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

- 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
+18
View File
@@ -477,6 +477,24 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max(ggml_me
return res;
}
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc(ggml_metal_library_t lib, ggml_op op) {
const char * name = nullptr;
switch (op) {
case GGML_OP_DSV4_HC_COMB: name = "kernel_dsv4_hc_comb_f32"; break;
case GGML_OP_DSV4_HC_PRE: name = "kernel_dsv4_hc_pre_f32"; break;
case GGML_OP_DSV4_HC_POST: name = "kernel_dsv4_hc_post_f32"; break;
default: GGML_ABORT("fatal error");
}
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
if (!res.pipeline) {
res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
}
return res;
}
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_library_t lib, const ggml_tensor * op) {
GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);