Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

metal: implement DeepSeek V4 hyper-connections (#26459)

Browse Source 
- Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and
  GGML_OP_DSV4_HC_POST with SIMDgroup register and shuffle optimized kernels.
- Add Metal dispatch and support plumbing and test the production Sinkhorn
  iteration count and embedding width.

Assisted-by: Codex

Co-authored-by: Thiago Padilha <thiago@padilha.cc>
This commit is contained in:
Georgi Gerganov
2026-08-02 21:06:02 +03:00
committed by GitHub
co-authored by Thiago Padilha
parent bb4e0e1b3f
commit fffbcbdb9d
8 changed files with 397 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+2
View File
@@ -8069,6 +8069,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_dsv4_hc_comb(1, 1));
test_cases.emplace_back(new test_dsv4_hc_comb(17, 4));
test_cases.emplace_back(new test_dsv4_hc_comb(257, 8));
test_cases.emplace_back(new test_dsv4_hc_comb(17, 20));
test_cases.emplace_back(new test_dsv4_hc_pre(1, 1));
test_cases.emplace_back(new test_dsv4_hc_pre(31, 17));
@@ -8078,6 +8079,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_dsv4_hc_post(1, 1));
test_cases.emplace_back(new test_dsv4_hc_post(31, 17));
test_cases.emplace_back(new test_dsv4_hc_post(128, 257));
test_cases.emplace_back(new test_dsv4_hc_post(4096, 21));
// glu ops
for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {