biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel] Add more dtype support for GGUF dequantization (#15879)

Browse Source 
Signed-off-by: lukas.bluebaum <lukas.bluebaum@aleph-alpha.com>
This commit is contained in:
LukasBluebaum
2025-04-02 10:58:48 +02:00
committed by GitHub
parent 101f1481f9
commit 90969fb39a
9 changed files with 80 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
csrc/quantization/gguf/dequantize.cuh
View File

@@ -94,8 +94,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
dfloat2 v;
dequantize_kernel(vx, ib, iqs, v);
y[iybs + iqs + 0] = v.x;
y[iybs + iqs + y_offset] = v.y;
y[iybs + iqs + 0] = convert_from_half<dst_t>(v.x);
y[iybs + iqs + y_offset] = convert_from_half<dst_t>(v.y);
}
template<typename dst_t>
@@ -114,10 +114,10 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
half dall = __low2half(x[i].dm);
half dmin = __high2half(x[i].dm);
y[l+ 0] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+0] >> 4)));
y[l+32] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+2] >> 4)));
y[l+64] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+4] >> 4)));
y[l+96] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+6] >> 4)));
y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+0] >> 4))));
y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+2] >> 4))));
y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+4] >> 4))));
y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+6] >> 4))));
}
template<typename dst_t>
@@ -148,7 +148,9 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
const uint8_t * q = x[i].qs + 32*n;
const uint8_t * hm = x[i].hmask;
for (int l = l0; l < l0+4; ++l) y[l] = __hmul(dl, __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
for (int l = l0; l < l0+4; ++l) {
y[l] = convert_from_half<dst_t>(__hmul(dl, __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
}
}
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
@@ -188,8 +190,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
const half d2 = __hmul(dall, __int2half_rn(sc));
const half m2 = __hmul(dmin, __int2half_rn(m));
for (int l = 0; l < n; ++l) {
y[l + 0] = __hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1);
y[l +32] = __hsub(__hmul(d2, __int2half_rn(q[l] >> 4)), m2);
y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn(q[l] >> 4)), m2));
}
}
@@ -220,11 +222,11 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
uint8_t hm = 1 << (2*il);
y[ 0] = __hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1);
y[ 1] = __hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1);
y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
hm <<= 1;
y[32] = __hsub(__hmul(d2, __int2half_rn((ql[0] >> 4) + (qh[0] & hm ? 16 : 0))), m2);
y[33] = __hsub(__hmul(d2, __int2half_rn((ql[1] >> 4) + (qh[1] & hm ? 16 : 0))), m2);
y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >> 4) + (qh[0] & hm ? 16 : 0))), m2));
y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >> 4) + (qh[1] & hm ? 16 : 0))), m2));
}
template<typename dst_t>
@@ -247,10 +249,10 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
const uint8_t qh = x[i].qh[32*ip + il];
const int8_t * sc = x[i].scales + is;
y[ 0] = __hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
y[32] = __hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
y[64] = __hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32)));
y[96] = __hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32)));
y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
}
template<typename dst_t>
@@ -269,7 +271,7 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
const uint32_t aux32 = q2[2] | (q2[3] << 16);
const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
template<typename dst_t>
@@ -286,7 +288,7 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
@@ -303,7 +305,7 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
template<typename dst_t>
@@ -324,8 +326,8 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.5f;
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
for (int j = 0; j < 4; ++j) {
y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
}
@@ -345,8 +347,8 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
const uint8_t signs = x[i].signs[4*ib + il];
for (int j = 0; j < 4; ++j) {
y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
}
@@ -367,7 +369,7 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
grid32[0] &= 0x0f0f0f0f;
for (int j = 0; j < 8; ++j) {
y[j] = __float2half(d * (q[j] + delta));
y[j] = d * (q[j] + delta);
}
}
@@ -392,7 +394,7 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
grid32[0] &= 0x0f0f0f0f;
for (int j = 0; j < 8; ++j) {
y[j] = __float2half(d * (q[j] + delta));
y[j] = d * (q[j] + delta);
}
}
@@ -409,8 +411,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
const uint8_t * q4 = x[ib].qs + 4*il;
const float d = __half2float(x[ib].d);
for (int j = 0; j < 4; ++j) {
y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >> 4]);
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
@@ -427,8 +429,8 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
const uint8_t * q4 = x[i].qs + 16*ib + 4*il;
const float d = __half2float(x[i].d) * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
for (int j = 0; j < 4; ++j) {
y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >> 4]);
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
}
}
@@ -522,7 +524,8 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k,
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
}
static to_fp16_cuda_t ggml_get_to_fp16_cuda(int64_t type) {
template<typename dst_t>
static to_cuda_ggml_t<dst_t> ggml_get_to_cuda(int64_t type) {
switch (type) {
case 2:
return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;