File size: 13,951 Bytes
3087568 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 | """TP-aware linear layers (L2 operators).
TODO: Consider demoting to L1 -- these are TP-aware wrappers around L1
primitives (Linear, Fp8Linear, AllReduce) and are used by other L2 ops,
creating same-level dependencies. Moving them to L1 would make each L2
operator independently benchmarkable.
"""
from __future__ import annotations
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from ....infra.tp import _tp_size, _tp_rank
from ..L1.allreduce import AllReduce
def _get_fp8_linear_cls():
from ..L1.fp8_linear import Fp8Linear
return Fp8Linear
_FP8_BLOCK = 128
def _scale_shape(out_dim: int, in_dim: int) -> tuple[int, int]:
return (math.ceil(out_dim / _FP8_BLOCK), math.ceil(in_dim / _FP8_BLOCK))
class ColumnParallelLinear(nn.Module):
"""Splits output dim across TP ranks."""
def __init__(self, input_size: int, output_size: int, bias: bool = False,
quant_config: dict | None = None):
super().__init__()
tp = _tp_size()
assert output_size % tp == 0
self.output_size_per_partition = output_size // tp
self.use_fp8 = quant_config is not None
if self.use_fp8:
self.weight = nn.Parameter(
torch.empty(self.output_size_per_partition, input_size,
dtype=torch.float8_e4m3fn),
requires_grad=False,
)
self.weight_scale_inv = nn.Parameter(
torch.empty(*_scale_shape(self.output_size_per_partition, input_size),
dtype=torch.float32),
requires_grad=False,
)
self.weight.weight_loader = self._weight_loader
self.weight_scale_inv.weight_loader = self._scale_loader
self.linear_op = _get_fp8_linear_cls()()
else:
self.weight = nn.Parameter(torch.empty(self.output_size_per_partition, input_size))
self.weight.weight_loader = self._weight_loader
self.bias = nn.Parameter(torch.empty(self.output_size_per_partition)) if bias else None
if self.bias is not None:
self.bias.weight_loader = self._weight_loader
def _weight_loader(self, param, loaded_weight):
tp, rank = _tp_size(), _tp_rank()
shard = param.data.size(0)
loaded_weight = loaded_weight.narrow(0, rank * shard, shard)
param.data.copy_(loaded_weight)
def _scale_loader(self, param, loaded_weight):
tp, rank = _tp_size(), _tp_rank()
rows_per_shard = param.data.size(0)
loaded_weight = loaded_weight.narrow(0, rank * rows_per_shard, rows_per_shard)
param.data.copy_(loaded_weight)
def forward(self, x):
if self.use_fp8:
return self.linear_op(x, self.weight, self.weight_scale_inv, self.bias)
return F.linear(x, self.weight, self.bias)
class MergedColumnParallelLinear(nn.Module):
"""gate_proj + up_proj merged into one linear, sharded across TP."""
def __init__(self, input_size: int, output_sizes: list[int], bias: bool = False,
quant_config: dict | None = None, disable_tp: bool = False):
super().__init__()
tp = _tp_size()
self.disable_tp = disable_tp
self.output_sizes = output_sizes
total = sum(output_sizes)
if not disable_tp:
assert all(s % tp == 0 for s in output_sizes)
self.use_fp8 = quant_config is not None
effective_tp = 1 if disable_tp else tp
if self.use_fp8:
self.weight = nn.Parameter(
torch.empty(total // effective_tp, input_size, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
self.weight_scale_inv = nn.Parameter(
torch.empty(*_scale_shape(total // effective_tp, input_size), dtype=torch.float32),
requires_grad=False,
)
self.weight.weight_loader = self._weight_loader
self.weight_scale_inv.weight_loader = self._scale_loader
self.linear_op = _get_fp8_linear_cls()()
else:
self.weight = nn.Parameter(torch.empty(total // effective_tp, input_size))
self.weight.weight_loader = self._weight_loader
self.bias = None
if bias:
self.bias = nn.Parameter(torch.empty(total // tp))
self.bias.weight_loader = self._weight_loader
def _weight_loader(self, param, loaded_weight, shard_id: int | None = None):
tp, rank = _tp_size(), _tp_rank()
if shard_id is None:
# Fused weight: ``loaded_weight`` is the full ``[sum(output_sizes), in]``
# tensor. Recurse per-shard so each output block is sharded across
# TP ranks independently (mirrors vLLM's ``MergedColumnParallelLinear``
# weight loader when called without an explicit shard id).
offset = 0
for sid, sz in enumerate(self.output_sizes):
self._weight_loader(
param, loaded_weight.narrow(0, offset, sz), sid,
)
offset += sz
return
effective_tp = 1 if self.disable_tp else tp
shard_offset = sum(self.output_sizes[:shard_id]) // effective_tp
shard_size = self.output_sizes[shard_id] // effective_tp
dst = param.data.narrow(0, shard_offset, shard_size)
if self.disable_tp:
dst.copy_(loaded_weight)
else:
src = loaded_weight.chunk(tp, 0)[rank]
dst.copy_(src)
def _scale_loader(self, param, loaded_weight, shard_id: int):
tp, rank = _tp_size(), _tp_rank()
effective_tp = 1 if self.disable_tp else tp
shard_size_out = self.output_sizes[shard_id] // effective_tp
scale_rows = math.ceil(shard_size_out / _FP8_BLOCK)
shard_offset_out = sum(self.output_sizes[:shard_id]) // effective_tp
scale_offset = math.ceil(shard_offset_out / _FP8_BLOCK)
if self.disable_tp:
param.data.narrow(0, scale_offset, scale_rows).copy_(loaded_weight)
else:
src = loaded_weight.chunk(tp, 0)[rank]
param.data.narrow(0, scale_offset, scale_rows).copy_(src)
def forward(self, x):
if self.use_fp8:
return self.linear_op(x, self.weight, self.weight_scale_inv, self.bias)
return F.linear(x, self.weight, self.bias)
class QKVParallelLinear(nn.Module):
"""Q, K, V projections merged and sharded across TP."""
def __init__(self, hidden_size: int, head_size: int,
total_num_heads: int, total_num_kv_heads: int,
bias: bool = False, quant_config: dict | None = None):
super().__init__()
tp = _tp_size()
self.head_size = head_size
self.num_heads = total_num_heads // tp
# Replicate KV heads when not evenly divisible by TP
if total_num_kv_heads % tp == 0:
self.num_kv_heads = total_num_kv_heads // tp
self._replicate_kv = False
else:
self.num_kv_heads = total_num_kv_heads
self._replicate_kv = True
output_size = (self.num_heads + 2 * self.num_kv_heads) * head_size
self.use_fp8 = quant_config is not None
if self.use_fp8:
self.weight = nn.Parameter(
torch.empty(output_size, hidden_size, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
self.weight_scale_inv = nn.Parameter(
torch.empty(*_scale_shape(output_size, hidden_size), dtype=torch.float32),
requires_grad=False,
)
self.weight.weight_loader = self._weight_loader
self.weight_scale_inv.weight_loader = self._scale_loader
self.linear_op = _get_fp8_linear_cls()()
else:
self.weight = nn.Parameter(torch.empty(output_size, hidden_size))
self.weight.weight_loader = self._weight_loader
self.bias = None
if bias:
self.bias = nn.Parameter(torch.empty(output_size))
self.bias.weight_loader = self._weight_loader
def _weight_loader(self, param, loaded_weight, shard_id: str):
tp, rank = _tp_size(), _tp_rank()
if shard_id == "q":
shard_size = self.num_heads * self.head_size
shard_offset = 0
src = loaded_weight.chunk(tp, 0)[rank]
elif shard_id == "k":
shard_size = self.num_kv_heads * self.head_size
shard_offset = self.num_heads * self.head_size
src = loaded_weight if self._replicate_kv else loaded_weight.chunk(tp, 0)[rank]
else:
shard_size = self.num_kv_heads * self.head_size
shard_offset = self.num_heads * self.head_size + self.num_kv_heads * self.head_size
src = loaded_weight if self._replicate_kv else loaded_weight.chunk(tp, 0)[rank]
dst = param.data.narrow(0, shard_offset, shard_size)
dst.copy_(src)
def _scale_loader(self, param, loaded_weight, shard_id: str):
tp, rank = _tp_size(), _tp_rank()
if shard_id == "q":
shard_size = self.num_heads * self.head_size
shard_offset = 0
elif shard_id == "k":
shard_size = self.num_kv_heads * self.head_size
shard_offset = self.num_heads * self.head_size
else:
shard_size = self.num_kv_heads * self.head_size
shard_offset = self.num_heads * self.head_size + self.num_kv_heads * self.head_size
scale_rows = math.ceil(shard_size / _FP8_BLOCK)
scale_offset = math.ceil(shard_offset / _FP8_BLOCK)
src = loaded_weight.chunk(tp, 0)[rank]
param.data.narrow(0, scale_offset, scale_rows).copy_(src)
def forward(self, x):
if self.use_fp8:
return self.linear_op(x, self.weight, self.weight_scale_inv, self.bias)
return F.linear(x, self.weight, self.bias)
class ReplicatedLinear(nn.Module):
"""Full weight replicated on every TP rank (no sharding, no all-reduce)."""
def __init__(self, input_size: int, output_size: int, bias: bool = True,
quant_config: dict | None = None):
super().__init__()
self.use_fp8 = quant_config is not None
if self.use_fp8:
self.weight = nn.Parameter(
torch.empty(output_size, input_size, dtype=torch.float8_e4m3fn),
requires_grad=False,
)
self.weight_scale_inv = nn.Parameter(
torch.empty(*_scale_shape(output_size, input_size),
dtype=torch.float32),
requires_grad=False,
)
self.weight.weight_loader = lambda p, w: p.data.copy_(w)
self.weight_scale_inv.weight_loader = lambda p, w: p.data.copy_(w)
self.linear_op = _get_fp8_linear_cls()()
else:
self.weight = nn.Parameter(torch.empty(output_size, input_size))
self.weight.weight_loader = lambda p, w: p.data.copy_(w)
self.bias = nn.Parameter(torch.empty(output_size)) if bias else None
if self.bias is not None:
self.bias.weight_loader = lambda p, w: p.data.copy_(w)
def forward(self, x):
if self.use_fp8:
return self.linear_op(x, self.weight, self.weight_scale_inv, self.bias)
return F.linear(x, self.weight, self.bias)
class RowParallelLinear(nn.Module):
"""Splits input dim across TP ranks, all-reduces output."""
def __init__(self, input_size: int, output_size: int, bias: bool = False,
quant_config: dict | None = None, reduce_results: bool = True):
super().__init__()
tp = _tp_size()
assert input_size % tp == 0
self.input_size_per_partition = input_size // tp
self.tp_size = tp
self.tp_rank = _tp_rank()
self.reduce_results = reduce_results
self.use_fp8 = quant_config is not None
if self.use_fp8:
self.weight = nn.Parameter(
torch.empty(output_size, self.input_size_per_partition,
dtype=torch.float8_e4m3fn),
requires_grad=False,
)
self.weight_scale_inv = nn.Parameter(
torch.empty(*_scale_shape(output_size, self.input_size_per_partition),
dtype=torch.float32),
requires_grad=False,
)
self.weight.weight_loader = self._weight_loader
self.weight_scale_inv.weight_loader = self._scale_loader
self.linear_op = _get_fp8_linear_cls()()
else:
self.weight = nn.Parameter(torch.empty(output_size, self.input_size_per_partition))
self.weight.weight_loader = self._weight_loader
self.bias = nn.Parameter(torch.empty(output_size)) if bias else None
if self.bias is not None:
self.bias.weight_loader = lambda p, w: p.data.copy_(w)
self.allreduce = AllReduce()
def _weight_loader(self, param, loaded_weight):
tp, rank = _tp_size(), _tp_rank()
shard = param.data.size(1)
loaded_weight = loaded_weight.narrow(1, rank * shard, shard)
param.data.copy_(loaded_weight)
def _scale_loader(self, param, loaded_weight):
tp, rank = _tp_size(), _tp_rank()
cols_per_shard = param.data.size(1)
loaded_weight = loaded_weight.narrow(1, rank * cols_per_shard, cols_per_shard)
param.data.copy_(loaded_weight)
def forward(self, x):
if self.use_fp8:
y = self.linear_op(x, self.weight, self.weight_scale_inv,
self.bias if self.tp_rank == 0 else None)
else:
y = F.linear(x, self.weight, self.bias if self.tp_rank == 0 else None)
if self.reduce_results and self.tp_size > 1:
y = self.allreduce(y)
return y
|