vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

IQ1_M: 1.75 bpw quantization (#6302)

Browse source 
* iq1_m: basics

* iq1_m: basics-2

* iq1_m: CUDA dequantize works

Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B.

* iq1_m: separate shifts for each group of 8 in a block

We get
PPL(LLaMA-v2-7B ) = 9.2810
PPL(LLaMA-v2-13B) = 6.8105

Not bad, but slightly higher than
  sqrt(PPL(IQ1_S) * PPL(IQ2_XXS))
which is the expected outcome given that IQ1_M is
halfway between IQ1_S and IQ2_XXS in terms of bpw.
From this, we would expect
 PPL = 9.14 for LLaMA-v2-7B
 PPL = 6.63 for LLaMA-v2-13B

* iq1_m: go to 3-bit scales

There is slight increase in PPL, but the 0.0625 bpw reduction
in size is totally worth it.

We now have
PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw
PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw
PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw

* iq1_m: scalar dot product

* iq1_m: AVX2 dot product

* iq1_m: very slightly faster AVX2 dot product

* iq1_m: ARM_NEON dot product

Works, but very slow (10.5 t/s)

* iq1_m: Metal - dequantize works, dot product does not

* iq1_m: Metal now works

About the same performance as iq1_s.

* iq1_m: minor

* iq1_m: checking pure iq1_m quantization

It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight
with Q4_K.

* iiq1_m: slightly faster ARM_NEON dot product

10.5 t/s -> 11.65 t/s

* iq1_m: faster ARM_NEON dot product

11.65 t/s -> 14.9 t/s

* iq1_m: another minor ARM_NEON dot product improvement

14.9 -> 15.0 t/s

* iq1_m: small PPL improvement via super-block scale adjustment

After quantizing block scales redo the super-block scale fit.

PPL(LLaMA-v2-7B ) = 9.3346
PPL(LLaMA-v2-13B) = 6.8419
PPL(LLaMA-v2-70B) = 4.8294
PPL(Mistral-7B  ) = 8.1624

* iq1_m: adapt to CUDA refactoring

* iq1_m: remove unused variable

We have progressed to warnings being errors.

* iq1_m: add to backend-ops tests

* iq1_m: fix Windows ARM

* iq1_m: use common definition of iq1m_scale_t

* cuda: assert -> NO_DEVICE_CODE

* iq1_M: PR comments

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2024-03-26 15:21:27 +01:00 • committed by GitHub
parent e097633f63
commit 55c1b2a3bb
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 1006 additions and 125 deletions
Download patch file Download diff file Expand all files Collapse all files

112
ggml-cuda/vecdotq.cuh
View file

@ -961,8 +961,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
return d * (sumi1 + sumi2);
#endif
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
@ -1001,13 +1000,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
#else
GGML_UNUSED(ksigns64);
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
#else
GGML_UNUSED(ksigns64);
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
@ -1049,13 +1046,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
#else
GGML_UNUSED(ksigns64);
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
#else
GGML_UNUSED(ksigns64);
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
@ -1085,12 +1080,10 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
return d * sumi;
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
@ -1119,12 +1112,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds);
return d * sumi;
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
@ -1159,8 +1150,50 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
const float m = d1q * __high2float(bq8_1[ib32].ds);
return d * sumi + m * delta;
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
const int ib32 = iqs;
int sumi[2] = {0, 0};
float sumf[2] = {0.f, 0.f};
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const int * q8 = (const int *)bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8)));
int grid0 = grid[0] & 0x0f0f0f0f;
int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2]));
const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0));
sumf[l/2] += delta*sumy;
}
#else
const int8_t * q8 = bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
int sumy = 0;
for (int j = 0; j < 4; ++j) {
sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
sumy += q8[j] + q8[j+4];
}
const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
sumf[l/2] += delta*sumy;
q8 += 8;
}
#endif
iq1m_scale_t scale;
const uint16_t * sc = (const uint16_t *)bq1->scales;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
#else
NO_DEVICE_CODE;
#endif
}
@ -1223,27 +1256,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
//// iqs is 0...7
//const int ib64 = iqs/2;
//const int il = iqs%2;
//const int32_t * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il;
//const int32_t * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il;
//const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il;
//const uint32_t * q4_2 = q4_1 + 4;
//const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4);
//const int8_t ls2 = (bq4->scales_l[ib64] >> 4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4);
//const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds);
//const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds);
//int v1, v2;
//int sumi1 = 0, sumi2 = 0;
//for (int j = 0; j < 2; ++j) {
// get_int_from_table_16(q4_1[j], values, v1, v2);
// sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1));
// get_int_from_table_16(q4_2[j], values, v1, v2);
// sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2));
//}
//return d1 * sumi1 + d2 * sumi2;
// iqs is 0...7
const int ib32 = iqs;
const int32_t * q8 = (const int *)bq8_1[ib32].qs;
@ -1259,24 +1271,8 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
}
return d * (sumi1 + sumi2);
//// iqs is 0...15
//const int ib32 = iqs/2;
//const int il = iqs%2;
//const int32_t * q8 = (const int *)bq8_1[ib32].qs + 2*il;
//const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il;
//const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
//const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
//int v1, v2;
//int sumi1 = 0, sumi2 = 0;
//for (int j = 0; j < 2; ++j) {
// get_int_from_table_16(q4[j], values, v1, v2);
// sumi1 = __dp4a(v1, q8[j+0], sumi1);
// sumi2 = __dp4a(v2, q8[j+4], sumi2);
//}
//return d * (sumi1 + sumi2);
#else
assert(false);
return 0.f;
NO_DEVICE_CODE;
#endif
#else
return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);