From 3f6205741de5bf42a38173df5295ee330007bdee Mon Sep 17 00:00:00 2001
From: Xuan Son Nguyen <son@huggingface.co>
Date: Mon, 7 Sep 2026 01:00:51 +0200
Subject: [PATCH] llama: properly handle KV on training

---
 src/llama-context.cpp | 21 +++++++++++++++++----
 src/llama-cparams.h   |  1 +
 src/llama-graph.cpp   | 10 ++++++++++
 3 files changed, 28 insertions(+), 4 deletions(-)

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index c1ef12f56b..bbf15b2110 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -274,6 +274,7 @@ llama_context::llama_context(
 
     // initialized later
     cparams.pipeline_parallel = false;
+    cparams.training = false;
 
     {
         const char * LLAMA_GRAPH_REUSE_DISABLE = getenv("LLAMA_GRAPH_REUSE_DISABLE");
@@ -2349,6 +2350,11 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
     if (n_sampling_outputs_max > 1) {
         res += (n_sampling_outputs_max - 1) * n_sampling_nodes_max;
     }
+
+    if (cparams.training) {
+        res *= 4;
+    }
+
     return res;
 }
 
@@ -3413,12 +3419,19 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     if (cparams.flash_attn) {
         LLAMA_LOG_INFO("%s: disabling flash attention, FLASH_ATTN_EXT has no backward pass\n", __func__);
         cparams.flash_attn = false;
-
-        // the graph changes without flash attention, need to reserve again
-        sched_need_reserve = true;
-        sched_reserve();
     }
 
+    // gradients cannot flow through the KV cache, so the attention reads the K and V of the current ubatch directly
+    if (n_ubatch == cparams.n_ctx) {
+        cparams.training = true;
+    } else {
+        LLAMA_LOG_WARN("%s: n_ubatch (%u) != n_ctx (%u), the K and V projections will not receive gradients\n", __func__, n_ubatch, cparams.n_ctx);
+    }
+
+    // the training graph is different, need to reserve again
+    sched_need_reserve = true;
+    sched_reserve();
+
     ggml_opt_params opt_params = ggml_opt_default_params(sched.get(), GGML_OPT_LOSS_TYPE_CROSS_ENTROPY);
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;
diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index b592de18c7..004fec5a6d 100644
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -53,6 +53,7 @@ struct llama_cparams {
     bool op_offload;
     bool kv_unified;
     bool pipeline_parallel;
+    bool training;           // set by llama_opt_init()
 
     std::vector<bool> embeddings_layer_inp; // [n_layer()] extract input embeddings for layer
 
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 4cbd5fe218..b13852dd89 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2884,6 +2884,11 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
+    if (cparams.training && mctx_cur->get_n_kv() == n_tokens) {
+        k = k_cur;
+        v = v_cur;
+    }
+
     ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, 0, kq_scale, il);
     cb(cur, "kqv_out", il);
 
@@ -3139,6 +3144,11 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
+    if (cparams.training && k_cur && v_cur && mctx_cur->get_n_kv() == n_tokens) {
+        k = k_cur;
+        v = v_cur;
+    }
+
     ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, 0, kq_scale, il);
     cb(cur, "kqv_out", il);