models : add Flint model (FlintForCausalLM) (#30928)

* Add Flint model (FlintForCausalLM). Assisted by Codex

* cleanup

* one more

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
alexey-rsiandSigbjørn Skjæret authored and GitHub committed 2026-10-21 23:26:22 +02:00
1 parent 99fb997184
commit a37cc46104
18 files changed
+319 -7

No files matched your search

+5 -2
View File
@@ -1684,7 +1684,10 @@ llm_graph_amax_src1 llm_graph_context::build_amax_src1_f16(
constexpr float k_ln_2 = 0.6931471805599453f;
ggml_tensor * e;
e = ggml_scale(ctx0, amax, 1.0f/32768.0f);
// note: this can technically be:
// e = ggml_scale(ctx0, amax, 1.0f/32768.0f);
// but we leave x32 headroom for the F16 sums of 32 values in Q8 activation blocks
e = ggml_scale(ctx0, amax, 1.0f/1024.0f);
e = ggml_log (ctx0, e);
e = ggml_scale(ctx0, e, k_log2_e);
e = ggml_ceil (ctx0, e);
@@ -2505,7 +2508,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
// ref: https://github.com/ggml-org/llama.cpp/pull/30656#issuecomment-6269597922
llm_graph_amax_src1 amax_src1 = { nullptr, nullptr };
if (arch == LLM_ARCH_MISTRAL4 || arch == LLM_ARCH_MELLUM) {
if (arch == LLM_ARCH_MISTRAL4 || arch == LLM_ARCH_MELLUM || arch == LLM_ARCH_FLINT) {
amax_src1 = build_amax_src1_f16(cur);
cur = amax_src1.res;
}