llama_cpp 0.14.5 → 0.14.7

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 7d80abb57b135ff04718e34099accaaabf3358553b0f061d79b195a99386739d
-  data.tar.gz: 5b24a9b7846b962f4063a0e50f15c6d9a9c874d1931ed32c200f3383869a2fd9
+  metadata.gz: 243241c78383cb68d4fb5027ffc54ea7f6789bd74bfe85fae8e62d45e7c3145d
+  data.tar.gz: b7c792c6fb2287b71a72ff823a31706dc0830aa704d86e6f8a92d1d0630649d9
 SHA512:
-  metadata.gz: dfb20e108a57b65ff624db1e2ee37034ffca406d906268d89ff441099a02c00fd67743a786a0353df2368614003604a4bf5982089024f14aee2e0f95e210e297
-  data.tar.gz: 0a0bbd93dfe57e033f25e5c3e3d61fb568362aa2d317851dbb69fe620e5e30bc8b08c27272579e7841c50b87984abf70ade4a9e7e34fb2615e106a5c2474b79e
+  metadata.gz: 59565cd5e6bd79d98d31dcf1ce505c8388a97296f607c2a114cf92a614a2cd39291a8a18a3f58993606ea3f0970d1eadbfe670280c5261c5826a54d77a2eb85d
+  data.tar.gz: 228bc19181b0163ef922e847f67e7b6a52dc1311c4e8173586dfca82eb402c5a08c104b5bac5ba0eee4772f615f8fd17f2d06cbc6db5323d133a46d3de85eeb4

data/CHANGELOG.md

@@ -1,3 +1,16 @@
+## [[0.14.7](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.6...v0.14.7)] - 2024-04-27
+
+- Bump llama.cpp from b2698 to b2740.
+  - Add `keep_split` accessor to `ModelQuantizeParams`.
+  - Add `pooling_type` method to `Context`.
+  - Add `token_is_eog?` method to `Model`.
+
+Implementation binding for llama_sample_token_with_rng has been skipped.
+
+## [[0.14.6](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.5...v0.14.6)] - 2024-04-20
+
+- Bump llama.cpp from b2658 to b2698.
+
 ## [[0.14.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.4...v0.14.5)] - 2024-04-13
 
 - Bump llama.cpp from b2608 to b2658.

data/ext/llama_cpp/llama_cpp.cpp

@@ -1321,6 +1321,8 @@ public:
     rb_define_method(rb_cLLaMAModelQuantizeParams, "only_copy", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_only_copy), 0);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "pure=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_pure), 1);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "pure", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_pure), 0);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "keep_split=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_keep_split), 1);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "keep_split", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_keep_split), 0);
   }
 
 private:
@@ -1405,6 +1407,18 @@ private:
     LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
     return ptr->params.pure ? Qtrue : Qfalse;
   }
+
+  // keep_split
+  static VALUE _llama_model_quantize_params_set_keep_split(VALUE self, VALUE keep_split) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    ptr->params.keep_split = RTEST(keep_split) ? true : false;
+    return ptr->params.keep_split ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_model_quantize_params_get_keep_split(VALUE self) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    return ptr->params.keep_split ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAModelQuantizeParams::llama_model_quantize_params_type = {
@@ -1487,6 +1501,7 @@ public:
     rb_define_method(rb_cLLaMAModel, "token_middle", RUBY_METHOD_FUNC(_llama_model_token_middle), 0);
     rb_define_method(rb_cLLaMAModel, "token_suffix", RUBY_METHOD_FUNC(_llama_model_token_suffix), 0);
     rb_define_method(rb_cLLaMAModel, "token_eot", RUBY_METHOD_FUNC(_llama_model_token_eot), 0);
+    rb_define_method(rb_cLLaMAModel, "token_is_eog?", RUBY_METHOD_FUNC(_llama_model_token_is_eog), 1);
   }
 
 private:
@@ -1634,10 +1649,10 @@ private:
     const llama_token token = NUM2INT(token_);
     LLaMAModelWrapper* ptr = get_llama_model(self);
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(ptr->model, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(ptr->model, token, result.data(), result.size(), false);
     if (n_tokens < 0) {
       result.resize(-n_tokens);
-      const int check = llama_token_to_piece(ptr->model, token, result.data(), result.size());
+      const int check = llama_token_to_piece(ptr->model, token, result.data(), result.size(), false);
       if (check != -n_tokens) {
         rb_raise(rb_eRuntimeError, "failed to convert");
         return Qnil;
@@ -1789,6 +1804,16 @@ private:
     LLaMAModelWrapper* ptr = get_llama_model(self);
     return INT2NUM(llama_token_eot(ptr->model));
   }
+
+  static VALUE _llama_model_token_is_eog(VALUE self, VALUE token_) {
+    if (!RB_INTEGER_TYPE_P(token_)) {
+      rb_raise(rb_eArgError, "token must be an integer");
+      return Qnil;
+    }
+    const llama_token token = NUM2INT(token_);
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return llama_token_is_eog(ptr->model, token) ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAModel::llama_model_type = {
@@ -2102,6 +2127,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "sample_grammar", RUBY_METHOD_FUNC(_llama_context_sample_grammar), -1);
     rb_define_method(rb_cLLaMAContext, "grammar_accept_token", RUBY_METHOD_FUNC(_llama_context_grammar_accept_token), -1);
     rb_define_method(rb_cLLaMAContext, "apply_control_vector", RUBY_METHOD_FUNC(_llama_context_apply_control_vector), -1);
+    rb_define_method(rb_cLLaMAContext, "pooling_type", RUBY_METHOD_FUNC(_llama_context_pooling_type), 0);
   }
 
 private:
@@ -3225,6 +3251,15 @@ private:
 
     return Qnil;
   }
+
+  static VALUE _llama_context_pooling_type(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    return INT2NUM(static_cast<int>(llama_pooling_type(ptr->ctx)));
+  }
 };
 
 const rb_data_type_t RbLLaMAContext::llama_context_type = {

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.14.5'
+  VERSION = '0.14.7'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2658'
+  LLAMA_CPP_VERSION = 'b2740'
 end

data/sig/llama_cpp.rbs

@@ -141,6 +141,7 @@ module LLaMACpp
     def token_middle: () -> Integer
     def token_suffix: () -> Integer
     def token_eot: () -> Integer
+    def token_is_eog?: (Integer) -> bool
   end
 
   class Timings
@@ -260,6 +261,7 @@ module LLaMACpp
     def sample_grammar: (::LLaMACpp::TokenDataArray, grammar: ::LLaMACpp::Grammar) -> void
     def grammar_accept_token: (grammar: ::LLaMACpp::Grammar, token: Integer) -> void
     def apply_control_vector: (data: Array[Float], n_embd: Integer, il_start: Integer, il_end: Integer) -> void
+    def pooling_type: () -> Integer
   end
 
   class ContextParams
@@ -328,6 +330,8 @@ module LLaMACpp
     def only_copy=: (bool) -> bool
     def pure: () -> bool
     def pure=: (bool) -> bool
+    def keep_split: () -> bool
+    def keep_split=: (bool) -> bool
   end
 
   class Params = ContextParams

data/vendor/tmp/llama.cpp/Makefile

@@ -386,6 +386,11 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
+ifndef LLAMA_NO_LLAMAFILE
+	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
+	OBJS        += sgemm.o
+endif
+
 ifdef LLAMA_BLIS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
@@ -482,11 +487,9 @@ ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/com
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(NVCC_COMPILE)
-
 endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
-
 	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
 	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
@@ -605,6 +608,11 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
+ifndef LLAMA_NO_LLAMAFILE
+sgemm.o: sgemm.cpp sgemm.h ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -689,8 +697,8 @@ OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -724,7 +732,7 @@ lib: llama.o ggml.o $(OBJS)
 	ar rcs libllama.a $^
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll *.dylib benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 	rm -vrf ggml-cuda/*.o
 
 #
@@ -761,7 +769,7 @@ batched: examples/batched/batched.cpp                         ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -793,10 +801,19 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% Makefile
+	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
+		echo "unsigned char $${NAME}[] = {" && \
+		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
+		echo "};" && \
+		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
+	) > $@
+
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

data/vendor/tmp/llama.cpp/ggml-alloc.c

@@ -371,16 +371,16 @@ struct ggml_gallocr {
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
     GGML_ASSERT(galloc != NULL);
 
-    galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
+    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
     GGML_ASSERT(galloc->bufts != NULL);
 
-    galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
     GGML_ASSERT(galloc->buffers != NULL);
 
-    galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
     GGML_ASSERT(galloc->buf_tallocs != NULL);
 
     for (int i = 0; i < n_bufs; i++) {
@@ -646,8 +646,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         free(galloc->hash_set.keys);
         free(galloc->hash_values);
         galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
-        galloc->hash_values   = calloc(sizeof(struct hash_node), hash_size);
+        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
+        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
         GGML_ASSERT(galloc->hash_set.keys != NULL);
         GGML_ASSERT(galloc->hash_values != NULL);
     } else {
@@ -667,7 +667,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     // set the node_allocs from the hash table
     if (galloc->n_nodes < graph->n_nodes) {
         free(galloc->node_allocs);
-        galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
         GGML_ASSERT(galloc->node_allocs != NULL);
     }
     galloc->n_nodes = graph->n_nodes;
@@ -697,7 +697,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
     if (galloc->n_leafs < graph->n_leafs) {
         free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
         GGML_ASSERT(galloc->leaf_allocs != NULL);
     }
     galloc->n_leafs = graph->n_leafs;

data/vendor/tmp/llama.cpp/ggml-backend.c

@@ -822,7 +822,11 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1721,23 +1725,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
 
-    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
+    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
     // initialize hash table
     sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
-    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
+    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
+    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
 
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
-    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
+    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
 
     sched->n_backends = n_backends;
 
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     const int initial_splits_capacity = 16;
-    sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
+    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
@@ -1968,10 +1972,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
     struct ggml_hash_set hash_set = {
         /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
+        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
     };
-    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
-    bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);
+    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
 
     struct ggml_init_params params = {
         /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),

data/vendor/tmp/llama.cpp/ggml-cuda.cu

@@ -1231,7 +1231,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
-        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool());
+        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
             const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
             GGML_ASSERT(to_fp16_cuda != nullptr);
@@ -1241,7 +1241,7 @@ static void ggml_cuda_op_mul_mat_cublas(
         }
         const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
 
-        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool());
+        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool(id));
         if (src1->type != GGML_TYPE_F16) {
             const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
             GGML_ASSERT(to_fp16_cuda != nullptr);
@@ -1250,7 +1250,7 @@ static void ggml_cuda_op_mul_mat_cublas(
             to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
         }
         const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
-        ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(), row_diff*src1_ncols);
+        ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(id), row_diff*src1_ncols);
 
         const half alpha_f16 = 1.0f;
         const half beta_f16 = 0.0f;
@@ -1946,7 +1946,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     } else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || fp16_performance_good) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
@@ -1960,20 +1960,73 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
+                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
+                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
+                                                 int64_t ne11, int64_t ne10,
+                                                 size_t nb11, size_t nb12) {
+    int32_t iid1 = blockIdx.x;
+    int32_t id = blockIdx.y;
+
+    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
+
+    if (row_id_i != i02) {
+        return;
+    }
+
+    const int64_t i11 = id % ne11;
+    const int64_t i12 = iid1;
+
+    __shared__ int src1_row;
+    if (threadIdx.x == 0) {
+        src1_row = atomicAdd(cur_src1_row, 1);
+        row_mapping[src1_row] = {id, iid1};
+    }
+    __syncthreads();
+
+    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
+    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
+
+    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
+        src1_row_contiguous[i] = src1_row_original[i];
+    }
+}
+
+static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
+                                                  const mmid_row_mapping * __restrict__ row_mapping,
+                                                  int64_t ne0,
+                                                  size_t nb1, size_t nb2) {
+    int32_t i = blockIdx.x;
+
+    const int32_t i1 = row_mapping[i].i1;
+    const int32_t i2 = row_mapping[i].i2;
+
+    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
+    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
+
+    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
+        dst_row_original[j] = dst_row_contiguous[j];
+    }
+}
+
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
 
     cudaStream_t stream = ctx.stream();
 
-    const size_t nb11 = src1->nb[1];
-    const size_t nb1  =  dst->nb[1];
-
-    const int32_t id = ((int32_t *) dst->op_params)[0];
-    const int32_t n_as = src0->ne[2];
+    const int64_t n_as = ne02;
+    const int64_t n_ids = ids->ne[0];
 
     std::vector<char> ids_host(ggml_nbytes(ids));
     const char * ids_dev = (const char *) ids->data;
@@ -1982,7 +2035,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     ggml_tensor src0_row = *src0;
     ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row = *dst;
+    ggml_tensor dst_row  = *dst;
 
     char * src0_original = (char *) src0->data;
     char * src1_original = (char *) src1->data;
@@ -1990,19 +2043,39 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     src0_row.ne[2] = 1;
     src0_row.ne[3] = 1;
-    src0_row.nb[3] = src0->nb[2];
+    src0_row.nb[3] = nb02;
 
-    if (src1->ne[1] == 1) {
-        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-            const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
 
-            GGML_ASSERT(row_id >= 0 && row_id < n_as);
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
 
-            src0_row.data = src0_original + row_id*src0->nb[2];
-            src1_row.data = src1_original + i01*src1->nb[1];
-            dst_row.data  =  dst_original + i01*dst->nb[1];
+    if (ne12 == 1) {
+        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+            for (int64_t id = 0; id < n_ids; id++) {
+                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
 
-            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+                GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                const int64_t i11 = id % ne11;
+                const int64_t i12 = iid1;
+
+                const int64_t i1 = id;
+                const int64_t i2 = i12;
+
+                src0_row.data = src0_original + i02*nb02;
+                src1_row.data = src1_original + i11*nb11 + i12*nb12;
+                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
+
+                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+            }
         }
     } else {
         ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
@@ -2011,54 +2084,69 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         src1_row.data = src1_contiguous.get();
         dst_row.data  =  dst_contiguous.get();
 
-        for (int32_t row_id = 0; row_id < n_as; ++row_id) {
+        for (int64_t i02 = 0; i02 < n_as; i02++) {
             int64_t num_src1_rows = 0;
-            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
 
-                if (row_id_i != row_id) {
-                    continue;
-                }
+            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+                for (int64_t id = 0; id < n_ids; id++) {
+                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
 
-                GGML_ASSERT(row_id >= 0 && row_id < n_as);
+                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
 
-                CUDA_CHECK(cudaMemcpyAsync(src1_contiguous.get() + num_src1_rows*nb11, src1_original + i01*nb11,
-                                        nb11, cudaMemcpyDeviceToDevice, stream));
-                num_src1_rows++;
+                    if (row_id_i != i02) {
+                        continue;
+                    }
+
+                    num_src1_rows++;
+                }
             }
 
             if (num_src1_rows == 0) {
                 continue;
             }
 
-            src0_row.data = src0_original + row_id*src0->nb[2];
+            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
+            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
+            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
 
-            src1_row.ne[1] = num_src1_rows;
-            dst_row.ne[1] = num_src1_rows;
+            {
+                dim3 block_dims(std::min((unsigned int)ne10, 768u));
+                dim3 grid_dims(ids->ne[1], n_ids);
+                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                        src1_original, src1_contiguous.get(),
+                        dev_cur_src1_row.get(), dev_row_mapping.get(),
+                        ids_dev, i02, ids->nb[1], ids->nb[0],
+                        ne11, ne10,
+                        nb11, nb12);
+                CUDA_CHECK(cudaGetLastError());
+            }
+
+            src0_row.data = src0_original + i02*nb02;
 
+            GGML_ASSERT(nb11 == sizeof(float)*ne10);
+            GGML_ASSERT(nb1 == sizeof(float)*ne0);
+
+            src1_row.ne[1] = num_src1_rows;
             src1_row.nb[1] = nb11;
             src1_row.nb[2] = num_src1_rows*nb11;
             src1_row.nb[3] = num_src1_rows*nb11;
 
+            dst_row.ne[1] = num_src1_rows;
             dst_row.nb[1] = nb1;
             dst_row.nb[2] = num_src1_rows*nb1;
             dst_row.nb[3] = num_src1_rows*nb1;
 
             ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
-            num_src1_rows = 0;
-            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
-
-                if (row_id_i != row_id) {
-                    continue;
-                }
-
-                GGML_ASSERT(row_id >= 0 && row_id < n_as);
-
-                CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous.get() + num_src1_rows*nb1,
-                                        nb1, cudaMemcpyDeviceToDevice, stream));
-                num_src1_rows++;
+            {
+                dim3 block_dims(std::min((unsigned int)ne0, 768u));
+                dim3 grid_dims(num_src1_rows);
+                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                        dst_original, dst_contiguous.get(),
+                        dev_row_mapping.get(),
+                        ne0,
+                        nb1, nb2);
+                CUDA_CHECK(cudaGetLastError());
             }
         }
     }
@@ -2487,7 +2575,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
 GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
-    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
 
     GGML_UNUSED(backend);
 }