mtfyun/ds4
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Keep the DeepSeek V4.1 wide index scorer exact and off the GLM path

Browse Source
This commit is contained in:
Tarjei Mandt committed 2026-10-05 11:21:26 +11:00
1 parent 3d3c83bbbd
commit 4f806387f7
4 files changed
+48 -6
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

ds4.c
+1 -1
View File
@@ -40927,7 +40927,7 @@ static bool ds41_attention_select_published(ds41_gpu_graph *g, const ds4_model *
!ds4_gpu_dsv41_quantize(g->index_q, 128, DS4_N_INDEXER_HEAD, DS4_V41_FP4_E8M0) ||
!ds41_matmul(g->index_weights, m, l->indexer_proj, g->norm, true) ||
#ifdef __APPLE__
!ds4_gpu_glm_indexer_score_one_tensor(g->index_scores, g->index_q, g->index_weights,
!ds4_gpu_dsv41_indexer_score_one_tensor(g->index_scores, g->index_q, g->index_weights,
g->index_cache[owner], n_comp, DS4_N_INDEXER_HEAD, 128, 1.0f / 64.0f, false)) return false;
#else
!ds4_gpu_dsv41_indexer_scores_batch(g->index_scores, g->index_q, g->index_weights,