nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Remove padding and multiple D2D copies for MTP (#24086)

Browse Source 
* Make ggml_gated_delta_net take only the initial recurrent state (D, 1, n_seqs) and passes the snapshot count K as an op parameter instead of inferring it from state->ne[1].

Remove the padding hack and copy all emitted snapshots into the recurrent cache with a single strided ggml_cpy

* Make GDN changes in all backends. Address review comments.

* Fix CI build errors
This commit is contained in:
Gaurav Garg
2026-06-10 23:21:16 +05:30
committed by GitHub
parent d2462f8f7a
commit e95dae18d6
20 changed files with 118 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-backend-meta.cpp
+2 -2
View File
@@ -776,8 +776,8 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
GGML_ASSERT(src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_1);
GGML_ASSERT(src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_1);
GGML_ASSERT(src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_1);
// state shape is (S_v*S_v*H, K, n_seqs); the heads dim is nested inside axis 0,
// so a head-aligned split on the input cache reshapes to axis 0 here (not axis 2).
// state shape is [S_v, S_v, H_v, n_seqs] (s0 only); the heads dim is its own axis 2,
// so a head-aligned split on the input cache lands on axis 2 here.
GGML_ASSERT(src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_2 || src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_1 || src_ss[5].axis == GGML_BACKEND_SPLIT_AXIS_0);
return {GGML_BACKEND_SPLIT_AXIS_0, {0}, {1}, 1};
};