diff --git a/common/ngram-map.cpp b/common/ngram-map.cpp index cab231bad..c5b8fc75e 100644 --- a/common/ngram-map.cpp +++ b/common/ngram-map.cpp @@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, * @return Vector of draft tokens, empty if no matching pattern is found */ llama_tokens common_ngram_simple_draft( - common_ngram_simple_state & state, + const common_ngram_simple_config & config, const llama_tokens & tokens, llama_token sampled) { // Simple implementation of self-speculative decoding without a draft model. // const size_t cur_len = tokens.size(); - // Only check every check_rate tokens to save compute - // i.e., perform check if (cur_len - idx_last_check) >= check_rate - if (state.idx_last_check + state.config.check_rate > cur_len) { - llama_tokens draft_tokens; - return draft_tokens; - } - size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history - size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft + const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history + const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft // vector for tokens we want to verify. // return empty vector if there is no match. @@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft( } pattern.push_back(sampled); // add the last token to the pattern - // We do a search in the token history. - state.idx_last_check = cur_len; - size_t match_pos = 0; // we ignore position 0, position 0 == no match // search backwards, but skip the current match (we are currently there) for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) { diff --git a/common/ngram-map.h b/common/ngram-map.h index c094d513d..9668bd5a7 100644 --- a/common/ngram-map.h +++ b/common/ngram-map.h @@ -27,23 +27,9 @@ struct common_ngram_simple_config { uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token }; -// current state (and config) of n-gram simple. -struct common_ngram_simple_state { - common_ngram_simple_config config; - - size_t idx_last_check = 0; // index of last check in context history (mutable) - - common_ngram_simple_state(const common_ngram_simple_config & config) - : config(config) {} -}; - // Searches for a n-gram in the history and checks whether a draft sequence should be generated. -// state: the ngram simple state to search in. -// inp: the tokens generated so far. -// sampled: the token that was just sampled. -// draft: vector to store the draft tokens, initially empty. llama_tokens common_ngram_simple_draft( - common_ngram_simple_state & state, + const common_ngram_simple_config & config, const llama_tokens & tokens, llama_token sampled); diff --git a/common/speculative.cpp b/common/speculative.cpp index 80cd31e35..c99b19dbf 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -463,12 +463,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state { // state of self-speculation (simple implementation, not ngram-map) struct common_speculative_state_ngram_simple : public common_speculative_state { - common_ngram_simple_state state; + common_ngram_simple_config config; + + uint16_t check_id = 0; // used to control the frequency of generating drafts common_speculative_state_ngram_simple( enum common_speculative_type type, - common_ngram_simple_state state) - : common_speculative_state(type), state(state) {} + common_ngram_simple_config config) + : common_speculative_state(type), config(config) {} void begin(const llama_tokens & prompt) override { GGML_UNUSED(prompt); @@ -479,7 +481,13 @@ struct common_speculative_state_ngram_simple : public common_speculative_state { const llama_tokens & prompt_tgt, llama_token id_last, llama_tokens & result) override { - result = common_ngram_simple_draft(state, prompt_tgt, id_last); + ++check_id; + if (check_id < config.check_rate) { + return; + } + check_id = 0; + + result = common_ngram_simple_draft(config, prompt_tgt, id_last); GGML_UNUSED(params); } @@ -889,14 +897,14 @@ common_speculative * common_speculative_init( uint16_t mgram_size_value = ngram_map.size_value; uint16_t check_rate = ngram_map.check_rate; - auto config_simple = common_ngram_simple_config{ + auto config_simple = common_ngram_simple_config { /* .size_ngram = */ ngram_size_key, /* .size_mgram = */ mgram_size_value, /* .check_rate = */ check_rate }; auto state = std::make_unique( /* .type = */ config.type, - /* .state = */ common_ngram_simple_state(config_simple) + /* .state = */ config_simple ); impls.push_back(std::move(state)); break; diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp index 57b6659ba..06d946c5f 100644 --- a/src/models/qwen3next.cpp +++ b/src/models/qwen3next.cpp @@ -265,9 +265,15 @@ std::pair llm_build_qwen3next::build_delta_net_chu cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); - ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp); + ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, + 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); + cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + // state to be updated per chunk ggml_tensor * new_state = state; // ggml_dup(ctx0, state); @@ -322,9 +328,9 @@ std::pair llm_build_qwen3next::build_delta_net_chu : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new - ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk)); + ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? - ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff))); + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));