Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

llm(llama): pass rope factors (#5924)

Michael Yangbb46bbcf5e90
diff --git a/llm/patches/0001-llama-3.1-rope-scaling.diff b/llm/patches/0001-llama-3.1-rope-scaling.diff
new file mode 100644
index 000000000..45dcb4f50
--- /dev/null
+++ b/llm/patches/0001-llama-3.1-rope-scaling.diff
@@ -0,0 +1,71 @@
+From 2f872f294fb6f5c6e8f983b68c40ea656053dd92 Mon Sep 17 00:00:00 2001
+From: Michael Yang <[email protected]>
+Date: Tue, 23 Jul 2024 14:33:29 -0700
+Subject: [PATCH] llama 3.1 rope scaling
+
+---
+ src/llama.cpp | 14 ++++++++++++--
+ 1 file changed, 12 insertions(+), 2 deletions(-)
+
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 8fe51971..a9969df8 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -2472,6 +2472,7 @@ struct llama_layer {
+ // long rope factors
+ struct ggml_tensor * rope_long = nullptr;
+ struct ggml_tensor * rope_short = nullptr;
++ struct ggml_tensor * rope_freqs = nullptr;
+
+ // bitnet scale
+ struct ggml_tensor * wq_scale;
+@@ -6143,6 +6144,8 @@ static bool llm_load_tensors(
+
+ layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
++ layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), { n_embd/n_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
++
+ if (n_expert == 0) {
+ layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+ layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
+@@ -8620,6 +8623,10 @@ struct llm_build_context {
+ // choose long/short freq factors based on the context size
+ const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
+
++ if (model.layers[il].rope_freqs != nullptr) {
++ return model.layers[il].rope_freqs;
++ }
++
+ if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
+ return model.layers[il].rope_long;
+ }
+@@ -8814,6 +8821,9 @@ struct llm_build_context {
+
+ // self-attention
+ {
++ // rope freq factors for llama3; may return nullptr for llama2 and other models
++ struct ggml_tensor * rope_factors = build_rope_factors(il);
++
+ // compute Q and K and RoPE them
+ struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+ cb(Qcur, "Qcur", il);
+@@ -8837,14 +8847,14 @@ struct llm_build_context {
+ }
+
+ Qcur = ggml_rope_ext(
+- ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
++ ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+ cb(Qcur, "Qcur", il);
+
+ Kcur = ggml_rope_ext(
+- ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
++ ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+ n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+--
+2.45.2
+