File size: 6,486 Bytes
40ef767 | 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 | """Python autograd wrapper for selective-update Triton kernel.
Matches upstream ``mamba_ssm.ops.triton.selective_state_update`` ABI exactly:
state: (B, D, N) or (B, H, D, N)
x: (B, D) or (B, H, D)
dt: (B, D) or (B, H, D)
A: (D, N) or (H, D, N)
B: (B, N) or (B, G, N)
C: (B, N) or (B, G, N)
D: (D,) or (H, D) or None
z: (B, D) or (B, H, D) or None
dt_bias: (D,) or (H, D) or None
state_batch_indices: (B,) or None
"""
from __future__ import annotations
import importlib.util
from pathlib import Path
import torch
import triton
_TRITON_KERNELS_PATH = (
Path(__file__).resolve().parent / "triton_kernels.py"
)
def _load_triton_kernel():
spec = importlib.util.spec_from_file_location(
"selective_update.triton_kernels",
_TRITON_KERNELS_PATH,
)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module.selective_state_update_kernel
selective_state_update_kernel = None # populated on first call
def _get_kernel():
global selective_state_update_kernel
if selective_state_update_kernel is None:
selective_state_update_kernel = _load_triton_kernel()
return selective_state_update_kernel
def _resolve_tie_hdim(A, dt, dt_bias):
if A.dim() == 2:
a0 = A[0]
return a0.stride(0) == 0 and a0.stride(1) == 0 and dt.stride(-1) == 0 and (dt_bias.stride(-1) == 0 if dt_bias is not None else True)
return A.stride(-1) == 0 and A.stride(-2) == 0 and dt.stride(-1) == 0 and (dt_bias.stride(-1) == 0 if dt_bias is not None else True)
class SelectiveUpdateFunction(torch.autograd.Function):
"""Differentiable single-token selective state update.
Mirrors upstream ``selective_state_update`` shape contract exactly.
Backward raises NotImplementedError — training uses the patched
selective_scan reverse path instead.
"""
@staticmethod
def forward(ctx, state, x, dt, A, B, C, D=None, z=None, dt_bias=None, dt_softplus=False, state_batch_indices=None):
orig_state = state
has_heads = state.dim() > 3
if state.dim() == 3:
state = state.unsqueeze(1)
if x.dim() == 2:
x = x.unsqueeze(1)
if dt.dim() == 2:
dt = dt.unsqueeze(1)
if A.dim() == 2:
A = A.unsqueeze(0)
if B.dim() == 2:
B = B.unsqueeze(1)
if C.dim() == 2:
C = C.unsqueeze(1)
if D is not None and D.dim() == 1:
D = D.unsqueeze(0)
if z is not None and z.dim() == 2:
z = z.unsqueeze(1)
if dt_bias is not None and dt_bias.dim() == 1:
dt_bias = dt_bias.unsqueeze(0)
B_sz, H, D_sz, N_sz = state.shape
x_sz = x.shape[0]
if x.shape != (x_sz, H, D_sz):
raise ValueError(f"x shape {x.shape} does not match state batch/heads/dim")
if dt.shape != (x_sz, H, D_sz):
raise ValueError(f"dt shape {dt.shape} does not match x shape")
if A.shape != (H, D_sz, N_sz):
raise ValueError(f"A shape {A.shape} does not match (H, D, N)")
ngroups = B.shape[1]
if H % ngroups != 0:
raise ValueError(f"nheads {H} must be divisible by ngroups {ngroups}")
if B.shape[0] != x_sz or B.shape[2] != N_sz or C.shape[0] != x_sz or C.shape[2] != N_sz:
raise ValueError(f"B/C batch or dstate mismatch")
if D is not None and D.shape not in {(H, D_sz), (H,)}:
raise ValueError(f"D shape {D.shape} does not match (H, D) or (H,)")
if z is not None and z.shape != (x_sz, H, D_sz):
raise ValueError(f"z shape {z.shape} does not match x shape")
if dt_bias is not None and dt_bias.shape not in {(H, D_sz), (H,)}:
raise ValueError(f"dt_bias shape {dt_bias.shape} does not match (H, D) or (H,)")
if state_batch_indices is not None and state_batch_indices.shape != (x_sz,):
raise ValueError(f"state_batch_indices shape {state_batch_indices.shape} does not match (B,)")
tie_hdim = _resolve_tie_hdim(A, dt, dt_bias)
nheads_ratio = H // ngroups
out = torch.empty_like(x)
grid = lambda META: (triton.cdiv(D_sz, META["BLOCK_SIZE_M"]), x_sz, H)
BLOCK_M = 32 if N_sz <= 16 else (16 if N_sz <= 32 else (8 if N_sz <= 64 else (4 if N_sz <= 128 else 4)))
num_warps = 4 if N_sz <= 64 else 8
z_strides = ((z.stride(0), z.stride(1), z.stride(2)) if z is not None else (0, 0, 0))
_get_kernel()[grid](
state, x, dt, dt_bias, A, B, C, D, z, out, state_batch_indices,
x_sz, H, D_sz, N_sz, nheads_ratio,
state.stride(0), state.stride(1), state.stride(2), state.stride(3),
x.stride(0), x.stride(1), x.stride(2),
dt.stride(0), dt.stride(1), dt.stride(2),
dt_bias.stride(0) if dt_bias is not None else 0, dt_bias.stride(1) if dt_bias is not None else 0,
A.stride(0), A.stride(1), A.stride(2),
B.stride(0), B.stride(1), B.stride(2),
C.stride(0), C.stride(1), C.stride(2),
D.stride(0) if D is not None else 0, D.stride(1) if D is not None else 0,
z_strides[0], z_strides[1], z_strides[2],
out.stride(0), out.stride(1), out.stride(2),
dt_softplus,
tie_hdim,
BLOCK_M,
num_warps=num_warps,
num_stages=2,
)
if not has_heads:
out = out.squeeze(1)
state = state.squeeze(1)
ctx.save_for_backward(orig_state, x, dt, A, B, C, D, z, dt_bias, state, state_batch_indices)
ctx.tie_hdim = tie_hdim
ctx.dt_softplus = dt_softplus
ctx.has_heads = has_heads
ctx.nheads_ratio = nheads_ratio
return out
@staticmethod
def backward(ctx, grad_out):
raise NotImplementedError(
"SelectiveUpdateFunction backward is not implemented. "
"Use autograd through the patched selective_scan path for training."
)
def selective_state_update(state, x, dt, A, B, C, D=None, z=None, dt_bias=None, dt_softplus=False, state_batch_indices=None, **kwargs):
"""Dispatches single-token state updates to the custom Triton backend.
Shapes mirror upstream ``mamba_ssm.ops.triton.selective_state_update`` exactly.
"""
return SelectiveUpdateFunction.apply(
state, x, dt, A, B, C, D, z, dt_bias, dt_softplus, state_batch_indices
)
|