From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001 From: jmorganca Date: Thu, 17 Oct 2024 15:18:22 -0700 Subject: [PATCH] add mllama support mllama adds cross-attention layers to the standard llama architecture it also requires a way to input a new tensor: cross_attention_state once per generation cross-attention layers don't change and so they are cached in the kv cache once per run remaining is to implement the cross attention mask --- examples/llava/llava.cpp | 5 +- ggml/src/ggml-backend-reg.cpp | 6 +- include/llama.h | 6 + src/llama-arch.cpp | 44 ++++++ src/llama-arch.h | 10 ++ src/llama-batch.cpp | 3 + src/llama-context.cpp | 28 ++-- src/llama-context.h | 2 + src/llama-cparams.h | 1 + src/llama-hparams.cpp | 6 + src/llama-hparams.h | 5 + src/llama-kv-cache.cpp | 13 +- src/llama-model-loader.cpp | 2 + src/llama-model.cpp | 65 ++++++++- src/llama-model.h | 12 ++ src/llama-quant.cpp | 4 +- src/llama.cpp | 262 +++++++++++++++++++++++++++++++++- 17 files changed, 452 insertions(+), 22 deletions(-) diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index 518aad3f..f0e484a1 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -445,7 +445,7 @@ struct llava_embd_batch { std::vector seq_ids; std::vector logits; llama_batch batch; - llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { + llava_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { pos .resize(n_tokens); n_seq_id.resize(n_tokens); seq_ids .resize(n_tokens + 1); @@ -457,6 +457,7 @@ struct llava_embd_batch { /*n_tokens =*/ n_tokens, /*tokens =*/ nullptr, /*embd =*/ embd, + /*n_embd =*/ n_embd, /*pos =*/ pos.data(), /*n_seq_id =*/ n_seq_id.data(), /*seq_id =*/ seq_ids.data(), @@ -480,7 +481,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_ n_eval = n_batch; } float * embd = image_embed->embed+i*n_embd; - llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0); + llava_embd_batch llava_batch = llava_embd_batch(embd, n_embd, n_eval, *n_past, 0); if (llama_decode(ctx_llama, llava_batch.batch)) { LOG_ERR("%s : failed to eval\n", __func__); return false; diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 955ed505..95036ef8 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -171,9 +171,9 @@ struct ggml_backend_registry { #ifdef GGML_USE_CANN register_backend(ggml_backend_cann_reg()); #endif -#ifdef GGML_USE_BLAS - register_backend(ggml_backend_blas_reg()); -#endif +// #ifdef GGML_USE_BLAS +// register_backend(ggml_backend_blas_reg()); +// #endif #ifdef GGML_USE_RPC register_backend(ggml_backend_rpc_reg()); #endif diff --git a/include/llama.h b/include/llama.h index 47919602..cc948005 100644 --- a/include/llama.h +++ b/include/llama.h @@ -249,6 +249,7 @@ extern "C" { llama_token * token; float * embd; + int32_t n_embd; llama_pos * pos; int32_t * n_seq_id; llama_seq_id ** seq_id; @@ -343,6 +344,7 @@ extern "C" { bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU bool flash_attn; // whether to use flash attention [EXPERIMENTAL] bool no_perf; // whether to measure performance timings + bool cross_attn; // whether to use cross attention // Abort callback // if it returns true, execution of llama_decode() will be aborted @@ -443,6 +445,10 @@ extern "C" { struct llama_context_params params), "use llama_init_from_model instead"); + // TODO (jmorganca): this should most likely be passed in as part of a batch + // and not set on the context for all batches. + LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state); + // Frees all allocated memory LLAMA_API void llama_free(struct llama_context * ctx); diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index a1e0ebcc..b6f20286 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -6,6 +6,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_MLLAMA, "mllama" }, { LLM_ARCH_DECI, "deci" }, { LLM_ARCH_FALCON, "falcon" }, { LLM_ARCH_GROK, "grok" }, @@ -127,6 +128,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, "%s.attention.block_skip_connection" }, + { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -225,6 +227,40 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, }, }, + { + LLM_ARCH_MLLAMA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { 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_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { 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_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_CROSS_ATTN_K_NORM, "blk.%d.cross_attn_k_norm" }, + { LLM_TENSOR_CROSS_ATTN_K_PROJ, "blk.%d.cross_attn_k_proj" }, + { LLM_TENSOR_CROSS_ATTN_O_PROJ, "blk.%d.cross_attn_o_proj" }, + { LLM_TENSOR_CROSS_ATTN_Q_NORM, "blk.%d.cross_attn_q_norm" }, + { LLM_TENSOR_CROSS_ATTN_Q_PROJ, "blk.%d.cross_attn_q_proj" }, + { LLM_TENSOR_CROSS_ATTN_V_PROJ, "blk.%d.cross_attn_v_proj" }, + { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" }, + { LLM_TENSOR_CROSS_ATTN_MLP_GATE, "blk.%d.cross_attn_mlp_gate" }, + }, + }, { LLM_ARCH_DECI, { @@ -1450,6 +1486,14 @@ static const std::map LLM_TENSOR_INFOS = { // this tensor is loaded for T5, but never used {LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}}, {LLM_TENSOR_BSKCN_TV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_CROSS_ATTN_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_CROSS_ATTN_K_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_CROSS_ATTN_O_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_CROSS_ATTN_Q_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_CROSS_ATTN_Q_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_CROSS_ATTN_V_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_CROSS_ATTN_ATTN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_CROSS_ATTN_MLP_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_INPUT, GGML_OP_IM2COL}}, {LLM_TENSOR_POS_NET_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_POS_NET_NORM1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 77919578..ec742224 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -10,6 +10,7 @@ enum llm_arch { LLM_ARCH_LLAMA, + LLM_ARCH_MLLAMA, LLM_ARCH_DECI, LLM_ARCH_FALCON, LLM_ARCH_BAICHUAN, @@ -131,6 +132,7 @@ enum llm_kv { LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ATTENTION_SCALE, LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, + LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_DIMENSION_SECTIONS, @@ -314,6 +316,14 @@ enum llm_tensor { LLM_TENSOR_CLS, LLM_TENSOR_CLS_OUT, LLM_TENSOR_BSKCN_TV, + LLM_TENSOR_CROSS_ATTN_K_NORM, + LLM_TENSOR_CROSS_ATTN_K_PROJ, + LLM_TENSOR_CROSS_ATTN_O_PROJ, + LLM_TENSOR_CROSS_ATTN_Q_NORM, + LLM_TENSOR_CROSS_ATTN_Q_PROJ, + LLM_TENSOR_CROSS_ATTN_V_PROJ, + LLM_TENSOR_CROSS_ATTN_ATTN_GATE, + LLM_TENSOR_CROSS_ATTN_MLP_GATE, LLM_TENSOR_CONV1D, LLM_TENSOR_CONVNEXT_DW, LLM_TENSOR_CONVNEXT_NORM, diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 01d5ca57..8682b0e6 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -316,6 +316,7 @@ struct llama_batch llama_batch_get_one( /*n_tokens =*/ n_tokens, /*tokens =*/ tokens, /*embd =*/ nullptr, + /*n_embd =*/ 0, /*pos =*/ nullptr, /*n_seq_id =*/ nullptr, /*seq_id =*/ nullptr, @@ -328,6 +329,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_ /*n_tokens =*/ 0, /*tokens =*/ nullptr, /*embd =*/ nullptr, + /*n_embd =*/ 0, /*pos =*/ nullptr, /*n_seq_id =*/ nullptr, /*seq_id =*/ nullptr, @@ -336,6 +338,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_ if (embd) { batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd); + batch.n_embd = embd; } else { batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc); } diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 47e79ed4..7b22fe13 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -74,10 +74,19 @@ void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) { } if (ubatch.embd) { - const int64_t n_embd = hparams.n_embd; - const int64_t n_tokens = ubatch.n_tokens; + if (lctx.inp_cross_attn_state && lctx.inp_cross_attn_state->buffer) { + ggml_backend_tensor_set(lctx.inp_cross_attn_state, ubatch.embd, 0, ggml_nbytes(lctx.inp_cross_attn_state)); + // zero out inp_embd since it's not used + float * inp_embd_data = (float *)lctx.inp_embd->data; + for (int i = 0; i < ggml_nelements(lctx.inp_embd); ++i) { + inp_embd_data[i] = 0.0f; + } + } else { + const int64_t n_embd = hparams.n_embd; + const int64_t n_tokens = ubatch.n_tokens; - ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); + ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); + } } if (ubatch.pos && lctx.inp_pos) { @@ -470,12 +479,11 @@ void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) { size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) { const auto & cparams = lctx.cparams; const auto & hparams = lctx.model.hparams; - const auto & vocab = lctx.model.vocab; const size_t n_outputs_max = std::max(n_outputs, (size_t) cparams.n_seq_max); const auto n_batch = cparams.n_batch; - const auto n_vocab = vocab.n_tokens(); + const auto n_vocab = hparams.n_vocab; const auto n_embd = hparams.n_embd; // TODO: use a per-batch flag for logits presence instead @@ -542,7 +550,7 @@ size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) { void llama_output_reorder(struct llama_context & ctx) { std::vector & out_ids = ctx.sbatch.out_ids; if (!out_ids.empty()) { - const uint32_t n_vocab = ctx.model.vocab.n_tokens(); + const uint32_t n_vocab = ctx.model.hparams.n_vocab; const uint32_t n_embd = ctx.model.hparams.n_embd; const int32_t n_outputs = ctx.n_outputs; @@ -657,6 +665,10 @@ void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) { ctx->cparams.causal_attn = causal_attn; } +void llama_set_cross_attention(struct llama_context * ctx, bool cross_attention) { + ctx->cparams.cross_attn = cross_attention; +} + void llama_synchronize(struct llama_context * ctx) { ggml_backend_sched_synchronize(ctx->sched.get()); @@ -726,7 +738,7 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); } - return ctx->logits + j*ctx->model.vocab.n_tokens(); + return ctx->logits + j*ctx->model.hparams.n_vocab; } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); #ifndef NDEBUG @@ -886,7 +898,7 @@ struct llama_data_write { } void write_logits(const struct llama_context * ctx) { - const uint64_t logits_size = std::min((uint64_t) ctx->logits_size, (uint64_t) ctx->n_outputs * ctx->model.vocab.n_tokens()); + const uint64_t logits_size = std::min((uint64_t) ctx->logits_size, (uint64_t) ctx->n_outputs * ctx->model.hparams.n_vocab); write(&logits_size, sizeof(logits_size)); diff --git a/src/llama-context.h b/src/llama-context.h index a9268b29..cf12c9d7 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -107,6 +107,8 @@ struct llama_context { struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] + + struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061] }; // TODO: make these methods of llama_context diff --git a/src/llama-cparams.h b/src/llama-cparams.h index 252012f3..9681e5a0 100644 --- a/src/llama-cparams.h +++ b/src/llama-cparams.h @@ -29,6 +29,7 @@ struct llama_cparams { bool offload_kqv; bool flash_attn; bool no_perf; + bool cross_attn; enum llama_pooling_type pooling_type; diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index f3955de9..0b841028 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -2,6 +2,8 @@ #include "ggml.h" +#include + uint32_t llama_hparams::n_head(uint32_t il) const { if (il < n_layer) { return n_head_arr[il]; @@ -76,4 +78,8 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const { } GGML_ABORT("fatal error"); +} + +bool llama_hparams::cross_attention_layers(uint32_t il) const { + return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end(); } \ No newline at end of file diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 1bdcdfd5..05383046 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -41,6 +41,7 @@ struct llama_hparams { uint32_t n_expert = 0; uint32_t n_expert_used = 0; uint32_t n_rel_attn_bkts = 0; + uint32_t n_vocab = 0; // for WavTokenizer struct llama_hparams_posnet posnet; @@ -51,6 +52,7 @@ struct llama_hparams { std::array n_ff_arr; std::array, 4> n_bskcn_arr = {}; + std::array cross_attn_layers; uint32_t n_layer_dense_lead = 0; uint32_t n_lora_q = 0; @@ -138,6 +140,9 @@ struct llama_hparams { // Block skip connection bool n_bskcn(uint32_t n, uint32_t il) const; + + // cross attention layers + bool cross_attention_layers(uint32_t il) const; }; static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index feffdf0d..b541c5a3 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -91,8 +91,17 @@ bool llama_kv_cache_init( return false; } - ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); - ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); + ggml_tensor * k, *v; + + // for cross attention layers + if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) { + k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i)); + v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i)); + } else { + k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); + v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); + } + ggml_format_name(k, "cache_k_l%d", i); ggml_format_name(v, "cache_v_l%d", i); cache.k_l.push_back(k); diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index 1252aca1..45d08721 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -315,6 +315,8 @@ namespace GGUFMeta { return true; } + template bool llama_model_loader::get_arr>(enum llm_kv kid, std::array& result, bool required); + template bool llama_model_loader::get_arr(const std::string & key, std::array & result, bool required) { const int kid = gguf_find_key(meta.get(), key.c_str()); diff --git a/src/llama-model.cpp b/src/llama-model.cpp index ad1315c6..21819080 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -401,6 +401,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { // get general kv ml.get_key(LLM_KV_GENERAL_NAME, name, false); + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab, false); // everything past this point is not vocab-related if (hparams.vocab_only) { @@ -412,6 +413,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false); if (arch == LLM_ARCH_WAVTOKENIZER_DEC) { ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features); @@ -435,9 +437,11 @@ void llama_model::load_hparams(llama_model_loader & ml) { std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0); std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0); std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0); + std::fill(hparams.cross_attn_layers.begin(), hparams.cross_attn_layers.end(), -1); ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer, false); ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer, false); + ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false); // n_head_kv is optional, default to n_head hparams.n_head_kv_arr = hparams.n_head_arr; @@ -486,7 +490,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); - if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) { + if (arch == LLM_ARCH_LLAMA || arch == LLM_ARCH_MLLAMA || arch == LLM_ARCH_DECI || arch == LLM_ARCH_FALCON) { if (hparams.n_rot != hparams.n_embd_head_k) { throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k)); } @@ -530,6 +534,16 @@ void llama_model::load_hparams(llama_model_loader & ml) { } } } break; + case LLM_ARCH_MLLAMA: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 40: type = LLM_TYPE_11B; break; + case 100: type = LLM_TYPE_90B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_DECI: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -1398,7 +1412,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { const int64_t n_embd_head_v = hparams.n_embd_head_v; const int64_t n_ff = hparams.n_ff(); const int64_t n_embd_gqa = n_embd_v_gqa; - const int64_t n_vocab = vocab.n_tokens(); + const int64_t n_vocab = hparams.n_vocab; const int64_t n_token_types = vocab.n_token_types(); const int64_t n_rot = hparams.n_rot; const int64_t n_expert = hparams.n_expert; @@ -1581,6 +1595,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } } break; + case LLM_ARCH_MLLAMA: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8}, 0); + + // output + { + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + if (hparams.cross_attention_layers(i)) { + layer.cross_attn_k_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_NORM, "weight", i), {128}, 0); + layer.cross_attn_k_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_K_PROJ, "weight", i), {n_embd, 1024}, 0); + layer.cross_attn_o_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_O_PROJ, "weight", i), {n_embd, n_embd}, 0); + layer.cross_attn_q_norm = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128}, 0); + layer.cross_attn_q_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd}, 0); + layer.cross_attn_v_proj = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024}, 0); + layer.cross_attn_attn_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1}, 0); + layer.cross_attn_mlp_gate = create_tensor(tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1}, 0); + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + } else { + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } + } break; case LLM_ARCH_DECI: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3925,6 +3985,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { // use what we call a normal RoPE, operating on pairs of consecutive head values case LLM_ARCH_LLAMA: + case LLM_ARCH_MLLAMA: case LLM_ARCH_DECI: case LLM_ARCH_BAICHUAN: case LLM_ARCH_STARCODER: diff --git a/src/llama-model.h b/src/llama-model.h index 1afb0024..7cf57587 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -9,6 +9,7 @@ #include #include #include +#include struct llama_model_loader; @@ -63,6 +64,7 @@ enum llm_type { LLM_TYPE_40B, LLM_TYPE_65B, LLM_TYPE_70B, + LLM_TYPE_90B, LLM_TYPE_236B, LLM_TYPE_314B, LLM_TYPE_671B, @@ -284,6 +286,16 @@ struct llama_layer { struct ggml_tensor * bskcn_tv = nullptr; + // cross attention + struct ggml_tensor * cross_attn_k_norm = nullptr; + struct ggml_tensor * cross_attn_k_proj = nullptr; + struct ggml_tensor * cross_attn_o_proj = nullptr; + struct ggml_tensor * cross_attn_q_norm = nullptr; + struct ggml_tensor * cross_attn_q_proj = nullptr; + struct ggml_tensor * cross_attn_v_proj = nullptr; + struct ggml_tensor * cross_attn_attn_gate = nullptr; + struct ggml_tensor * cross_attn_mlp_gate = nullptr; + struct llama_layer_posnet posnet; struct llama_layer_convnext convnext; diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index fb798265..6eb1da08 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -632,7 +632,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: if (llama_model_has_encoder(&model)) { n_attn_layer *= 3; } - GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected"); + if (qs.n_attention_wv != n_attn_layer) { + LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv); + } } size_t total_size_org = 0; diff --git a/src/llama.cpp b/src/llama.cpp index 6d320ea4..8f7902df 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -154,6 +154,21 @@ static struct ggml_tensor * llm_build_inp_embd( return inpL; } +static struct ggml_tensor * llm_build_inp_cross_attn_state( + struct ggml_context * ctx, + struct llama_context & lctx, + const llama_hparams & hparams, + const llm_build_cb & cb) { + const int64_t n_embd = hparams.n_embd; + + struct ggml_tensor * inpCAS = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4); + cb(inpCAS, "inp_cross_attn_state", -1); + ggml_set_input(inpCAS); + lctx.inp_cross_attn_state = inpCAS; + + return inpCAS; +} + static void llm_build_kv_store( struct ggml_context * ctx, const llama_hparams & hparams, @@ -1157,6 +1172,7 @@ struct llm_build_context { lctx.inp_pos_bucket = nullptr; lctx.inp_embd_enc = nullptr; lctx.inp_KQ_mask_cross = nullptr; + lctx.inp_cross_attn_state = nullptr; } void free() { @@ -1639,6 +1655,240 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_mllama() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + struct ggml_tensor * inpCAS; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); + inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb); + + // 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) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + if (hparams.cross_attention_layers(il)) { + if (!ubatch.embd && !cparams.cross_attn) { + continue; + } + + // cross attention layer + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur); + cb(Qcur, "Qcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + + Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 2, 1, 3)); + cb(Qcur, "Qcur", il); + + Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur, * Vcur; + if (ubatch.embd) { + Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS); + cb(Kcur, "Kcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404); + cb(Kcur, "Kcur", il); + + Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + cb(Kcur, "Kcur", il); + + Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il); + cb(Kcur, "Kcur", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il])); + + Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS); + cb(Vcur, "Vcur", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404); + cb(Vcur, "Vcur", il); + + Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3); + cb(Vcur, "Vcur", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il])); + } else { + Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]); + cb(Kcur, "Kcur (view)", il); + + Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]); + cb(Vcur, "Vcur (view)", il); + } + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur); + cb(kq, "kq", il); + + // TODO: apply causal masks + struct ggml_tensor * kq_soft_max = ggml_soft_max_ext(ctx0, kq, nullptr, 1.f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias); + cb(kq_soft_max, "kq_soft_max", il); + + Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur)); + cb(Vcur, "Vcur", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens); + cb(cur, "kqv_merged_cont", il); + + cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur); + cb(cur, "cur", il); + + // TODO: do this in place once? + cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate)); + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + // TODO: do this inplace once? + cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp); + cb(cur, "ffn_out", il); + + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } else { + // self attention layer + + // 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); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + 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, rope_factors, + n_rot, 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, lctx, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = lctx.cvec.apply_to(ctx0, cur, il); + 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 = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_deci() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); @@ -8344,6 +8594,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_llama(); } break; + case LLM_ARCH_MLLAMA: + { + result = llm.build_mllama(); + } break; case LLM_ARCH_DECI: { result = llm.build_deci(); @@ -8634,7 +8888,7 @@ static int llama_prepare_sbatch( n_outputs = 1; } - lctx.sbatch.from_batch(batch, n_embd, + lctx.sbatch.from_batch(batch, batch.n_embd, /* simple_split */ !lctx.kv_self.recurrent, /* logits_all */ n_outputs == n_tokens_all); @@ -8749,7 +9003,6 @@ static int llama_decode_impl( const llama_batch & batch = batch_allocr.batch; const auto & model = lctx.model; - const auto & vocab = model.vocab; const auto & hparams = model.hparams; const auto & cparams = lctx.cparams; @@ -8760,7 +9013,7 @@ static int llama_decode_impl( llama_kv_slot_restorer kv_slot_restorer(kv_self); const int64_t n_embd = hparams.n_embd; - const int64_t n_vocab = vocab.n_tokens(); + const int64_t n_vocab = hparams.n_vocab; uint32_t n_outputs = 0; uint32_t n_outputs_prev = 0; @@ -9025,7 +9278,7 @@ static int llama_encode_impl( const int64_t n_embd = hparams.n_embd; - lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); + lctx.sbatch.from_batch(batch, batch.n_embd, /* simple_split */ true, /* logits_all */ true); const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens); @@ -9511,6 +9764,7 @@ struct llama_context_params llama_context_default_params() { /*.offload_kqv =*/ true, /*.flash_attn =*/ false, /*.no_perf =*/ true, + /*.cross_attn =*/ false, /*.abort_callback =*/ nullptr, /*.abort_callback_data =*/ nullptr, };