nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 17 Packages Projects Releases Wiki Activity

model : Add LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M (#24913)

Browse Source 
* model : Add LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M

* Restore LFM2 models in README.md
This commit is contained in:
Tarek Dakhran
2026-06-24 09:49:46 +03:00
committed by GitHub
parent ac4105d68b
commit 88636e178f
4 changed files with 28 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/lfm2.cpp
+14 -4
View File
@@ -190,7 +190,15 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
auto * conv_rs = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
auto * conv = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
bx = ggml_concat(ctx0, conv, bx, 0);
// causal prepends the state, non-causal pads symmetrically for a centered window
if (hparams.causal_attn) {
bx = ggml_concat(ctx0, conv, bx, 0);
} else {
const int64_t pad = (hparams.n_shortconv_l_cache - 1) / 2;
auto * left = ggml_cont(ctx0,
ggml_view_3d(ctx0, conv, pad, hparams.n_embd, n_seqs, conv->nb[1], conv->nb[2], (d_conv - pad) * conv->nb[0]));
bx = ggml_pad_ext(ctx0, ggml_concat(ctx0, left, bx, 0), 0, pad, 0, 0, 0, 0, 0, 0);
}
GGML_ASSERT(bx->ne[0] > conv->ne[0]);
// last d_conv columns is a new conv state
@@ -266,10 +274,12 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
cb(cur, "result_norm", -1);
res->t_embd = cur;
cur = build_lora_mm(model.output, cur, model.output_s);
cb(cur, "result_output", -1);
if (!cparams.embeddings) {
cur = build_lora_mm(model.output, cur, model.output_s);
cb(cur, "result_output", -1);
res->t_logits = cur;
res->t_logits = cur;
}
ggml_build_forward_expand(gf, cur);
}