diff --git a/llm/patches/07-gemma.diff b/llm/patches/07-gemma.diff new file mode 100644 index 00000000..86eac3d1 --- /dev/null +++ b/llm/patches/07-gemma.diff @@ -0,0 +1,305 @@ +From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001 +From: Ollama maintainers +Date: Wed, 26 Jun 2024 16:18:09 -0700 +Subject: [PATCH] Architecture support + +--- + llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++- + 1 file changed, 193 insertions(+), 1 deletion(-) + +diff --git a/llama.cpp b/llama.cpp +index 61948751..3b4196f5 100644 +--- a/llama.cpp ++++ b/llama.cpp +@@ -217,6 +217,7 @@ enum llm_arch { + LLM_ARCH_INTERNLM2, + LLM_ARCH_MINICPM, + LLM_ARCH_GEMMA, ++ LLM_ARCH_GEMMA2, + LLM_ARCH_STARCODER2, + LLM_ARCH_MAMBA, + LLM_ARCH_XVERSE, +@@ -255,6 +256,7 @@ static const std::map LLM_ARCH_NAMES = { + { LLM_ARCH_INTERNLM2, "internlm2" }, + { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_GEMMA, "gemma" }, ++ { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_XVERSE, "xverse" }, +@@ -464,10 +466,12 @@ enum llm_tensor { + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_NORM_2, + LLM_TENSOR_ATTN_OUT_NORM, ++ LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_ROT_EMBD, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_NORM, ++ LLM_TENSOR_FFN_POST_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, +@@ -960,6 +964,24 @@ static const std::map> LLM_TENSOR_NA + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, ++ { ++ LLM_ARCH_GEMMA2, ++ { ++ { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, ++ { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, ++ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, ++ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, ++ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, ++ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, ++ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, ++ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, ++ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, ++ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, ++ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, ++ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, ++ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, ++ }, ++ }, + { + LLM_ARCH_STARCODER2, + { +@@ -1941,6 +1963,8 @@ enum e_model { + MODEL_8x22B, + MODEL_16x12B, + MODEL_10B_128x3_66B, ++ MODEL_9B, ++ MODEL_27B, + }; + + static const size_t kiB = 1024; +@@ -2114,6 +2138,7 @@ struct llama_layer { + struct ggml_tensor * attn_out_norm_b; + struct ggml_tensor * attn_q_a_norm; + struct ggml_tensor * attn_kv_a_norm; ++ struct ggml_tensor * attn_post_norm; + + // attention + struct ggml_tensor * wq; +@@ -2136,6 +2161,7 @@ struct llama_layer { + // normalization + struct ggml_tensor * ffn_norm; + struct ggml_tensor * ffn_norm_b; ++ struct ggml_tensor * ffn_post_norm; + struct ggml_tensor * layer_out_norm; + struct ggml_tensor * layer_out_norm_b; + struct ggml_tensor * ffn_norm_exps; +@@ -4529,6 +4555,16 @@ static void llm_load_hparams( + } + } break; + case LLM_ARCH_GEMMA: ++ { ++ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ++ ++ switch (hparams.n_layer) { ++ case 18: model.type = e_model::MODEL_9B; break; ++ case 28: model.type = e_model::MODEL_27B; break; ++ default: model.type = e_model::MODEL_UNKNOWN; ++ } ++ } break; ++ case LLM_ARCH_GEMMA2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + +@@ -6305,6 +6341,40 @@ static bool llm_load_tensors( + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + } + } break; ++ case LLM_ARCH_GEMMA2: ++ { ++ model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); ++ ++ // output ++ model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); ++ model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading ++ ++ const int64_t n_ff = hparams.n_ff; ++ const int64_t n_embd_head_k = hparams.n_embd_head_k; ++ const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); ++ const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); ++ ++ for (uint32_t i = 0; i < n_layer; ++i) { ++ ggml_context * ctx_layer = ctx_for_layer(i); ++ ggml_context * ctx_split = ctx_for_layer_split(i); ++ ++ auto & layer = model.layers[i]; ++ ++ layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); ++ ++ layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); ++ layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); ++ layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); ++ layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); ++ layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}); ++ ++ layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); ++ layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); ++ layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "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}); ++ layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}); ++ } ++ } break; + case LLM_ARCH_STARCODER2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); +@@ -10614,6 +10684,123 @@ struct llm_build_context { + return gf; + } + ++ struct ggml_cgraph * build_gemma2() { ++ struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); ++ ++ const int64_t n_embd_head_k = hparams.n_embd_head_k; ++ ++ struct ggml_tensor * cur; ++ struct ggml_tensor * inpL; ++ ++ inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); ++ ++ inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); ++ cb(inpL, "inp_scaled", -1); ++ ++ // inp_pos - contains the positions ++ struct ggml_tensor * inp_pos = build_inp_pos(); ++ ++ // KQ_mask (mask for 1 head, it will be broadcasted to all heads) ++ struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); ++ ++ for (int il = 0; il < n_layer; ++il) { ++ // norm ++ cur = llm_build_norm(ctx0, inpL, hparams, ++ model.layers[il].attn_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "attn_norm", il); ++ ++ // self-attention ++ { ++ // compute Q and K and RoPE them ++ struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); ++ cb(Qcur, "Qcur", il); ++ ++ struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); ++ cb(Kcur, "Kcur", il); ++ ++ struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); ++ cb(Vcur, "Vcur", il); ++ ++ Qcur = ggml_rope_ext( ++ ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, ++ n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale, ++ ext_factor, attn_factor, beta_fast, beta_slow); ++ cb(Qcur, "Qcur", il); ++ ++ Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); ++ cb(Qcur, "Qcur_scaled", il); ++ ++ Kcur = ggml_rope_ext( ++ ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, ++ n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale, ++ ext_factor, attn_factor, beta_fast, beta_slow); ++ cb(Kcur, "Kcur", il); ++ ++ cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, ++ model.layers[il].wo, NULL, ++ Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); ++ } ++ ++ if (il == n_layer - 1) { ++ // skip computing output for unused tokens ++ struct ggml_tensor * inp_out_ids = build_inp_out_ids(); ++ cur = ggml_get_rows(ctx0, cur, inp_out_ids); ++ inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ++ } ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.layers[il].attn_post_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "attn_post_norm", il); ++ ++ struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); ++ cb(sa_out, "sa_out", il); ++ ++ cur = llm_build_norm(ctx0, sa_out, hparams, ++ model.layers[il].ffn_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "ffn_norm", il); ++ ++ // feed-forward network ++ { ++ cur = llm_build_ffn(ctx0, cur, ++ model.layers[il].ffn_up, NULL, ++ model.layers[il].ffn_gate, NULL, ++ model.layers[il].ffn_down, NULL, ++ NULL, ++ LLM_FFN_GELU, LLM_FFN_PAR, cb, il); ++ cb(cur, "ffn_out", il); ++ } ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.layers[il].ffn_post_norm, NULL, ++ LLM_NORM_RMS, cb, -1); ++ cb(cur, "ffn_post_norm", -1); ++ ++ cur = ggml_add(ctx0, cur, sa_out); ++ cb(cur, "l_out", il); ++ ++ // input for next layer ++ inpL = cur; ++ } ++ ++ cur = inpL; ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.output_norm, NULL, ++ LLM_NORM_RMS, cb, -1); ++ cb(cur, "result_norm", -1); ++ ++ // lm_head ++ cur = ggml_mul_mat(ctx0, model.output, cur); ++ cb(cur, "result_output", -1); ++ ++ ggml_build_forward_expand(gf, cur); ++ ++ return gf; ++ } ++ + struct ggml_cgraph * build_starcoder2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + +@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph( + { + result = llm.build_gemma(); + } break; ++ case LLM_ARCH_GEMMA2: ++ { ++ result = llm.build_gemma2(); ++ } break; + case LLM_ARCH_STARCODER2: + { + result = llm.build_starcoder2(); +@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { + case LLM_ARCH_PHI2: + case LLM_ARCH_PHI3: + case LLM_ARCH_GEMMA: ++ case LLM_ARCH_GEMMA2: + case LLM_ARCH_STARCODER2: + case LLM_ARCH_GPTNEOX: + return LLAMA_ROPE_TYPE_NEOX; +@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal( + if (add_ass) { + ss << "assistant\n"; + } +- } else if (tmpl == "gemma" || tmpl.find("") != std::string::npos) { ++ } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("") != std::string::npos) { + // google/gemma-7b-it + std::string system_prompt = ""; + for (auto message : chat) { +-- +2.45.2 +