File size: 14,227 Bytes
919fd68 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 | """C/R-conditioned KDA recurrent linear attention expert.
Keeps exact Q/K/V/C/R attention as the global path; this expert carries
long-running sequence state. Forget/write gates are conditioned on intent
context ``C`` and relation ``R`` glyphs. Output blend is zero-init so existing
checkpoints remain identity-compatible.
Uses ``fla.ops.kda.chunk_kda`` for CUDA execution and a tensor-native reference
recurrence for CPU execution. A CUDA kernel/import failure is surfaced instead
of silently changing the production algorithm.
"""
from __future__ import annotations
from typing import cast
import torch
import torch.nn as nn
import torch.nn.functional as F
from resynthesis.config import GLYPH_DIM
RESYNTHESIS_KDA_LOG_DECAY_FLOOR = -5.0
class CRConditionedKDAExpert(nn.Module):
"""Channel-decayed recurrent attention conditioned on C/R planes."""
def __init__(
self,
hidden_size: int,
num_heads: int,
*,
glyph_dim: int = GLYPH_DIM,
head_dim: int | None = None,
) -> None:
super().__init__()
heads = max(1, int(num_heads))
width = int(hidden_size)
if width % heads != 0 and head_dim is None:
while heads > 1 and width % heads != 0:
heads -= 1
self.hidden_size = width
self.num_heads = heads
self.head_dim = int(head_dim) if head_dim is not None else width // heads
self.glyph_dim = int(glyph_dim)
inner = self.num_heads * self.head_dim
# Head-grouped projections retain the full recurrent state width without
# duplicating four dense hidden-by-hidden matrices per science layer.
self.q_proj = nn.Conv1d(
width,
inner,
kernel_size=1,
groups=self.num_heads,
bias=False,
)
self.k_proj = nn.Conv1d(
width,
inner,
kernel_size=1,
groups=self.num_heads,
bias=False,
)
self.v_proj = nn.Conv1d(
width,
inner,
kernel_size=1,
groups=self.num_heads,
bias=False,
)
self.out_proj = nn.Conv1d(
inner,
width,
kernel_size=1,
groups=self.num_heads,
bias=False,
)
# Keep the recurrent output head-normalized and apply an
# input-dependent full-rank gate before the output projection. This is
# intentionally a dense projection: grouping it by recurrent head would
# prevent the gate from coordinating channels learned by different
# heads.
self.output_gate_proj = nn.Linear(width, inner, bias=False)
self.output_norm = nn.RMSNorm(self.head_dim)
# Channel-wise forget (log-space raw) + scalar write, C/R conditioned.
self.forget_hidden_proj = nn.Conv1d(
width,
inner,
kernel_size=1,
groups=self.num_heads,
bias=True,
)
self.forget_intent_proj = nn.Linear(
self.glyph_dim,
inner,
bias=False,
)
self.forget_relation_proj = nn.Linear(
self.glyph_dim,
inner,
bias=False,
)
self.write_hidden_proj = nn.Conv1d(
width,
self.num_heads,
kernel_size=1,
groups=self.num_heads,
bias=True,
)
self.write_intent_proj = nn.Linear(
self.glyph_dim,
self.num_heads,
bias=False,
)
self.write_relation_proj = nn.Linear(
self.glyph_dim,
self.num_heads,
bias=False,
)
self.short_conv = nn.Conv1d(
width,
width,
kernel_size=3,
padding=2,
groups=width,
bias=False,
)
# Parameterize the finite log decay as
# ``g_min * sigmoid(exp(A_h) * z_h)``. ``A_h`` is one learned scalar per
# recurrent head, matching FLA's fused safe-gate contract. Its
# one-dimensional zero initialization is also important for checkpoint
# adoption: native graph migration preserves constructor values for
# vector parameters, so an older checkpoint resumes with exp(A_h)=1.
self.decay_log_scale = nn.Parameter(torch.zeros(self.num_heads))
self.blend_scale = nn.Parameter(torch.zeros(()))
self._reset()
def _reset(self) -> None:
for module in (
self.q_proj,
self.k_proj,
self.v_proj,
self.out_proj,
):
nn.init.xavier_uniform_(module.weight)
nn.init.xavier_uniform_(self.output_gate_proj.weight)
nn.init.ones_(self.output_norm.weight)
nn.init.zeros_(self.decay_log_scale)
nn.init.xavier_uniform_(self.forget_hidden_proj.weight)
forget_bias_t = self.forget_hidden_proj.bias
if forget_bias_t is None:
raise RuntimeError("KDA forget projection has no trained bias")
nn.init.zeros_(forget_bias_t)
nn.init.xavier_uniform_(self.forget_intent_proj.weight)
nn.init.xavier_uniform_(self.forget_relation_proj.weight)
nn.init.xavier_uniform_(self.write_hidden_proj.weight)
write_bias_t = self.write_hidden_proj.bias
if write_bias_t is None:
raise RuntimeError("KDA write projection has no trained bias")
nn.init.zeros_(write_bias_t)
nn.init.xavier_uniform_(self.write_intent_proj.weight)
nn.init.xavier_uniform_(self.write_relation_proj.weight)
nn.init.dirac_(self.short_conv.weight)
nn.init.zeros_(self.blend_scale)
@staticmethod
def _project_sequence(
projection: nn.Conv1d,
tensor: torch.Tensor,
) -> torch.Tensor:
projected_t = cast(torch.Tensor, projection(tensor.transpose(1, 2)))
return projected_t.transpose(1, 2)
def _reshape_heads(self, tensor: torch.Tensor) -> torch.Tensor:
batch, seq, _ = tensor.shape
return tensor.view(batch, seq, self.num_heads, self.head_dim)
def _bounded_log_decay(
self,
forget_raw: torch.Tensor,
) -> torch.Tensor:
"""Map learned decay logits into the finite recurrent log range."""
if (
forget_raw.ndim != 4
or forget_raw.shape[-2:] != (
self.num_heads,
self.head_dim,
)
):
raise ValueError("KDA forget-logit geometry differs")
decay_scale_t = self.decay_log_scale.exp().view(
1,
1,
self.num_heads,
1,
)
return torch.sigmoid(decay_scale_t * forget_raw).mul(
RESYNTHESIS_KDA_LOG_DECAY_FLOOR
)
def forward(
self,
hidden: torch.Tensor,
*,
intent_glyph_context: torch.Tensor,
relation_glyph_context: torch.Tensor,
) -> torch.Tensor:
if hidden.ndim != 3 or hidden.shape[-1] != self.hidden_size:
raise ValueError("KDA expert hidden geometry differs")
if intent_glyph_context.shape[:2] != hidden.shape[:2]:
raise ValueError("KDA intent context geometry differs")
if relation_glyph_context.shape[:2] != hidden.shape[:2]:
raise ValueError("KDA relation context geometry differs")
# Depthwise short conv on sequence (B,H,T) — causal via left pad trim.
conv_in = hidden.transpose(1, 2)
conv_out = self.short_conv(conv_in)[..., : hidden.shape[1]].transpose(1, 2)
x = F.silu(conv_out)
intent_t = intent_glyph_context.to(dtype=x.dtype)
relation_t = relation_glyph_context.to(dtype=x.dtype)
q = self._reshape_heads(self._project_sequence(self.q_proj, x))
k = self._reshape_heads(self._project_sequence(self.k_proj, x))
v = self._reshape_heads(self._project_sequence(self.v_proj, x))
q = F.normalize(q, dim=-1)
k = F.normalize(k, dim=-1)
# fla chunk_kda: g is forget in log space [B,T,H,K]; beta write [B,T,H]
forget_raw = (
self._project_sequence(self.forget_hidden_proj, x)
+ self.forget_intent_proj(intent_t)
+ self.forget_relation_proj(relation_t)
).view(
hidden.shape[0],
hidden.shape[1],
self.num_heads,
self.head_dim,
)
write_logits_t = (
self._project_sequence(self.write_hidden_proj, x)
+ self.write_intent_proj(intent_t)
+ self.write_relation_proj(relation_t)
)
out = self._run_kda_from_logits(
q,
k,
v,
forget_raw,
write_logits_t,
)
normalized_out = self.output_norm(out)
flat = normalized_out.reshape(hidden.shape[0], hidden.shape[1], -1)
output_gate_t = torch.sigmoid(self.output_gate_proj(x))
gated_flat = output_gate_t * flat
projected = self._project_sequence(
self.out_proj,
gated_flat.to(dtype=hidden.dtype),
)
return torch.tanh(self.blend_scale) * projected
def _run_kda_from_logits(
self,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
forget_raw: torch.Tensor,
write_logits: torch.Tensor,
) -> torch.Tensor:
"""Run KDA while retaining the native fused safe-gate CUDA path.
FLA's KDA backend owns the same lower-bounded recurrent activation.
Supplying the logits and per-head ``A_h`` directly lets it keep gate
activation, 16-token safe rescaling, and write sigmoid inside the
kernel. CPU and the one-token zero-state shortcut materialize the
identical equations explicitly.
"""
if q.is_cuda and q.shape[1] > 1:
from fla.ops.kda import chunk_kda # type: ignore[import-untyped]
out, _state = chunk_kda(
q.contiguous(),
k.contiguous(),
v.contiguous(),
forget_raw.contiguous(),
write_logits.contiguous(),
use_gate_in_kernel=True,
use_beta_sigmoid_in_kernel=True,
safe_gate=True,
lower_bound=RESYNTHESIS_KDA_LOG_DECAY_FLOOR,
A_log=self.decay_log_scale.contiguous(),
)
if not isinstance(out, torch.Tensor):
raise RuntimeError("FLA KDA returned a non-tensor output")
return out
return self._run_kda(
q,
k,
v,
self._bounded_log_decay(forget_raw),
torch.sigmoid(write_logits),
)
def _run_kda(
self,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
) -> torch.Tensor:
if q.shape[1] == 1:
# CAS bulk science attempts are isolated single-token recurrent
# sessions. With the exact zero initial state, decay and delta
# correction multiply zero, so KDA reduces to beta*v*(k·q).
# Keep a zero dependency on g because the reference recurrence
# produces a zero (not absent) forget-gate gradient.
q_t = q[:, 0]
k_t = k[:, 0]
v_t = v[:, 0]
beta_t = beta[:, 0].unsqueeze(-1).float()
scale = q_t.shape[-1] ** -0.5
alignment_t = (
(k_t.float() * q_t.float()).sum(dim=-1, keepdim=True)
* scale
)
forget_zero_t = g[:, :1].sum(dim=-1, keepdim=True).mul(0)
return (
beta_t.mul(v_t.float())
.mul(alignment_t)
.to(dtype=q.dtype)
.unsqueeze(1)
+ forget_zero_t.to(dtype=q.dtype)
)
# Triton/FLA kernels require CUDA tensors; CPU has an explicit reference
# implementation. The historical sequence-parallel canary split this
# recurrence and carried state in a Python loop. That was sequential,
# not context parallel. A full native launch is both faster and exact;
# real KDA context-parallel owners use the associative tensor packet in
# ``sequence_parallel.py`` rather than a host-flag simulation.
if q.is_cuda:
from fla.ops.kda import chunk_kda # type: ignore[import-untyped]
out, _state = chunk_kda(
q.contiguous(),
k.contiguous(),
v.contiguous(),
g.contiguous(),
beta.contiguous(),
)
if not isinstance(out, torch.Tensor):
raise RuntimeError("FLA KDA returned a non-tensor output")
return out
return self._reference_kda(q, k, v, g, beta)
@staticmethod
def _reference_kda(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
) -> torch.Tensor:
"""O(T) reference KDA / gated delta-rule recurrence for CPU."""
batch, seq, heads, dim = q.shape
value_dim = v.shape[-1]
state = q.new_zeros(batch, heads, dim, value_dim)
outputs = q.new_empty(batch, seq, heads, value_dim)
scale = dim ** -0.5
for t in range(seq):
alpha = torch.exp(g[:, t]).clamp(0.0, 1.0) # [B,H,K]
bt = beta[:, t].unsqueeze(-1).unsqueeze(-1) # [B,H,1,1]
kt = k[:, t].unsqueeze(-1) # [B,H,K,1]
vt = v[:, t].unsqueeze(-2) # [B,H,1,V]
state = state * alpha.unsqueeze(-1)
# Delta-rule correction then write.
read = torch.matmul(state.transpose(-1, -2), k[:, t].unsqueeze(-1))
state = state - bt * torch.matmul(kt, read.transpose(-1, -2))
state = state + bt * torch.matmul(kt, vt)
ot = torch.matmul(
state.transpose(-1, -2),
q[:, t].mul(scale).unsqueeze(-1),
).squeeze(-1)
outputs[:, t] = ot
return outputs
|