llama : remove non-fused GDN and ubatch-dependent graph nodes (#32078)

- remove build_delta_net_chunking and build_delta_net_autoregressive
- always use the fused gated delta net
- make graph_max_nodes independent of the ubatch size
- drop the fused GDN probes and fused node tags

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2026-11-11 15:16:16 +02:00
1 parent b326dd6752
commit ccd360f060
12 files changed
+13 -510

No files matched your search

+1 -8
View File
@@ -457,16 +457,9 @@ ggml_tensor * llama_model_clef::graph::build_layer_attn_linear(
q_conv = build_gdn_l2_norm(ctx0, q_conv, hparams.f_norm_rms_eps);
k_conv = build_gdn_l2_norm(ctx0, k_conv, hparams.f_norm_rms_eps);
// note: need explicit repeat only if we are not using the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_tokens, n_seqs);
}
ggml_tensor * state = ggml_fill(ctx0, ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_v_dim, head_v_dim, num_v_heads, n_seqs), 0.0f);
ggml_tensor * output = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il).first;
ggml_tensor * output = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state).first;
cb(output, "attn_output", il);
// gated normalization