3 {
4 {
"token_emb",
CK_SCOPE_GLOBAL,
CK_ROLE_WEIGHT, { {
CK_DIM_VOCAB, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
5 {
"pos_emb",
CK_SCOPE_GLOBAL,
CK_ROLE_WEIGHT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
6 {
"embedded_input",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
7 {
"rope_cos_cache",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_HEAD_DIM, 1, 2 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL,
"rope_theta",
CK_DT_FP32},
8 {
"rope_sin_cache",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_HEAD_DIM, 1, 2 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL,
"rope_theta",
CK_DT_FP32},
9 {
"final_ln_weight",
CK_SCOPE_GLOBAL,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
10 {
"final_ln_bias",
CK_SCOPE_GLOBAL,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
11 {
"final_ln_mean",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
12 {
"final_ln_rstd",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
13 {
"final_output",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
14 {
"lm_head_weight",
CK_SCOPE_GLOBAL,
CK_ROLE_WEIGHT, { {
CK_DIM_VOCAB, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0,
"token_emb", NULL,
CK_DT_FP32},
15 {
"logits",
CK_SCOPE_GLOBAL,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_VOCAB, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
16 {
"d_token_emb",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_VOCAB, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
17 {
"d_pos_emb",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
18 {
"d_final_output",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
19 {
"d_final_input",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
20 {
"d_final_ln_weight",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
21 {
"d_logits",
CK_SCOPE_GLOBAL,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_VOCAB, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
22 {
"input",
CK_SCOPE_LAYER,
CK_ROLE_INPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
23 {
"ln1_gamma",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
24 {
"ln1_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
25 {
"ln1_rstd",
CK_SCOPE_LAYER,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 1, NULL, NULL,
CK_DT_FP32},
26 {
"wq",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
27 {
"bq",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
28 {
"wk",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
29 {
"bk",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
30 {
"wv",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
31 {
"bv",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
32 {
"q",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
33 {
"k",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
34 {
"v",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
35 {
"scores",
CK_SCOPE_LAYER,
CK_ROLE_ACTIVATION, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_CTX, 1, 1 }, {
CK_DIM_ALIGNED_CTX, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL,
"training_enabled",
CK_DT_FP32},
36 {
"attn_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
37 {
"wo",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
38 {
"bo",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
39 {
"proj_tmp",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
40 {
"proj_scratch",
CK_SCOPE_LAYER,
CK_ROLE_SCRATCH, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
41 {
"residual1",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
42 {
"ln2_gamma",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
43 {
"ln2_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
44 {
"ln2_rstd",
CK_SCOPE_LAYER,
CK_ROLE_ACTIVATION, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 1, NULL, NULL,
CK_DT_FP32},
45 {
"w1",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
46 {
"b1",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
47 {
"fc1_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
48 {
"swiglu_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
49 {
"w2",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
50 {
"b2",
CK_SCOPE_LAYER,
CK_ROLE_WEIGHT, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
51 {
"mlp_out",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
52 {
"output",
CK_SCOPE_LAYER,
CK_ROLE_OUTPUT, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
53 {
"d_output",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
54 {
"d_residual1",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
55 {
"d_mlp_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
56 {
"d_swiglu_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
57 {
"d_w2",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
58 {
"d_b2",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
59 {
"d_fc1_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
60 {
"d_ln2_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
61 {
"d_w1",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
62 {
"d_b1",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_INTERMEDIATE, 2, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
63 {
"d_ln2_gamma",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
64 {
"d_input",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
65 {
"d_proj_tmp",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
66 {
"d_attn_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
67 {
"d_wo",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
68 {
"d_bo",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
69 {
"d_q",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
70 {
"d_k",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
71 {
"d_v",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
72 {
"d_scores",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_CTX, 1, 1 }, {
CK_DIM_ALIGNED_CTX, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
73 {
"d_ln1_out",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_TOKENS, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
74 {
"d_wq",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
75 {
"d_bq",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
76 {
"d_wk",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
77 {
"d_bk",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
78 {
"d_wv",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
79 {
"d_bv",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_NUM_KV_HEADS, 1, 1 }, {
CK_DIM_ALIGNED_HEAD, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
80 {
"d_ln1_gamma",
CK_SCOPE_LAYER,
CK_ROLE_GRAD, { {
CK_DIM_ALIGNED_EMBED, 1, 1 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 }, {
CK_DIM_END, 0, 0 } }, 0, NULL, NULL,
CK_DT_FP32},
81};