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

vulkan: Support GGML_TYPE_NVFP4 (#21455)

Browse Source 
This adds nvfp4 support for get_rows, dequant, and mul_mat(_id). For
mul_mat, it does not add support for the dp4/q8_1 path, it's all via
fp16/fp32.
This commit is contained in:
Jeff Bolz
2026-04-14 11:34:23 +02:00
committed by GitHub
parent e489a5ca0e
commit 6a6780a232
8 changed files with 171 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+2 -1
View File
@@ -66,6 +66,7 @@ const std::vector<std::string> type_names = {
"iq4_xs",
"iq4_nl",
"mxfp4",
"nvfp4",
"bf16",
};
@@ -556,7 +557,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
std::string load_vec_quant = "2";
if ((tname == "q1_0") || (tname == "q4_0") || (tname == "q4_1") || (tname == "q5_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
load_vec_quant = "8";
else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_xs") || (tname == "iq4_nl") || (tname == "mxfp4"))
else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_xs") || (tname == "iq4_nl") || (tname == "mxfp4") || (tname == "nvfp4"))
load_vec_quant = "4";
if (tname == "bf16") {