llama.cpp eval-callback — ggml op level
Every row below is a real ggml tensor operation that fired inside llama.cpp's own inference engine for this exact forward pass — not module-level, the actual matmuls, norms, views and the fused flash-attention kernel, in firing order.
| # | Tensor | Op | Operands | Output shape | |
|---|---|---|---|---|---|
| 1 | embd | GET_ROWS | token_embd.weight{384, 65, 1, 1} , inp_tokens{6, 1, 1, 1} | → | {384, 6, 1, 1} |
| 2 | pos_embd | GET_ROWS | position_embd.weight{384, 256, 1, 1} , leaf_5{6, 1, 1, 1} | → | {384, 6, 1, 1} |
| 3 | inpL | ADD | embd{384, 6, 1, 1} , pos_embd{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 4 | norm-0 | NORM | inpL{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 5 | norm_w-0 | MUL | norm-0{384, 6, 1, 1} , blk.0.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 6 | attn_norm-0 | ADD | norm_w-0{384, 6, 1, 1} , blk.0.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 7 | wqkv-0 | MUL_MAT | blk.0.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-0{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 8 | wqkv_b-0 | ADD | wqkv-0{1152, 6, 1, 1} , blk.0.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 9 | Qcur-0 | VIEW | wqkv_b-0{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 10 | Vcur-0 | VIEW | wqkv_b-0{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 11 | Kcur-0 | VIEW | wqkv_b-0{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 12 | Kcur-0 (view) | VIEW | Kcur-0{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 13 | cache_k_l0 (view) | SET_ROWS | Kcur-0 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 14 | Vcur-0 (view) | VIEW | Vcur-0{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 15 | cache_v_l0 (view) | SET_ROWS | Vcur-0 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 16 | Qcur-0 (view) | VIEW | Qcur-0{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 17 | Qcur-0 (view) (permuted) | PERMUTE | Qcur-0 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 18 | cache_k_l0 (view) | VIEW | cache_k_l0{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 19 | cache_k_l0 (view) (permuted) | PERMUTE | cache_k_l0 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 20 | cache_v_l0 (view) | VIEW | cache_v_l0{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 21 | cache_v_l0 (view) (permuted) | PERMUTE | cache_v_l0 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 22 | node_21 | FLASH_ATTN_EXT | Qcur-0 (view) (permuted){64, 6, 6, 1} , cache_k_l0 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 23 | kqv_out-0 | RESHAPE | node_21{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 24 | node_23 | MUL_MAT | blk.0.attn_output.weight{384, 384, 1, 1} , kqv_out-0{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 25 | node_24 | ADD | node_23{384, 6, 1, 1} , blk.0.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 26 | ffn_inp-0 | ADD | node_24{384, 6, 1, 1} , inpL{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 27 | norm-0 | NORM | ffn_inp-0{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 28 | norm_w-0 | MUL | norm-0{384, 6, 1, 1} , blk.0.ffn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 29 | ffn_norm-0 | ADD | norm_w-0{384, 6, 1, 1} , blk.0.ffn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 30 | ffn_up-0 | MUL_MAT | blk.0.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-0{384, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 31 | ffn_up_b-0 | ADD | ffn_up-0{1536, 6, 1, 1} , blk.0.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 6, 1, 1} |
| 32 | ffn_gelu-0 | GELU | ffn_up_b-0{1536, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 33 | ffn_down-0 | MUL_MAT | blk.0.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-0{1536, 6, 1, 1} | → | {384, 6, 1, 1} |
| 34 | ffn_out-0 | ADD | ffn_down-0{384, 6, 1, 1} , blk.0.ffn_down.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 35 | l_out-0 | ADD | ffn_out-0{384, 6, 1, 1} , ffn_inp-0{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 36 | norm-1 | NORM | l_out-0{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 37 | norm_w-1 | MUL | norm-1{384, 6, 1, 1} , blk.1.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 38 | attn_norm-1 | ADD | norm_w-1{384, 6, 1, 1} , blk.1.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 39 | wqkv-1 | MUL_MAT | blk.1.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-1{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 40 | wqkv_b-1 | ADD | wqkv-1{1152, 6, 1, 1} , blk.1.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 41 | Qcur-1 | VIEW | wqkv_b-1{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 42 | Vcur-1 | VIEW | wqkv_b-1{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 43 | Kcur-1 | VIEW | wqkv_b-1{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 44 | Kcur-1 (view) | VIEW | Kcur-1{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 45 | cache_k_l1 (view) | SET_ROWS | Kcur-1 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 46 | Vcur-1 (view) | VIEW | Vcur-1{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 47 | cache_v_l1 (view) | SET_ROWS | Vcur-1 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 48 | Qcur-1 (view) | VIEW | Qcur-1{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 49 | Qcur-1 (view) (permuted) | PERMUTE | Qcur-1 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 50 | cache_k_l1 (view) | VIEW | cache_k_l1{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 51 | cache_k_l1 (view) (permuted) | PERMUTE | cache_k_l1 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 52 | cache_v_l1 (view) | VIEW | cache_v_l1{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 53 | cache_v_l1 (view) (permuted) | PERMUTE | cache_v_l1 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 54 | node_53 | FLASH_ATTN_EXT | Qcur-1 (view) (permuted){64, 6, 6, 1} , cache_k_l1 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 55 | kqv_out-1 | RESHAPE | node_53{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 56 | node_55 | MUL_MAT | blk.1.attn_output.weight{384, 384, 1, 1} , kqv_out-1{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 57 | node_56 | ADD | node_55{384, 6, 1, 1} , blk.1.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 58 | ffn_inp-1 | ADD | node_56{384, 6, 1, 1} , l_out-0{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 59 | norm-1 | NORM | ffn_inp-1{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 60 | norm_w-1 | MUL | norm-1{384, 6, 1, 1} , blk.1.ffn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 61 | ffn_norm-1 | ADD | norm_w-1{384, 6, 1, 1} , blk.1.ffn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 62 | ffn_up-1 | MUL_MAT | blk.1.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-1{384, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 63 | ffn_up_b-1 | ADD | ffn_up-1{1536, 6, 1, 1} , blk.1.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 6, 1, 1} |
| 64 | ffn_gelu-1 | GELU | ffn_up_b-1{1536, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 65 | ffn_down-1 | MUL_MAT | blk.1.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-1{1536, 6, 1, 1} | → | {384, 6, 1, 1} |
| 66 | ffn_out-1 | ADD | ffn_down-1{384, 6, 1, 1} , blk.1.ffn_down.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 67 | l_out-1 | ADD | ffn_out-1{384, 6, 1, 1} , ffn_inp-1{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 68 | norm-2 | NORM | l_out-1{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 69 | norm_w-2 | MUL | norm-2{384, 6, 1, 1} , blk.2.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 70 | attn_norm-2 | ADD | norm_w-2{384, 6, 1, 1} , blk.2.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 71 | wqkv-2 | MUL_MAT | blk.2.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-2{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 72 | wqkv_b-2 | ADD | wqkv-2{1152, 6, 1, 1} , blk.2.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 73 | Qcur-2 | VIEW | wqkv_b-2{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 74 | Vcur-2 | VIEW | wqkv_b-2{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 75 | Kcur-2 | VIEW | wqkv_b-2{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 76 | Kcur-2 (view) | VIEW | Kcur-2{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 77 | cache_k_l2 (view) | SET_ROWS | Kcur-2 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 78 | Vcur-2 (view) | VIEW | Vcur-2{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 79 | cache_v_l2 (view) | SET_ROWS | Vcur-2 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 80 | Qcur-2 (view) | VIEW | Qcur-2{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 81 | Qcur-2 (view) (permuted) | PERMUTE | Qcur-2 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 82 | cache_k_l2 (view) | VIEW | cache_k_l2{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 83 | cache_k_l2 (view) (permuted) | PERMUTE | cache_k_l2 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 84 | cache_v_l2 (view) | VIEW | cache_v_l2{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 85 | cache_v_l2 (view) (permuted) | PERMUTE | cache_v_l2 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 86 | node_85 | FLASH_ATTN_EXT | Qcur-2 (view) (permuted){64, 6, 6, 1} , cache_k_l2 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 87 | kqv_out-2 | RESHAPE | node_85{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 88 | node_87 | MUL_MAT | blk.2.attn_output.weight{384, 384, 1, 1} , kqv_out-2{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 89 | node_88 | ADD | node_87{384, 6, 1, 1} , blk.2.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 90 | ffn_inp-2 | ADD | node_88{384, 6, 1, 1} , l_out-1{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 91 | norm-2 | NORM | ffn_inp-2{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 92 | norm_w-2 | MUL | norm-2{384, 6, 1, 1} , blk.2.ffn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 93 | ffn_norm-2 | ADD | norm_w-2{384, 6, 1, 1} , blk.2.ffn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 94 | ffn_up-2 | MUL_MAT | blk.2.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-2{384, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 95 | ffn_up_b-2 | ADD | ffn_up-2{1536, 6, 1, 1} , blk.2.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 6, 1, 1} |
| 96 | ffn_gelu-2 | GELU | ffn_up_b-2{1536, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 97 | ffn_down-2 | MUL_MAT | blk.2.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-2{1536, 6, 1, 1} | → | {384, 6, 1, 1} |
| 98 | ffn_out-2 | ADD | ffn_down-2{384, 6, 1, 1} , blk.2.ffn_down.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 99 | l_out-2 | ADD | ffn_out-2{384, 6, 1, 1} , ffn_inp-2{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 100 | norm-3 | NORM | l_out-2{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 101 | norm_w-3 | MUL | norm-3{384, 6, 1, 1} , blk.3.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 102 | attn_norm-3 | ADD | norm_w-3{384, 6, 1, 1} , blk.3.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 103 | wqkv-3 | MUL_MAT | blk.3.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-3{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 104 | wqkv_b-3 | ADD | wqkv-3{1152, 6, 1, 1} , blk.3.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 105 | Qcur-3 | VIEW | wqkv_b-3{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 106 | Vcur-3 | VIEW | wqkv_b-3{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 107 | Kcur-3 | VIEW | wqkv_b-3{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 108 | Kcur-3 (view) | VIEW | Kcur-3{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 109 | cache_k_l3 (view) | SET_ROWS | Kcur-3 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 110 | Vcur-3 (view) | VIEW | Vcur-3{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 111 | cache_v_l3 (view) | SET_ROWS | Vcur-3 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 112 | Qcur-3 (view) | VIEW | Qcur-3{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 113 | Qcur-3 (view) (permuted) | PERMUTE | Qcur-3 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 114 | cache_k_l3 (view) | VIEW | cache_k_l3{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 115 | cache_k_l3 (view) (permuted) | PERMUTE | cache_k_l3 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 116 | cache_v_l3 (view) | VIEW | cache_v_l3{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 117 | cache_v_l3 (view) (permuted) | PERMUTE | cache_v_l3 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 118 | node_117 | FLASH_ATTN_EXT | Qcur-3 (view) (permuted){64, 6, 6, 1} , cache_k_l3 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 119 | kqv_out-3 | RESHAPE | node_117{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 120 | node_119 | MUL_MAT | blk.3.attn_output.weight{384, 384, 1, 1} , kqv_out-3{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 121 | node_120 | ADD | node_119{384, 6, 1, 1} , blk.3.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 122 | ffn_inp-3 | ADD | node_120{384, 6, 1, 1} , l_out-2{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 123 | norm-3 | NORM | ffn_inp-3{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 124 | norm_w-3 | MUL | norm-3{384, 6, 1, 1} , blk.3.ffn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 125 | ffn_norm-3 | ADD | norm_w-3{384, 6, 1, 1} , blk.3.ffn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 126 | ffn_up-3 | MUL_MAT | blk.3.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-3{384, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 127 | ffn_up_b-3 | ADD | ffn_up-3{1536, 6, 1, 1} , blk.3.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 6, 1, 1} |
| 128 | ffn_gelu-3 | GELU | ffn_up_b-3{1536, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 129 | ffn_down-3 | MUL_MAT | blk.3.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-3{1536, 6, 1, 1} | → | {384, 6, 1, 1} |
| 130 | ffn_out-3 | ADD | ffn_down-3{384, 6, 1, 1} , blk.3.ffn_down.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 131 | l_out-3 | ADD | ffn_out-3{384, 6, 1, 1} , ffn_inp-3{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 132 | norm-4 | NORM | l_out-3{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 133 | norm_w-4 | MUL | norm-4{384, 6, 1, 1} , blk.4.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 134 | attn_norm-4 | ADD | norm_w-4{384, 6, 1, 1} , blk.4.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 135 | wqkv-4 | MUL_MAT | blk.4.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-4{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 136 | wqkv_b-4 | ADD | wqkv-4{1152, 6, 1, 1} , blk.4.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 137 | Qcur-4 | VIEW | wqkv_b-4{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 138 | Vcur-4 | VIEW | wqkv_b-4{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 139 | Kcur-4 | VIEW | wqkv_b-4{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 140 | Kcur-4 (view) | VIEW | Kcur-4{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 141 | cache_k_l4 (view) | SET_ROWS | Kcur-4 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 142 | Vcur-4 (view) | VIEW | Vcur-4{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 143 | cache_v_l4 (view) | SET_ROWS | Vcur-4 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 144 | Qcur-4 (view) | VIEW | Qcur-4{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 145 | Qcur-4 (view) (permuted) | PERMUTE | Qcur-4 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 146 | cache_k_l4 (view) | VIEW | cache_k_l4{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 147 | cache_k_l4 (view) (permuted) | PERMUTE | cache_k_l4 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 148 | cache_v_l4 (view) | VIEW | cache_v_l4{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 149 | cache_v_l4 (view) (permuted) | PERMUTE | cache_v_l4 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 150 | node_149 | FLASH_ATTN_EXT | Qcur-4 (view) (permuted){64, 6, 6, 1} , cache_k_l4 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 151 | kqv_out-4 | RESHAPE | node_149{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 152 | node_151 | MUL_MAT | blk.4.attn_output.weight{384, 384, 1, 1} , kqv_out-4{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 153 | node_152 | ADD | node_151{384, 6, 1, 1} , blk.4.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 154 | ffn_inp-4 | ADD | node_152{384, 6, 1, 1} , l_out-3{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 155 | norm-4 | NORM | ffn_inp-4{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 156 | norm_w-4 | MUL | norm-4{384, 6, 1, 1} , blk.4.ffn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 157 | ffn_norm-4 | ADD | norm_w-4{384, 6, 1, 1} , blk.4.ffn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 158 | ffn_up-4 | MUL_MAT | blk.4.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-4{384, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 159 | ffn_up_b-4 | ADD | ffn_up-4{1536, 6, 1, 1} , blk.4.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 6, 1, 1} |
| 160 | ffn_gelu-4 | GELU | ffn_up_b-4{1536, 6, 1, 1} | → | {1536, 6, 1, 1} |
| 161 | ffn_down-4 | MUL_MAT | blk.4.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-4{1536, 6, 1, 1} | → | {384, 6, 1, 1} |
| 162 | ffn_out-4 | ADD | ffn_down-4{384, 6, 1, 1} , blk.4.ffn_down.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 163 | l_out-4 | ADD | ffn_out-4{384, 6, 1, 1} , ffn_inp-4{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 164 | norm-5 | NORM | l_out-4{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 165 | norm_w-5 | MUL | norm-5{384, 6, 1, 1} , blk.5.attn_norm.weight{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 166 | attn_norm-5 | ADD | norm_w-5{384, 6, 1, 1} , blk.5.attn_norm.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 167 | wqkv-5 | MUL_MAT | blk.5.attn_qkv.weight{384, 1152, 1, 1} , attn_norm-5{384, 6, 1, 1} | → | {1152, 6, 1, 1} |
| 168 | wqkv_b-5 | ADD | wqkv-5{1152, 6, 1, 1} , blk.5.attn_qkv.bias{1152, 1, 1, 1} | → | {1152, 6, 1, 1} |
| 169 | Qcur-5 | VIEW | wqkv_b-5{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 170 | Vcur-5 | VIEW | wqkv_b-5{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 171 | Kcur-5 | VIEW | wqkv_b-5{1152, 6, 1, 1} | → | {64, 6, 6, 1} |
| 172 | Kcur-5 (view) | VIEW | Kcur-5{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 173 | cache_k_l5 (view) | SET_ROWS | Kcur-5 (view){384, 6, 1, 1} , leaf_9{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 174 | Vcur-5 (view) | VIEW | Vcur-5{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 175 | cache_v_l5 (view) | SET_ROWS | Vcur-5 (view){384, 6, 1, 1} , leaf_11{6, 1, 1, 1} | → | {384, 256, 1, 1} |
| 176 | Qcur-5 (view) | VIEW | Qcur-5{64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 177 | Qcur-5 (view) (permuted) | PERMUTE | Qcur-5 (view){64, 6, 6, 1} | → | {64, 6, 6, 1} |
| 178 | cache_k_l5 (view) | VIEW | cache_k_l5{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 179 | cache_k_l5 (view) (permuted) | PERMUTE | cache_k_l5 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 180 | cache_v_l5 (view) | VIEW | cache_v_l5{384, 256, 1, 1} | → | {64, 6, 256, 1} |
| 181 | cache_v_l5 (view) (permuted) | PERMUTE | cache_v_l5 (view){64, 6, 256, 1} | → | {64, 256, 6, 1} |
| 182 | node_181 | FLASH_ATTN_EXT | Qcur-5 (view) (permuted){64, 6, 6, 1} , cache_k_l5 (view) (permuted){64, 256, 6, 1} | → | {64, 6, 6, 1} |
| 183 | kqv_out-5 | RESHAPE | node_181{64, 6, 6, 1} | → | {384, 6, 1, 1} |
| 184 | node_183 | MUL_MAT | blk.5.attn_output.weight{384, 384, 1, 1} , kqv_out-5{384, 6, 1, 1} | → | {384, 6, 1, 1} |
| 185 | node_184 | ADD | node_183{384, 6, 1, 1} , blk.5.attn_output.bias{384, 1, 1, 1} | → | {384, 6, 1, 1} |
| 186 | node_185 | GET_ROWS | node_184{384, 6, 1, 1} , leaf_89{1, 1, 1, 1} | → | {384, 1, 1, 1} |
| 187 | node_186 | GET_ROWS | l_out-4{384, 6, 1, 1} , leaf_89{1, 1, 1, 1} | → | {384, 1, 1, 1} |
| 188 | ffn_inp-5 | ADD | node_185{384, 1, 1, 1} , node_186{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 189 | norm-5 | NORM | ffn_inp-5{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 190 | norm_w-5 | MUL | norm-5{384, 1, 1, 1} , blk.5.ffn_norm.weight{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 191 | ffn_norm-5 | ADD | norm_w-5{384, 1, 1, 1} , blk.5.ffn_norm.bias{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 192 | ffn_up-5 | MUL_MAT | blk.5.ffn_up.weight{384, 1536, 1, 1} , ffn_norm-5{384, 1, 1, 1} | → | {1536, 1, 1, 1} |
| 193 | ffn_up_b-5 | ADD | ffn_up-5{1536, 1, 1, 1} , blk.5.ffn_up.bias{1536, 1, 1, 1} | → | {1536, 1, 1, 1} |
| 194 | ffn_gelu-5 | GELU | ffn_up_b-5{1536, 1, 1, 1} | → | {1536, 1, 1, 1} |
| 195 | ffn_down-5 | MUL_MAT | blk.5.ffn_down.weight{1536, 384, 1, 1} , ffn_gelu-5{1536, 1, 1, 1} | → | {384, 1, 1, 1} |
| 196 | ffn_out-5 | ADD | ffn_down-5{384, 1, 1, 1} , blk.5.ffn_down.bias{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 197 | l_out-5 | ADD | ffn_out-5{384, 1, 1, 1} , ffn_inp-5{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 198 | norm | NORM | l_out-5{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 199 | norm_w | MUL | norm{384, 1, 1, 1} , output_norm.weight{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 200 | result_norm | ADD | norm_w{384, 1, 1, 1} , output_norm.bias{384, 1, 1, 1} | → | {384, 1, 1, 1} |
| 201 | result_output | MUL_MAT | output.weight{384, 65, 1, 1} , result_norm{384, 1, 1, 1} | → | {65, 1, 1, 1} |
llama-eval-callback's debug output, built and run natively via WSL. GET_ROWS is an embedding lookup, MUL_MAT a matrix multiply, FLASH_ATTN_EXT llama.cpp's fused attention kernel (query, key-cache and value-cache in one op, rather than separate matmul-softmax-matmul steps).