Metal support

ggml and llama.cpp support Metal, do Apple Silicon users need to use LLaMA.cpp or can they use gglm.cpp with Falcon?

[pjezek]

I tried the following:

build: LLAMA_METAL=1 make falcon_main falcon_quantize falcon_perplexity

then run the model with: ./falcon_main -t 4 -ngl 100 -b 1 -m ../Models/WizardLM-Uncensored-Falcon-7B-GGML/wizardlm-7b-uncensored.ggccv1.q4_0.bin -enc -p "write a story about llamas"

It outputs:

main: build = 883 (2b487f2)
falcon.cpp: loading model from ../Models/WizardLM-Uncensored-Falcon-7B-GGML/wizardlm-7b-uncensored.ggccv1.q4_0.bin
falcon.cpp: file version 10
+---------------+------------+---------+---------+-------+--------+---------------+---------+--------+-------+--------+
|          Info |     format | n_vocab |   n_bpe | n_ctx | n_embd |   n_head ; kv | n_layer | falcon | ftype |   n_ff |
+---------------+------------+---------+---------+-------+--------+---------------+---------+--------+-------+--------+
|               |    ggcc v1 |   65024 |   64784 |  2048 |   4544 |      71 ;   1 |      32 |  7; 7B |     2 |  18176 |
+---------------+------------+---------+---------+-------+--------+---------------+---------+--------+-------+--------+
falcon_model_load_internal: ggml ctx size =    0.00 MB (mmap size = 3872.00 MB)
falcon.cpp: Special mode: Wizard-type finetuning - changing tensor shape
falcon.cpp: Special mode: Wizard-type finetuning - changing tensor shape
falcon_model_load_internal: mem required  = 4196.81 MB (+   48.00 MB per state)
[==================================================] 100%  Tensors populated
falcon_context_prepare: Context falcon_main RAM buffers - key_val =   16.00 MB, Compute =  160.00 MB, Scratch 0 =  124.00 MB, Scratch 1 =   40.14 MB
ggml_metal_init: allocating
ggml_metal_init: using MPS
ggml_metal_init: loading '/Volumes/SanDisk/ggllm.cpp/ggml-metal.metal'
ggml_metal_init: loaded kernel_add                            0x14160b850
ggml_metal_init: loaded kernel_mul                            0x14160bf70
ggml_metal_init: loaded kernel_mul_row                        0x14160c5a0
ggml_metal_init: loaded kernel_scale                          0x14160cac0
ggml_metal_init: loaded kernel_silu                           0x14160cfe0
ggml_metal_init: loaded kernel_relu                           0x14160d500
ggml_metal_init: loaded kernel_gelu                           0x14160da20
ggml_metal_init: loaded kernel_soft_max                       0x14160e0d0
ggml_metal_init: loaded kernel_diag_mask_inf                  0x14160e730
ggml_metal_init: loaded kernel_get_rows_f16                   0x14160edb0
ggml_metal_init: loaded kernel_get_rows_q4_0                  0x14160f430
ggml_metal_init: loaded kernel_get_rows_q4_1                  0x14160fc20
ggml_metal_init: loaded kernel_get_rows_q2_k                  0x1416102a0
ggml_metal_init: loaded kernel_get_rows_q3_k                  0x141610920
ggml_metal_init: loaded kernel_get_rows_q4_k                  0x141610fa0
ggml_metal_init: loaded kernel_get_rows_q5_k                  0x141611620
ggml_metal_init: loaded kernel_get_rows_q6_k                  0x141611ca0
ggml_metal_init: loaded kernel_rms_norm                       0x141612350
ggml_metal_init: loaded kernel_norm                           0x141612a00
ggml_metal_init: loaded kernel_mul_mat_f16_f32                0x1416133d0
ggml_metal_init: loaded kernel_mul_mat_q4_0_f32               0x141613ab0
ggml_metal_init: loaded kernel_mul_mat_q4_1_f32               0x141614190
ggml_metal_init: loaded kernel_mul_mat_q2_k_f32               0x141614870
ggml_metal_init: loaded kernel_mul_mat_q3_k_f32               0x1416150f0
ggml_metal_init: loaded kernel_mul_mat_q4_k_f32               0x1416157d0
ggml_metal_init: loaded kernel_mul_mat_q5_k_f32               0x141615eb0
ggml_metal_init: loaded kernel_mul_mat_q6_k_f32               0x141616590
ggml_metal_init: loaded kernel_rope                           0x141617080
ggml_metal_init: loaded kernel_alibi_f32                      0x141617940
ggml_metal_init: loaded kernel_cpy_f32_f16                    0x1416181d0
ggml_metal_init: loaded kernel_cpy_f32_f32                    0x141618a60
ggml_metal_init: loaded kernel_cpy_f16_f16                    0x1416192f0
ggml_metal_add_buffer: allocated 'data            ' buffer, size =  3874.44 MB
ggml_metal_add_buffer: allocated 'eval            ' buffer, size =   160.00 MB
ggml_metal_add_buffer: allocated 'kv              ' buffer, size =    48.02 MB
ggml_metal_add_buffer: allocated 'scr0            ' buffer, size =   124.00 MB
ggml_metal_add_buffer: allocated 'scr1            ' buffer, size =    40.14 MB

+------------+-----+------+--------+-------------+-------------+-----+------+---------+------+---------+------+------+------+-----+
| Syst. Info | AVX | AVX2 | AVX512 | AVX512_VBMI | AVX512_VNNI | FMA | NEON | ARM_FMA | F16C | FP16_VA | SIMD | BLAS | SSE3 | VSX |
+------------+-----+------+--------+-------------+-------------+-----+------+---------+------+---------+------+------+------+-----+
|  4/10 thrd | 0   | 0    | 0      | 0           | 0           | 0   | 1    | 1       | 0    | 1       | 0    | 1    | 0    | 0   |
+------------+-----+------+--------+-------------+-------------+-----+------+---------+------+---------+------+------+------+-----+

+------------+-------+-------+-------+-------+-------+-------+-------+-------+------+------+--------+---------+
|   Sampling | rpt_n | rpt_p | prs_p | frq_p | top_k | tfs_z | top_p | typ_p | temp | miro | mir_lr | mir_ent |
+------------+-------+-------+-------+-------+-------+-------+-------+-------+------+------+--------+---------+
|            |    64 | 1.100 | 0.000 | 0.000 |    40 | 1.000 | 0.950 | 1.000 | 0.80 |    0 | 0.1000 | 5.00000 |
+============+=======+=======+=======+=======+=======+=======+-------+-------+------+------+--------+---------+
| Generation |   Ctx | Batch |  Keep | Prom. |          Seed |             Finetune | Stop |
+------------+-------+-------+-------+-------+---------------+----------------------+------+
|            |  2048 |     1 |     0 |    10 |    1692449979 |               WIZARD | #  1 |
+------------+-------+-------+-------+-------+---------------+----------------------+------+


GGML_ASSERT: ggml-metal.m:530: ne02 == ne12
GGML_ASSERT: ggml-metal.m:530: ne02 == ne12
zsh: abort      ./falcon_main -t 4 -ngl 100 -b 1 -m  -enc -p "write a story about llamas"

[0x3333]

Same issue here... I'll try to convert the model to other types.