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
ggml/src/ggml-metal/ggml-metal-device.m
+36
View File
@@ -1299,6 +1299,42 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
return false;
}
return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
case GGML_OP_DSV4_HC_COMB:
return has_simdgroup_reduction &&
op->src[0]->type == GGML_TYPE_F32 &&
op->src[1]->type == GGML_TYPE_F32 &&
op->src[2]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32 &&
op->src[0]->ne[0] == 24 &&
op->src[1]->ne[0] >= 3 &&
op->src[2]->ne[0] == 24 &&
ggml_is_contiguous_rows(op->src[0]) &&
ggml_is_contiguous_rows(op->src[1]) &&
ggml_is_contiguous_rows(op->src[2]);
case GGML_OP_DSV4_HC_PRE:
return has_simdgroup_reduction &&
op->src[0]->type == GGML_TYPE_F32 &&
op->src[1]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32 &&
op->src[0]->ne[1] == 4 &&
op->src[1]->ne[0] == 4 &&
ggml_is_contiguous_rows(op->src[0]) &&
ggml_is_contiguous_rows(op->src[1]);
case GGML_OP_DSV4_HC_POST:
return has_simdgroup_reduction &&
op->src[0]->type == GGML_TYPE_F32 &&
op->src[1]->type == GGML_TYPE_F32 &&
op->src[2]->type == GGML_TYPE_F32 &&
op->src[3]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32 &&
op->src[1]->ne[1] == 4 &&
op->src[2]->ne[0] == 4 &&
op->src[3]->ne[0] == 4 &&
op->src[3]->ne[1] == 4 &&
ggml_is_contiguous_rows(op->src[0]) &&
ggml_is_contiguous_rows(op->src[1]) &&
ggml_is_contiguous_rows(op->src[2]) &&
ggml_is_contiguous_rows(op->src[3]);
case GGML_OP_SSM_CONV:
case GGML_OP_SSM_SCAN:
return has_simdgroup_reduction;