Skip to content

Commit d3bf1e2

Browse files
committed
mamba : remove some useless comments
No code change.
1 parent eb8cedd commit d3bf1e2

File tree

1 file changed

+2
-6
lines changed

1 file changed

+2
-6
lines changed

llama.cpp

Lines changed: 2 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -4688,9 +4688,6 @@ static bool llm_load_tensors(
46884688

46894689
auto & layer = model.layers[i];
46904690

4691-
// TODO: what's the difference between ctx_layer and ctx_split?
4692-
// A: It seems that ctx_split is for matrices (2d???) while ctx_layer is for other things (like 1D bias and norms, probably.)
4693-
46944691
// norm
46954692
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
46964693

@@ -7901,7 +7898,6 @@ struct llm_build_context {
79017898

79027899
const int32_t n_tok = batch.n_tokens;
79037900

7904-
// hopefully the compiler does constant folding
79057901
const int64_t d_model = n_embd;
79067902
const int64_t d_inner = n_head;
79077903
GGML_ASSERT(2 * d_model == d_inner);
@@ -7948,8 +7944,8 @@ struct llm_build_context {
79487944

79497945
// The following tensor is too big in order to avoid an assertion error when making an overlapping view.
79507946
// TODO: in ggml_new_tensor_impl, handle overlapping data range in data size calculation
7951-
// This could then be a tensor with ne[] = {(d_conv-1)+n_tok, d_inner}
7952-
// which is around (d_conv-1) times as small as its current size.
7947+
// This could then be a tensor with ne[] = {(d_conv-1)+n_tok, d_inner},
7948+
// but the size difference is not that big (d_conv is usually 4).
79537949
struct ggml_tensor * conv_x = ggml_new_tensor_1d(ctx0, conv_state->type, d_conv*d_inner*n_tok);
79547950
const size_t conv_x_nb1 = (d_conv - 1 + n_tok) * ggml_element_size(conv_x);
79557951

0 commit comments

Comments
 (0)