Skip to content

qwen35: apply the Hadamard inverse to token embeddings in the MTP draft graph #217

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
sudoingX wants to merge 1 commit into from
Closed

qwen35: apply the Hadamard inverse to token embeddings in the MTP draft graph #217

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions src/models/qwen35.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
#include "models.h"
#include "llama-memory-recurrent.h"
#include "llama-impl.h"

void llama_model_qwen35::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
Expand Down Expand Up @@ -634,6 +635,20 @@ llama_model_qwen35::graph_mtp::graph_mtp(const llama_model & model, const llm_gr
ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;

tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);

// a Hadamard-latent embedding table (prism.hadamard.inverse_weight_names)
// stores rotated rows; restore the primal basis right after the lookup,
// the same way llm_graph_context::build_inp_embd() does, so an MTP head
// that shares the trunk's token_embd reads standard-basis embeddings
Comment on lines +639 to +642
if (hadamard_inverses) {
const auto it = hadamard_inverses->find(tok_embd_w);
if (it != hadamard_inverses->end()) {
tok_embd = llama_mul_mat_hadamard(ctx0, tok_embd, it->second.rot);
if (it->second.signs) {
tok_embd = ggml_mul(ctx0, tok_embd, it->second.signs);
}
}
}
} else {
tok_embd = inp->embd;
}
Expand Down