upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-10-02 19:07:25 -05:00
Code Issues Packages Projects Releases Wiki Activity

consistent naming + less verbose comments

Browse Source
This commit is contained in:
Aman Gupta
2026-10-01 18:45:03 +08:00
committed by GitHub
parent 6880e6d2ec
commit ed8f16db92
2 changed files with 3 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
conversion/qwen4exp.py
+1 -2
View File
@@ -28,8 +28,7 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
# the MTP head: one full-attention QSA block after the trunk, fed by the trunk's hc-wide residual
supports_mtp_export = True
# MTP names the shared Qwen remapper does not know: fc_embedding and fc_hidden act on the
# embedding and on each hc stream, and the block collapses its streams with its own mixer
# MTP tensors the shared Qwen remapper does not know
_MTP_EXTRA = {
"fc_embedding": "nextn_fc_embedding",
"fc_hidden": "nextn_fc_hidden",
src/models/qwen4exp.cpp
+2 -12
View File
@@ -450,9 +450,6 @@ llama_model_qwen4exp::graph::graph(const llama_model & model, const llm_graph_pa
// make sure hc_init is in the same graph split as the first layer (-sm tensor)
ggml_build_forward_expand(gf, res_hc);
// the MTP head reads the residual of every row unless its rows are masked to the outputs
const bool narrow_early = !cparams.embeddings_nextn || cparams.embeddings_nextn_masked;
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = res_hc;
@@ -476,7 +473,7 @@ llama_model_qwen4exp::graph::graph(const llama_model & model, const llm_graph_pa
cur = build_layer_attn(inp->get_attn(), mctx_hyb, inp_kpool, cur, inp_pos, sections, il);
}
if (il == n_layer - 1 && inp_out_ids && narrow_early) {
if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
// everything below is per token, so drop the rows that produce no output
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inject = ggml_get_rows(ctx0, inject, inp_out_ids);
@@ -511,7 +508,7 @@ llama_model_qwen4exp::graph::graph(const llama_model & model, const llm_graph_pa
ggml_build_forward_expand(gf, res->t_h_nextn);
}
if (inp_out_ids && !narrow_early) {
if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
res_hc = ggml_reshape_2d(ctx0, res_hc, n_embd*hc, res_hc->ne[2]);
res_hc = ggml_get_rows(ctx0, res_hc, inp_out_ids);
res_hc = ggml_reshape_3d(ctx0, res_hc, n_embd, hc, res_hc->ne[1]);
@@ -532,13 +529,6 @@ llama_model_qwen4exp::graph::graph(const llama_model & model, const llm_graph_pa
ggml_build_forward_expand(gf, cur);
}
// MTP draft head, as in the vLLM and SGLang qwen4_exp MTP models:
// x_s = eh_proj([enorm(e) ; hnorm(h_s)]) = fc_embedding(enorm(e)) + fc_hidden(hnorm(h_s)) for every hc stream s
// x = one decoder layer (QSA attention, MoE) on the hc-wide residual
// logits = lm_head(own mixer(x)), and x is the h of the next draft step
// h is the trunk's hc-wide residual before its final mixer, e is the embedding of the next token
// hnorm takes the RMS of each stream: vLLM and SGLang take one RMS over all hc*n_embd, but the streams
// differ ~7x in scale and the per-stream norm drafts better (+1-3 points acceptance on Qwen3.8-Flash-Next)
llama_model_qwen4exp::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) :
graph(model, params, no_build{}) {
GGML_ASSERT(hparams.n_layer_nextn == 1 && "qwen4exp MTP has a single block");