From: Georgi Gerganov <redacted>
Date: Tue, 24 Feb 2026 22:01:13 +0000 (+0200)
Subject: models : fix graph splits (#19866)
X-Git-Tag: gguf-v0.18.0~26
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=244641955f6146f7e8474afff7772d427593a534;p=pkg%2Fggml%2Fsources%2Fllama.cpp

models : fix graph splits (#19866)
---

diff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cpp
index 4d6bb83c1..83d11241f 100644
--- a/src/models/kimi-linear.cpp
+++ b/src/models/kimi-linear.cpp
@@ -116,6 +116,8 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
         cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Check layer type by checking which tensors exist
         // KDA layers have ssm_a_log tensor, MLA layers have wkv_a_mqa tensor
         bool is_kda = (layer.ssm_a != nullptr);
diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp
index 56eefd7de..bacf7a4c2 100644
--- a/src/models/qwen35.cpp
+++ b/src/models/qwen35.cpp
@@ -29,6 +29,8 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp
index c7295e336..77f18b5ae 100644
--- a/src/models/qwen35moe.cpp
+++ b/src/models/qwen35moe.cpp
@@ -29,6 +29,8 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -269,7 +271,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp
index 974120ea6..9d3a68cfe 100644
--- a/src/models/qwen3next.cpp
+++ b/src/models/qwen3next.cpp
@@ -21,6 +21,8 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);
 
+        ggml_build_forward_expand(gf, cur);
+
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
@@ -354,7 +356,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(state_update_target, "state_update_target", il);
 
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
-    cb(conv_states_all, "conv_states_updated", il);
 
     ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
     state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);