From 5fdfa6282936576d2f352d4b97f397a109f207a6 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Sun, 6 Sep 2026 09:46:21 -0700 Subject: [PATCH] models : fix GDN normalization from `max` to `rsqrt` (#28068) * models: use flash-linear-attention's l2norm for gated delta net q/k The GDN q/k normalization is defined by flash-linear-attention as l2norm(x) = x * rsqrt(sum(x*x) + eps) with eps inside the root. Every GDN call site in the tree uses ggml_l2_norm instead, which is x / max(sqrt(sum(x*x)), eps), i.e. torch.nn.functional.normalize - its CUDA kernel cites that page. The clamp never engages at these magnitudes, so in practice llama.cpp normalizes with no epsilon at all where the reference has one inside the root. transformers made the same substitution when it first added Qwen3-Next and corrected it three days later in huggingface/transformers#40842, 'Fix the misalignment between the l2norm in GDN of Qwen3-Next and the implementation in the FLA library'. vLLM and SGLang vendor FLA rather than reimplementing it, so neither ever had the clamp. eps keeps coming from the checkpoint, exactly as every call site already passed it. The references hardcode 1e-6 for this norm; that is a separate question and the two agree on every GDN checkpoint in the wild. ggml_l2_norm itself is correct and unchanged, as is rwkv7-base, its original caller, which passes normalize's own default eps of 1e-12. No new ggml op: rms_norm already carries eps inside the root, so rms_norm(x, eps/n) * (1/sqrt(n)) is exactly x * rsqrt(sum(x*x) + eps). * Update src/models/models.h Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- src/models/bailingmoe3.cpp | 4 ++-- src/models/kimi-k3.cpp | 6 +++--- src/models/kimi-linear.cpp | 5 +++-- src/models/models.h | 7 +++++++ src/models/qwen35.cpp | 5 +++-- src/models/qwen35moe.cpp | 5 +++-- src/models/qwen3next.cpp | 5 +++-- src/models/qwen4exp.cpp | 5 +++-- 8 files changed, 27 insertions(+), 15 deletions(-) diff --git a/src/models/bailingmoe3.cpp b/src/models/bailingmoe3.cpp index 1f2592cfa..e208c7d5a 100644 --- a/src/models/bailingmoe3.cpp +++ b/src/models/bailingmoe3.cpp @@ -280,8 +280,8 @@ llama_model_bailingmoe3::graph::graph(const llama_model & model, const llm_graph ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur); beta = ggml_sigmoid(ctx0, ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs)); - q = ggml_l2_norm(ctx0, q, hparams.f_norm_rms_eps); - k = ggml_l2_norm(ctx0, k, hparams.f_norm_rms_eps); + q = build_gdn_l2_norm(ctx0, q, hparams.f_norm_rms_eps); + k = build_gdn_l2_norm(ctx0, k, hparams.f_norm_rms_eps); ggml_tensor * states_all = mctx_cur->get_s_l(il); ggml_tensor * state = build_rs(inp_rs, states_all, hparams.n_embd_s(), n_seqs); diff --git a/src/models/kimi-k3.cpp b/src/models/kimi-k3.cpp index b061093eb..b7604cbf2 100644 --- a/src/models/kimi-k3.cpp +++ b/src/models/kimi-k3.cpp @@ -441,9 +441,9 @@ ggml_tensor * llama_model_kimi_k3::graph::build_kda_layer( ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs); state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head_kda, n_seqs); - const float eps = hparams.f_norm_rms_eps; - Qcur = ggml_l2_norm(ctx0, Qcur, eps); - Kcur = ggml_l2_norm(ctx0, Kcur, eps); + const float eps_norm = hparams.f_norm_rms_eps; + Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm); + Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm); auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il); diff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cpp index 601d1d9be..f391f5f50 100644 --- a/src/models/kimi-linear.cpp +++ b/src/models/kimi-linear.cpp @@ -331,10 +331,11 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs); state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head, n_seqs); + const float eps_norm = hparams.f_norm_rms_eps; - Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm); - Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm); + Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm); + Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm); // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il); diff --git a/src/models/models.h b/src/models/models.h index 50e9a235c..87195fddd 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -10,6 +10,13 @@ class llama_memory_hybrid_idx_context; +// ref: https://github.com/ggml-org/llama.cpp/pull/28068 +static inline ggml_tensor * build_gdn_l2_norm(ggml_context * ctx, ggml_tensor * x, float eps) { + const float n = x->ne[0]; + + return ggml_scale(ctx, ggml_rms_norm(ctx, x, eps/n), 1.0f/sqrtf(n)); +} + // // base classes // diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp index 0b9210981..478f9ebea 100644 --- a/src/models/qwen35.cpp +++ b/src/models/qwen35.cpp @@ -423,10 +423,11 @@ ggml_tensor * llama_model_qwen35::graph::build_layer_attn_linear( cb(k_conv, "k_conv", il); cb(v_conv, "v_conv", il); + const float eps_norm = hparams.f_norm_rms_eps; - q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); - k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm); + k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm); //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp index ed4083f12..488c7d357 100644 --- a/src/models/qwen35moe.cpp +++ b/src/models/qwen35moe.cpp @@ -447,10 +447,11 @@ ggml_tensor * llama_model_qwen35moe::graph::build_layer_attn_linear( cb(k_conv, "k_conv", il); cb(v_conv, "v_conv", il); + const float eps_norm = hparams.f_norm_rms_eps; - q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); - k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm); + k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm); //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp index eb823b8ea..222c0acf0 100644 --- a/src/models/qwen3next.cpp +++ b/src/models/qwen3next.cpp @@ -503,10 +503,11 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn_linear( cb(k_conv, "k_conv", il); cb(v_conv, "v_conv", il); + const float eps_norm = hparams.f_norm_rms_eps; - q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); - k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm); + k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm); //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); diff --git a/src/models/qwen4exp.cpp b/src/models/qwen4exp.cpp index 1484c9b07..8ace95f73 100644 --- a/src/models/qwen4exp.cpp +++ b/src/models/qwen4exp.cpp @@ -936,10 +936,11 @@ ggml_tensor * llama_model_qwen4exp::graph::build_layer_attn_linear( cb(k_conv, "k_conv", il); cb(v_conv, "v_conv", il); + const float eps_norm = hparams.f_norm_rms_eps; - q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); - k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm); + k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm); // repeat to match shapes when head keys != value keys; unneeded with the fused GDN if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {