Skip to content

Commit

Permalink
1.5 bit quantization (ggerganov#5453)
Browse files Browse the repository at this point in the history 
* iq1_s: WIP basics

* iq1_s: CUDA is working

* iq1_s: scalar CPU dot product

* iq1_s: WIP AVX2 dot product - something is not right

* Fix tests

* Fix shadow warnings

* Fix after merge with latest master

* iq1_s: AVX2 finally works

* iq1_s: ARM_NEON dot product. Works, but not very fast

* iq1_s: better grid

* iq1_s: use IQ2_XXS for attn_output

At a cost of 0.04 extra bpw this gives a big improvement in PPL.

* iq1_s: Metal basics

Dequantize works, but not dot product

* iq1_s: Metal works, but quite slow

As usual, Apple Silicon does not like the code I write.

* iq1_s: Tests

* iq1_s: slightly faster dot product

---------

Co-authored-by: Iwan Kawrakow <[email protected]>
  • Loading branch information
@ikawrakow @Kawrakow @hodlen
2 people authored and hodlen committed Apr 1, 2024
1 parent 7189c5e commit c71b9bf
Show file tree
Hide file tree
Showing 12 changed files with 1,286 additions and 48 deletions.
6 changes: 4 additions & 2 deletions examples/quantize/quantize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
{ "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization", },
{ "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", },
{ "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
{ "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", },
Expand Down Expand Up @@ -287,9 +288,10 @@ int main(int argc, char ** argv) {
}
}

if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
fprintf(stderr, "\n===============================================================================================\n");
fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
fprintf(stderr, "===============================================================================================\n\n\n");
return 1;
}
Expand Down
2 changes: 1 addition & 1 deletion ggml-backend.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -756,7 +756,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_CPY:
return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS; // missing type_traits.from_float
return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
default:
Expand Down
224 changes: 223 additions & 1 deletion ggml-cuda.cu
View file

Large diffs are not rendered by default.

29 changes: 27 additions & 2 deletions ggml-metal.m
View file
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,7 @@
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
GGML_METAL_KERNEL_TYPE_RMS_NORM,
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
Expand All @@ -83,6 +84,7 @@
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
Expand All @@ -101,6 +103,7 @@
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
Expand All @@ -116,6 +119,7 @@
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
Expand All @@ -131,6 +135,7 @@
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F16,
GGML_METAL_KERNEL_TYPE_ALIBI_F32,
Expand Down Expand Up @@ -433,6 +438,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
Expand All @@ -455,6 +461,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
Expand All @@ -473,6 +480,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
Expand All @@ -488,6 +496,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
Expand All @@ -503,6 +512,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true);
Expand Down Expand Up @@ -1318,6 +1328,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}

Expand Down Expand Up @@ -1452,6 +1463,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
Expand Down Expand Up @@ -1486,7 +1503,7 @@ static bool ggml_metal_graph_compute(

if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
Expand Down Expand Up @@ -1592,6 +1609,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}

Expand Down Expand Up @@ -1729,6 +1747,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
Expand Down Expand Up @@ -1779,7 +1803,7 @@ static bool ggml_metal_graph_compute(

if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
Expand Down Expand Up @@ -1833,6 +1857,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
Expand Down
Loading

0 comments on commit c71b9bf

Please sign in to comment.