happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

models : fix qwen3.5 beta/gate shapes (#19730) 

* models : fix qwen3.5 beta/gate shapes

* cont : avoid extra reshapes
This commit is contained in:
Georgi Gerganov 2026-02-19 15:19:53 +02:00 committed by GitHub
parent e6267a9359
commit da348c9dfb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 11 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/models/kimi-linear.cpp
View File

@ -149,17 +149,19 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
g1 = ggml_mul(ctx0, g1, A);
cb(g1, "kda_g1", il);
g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
// Compute beta (mixing coefficient)
ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, n_head, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs);
cb(beta, "kda_beta", il);
beta = ggml_sigmoid(ctx0, beta);
// Reshape for KDA recurrence
// {n_embd, n_tokens} -> {n_embd, n_seq_tokens, n_seqs}
cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
// Get SSM state and compute KDA recurrence using ggml_kda_scan
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
@ -169,10 +171,6 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm);
Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
beta = ggml_sigmoid(ctx0, beta);
beta = ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs);
g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
// Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?

4
src/models/qwen35.cpp
View File

@ -216,7 +216,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
beta = ggml_sigmoid(ctx0, beta);
@ -232,6 +232,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);

4
src/models/qwen35moe.cpp
View File

@ -216,7 +216,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
beta = ggml_sigmoid(ctx0, beta);
@ -232,6 +232,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);