PDL (#27402)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li 2025-11-07 16:05:48 +08:00 committed by GitHub
parent a736e5ff77
commit 21b82f4ea2
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 68 additions and 17 deletions

View File

@ -6,6 +6,8 @@ import torch
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import direct_register_custom_op from vllm.utils.torch_utils import direct_register_custom_op
from .utils import supports_pdl
_LORA_PTR_DICT: dict[tuple[int, ...], torch.tensor] = {} _LORA_PTR_DICT: dict[tuple[int, ...], torch.tensor] = {}
@ -82,6 +84,8 @@ def _fused_moe_lora_kernel(
BLOCK_SIZE_K: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr, GROUP_SIZE_M: tl.constexpr,
SPLIT_K: tl.constexpr, SPLIT_K: tl.constexpr,
USE_GDC: tl.constexpr,
IS_PRIMARY: tl.constexpr,
): ):
pid = tl.program_id(axis=0) pid = tl.program_id(axis=0)
slice_id = tl.program_id(axis=1) slice_id = tl.program_id(axis=1)
@ -110,13 +114,11 @@ def _fused_moe_lora_kernel(
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr + lora_id) num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr + lora_id)
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
return return
# get the expert_id to process curr shard # get the expert_id to process curr shard
ind = lora_id * stride_el + pid_m ind = lora_id * stride_el + pid_m
expert_id = tl.load(expert_ids_ptr + ind, ind < max_loras * stride_el, -1) expert_id = tl.load(expert_ids_ptr + ind, ind < max_loras * stride_el, -1)
if expert_id == -1: if expert_id == -1:
return return
# get a_ptr,b_ptr,c_ptr # get a_ptr,b_ptr,c_ptr
cur_a_ptr = a_ptr + (slice_id % num_slice_a) * slice_a_size cur_a_ptr = a_ptr + (slice_id % num_slice_a) * slice_a_size
cur_b_ptr = tl.load(b_ptr + slice_id).to(tl.pointer_type(c_ptr.dtype.element_ty)) cur_b_ptr = tl.load(b_ptr + slice_id).to(tl.pointer_type(c_ptr.dtype.element_ty))
@ -149,12 +151,17 @@ def _fused_moe_lora_kernel(
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, grid_k): for k in range(0, grid_k):
k_remaining = K - k * (BLOCK_SIZE_K * SPLIT_K) k_remaining = K - k * (BLOCK_SIZE_K * SPLIT_K)
# pre-fetch lora weight
b = tl.load(b_ptrs, mask=offs_k[:, None] < k_remaining, other=0.0)
# GDC wait waits for ALL programs in the the prior kernel to complete
# before continuing.
if USE_GDC and not IS_PRIMARY:
tl.extra.cuda.gdc_wait()
a = tl.load( a = tl.load(
a_ptrs, a_ptrs,
mask=token_mask[:, None] & (offs_k[None, :] < k_remaining), mask=token_mask[:, None] & (offs_k[None, :] < k_remaining),
other=0.0, other=0.0,
) )
b = tl.load(b_ptrs, mask=offs_k[:, None] < k_remaining, other=0.0)
accumulator += tl.dot(a, b) accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block. # Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * SPLIT_K * stride_ak a_ptrs += BLOCK_SIZE_K * SPLIT_K * stride_ak
@ -163,12 +170,15 @@ def _fused_moe_lora_kernel(
if MUL_ROUTED_WEIGHT: if MUL_ROUTED_WEIGHT:
moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
accumulator = accumulator * moe_weight[:, None] accumulator = accumulator * moe_weight[:, None]
if USE_GDC and IS_PRIMARY:
# GDC launch dependents hints the runtime system to launch dependent kernels.
tl.extra.cuda.gdc_launch_dependents()
accumulator = accumulator.to(c_ptr.dtype.element_ty) accumulator = accumulator.to(c_ptr.dtype.element_ty)
# Write back the block of the output # Write back the block of the output
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
c_ptrs = cur_c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] c_ptrs = cur_c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
c_mask = token_mask[:, None] & (offs_cn[None, :] < N) c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
if SPLIT_K == 1: if SPLIT_K == 1:
tl.store(c_ptrs, accumulator, mask=c_mask) tl.store(c_ptrs, accumulator, mask=c_mask)
else: else:
@ -209,7 +219,7 @@ def _fused_moe_lora_shrink(
mul_routed_weight: bool = False, mul_routed_weight: bool = False,
) -> None: ) -> None:
w1_lora_a_stacked = lora_a_stacked[0] w1_lora_a_stacked = lora_a_stacked[0]
use_gdc = supports_pdl(qcurr_hidden_states.device)
shrink_config = { shrink_config = {
"BLOCK_SIZE_M": block_size_m, "BLOCK_SIZE_M": block_size_m,
"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_N": block_size_n,
@ -218,6 +228,8 @@ def _fused_moe_lora_shrink(
"num_warps": num_warps, "num_warps": num_warps,
"num_stages": num_stages, "num_stages": num_stages,
"SPLIT_K": split_k, "SPLIT_K": split_k,
"USE_GDC": use_gdc,
"launch_pdl": use_gdc, # triton kernel metadata
} }
b_ptr = _get_ptr(lora_a_stacked, device) b_ptr = _get_ptr(lora_a_stacked, device)
@ -229,7 +241,6 @@ def _fused_moe_lora_shrink(
len(lora_a_stacked), len(lora_a_stacked),
lora_a_stacked[0].shape[0], lora_a_stacked[0].shape[0],
) )
_fused_moe_lora_kernel[grid]( _fused_moe_lora_kernel[grid](
qcurr_hidden_states, qcurr_hidden_states,
b_ptr, b_ptr,
@ -261,6 +272,7 @@ def _fused_moe_lora_shrink(
num_slice_c=num_slices, num_slice_c=num_slices,
top_k=1 if mul_routed_weight else top_k_num, top_k=1 if mul_routed_weight else top_k_num,
MUL_ROUTED_WEIGHT=False, MUL_ROUTED_WEIGHT=False,
IS_PRIMARY=True,
**shrink_config, **shrink_config,
) )
@ -314,7 +326,7 @@ def _fused_moe_lora_expand(
dtype=output.dtype, dtype=output.dtype,
device=device, device=device,
) )
use_gdc = supports_pdl(a_intermediate_cache1.device)
expand_config = { expand_config = {
"BLOCK_SIZE_M": block_size_m, "BLOCK_SIZE_M": block_size_m,
"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_N": block_size_n,
@ -323,6 +335,8 @@ def _fused_moe_lora_expand(
"num_warps": num_warps, "num_warps": num_warps,
"num_stages": num_stages, "num_stages": num_stages,
"SPLIT_K": split_k, # Set split_k = 1 for expand calls "SPLIT_K": split_k, # Set split_k = 1 for expand calls
"USE_GDC": use_gdc,
"launch_pdl": use_gdc, # triton kernel metadata
} }
grid = lambda META: ( grid = lambda META: (
@ -361,6 +375,7 @@ def _fused_moe_lora_expand(
num_slice_c=num_slices, num_slice_c=num_slices,
top_k=1, top_k=1,
MUL_ROUTED_WEIGHT=mul_routed_weight, MUL_ROUTED_WEIGHT=mul_routed_weight,
IS_PRIMARY=False,
**expand_config, **expand_config,
) )
for i in range(num_slices): for i in range(num_slices):

View File

@ -22,6 +22,7 @@ def mm_k(
SPLIT_K: tl.constexpr, SPLIT_K: tl.constexpr,
CAST_TYPE: tl.constexpr, CAST_TYPE: tl.constexpr,
b_dtype: tl.constexpr, b_dtype: tl.constexpr,
USE_GDC: tl.constexpr,
): ):
""" """
Given a_ptr and b_ptr, that identify the rows of A (m x k) and columns of Given a_ptr and b_ptr, that identify the rows of A (m x k) and columns of
@ -45,19 +46,25 @@ def mm_k(
CAST_TYPE: if True, cast the values from the A matrix to the B CAST_TYPE: if True, cast the values from the A matrix to the B
matrix dtype. matrix dtype.
b_dtype: datatype of the B matrix b_dtype: datatype of the B matrix
USE_GDC: Whether to use PDL. True indicates use.
""" """
accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32) accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
for k in range(tl.cdiv(K, BLOCK_K * SPLIT_K)): for k in range(tl.cdiv(K, BLOCK_K * SPLIT_K)):
if EVEN_K: if EVEN_K:
tiled_a = tl.load(a_ptr) # pre-fetech lora weight
tiled_b = tl.load(b_ptr) tiled_b = tl.load(b_ptr)
if USE_GDC:
tl.extra.cuda.gdc_wait()
tiled_a = tl.load(a_ptr)
else: else:
tiled_a = tl.load(
a_ptr, mask=offset_k[None, :] < K - k * (BLOCK_K * SPLIT_K), other=0
)
tiled_b = tl.load( tiled_b = tl.load(
b_ptr, mask=offset_k[:, None] < K - k * (BLOCK_K * SPLIT_K), other=0 b_ptr, mask=offset_k[:, None] < K - k * (BLOCK_K * SPLIT_K), other=0
) )
if USE_GDC:
tl.extra.cuda.gdc_wait()
tiled_a = tl.load(
a_ptr, mask=offset_k[None, :] < K - k * (BLOCK_K * SPLIT_K), other=0
)
if CAST_TYPE: if CAST_TYPE:
tiled_a = tiled_a.to(b_dtype) tiled_a = tiled_a.to(b_dtype)
accumulator += tl.dot( accumulator += tl.dot(
@ -102,6 +109,7 @@ def do_expand_kernel(
EVEN_K: tl.constexpr, EVEN_K: tl.constexpr,
CAST_TYPE: tl.constexpr, CAST_TYPE: tl.constexpr,
ADD_INPUTS: tl.constexpr, ADD_INPUTS: tl.constexpr,
USE_GDC: tl.constexpr,
): ):
""" """
Given an array of integers that identifies the rows of A, ram, Given an array of integers that identifies the rows of A, ram,
@ -154,6 +162,7 @@ def do_expand_kernel(
# Compute the block matrix product. # Compute the block matrix product.
SPLIT_K = 1 SPLIT_K = 1
accumulator = mm_k( accumulator = mm_k(
a_ptr, a_ptr,
b_ptr, b_ptr,
@ -168,6 +177,7 @@ def do_expand_kernel(
SPLIT_K, SPLIT_K,
CAST_TYPE, CAST_TYPE,
cur_lora_ptr.dtype.element_ty, cur_lora_ptr.dtype.element_ty,
USE_GDC,
) )
tiled_c = accumulator.to(cur_lora_ptr.dtype.element_ty) tiled_c = accumulator.to(cur_lora_ptr.dtype.element_ty)
@ -223,6 +233,7 @@ def do_shrink_kernel(
EVEN_K: tl.constexpr, EVEN_K: tl.constexpr,
SPLIT_K: tl.constexpr, SPLIT_K: tl.constexpr,
SLICE_NUM: tl.constexpr, SLICE_NUM: tl.constexpr,
USE_GDC: tl.constexpr,
): ):
""" """
Given an array of integers that identifies the rows of A, ram, Given an array of integers that identifies the rows of A, ram,
@ -272,8 +283,11 @@ def do_shrink_kernel(
SPLIT_K, SPLIT_K,
False, False,
cur_lora_ptr.dtype.element_ty, cur_lora_ptr.dtype.element_ty,
False, # USE_GDC is always False in shrink kernel
) )
# GDC launch dependents hints the runtime system to launch dependent kernels.
if USE_GDC:
tl.extra.cuda.gdc_launch_dependents()
# Identify the C output pointers to store the results of the accumulator. # Identify the C output pointers to store the results of the accumulator.
offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
offset_cm = tl.arange(0, BLOCK_M) offset_cm = tl.arange(0, BLOCK_M)
@ -284,10 +298,10 @@ def do_shrink_kernel(
+ offset_cn[None, :] * output_d2_stride + offset_cn[None, :] * output_d2_stride
) )
c_mask = (offset_cm[:, None] < M_LEN) & (offset_cn[None, :] < N) c_mask = (offset_cm[:, None] < M_LEN) & (offset_cn[None, :] < N)
accumulator *= scaling accumulator *= scaling
# handles write-back with reduction-splitting # handles write-back with reduction-splitting
if SPLIT_K == 1: if SPLIT_K == 1:
tl.store(c_ptr, accumulator, mask=c_mask) tl.store(c_ptr, accumulator, mask=c_mask)
else: else:
tl.atomic_add(c_ptr, accumulator, mask=c_mask) tl.atomic_add(c_ptr, accumulator, mask=c_mask, sem="relaxed")

View File

@ -14,6 +14,8 @@ from vllm.lora.ops.triton_ops.utils import _get_lora_b_ptr, get_lora_op_configs
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import direct_register_custom_op from vllm.utils.torch_utils import direct_register_custom_op
from .utils import supports_pdl
@triton.jit @triton.jit
def _lora_expand_kernel( def _lora_expand_kernel(
@ -45,6 +47,7 @@ def _lora_expand_kernel(
CAST_TYPE: tl.constexpr, CAST_TYPE: tl.constexpr,
SLICE_NUM: tl.constexpr, SLICE_NUM: tl.constexpr,
SAME_STRIDE: tl.constexpr, SAME_STRIDE: tl.constexpr,
USE_GDC: tl.constexpr,
): ):
cta_n_num = tl.cdiv(N, BLOCK_N) cta_n_num = tl.cdiv(N, BLOCK_N)
cta_m_num = tl.cdiv(M, BLOCK_M) cta_m_num = tl.cdiv(M, BLOCK_M)
@ -121,6 +124,7 @@ def _lora_expand_kernel(
EVEN_K, EVEN_K,
CAST_TYPE, CAST_TYPE,
ADD_INPUTS, ADD_INPUTS,
USE_GDC,
) )
@ -236,7 +240,7 @@ def _lora_expand(
# thread blocks simply exit. # thread blocks simply exit.
MAX_LORAS, MAX_LORAS,
) )
use_gdc = supports_pdl(inputs.device)
_lora_expand_kernel[grid]( _lora_expand_kernel[grid](
inputs, inputs,
lora_ptr_tensor, lora_ptr_tensor,
@ -266,9 +270,11 @@ def _lora_expand(
CAST_TYPE, CAST_TYPE,
NUM_SLICES, NUM_SLICES,
same_stride, same_stride,
use_gdc,
num_warps=NUM_WARPS, num_warps=NUM_WARPS,
num_ctas=NUM_CTAS, num_ctas=NUM_CTAS,
num_stages=NUM_STAGES, num_stages=NUM_STAGES,
launch_pdl=use_gdc,
) )
return return

View File

@ -14,6 +14,8 @@ from vllm.lora.ops.triton_ops.utils import _get_lora_a_ptr, get_lora_op_configs
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import direct_register_custom_op from vllm.utils.torch_utils import direct_register_custom_op
from .utils import supports_pdl
@triton.jit @triton.jit
def _lora_shrink_kernel( def _lora_shrink_kernel(
@ -43,6 +45,7 @@ def _lora_shrink_kernel(
SPLIT_K: tl.constexpr, SPLIT_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr, GROUP_SIZE_M: tl.constexpr,
SLICE_NUM: tl.constexpr, SLICE_NUM: tl.constexpr,
USE_GDC: tl.constexpr,
): ):
cta_n_num = tl.cdiv(N, BLOCK_N) cta_n_num = tl.cdiv(N, BLOCK_N)
cta_m_num = tl.cdiv(M, BLOCK_M) cta_m_num = tl.cdiv(M, BLOCK_M)
@ -83,7 +86,6 @@ def _lora_shrink_kernel(
cta_lora_seq_indices = ( cta_lora_seq_indices = (
token_indices_sorted_by_lora_ids + lora_m_indices_start + cta_m_offset token_indices_sorted_by_lora_ids + lora_m_indices_start + cta_m_offset
) )
# Load all relevant row indices. # Load all relevant row indices.
offset_m = tl.arange(0, BLOCK_M) % cta_m_len offset_m = tl.arange(0, BLOCK_M) % cta_m_len
ram = tl.load(cta_lora_seq_indices + offset_m) ram = tl.load(cta_lora_seq_indices + offset_m)
@ -118,6 +120,7 @@ def _lora_shrink_kernel(
EVEN_K, EVEN_K,
SPLIT_K, SPLIT_K,
SLICE_NUM, SLICE_NUM,
USE_GDC,
) )
@ -217,7 +220,7 @@ def _lora_shrink(
# thread blocks exit early. # thread blocks exit early.
MAX_LORAS, MAX_LORAS,
) )
use_gdc = supports_pdl(inputs.device)
_lora_shrink_kernel[grid]( _lora_shrink_kernel[grid](
inputs, inputs,
lora_ptr_tensor, lora_ptr_tensor,
@ -245,9 +248,11 @@ def _lora_shrink(
SPLIT_K, SPLIT_K,
GROUP_SIZE_M, GROUP_SIZE_M,
NUM_SLICES, NUM_SLICES,
use_gdc,
num_warps=NUM_WARPS, num_warps=NUM_WARPS,
num_ctas=NUM_CTAS, num_ctas=NUM_CTAS,
num_stages=NUM_STAGES, num_stages=NUM_STAGES,
launch_pdl=use_gdc,
) )
return return

View File

@ -3,6 +3,7 @@
import functools import functools
import json import json
from functools import lru_cache
from pathlib import Path from pathlib import Path
from typing import Any from typing import Any
@ -10,6 +11,7 @@ import torch
from vllm import envs from vllm import envs
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.platforms import current_platform
logger = init_logger(__name__) logger = init_logger(__name__)
@ -282,3 +284,12 @@ def get_lora_op_configs(
assert config_data is not None assert config_data is not None
return config_data return config_data
@lru_cache
def supports_pdl(device: torch.device | None = None) -> bool:
"""
Refer to: https://github.com/triton-lang/triton/blob/v3.5.0/python/tutorials/11-programmatic-dependent-launch.py
"""
# PDL requires compute capability SM90 or above
return current_platform.is_cuda() and current_platform.has_device_capability(90)