From 4c4cb30736582cacb1a164a9d4bc8e17b1014be7 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sat, 24 Feb 2024 16:23:52 +0200 Subject: IQ3_S: a much better alternative to Q3_K (#5676) * 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 --- ggml-quants.h | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) (limited to 'ggml-quants.h') diff --git a/ggml-quants.h b/ggml-quants.h index 113623b6..303b0b6f 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -191,6 +191,21 @@ typedef struct { } block_iq3_xxs; static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding"); +// 3.4375 bpw +#if QK_K == 64 +#define IQ3S_N_SCALE 2 +#else +#define IQ3S_N_SCALE QK_K/64 +#endif +typedef struct { + ggml_fp16_t d; + uint8_t qs[QK_K/4]; + uint8_t qh[QK_K/32]; + uint8_t signs[QK_K/8]; + uint8_t scales[IQ3S_N_SCALE]; +} block_iq3_s; +static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding"); + typedef struct { ggml_fp16_t d; uint8_t qs[QK_K/8]; @@ -226,6 +241,7 @@ void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGM void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k); void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k); void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int k); +void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int k); void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); @@ -242,6 +258,7 @@ void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); +void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); // Dequantization void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); @@ -262,6 +279,7 @@ void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_ void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); +void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); // Dot product void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); @@ -280,6 +298,7 @@ void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); // // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") @@ -289,6 +308,7 @@ size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); +size_t quantize_iq3_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -- cgit v1.2.3