model : fix DFlash output head sharing (#30111)

* llama : fix DFlash output head sharing

Assisted-by: Codex

* dflash : read tied output weights from GGUF metadata

Assisted-by: Codex

* llama : share tied word embedding metadata

Assisted-by: Codex

* llama : remove DFlash embedding head fallback

Assisted-by: Codex
This commit is contained in:
Hrishith Thadicherla authored and GitHub committed 2026-10-08 17:42:28 +02:00
1 parent fc9ce6b9d5
commit a11f57ba93
2 files changed
+2 -5

No files matched your search

-3
View File
@@ -167,9 +167,6 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
// 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 (output == nullptr && tok_embd != nullptr) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab_draft }, TENSOR_DUPLICATED);
}
if (hparams.dsv4_hc_mult > 0) {
const int64_t q_lora_rank = hparams.n_lora_q;