diff --git a/src/llama-kv-cache-dsa-iswa.cpp b/src/llama-kv-cache-dsa-iswa.cpp index 02d14a344..3d91228f3 100644 --- a/src/llama-kv-cache-dsa-iswa.cpp +++ b/src/llama-kv-cache-dsa-iswa.cpp @@ -23,7 +23,6 @@ llama_kv_cache_dsa_iswa::llama_kv_cache_dsa_iswa( uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, const layer_filter_cb & filter_mla, const layer_filter_cb & filter_lid, const layer_reuse_cb & reuse) : unified(unified) { @@ -65,14 +64,14 @@ llama_kv_cache_dsa_iswa::llama_kv_cache_dsa_iswa( kv_dsa = std::make_unique( model, no_alloc, type_k, type_v, - v_trans, offload, unified, size_dsa, n_seq_max, n_pad, + v_trans, offload, unified, size_dsa, n_seq_max, 0, LLAMA_SWA_TYPE_NONE, filter_dsa, filter_lid, reuse); LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa); kv_swa = std::make_unique( model, no_alloc, hparams, type_k, type_v, - v_trans, offload, unified, size_swa, n_seq_max, n_pad, + v_trans, offload, unified, size_swa, n_seq_max, hparams.n_swa, hparams.swa_type, nullptr, filter_swa, reuse, nullptr); } diff --git a/src/llama-kv-cache-dsa-iswa.h b/src/llama-kv-cache-dsa-iswa.h index 1aae269da..73fbaa7df 100644 --- a/src/llama-kv-cache-dsa-iswa.h +++ b/src/llama-kv-cache-dsa-iswa.h @@ -24,7 +24,6 @@ public: uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, const layer_filter_cb & filter_mla, const layer_filter_cb & filter_lid, const layer_reuse_cb & reuse); diff --git a/src/llama-kv-cache-dsa.cpp b/src/llama-kv-cache-dsa.cpp index 307ff31dd..2602ae3c3 100644 --- a/src/llama-kv-cache-dsa.cpp +++ b/src/llama-kv-cache-dsa.cpp @@ -21,7 +21,6 @@ llama_kv_cache_dsa::llama_kv_cache_dsa( bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, const layer_filter_cb & filter_mla, @@ -33,7 +32,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa( kv_mla = std::make_unique( model, no_alloc, model.hparams, type_k, type_v, - v_trans, offload, unified, kv_size, n_seq_max, n_pad, + v_trans, offload, unified, kv_size, n_seq_max, n_swa, swa_type, nullptr, filter_mla, reuse, nullptr); // we use llama_kv_cache for caching indexer keys @@ -50,7 +49,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa( kv_lid = std::make_unique( model, no_alloc, hparams_lid, type_k, type_v, - v_trans, offload, unified, kv_size, n_seq_max, n_pad, + v_trans, offload, unified, kv_size, n_seq_max, n_swa, swa_type, nullptr, filter_lid, reuse, nullptr); } diff --git a/src/llama-kv-cache-dsa.h b/src/llama-kv-cache-dsa.h index 584ab40b9..13a59047c 100644 --- a/src/llama-kv-cache-dsa.h +++ b/src/llama-kv-cache-dsa.h @@ -24,7 +24,6 @@ public: bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, const layer_filter_cb & filter_mla, diff --git a/src/llama-kv-cache-dsv4.cpp b/src/llama-kv-cache-dsv4.cpp index 70bfee8f9..f0ab9f35d 100644 --- a/src/llama-kv-cache-dsv4.cpp +++ b/src/llama-kv-cache-dsv4.cpp @@ -1252,7 +1252,6 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4( uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, uint32_t n_rs_seq, const layer_filter_cb & filter, const layer_reuse_cb & reuse) : @@ -1288,7 +1287,7 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4( kv_raw = std::make_unique( model, no_alloc, hparams_raw, type_k, type_v, - v_trans, offload, swa_full, unified_raw, kv_size, n_seq_max, n_ubatch, n_pad, + v_trans, offload, swa_full, unified_raw, kv_size, n_seq_max, n_ubatch, nullptr, filter_raw, reuse, nullptr); // V4.1: CSA slot = ratio 2, HCA slot = ratio 1, no overlap and no indexer compressor; the index K is stored @@ -1392,7 +1391,7 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4( kv_csa = std::make_unique( model, no_alloc, hparams_csa, type_k, type_v, - v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_csa), 256u), n_seq_max, n_pad, + v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_csa), 256u), n_seq_max, 0, LLAMA_SWA_TYPE_NONE, nullptr, filter_csa, nullptr, nullptr); LLAMA_LOG_INFO("%s: creating DSV4 HCA compressed KV cache, size = %u cells\n", @@ -1400,7 +1399,7 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4( kv_hca = std::make_unique( model, no_alloc, hparams_hca, type_k, type_v, - v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_hca), 256u), n_seq_max, n_pad, + v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_hca), 256u), n_seq_max, 0, LLAMA_SWA_TYPE_NONE, nullptr, filter_hca, nullptr, nullptr); LLAMA_LOG_INFO("%s: creating DSV4 lightning-indexer KV cache, size = %u cells\n", @@ -1408,7 +1407,7 @@ llama_kv_cache_dsv4::llama_kv_cache_dsv4( kv_lid = std::make_unique( model, no_alloc, hparams_lid, type_k, type_v, - v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_csa), 256u), n_seq_max, n_pad, + v_trans, offload, unified_compressed, GGML_PAD(dsv4_comp_size(kv_size, ratio_csa), 256u), n_seq_max, 0, LLAMA_SWA_TYPE_NONE, nullptr, filter_lid, nullptr, nullptr); LLAMA_LOG_INFO("%s: creating DSV4 CSA compressor state\n", __func__); diff --git a/src/llama-kv-cache-dsv4.h b/src/llama-kv-cache-dsv4.h index 3e97c46e2..44bf95955 100644 --- a/src/llama-kv-cache-dsv4.h +++ b/src/llama-kv-cache-dsv4.h @@ -101,7 +101,6 @@ public: uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, uint32_t n_rs_seq, const layer_filter_cb & filter, const layer_reuse_cb & reuse); diff --git a/src/llama-kv-cache-iswa.cpp b/src/llama-kv-cache-iswa.cpp index 4606bcf98..8ce8e0165 100644 --- a/src/llama-kv-cache-iswa.cpp +++ b/src/llama-kv-cache-iswa.cpp @@ -23,13 +23,12 @@ llama_kv_cache_iswa::llama_kv_cache_iswa( uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, llama_memory_t mem_other, const layer_filter_cb & filter, const layer_reuse_cb & reuse, const layer_share_cb & share) : llama_kv_cache_iswa(model, no_alloc, model.hparams, type_k, type_v, v_trans, offload, swa_full, unified, - kv_size, n_seq_max, n_ubatch, n_pad, mem_other, filter, reuse, share) { + kv_size, n_seq_max, n_ubatch, mem_other, filter, reuse, share) { } llama_kv_cache_iswa::llama_kv_cache_iswa( @@ -45,7 +44,6 @@ llama_kv_cache_iswa::llama_kv_cache_iswa( uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, llama_memory_t mem_other, const layer_filter_cb & filter, const layer_reuse_cb & reuse, @@ -96,14 +94,14 @@ llama_kv_cache_iswa::llama_kv_cache_iswa( kv_base = std::make_unique( model, no_alloc, hparams, type_k, type_v, - v_trans, offload, unified, size_base, n_seq_max, n_pad, + v_trans, offload, unified, size_base, n_seq_max, 0, LLAMA_SWA_TYPE_NONE, mem_other_base, filter_base, reuse, share); LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa); kv_swa = std::make_unique( model, no_alloc, hparams, type_k, type_v, - v_trans, offload, unified, size_swa, n_seq_max, n_pad, + v_trans, offload, unified, size_swa, n_seq_max, hparams.n_swa, hparams.swa_type, mem_other_swa, filter_swa, reuse, share); } diff --git a/src/llama-kv-cache-iswa.h b/src/llama-kv-cache-iswa.h index 8352669dd..65aba2af7 100644 --- a/src/llama-kv-cache-iswa.h +++ b/src/llama-kv-cache-iswa.h @@ -25,7 +25,6 @@ public: uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, llama_memory_t mem_other, const layer_filter_cb & filter, const layer_reuse_cb & reuse, @@ -44,7 +43,6 @@ public: uint32_t kv_size, uint32_t n_seq_max, uint32_t n_ubatch, - uint32_t n_pad, llama_memory_t mem_other, const layer_filter_cb & filter, const layer_reuse_cb & reuse, diff --git a/src/llama-kv-cache-msa.cpp b/src/llama-kv-cache-msa.cpp index a57937ff4..2d6bac0cb 100644 --- a/src/llama-kv-cache-msa.cpp +++ b/src/llama-kv-cache-msa.cpp @@ -20,21 +20,20 @@ llama_kv_cache_msa::llama_kv_cache_msa( bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, const layer_filter_cb & filter, const layer_filter_cb & filter_idx, const layer_reuse_cb & reuse) : hparams_idx(model.hparams), - n_stream(unified ? 1 : n_seq_max), n_seq_max(n_seq_max), n_pad(n_pad), + n_stream(unified ? 1 : n_seq_max), n_seq_max(n_seq_max), n_swa(n_swa), swa_type(swa_type) { LLAMA_LOG_INFO("%s: creating main KV cache, size = %u cells\n", __func__, kv_size); kv_base = std::make_unique( model, no_alloc, model.hparams, type_k, type_v, - v_trans, offload, unified, kv_size, n_seq_max, n_pad, + v_trans, offload, unified, kv_size, n_seq_max, n_swa, swa_type, nullptr, filter, reuse, nullptr); // the MSA indexer uses a single key head per layer @@ -46,7 +45,7 @@ llama_kv_cache_msa::llama_kv_cache_msa( kv_idx = std::make_unique( model, no_alloc, hparams_idx, type_k, type_v, - v_trans, offload, unified, kv_size, n_seq_max, n_pad, + v_trans, offload, unified, kv_size, n_seq_max, n_swa, swa_type, nullptr, filter_idx, reuse, nullptr); } @@ -262,7 +261,7 @@ const llama_kv_cache_context * llama_kv_cache_msa_context::get_idx() const { uint32_t llama_kv_cache_msa_context::get_n_pos() const { // pad the value so that the graph remains constant across batches and can be reused - const uint32_t n_pad_cur = std::max(kv->get_n_pad(), 256u); + const uint32_t n_pad = 256; llama_pos pos_max = -1; @@ -270,7 +269,7 @@ uint32_t llama_kv_cache_msa_context::get_n_pos() const { pos_max = std::max(pos_max, kv->seq_pos_max(seq_id)); } - return std::max(n_pad_cur, GGML_PAD((uint32_t) (pos_max + 1), n_pad_cur)); + return std::max(n_pad, GGML_PAD((uint32_t) (pos_max + 1), n_pad)); } void llama_kv_cache_msa_context::set_input_cell_pos(ggml_tensor * dst, const llama_ubatch * ubatch, int32_t div) const { diff --git a/src/llama-kv-cache-msa.h b/src/llama-kv-cache-msa.h index 3b1477c83..81c6827f3 100644 --- a/src/llama-kv-cache-msa.h +++ b/src/llama-kv-cache-msa.h @@ -23,7 +23,6 @@ public: bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, const layer_filter_cb & filter, @@ -68,7 +67,6 @@ public: llama_kv_cache * get_base() const; llama_kv_cache * get_idx () const; - uint32_t get_n_pad() const { return n_pad; } uint32_t get_n_seq_max() const { return n_seq_max; } uint32_t get_n_swa() const { return n_swa; } llama_swa_type get_swa_type() const { return swa_type; } @@ -79,7 +77,6 @@ private: const uint32_t n_stream = 1; const uint32_t n_seq_max = 1; - const uint32_t n_pad = 1; const uint32_t n_swa = 0; const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE; diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index 61b5cb370..e108f0423 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -73,7 +73,6 @@ llama_kv_cache::llama_kv_cache( bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, llama_memory_t mem_other, @@ -82,7 +81,7 @@ llama_kv_cache::llama_kv_cache( const layer_share_cb & share, const char * name_tag) : model(model), hparams(hparams), v_trans(v_trans), - n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type), + n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_swa(n_swa), swa_type(swa_type), ctxs_bufs(no_alloc), other(static_cast(mem_other)), v_cells_impl(other ? other->v_cells_impl : std::make_shared()), @@ -99,8 +98,6 @@ llama_kv_cache::llama_kv_cache( } } - GGML_ASSERT(kv_size % n_pad == 0); - const uint32_t n_layer = hparams.n_layer_all; // define a comparator for the buft -> ctx map to ensure that the order is well-defined: @@ -1260,12 +1257,12 @@ uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const { // pad the n_kv value so that the graph remains constant across batches and can be reused // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220) - const uint32_t n_pad_cur = std::max(n_pad, 256u); + const uint32_t n_pad = 256; for (uint32_t s = 0; s < sinfo.n_stream(); ++s) { const auto & cells = v_cells[sinfo.strm[s]]; - result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result); + result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result); } return result; diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h index b31cc226d..60d2e183c 100644 --- a/src/llama-kv-cache.h +++ b/src/llama-kv-cache.h @@ -107,7 +107,6 @@ public: bool unified, uint32_t kv_size, uint32_t n_seq_max, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, llama_memory_t mem_other, @@ -272,9 +271,6 @@ private: const uint32_t n_seq_max = 1; const uint32_t n_stream = 1; - // required padding - const uint32_t n_pad = 1; - // SWA const uint32_t n_swa = 0; diff --git a/src/llama-memory-hybrid-idx.cpp b/src/llama-memory-hybrid-idx.cpp index 03f0c342d..da1747649 100644 --- a/src/llama-memory-hybrid-idx.cpp +++ b/src/llama-memory-hybrid-idx.cpp @@ -24,7 +24,6 @@ llama_memory_hybrid_idx::llama_memory_hybrid_idx( ggml_type type_v, bool v_trans, uint32_t kv_size, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, /* recurrent */ @@ -43,7 +42,7 @@ llama_memory_hybrid_idx::llama_memory_hybrid_idx( llama_memory_hybrid( model, no_alloc, - type_k, type_v, v_trans, kv_size, n_pad, n_swa, swa_type, + type_k, type_v, v_trans, kv_size, n_swa, swa_type, type_r, type_s, rs_size, n_seq_max, n_rs_seq, offload, unified, filter_attn, filter_recr), @@ -67,7 +66,7 @@ llama_memory_hybrid_idx::llama_memory_hybrid_idx( return new llama_kv_cache( model, no_alloc, hparams_idx, type_k, type_v, v_trans, offload, unified, - kv_size, n_seq_max, n_pad, n_swa, swa_type, + kv_size, n_seq_max, n_swa, swa_type, nullptr, filter_idx, nullptr, nullptr, "idx_"); }()) {} diff --git a/src/llama-memory-hybrid-idx.h b/src/llama-memory-hybrid-idx.h index faa1724b0..59940dbb8 100644 --- a/src/llama-memory-hybrid-idx.h +++ b/src/llama-memory-hybrid-idx.h @@ -24,7 +24,6 @@ public: ggml_type type_v, bool v_trans, uint32_t kv_size, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, /* recurrent */ diff --git a/src/llama-memory-hybrid-iswa.cpp b/src/llama-memory-hybrid-iswa.cpp index 4c0f14c22..6d843c76e 100644 --- a/src/llama-memory-hybrid-iswa.cpp +++ b/src/llama-memory-hybrid-iswa.cpp @@ -18,7 +18,6 @@ llama_memory_hybrid_iswa::llama_memory_hybrid_iswa( bool swa_full, uint32_t kv_size, uint32_t n_ubatch, - uint32_t n_pad, /* recurrent */ ggml_type type_r, ggml_type type_s, @@ -44,7 +43,6 @@ llama_memory_hybrid_iswa::llama_memory_hybrid_iswa( kv_size, n_seq_max, n_ubatch, - n_pad, nullptr, filter_attn == nullptr ? [&](int32_t il) { return !hparams.is_recr(il); } diff --git a/src/llama-memory-hybrid-iswa.h b/src/llama-memory-hybrid-iswa.h index 6e68f3f44..52c885ac9 100644 --- a/src/llama-memory-hybrid-iswa.h +++ b/src/llama-memory-hybrid-iswa.h @@ -28,7 +28,6 @@ public: bool swa_full, uint32_t kv_size, uint32_t n_ubatch, - uint32_t n_pad, /* recurrent */ ggml_type type_r, ggml_type type_s, diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp index 0361b23ca..7d3b27908 100644 --- a/src/llama-memory-hybrid.cpp +++ b/src/llama-memory-hybrid.cpp @@ -16,7 +16,6 @@ llama_memory_hybrid::llama_memory_hybrid( ggml_type type_v, bool v_trans, uint32_t kv_size, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, /* recurrent */ @@ -43,7 +42,6 @@ llama_memory_hybrid::llama_memory_hybrid( unified, kv_size, n_seq_max, - n_pad, n_swa, swa_type, nullptr, diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h index 74fa89ea6..2b4d1269c 100644 --- a/src/llama-memory-hybrid.h +++ b/src/llama-memory-hybrid.h @@ -26,7 +26,6 @@ public: ggml_type type_v, bool v_trans, uint32_t kv_size, - uint32_t n_pad, uint32_t n_swa, llama_swa_type swa_type, /* recurrent */ diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 78a387e5d..0220c5ab8 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2842,7 +2842,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr, @@ -2870,7 +2869,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr, @@ -2896,7 +2894,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, filter_mla, @@ -2935,7 +2932,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* attn_type_v */ params.type_v, /* attn_v_trans */ !cparams.flash_attn, /* attn_kv_size */ cparams.n_ctx_seq, - /* attn_n_pad */ 1, /* attn_n_swa */ hparams.n_swa, /* attn_swa_type */ hparams.swa_type, /* recurrent_type_r */ GGML_TYPE_F32, @@ -2964,7 +2960,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr, @@ -2985,7 +2980,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr, @@ -3013,7 +3007,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr, @@ -3040,7 +3033,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, filter_mla, filter_lid, nullptr); @@ -3068,7 +3060,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, nullptr, filter_mtp, nullptr, @@ -3086,7 +3077,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, cparams.n_rs_seq, nullptr, nullptr); @@ -3110,7 +3100,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, nullptr, nullptr, nullptr, @@ -3198,7 +3187,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* attn_swa_full */ params.swa_full, /* attn_kv_size */ cparams.n_ctx_seq, /* attn_n_ubatch */ cparams.n_ubatch, - /* attn_n_pad */ 1, /* recurrent_type_r */ GGML_TYPE_F32, /* recurrent_type_s */ GGML_TYPE_F32, /* recurrent_rs_size */ std::max((uint32_t) 1, cparams.n_seq_max), @@ -3217,7 +3205,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* attn_type_v */ params.type_v, /* attn_v_trans */ !cparams.flash_attn, /* attn_kv_size */ cparams.n_ctx_seq, - /* attn_n_pad */ 1, /* attn_n_swa */ hparams.n_swa, /* attn_swa_type */ hparams.swa_type, /* recurrent_type_k */ GGML_TYPE_F32, @@ -3238,7 +3225,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* attn_type_v */ params.type_v, /* attn_v_trans */ !cparams.flash_attn, /* attn_kv_size */ cparams.n_ctx_seq, - /* attn_n_pad */ 1, /* attn_n_swa */ hparams.n_swa, /* attn_swa_type */ hparams.swa_type, /* recurrent_type_k */ GGML_TYPE_F32, @@ -3310,7 +3296,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, mem_other, filter, reuse, @@ -3328,7 +3313,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_ctx_seq, cparams.n_seq_max, cparams.n_ubatch, - 1, nullptr, filter, reuse, @@ -3348,7 +3332,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.kv_unified, cparams.n_ctx_seq, cparams.n_seq_max, - 1, hparams.n_swa, hparams.swa_type, nullptr,