diff --git a/examples/talk-llama/llama-arch.cpp b/examples/talk-llama/llama-arch.cpp index aa21108a..e63ab284 100644 --- a/examples/talk-llama/llama-arch.cpp +++ b/examples/talk-llama/llama-arch.cpp @@ -45,6 +45,9 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_GEMMA3N, "gemma3n" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_MAMBA2, "mamba2" }, + { LLM_ARCH_JAMBA, "jamba" }, + { LLM_ARCH_FALCON_H1, "falcon-h1" }, { LLM_ARCH_XVERSE, "xverse" }, { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_COHERE2, "cohere2" }, @@ -70,6 +73,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_ARWKV7, "arwkv7" }, { LLM_ARCH_GRANITE, "granite" }, { LLM_ARCH_GRANITE_MOE, "granitemoe" }, + { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" }, { LLM_ARCH_CHAMELEON, "chameleon" }, { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" }, { LLM_ARCH_PLM, "plm" }, @@ -77,6 +81,9 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_DOTS1, "dots1" }, { LLM_ARCH_ARCEE, "arcee" }, { LLM_ARCH_ERNIE4_5, "ernie4_5" }, + { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" }, + { LLM_ARCH_SMOLLM3, "smollm3" }, + { LLM_ARCH_LFM2, "lfm2" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -149,7 +156,6 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" }, { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" }, - { LLM_KV_ATTENTION_LAYER_INDICES, "%s.attention.layer_indices" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -170,6 +176,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, + { LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" }, { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" }, { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" }, @@ -182,6 +189,8 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" }, + { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" }, + { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" }, { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, @@ -1004,6 +1013,77 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, }, }, + { + LLM_ARCH_MAMBA2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + }, + }, + { + LLM_ARCH_JAMBA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" }, + { LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + { 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_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_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_ARCH_FALCON_H1, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + { 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_ARCH_XVERSE, { @@ -1564,6 +1644,43 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, }, }, + { + LLM_ARCH_GRANITE_HYBRID, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + // mamba(2) ssm layers + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + // attention layers + { 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" }, + // dense FFN + { 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" }, + // moe FFN + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { 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" }, + // shared expert + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_CHAMELEON, { @@ -1676,6 +1793,67 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_HUNYUAN_MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + { 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_ARCH_SMOLLM3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_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_ARCH_LFM2, + { + { 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_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" }, + { LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" }, + { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" }, + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + } + }, { LLM_ARCH_UNKNOWN, { @@ -1760,7 +1938,11 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_FFN_ACT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}}, {LLM_TENSOR_SSM_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}}, {LLM_TENSOR_SSM_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}}, + {LLM_TENSOR_SSM_DT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_SSM_B_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_SSM_C_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, @@ -1839,6 +2021,9 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_CONVNEXT_PW1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_CONVNEXT_PW2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_CONVNEXT_GAMMA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}}, + {LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, }; LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {} @@ -1894,6 +2079,7 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) { bool llm_arch_is_recurrent(const llm_arch & arch) { switch (arch) { case LLM_ARCH_MAMBA: + case LLM_ARCH_MAMBA2: case LLM_ARCH_RWKV6: case LLM_ARCH_RWKV6QWEN2: case LLM_ARCH_RWKV7: @@ -1905,9 +2091,12 @@ bool llm_arch_is_recurrent(const llm_arch & arch) { } bool llm_arch_is_hybrid(const llm_arch & arch) { - // TODO: There are currently no hybrid models! Once there are, this will be - // the place to identify them switch (arch) { + case LLM_ARCH_JAMBA: + case LLM_ARCH_FALCON_H1: + case LLM_ARCH_GRANITE_HYBRID: + case LLM_ARCH_LFM2: + return true; default: return false; } diff --git a/examples/talk-llama/llama-arch.h b/examples/talk-llama/llama-arch.h index 0771ec3e..1f973259 100644 --- a/examples/talk-llama/llama-arch.h +++ b/examples/talk-llama/llama-arch.h @@ -49,6 +49,9 @@ enum llm_arch { LLM_ARCH_GEMMA3N, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, + LLM_ARCH_MAMBA2, + LLM_ARCH_JAMBA, + LLM_ARCH_FALCON_H1, LLM_ARCH_XVERSE, LLM_ARCH_COMMAND_R, LLM_ARCH_COHERE2, @@ -74,6 +77,7 @@ enum llm_arch { LLM_ARCH_ARWKV7, LLM_ARCH_GRANITE, LLM_ARCH_GRANITE_MOE, + LLM_ARCH_GRANITE_HYBRID, LLM_ARCH_CHAMELEON, LLM_ARCH_WAVTOKENIZER_DEC, LLM_ARCH_PLM, @@ -81,6 +85,9 @@ enum llm_arch { LLM_ARCH_DOTS1, LLM_ARCH_ARCEE, LLM_ARCH_ERNIE4_5, + LLM_ARCH_HUNYUAN_MOE, + LLM_ARCH_SMOLLM3, + LLM_ARCH_LFM2, LLM_ARCH_UNKNOWN, }; @@ -153,7 +160,6 @@ enum llm_kv { LLM_KV_ATTENTION_SCALE, LLM_KV_ATTENTION_KEY_LENGTH_MLA, LLM_KV_ATTENTION_VALUE_LENGTH_MLA, - LLM_KV_ATTENTION_LAYER_INDICES, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_DIMENSION_SECTIONS, @@ -174,6 +180,7 @@ enum llm_kv { LLM_KV_SSM_CONV_KERNEL, LLM_KV_SSM_STATE_SIZE, LLM_KV_SSM_TIME_STEP_RANK, + LLM_KV_SSM_GROUP_COUNT, LLM_KV_SSM_DT_B_C_RMS, LLM_KV_WKV_HEAD_SIZE, @@ -221,6 +228,8 @@ enum llm_kv { LLM_KV_CLASSIFIER_OUTPUT_LABELS, + LLM_KV_SHORTCONV_L_CACHE, + // deprecated: LLM_KV_TOKENIZER_PREFIX_ID, LLM_KV_TOKENIZER_SUFFIX_ID, @@ -291,8 +300,12 @@ enum llm_tensor { LLM_TENSOR_SSM_CONV1D, LLM_TENSOR_SSM_X, LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_DT_NORM, LLM_TENSOR_SSM_A, + LLM_TENSOR_SSM_B_NORM, + LLM_TENSOR_SSM_C_NORM, LLM_TENSOR_SSM_D, + LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, LLM_TENSOR_TIME_MIX_W0, LLM_TENSOR_TIME_MIX_W1, @@ -386,6 +399,9 @@ enum llm_tensor { LLM_TENSOR_POS_NET_ATTN_K, LLM_TENSOR_POS_NET_ATTN_V, LLM_TENSOR_POS_NET_ATTN_OUT, + LLM_TENSOR_SHORTCONV_CONV, + LLM_TENSOR_SHORTCONV_INPROJ, + LLM_TENSOR_SHORTCONV_OUTPROJ, }; enum llm_tensor_layer { diff --git a/examples/talk-llama/llama-batch.cpp b/examples/talk-llama/llama-batch.cpp index 91b1d607..3bc8554e 100644 --- a/examples/talk-llama/llama-batch.cpp +++ b/examples/talk-llama/llama-batch.cpp @@ -166,6 +166,8 @@ bool llama_batch_allocr::init( // note: tracking the other way around is not necessary for now //seq_cpl[s0][s1] = true; + + has_cpl = true; } } } @@ -405,6 +407,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const { return n_outputs; } +uint32_t llama_batch_allocr::get_n_used() const { + return n_used; +} + std::vector & llama_batch_allocr::get_out_ids() { return out_ids; } @@ -420,6 +426,8 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const { void llama_batch_allocr::split_reset() { out_ids.clear(); + n_used = 0; + used.clear(); used.resize(get_n_tokens(), false); @@ -444,6 +452,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) { idxs.push_back(cur_idx); used[cur_idx] = true; + ++n_used; ++cur_idx; @@ -459,9 +468,17 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) { return ubatch_add(idxs, idxs.size(), false); } -llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) { +llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) { + if (sequential && has_cpl) { + LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__); + + return {}; + } + std::vector cur_seq_set; + llama_seq_id last_seq_id = -1; + // determine the non-overlapping sequence sets participating in this ubatch for (int32_t i = 0; i < batch.n_tokens; ++i) { if (used[i]) { @@ -478,9 +495,16 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) { } } + // accept only increasing sequence ids + if (sequential) { + add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1); + } + if (add) { cur_seq_set.push_back(seq_set[i]); + last_seq_id = batch.seq_id[i][0]; + if (cur_seq_set.size() > n_ubatch) { break; } @@ -529,6 +553,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) { idxs_per_seq[s].push_back(idx); used[idx] = true; + ++n_used; ++cur_idx[s]; } @@ -570,6 +595,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) { idxs.push_back(cur_idx); used[cur_idx] = true; + ++n_used; if (idxs.size() >= n_ubatch) { break; diff --git a/examples/talk-llama/llama-batch.h b/examples/talk-llama/llama-batch.h index d2c53761..3420803f 100644 --- a/examples/talk-llama/llama-batch.h +++ b/examples/talk-llama/llama-batch.h @@ -54,6 +54,7 @@ public: uint32_t get_n_tokens() const; uint32_t get_n_outputs() const; + uint32_t get_n_used() const; // the array of output indices in the order they were encountered during the ubatch splitting std::vector & get_out_ids(); @@ -69,7 +70,8 @@ public: llama_ubatch split_simple(uint32_t n_ubatch); // make ubatches of equal-length sequences sets - llama_ubatch split_equal(uint32_t n_ubatch); + // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids + llama_ubatch split_equal(uint32_t n_ubatch, bool sequential); // sequence-set-wise split - each ubatch contains a single sequence-set llama_ubatch split_seq(uint32_t n_ubatch); @@ -112,6 +114,9 @@ private: using pos_set_t = std::set; using seq_cpl_t = std::vector; + // helper flag to quickly determine if there are any coupled sequences in the batch + bool has_cpl; + std::vector seq_pos; // seq_pos[s]: the set of positions in sequence s std::vector seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1 @@ -125,6 +130,8 @@ private: // batch indices of the output std::vector out_ids; + uint32_t n_used; + // used[i] indicates if token i has already been used in a previous ubatch std::vector used; diff --git a/examples/talk-llama/llama-chat.cpp b/examples/talk-llama/llama-chat.cpp index 5d317f4e..cbc19d3c 100644 --- a/examples/talk-llama/llama-chat.cpp +++ b/examples/talk-llama/llama-chat.cpp @@ -64,6 +64,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "bailing", LLM_CHAT_TEMPLATE_BAILING }, { "llama4", LLM_CHAT_TEMPLATE_LLAMA4 }, { "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM }, + { "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE }, }; llm_chat_template llm_chat_template_from_str(const std::string & name) { @@ -185,6 +186,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_LLAMA4; } else if (tmpl_contains("<|endofuserprompt|>")) { return LLM_CHAT_TEMPLATE_DOTS1; + } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) { + return LLM_CHAT_TEMPLATE_HUNYUAN_MOE; } return LLM_CHAT_TEMPLATE_UNKNOWN; } @@ -665,6 +668,18 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|response|>"; } + } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_MOE) { + // tencent/Hunyuan-A13B-Instruct + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + ss << "<|startoftext|>" << message->content << "<|extra_4|>"; + } else if (role == "assistant") { + ss << "<|startoftext|>" << message->content << "<|eos|>"; + } else { + ss << "<|startoftext|>" << message->content << "<|extra_0|>"; + } + } } else { // template not supported return -1; diff --git a/examples/talk-llama/llama-chat.h b/examples/talk-llama/llama-chat.h index 38800010..b621fda2 100644 --- a/examples/talk-llama/llama-chat.h +++ b/examples/talk-llama/llama-chat.h @@ -44,6 +44,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_LLAMA4, LLM_CHAT_TEMPLATE_SMOLVLM, LLM_CHAT_TEMPLATE_DOTS1, + LLM_CHAT_TEMPLATE_HUNYUAN_MOE, LLM_CHAT_TEMPLATE_UNKNOWN, }; diff --git a/examples/talk-llama/llama-graph.cpp b/examples/talk-llama/llama-graph.cpp index 010300df..a248a7ec 100644 --- a/examples/talk-llama/llama-graph.cpp +++ b/examples/talk-llama/llama-graph.cpp @@ -281,19 +281,22 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) { - if (self_kq_mask) { - mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); - } + mctx->set_input_k_idxs(self_k_idxs, ubatch); + mctx->set_input_v_idxs(self_v_idxs, ubatch); + + mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); } void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) { - if (self_kq_mask) { - mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); - } + mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch); + mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch); - if (self_kq_mask_swa) { - mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn); - } + mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); + + mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch); + mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch); + + mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn); } void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { @@ -333,27 +336,8 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) { - if (self_kq_mask) { - mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); - } - - const int64_t n_rs = mctx->get_recr()->get_n_rs(); - - if (s_copy) { - GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); - int32_t * data = (int32_t *) s_copy->data; - - // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n - for (uint32_t i = 0; i < n_rs; ++i) { - data[i] = mctx->get_recr()->s_copy(i); - } - } -} - -void llm_graph_input_one::set_input(const llama_ubatch *) { - GGML_ASSERT(one && ggml_nelements(one) == 1); - float f_one = 1.0f; - ggml_backend_tensor_set(one, &f_one, 0, sizeof(float)); + inp_attn->set_input(ubatch); + inp_rs->set_input(ubatch); } // @@ -987,33 +971,6 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t return pos_bias; } -llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const { - const auto * mctx_cur = static_cast(mctx); - - auto inp = std::make_unique(hparams, cparams, mctx_cur); - - { - GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers"); - - const auto n_kv = inp->mctx->get_attn()->get_n_kv(); - - inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - //cb(inp->self_kq_mask, "KQ_mask", -1); - ggml_set_input(inp->self_kq_mask); - - inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; - } - - { - const auto n_rs = mctx_cur->get_recr()->get_n_rs(); - - inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs); - ggml_set_input(inp->s_copy); - } - - return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp)); -} - ggml_tensor * llm_graph_context::build_attn_mha( ggml_cgraph * gf, ggml_tensor * q, @@ -1135,8 +1092,7 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con auto inp = std::make_unique(hparams, cparams); // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch - inp->kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - //cb(inp_kq_mask, "KQ_mask", -1); + inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1); ggml_set_input(inp->kq_mask); inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask; @@ -1188,8 +1144,12 @@ ggml_tensor * llm_graph_context::build_attn( return cur; } -llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const { - const auto * mctx_cur = static_cast(mctx); +static std::unique_ptr build_attn_inp_kv_unified_impl( + ggml_context * ctx0, + const llama_ubatch & ubatch, + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache_unified_context * mctx_cur) { auto inp = std::make_unique(hparams, cparams, mctx_cur); @@ -1197,14 +1157,25 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA"); const auto n_kv = mctx_cur->get_n_kv(); + const auto n_tokens = ubatch.n_tokens; - inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - //cb(inp->self_kq_mask, "KQ_mask", -1); + inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch); + inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch); + + inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1); ggml_set_input(inp->self_kq_mask); inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; } + return inp; +} + +llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const { + const auto * mctx_cur = static_cast(mctx); + + auto inp = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur); + return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp)); } @@ -1226,12 +1197,15 @@ ggml_tensor * llm_graph_context::build_attn( ggml_build_forward_expand(gf, k_cur); ggml_build_forward_expand(gf, v_cur); - const auto * mctx_cur = static_cast(mctx); + const auto * mctx_cur = inp->mctx; // store to KV cache { - ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); - ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); + const auto & k_idxs = inp->get_k_idxs(); + const auto & v_idxs = inp->get_v_idxs(); + + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il)); + ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il)); } const auto & kq_mask = inp->get_kq_mask(); @@ -1282,7 +1256,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_build_forward_expand(gf, v_cur); } - const auto * mctx_iswa = static_cast(mctx); + const auto * mctx_iswa = inp->mctx; const bool is_swa = hparams.is_swa(il); @@ -1290,11 +1264,15 @@ ggml_tensor * llm_graph_context::build_attn( // optionally store to KV cache if (k_cur) { - ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); + const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs(); + + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il)); } if (v_cur) { - ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); + const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs(); + + ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il)); } const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask(); @@ -1326,7 +1304,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const { const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; - inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1); ggml_set_input(inp->cross_kq_mask); inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask; @@ -1376,56 +1354,9 @@ ggml_tensor * llm_graph_context::build_attn( return cur; } -ggml_tensor * llm_graph_context::build_attn( - llm_graph_input_mem_hybrid * inp, - ggml_cgraph * gf, - ggml_tensor * wo, - ggml_tensor * wo_b, - ggml_tensor * q_cur, - ggml_tensor * k_cur, - ggml_tensor * v_cur, - ggml_tensor * kq_b, - ggml_tensor * v_mla, - float kq_scale, - int il) const { - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(gf, q_cur); - ggml_build_forward_expand(gf, k_cur); - ggml_build_forward_expand(gf, v_cur); - - const auto * mctx_cur = static_cast(mctx)->get_attn(); - - // store to KV cache - { - ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); - ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); - } - - const auto & kq_mask = inp->get_kq_mask(); - - ggml_tensor * q = q_cur; - ggml_tensor * k = mctx_cur->get_k(ctx0, il); - ggml_tensor * v = mctx_cur->get_v(ctx0, il); - - ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale); - cb(cur, "kqv_out", il); - - if (wo) { - cur = build_lora_mm(wo, cur); - if (arch == LLM_ARCH_GLM4) { - // GLM4 seems to have numerical issues with half-precision accumulators - ggml_mul_mat_set_prec(cur, GGML_PREC_F32); - } - } - - if (wo_b) { - cur = ggml_add(ctx0, cur, wo_b); - } - - return cur; -} - +// TODO: maybe separate the inner implementation into a separate function +// like with the non-sliding window equivalent +// once sliding-window hybrid caches are a thing. llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const { const auto * mctx_cur = static_cast(mctx); @@ -1434,8 +1365,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif { const auto n_kv = mctx_cur->get_base()->get_n_kv(); - inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - //cb(inp->self_kq_mask, "KQ_mask", -1); + inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch); + inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch); + + inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1); ggml_set_input(inp->self_kq_mask); inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; @@ -1446,8 +1379,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif const auto n_kv = mctx_cur->get_swa()->get_n_kv(); - inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1); + inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch); + inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch); + + inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1); ggml_set_input(inp->self_kq_mask_swa); inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa; @@ -1466,7 +1401,7 @@ ggml_tensor * llm_graph_context::build_rs( uint32_t kv_head, uint32_t kv_size, int32_t rs_zero, - bool avoid_copies) const { + const llm_graph_get_rows_fn & get_state_rows) const { ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size); @@ -1475,19 +1410,11 @@ ggml_tensor * llm_graph_context::build_rs( ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0)); ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0)); - ggml_tensor * output_states; - - if (!avoid_copies) { - // copy states - // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv - // {state_size, kv_size} -> {state_size, n_seqs} - output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0)); - ggml_build_forward_expand(gf, output_states); - } else { - // FIXME: make the gathering operation happen before the copy below - // (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?) - output_states = states; - } + // copy states + // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv + // {state_size, kv_size} -> {state_size, n_seqs} + ggml_tensor * output_states = get_state_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0)); + ggml_build_forward_expand(gf, output_states); // copy extra states which won't be changed further (between n_seqs and n_kv) ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0])); @@ -1499,8 +1426,9 @@ ggml_tensor * llm_graph_context::build_rs( return output_states; } -llm_graph_input_rs * llm_graph_context::build_rs_inp() const { - const auto * mctx_cur = static_cast(mctx); +static std::unique_ptr build_rs_inp_impl( + ggml_context * ctx0, + const llama_memory_recurrent_context * mctx_cur) { auto inp = std::make_unique(mctx_cur); @@ -1509,6 +1437,14 @@ llm_graph_input_rs * llm_graph_context::build_rs_inp() const { inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs); ggml_set_input(inp->s_copy); + return inp; +} + +llm_graph_input_rs * llm_graph_context::build_rs_inp() const { + const auto * mctx_cur = static_cast(mctx); + + auto inp = build_rs_inp_impl(ctx0, mctx_cur); + return (llm_graph_input_rs *) res->add_input(std::move(inp)); } @@ -1518,22 +1454,10 @@ ggml_tensor * llm_graph_context::build_rs( ggml_tensor * s, int32_t state_size, int32_t n_seqs, - bool avoid_copies) const { - const auto * mctx_cur = static_cast(mctx); + const llm_graph_get_rows_fn & get_state_rows) const { + const auto * kv_state = inp->mctx; - return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies); -} - -ggml_tensor * llm_graph_context::build_rs( - llm_graph_input_mem_hybrid * inp, - ggml_cgraph * gf, - ggml_tensor * s, - int32_t state_size, - int32_t n_seqs, - bool avoid_copies) const { - const auto * mctx_cur = static_cast(mctx)->get_recr(); - - return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies); + return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows); } ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( @@ -1578,6 +1502,17 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( ); } +llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const { + const auto * mctx_cur = static_cast(mctx); + + auto inp_rs = build_rs_inp_impl(ctx0, mctx_cur->get_recr()); + auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn()); + + auto inp = std::make_unique(std::move(inp_attn), std::move(inp_rs), mctx_cur); + + return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp)); +} + void llm_graph_context::build_pooling( ggml_cgraph * gf, ggml_tensor * cls, diff --git a/examples/talk-llama/llama-graph.h b/examples/talk-llama/llama-graph.h index ceddb602..fbf8e288 100644 --- a/examples/talk-llama/llama-graph.h +++ b/examples/talk-llama/llama-graph.h @@ -228,8 +228,8 @@ public: ggml_tensor * get_kq_mask() const { return kq_mask_cnv; } - ggml_tensor * kq_mask = nullptr; // F32 [n_tokens, n_batch] - ggml_tensor * kq_mask_cnv = nullptr; // [n_tokens, n_batch] + ggml_tensor * kq_mask = nullptr; // F32 [n_tokens, n_batch, 1, 1] + ggml_tensor * kq_mask_cnv = nullptr; // [n_tokens, n_batch, 1, 1] const llama_hparams & hparams; const llama_cparams & cparams; @@ -249,10 +249,16 @@ public: void set_input(const llama_ubatch * ubatch) override; + ggml_tensor * get_k_idxs() const { return self_k_idxs; } + ggml_tensor * get_v_idxs() const { return self_v_idxs; } + ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } - ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch] - ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch] + ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch] + ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] + + ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch, 1, 1] + ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch, 1, 1] const llama_hparams & hparams; const llama_cparams & cparams; @@ -274,13 +280,23 @@ public: void set_input(const llama_ubatch * ubatch) override; + ggml_tensor * get_k_idxs() const { return self_k_idxs; } + ggml_tensor * get_v_idxs() const { return self_v_idxs; } + ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; } + ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; } + ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; } - ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch] - ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch] - ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch] - ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch] + ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch] + ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] + ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch] + ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] + + ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch, 1, 1] + ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch, 1, 1] + ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch, 1, 1] + ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch, 1, 1] const llama_hparams & hparams; const llama_cparams & cparams; @@ -297,8 +313,8 @@ public: ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; } - ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch] - ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch] + ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1] + ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1] const llama_cross * cross = nullptr; }; @@ -306,41 +322,25 @@ public: class llm_graph_input_mem_hybrid : public llm_graph_input_i { public: llm_graph_input_mem_hybrid( - const llama_hparams & hparams, - const llama_cparams & cparams, - const llama_memory_hybrid_context * mctx) : - hparams(hparams), - cparams(cparams), - mctx(mctx) { - } + std::unique_ptr inp_attn, + std::unique_ptr inp_rs, + const llama_memory_hybrid_context * mctx) : + inp_attn(std::move(inp_attn)), + inp_rs(std::move(inp_rs)), + mctx(mctx) { } virtual ~llm_graph_input_mem_hybrid() = default; void set_input(const llama_ubatch * ubatch) override; - ggml_tensor * s_copy; // I32 [kv_size] + std::unique_ptr inp_attn; + std::unique_ptr inp_rs; - ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } - - ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch] - ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch] - - const llama_hparams & hparams; - const llama_cparams & cparams; + llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); } + llm_graph_input_rs * get_recr() const { return inp_rs.get(); } const llama_memory_hybrid_context * mctx; }; -// TODO: remove this when ggml_scale_add is implemented -class llm_graph_input_one : public llm_graph_input_i { -public: - llm_graph_input_one() {} - virtual ~llm_graph_input_one() = default; - - void set_input(const llama_ubatch *) override; - - ggml_tensor * one = nullptr; // F32 -}; - // // llm_graph_result // @@ -424,6 +424,9 @@ struct llm_graph_params { const llm_graph_cb & cb; }; +// used in build_rs to properly order writes and avoid unnecessary copies +using llm_graph_get_rows_fn = std::function; + struct llm_graph_context { const llm_arch arch; @@ -554,8 +557,6 @@ struct llm_graph_context { ggml_tensor * build_inp_pos_bucket_dec() const; ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const; - llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const; - // // attention // @@ -631,18 +632,6 @@ struct llm_graph_context { float kq_scale, int il) const; - ggml_tensor * build_attn( - llm_graph_input_mem_hybrid * inp, - ggml_cgraph * gf, - ggml_tensor * wo, - ggml_tensor * wo_b, - ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] - ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] - ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] - ggml_tensor * kq_b, - ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] - float kq_scale, - int il) const; // // recurrent // @@ -663,7 +652,7 @@ struct llm_graph_context { uint32_t kv_head, uint32_t kv_size, int32_t rs_zero, - bool avoid_copies = false) const; + const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const; llm_graph_input_rs * build_rs_inp() const; @@ -673,15 +662,7 @@ struct llm_graph_context { ggml_tensor * s, int32_t state_size, int32_t n_seqs, - bool avoid_copies = false) const; - - ggml_tensor * build_rs( - llm_graph_input_mem_hybrid * inp, - ggml_cgraph * gf, - ggml_tensor * s, - int32_t state_size, - int32_t n_seqs, - bool avoid_copies = false) const; + const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const; ggml_tensor * build_rwkv_token_shift_load( llm_graph_input_rs * inp, @@ -693,6 +674,11 @@ struct llm_graph_context { ggml_tensor * token_shift, const llama_ubatch & ubatch, int il) const; + // + // hybrid + // + + llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const; // // pooling diff --git a/examples/talk-llama/llama-hparams.cpp b/examples/talk-llama/llama-hparams.cpp index bba7a12d..7aa736e2 100644 --- a/examples/talk-llama/llama-hparams.cpp +++ b/examples/talk-llama/llama-hparams.cpp @@ -71,9 +71,15 @@ uint32_t llama_hparams::n_embd_r() const { return token_shift_count * n_embd; } + if (n_shortconv_l_cache != 0) { + // for LFM2 models + return n_embd * (n_shortconv_l_cache - 1); + } + // TODO: maybe support other convolution strides than 1 // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed - return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner; + // Corresponds to Mamba's conv_states size + return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * (ssm_d_inner + 2*ssm_n_group*ssm_d_state); } uint32_t llama_hparams::n_embd_s() const { diff --git a/examples/talk-llama/llama-hparams.h b/examples/talk-llama/llama-hparams.h index e85afe14..d0500e4d 100644 --- a/examples/talk-llama/llama-hparams.h +++ b/examples/talk-llama/llama-hparams.h @@ -55,6 +55,8 @@ struct llama_hparams { struct llama_hparams_posnet posnet; struct llama_hparams_convnext convnext; + uint32_t n_shortconv_l_cache = 0; + std::array n_head_arr; std::array n_head_kv_arr; std::array n_ff_arr; @@ -114,6 +116,7 @@ struct llama_hparams { uint32_t ssm_d_inner = 0; uint32_t ssm_d_state = 0; uint32_t ssm_dt_rank = 0; + uint32_t ssm_n_group = 0; // for hybrid state space models std::array recurrent_layer_arr; diff --git a/examples/talk-llama/llama-kv-cache-unified-iswa.cpp b/examples/talk-llama/llama-kv-cache-unified-iswa.cpp index d1f839b6..fe207ad5 100644 --- a/examples/talk-llama/llama-kv-cache-unified-iswa.cpp +++ b/examples/talk-llama/llama-kv-cache-unified-iswa.cpp @@ -113,20 +113,25 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all ubatches.push_back(std::move(ubatch)); // NOLINT } - auto heads_base = kv_base->prepare(ubatches); - if (heads_base.empty()) { + if (balloc.get_n_used() < balloc.get_n_tokens()) { + // failed to find a suitable split break; } - auto heads_swa = kv_swa->prepare(ubatches); - if (heads_swa.empty()) { + auto sinfos_base = kv_base->prepare(ubatches); + if (sinfos_base.empty()) { break; } - assert(heads_base.size() == heads_swa.size()); + auto sinfos_swa = kv_swa->prepare(ubatches); + if (sinfos_swa.empty()) { + break; + } + + assert(sinfos_base.size() == sinfos_swa.size()); return std::make_unique( - this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches)); } while (false); // if it fails, try equal split @@ -135,7 +140,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all std::vector ubatches; while (true) { - auto ubatch = balloc.split_equal(n_ubatch); + auto ubatch = balloc.split_equal(n_ubatch, false); if (ubatch.n_tokens == 0) { break; @@ -144,20 +149,25 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all ubatches.push_back(std::move(ubatch)); // NOLINT } - auto heads_base = kv_base->prepare(ubatches); - if (heads_base.empty()) { + if (balloc.get_n_used() < balloc.get_n_tokens()) { + // failed to find a suitable split break; } - auto heads_swa = kv_swa->prepare(ubatches); - if (heads_swa.empty()) { + auto sinfos_base = kv_base->prepare(ubatches); + if (sinfos_base.empty()) { break; } - assert(heads_base.size() == heads_swa.size()); + auto sinfos_swa = kv_swa->prepare(ubatches); + if (sinfos_swa.empty()) { + break; + } + + assert(sinfos_base.size() == sinfos_swa.size()); return std::make_unique( - this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches)); } while (false); // TODO: if we fail again, we should attempt different splitting strategies @@ -220,13 +230,13 @@ llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, - std::vector heads_base, - std::vector heads_swa, + slot_info_vec_t sinfos_base, + slot_info_vec_t sinfos_swa, std::vector ubatches) : ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(heads_base), this->ubatches)), - ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(heads_swa), this->ubatches)), + ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)), + ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa), this->ubatches)), status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) { } diff --git a/examples/talk-llama/llama-kv-cache-unified-iswa.h b/examples/talk-llama/llama-kv-cache-unified-iswa.h index 46c1ed61..23205d82 100644 --- a/examples/talk-llama/llama-kv-cache-unified-iswa.h +++ b/examples/talk-llama/llama-kv-cache-unified-iswa.h @@ -74,6 +74,8 @@ private: class llama_kv_cache_unified_iswa_context : public llama_memory_context_i { public: + using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t; + // used for errors llama_kv_cache_unified_iswa_context(llama_memory_status status); @@ -90,8 +92,8 @@ public: // used to create a batch processing context from a batch llama_kv_cache_unified_iswa_context( llama_kv_cache_unified_iswa * kv, - std::vector heads_base, - std::vector heads_swa, + slot_info_vec_t sinfos_base, + slot_info_vec_t sinfos_swa, std::vector ubatches); virtual ~llama_kv_cache_unified_iswa_context(); diff --git a/examples/talk-llama/llama-kv-cache-unified.cpp b/examples/talk-llama/llama-kv-cache-unified.cpp index 7f7b162f..d3129cc5 100644 --- a/examples/talk-llama/llama-kv-cache-unified.cpp +++ b/examples/talk-llama/llama-kv-cache-unified.cpp @@ -156,6 +156,13 @@ llama_kv_cache_unified::llama_kv_cache_unified( const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG"); debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0; + + const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS"); + supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0; + + if (!supports_set_rows) { + LLAMA_LOG_WARN("%s: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility\n", __func__); + } } void llama_kv_cache_unified::clear(bool data) { @@ -353,13 +360,18 @@ llama_memory_context_ptr llama_kv_cache_unified::init_batch( ubatches.push_back(std::move(ubatch)); // NOLINT } - auto heads = prepare(ubatches); - if (heads.empty()) { + if (balloc.get_n_used() < balloc.get_n_tokens()) { + // failed to find a suitable split + break; + } + + auto sinfos = prepare(ubatches); + if (sinfos.empty()) { break; } return std::make_unique( - this, std::move(heads), std::move(ubatches)); + this, std::move(sinfos), std::move(ubatches)); } while (false); return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); @@ -402,12 +414,13 @@ llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lct return std::make_unique(this, lctx, do_shift, std::move(dinfo)); } -llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::vector & ubatches) { - llama_kv_cache_unified::ubatch_heads res; +llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const std::vector & ubatches) { + llama_kv_cache_unified::slot_info_vec_t res; struct state { uint32_t head_old; // old position of the head, before placing the ubatch - uint32_t head_new; // new position of the head, after placing the ubatch + + slot_info sinfo; // slot info for the ubatch llama_kv_cells_unified cells; // copy of the old cells, before placing the ubatch }; @@ -418,26 +431,29 @@ llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std:: bool success = true; for (const auto & ubatch : ubatches) { + // non-continuous slots require support for ggml_set_rows() + const bool cont = supports_set_rows ? false : true; + // only find a suitable slot for the ubatch. don't modify the cells yet - const int32_t head_new = find_slot(ubatch); - if (head_new < 0) { + const auto sinfo_new = find_slot(ubatch, cont); + if (sinfo_new.empty()) { success = false; break; } // remeber the position that we found - res.push_back(head_new); + res.push_back(sinfo_new); // store the old state of the cells in the recovery stack - states.push_back({head, (uint32_t) head_new, cells.cp(head_new, ubatch.n_tokens)}); + states.push_back({head, sinfo_new, cells.cp(sinfo_new.idxs)}); // now emplace the ubatch - apply_ubatch(head_new, ubatch); + apply_ubatch(sinfo_new, ubatch); } // iterate backwards and restore the cells to their original state for (auto it = states.rbegin(); it != states.rend(); ++it) { - cells.set(it->head_new, it->cells); + cells.set(it->sinfo.idxs, it->cells); head = it->head_old; } @@ -539,7 +555,7 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d return updated; } -int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { +llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const { const uint32_t n_tokens = ubatch.n_tokens; uint32_t head_cur = this->head; @@ -552,7 +568,7 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { if (n_tokens > cells.size()) { LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size()); - return -1; + return { }; } if (debug > 0) { @@ -615,15 +631,26 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { uint32_t n_tested = 0; + // for continuous slots, we test that all tokens in the ubatch fit, starting from the current head + // for non-continuous slots, we test the tokens one by one + const uint32_t n_test = cont ? n_tokens : 1; + + slot_info res; + + auto & idxs = res.idxs; + + idxs.reserve(n_tokens); + while (true) { - if (head_cur + n_tokens > cells.size()) { + if (head_cur + n_test > cells.size()) { n_tested += cells.size() - head_cur; head_cur = 0; continue; } - bool found = true; - for (uint32_t i = 0; i < n_tokens; i++) { + for (uint32_t i = 0; i < n_test; i++) { + const auto idx = head_cur; + //const llama_pos pos = ubatch.pos[i]; //const llama_seq_id seq_id = ubatch.seq_id[i][0]; @@ -633,19 +660,19 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { // - (disabled) mask causally, if the sequence is the same as the one we are inserting // - mask SWA, using current max pos for that sequence in the cache // always insert in the cell with minimum pos - bool can_use = cells.is_empty(head_cur + i); + bool can_use = cells.is_empty(idx); - if (!can_use && cells.seq_count(head_cur + i) == 1) { - const llama_pos pos_cell = cells.pos_get(head_cur + i); + if (!can_use && cells.seq_count(idx) == 1) { + const llama_pos pos_cell = cells.pos_get(idx); // (disabled) causal mask // note: it's better to purge any "future" tokens beforehand - //if (cells.seq_has(head_cur + i, seq_id)) { + //if (cells.seq_has(idx, seq_id)) { // can_use = pos_cell >= pos; //} if (!can_use) { - const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i); + const llama_seq_id seq_id_cell = cells.seq_get(idx); // SWA mask if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) { @@ -654,28 +681,39 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { } } - if (!can_use) { - found = false; - head_cur += i + 1; - n_tested += i + 1; + head_cur++; + n_tested++; + + if (can_use) { + idxs.push_back(idx); + } else { break; } } - if (found) { + if (idxs.size() == n_tokens) { break; } + if (cont) { + idxs.clear(); + } + if (n_tested >= cells.size()) { //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); - return -1; + return { }; } } - return head_cur; + // we didn't find a suitable slot - return empty result + if (idxs.size() < n_tokens) { + res.clear(); + } + + return res; } -void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) { +void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) { // keep track of the max sequence position that we would overwrite with this ubatch // for non-SWA cache, this would be always empty llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ]; @@ -683,22 +721,26 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch seq_pos_max_rm[s] = -1; } - for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { - if (!cells.is_empty(head_cur + i)) { - assert(cells.seq_count(head_cur + i) == 1); + assert(ubatch.n_tokens == sinfo.idxs.size()); - const llama_seq_id seq_id = cells.seq_get(head_cur + i); - const llama_pos pos = cells.pos_get(head_cur + i); + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + const auto idx = sinfo.idxs.at(i); + + if (!cells.is_empty(idx)) { + assert(cells.seq_count(idx) == 1); + + const llama_seq_id seq_id = cells.seq_get(idx); + const llama_pos pos = cells.pos_get(idx); seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos); - cells.rm(head_cur + i); + cells.rm(idx); } - cells.pos_set(head_cur + i, ubatch.pos[i]); + cells.pos_set(idx, ubatch.pos[i]); for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) { - cells.seq_add(head_cur + i, ubatch.seq_id[i][s]); + cells.seq_add(idx, ubatch.seq_id[i][s]); } } @@ -719,7 +761,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch } // move the head at the end of the slot - head = head_cur + ubatch.n_tokens; + head = sinfo.idxs.back() + 1; } bool llama_kv_cache_unified::get_can_shift() const { @@ -772,47 +814,133 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint 0); } -ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const { +ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const { const int32_t ikv = map_layer_ids.at(il); auto * k = layers[ikv].k; + const int64_t n_embd_k_gqa = k->ne[0]; const int64_t n_tokens = k_cur->ne[2]; + k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens); + + if (k_idxs && supports_set_rows) { + return ggml_set_rows(ctx, k, k_cur, k_idxs); + } + + // TODO: fallback to old ggml_cpy() method for backwards compatibility + // will be removed when ggml_set_rows() is adopted by all backends + ggml_tensor * k_view = ggml_view_1d(ctx, k, - n_tokens*hparams.n_embd_k_gqa(il), - ggml_row_size(k->type, hparams.n_embd_k_gqa(il))*head_cur); + n_tokens*n_embd_k_gqa, + ggml_row_size(k->type, n_embd_k_gqa)*sinfo.head()); return ggml_cpy(ctx, k_cur, k_view); } -ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const { +ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const { const int32_t ikv = map_layer_ids.at(il); auto * v = layers[ikv].v; + const int64_t n_embd_v_gqa = v->ne[0]; const int64_t n_tokens = v_cur->ne[2]; - v_cur = ggml_reshape_2d(ctx, v_cur, hparams.n_embd_v_gqa(il), n_tokens); + v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens); + + if (v_idxs && supports_set_rows) { + if (!v_trans) { + return ggml_set_rows(ctx, v, v_cur, v_idxs); + } + + // the row becomes a single element + ggml_tensor * v_view = ggml_reshape_3d(ctx, v, 1, v->ne[1], v->ne[0]); + + // note: the V cache is transposed when not using flash attention + v_cur = ggml_permute(ctx, ggml_reshape_3d(ctx, v_cur, v_cur->ne[0], 1, v_cur->ne[1]), 2, 0, 1, 3); + + // note: we can be more explicit here at the cost of extra cont + // however, above we take advantage that a row of single element is always continuous regardless of the row stride + //v_cur = ggml_transpose(ctx, v_cur); + //v_cur = ggml_cont_3d(ctx, v_cur, 1, v_cur->ne[0], v_cur->ne[1]); + + // we broadcast the KV indices n_embd_v_gqa times + // v [1, n_kv, n_embd_v_gqa] + // v_cur [1, n_tokens, n_embd_v_gqa] + // v_idxs [n_tokens, 1, 1] + return ggml_set_rows(ctx, v_view, v_cur, v_idxs); + } + + // TODO: fallback to old ggml_cpy() method for backwards compatibility + // will be removed when ggml_set_rows() is adopted by all backends ggml_tensor * v_view = nullptr; if (!v_trans) { v_view = ggml_view_1d(ctx, v, - n_tokens*hparams.n_embd_v_gqa(il), - ggml_row_size(v->type, hparams.n_embd_v_gqa(il))*head_cur); + n_tokens*n_embd_v_gqa, + ggml_row_size(v->type, n_embd_v_gqa)*sinfo.head()); } else { - // note: the V cache is transposed when not using flash attention - v_view = ggml_view_2d(ctx, v, n_tokens, hparams.n_embd_v_gqa(il), - (v->ne[1])*ggml_element_size(v), - (head_cur)*ggml_element_size(v)); - v_cur = ggml_transpose(ctx, v_cur); + + v_view = ggml_view_2d(ctx, v, n_tokens, n_embd_v_gqa, + (v->ne[1] )*ggml_element_size(v), + (sinfo.head())*ggml_element_size(v)); } return ggml_cpy(ctx, v_cur, v_view); } +ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const { + const uint32_t n_tokens = ubatch.n_tokens; + + ggml_tensor * k_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens); + + ggml_set_input(k_idxs); + + return k_idxs; +} + +ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const { + const uint32_t n_tokens = ubatch.n_tokens; + + ggml_tensor * v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens); + + ggml_set_input(v_idxs); + + return v_idxs; +} + +void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const { + if (!supports_set_rows) { + return; + } + + const uint32_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer)); + int64_t * data = (int64_t *) dst->data; + + for (int64_t i = 0; i < n_tokens; ++i) { + data[i] = sinfo.idxs.at(i); + } +} + +void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const { + if (!supports_set_rows) { + return; + } + + const uint32_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer)); + int64_t * data = (int64_t *) dst->data; + + for (int64_t i = 0; i < n_tokens; ++i) { + data[i] = sinfo.idxs.at(i); + } +} + void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { const uint32_t n_tokens = ubatch->n_tokens; @@ -1552,13 +1680,15 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell ubatch.seq_id[i] = &dest_seq_id; } - const auto head_cur = find_slot(ubatch); - if (head_cur < 0) { + const auto sinfo = find_slot(ubatch, true); + if (sinfo.empty()) { LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); return false; } - apply_ubatch(head_cur, ubatch); + apply_ubatch(sinfo, ubatch); + + const auto head_cur = sinfo.head(); // keep the head at the old position because we will read the KV data into it in state_read_data() head = head_cur; @@ -1744,7 +1874,11 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_stat llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) { n_kv = kv->get_size(); - head = 0; + + // create a dummy slot info - the actual data is irrelevant. we just need to build the graph + sinfos.resize(1); + sinfos[0].idxs.resize(1); + sinfos[0].idxs[0] = 0; } llama_kv_cache_unified_context::llama_kv_cache_unified_context( @@ -1759,8 +1893,8 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified_context::llama_kv_cache_unified_context( llama_kv_cache_unified * kv, - llama_kv_cache_unified::ubatch_heads heads, - std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) { + llama_kv_cache_unified::slot_info_vec_t sinfos, + std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sinfos(std::move(sinfos)), ubatches(std::move(ubatches)) { } llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default; @@ -1768,7 +1902,7 @@ llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default; bool llama_kv_cache_unified_context::next() { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - if (++i_next >= ubatches.size()) { + if (++i_cur >= ubatches.size()) { return false; } @@ -1785,10 +1919,9 @@ bool llama_kv_cache_unified_context::apply() { return true; } - kv->apply_ubatch(heads[i_next], ubatches[i_next]); + kv->apply_ubatch(sinfos[i_cur], ubatches[i_cur]); n_kv = kv->get_n_kv(); - head = heads[i_next]; return true; } @@ -1800,7 +1933,7 @@ llama_memory_status llama_kv_cache_unified_context::get_status() const { const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const { assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - return ubatches[i_next]; + return ubatches[i_cur]; } uint32_t llama_kv_cache_unified_context::get_n_kv() const { @@ -1815,18 +1948,34 @@ ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t return kv->get_v(ctx, il, n_kv); } -ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const { - return kv->cpy_k(ctx, k_cur, il, head); +ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const { + return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]); } -ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const { - return kv->cpy_v(ctx, v_cur, il, head); +ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const { + return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]); +} + +ggml_tensor * llama_kv_cache_unified_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const { + return kv->build_input_k_idxs(ctx, ubatch); +} + +ggml_tensor * llama_kv_cache_unified_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const { + return kv->build_input_v_idxs(ctx, ubatch); } void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const { kv->set_input_k_shift(dst); } +void llama_kv_cache_unified_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const { + kv->set_input_k_idxs(dst, ubatch, sinfos[i_cur]); +} + +void llama_kv_cache_unified_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const { + kv->set_input_v_idxs(dst, ubatch, sinfos[i_cur]); +} + void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { kv->set_input_kq_mask(dst, ubatch, causal_attn); } diff --git a/examples/talk-llama/llama-kv-cache-unified.h b/examples/talk-llama/llama-kv-cache-unified.h index 4c53f127..b8b0356e 100644 --- a/examples/talk-llama/llama-kv-cache-unified.h +++ b/examples/talk-llama/llama-kv-cache-unified.h @@ -24,8 +24,6 @@ public: // this callback is used to filter out layers that should not be included in the cache using layer_filter_cb = std::function; - using ubatch_heads = std::vector; - struct defrag_info { bool empty() const { return ids.empty(); @@ -37,6 +35,32 @@ public: std::vector ids; }; + // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the + // KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]] + struct slot_info { + // data for ggml_set_rows + using idx_vec_t = std::vector; + + idx_vec_t idxs; + + uint32_t head() const { + return idxs.at(0); + } + + bool empty() const { + return idxs.empty(); + } + + void clear() { + idxs.clear(); + } + + // TODO: implement + //std::vector seq_idxs; + }; + + using slot_info_vec_t = std::vector; + llama_kv_cache_unified( const llama_model & model, layer_filter_cb && filter, @@ -102,30 +126,37 @@ public: ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const; // store k_cur and v_cur in the cache based on the provided head location - ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const; - ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const; + ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const; + ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const; // // preparation API // - // find places for the provided ubatches in the cache, returns the head locations + // find places for the provided ubatches in the cache, returns the slot infos // return empty vector on failure - ubatch_heads prepare(const std::vector & ubatches); + slot_info_vec_t prepare(const std::vector & ubatches); bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo); - // return the cell position where we can insert the ubatch - // return -1 on failure to find a contiguous slot of kv cells - int32_t find_slot(const llama_ubatch & ubatch) const; + // find a slot of kv cells that can hold the ubatch + // if cont == true, then the slot must be continuous + // return empty slot_info on failure + slot_info find_slot(const llama_ubatch & ubatch, bool cont) const; - // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens) - void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch); + // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]] + void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch); // - // set_input API + // input API // + ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const; + ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const; + + void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const; + void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const; + void set_input_kq_mask (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const; void set_input_k_shift (ggml_tensor * dst) const; void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const; @@ -157,8 +188,13 @@ private: // SWA const uint32_t n_swa = 0; + // env: LLAMA_KV_CACHE_DEBUG int debug = 0; + // env: LLAMA_SET_ROWS (temporary) + // ref: https://github.com/ggml-org/llama.cpp/pull/14285 + int supports_set_rows = false; + const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE; std::vector ctxs; @@ -211,8 +247,8 @@ private: class llama_kv_cache_unified_context : public llama_memory_context_i { public: // some shorthands - using ubatch_heads = llama_kv_cache_unified::ubatch_heads; - using defrag_info = llama_kv_cache_unified::defrag_info; + using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t; + using defrag_info = llama_kv_cache_unified::defrag_info; // used for errors llama_kv_cache_unified_context(llama_memory_status status); @@ -231,7 +267,7 @@ public: // used to create a batch procesing context from a batch llama_kv_cache_unified_context( llama_kv_cache_unified * kv, - ubatch_heads heads, + slot_info_vec_t sinfos, std::vector ubatches); virtual ~llama_kv_cache_unified_context(); @@ -257,11 +293,16 @@ public: ggml_tensor * get_v(ggml_context * ctx, int32_t il) const; // store k_cur and v_cur in the cache based on the provided head location - ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const; - ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const; + ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const; + ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const; - void set_input_k_shift(ggml_tensor * dst) const; + ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const; + ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const; + void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const; + void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const; + + void set_input_k_shift (ggml_tensor * dst) const; void set_input_kq_mask (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const; void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const; @@ -283,10 +324,10 @@ private: // batch processing context // - // the index of the next ubatch to process - size_t i_next = 0; + // the index of the cur ubatch to process + size_t i_cur = 0; - ubatch_heads heads; + slot_info_vec_t sinfos; std::vector ubatches; @@ -297,7 +338,4 @@ private: // a heuristic, to avoid attending the full cache if it is not yet utilized // as the cache gets filled, the benefit from this heuristic disappears int32_t n_kv; - - // the beginning of the current slot in which the ubatch will be inserted - int32_t head; }; diff --git a/examples/talk-llama/llama-kv-cells.h b/examples/talk-llama/llama-kv-cells.h index c95d6359..0d0dd316 100644 --- a/examples/talk-llama/llama-kv-cells.h +++ b/examples/talk-llama/llama-kv-cells.h @@ -105,10 +105,30 @@ public: res.resize(n); for (uint32_t j = 0; j < n; ++j) { - res.pos[j] = pos[i + j]; - res.seq[j] = seq[i + j]; + const auto idx = i + j; - assert(shift[i + j] == 0); + res.pos[j] = pos[idx]; + res.seq[j] = seq[idx]; + + assert(shift[idx] == 0); + } + + return res; + } + + // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1]) + llama_kv_cells_unified cp(const std::vector & idxs) const { + llama_kv_cells_unified res; + + res.resize(idxs.size()); + + for (uint32_t j = 0; j < idxs.size(); ++j) { + const auto idx = idxs[j]; + + res.pos[j] = pos[idx]; + res.seq[j] = seq[idx]; + + assert(shift[idx] == 0); } return res; @@ -119,26 +139,58 @@ public: assert(i + other.pos.size() <= pos.size()); for (uint32_t j = 0; j < other.pos.size(); ++j) { - if (pos[i + j] == -1 && other.pos[j] != -1) { + const auto idx = i + j; + + if (pos[idx] == -1 && other.pos[j] != -1) { used.insert(i + j); } - if (pos[i + j] != -1 && other.pos[j] == -1) { + if (pos[idx] != -1 && other.pos[j] == -1) { used.erase(i + j); } - if (pos[i + j] != -1) { + if (pos[idx] != -1) { seq_pos_rm(i + j); } - pos[i + j] = other.pos[j]; - seq[i + j] = other.seq[j]; + pos[idx] = other.pos[j]; + seq[idx] = other.seq[j]; - if (pos[i + j] != -1) { + if (pos[idx] != -1) { seq_pos_add(i + j); } - assert(shift[i + j] == 0); + assert(shift[idx] == 0); + } + } + + // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1]) + void set(const std::vector & idxs, const llama_kv_cells_unified & other) { + assert(idxs.size() == other.pos.size()); + + for (uint32_t j = 0; j < other.pos.size(); ++j) { + const auto idx = idxs[j]; + + if (pos[idx] == -1 && other.pos[j] != -1) { + used.insert(idx); + } + + if (pos[idx] != -1 && other.pos[j] == -1) { + used.erase(idx); + } + + if (pos[idx] != -1) { + seq_pos_rm(idx); + } + + pos[idx] = other.pos[j]; + seq[idx] = other.seq[j]; + + if (pos[idx] != -1) { + seq_pos_add(idx); + } + + assert(shift[idx] == 0); } } diff --git a/examples/talk-llama/llama-memory-hybrid.cpp b/examples/talk-llama/llama-memory-hybrid.cpp index 67cbf955..6cd10db0 100644 --- a/examples/talk-llama/llama-memory-hybrid.cpp +++ b/examples/talk-llama/llama-memory-hybrid.cpp @@ -70,7 +70,7 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba // if all tokens are output, split by sequence ubatch = balloc.split_seq(n_ubatch); } else { - ubatch = balloc.split_equal(n_ubatch); + ubatch = balloc.split_equal(n_ubatch, false); } if (ubatch.n_tokens == 0) { @@ -80,6 +80,11 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba ubatches.push_back(std::move(ubatch)); // NOLINT } + if (balloc.get_n_used() < balloc.get_n_tokens()) { + // failed to find a suitable split + break; + } + // prepare the recurrent batches first if (!mem_recr->prepare(ubatches)) { // TODO: will the recurrent cache be in an undefined context at this point? @@ -195,11 +200,11 @@ llama_memory_hybrid_context::llama_memory_hybrid_context( llama_memory_hybrid_context::llama_memory_hybrid_context( llama_memory_hybrid * mem, - std::vector heads_attn, + slot_info_vec_t sinfos_attn, std::vector ubatches) : ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)), + ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)), ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)), status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) { } diff --git a/examples/talk-llama/llama-memory-hybrid.h b/examples/talk-llama/llama-memory-hybrid.h index f0c2420e..4ac31817 100644 --- a/examples/talk-llama/llama-memory-hybrid.h +++ b/examples/talk-llama/llama-memory-hybrid.h @@ -92,6 +92,8 @@ private: class llama_memory_hybrid_context : public llama_memory_context_i { public: + using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t; + // init failure explicit llama_memory_hybrid_context(llama_memory_status status); @@ -107,7 +109,7 @@ public: // init success llama_memory_hybrid_context( llama_memory_hybrid * mem, - std::vector heads_attn, + slot_info_vec_t sinfos_attn, std::vector ubatches); ~llama_memory_hybrid_context() = default; diff --git a/examples/talk-llama/llama-memory-recurrent.cpp b/examples/talk-llama/llama-memory-recurrent.cpp index 6ed84057..2c1ae670 100644 --- a/examples/talk-llama/llama-memory-recurrent.cpp +++ b/examples/talk-llama/llama-memory-recurrent.cpp @@ -25,9 +25,6 @@ llama_memory_recurrent::llama_memory_recurrent( uint32_t n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) { const int32_t n_layer = hparams.n_layer; - LLAMA_LOG_INFO("%s: mem_size = %u, n_seq_max = %u, type_r = '%s', type_s = '%s', n_layer = %d\n", - __func__, mem_size, n_seq_max, ggml_type_name(type_r), ggml_type_name(type_s), n_layer); - head = 0; size = mem_size; used = 0; @@ -84,7 +81,7 @@ llama_memory_recurrent::llama_memory_recurrent( ggml_context * ctx = ctx_for_buft(buft); if (!ctx) { - throw std::runtime_error("failed to create ggml context for kv cache"); + throw std::runtime_error("failed to create ggml context for rs cache"); } ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size); @@ -102,10 +99,10 @@ llama_memory_recurrent::llama_memory_recurrent( ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (!buf) { - throw std::runtime_error("failed to allocate buffer for kv cache"); + throw std::runtime_error("failed to allocate buffer for rs cache"); } ggml_backend_buffer_clear(buf, 0); - LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); + LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); bufs.emplace_back(buf); } @@ -113,8 +110,8 @@ llama_memory_recurrent::llama_memory_recurrent( const size_t memory_size_r = size_r_bytes(); const size_t memory_size_s = size_s_bytes(); - LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__, - (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f), + LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u seqs), R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__, + (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f), mem_size, n_layer, n_seq_max, ggml_type_name(type_r), (float)memory_size_r / (1024.0f * 1024.0f), ggml_type_name(type_s), (float)memory_size_s / (1024.0f * 1024.0f)); } @@ -374,7 +371,7 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & // if all tokens are output, split by sequence ubatch = balloc.split_seq(n_ubatch); } else { - ubatch = balloc.split_equal(n_ubatch); + ubatch = balloc.split_equal(n_ubatch, false); } if (ubatch.n_tokens == 0) { @@ -384,6 +381,11 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & ubatches.push_back(std::move(ubatch)); // NOLINT } + if (balloc.get_n_used() < balloc.get_n_tokens()) { + // failed to find a suitable split + break; + } + if (!prepare(ubatches)) { break; } diff --git a/examples/talk-llama/llama-model.cpp b/examples/talk-llama/llama-model.cpp index b15bf73c..a322fc39 100644 --- a/examples/talk-llama/llama-model.cpp +++ b/examples/talk-llama/llama-model.cpp @@ -40,17 +40,21 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_190M: return "190M"; case LLM_TYPE_220M: return "220M"; case LLM_TYPE_250M: return "250M"; + case LLM_TYPE_256M: return "256M"; case LLM_TYPE_270M: return "270M"; case LLM_TYPE_335M: return "335M"; + case LLM_TYPE_350M: return "350M"; case LLM_TYPE_410M: return "410M"; case LLM_TYPE_450M: return "450M"; case LLM_TYPE_475M: return "475M"; + case LLM_TYPE_700M: return "700M"; case LLM_TYPE_770M: return "770M"; case LLM_TYPE_780M: return "780M"; case LLM_TYPE_0_3B: return "0.3B"; case LLM_TYPE_0_5B: return "0.5B"; case LLM_TYPE_0_6B: return "0.6B"; case LLM_TYPE_1B: return "1B"; + case LLM_TYPE_1_2B: return "1.2B"; case LLM_TYPE_1_3B: return "1.3B"; case LLM_TYPE_1_4B: return "1.4B"; case LLM_TYPE_1_5B: return "1.5B"; @@ -102,6 +106,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_57B_A14B: return "57B.A14B"; case LLM_TYPE_17B_16E: return "17Bx16E (Scout)"; case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)"; + case LLM_TYPE_A13B: return "A13B"; case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; case LLM_TYPE_E2B: return "E2B"; @@ -208,23 +213,27 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_SSM_CONV: { - // FIXME - ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 12345, w->ne[1], 6789); + const int64_t n_seq_tokens = 512; + const int64_t n_seqs = 3; + ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0] - 1 + n_seq_tokens, w->ne[1], n_seqs); op_tensor = ggml_ssm_conv(ctx, conv_x, w); } break; case GGML_OP_SSM_SCAN: { - // FIXME - const int64_t d_state = w->ne[0]; - const int64_t d_inner = w->ne[1]; + // w is ssm_a, which is used to distinguish Mamba-1 and Mamba-2 + const int64_t d_state = w->ne[0] == 1 ? hparams.ssm_d_state : w->ne[0]; + const int64_t n_head = w->ne[1]; + const int64_t head_dim = hparams.ssm_d_inner / n_head; + const int64_t n_group = hparams.ssm_n_group ? hparams.ssm_n_group : 1; const int64_t n_seq_tokens = 512; - const int64_t n_seqs = 1; - ggml_tensor * s = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, d_inner, n_seqs); - ggml_tensor * x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs); - ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs); - ggml_tensor * B = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs); - ggml_tensor * C = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs); - op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C); + const int64_t n_seqs = 3; + ggml_tensor * s = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, head_dim, n_head, n_seqs); + ggml_tensor * x = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_dim, n_head, n_seq_tokens, n_seqs); + ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_head, n_seq_tokens, n_seqs); + ggml_tensor * B = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs); + ggml_tensor * C = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs); + ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_seqs); + op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C, ids); } break; case GGML_OP_RWKV_WKV6: { @@ -576,6 +585,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { case 22: type = LLM_TYPE_1B; break; case 26: type = LLM_TYPE_3B; break; case 28: type = LLM_TYPE_3B; break; // Llama 3.2 3B + case 30: type = LLM_TYPE_256M; break; // smoldocling 256M // granite uses a vocab with len 49152 case 32: type = n_vocab == 49152 ? LLM_TYPE_3B : (n_vocab < 40000 ? LLM_TYPE_7B : LLM_TYPE_8B); break; case 36: type = LLM_TYPE_8B; break; // granite @@ -1081,6 +1091,58 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_MAMBA2: + { + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 24: + switch (hparams.n_embd) { + case 768: type = LLM_TYPE_SMALL; break; + default: type = LLM_TYPE_UNKNOWN; + } break; + case 48: + switch (hparams.n_embd) { + case 1024: type = LLM_TYPE_MEDIUM; break; + case 1536: type = LLM_TYPE_LARGE; break; + case 2048: type = LLM_TYPE_XL; break; + default: type = LLM_TYPE_UNKNOWN; + } break; + case 64: + switch (hparams.n_embd) { + case 2560: type = LLM_TYPE_3B; break; + case 4096: type = LLM_TYPE_7B; break; + default: type = LLM_TYPE_UNKNOWN; + } break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_JAMBA: + { + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0; + } + + switch (hparams.n_layer) { + // TODO: Jamba layers are a bit heterogenous, so naming this is hard. + case 12: // 900M 8x???M + case 32: // 51B 16x?B + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_XVERSE: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -1447,6 +1509,11 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale); ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale); + // Granite uses rope_finetuned as a switch for rope, so default to true + bool rope_finetuned = true; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; + switch (hparams.n_layer) { case 32: type = LLM_TYPE_3B; break; case 40: type = LLM_TYPE_3B; break; @@ -1454,6 +1521,40 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } + // For Granite MoE Shared + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false); + } break; + case LLM_ARCH_GRANITE_HYBRID: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale, /* required */ false); + ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale, /* required */ false); + ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, /* required */ false); + ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale, /* required */ false); + + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Granite uses rope_finetuned as a switch for rope, so default to true + bool rope_finetuned = true; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; + + // A layer is recurrent IFF the n_head_kv value is set to 0 + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0; + } + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + // TODO: Add llm type label (not sure this is useful) + default: type = LLM_TYPE_UNKNOWN; + } + // For Granite MoE Shared ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false); } break; @@ -1513,6 +1614,72 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_FALCON_H1: + { + // Common parameters + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + // SSM parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true); + + switch (hparams.n_layer) { + case 36: + type = LLM_TYPE_0_5B; break; + case 24: + type = LLM_TYPE_1_5B; break; + case 66: + type = LLM_TYPE_1B; break; + case 32: + type = LLM_TYPE_3B; break; + case 44: + type = LLM_TYPE_7B; break; + case 72: + type = LLM_TYPE_34B; break; + default: + type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_HUNYUAN_MOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); + + switch (hparams.n_layer) { + case 32: type = LLM_TYPE_A13B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_SMOLLM3: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + hparams.n_no_rope_layer_step = 4; + + switch (hparams.n_layer) { + case 36: type = LLM_TYPE_3B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_LFM2: + { + ml.get_key(LLM_KV_SHORTCONV_L_CACHE, hparams.n_shortconv_l_cache); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + for (uint32_t il = 0; il < hparams.n_layer; ++il) { + hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0; + } + switch (hparams.n_embd) { + case 1024: type = LLM_TYPE_350M; break; + case 1536: type = LLM_TYPE_700M; break; + case 2048: type = LLM_TYPE_1_2B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -3124,6 +3291,228 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0); } } break; + case LLM_ARCH_MAMBA2: + { + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_head = hparams.ssm_dt_rank; + const int64_t n_group = hparams.ssm_n_group; + const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head; + + // only an expansion factor of 2 is supported for now + GGML_ASSERT(2 * n_embd == d_inner); + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed, duplicated to allow offloading + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + // norm + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0); + + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0); + layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0); + + layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_head}, 0); + + // no "weight" suffix for these + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0); + layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_head}, 0); + + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0); + + // out_proj + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0); + } + } break; + case LLM_ARCH_JAMBA: + { + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + + // only an expansion factor of 2 is supported for now + GGML_ASSERT(2 * n_embd == d_inner); + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed, duplicated to allow offloading + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + const int64_t n_head_kv = hparams.n_head_kv(i); + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(i); + + auto & layer = layers[i]; + + // norm + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + if (n_head_kv == 0) { + // Mamba layer + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}, 0); + + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}, 0); + layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}, 0); + + layer.ssm_x = create_tensor(tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}, 0); + + layer.ssm_dt_norm = create_tensor(tn(LLM_TENSOR_SSM_DT_NORM, "weight", i), {dt_rank}, 0); + + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}, 0); + layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0); + + layer.ssm_b_norm = create_tensor(tn(LLM_TENSOR_SSM_B_NORM, "weight", i), {d_state}, 0); + layer.ssm_c_norm = create_tensor(tn(LLM_TENSOR_SSM_C_NORM, "weight", i), {d_state}, 0); + + // no "weight" suffix for these + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0); + layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0); + + // out_proj + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0); + } else { + // Attention layers + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); + } + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); + + if (layer.ffn_gate_inp) { + // MoE + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + } else { + // FFN (no MoE) + 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_GRANITE_HYBRID: + { + // mamba2 Mixer SSM params + // NOTE: int64_t for tensor dimensions + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_ssm_head = hparams.ssm_dt_rank; + const int64_t n_group = hparams.ssm_n_group; + const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_ssm_head; + + // only an expansion factor of 2 is supported for now + GGML_ASSERT(2 * n_embd == d_inner); + + // embeddings + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed, duplicated to allow offloading + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + // norm + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + if (hparams.is_recurrent(i)) { + // ssm layers + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0); + + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0); + layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED); + + layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0); + + // no "weight" suffix for these + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0); + layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0); + + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0); + + // out_proj + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0); + } else { + // attention layers (with optional bias) + const int64_t n_head_i = hparams.n_head(i); + const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i); + const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i); + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head_i}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa_i}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa_i}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED); + layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + } + + // feed forward (w/ optional biases) + if (n_expert > 0) { + // MoE FFN + 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}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0); + } + } else { + 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}, TENSOR_NOT_REQUIRED | (i != 0 ? 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); + layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED); + layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED); + } + } + } break; case LLM_ARCH_XVERSE: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -4391,6 +4780,182 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; + case LLM_ARCH_FALCON_H1: + { + // Common + const int64_t hidden_size = hparams.n_embd; // hidden_size + + // mamba2 Mixer SSM params + const int64_t ssm_conv_kernel_size = hparams.ssm_d_conv; // ssm_conv_kernel_size + const int64_t ssm_n_groups = hparams.ssm_n_group; // ssm_n_groups + const int64_t ssm_state_size = hparams.ssm_d_state; // ssm_state_size + const int64_t ssm_intermediate_size = hparams.ssm_d_inner; // TODO expand + const int64_t ssm_num_heads = hparams.ssm_dt_rank; // ssm_num_heads + const int64_t ssm_conv_dim = ssm_intermediate_size + 2 * ssm_n_groups * ssm_state_size; + const int64_t ssm_projection_size = ssm_intermediate_size + ssm_conv_dim + ssm_num_heads; + + // attn params + const int64_t attn_num_attention_head = hparams.n_head(0); // rename to: attn_num_attention_head + const int64_t attn_num_key_value_head = hparams.n_head_kv(0); + + // ffn params + const int64_t ffn_intermediate_size = hparams.n_ff(0); + + // embeddings + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hidden_size, n_vocab}, 0); + + // output + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hidden_size, n_vocab}, TENSOR_NOT_REQUIRED); + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {hidden_size}, 0); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hidden_size, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + /*SSM LAYERS*/ + // ssm in + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {hidden_size, ssm_projection_size}, 0); + // ssm 1d conv + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {ssm_conv_kernel_size, ssm_conv_dim}, 0); + layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {ssm_conv_dim}, TENSOR_NOT_REQUIRED); + // ssm_dt + layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {ssm_num_heads}, 0); + // no "weight" suffix for these + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, ssm_num_heads}, 0); + layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, ssm_num_heads}, 0); + // ssm_norm + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {ssm_intermediate_size / ssm_n_groups, ssm_n_groups}, TENSOR_NOT_REQUIRED); + // out_proj + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {ssm_intermediate_size, hidden_size}, 0); + + /*ATTENTION LAYERS*/ + // attention layers (with optional bias) + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {hidden_size, n_embd_head_k * attn_num_attention_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {hidden_size, attn_num_key_value_head * n_embd_head_k}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {hidden_size, attn_num_key_value_head * n_embd_head_v}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * attn_num_attention_head, hidden_size}, 0); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {attn_num_key_value_head * n_embd_head_k}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {attn_num_key_value_head * n_embd_head_v}, TENSOR_NOT_REQUIRED); + layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED); + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {hidden_size}, 0); + + + // feed forward (w/ optional biases) + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, i), {hidden_size}, 0); + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {hidden_size, ffn_intermediate_size}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { ffn_intermediate_size, hidden_size}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {hidden_size, ffn_intermediate_size}, 0); + + layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {ffn_intermediate_size}, TENSOR_NOT_REQUIRED); + layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED); + layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {ffn_intermediate_size}, TENSOR_NOT_REQUIRED); + } + } break; + case LLM_ARCH_HUNYUAN_MOE: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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}, 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}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + 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.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0); + } + } break; + case LLM_ARCH_SMOLLM3: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 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}, 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}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + 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.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_LFM2: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + // ffn is same for transformer and conv layers + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 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); + + // for operator_norm + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + if (!hparams.is_recurrent(i)) { + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, hparams.n_embd_k_gqa(i)}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, hparams.n_embd_v_gqa(i)}, 0); + + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); + } else { + layer.shortconv.conv = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV, "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0); + layer.shortconv.in_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ, "weight", i), {n_embd, 3 * n_embd}, 0); + layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0); + } + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -4630,12 +5195,6 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); - LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); - LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); - LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); - LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); - LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms = %d\n", __func__, hparams.ssm_dt_b_c_rms); - if (!classifier_labels.empty()) { LLAMA_LOG_INFO("%s: n_cls_out = %u\n", __func__, hparams.n_cls_out); @@ -4646,6 +5205,19 @@ void llama_model::print_info() const { } } + if (arch == LLM_ARCH_MAMBA || + arch == LLM_ARCH_MAMBA2 || + arch == LLM_ARCH_JAMBA || + arch == LLM_ARCH_FALCON_H1 || + arch == LLM_ARCH_GRANITE_HYBRID) { + LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); + LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); + LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); + LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + LLAMA_LOG_INFO("%s: ssm_n_group = %u\n", __func__, hparams.ssm_n_group); + LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms = %d\n", __func__, hparams.ssm_dt_b_c_rms); + } + LLAMA_LOG_INFO("%s: model type = %s\n", __func__, type_name().c_str()); if (pimpl->n_elements >= 1e12) { LLAMA_LOG_INFO("%s: model params = %.2f T\n", __func__, pimpl->n_elements*1e-12); @@ -4692,7 +5264,8 @@ void llama_model::print_info() const { if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || - arch == LLM_ARCH_GRANITE_MOE) { + arch == LLM_ARCH_GRANITE_MOE || + arch == LLM_ARCH_GRANITE_HYBRID) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); @@ -5582,12 +6155,10 @@ struct llm_build_falcon : public llm_graph_context { cur = build_lora_mm(model.layers[il].wqkv, cur); cb(cur, "wqkv", il); - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode @@ -5864,12 +6435,10 @@ struct llm_build_dbrx : public llm_graph_context { cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); cb(cur, "wqkv_clamped", il); - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -6380,12 +6949,10 @@ struct llm_build_neo_bert : public llm_graph_context { cur = build_lora_mm(model.layers[il].wqkv, cur); cb(cur, "wqkv", il); - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); // RoPE @@ -6615,8 +7182,8 @@ struct llm_build_mpt : public llm_graph_context { cb(cur, "wqkv_clamped", il); } - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); cb(Qcur, "Qcur", il); @@ -6636,6 +7203,12 @@ struct llm_build_mpt : public llm_graph_context { model.layers[il].attn_k_norm_b, LLM_NORM, il); cb(Kcur, "Kcur", il); + } else { + Qcur = ggml_cont(ctx0, Qcur); + cb(Qcur, "Qcur", il); + + Kcur = ggml_cont(ctx0, Kcur); + cb(Kcur, "Kcur", il); } Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); @@ -6890,12 +7463,10 @@ struct llm_build_qwen : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); // using mode = 2 for neox mode @@ -7660,21 +8231,21 @@ struct llm_build_phi2 : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); } else { Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -7798,21 +8369,21 @@ struct llm_build_phi3 : public llm_graph_context { cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); cb(cur, "wqkv", il); - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa))); } else { Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -8168,12 +8739,10 @@ struct llm_build_codeshell : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -8589,8 +9158,6 @@ struct llm_build_minicpm3 : public llm_graph_context { ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); cb(k_pe, "k_pe", il); - // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont - kv_compressed = ggml_cont(ctx0, kv_compressed); kv_compressed = build_norm(kv_compressed, model.layers[il].attn_kv_a_norm, NULL, LLM_NORM_RMS, il); @@ -8617,12 +9184,6 @@ struct llm_build_minicpm3 : public llm_graph_context { v_states = ggml_cont(ctx0, v_states); cb(v_states, "v_states", il); - v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, - ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), - 0); - cb(v_states, "v_states", il); - - q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this q_pe = ggml_rope_ext( ctx0, q_pe, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, @@ -8631,7 +9192,6 @@ struct llm_build_minicpm3 : public llm_graph_context { cb(q_pe, "q_pe", il); // shared RoPE key - k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this k_pe = ggml_rope_ext( ctx0, k_pe, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, @@ -9111,8 +9671,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context { const int n_layer_sparsity = 10; // number of layers using activation sparsity const float f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95) - ggml_tensor * one; // containing single element 1.0f - llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model), @@ -9124,14 +9682,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context { ggml_tensor * cur; ggml_tensor * inpL; - // TODO: remove this when ggml_scale_add is implemented - one = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - { - auto inp = std::make_unique(); - inp->one = one; - res->add_input(std::move(inp)); - } - inpL = build_inp_embd(model.tok_embd); // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) @@ -9521,7 +10071,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context { cb(innovation, "innovation", il); ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens] - all_coefs = ggml_add(ctx0, all_coefs, one); + all_coefs = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f); // + 1.0 cb(all_coefs, "all_coefs", il); all_coefs = ggml_cont(ctx0, ggml_transpose(ctx0, all_coefs)); // [n_tokens, n_altup] all_coefs = ggml_reshape_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup] @@ -9664,81 +10214,32 @@ struct llm_build_starcoder2 : public llm_graph_context { } }; -struct llm_build_mamba : public llm_graph_context { - const llama_model & model; +struct llm_graph_context_mamba : public llm_graph_context { + llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {} - llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) { - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - - auto * rs_inp = build_rs_inp(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - cur = build_mamba_layer(rs_inp, gf, cur, ubatch, il); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // residual - cur = ggml_add(ctx0, cur, inpL); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - // final rmsnorm - cur = build_norm(inpL, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - // TODO: split ggml_tensor * build_mamba_layer( llm_graph_input_rs * inp, ggml_cgraph * gf, ggml_tensor * cur, + const llama_model & model, const llama_ubatch & ubatch, - int il) const { - const auto * mctx_cur = static_cast(mctx); + int il) { + + const auto * mctx_cur = inp->mctx; const auto kv_head = mctx_cur->get_head(); + const auto & layer = model.layers[il]; + const int64_t d_conv = hparams.ssm_d_conv; const int64_t d_inner = hparams.ssm_d_inner; const int64_t d_state = hparams.ssm_d_state; const int64_t dt_rank = hparams.ssm_dt_rank; + const int64_t n_head = d_inner; + const int64_t head_dim = 1; const int64_t n_seqs = ubatch.n_seqs; // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; - // Use the same RMS norm as the final layer norm - const float norm_rms_eps = hparams.f_norm_rms_eps; const int64_t n_seq_tokens = ubatch.n_seq_tokens; @@ -9749,21 +10250,14 @@ struct llm_build_mamba : public llm_graph_context { ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - // (ab)using the KV cache to store the states - ggml_tensor * conv = build_rs( - inp, gf, conv_states_all, - hparams.n_embd_r(), n_seqs); + ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), n_seqs); conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs); - ggml_tensor * ssm = build_rs( - inp, gf, ssm_states_all, - hparams.n_embd_s(), n_seqs); - ssm = ggml_reshape_3d(ctx0, ssm, d_state, d_inner, n_seqs); // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} - ggml_tensor * xz = build_lora_mm(model.layers[il].ssm_in, cur); + ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur); // split the above in two // => {d_inner, n_seq_tokens, n_seqs} ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); @@ -9792,10 +10286,10 @@ struct llm_build_mamba : public llm_graph_context { // then permute away the ne[0] dimension, // and then you're left with the resulting x tensor. // For simultaneous sequences, all sequences need to have the same length. - x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d); // bias - x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b); + x = ggml_add(ctx0, x, layer.ssm_conv1d_b); x = ggml_silu(ctx0, x); } @@ -9803,41 +10297,186 @@ struct llm_build_mamba : public llm_graph_context { // ssm { // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} - ggml_tensor * x_db = build_lora_mm(model.layers[il].ssm_x, x); + ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x); // split ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); - ggml_tensor * B = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); - ggml_tensor * C = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); + ggml_tensor * B = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); + ggml_tensor * C = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); - // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers - if (ssm_dt_b_c_rms) { - dt = ggml_rms_norm(ctx0, dt, norm_rms_eps); - B = ggml_rms_norm(ctx0, B, norm_rms_eps); - C = ggml_rms_norm(ctx0, C, norm_rms_eps); + // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers + if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) { + dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il); + B = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il); + C = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il); } // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} - dt = build_lora_mm(model.layers[il].ssm_dt, dt); - dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); + dt = build_lora_mm(layer.ssm_dt, dt); + dt = ggml_add(ctx0, dt, layer.ssm_dt_b); - // Custom operator to optimize the parallel associative scan - // as described in the Annex D of the Mamba paper. - // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} - ggml_tensor * y_ssm = ggml_ssm_scan(ctx0, ssm, x, dt, model.layers[il].ssm_a, B, C); + cur = x; + x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs); + + ggml_tensor * A = layer.ssm_a; + + // use the states and the indices provided by build_recurrent_state + // (this is necessary in order to properly use the states before they are overwritten, + // while avoiding to make unnecessary copies of the states) + auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { + ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); + }; + + ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); // store last states ggml_build_forward_expand(gf, ggml_cpy(ctx0, - ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]), + ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]), ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); - ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0); + ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0); + + // TODO: skip computing output earlier for unused tokens + + y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d)); + y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = build_lora_mm(layer.ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + + return cur; + } + + ggml_tensor * build_mamba2_layer( + llm_graph_input_rs * inp, + ggml_cgraph * gf, + ggml_tensor * cur, + const llama_model & model, + const llama_ubatch & ubatch, + int il) const { + + const auto * mctx_cur = inp->mctx; + + const auto kv_head = mctx_cur->get_head(); + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_head = hparams.ssm_dt_rank; + const int64_t head_dim = d_inner / n_head; + const int64_t n_group = hparams.ssm_n_group; + const int64_t n_seqs = ubatch.n_seqs; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads + + // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs} + ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur); + + // split the above in three + ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*zxBCdt->nb[0], zxBCdt->nb[1], zxBCdt->nb[2], 0); + ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2*n_group*d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], d_inner*ggml_element_size(zxBCdt)); + ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], (2*d_inner + 2*n_group*d_state)*ggml_element_size(zxBCdt)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); + + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, + (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs), + kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + + // bias + xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b); + + xBC = ggml_silu(ctx0, xBC); + } + + // ssm + { + // These correspond to V K Q in SSM/attention duality + ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*xBC->nb[0], xBC->nb[1], xBC->nb[2], 0); + ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], d_inner*ggml_element_size(xBC)); + ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC)); + + // {n_head, n_seq_tokens, n_seqs} + dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b); + + ggml_tensor * A = model.layers[il].ssm_a; + + // use the states and the indices provided by build_recurrent_state + // (this is necessary in order to properly use the states before they are overwritten, + // while avoiding to make unnecessary copies of the states) + auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { + ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); + + // TODO: use semistructured matrices to implement state-space duality + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); + }; + + ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); + + // store last states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, ggml_nelements(x)*x->nb[0]), + ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); + + ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head*x->nb[1], n_seq_tokens*n_head*x->nb[1], 0); // TODO: skip computing output earlier for unused tokens - // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); - y = ggml_mul(ctx0, y, ggml_silu(ctx0, ggml_cont(ctx0, z))); + y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); + + // grouped RMS norm + if (model.layers[il].ssm_norm) { + y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs); + y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il); + } + + y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs); // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} cur = build_lora_mm(model.layers[il].ssm_out, y); @@ -9845,12 +10484,178 @@ struct llm_build_mamba : public llm_graph_context { // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); - //cb(cur, "mamba_out", il); + cb(cur, "mamba_out", il); return cur; } }; +struct llm_build_mamba : public llm_graph_context_mamba { + llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context_mamba(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + auto * rs_inp = build_rs_inp(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (model.arch == LLM_ARCH_MAMBA2) { + cur = build_mamba2_layer(rs_inp, gf, cur, model, ubatch, il); + } else { + cur = build_mamba_layer(rs_inp, gf, cur, model, ubatch, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // residual + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + +}; + +struct llm_build_jamba : public llm_graph_context_mamba { + llm_build_jamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + auto * inp_hybrid = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (n_head_kv == 0) { + cur = build_mamba_layer(inp_hybrid->get_recr(), gf, cur, model, ubatch, il); + } else { + // Attention + + struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // No RoPE :) + cur = build_attn(inp_hybrid->get_attn(), gf, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // residual + struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur); + cb(cur, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + // FFN + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + + // residual + cur = ggml_add(ctx0, ffn_inp, cur); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + struct llm_build_command_r : public llm_graph_context { llm_build_command_r(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; @@ -10557,10 +11362,10 @@ struct llm_build_openelm : public llm_graph_context { cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0); cb(Qcur, "Qcur", il); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head); cb(Kcur, "Kcur", il); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); @@ -10682,12 +11487,10 @@ struct llm_build_gptneox : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); - ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -11932,6 +12735,8 @@ struct llm_build_chatglm : public llm_graph_context { if (model.layers[il].bv) { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } else { cur = build_lora_mm(model.layers[il].wqkv, cur); cb(cur, "wqkv", il); @@ -11939,13 +12744,11 @@ struct llm_build_chatglm : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); @@ -12066,6 +12869,8 @@ struct llm_build_glm4 : public llm_graph_context { if (model.layers[il].bv) { Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } else { cur = build_lora_mm(model.layers[il].wqkv, cur); cb(cur, "wqkv", il); @@ -12073,13 +12878,11 @@ struct llm_build_glm4 : public llm_graph_context { cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cb(cur, "bqkv", il); } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext( @@ -13178,13 +13981,11 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { } }; - struct llm_build_granite : public llm_graph_context { llm_build_granite( const llama_model & model, const llm_graph_params & params, - ggml_cgraph * gf, - const bool use_rope = true) + ggml_cgraph * gf) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; @@ -13199,14 +14000,12 @@ struct llm_build_granite : public llm_graph_context { // inp_pos - built only if rope enabled ggml_tensor * inp_pos = nullptr; - if (use_rope) { + if (hparams.rope_finetuned) { inp_pos = build_inp_pos(); } auto * inp_attn = build_attn_inp_kv_unified(); - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - ggml_tensor * inp_out_ids = build_inp_out_ids(); for (int il = 0; il < n_layer; ++il) { @@ -13219,128 +14018,17 @@ struct llm_build_granite : public llm_graph_context { cb(cur, "attn_norm", il); // self-attention - { - // compute Q and K and (optionally) RoPE them - ggml_tensor * Qcur = build_lora_mm(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); - } - - ggml_tensor * Kcur = build_lora_mm(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); - } - - ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (use_rope) { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, 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); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, gf, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } + cur = build_attention_layer( + gf, cur, inp_pos, inp_attn, + model, n_embd_head, il); if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - // For Granite architectures - scale residual - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(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, il); - cb(cur, "ffn_out", il); - - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // For Granite MoE Shared - if (hparams.n_ff_shexp > 0) { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } else { - cur = moe_out; - } - } - - // For Granite architectures - scale residual - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); // input for next layer inpL = cur; @@ -13365,6 +14053,370 @@ struct llm_build_granite : public llm_graph_context { ggml_build_forward_expand(gf, cur); } + + ggml_tensor * build_attention_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv_unified * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(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); + } + + ggml_tensor * Kcur = build_lora_mm(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); + } + + ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, 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); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; + } + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; + } +}; + +struct llm_build_granite_hybrid : public llm_graph_context_mamba { + + llm_build_granite_hybrid( + const llama_model & model, + const llm_graph_params & params, + ggml_cgraph * gf) : + llm_graph_context_mamba(params) { + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + // Positional embeddings populated if rope enabled + ggml_tensor * inp_pos = nullptr; + if (hparams.rope_finetuned) { + inp_pos = build_inp_pos(); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (hparams.is_recurrent(il)) { + // ssm layer // + cur = build_mamba2_layer(inp->get_recr(), gf, cur, model, ubatch, il); + } else { + // attention layer // + cur = build_attention_layer( + gf, cur, inp_pos, inp->get_attn(), model, + n_embd_head, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architectures - scale logits + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + ggml_tensor * build_attention_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv_unified * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(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); + } + + ggml_tensor * Kcur = build_lora_mm(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); + } + + ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, 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); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; + } + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; + } }; // ref: https://github.com/facebookresearch/chameleon @@ -14298,6 +15350,132 @@ struct llm_build_ernie4_5 : public llm_graph_context { } }; +struct llm_build_falcon_h1 : public llm_graph_context_mamba { + llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // Build the inputs in the recurrent & kv cache + auto * inp = build_inp_mem_hybrid(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur-post-rope", il); + cb(Kcur, "Kcur-post-rope", il); + cb(Vcur, "Vcur-post-rope", il); + + ggml_tensor * attn_out = build_attn(inp->get_attn(), gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(attn_out, "attn_out", il); + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + // Mamba2 layer + cb(cur, "ssm_in", il); + + ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), gf, cur, model, ubatch, il); + cb(ssm_out, "ssm_out", il); + + // // Aggregation + cur = ggml_add(ctx0, attn_out, ssm_out); + inpSA = ggml_add(ctx0, cur, inpSA); + cb(cur, "layer_out", il); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = inpSA; + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, inpSA); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + struct llm_build_arcee : public llm_graph_context { llm_build_arcee(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; @@ -14433,6 +15611,461 @@ struct llm_build_arcee : public llm_graph_context { } }; +struct llm_build_hunyuan_moe : public llm_graph_context { + llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + 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); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(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); + } + + ggml_tensor * Kcur = build_lora_mm(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); + } + + ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, 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); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, nullptr, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_norm", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, nullptr, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_norm", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network (non-MoE) + ggml_tensor * cur_mlp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_mlp, "ffn_mlp", il); + + // MoE branch + ggml_tensor * cur_moe = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, + true, // norm_topk_prob + false, + 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur_moe, "ffn_moe_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp); + cb(ffn_out, "ffn_out", il); + + cur = ggml_add(ctx0, ffn_out, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_smollm3 : public llm_graph_context { + llm_build_smollm3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + 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); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(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); + } + + ggml_tensor * Kcur = build_lora_mm(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); + } + + ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (use_rope) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_lfm2 : public llm_graph_context { + const llama_model & model; + + llm_build_lfm2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) { + + ggml_tensor * cur = build_inp_embd(model.tok_embd); + cb(cur, "model.embed_tokens", -1); + + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_hybrid = build_inp_mem_hybrid(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + auto * prev_cur = cur; + cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "model.layers.{}.operator_norm", il); + + cur = hparams.is_recurrent(il) ? + build_shortconv_block(gf, cur, inp_hybrid->get_recr(), il) : + build_attn_block(gf, cur, inp_pos, inp_hybrid->get_attn(), il) ; + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids); + } + + cur = ggml_add(ctx0, prev_cur, cur); + cur = ggml_add(ctx0, cur, build_feed_forward(cur, il)); + } + + cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "model.embedding_norm", -1); + res->t_embd = cur; + + // lm_head is tied with embeddings + cur = build_lora_mm(model.tok_embd, cur); + cb(cur, "lm_head", -1); + + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + ggml_tensor * build_feed_forward(ggml_tensor * cur, + int il) const { + cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "model.layers.{}.ffn_norm", il); + + GGML_ASSERT(!model.layers[il].ffn_up_b); + GGML_ASSERT(!model.layers[il].ffn_gate_b); + GGML_ASSERT(!model.layers[il].ffn_down_b); + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "model.layers.{}.feed_forward.w2", il); + + return cur; + } + + ggml_tensor * build_attn_block(ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv_unified * inp_attn, + int il) const { + GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il)); + auto const n_embd_head = hparams.n_embd_head_v; + auto const n_head_kv = hparams.n_head_kv(il); + + auto * q = build_lora_mm(model.layers[il].wq, cur); + cb(q, "model.layers.{}.self_attn.q_proj", il); + auto * k = build_lora_mm(model.layers[il].wk, cur); + cb(k, "model.layers.{}.self_attn.k_proj", il); + auto * v = build_lora_mm(model.layers[il].wv, cur); + cb(v, "model.layers.{}.self_attn.v_proj", il); + + q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens); + k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens); + v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens); + + // qk norm + q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(q, "model.layers.{}.self_attn.q_layernorm", il); + k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(k, "model.layers.{}.self_attn.k_layernorm", il); + + // RoPE + q = ggml_rope_ext( + ctx0, q, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + k = ggml_rope_ext( + ctx0, k, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, + q, k, v, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + cb(cur, "model.layers.{}.self_attn.out_proj", il); + + return cur; + } + + ggml_tensor * build_shortconv_block(ggml_cgraph * gf, + ggml_tensor * cur, + llm_graph_input_rs * inp_recr, + int il) { + const auto * mctx_cur = static_cast(mctx)->get_recr(); + + auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur); + cb(bcx, "model.layers.{}.conv.in_proj", il); + + constexpr auto n_chunks = 3; + GGML_ASSERT(bcx->ne[0] % n_chunks == 0); + auto const chunk_size = bcx->ne[0] / n_chunks; + auto * b = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 0 * chunk_size * ggml_element_size(bcx)); + auto * c = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 1 * chunk_size * ggml_element_size(bcx)); + auto * x = ggml_view_2d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->nb[1], 2 * chunk_size * ggml_element_size(bcx)); + + auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x)); + + // read conv state directly, with build_rs generation is slower + ggml_tensor * conv_state = mctx_cur->get_r_l(il); + const int64_t n_seqs = ubatch.n_seqs; + ggml_tensor * conv = build_rs(inp_recr, gf, conv_state, hparams.n_embd_r(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv_state, hparams.n_shortconv_l_cache - 1, hparams.n_embd, n_seqs); + + bx = ggml_concat(ctx0, conv, bx, 0); + GGML_ASSERT(bx->ne[0] > conv->ne[0]); + + auto * new_conv = ggml_view_2d(ctx0, bx, conv->ne[0], bx->ne[1], bx->nb[1], (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx)); + GGML_ASSERT(ggml_are_same_shape(conv, new_conv)); + + // write conv state + ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv, conv_state)); + + auto * conv_kernel = model.layers[il].shortconv.conv; + GGML_ASSERT(hparams.n_shortconv_l_cache > 0); + + // construct ssm_conv op + ggml_tensor * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel); + cb(conv_out, "model.layers.{}.conv.conv", il); + + auto * y = ggml_mul(ctx0, c, conv_out); + + y = build_lora_mm(model.layers[il].shortconv.out_proj, y); + cb(y, "model.layers.{}.conv.out_proj", il); + + return y; + } +}; + llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const { llama_memory_i * res; @@ -14479,7 +16112,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* recurrent_type_v */ GGML_TYPE_F32, /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max), /* n_seq_max */ cparams.n_seq_max, - /* offload */ cparams.offload_kqv); + /* offload */ cparams.offload_kqv, + /* filter_attn */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr, + /* filter_recr */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr); } else { const auto padding = llama_kv_cache_unified::get_padding(cparams); @@ -14668,9 +16303,14 @@ llm_graph_result_ptr llama_model::build_graph( llm = std::make_unique(*this, params, gf); } break; case LLM_ARCH_MAMBA: + case LLM_ARCH_MAMBA2: { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_JAMBA: + { + llm = std::make_unique(*this, params, gf); + } break; case LLM_ARCH_XVERSE: { llm = std::make_unique(*this, params, gf); @@ -14784,6 +16424,10 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_GRANITE_HYBRID: + { + llm = std::make_unique(*this, params, gf); + } break; case LLM_ARCH_CHAMELEON: { llm = std::make_unique(*this, params, gf); @@ -14812,6 +16456,22 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_HUNYUAN_MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_SMOLLM3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_FALCON_H1: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_LFM2: + { + llm = std::make_unique(*this, params, gf); + } break; default: GGML_ABORT("fatal error"); } @@ -14928,6 +16588,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_REFACT: case LLM_ARCH_BLOOM: case LLM_ARCH_MAMBA: + case LLM_ARCH_MAMBA2: + case LLM_ARCH_JAMBA: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_T5: case LLM_ARCH_T5ENCODER: @@ -14959,15 +16621,18 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_GLM4: case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: + case LLM_ARCH_GRANITE_HYBRID: case LLM_ARCH_CHAMELEON: case LLM_ARCH_BAILINGMOE: case LLM_ARCH_NEO_BERT: + case LLM_ARCH_SMOLLM3: case LLM_ARCH_ARCEE: case LLM_ARCH_ERNIE4_5: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 case LLM_ARCH_FALCON: + case LLM_ARCH_FALCON_H1: case LLM_ARCH_GROK: case LLM_ARCH_DBRX: case LLM_ARCH_BERT: @@ -14999,6 +16664,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_EXAONE: case LLM_ARCH_MINICPM3: case LLM_ARCH_DOTS1: + case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_LFM2: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: diff --git a/examples/talk-llama/llama-model.h b/examples/talk-llama/llama-model.h index a958c599..027a7f0c 100644 --- a/examples/talk-llama/llama-model.h +++ b/examples/talk-llama/llama-model.h @@ -32,17 +32,21 @@ enum llm_type { LLM_TYPE_190M, LLM_TYPE_220M, LLM_TYPE_250M, + LLM_TYPE_256M, LLM_TYPE_270M, LLM_TYPE_335M, + LLM_TYPE_350M, LLM_TYPE_410M, LLM_TYPE_450M, LLM_TYPE_475M, + LLM_TYPE_700M, LLM_TYPE_770M, LLM_TYPE_780M, LLM_TYPE_0_3B, LLM_TYPE_0_5B, LLM_TYPE_0_6B, LLM_TYPE_1B, + LLM_TYPE_1_2B, LLM_TYPE_1_3B, LLM_TYPE_1_4B, LLM_TYPE_1_5B, @@ -94,6 +98,7 @@ enum llm_type { LLM_TYPE_57B_A14B, LLM_TYPE_17B_16E, // llama4 Scout LLM_TYPE_17B_128E, // llama4 Maverick + LLM_TYPE_A13B, LLM_TYPE_30B_A3B, LLM_TYPE_235B_A22B, LLM_TYPE_E2B, @@ -153,6 +158,12 @@ struct llama_layer_convnext { struct ggml_tensor * gamma = nullptr; }; +struct llama_layer_shortconv { + struct ggml_tensor * in_proj = nullptr; + struct ggml_tensor * conv = nullptr; + struct ggml_tensor * out_proj = nullptr; +}; + struct llama_layer { // normalization struct ggml_tensor * attn_norm = nullptr; @@ -172,6 +183,10 @@ struct llama_layer { struct ggml_tensor * ffn_sub_norm = nullptr; struct ggml_tensor * attn_norm_cross = nullptr; struct ggml_tensor * attn_norm_enc = nullptr; + struct ggml_tensor * ssm_norm = nullptr; + struct ggml_tensor * ssm_dt_norm = nullptr; + struct ggml_tensor * ssm_b_norm = nullptr; + struct ggml_tensor * ssm_c_norm = nullptr; // attention struct ggml_tensor * wq = nullptr; @@ -335,6 +350,8 @@ struct llama_layer { struct llama_layer_posnet posnet; struct llama_layer_convnext convnext; + + struct llama_layer_shortconv shortconv; }; struct llama_model { diff --git a/examples/talk-llama/llama-quant.cpp b/examples/talk-llama/llama-quant.cpp index f4b5713d..4dbd1e30 100644 --- a/examples/talk-llama/llama-quant.cpp +++ b/examples/talk-llama/llama-quant.cpp @@ -844,6 +844,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: // do not quantize Mamba's small yet 2D weights // NOTE: can't use LLM_TN here because the layer number is not known quantize &= name.find("ssm_conv1d.weight") == std::string::npos; + quantize &= name.find("shortconv.conv.weight") == std::string::npos; // do not quantize RWKV's small yet 2D weights quantize &= name.find("time_mix_first.weight") == std::string::npos; diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp index 5c9eb875..e0e578d6 100644 --- a/examples/talk-llama/llama-vocab.cpp +++ b/examples/talk-llama/llama-vocab.cpp @@ -351,6 +351,7 @@ struct llm_tokenizer_bpe : llm_tokenizer { break; case LLAMA_VOCAB_PRE_TYPE_STABLELM2: case LLAMA_VOCAB_PRE_TYPE_QWEN2: + case LLAMA_VOCAB_PRE_TYPE_HUNYUAN: regex_exprs = { // original regex from tokenizer.json // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" @@ -1522,7 +1523,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "llama-v3" || tokenizer_pre == "llama-bpe"|| tokenizer_pre == "falcon3" || - tokenizer_pre == "pixtral") { + tokenizer_pre == "falcon-h1" || + tokenizer_pre == "pixtral" || + tokenizer_pre == "midm-2.0" || + tokenizer_pre == "lfm2") { pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3; ignore_merges = true; add_bos = true; @@ -1554,7 +1558,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "jina-de" || tokenizer_pre == "gigachat" || tokenizer_pre == "jina-v2-es" || - tokenizer_pre == "jina-v2-de") { + tokenizer_pre == "jina-v2-de" || + tokenizer_pre == "a.x-4.0") { pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; } else if ( tokenizer_pre == "jina-v1-en" || @@ -1656,6 +1661,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "seed-coder") { pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER; clean_spaces = false; + } else if ( + tokenizer_pre == "hunyuan") { + pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -1839,6 +1848,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { || t.first == "" || t.first == "_" || t.first == "<｜end▁of▁sentence｜>" // DeepSeek + || t.first == "" // smoldocling ) { special_eot_id = t.second; if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) { @@ -1998,6 +2008,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { || t.first == "" || t.first == "_" || t.first == "<|end_of_text|>" + || t.first == "" // smoldocling ) { special_eog_ids.insert(t.second); if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) { diff --git a/examples/talk-llama/llama-vocab.h b/examples/talk-llama/llama-vocab.h index 40e4d1c0..46a1ccec 100644 --- a/examples/talk-llama/llama-vocab.h +++ b/examples/talk-llama/llama-vocab.h @@ -6,6 +6,47 @@ #include #include +// pre-tokenization types +enum llama_vocab_pre_type { + LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0, + LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1, + LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2, + LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3, + LLAMA_VOCAB_PRE_TYPE_FALCON = 4, + LLAMA_VOCAB_PRE_TYPE_MPT = 5, + LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, + LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, + LLAMA_VOCAB_PRE_TYPE_REFACT = 8, + LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, + LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10, + LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, + LLAMA_VOCAB_PRE_TYPE_OLMO = 12, + LLAMA_VOCAB_PRE_TYPE_DBRX = 13, + LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, + LLAMA_VOCAB_PRE_TYPE_PORO = 15, + LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16, + LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17, + LLAMA_VOCAB_PRE_TYPE_VIKING = 18, + LLAMA_VOCAB_PRE_TYPE_JAIS = 19, + LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, + LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, + LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, + LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, + LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, + LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, + LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, + LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, + LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, + LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, + LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, + LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, + LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, + LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, + LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, + LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35, + LLAMA_VOCAB_PRE_TYPE_HUNYUAN = 36, +}; + struct LLM_KV; struct llama_model_loader; diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index 3eda9bc6..f73b1ab6 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -79,46 +79,6 @@ extern "C" { LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization }; - // pre-tokenization types - enum llama_vocab_pre_type { - LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0, - LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1, - LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM = 2, - LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3, - LLAMA_VOCAB_PRE_TYPE_FALCON = 4, - LLAMA_VOCAB_PRE_TYPE_MPT = 5, - LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, - LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, - LLAMA_VOCAB_PRE_TYPE_REFACT = 8, - LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9, - LLAMA_VOCAB_PRE_TYPE_STABLELM2 = 10, - LLAMA_VOCAB_PRE_TYPE_QWEN2 = 11, - LLAMA_VOCAB_PRE_TYPE_OLMO = 12, - LLAMA_VOCAB_PRE_TYPE_DBRX = 13, - LLAMA_VOCAB_PRE_TYPE_SMAUG = 14, - LLAMA_VOCAB_PRE_TYPE_PORO = 15, - LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16, - LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17, - LLAMA_VOCAB_PRE_TYPE_VIKING = 18, - LLAMA_VOCAB_PRE_TYPE_JAIS = 19, - LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, - LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, - LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, - LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, - LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, - LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, - LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, - LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, - LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, - LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, - LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, - LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, - LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, - LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, - LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, - LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35, - }; - enum llama_rope_type { LLAMA_ROPE_TYPE_NONE = -1, LLAMA_ROPE_TYPE_NORM = 0,