Skip to content

Commit

Permalink
ggml : SOTA 2-bit quants (add IQ2_XS) (#4856)
Browse files Browse the repository at this point in the history 
* iq2_xs: basics

* iq2_xs: this should have been in the basics

* iq2_xs: CUDA and scalar CPU works

* iq2_xs: WIP Metal

* iq2_xs: Metal now works

* iq2_xs: working, but dog slow, ARM_NEON dot product

* iq2_xs: better ARM_NEON dot product

We are now at 19.5 t/s for TG-128 and 61 t/s for PP-512 when
running on the CPU.

* iq2_xs: AVX2 dot product - 19.5 t/s

* iq2_xs: faster AVX2 dit product

21.4 t/s for TG-128, 59.2 t/s for PP-512.
The latter is 2x compared to the previous version.

* iq2_xs: had forgotten to delete iq2-data.h

* Add llama enum for IQ2_XS

---------

Co-authored-by: Iwan Kawrakow <[email protected]>
  • Loading branch information
@ikawrakow @Kawrakow
ikawrakow and Kawrakow committed Jan 11, 2024
1 parent 3ba5b8c commit 49662cb
Show file tree
Hide file tree
Showing 10 changed files with 1,038 additions and 28 deletions.
232 changes: 227 additions & 5 deletions ggml-cuda.cu
View file

Large diffs are not rendered by default.

42 changes: 36 additions & 6 deletions ggml-metal.m
View file
Original file line number Diff line number Diff line change
Expand Up @@ -89,6 +89,7 @@
GGML_METAL_DECL_KERNEL(get_rows_q6_K);
GGML_METAL_DECL_KERNEL(get_rows_i32);
GGML_METAL_DECL_KERNEL(get_rows_iq2_xxs);
GGML_METAL_DECL_KERNEL(get_rows_iq2_xs);
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(group_norm);
GGML_METAL_DECL_KERNEL(norm);
Expand All @@ -108,6 +109,7 @@
GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_iq2_xxs_f32);
GGML_METAL_DECL_KERNEL(mul_mv_iq2_xs_f32);
GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32);
//GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16);
GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32);
Expand All @@ -124,6 +126,7 @@
GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xxs_f32);
GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xs_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
Expand All @@ -137,6 +140,7 @@
GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_iq2_xxs_f32);
GGML_METAL_DECL_KERNEL(mul_mm_iq2_xs_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32);
Expand All @@ -150,6 +154,7 @@
GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xxs_f32);
GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xs_f32);
GGML_METAL_DECL_KERNEL(rope_f32);
GGML_METAL_DECL_KERNEL(rope_f16);
GGML_METAL_DECL_KERNEL(alibi_f32);
Expand Down Expand Up @@ -385,6 +390,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(get_rows_q6_K);
GGML_METAL_ADD_KERNEL(get_rows_i32);
GGML_METAL_ADD_KERNEL(get_rows_iq2_xxs);
GGML_METAL_ADD_KERNEL(get_rows_iq2_xs);
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(group_norm);
GGML_METAL_ADD_KERNEL(norm);
Expand All @@ -404,6 +410,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_iq2_xxs_f32);
GGML_METAL_ADD_KERNEL(mul_mv_iq2_xs_f32);
GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32);
//GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16);
GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32);
Expand All @@ -420,6 +427,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xxs_f32);
GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xs_f32);
if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
Expand All @@ -434,6 +442,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_iq2_xxs_f32);
GGML_METAL_ADD_KERNEL(mul_mm_iq2_xs_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32);
Expand All @@ -447,6 +456,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xxs_f32);
GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xs_f32);
}
GGML_METAL_ADD_KERNEL(rope_f32);
GGML_METAL_ADD_KERNEL(rope_f16);
Expand Down Expand Up @@ -513,6 +523,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(get_rows_q6_K);
GGML_METAL_DEL_KERNEL(get_rows_i32);
GGML_METAL_DEL_KERNEL(get_rows_iq2_xxs);
GGML_METAL_DEL_KERNEL(get_rows_iq2_xs);
GGML_METAL_DEL_KERNEL(rms_norm);
GGML_METAL_DEL_KERNEL(group_norm);
GGML_METAL_DEL_KERNEL(norm);
Expand All @@ -532,6 +543,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_iq2_xxs_f32);
GGML_METAL_DEL_KERNEL(mul_mv_iq2_xs_f32);
GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32);
//GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16);
GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32);
Expand All @@ -548,6 +560,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xxs_f32);
GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xs_f32);
if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
Expand All @@ -562,6 +575,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_iq2_xxs_f32);
GGML_METAL_DEL_KERNEL(mul_mm_iq2_xs_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32);
Expand All @@ -575,6 +589,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xxs_f32);
GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xs_f32);
}
GGML_METAL_DEL_KERNEL(rope_f32);
GGML_METAL_DEL_KERNEL(rope_f16);
Expand Down Expand Up @@ -1561,6 +1576,7 @@ bool ggml_metal_graph_compute(
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xxs_f32]; break;
case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xs_f32]; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
Expand Down Expand Up @@ -1679,6 +1695,12 @@ bool ggml_metal_graph_compute(
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xxs_f32];
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xs_f32];
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
Expand Down Expand Up @@ -1712,12 +1734,12 @@ 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_IQ2_XXS ||
src0t == GGML_TYPE_Q2_K) { // || 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) {
[encoder setThreadgroupMemoryLength:(256*8+128) atIndex:0];
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q4_K) {
Expand Down Expand Up @@ -1810,6 +1832,7 @@ bool ggml_metal_graph_compute(
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break;
case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xxs_f32]; break;
case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xs_f32]; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
Expand Down Expand Up @@ -1931,6 +1954,12 @@ bool ggml_metal_graph_compute(
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xxs_f32];
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xs_f32];
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
Expand Down Expand Up @@ -1980,12 +2009,12 @@ 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_IQ2_XXS ||
src2t == GGML_TYPE_Q2_K) { // || 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) {
[encoder setThreadgroupMemoryLength:(256*8+128) atIndex:0];
else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
const int mem_size = src2t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_Q4_K) {
Expand Down Expand Up @@ -2026,6 +2055,7 @@ bool ggml_metal_graph_compute(
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
case GGML_TYPE_I32: [encoder setComputePipelineState:ctx->pipeline_get_rows_i32]; break;
case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xxs]; break;
case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xs]; break;
default: GGML_ASSERT(false && "not implemented");
}

Expand Down
Loading

0 comments on commit 49662cb

Please sign in to comment.