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

Deepseek 4: -sm tensor (#26490)

Browse Source 
* DSV4: sm tensor

* set coarser granularity for head splits

* fix dspark

* add model saving for dsv4 + allow dflash to return on specific device

* add comment about dsv4 seq_rm

* simplify

* add shared expert delayed allreduce

* remove special test for dsv4
This commit is contained in:
Aman Gupta
2026-08-24 09:20:25 +03:00
committed by GitHub
parent c060ca974c
commit bf0a29cc16
7 changed files with 314 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/dflash.cpp
+4 -3
View File
@@ -117,6 +117,10 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); // decoder final norm
// optional: reduced-vocab drafts ship their own lm head, full-vocab drafts can share the target's via ctx_other
// a draft with its own embeddings + head references no target tensors and can run on devices the target does not use (e.g. -devd with a tensor-split target)
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
if (hparams.dsv4_hc_mult > 0) {
const int64_t q_lora_rank = hparams.n_lora_q;
const int64_t n_ff_exp = hparams.n_ff_exp;
@@ -167,9 +171,6 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
return;
}
// optional: reduced-vocab drafts ship their own, full-vocab drafts share the target's via ctx_other
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];