IQ1_M: 1.75 bpw quantization (#6302)
* 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
GitHub
parent
e097633f63
commit
55c1b2a3bb
16 changed files with 1006 additions and 125 deletions
37
ggml-metal.m
37
ggml-metal.m
|
|
@ -64,6 +64,7 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
|
||||
|
|
@ -91,6 +92,7 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
|
||||
|
|
@ -114,6 +116,7 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
|
||||
|
|
@ -134,6 +137,7 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
|
||||
|
|
@ -154,6 +158,7 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_ROPE_F32,
|
||||
|
|
@ -490,6 +495,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, 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_IQ1_M, get_rows_iq1_m, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
|
||||
|
|
@ -517,6 +523,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_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_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_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);
|
||||
|
|
@ -540,6 +547,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_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_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_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);
|
||||
|
|
@ -560,6 +568,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_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_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_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);
|
||||
|
|
@ -580,6 +589,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_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_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
|
||||
|
|
@ -1421,6 +1431,7 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
|
||||
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
|
||||
|
|
@ -1575,6 +1586,12 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ1_M:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
{
|
||||
nth0 = 4;
|
||||
|
|
@ -1619,9 +1636,9 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
|
||||
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
|
||||
|
||||
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_IQ1_S || src0t == GGML_TYPE_IQ2_S) {
|
||||
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_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
|
||||
[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) {
|
||||
|
|
@ -1743,6 +1760,7 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
|
||||
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
|
||||
|
|
@ -1900,6 +1918,12 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ1_M:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
{
|
||||
nth0 = 4;
|
||||
|
|
@ -1960,9 +1984,9 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
[encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
|
||||
}
|
||||
|
||||
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_IQ1_S || src2t == GGML_TYPE_IQ2_S) {
|
||||
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_IQ1_S || src2t == GGML_TYPE_IQ1_M || src2t == GGML_TYPE_IQ2_S) {
|
||||
[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) {
|
||||
|
|
@ -2024,6 +2048,7 @@ static enum ggml_status ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
|
||||
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue