File size: 13,156 Bytes
6fda0c2 | 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 | #!/usr/bin/env python3
"""PAIRED-NIBBLE {0,4,8}: store the {0,4,8} quantization as a UNIFORM 4-bit (nibble) layout by splitting each
INT8 outlier into two nibbles -- low nibble in its own slot, high nibble in a paired slot -- and SKIPPING the
dead positions from storage (they are always 0 at storage, so they cost 0 bytes; they are still SCANNED as a
constant-0 state that accumulates one step of dB*x and contributes to the readout -- exact {0,4,8} semantics).
A static per-GROUP d_state permutation orders positions [outliers | int4 | dead] so every load is coalesced
(no gather). Dequant is bit-identical to {0,4,8}, so accuracy == {0,4,8} (no new approximation); the novelty
is the storage/kernel. Config dead=48/int4=64/int8=16: c_lo_out 8B + c_hi_out 8B + c_int4 32B = 48B = 3.0 b/elem
payload + fp16 s4,s8 (0.25) = 3.25 b/elem -- below PERM-PREFIX (5.25b) at full {0,4,8} accuracy, single uniform
nibble representation. Correctness-gated vs the identical PyTorch dequant->scan->requant."""
import os, torch
import triton
import triton.language as tl
D_STATE = 128
K8 = 16 # INT8 outliers (stored as 2 nibbles each)
K4 = 64 # INT4
KD = D_STATE - K8 - K4 # 48 dead (skipped from storage, scanned as 0)
@triton.jit
def _pairnib_kernel(
clo_ptr, chi_ptr, c4_ptr, s8_ptr, s4_ptr, # outlier-lo nibbles, outlier-hi nibbles, int4 nibbles, scales
x_ptr, dt_ptr, dtb_ptr, A_ptr, B_ptr, C_ptr, D_ptr, out_ptr,
batch, nheads, dim, dstate, k8, k4, kd, nhg,
stride_clo_b, stride_clo_h, stride_clo_d, stride_clo_n,
stride_chi_b, stride_chi_h, stride_chi_d, stride_chi_n,
stride_c4_b, stride_c4_h, stride_c4_d, stride_c4_n,
stride_sc_b, stride_sc_h, stride_sc_d,
stride_x_b, stride_x_h, stride_x_d,
stride_dt_b, stride_dt_h, stride_dt_d,
stride_dtb_h, stride_dtb_d,
stride_A_h, stride_A_d, stride_A_n,
stride_B_b, stride_B_g, stride_B_n,
stride_C_b, stride_C_g, stride_C_n,
stride_D_h, stride_D_d,
stride_out_b, stride_out_h, stride_out_d,
DT_SOFTPLUS: tl.constexpr, BLOCK_M: tl.constexpr,
BLOCK_O: tl.constexpr, BLOCK_4: tl.constexpr, BLOCK_D: tl.constexpr,
):
pid_m = tl.program_id(0); pid_b = tl.program_id(1); pid_h = tl.program_id(2)
offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M); mask_m = offs_m < dim
clo_ptr += pid_b * stride_clo_b + pid_h * stride_clo_h
chi_ptr += pid_b * stride_chi_b + pid_h * stride_chi_h
c4_ptr += pid_b * stride_c4_b + pid_h * stride_c4_h
s8_ptr += pid_b * stride_sc_b + pid_h * stride_sc_h
s4_ptr += pid_b * stride_sc_b + pid_h * stride_sc_h
x_ptr += pid_b * stride_x_b + pid_h * stride_x_h
dt_ptr += pid_b * stride_dt_b + pid_h * stride_dt_h
dtb_ptr += pid_h * stride_dtb_h
A_ptr += pid_h * stride_A_h
B_ptr += pid_b * stride_B_b + (pid_h // nhg) * stride_B_g
C_ptr += pid_b * stride_C_b + (pid_h // nhg) * stride_C_g
D_ptr += pid_h * stride_D_h
out_ptr += pid_b * stride_out_b + pid_h * stride_out_h
s8 = tl.load(s8_ptr + offs_m * stride_sc_d, mask=mask_m, other=0.0).to(tl.float32)
s4 = tl.load(s4_ptr + offs_m * stride_sc_d, mask=mask_m, other=0.0).to(tl.float32)
x = tl.load(x_ptr + offs_m * stride_x_d, mask=mask_m, other=0.0).to(tl.float32)
dt = tl.load(dt_ptr + offs_m * stride_dt_d, mask=mask_m, other=0.0).to(tl.float32)
dt += tl.load(dtb_ptr + offs_m * stride_dtb_d, mask=mask_m, other=0.0).to(tl.float32)
if DT_SOFTPLUS:
dt = tl.where(dt <= 20.0, tl.log(1.0 + tl.exp(dt)), dt)
D = tl.load(D_ptr + offs_m * stride_D_d, mask=mask_m, other=0.0).to(tl.float32)
# ---- OUTLIERS [0:k8): INT8 from two nibbles (lo own slot + hi paired slot), coalesced ----
o = tl.arange(0, BLOCK_O); mo = o < k8; mmo = mask_m[:, None] & mo[None, :]
bo = o // 2; sho = (o % 2) * 4
lo = (tl.load(clo_ptr + (offs_m[:, None] * stride_clo_d + bo[None, :] * stride_clo_n), mask=mmo, other=0) >> sho[None, :]) & 0xF
hi = (tl.load(chi_ptr + (offs_m[:, None] * stride_chi_d + bo[None, :] * stride_chi_n), mask=mmo, other=0) >> sho[None, :]) & 0xF
ub = (hi.to(tl.int32) << 4) | lo.to(tl.int32) # 0..255
v8 = ub - 256 * (ub > 127).to(tl.int32) # signed int8
Ao = tl.load(A_ptr + (offs_m[:, None] * stride_A_d + o[None, :] * stride_A_n), mask=mmo, other=0.0).to(tl.float32)
Bo = tl.load(B_ptr + o * stride_B_n, mask=mo, other=0.0).to(tl.float32)
Co = tl.load(C_ptr + o * stride_C_n, mask=mo, other=0.0).to(tl.float32)
sto = v8.to(tl.float32) * s8[:, None] * tl.exp(Ao * dt[:, None]) + (Bo[None, :] * dt[:, None]) * x[:, None]
out = tl.sum(tl.where(mmo, sto * Co[None, :], 0.0), axis=1)
# ---- INT4 [k8:k8+k4): one nibble, coalesced ----
f = tl.arange(0, BLOCK_4); mf = f < k4; mmf = mask_m[:, None] & mf[None, :]
p4 = k8 + f; b4 = f // 2; sh4 = (f % 2) * 4
nib = (tl.load(c4_ptr + (offs_m[:, None] * stride_c4_d + b4[None, :] * stride_c4_n), mask=mmf, other=0) >> sh4[None, :]) & 0xF
v4 = nib.to(tl.int32) - 16 * (nib > 7).to(tl.int32)
A4 = tl.load(A_ptr + (offs_m[:, None] * stride_A_d + p4[None, :] * stride_A_n), mask=mmf, other=0.0).to(tl.float32)
B4 = tl.load(B_ptr + p4 * stride_B_n, mask=mf, other=0.0).to(tl.float32)
C4 = tl.load(C_ptr + p4 * stride_C_n, mask=mf, other=0.0).to(tl.float32)
st4 = v4.to(tl.float32) * s4[:, None] * tl.exp(A4 * dt[:, None]) + (B4[None, :] * dt[:, None]) * x[:, None]
out += tl.sum(tl.where(mmf, st4 * C4[None, :], 0.0), axis=1)
# ---- DEAD [k8+k4:dstate): state==0 (not stored); scanned -> st = dB*x; contributes dB*x*C ----
d = tl.arange(0, BLOCK_D); md = d < kd; mmd = mask_m[:, None] & md[None, :]
pd = k8 + k4 + d
Bd = tl.load(B_ptr + pd * stride_B_n, mask=md, other=0.0).to(tl.float32)
Cd = tl.load(C_ptr + pd * stride_C_n, mask=md, other=0.0).to(tl.float32)
std = (Bd[None, :] * dt[:, None]) * x[:, None]
out += tl.sum(tl.where(mmd, std * Cd[None, :], 0.0), axis=1) + x * D
tl.store(out_ptr + offs_m * stride_out_d, out, mask=mask_m)
# ---- REQUANT + STORE (outliers -> 2 nibble bufs; int4 -> nibble buf; dead -> skip) ----
ns8 = tl.maximum(tl.max(tl.where(mmo, tl.abs(sto), 0.0), axis=1) / 127.0, 1e-8)
q8 = tl.minimum(tl.maximum(tl.extra.cuda.libdevice.round(sto / ns8[:, None]), -127.0), 127.0).to(tl.int32)
qlo = tl.where(mo[None, :], q8 & 0xF, 0).to(tl.uint8)
qhi = tl.where(mo[None, :], (q8 >> 4) & 0xF, 0).to(tl.uint8)
lo0, lo1 = tl.split(tl.reshape(qlo, (BLOCK_M, BLOCK_O // 2, 2))); plo = lo0 | (lo1 << 4)
hi0, hi1 = tl.split(tl.reshape(qhi, (BLOCK_M, BLOCK_O // 2, 2))); phi = hi0 | (hi1 << 4)
op = tl.arange(0, BLOCK_O // 2); mop = op < (k8 // 2)
tl.store(clo_ptr + (offs_m[:, None] * stride_clo_d + op[None, :] * stride_clo_n), plo, mask=mask_m[:, None] & mop[None, :])
tl.store(chi_ptr + (offs_m[:, None] * stride_chi_d + op[None, :] * stride_chi_n), phi, mask=mask_m[:, None] & mop[None, :])
ns4 = tl.maximum(tl.max(tl.where(mmf, tl.abs(st4), 0.0), axis=1) / 7.0, 1e-8)
q4 = tl.minimum(tl.maximum(tl.extra.cuda.libdevice.round(st4 / ns4[:, None]), -7.0), 7.0)
qn = tl.where(mf[None, :], q4.to(tl.int32) & 0xF, 0).to(tl.uint8)
f0, f1 = tl.split(tl.reshape(qn, (BLOCK_M, BLOCK_4 // 2, 2))); pf = f0 | (f1 << 4)
fp = tl.arange(0, BLOCK_4 // 2); mfp = fp < (k4 // 2)
tl.store(c4_ptr + (offs_m[:, None] * stride_c4_d + fp[None, :] * stride_c4_n), pf, mask=mask_m[:, None] & mfp[None, :])
tl.store(s8_ptr + offs_m * stride_sc_d, ns8.to(tl.float16), mask=mask_m)
tl.store(s4_ptr + offs_m * stride_sc_d, ns4.to(tl.float16), mask=mask_m)
def pairnib_scan(clo, chi, c4, s8, s4, x, dt, A, B, C, D, dt_bias, dstate=D_STATE, dt_softplus=True):
Bb, H, dim, _ = clo.shape; G = B.shape[1]
out = torch.empty(Bb, H, dim, device=clo.device, dtype=torch.float32)
BLOCK_M = 16
grid = (triton.cdiv(dim, BLOCK_M), Bb, H)
_pairnib_kernel[grid](
clo, chi, c4, s8, s4, x, dt, dt_bias, A, B, C, D, out,
Bb, H, dim, dstate, K8, K4, KD, H // G,
clo.stride(0), clo.stride(1), clo.stride(2), clo.stride(3),
chi.stride(0), chi.stride(1), chi.stride(2), chi.stride(3),
c4.stride(0), c4.stride(1), c4.stride(2), c4.stride(3),
s8.stride(0), s8.stride(1), s8.stride(2),
x.stride(0), x.stride(1), x.stride(2), dt.stride(0), dt.stride(1), dt.stride(2),
dt_bias.stride(0), dt_bias.stride(1),
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), D.stride(1),
out.stride(0), out.stride(1), out.stride(2),
DT_SOFTPLUS=dt_softplus, BLOCK_M=BLOCK_M,
BLOCK_O=triton.next_power_of_2(K8), BLOCK_4=triton.next_power_of_2(K4), BLOCK_D=triton.next_power_of_2(KD))
return out
def pack_pairnib(state_perm):
"""state_perm:[B,H,dim,N] in permuted order [outliers 0:K8 | int4 K8:K8+K4 | dead K8+K4:N] -> nibble bufs."""
pre = state_perm[..., :K8]; mid = state_perm[..., K8:K8 + K4]
s8 = (pre.abs().amax(-1) / 127.0).clamp(min=1e-8)
s4 = (mid.abs().amax(-1) / 7.0).clamp(min=1e-8)
q8 = (pre / s8[..., None]).round().clamp(-127, 127).to(torch.int32)
qlo = (q8 & 0xF).to(torch.uint8); qhi = ((q8 >> 4) & 0xF).to(torch.uint8)
clo = (qlo[..., 0::2] | (qlo[..., 1::2] << 4)).contiguous()
chi = (qhi[..., 0::2] | (qhi[..., 1::2] << 4)).contiguous()
q4 = ((mid / s4[..., None]).round().clamp(-7, 7).to(torch.int32) & 0xF).to(torch.uint8)
c4 = (q4[..., 0::2] | (q4[..., 1::2] << 4)).contiguous()
return clo, chi, c4, s8.half().contiguous(), s4.half().contiguous()
def dequant_pairnib(clo, chi, c4, s8, s4):
B, H, dim, _ = clo.shape
lo = torch.stack([(clo & 0xF), (clo >> 4) & 0xF], -1).reshape(B, H, dim, K8).to(torch.int32)
hi = torch.stack([(chi & 0xF), (chi >> 4) & 0xF], -1).reshape(B, H, dim, K8).to(torch.int32)
ub = (hi << 4) | lo; v8 = ub - 256 * (ub > 127).to(torch.int32)
pre = v8.float() * s8[..., None].float()
n4 = torch.stack([(c4 & 0xF), (c4 >> 4) & 0xF], -1).reshape(B, H, dim, K4).to(torch.int32)
v4 = n4 - 16 * (n4 > 7).to(torch.int32)
mid = v4.float() * s4[..., None].float()
dead = torch.zeros(B, H, dim, KD, device=clo.device)
return torch.cat([pre, mid, dead], -1)
if __name__ == "__main__":
os.environ.setdefault("TRITON_CACHE_DIR", "/tmp/triton_cache")
torch.manual_seed(0); dev = "cuda"
Bb, H, dim, N, G = 2, 128, 80, 128, 8
state = (torch.randn(Bb, H, dim, N, device=dev) * torch.rand(Bb, H, dim, 1, device=dev) * 3).float()
state[..., K8 + K4:] = 0.0 # dead region starts at 0 (the format)
clo, chi, c4, s8, s4 = pack_pairnib(state)
A = -torch.exp(torch.randn(H, dim, N, device=dev)); D = torch.randn(H, dim, device=dev)
x = torch.randn(Bb, H, dim, device=dev); dt = torch.rand(Bb, H, dim, device=dev) * 0.1
dtb = torch.randn(H, dim, device=dev) * 0.1
Bg = torch.randn(Bb, G, N, device=dev); Cg = torch.randn(Bb, G, N, device=dev)
Bm = Bg.repeat_interleave(H // G, 1); Cm = Cg.repeat_interleave(H // G, 1)
clo_k, chi_k, c4_k, s8_k, s4_k = clo.clone(), chi.clone(), c4.clone(), s8.clone(), s4.clone()
out_k = pairnib_scan(clo_k, chi_k, c4_k, s8_k, s4_k, x, dt, A, Bg, Cg, D, dtb)
deq = dequant_pairnib(clo, chi, c4, s8, s4)
dtf = dt.float() + dtb.float(); dtf = torch.where(dtf <= 20, torch.log1p(torch.exp(dtf)), dtf)
st = deq * torch.exp(A.float() * dtf[..., None]) + (Bm.float()[:, :, None, :] * dtf[..., None]) * x.float()[..., None]
out_r = (st * Cm.float()[:, :, None, :]).sum(-1) + x.float() * D.float()
rel = (out_k - out_r).abs().max().item() / (out_r.abs().max().item() + 1e-9)
rha = lambda z: torch.sign(z) * torch.floor(torch.abs(z) + 0.5)
pre, mid = st[..., :K8], st[..., K8:K8 + K4]
ns8 = (pre.abs().amax(-1) / 127.0).clamp(min=1e-8).half().float()
ns4 = (mid.abs().amax(-1) / 7.0).clamp(min=1e-8).half().float()
q8n = rha(pre / ns8[..., None]).clamp(-127, 127); q4n = rha(mid / ns4[..., None]).clamp(-7, 7)
lo = torch.stack([(clo_k & 0xF), (clo_k >> 4) & 0xF], -1).reshape(Bb, H, dim, K8).to(torch.int32)
hi = torch.stack([(chi_k & 0xF), (chi_k >> 4) & 0xF], -1).reshape(Bb, H, dim, K8).to(torch.int32)
ub = (hi << 4) | lo; q8k = (ub - 256 * (ub > 127).to(torch.int32)).float()
n4 = torch.stack([(c4_k & 0xF), (c4_k >> 4) & 0xF], -1).reshape(Bb, H, dim, K4).to(torch.int32)
q4k = (n4 - 16 * (n4 > 7).to(torch.int32)).float()
m8 = (q8k == q8n).float().mean().item(); md8 = (q8k - q8n).abs().max().item()
m4 = (q4k == q4n).float().mean().item(); md4 = (q4k - q4n).abs().max().item()
bpe = (K8 * 4 + K8 * 4 + K4 * 4) / N + 2 * 16.0 / N # lo+hi outliers + int4 nibbles + scales
ok = rel < 5e-3 and m8 > 0.99 and m4 > 0.99 and max(md8, md4) <= 1.0
print(f"[gate-pairnib] out rel={rel:.2e} int8-match={m8:.4f}(maxD{md8:.0f}) int4-match={m4:.4f}(maxD{md4:.0f}) "
f"({bpe:.2f} b/elem, dead={KD} skipped) -> {'PASS' if ok else 'FAIL'}", flush=True)
print("PAIRNIB_GATE_DONE", flush=True)
|