diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 1e4c2422756ac..f00700da71fcd 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -65,7 +65,7 @@ if (GGML_LTO) endif() endif() -if (GGML_CCACHE) +if (GGML_CCACHE AND NOT CMAKE_C_COMPILER_LAUNCHER AND NOT CMAKE_CXX_COMPILER_LAUNCHER) find_program(GGML_CCACHE_FOUND ccache) find_program(GGML_SCCACHE_FOUND sccache) diff --git a/ggml/src/ggml-cann/.clang-format b/ggml/src/ggml-cann/.clang-format deleted file mode 100644 index 2ad03d7391936..0000000000000 --- a/ggml/src/ggml-cann/.clang-format +++ /dev/null @@ -1,168 +0,0 @@ ---- -Language: Cpp -# BasedOnStyle: Google -AccessModifierOffset: -1 -AlignAfterOpenBracket: Align -AlignConsecutiveMacros: false -AlignConsecutiveAssignments: false -AlignConsecutiveDeclarations: false -AlignEscapedNewlines: Left -AlignOperands: true -AlignTrailingComments: true -AllowAllArgumentsOnNextLine: true -AllowAllConstructorInitializersOnNextLine: true -AllowAllParametersOfDeclarationOnNextLine: true -AllowShortBlocksOnASingleLine: Never -AllowShortCaseLabelsOnASingleLine: false -AllowShortFunctionsOnASingleLine: All -AllowShortLambdasOnASingleLine: All -AllowShortIfStatementsOnASingleLine: WithoutElse -AllowShortLoopsOnASingleLine: true -AlwaysBreakAfterDefinitionReturnType: None -AlwaysBreakAfterReturnType: None -AlwaysBreakBeforeMultilineStrings: true -AlwaysBreakTemplateDeclarations: Yes -BinPackArguments: true -BinPackParameters: true -BraceWrapping: - AfterCaseLabel: false - AfterClass: false - AfterControlStatement: false - AfterEnum: false - AfterFunction: false - AfterNamespace: false - AfterObjCDeclaration: false - AfterStruct: false - AfterUnion: false - AfterExternBlock: false - BeforeCatch: false - BeforeElse: false - IndentBraces: false - SplitEmptyFunction: true - SplitEmptyRecord: true - SplitEmptyNamespace: true -BreakBeforeBinaryOperators: None -BreakBeforeBraces: Attach -BreakBeforeInheritanceComma: false -BreakInheritanceList: BeforeColon -BreakBeforeTernaryOperators: true -BreakConstructorInitializersBeforeComma: false -BreakConstructorInitializers: BeforeColon -BreakAfterJavaFieldAnnotations: false -BreakStringLiterals: true -ColumnLimit: 80 -CommentPragmas: '^ IWYU pragma:' -CompactNamespaces: false -ConstructorInitializerAllOnOneLineOrOnePerLine: true -ConstructorInitializerIndentWidth: 4 -ContinuationIndentWidth: 4 -Cpp11BracedListStyle: true -DeriveLineEnding: true -DerivePointerAlignment: true -DisableFormat: false -ExperimentalAutoDetectBinPacking: false -FixNamespaceComments: true -ForEachMacros: - - foreach - - Q_FOREACH - - BOOST_FOREACH -IncludeBlocks: Regroup -IncludeCategories: - - Regex: '^' - Priority: 2 - SortPriority: 0 - - Regex: '^<.*\.h>' - Priority: 1 - SortPriority: 0 - - Regex: '^<.*' - Priority: 2 - SortPriority: 0 - - Regex: '.*' - Priority: 3 - SortPriority: 0 -IncludeIsMainRegex: '([-_](test|unittest))?$' -IncludeIsMainSourceRegex: '' -IndentCaseLabels: true -IndentGotoLabels: true -IndentPPDirectives: None -IndentWidth: 4 -IndentWrappedFunctionNames: false -JavaScriptQuotes: Leave -JavaScriptWrapImports: true -KeepEmptyLinesAtTheStartOfBlocks: false -MacroBlockBegin: '' -MacroBlockEnd: '' -MaxEmptyLinesToKeep: 1 -NamespaceIndentation: None -ObjCBinPackProtocolList: Never -ObjCBlockIndentWidth: 2 -ObjCSpaceAfterProperty: false -ObjCSpaceBeforeProtocolList: true -PenaltyBreakAssignment: 2 -PenaltyBreakBeforeFirstCallParameter: 1 -PenaltyBreakComment: 300 -PenaltyBreakFirstLessLess: 120 -PenaltyBreakString: 1000 -PenaltyBreakTemplateDeclaration: 10 -PenaltyExcessCharacter: 1000000 -PenaltyReturnTypeOnItsOwnLine: 200 -PointerAlignment: Left -RawStringFormats: - - Language: Cpp - Delimiters: - - cc - - CC - - cpp - - Cpp - - CPP - - 'c++' - - 'C++' - CanonicalDelimiter: '' - BasedOnStyle: google - - Language: TextProto - Delimiters: - - pb - - PB - - proto - - PROTO - EnclosingFunctions: - - EqualsProto - - EquivToProto - - PARSE_PARTIAL_TEXT_PROTO - - PARSE_TEST_PROTO - - PARSE_TEXT_PROTO - - ParseTextOrDie - - ParseTextProtoOrDie - CanonicalDelimiter: '' - BasedOnStyle: google -ReflowComments: true -SortIncludes: true -SortUsingDeclarations: true -SpaceAfterCStyleCast: false -SpaceAfterLogicalNot: false -SpaceAfterTemplateKeyword: true -SpaceBeforeAssignmentOperators: true -SpaceBeforeCpp11BracedList: false -SpaceBeforeCtorInitializerColon: true -SpaceBeforeInheritanceColon: true -SpaceBeforeParens: ControlStatements -SpaceBeforeRangeBasedForLoopColon: true -SpaceInEmptyBlock: false -SpaceInEmptyParentheses: false -SpacesBeforeTrailingComments: 2 -SpacesInAngles: false -SpacesInConditionalStatement: false -SpacesInContainerLiterals: true -SpacesInCStyleCastParentheses: false -SpacesInParentheses: false -SpacesInSquareBrackets: false -SpaceBeforeSquareBrackets: false -Standard: Auto -StatementMacros: - - Q_UNUSED - - QT_REQUIRE_VERSION -TabWidth: 8 -UseCRLF: false -UseTab: Never -... - diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 9467c3a010db4..3479a8cca3d64 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1317,8 +1317,8 @@ int llama_context::decode(llama_batch & inp_batch) { n_outputs = n_outputs_new; } - // non-causal masks do not use the KV cache - if (hparams.causal_attn) { + // find KV slot + { kv_self_update(); // if we have enough unused cells before the current head -> diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 0bd40174438cc..cec203df49268 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -402,120 +402,86 @@ 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 || self_kq_mask_swa) { - // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. - if (cparams.causal_attn) { - const int64_t n_kv = kv_self->n; - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - - float * data = nullptr; - float * data_swa = nullptr; - - if (self_kq_mask) { - GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); - data = (float *) self_kq_mask->data; - } + const int64_t n_kv = kv_self->n; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; - if (self_kq_mask_swa) { - GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer)); - data_swa = (float *) self_kq_mask_swa->data; - } + float * data = nullptr; + float * data_swa = nullptr; - // For causal attention, use only the previous KV cells - // of the correct sequence for each token of the ubatch. - // It's assumed that if a token in the batch has multiple sequences, they are equivalent. - for (int h = 0; h < 1; ++h) { - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + if (self_kq_mask) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); + data = (float *) self_kq_mask->data; + } - for (int j = 0; j < n_seq_tokens; ++j) { - const llama_pos pos = ubatch->pos[s*n_seq_tokens + j]; + if (self_kq_mask_swa) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer)); + data_swa = (float *) self_kq_mask_swa->data; + } - for (int i = 0; i < n_kv; ++i) { - float f; - if (!kv_self->cells[i].has_seq_id(seq_id) || kv_self->cells[i].pos > pos) { - f = -INFINITY; + // Use only the previous KV cells of the correct sequence for each token of the ubatch. + // It's assumed that if a token in the batch has multiple sequences, they are equivalent. + // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch: + // Causal mask: + // xxx------- + // xxxx------ + // xxxxx----- + // Non-causal mask: + // xxxxx----- + // xxxxx----- + // xxxxx----- + // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615 + for (int h = 0; h < 1; ++h) { + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + j]; + for (int i = 0; i < n_kv; ++i) { + float f; + // mask the token if: + if (!kv_self->cells[i].has_seq_id(seq_id) // not the correct sequence + || (cparams.causal_attn && kv_self->cells[i].pos > pos) // for causal, mask future tokens + ) { + f = -INFINITY; + } else { + if (hparams.use_alibi) { + f = -std::abs(kv_self->cells[i].pos - pos); } else { - if (hparams.use_alibi) { - f = -std::abs(kv_self->cells[i].pos - pos); - } else { - f = 0.0f; - } - } - - if (data) { - data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + f = 0.0f; } + } - // may need to cut off old tokens for sliding window - if (data_swa) { - if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) { - f = -INFINITY; - } - data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; - } + if (data) { + data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; } - } - } - if (data) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + // may need to cut off old tokens for sliding window + if (data_swa) { + if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; } } } + } - if (data_swa) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; - } + // mask padded tokens + if (data) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; } } } - } else { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - // when using kv cache, the mask needs to match the kv cache size - const int64_t n_stride = n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); - - float * data = (float *) self_kq_mask->data; - - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch->seq_id[s1][0]; - - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; - - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; - - for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { - if (ubatch->seq_id[s0][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); - } else { - f = 0.0f; - } - break; - } - } - - data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; - } - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; - } + // mask padded tokens + if (data_swa) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; } } } diff --git a/src/llama-model.cpp b/src/llama-model.cpp index a4f06112d2842..e712960f8c14a 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -256,7 +256,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara return nullptr; } -// CPU: ACCEL -> CPU extra -> GPU host -> CPU +// CPU: ACCEL -> GPU host -> CPU extra -> CPU static buft_list_t make_cpu_buft_list(const std::vector & devices) { buft_list_t buft_list; @@ -272,32 +272,6 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } } - bool has_gpu_device = false; - for (auto * dev : devices) { - if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) { - has_gpu_device = true; - break; - } - } - - // add extra buffer types, only if no GPU device is present - // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094 - if (!has_gpu_device) { - auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); - auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); - auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) - ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); - if (ggml_backend_dev_get_extra_bufts_fn) { - ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); - while (extra_bufts && *extra_bufts) { - buft_list.emplace_back(cpu_dev, *extra_bufts); - ++extra_bufts; - } - } - } else { - LLAMA_LOG_WARN("%s: disabling extra buffer types (i.e. repacking) since a GPU device is available\n", __func__); - } - // add a host buffer type // storing the tensors in a host buffer is useful when the processing of large batches // is offloaded to a GPU device, since it reduces the time spent on data transfers @@ -312,6 +286,20 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } } + // add extra buffer types, only if no GPU device is present + // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094 + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); + auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) + ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); + if (ggml_backend_dev_get_extra_bufts_fn) { + ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); + while (extra_bufts && *extra_bufts) { + buft_list.emplace_back(cpu_dev, *extra_bufts); + ++extra_bufts; + } + } + // add the CPU buffer type for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { ggml_backend_dev_t dev = ggml_backend_dev_get(i);