llama : drop the redundant cnv from llm_attn_mask (#31558)

The dspark block mask is now derived at input-fill time: llm_attn_mask::block_uniform
fills every token of a sequence with the mask row of its first token, so the
graph-level reshape/repeat in dflash is gone and cnv always aliases raw.

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB)
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2026-11-03 08:11:06 +02:00
1 parent 0f663d2471
commit ca8daad199
14 files changed
+46 -69

No files matched your search

-2
View File
@@ -352,8 +352,6 @@ llm_graph_input_attn_no_cache * llama_model_clef::graph::build_attn_inp_causal()
ggml_set_input(inp->self_attn_mask.raw);
cb(inp->self_attn_mask.raw, "self_kq_mask", -1);
inp->self_attn_mask.cnv = inp->self_attn_mask.raw;
return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
}