File size: 5,971 Bytes
d02d576 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | from typing import List, Optional, Tuple
import torch
if torch.version.hip is not None:
# ROCM custom allreduce
def init_custom_ar(
meta: torch.Tensor,
rank_data: torch.Tensor,
handles: List[str],
offsets: List[int],
rank: int,
full_nvlink: bool,
) -> int:
return torch.ops.sgl_kernel.init_custom_ar.default(
meta, rank_data, handles, offsets, rank, full_nvlink
)
def all_reduce_reg(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
torch.ops.sgl_kernel.all_reduce_reg.default(fa, inp, out)
def all_reduce_unreg(
fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor, out: torch.Tensor
) -> None:
torch.ops.sgl_kernel.all_reduce_unreg.default(fa, inp, reg_buffer, out)
def deterministic_all_reduce_reg(
fa: int, inp: torch.Tensor, out: torch.Tensor
) -> None:
torch.ops.sgl_kernel.deterministic_all_reduce_reg.default(fa, inp, out)
def deterministic_all_reduce_unreg(
fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor, out: torch.Tensor
) -> None:
torch.ops.sgl_kernel.deterministic_all_reduce_unreg.default(
fa, inp, reg_buffer, out
)
def dispose(fa: int) -> None:
torch.ops.sgl_kernel.dispose.default(fa)
def meta_size() -> int:
return torch.ops.sgl_kernel.meta_size.default()
def register_buffer(
fa: int, t: torch.Tensor, handles: List[str], offsets: List[int]
) -> None:
return torch.ops.sgl_kernel.register_buffer.default(fa, t, handles, offsets)
def get_graph_buffer_ipc_meta(fa: int) -> Tuple[torch.Tensor, List[int]]:
return torch.ops.sgl_kernel.get_graph_buffer_ipc_meta.default(fa)
def register_graph_buffers(
fa: int, handles: List[str], offsets: List[List[int]]
) -> None:
torch.ops.sgl_kernel.register_graph_buffers.default(fa, handles, offsets)
def allocate_meta_buffer(size: int) -> torch.Tensor:
return torch.ops.sgl_kernel.allocate_meta_buffer.default(size)
def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
return torch.ops.sgl_kernel.get_meta_buffer_ipc_handle.default(inp)
# ROCM quick allreduce
def init_custom_qr(
rank: int, world_size: int, qr_max_size: Optional[int] = None
) -> int:
return torch.ops.sgl_kernel.init_custom_qr.default(
world_size, rank, qr_max_size
)
def qr_get_handle(fa: int) -> torch.Tensor:
return torch.ops.sgl_kernel.qr_get_handle.default(fa)
def qr_open_handles(fa: int, handles: list[torch.Tensor]) -> None:
torch.ops.sgl_kernel.qr_open_handles.default(fa, handles)
def qr_all_reduce(
fa: int,
profile: int,
inp: torch.Tensor,
out: torch.Tensor,
cast_bf162half: bool,
) -> None:
torch.ops.sgl_kernel.qr_all_reduce.default(
fa, profile, inp, out, cast_bf162half
)
def qr_destroy(fa: int) -> None:
torch.ops.sgl_kernel.qr_destroy.default(fa)
def qr_max_size() -> int:
return torch.ops.sgl_kernel.qr_max_size.default()
# mscclpp
def mscclpp_generate_unique_id() -> bytes:
raise NotImplementedError()
def mscclpp_init_context(
unique_id: bytes,
rank: int,
world_size: int,
scratch: torch.Tensor,
put_buffer: torch.Tensor,
nranks_per_node: int,
rank_to_node: List[int],
rank_to_ib: List[int],
context_selection: int,
) -> int:
raise NotImplementedError()
def mscclpp_allreduce(
context: int, inp: torch.Tensor, out: torch.Tensor, nthreads: int, nblocks: int
) -> None:
raise NotImplementedError()
else:
def init_custom_ar(
ipc_tensors: List[int], rank_data: torch.Tensor, rank: int, full_nvlink: bool
) -> int:
return torch.ops.sgl_kernel.init_custom_ar.default(
ipc_tensors, rank_data, rank, full_nvlink
)
def dispose(fa: int) -> None:
torch.ops.sgl_kernel.dispose.default(fa)
def all_reduce(
fa: int,
inp: torch.Tensor,
out: torch.Tensor,
reg_buffer: int,
reg_buffer_sz_bytes: int,
) -> None:
torch.ops.sgl_kernel.all_reduce.default(
fa, inp, out, reg_buffer, reg_buffer_sz_bytes
)
def get_graph_buffer_ipc_meta(fa) -> Tuple[List[int], List[int]]:
return torch.ops.sgl_kernel.get_graph_buffer_ipc_meta.default(fa)
def register_buffer(fa: int, fake_ipc_ptrs: List[int]) -> None:
return torch.ops.sgl_kernel.register_buffer.default(fa, fake_ipc_ptrs)
def register_graph_buffers(
fa: int, handles: List[List[int]], offsets: List[List[int]]
) -> None:
torch.ops.sgl_kernel.register_graph_buffers.default(fa, handles, offsets)
def meta_size() -> int:
return torch.ops.sgl_kernel.meta_size.default()
def mscclpp_generate_unique_id() -> torch.Tensor:
return torch.ops.sgl_kernel.mscclpp_generate_unique_id.default()
def mscclpp_init_context(
unique_id: torch.Tensor,
rank: int,
world_size: int,
scratch: torch.Tensor,
put_buffer: torch.Tensor,
nranks_per_node: int,
rank_to_node: List[int],
rank_to_ib: List[int],
context_selection: int,
) -> int:
return torch.ops.sgl_kernel.mscclpp_init_context.default(
unique_id,
rank,
world_size,
scratch,
put_buffer,
nranks_per_node,
rank_to_node,
rank_to_ib,
context_selection,
)
def mscclpp_allreduce(
context: int, inp: torch.Tensor, out: torch.Tensor, nthreads: int, nblocks: int
) -> None:
torch.ops.sgl_kernel.mscclpp_allreduce.default(
context, inp, out, nthreads, nblocks
)
|