diff options
| author | Georgi Gerganov <ggerganov@gmail.com> | 2024-04-30 12:16:08 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-04-30 12:16:08 +0300 |
| commit | 9c67c2773d4b706cf71d70ecf4aa180b62501960 (patch) | |
| tree | be51cbda5b15ae1bb3a465a2551e7dbe6d3101d7 /ggml-metal.m | |
| parent | 952d03dbead16e4dbdd1d3458486340673cc2465 (diff) | |
ggml : add Flash Attention (#5021)
* ggml : add ggml_flash_attn_ext API
* ggml : fix GQA support in ggml_flash_attn_ext
* ggml : online attention (CPU)
* metal : initial implementation
* metal : f16 precision
* metal : reduce branches
* metal : specialize for head size
* wip : 8 rows per simd group
* wip : 4 rows per simd group
* wip : template for rows per warp
* metal : parallelize across KV size
* metal : parallel reduce across heads
* metal : efficient flash_attn_f16 implementation
* metal : avoid redundant loads of the attention
* metal : scale and mask in matrix form
* metal : fix comment
* llama : avoid ggml_cast, use F32 query
* metal : add parallel reduce version (disabled)
* metal : move output into local memory + optimize
- the result from each simdgroup now stays in the registers
- significantly reduced SRAM usage
- more efficient skipping of -INF blocks
- avoid simdgroup barrier in hot loop
- add comments
* metal : add tests, fix scaling, support C > 32
* metal : improve precision
* ggml : fix f16 mad
* metal : minor
* metal : support Q > 8
* tests : add ATTN tests
* metal : disable buffer allocation logs
* tests : more
* metal : faster inner loop for C == 32
* metal : fix array initialization
* tests : ifdef
* ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext
* ggml : fix ggml_soft_max mask requirement
* cuda : fix soft_max to use correct mask size
* cuda : add flash_attn kernel (wip)
* metal : optimize softmax for C > 32
* metal : optimize softmax
* tests : minor fix
* cuda : avoid zeroing fragments
* tests : update dims
* cuda : fix __hisinf() result check
* cuda : avoid warp_reduce for smax
* cuda : use int instead of int64_t
Noticeably improves performance (thanks to Johannes)
* cuda : make loops use the same loop values
Thanks Johannes again for the tip
* cuda : unroll some of the loops
* cuda : avoid __hisinf branches
* cuda : use half2 in softmax
* cuda : switch to 1 warp for bs > 16
* cuda : speed-up reduce part of the kernel
* cuda : unroll Q*K^T loop
* cuda : fix -INF block check
* cuda : simplify softmax
* cuda : fix matrix names
* cuda : minor
* llama : adapt to F16 KQ_pos
* llama : adapt new models to F16 KQ_mask
* ggml : fix F16 store (ARM NEON)
* llama : fix type of KQ_mask and KQ_pos
* ggml : fix CPU soft_max
* tests : add hs=256
* cuda : fix build
* metal : improve perf via smaller int registers
* cuda : adapt soft_max to F16 mask and pos
* CUDA: faster FlashAttention, kernel for bs == 1
* 16 cols for Phi-2
* no vec for hs, no hs==256 ncols==32 for Volta
* adjust kernel selection logic
* 4 warps, 256 stride for all D
* no ncols == 64
* Multiple parallel blocks for batch size 1
* fix compile warnings
* fix excessive KQ_b loads
* fix cmake build
* fix KV cache padding, NaN from INFINITY (#6438)
* llama : flash_attn cparam + fix defrag
* server: support flash_attn param
* server: bench: enable flash_attn param
* CUDA: refactor host code, dyn. par. blocks
* fix flash_attn_vec_f16 race condition
* flush softmax exp below threshold to 0
* store temp KQ in registers
* Calculate KQ as FP32 if KQV has GGML_PREC_F32
* Add __hgt2_mask implementation for CUDA 11
* fix KQ FP32 precision fpr parallel_blocks > 1
* llama-bench : add -fa,--flash-attn arg
* metal : add BS=1 kernel for flash attention (#6508)
* metal : add BS=1 kernel for flash attention (wip)
* metal : support more than 1 warps
* metal : opts
* metal : opt
* metal : switch to parallel reduce
* metal : reduce registers
* metal : simplify
* metal : initial FA vec kernel
* metal : use F32 attention accumulators
* batched-bench : add fattn arg
* llama : simplify llama_build_kv_store
ggml-ci
* llama : adapt build_olmo to changes
* ggml : fix arm fp16 store on windows
* metal : clean-up
* metal : clean-up kernel code
* metal : minor
* tests : remove benchmarks
ggml-ci
* ggml : fix avx512 const correctness
ggml-ci
* ggml : fix soft_max with bias on CPU
ggml-ci
* common : print --flash-attn in help
* ggml : fix num dimensions in ggml_flash_attn_ext
* llama : force disable flash attention for incompatible models
* ggml : ggml_soft_max support F16/F32 mask/pos
ggml-ci
* cuda : uint -> uint32_t
* cuda : "constexpr dim3" -> "const dim3"
ggml-ci
* cuda : try to fix __hgt2_mask
ggml-ci
* ggml : add TODO's for F16/F32 mask/pos support in other backends
* llama : replace bool need_kq_pos with use_alibi
* llama : prep ALiBi support for BERT models
ggml-ci
* llama : fix n_batch requirements
ggml-ci
* cont
* server : add help for --flash-attn arg
* llama : disable FA for AMD
* tests : remove TMP_ATTN_BENCH
ggml-ci
* llama : support save/load state with FA enabled
ggml-ci
* ci : add CUDA save-load-state tests
ggml-ci
* llama : llama_kv_cache_clear zeroes data + fix save-load seq
ggml-ci
* llama : fix copy-paste errors, add TODO
* llama : disallow incompatible states
* llama : update llama_state_get_size after v_trans field
* metal : remove tmp log
* llama : add static reminder for llama_state_get_size
* metal : fix max nsg
ggml-ci
* ci : fix arg order
ggml-ci
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
Diffstat (limited to 'ggml-metal.m')
| -rw-r--r-- | ggml-metal.m | 547 |
1 files changed, 371 insertions, 176 deletions
diff --git a/ggml-metal.m b/ggml-metal.m index 9cb42198..c6d580b8 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -46,8 +46,10 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, GGML_METAL_KERNEL_TYPE_SILU, GGML_METAL_KERNEL_TYPE_SILU_4, - GGML_METAL_KERNEL_TYPE_SOFT_MAX, - GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, @@ -177,6 +179,14 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, GGML_METAL_KERNEL_TYPE_CPY_F32_F16, GGML_METAL_KERNEL_TYPE_CPY_F32_F32, GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, @@ -443,7 +453,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { } /* - GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ + GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \ (int) kernel->pipeline.threadExecutionWidth); \ */ @@ -459,172 +469,182 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { return NULL; \ } \ } else { \ - GGML_METAL_LOG_WARN("%s: skipping %-32s (not supported)\n", __func__, "kernel_"#name); \ + GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \ } // simd_sum and simd_max requires MTLGPUFamilyApple7 - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX, soft_max, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, soft_max_4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); - 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_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); - 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); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction); - 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_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); - //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); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction); - 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_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); - 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); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm); - 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_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); - 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); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm); - 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_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); - 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); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); + 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_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); + 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); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction); + 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_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); + //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); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction); + 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_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); + 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); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm); + 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_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); + 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); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm); + 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_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); + 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); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); } [metal_library release]; @@ -743,6 +763,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: case GGML_OP_LEAKY_RELU: + case GGML_OP_FLASH_ATTN_EXT: return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: @@ -1326,20 +1347,33 @@ static enum ggml_status ggml_metal_graph_compute( } break; case GGML_OP_SOFT_MAX: { + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F16 || src2->type == GGML_TYPE_F32); + int nth = 32; // SIMD width id<MTLComputePipelineState> pipeline = nil; + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16); + if (ne00%4 == 0) { while (nth < ne00/4 && nth < 256) { nth *= 2; } - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_4].pipeline; + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline; + } } else { while (nth < ne00 && nth < 1024) { nth *= 2; } - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline; + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline; + } } float scale; @@ -2503,6 +2537,161 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; + case GGML_OP_FLASH_ATTN_EXT: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + struct ggml_tensor * src3 = gf->nodes[i]->src[3]; + + GGML_ASSERT(ggml_are_same_shape(src1, src2)); + GGML_ASSERT(src3); + + size_t offs_src3 = 0; + + id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; + + GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16); + GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) && + "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big"); + + const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30); + const int64_t ne31 = src3 ? src3->ne[1] : 0; + const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32); + const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33); + + const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30); + const uint64_t nb31 = src3 ? src3->nb[1] : 0; + const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32); + const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33); + + const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t); + + float scale; + memcpy(&scale, dst->op_params, sizeof(float)); + + id<MTLComputePipelineState> pipeline = nil; + + bool use_vec_kernel = false; + + if (ne01 >= 4 || (ne00%128 != 0)) { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; + default: + { + GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_METAL_LOG_ERROR("add template specialization for this size\n"); + GGML_ASSERT(false && "add template specialization for this size"); + } + } + } else { + use_vec_kernel = true; + + switch (ne00) { + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; + default: + { + GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_METAL_LOG_ERROR("add template specialization for this size\n"); + GGML_ASSERT(false && "add template specialization for this size"); + } + } + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:8]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:11]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:12]; + [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:14]; + [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:15]; + [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:16]; + [encoder setBytes:&nb10 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:19]; + [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:20]; + [encoder setBytes:&ne31 length:sizeof( int64_t) atIndex:21]; + [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:22]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:23]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:24]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:25]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:26]; + [encoder setBytes:&scale length:sizeof( float) atIndex:27]; + + if (!use_vec_kernel) { + // half8x8 kernel + const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 8 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + int64_t nsgmax = 2; + + while (true) { + const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2); + if (smem > ctx->device.maxThreadgroupMemoryLength) { + break; + } + nsgmax *= 2; + } + nsgmax /= 2; + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4; + + const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } else { + // half1x4 kernel + const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 1 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)); + + int64_t nsg = 1; + while (nsg <= nsgt) { + nsg *= 2; + } + nsg /= 2; + + const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } + } break; case GGML_OP_DUP: case GGML_OP_CPY: case GGML_OP_CONT: @@ -2706,10 +2895,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe UNUSED(buft); } -static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) { +static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) { +#ifndef GGML_METAL_NDEBUG #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)", + GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)", + __func__, + size_aligned / 1024.0 / 1024.0, device.currentAllocatedSize / 1024.0 / 1024.0, device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); @@ -2719,10 +2911,15 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) { GGML_METAL_LOG_INFO("\n"); } } else { - GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0); + GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n", + __func__, + size_aligned / 1024.0 / 1024.0, + device.currentAllocatedSize / 1024.0 / 1024.0); } #endif +#endif UNUSED(device); + UNUSED(size_aligned); } GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { @@ -2756,8 +2953,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff return NULL; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0); - ggml_backend_metal_log_allocated_size(device); + //ggml_backend_metal_log_allocated_size(device, size_aligned); return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size); } @@ -2844,7 +3040,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, return false; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0); + ggml_backend_metal_log_allocated_size(device, size_aligned); ++ctx->n_buffers; } else { @@ -2867,7 +3063,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, return false; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i); + ggml_backend_metal_log_allocated_size(device, size_step_aligned); + if (i + size_step < size) { GGML_METAL_LOG_INFO("\n"); } @@ -2876,8 +3073,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, } } - ggml_backend_metal_log_allocated_size(device); - return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size); } |