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

IQ3_S: a much better alternative to Q3_K (#5676)

Browse source 
* iq4_nl: squash commits for easier rebase

* Basics (quantize, dequantize)
* CUDA dequantize and dot product
* Slightly faster CUDA dot product (120 t/s)
* Switch to 6-bit scales
* Scalar dot product
* AVX2 dot product
* ARM_NEON dot product
* Works on metal, but still slow
* Slightly better Metal dot product
* Another small Metal improvement
* Metal dot product is getting there
* Faster CUDA dot product
* Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided
* Report the actual bpw
* Add _xs mix that is 4.05 bpw for non-MoE models
* Remove IQ4_XS for now, slightly adjust kvalues_iq4nl
* AVX2 dot product uses Q8_0 instead of Q8_K
* Add to test-backend-ops
* Minor fix
* Also use use Q5_K for attn_output in MoE models
* Fixes after merging latest master
* Switching to blocks of 32
* AVX2 for blocks of 32
* Scaler dot product for blocks of 32
* ARM_NEON dot product for blocks of 32
* Metal kernels for blocks of 32
* Slightly faster Metal kernels

* Resurrecting iq3_xs

After all the experimentation, nothing was better than this.

* Minor PPL improvement via a block scale fudge factor

* Minor improvement via 3 neighbours

* iq3_xs: working scalar and AVX2 dot products

* iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s)

* iq3_xs: working Metal implementation

* Adding IQ3_M - IQ3_XS mix with mostly Q4_K

* iiq3_xs: a 3.4375 bpw variant

* iq3_xs: make CUDA work for new version

* iq3_xs: make scalar and AVX2 work for new version

* iq3_s: make ARM_NEON work with new version

* iq3_xs: make new version work on metal

Performance is very similar to Q3_K_S

* iq3_xs: tiny Metal speed improvement

* iq3_xs: tiny Metal speed improvement

* Fix stupid warning

* Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS

* iq3_xs: rename to iq3_s

* iq3_s: make tests pass

* Move Q3_K_XS mix to 3.25 bpw

* Attempt to fix failing tests

* Another attempt to fix the Windows builds

* Attempt to fix ROCm

* ROCm again

* iq3_s: partial fix for QK_K = 64

* iq3_s: make it work on metal for QK_K = 64

Pleasent surprise: the coding was super-block size independent,
so all it took was to delete some QK_K == 256 guards.

* Will this fix ROCm?

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2024-02-24 16:23:52 +02:00 • committed by GitHub
parent 525213d2f5
commit 4c4cb30736
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 1211 additions and 84 deletions
Download patch file Download diff file Expand all files Collapse all files

31
ggml.c
View file

@ -678,6 +678,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ3_S] = {
.type_name = "iq3_s",
.blck_size = QK_K,
.type_size = sizeof(block_iq3_s),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq3_s,
.from_float = quantize_row_iq3_s,
.from_float_reference = (ggml_from_float_t)quantize_row_iq3_s_reference,
.vec_dot = ggml_vec_dot_iq3_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_S] = {
.type_name = "iq1_s",
.blck_size = QK_K,
@ -2304,6 +2316,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_IQ3_XXS: wtype = GGML_TYPE_IQ3_XXS; break;
case GGML_FTYPE_MOSTLY_IQ1_S: wtype = GGML_TYPE_IQ1_S; break;
case GGML_FTYPE_MOSTLY_IQ4_NL: wtype = GGML_TYPE_IQ4_NL; break;
case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break;
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
}
@ -7738,6 +7751,7 @@ static void ggml_compute_forward_add(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
{
ggml_compute_forward_add_q_f32(params, dst);
} break;
@ -8017,6 +8031,7 @@ static void ggml_compute_forward_add1(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
{
ggml_compute_forward_add1_q_f32(params, dst);
} break;
@ -8141,6 +8156,7 @@ static void ggml_compute_forward_acc(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
default:
{
GGML_ASSERT(false);
@ -11039,6 +11055,7 @@ static void ggml_compute_forward_out_prod(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
{
ggml_compute_forward_out_prod_q_f32(params, dst);
} break;
@ -11227,6 +11244,7 @@ static void ggml_compute_forward_set(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
default:
{
GGML_ASSERT(false);
@ -11429,6 +11447,7 @@ static void ggml_compute_forward_get_rows(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
{
ggml_compute_forward_get_rows_q(params, dst);
} break;
@ -12129,6 +12148,7 @@ static void ggml_compute_forward_alibi(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_Q8_K:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
@ -12212,6 +12232,7 @@ static void ggml_compute_forward_clamp(
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_Q8_K:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
@ -19463,6 +19484,7 @@ void ggml_quantize_init(enum ggml_type type) {
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ1_S: iq2xs_init_impl(type); break;
case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
case GGML_TYPE_IQ3_S: iq3xs_init_impl(512); break;
default: // nothing
break;
}
@ -19737,6 +19759,15 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ3_S:
{
GGML_ASSERT(start % QK_K == 0);
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq3_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ1_S:
{
GGML_ASSERT(start % QK_K == 0);