Commit 4783b1145 for llama.cpp
commit 4783b1145c7073bb809f4c282ff4853cbfe42954
Author: Aman Gupta <amangupta052@gmail.com>
Date: Fri Nov 20 18:49:16 2026 +0800
llama : move the indexer key rotation choice into the model classes (#32653)
* llama : move the indexer key rotation choice into the model classes
The KV cache forced a Hadamard rotation for the lightning indexer keys
based on a list of architectures. The virtual llama_model::indexer_k_rot()
now returns true for the same models. DeepSeek4.1 inherits from DeepSeek4
but was not in the list, so it overrides it back to false.
* llama : rename indexer_k_rot() to lid_k_hdmd()
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 5dd0e5f02..fbe422a27 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -314,11 +314,8 @@ llama_kv_cache::llama_kv_cache(
ggml_is_quantized(type_k) &&
hparams.n_embd_head_k() % 64 == 0;
- // always create Hadamard rotation tensors for DeepSeek lightning indexers
- if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4 ||
- model.arch == LLM_ARCH_GLM_DSA || model.arch == LLM_ARCH_DOTS3NOTE ||
- model.arch == LLM_ARCH_MISTRAL_LARGE4) &&
- hparams.n_embd_head_k_full == hparams.indexer_head_size) {
+ // always create Hadamard rotation tensors for lightning indexer keys
+ if (model.lid_k_hdmd() && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
attn_rot_k = true;
}
diff --git a/src/llama-model.h b/src/llama-model.h
index 1cc5d743a..e86e5e54f 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -1055,6 +1055,9 @@ struct llama_model {
// graph size budget: max graph nodes per model tensor
virtual uint32_t graph_nodes_per_tensor() const { return 8; }
+ // apply Hadamard transform to the lightning indexer keys
+ virtual bool lid_k_hdmd() const { return false; }
+
ggml_cgraph * build_graph(const llm_graph_params & params) const;
virtual void load_stats (llama_model_loader & ml) = 0;
diff --git a/src/models/models.h b/src/models/models.h
index 979ceffc2..822d707b9 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1246,6 +1246,7 @@ struct llama_model_deepseek32 : public llama_model_base {
llama_model_deepseek32(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ bool lid_k_hdmd() const override { return true; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -1262,6 +1263,7 @@ struct llama_model_dots3note : public llama_model_base {
llama_model_dots3note(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ bool lid_k_hdmd() const override { return true; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -1285,6 +1287,7 @@ struct llama_model_deepseek4 : public llama_model_base {
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
uint32_t graph_nodes_per_tensor() const override { return 32; }
+ bool lid_k_hdmd() const override { return true; }
struct graph : public llm_graph_context {
struct no_build {};
@@ -1485,6 +1488,7 @@ struct llama_model_deepseek41 : public llama_model_deepseek4 {
llama_model_deepseek41(const struct llama_model_params & params) : llama_model_deepseek4(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ bool lid_k_hdmd() const override { return false; }
// Engram: the hash constants are hparams.ple_* (one module per Engram layer), hashed on the host (ggml has no int64 mul/xor/mod)
int engram_index(int il) const;
@@ -1543,6 +1547,7 @@ struct llama_model_glm_dsa : public llama_model_base {
llama_model_glm_dsa(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ bool lid_k_hdmd() const override { return true; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);
@@ -1677,6 +1682,7 @@ struct llama_model_mistral_large4 : public llama_model_base {
llama_model_mistral_large4(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
+ bool lid_k_hdmd() const override { return true; }
struct graph : public llm_graph_context {
graph(const llm_graph_params & params);