Merge a279f17815 into 815fe72adc

convert_hf_to_gguf.py

@@ -357,6 +357,10 @@ class Model:
                         data_qtype = gguf.GGMLQuantizationType.TQ1_0
                     elif self.ftype == gguf.LlamaFileType.MOSTLY_TQ2_0:
                         data_qtype = gguf.GGMLQuantizationType.TQ2_0
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q4_0:
+                        data_qtype = gguf.GGMLQuantizationType.Q4_0
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q4_1:
+                        data_qtype = gguf.GGMLQuantizationType.Q4_1
                     else:
                         raise ValueError(f"Unknown file type: {self.ftype.name}")
 
@@ -4296,8 +4300,8 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q4_0", "q4_1", "q8_0", "tq1_0", "tq2_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q4_0, q4_1 , q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -4383,6 +4387,8 @@ def main() -> None:
         "f32": gguf.LlamaFileType.ALL_F32,
         "f16": gguf.LlamaFileType.MOSTLY_F16,
         "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q4_0": gguf.LlamaFileType.MOSTLY_Q4_0,
+        "q4_1": gguf.LlamaFileType.MOSTLY_Q4_1,
         "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
         "tq1_0": gguf.LlamaFileType.MOSTLY_TQ1_0,
         "tq2_0": gguf.LlamaFileType.MOSTLY_TQ2_0,

ggml/include/ggml.h

@@ -217,7 +217,6 @@
 
 #define GGML_MAX_DIMS           4
 #define GGML_MAX_PARAMS         2048
-#define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
 #define GGML_MAX_N_THREADS      512
 #define GGML_MAX_OP_PARAMS      64
@@ -657,6 +656,7 @@ extern "C" {
     };
 
     // scratch buffer
+    // TODO: deprecate and remove
     struct ggml_scratch {
         size_t offs;
         size_t size;
@@ -760,8 +760,9 @@ extern "C" {
 
     // main
 
-    GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void                  ggml_free(struct ggml_context * ctx);
+    GGML_API struct ggml_context * ggml_init (struct ggml_init_params params);
+    GGML_API void                  ggml_reset(struct ggml_context * ctx);
+    GGML_API void                  ggml_free (struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 

ggml/src/CMakeLists.txt

@@ -1402,7 +1402,7 @@ list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
 
 find_library(MATH_LIBRARY m)
 if (MATH_LIBRARY)
-    if (NOT WIN32 OR NOT GGML_SYCL)
+    if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT})
         list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
     endif()
 endif()

ggml/src/ggml.c

@@ -306,6 +306,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
 }
 
 #define GGML_DEBUG 0
+
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
 
@@ -2014,7 +2015,7 @@ static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
 
 struct ggml_context {
     size_t mem_size;
-    void* mem_buffer;
+    void * mem_buffer;
     bool   mem_buffer_owned;
     bool   no_alloc;
     bool   no_alloc_save; // this is used to save the no_alloc state when using scratch buffers
@@ -3263,7 +3264,6 @@ struct ggml_numa_nodes {
 //
 
 struct ggml_state {
-    struct ggml_context_container contexts[GGML_MAX_CONTEXTS];
     struct ggml_numa_nodes numa;
 };
 
@@ -3845,7 +3845,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
 
             g_state = (struct ggml_state) {
-                /*.contexts =*/ { { 0 } },
                 /*.numa =*/ {
                     .n_nodes = 0,
                     .total_cpus = 0,
@@ -3864,26 +3863,9 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         is_first_call = false;
     }
 
-    // find non-used context in g_state
-    struct ggml_context * ctx = NULL;
-
-    for (int i = 0; i < GGML_MAX_CONTEXTS; i++) {
-        if (!g_state.contexts[i].used) {
-            g_state.contexts[i].used = true;
-            ctx = &g_state.contexts[i].context;
-
-            GGML_PRINT_DEBUG("%s: found unused context %d\n", __func__, i);
-            break;
-        }
-    }
-
-    if (ctx == NULL) {
-        GGML_PRINT_DEBUG("%s: no unused context found\n", __func__);
-
     ggml_critical_section_end();
 
-        return NULL;
-    }
+    struct ggml_context * ctx = GGML_MALLOC(sizeof(struct ggml_context));
 
     // allow to call ggml_init with 0 size
     if (params.mem_size == 0) {
@@ -3911,42 +3893,31 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
-    ggml_critical_section_end();
-
     return ctx;
 }
 
+void ggml_reset(struct ggml_context * ctx) {
+    if (ctx == NULL) {
+        return;
+    }
+
+    ctx->n_objects     = 0;
+    ctx->objects_begin = NULL;
+    ctx->objects_end   = NULL;
+    ctx->scratch       = (struct ggml_scratch) { 0, 0, NULL, };
+    ctx->scratch_save  = (struct ggml_scratch) { 0, 0, NULL, };
+}
+
 void ggml_free(struct ggml_context * ctx) {
     if (ctx == NULL) {
         return;
     }
 
-    // make this function thread safe
-    ggml_critical_section_start();
-
-    bool found = false;
-
-    for (int i = 0; i < GGML_MAX_CONTEXTS; i++) {
-        if (&g_state.contexts[i].context == ctx) {
-            g_state.contexts[i].used = false;
-
-            GGML_PRINT_DEBUG("%s: context %d has been freed. memory used = %zu\n",
-                    __func__, i, ggml_used_mem(ctx));
-
     if (ctx->mem_buffer_owned) {
         ggml_aligned_free(ctx->mem_buffer, ctx->mem_size);
     }
 
-            found = true;
-            break;
-        }
-    }
-
-    if (!found) {
-        GGML_PRINT_DEBUG("%s: context not found\n", __func__);
-    }
-
-    ggml_critical_section_end();
+    GGML_FREE(ctx);
 }
 
 size_t ggml_used_mem(const struct ggml_context * ctx) {

scripts/sync-ggml.last

@@ -1 +1 @@
-162e232411ee98ceb0cccfa84886118d917d2123
+bb78a40dc60e04c626bac2b65840b509988e990d

src/llama.cpp

@@ -4860,19 +4860,12 @@ struct llama_model_loader {
         *last  = 0;
         *addr = mapping->addr;
         for (ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor; tensor = ggml_get_next_tensor(ctx, tensor)) {
-            try {
             const auto * weight = get_weight(ggml_get_name(tensor));
-                if (!weight) {
-                    continue;
-                }
-                if (weight->idx != idx) {
+            if (!weight || weight->idx != idx) {
                 continue;
             }
             *first = std::min(*first, weight->offs);
             *last  = std::max(*last,  weight->offs + ggml_nbytes(tensor));
-            } catch(...) {
-                // the tensor is not in the model
-            }
         }
     }
 
@@ -5049,7 +5042,6 @@ struct llama_model_loader {
                     ggml_backend_tensor_set(cur, data, 0, n_size);
                 }
             } else {
-                GGML_ASSERT(weight->idx < files.size());
                 const auto & file = files.at(weight->idx);
                 if (ggml_backend_buffer_is_host(cur->buffer)) {
                     file->seek(weight->offs, SEEK_SET);
@@ -17170,19 +17162,11 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
 
         auto * buft = ggml_backend_cpu_buffer_type();
         // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory
-        ggml_tensor * output_tensor = lctx.model.output;
-        if (!output_tensor) {
-            // bert models don't have an output tensor, use the last layer
-            output_tensor = lctx.model.layers.back().layer_out_norm;
-        }
-        if (output_tensor) {
-            auto * output_buft = ggml_backend_buffer_get_type(output_tensor->buffer);
-            auto * output_dev = ggml_backend_buft_get_device(output_buft);
-            auto * output_dev_host_buft = ggml_backend_dev_host_buffer_type(output_dev);
+        auto * output_dev = lctx.model.dev_output.dev;
+        auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr;
         if (output_dev_host_buft) {
             buft = output_dev_host_buft;
         }
-        }
         lctx.buf_output = ggml_backend_buft_alloc_buffer(buft, new_size);
         if (lctx.buf_output == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0));
@@ -18623,8 +18607,25 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         }
     }
 
+    // make a list of weights
+    std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
+    tensors.reserve(ml.weights_map.size());
     for (const auto & it : ml.weights_map) {
-        const struct ggml_tensor * tensor = it.second.tensor;
+        tensors.push_back(&it.second);
+    }
+
+    // keep_split requires that the weights are sorted by split index
+    if (params->keep_split) {
+        std::sort(tensors.begin(), tensors.end(), [](const llama_model_loader::llama_tensor_weight * a, const llama_model_loader::llama_tensor_weight * b) {
+            if (a->idx == b->idx) {
+                return a->offs < b->offs;
+            }
+            return a->idx < b->idx;
+        });
+    }
+
+    for (const auto * it : tensors) {
+        const struct ggml_tensor * tensor = it->tensor;
 
         const std::string name = ggml_get_name(tensor);
 
@@ -18664,22 +18665,20 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     std::vector<no_init<float>> f32_conv_buf;
 
     uint16_t n_split = 1;
-    const auto & weights_map = ml.weights_map;
 
     // Assume split index is continuous
     if (params->keep_split) {
-        for (const auto & it : weights_map) {
-            n_split = std::max(uint16_t(it.second.idx + 1), n_split);
+        for (const auto * it : tensors) {
+            n_split = std::max(uint16_t(it->idx + 1), n_split);
         }
-
     }
     std::vector<gguf_context*> ctx_outs(n_split, NULL);
     ctx_outs[0] = ctx_out;
 
     // populate the original tensors so we get an initial meta data
-    for (const auto & it : weights_map) {
-        uint16_t i_split = params->keep_split ? it.second.idx : 0;
-        struct ggml_tensor * tensor = it.second.tensor;
+    for (const auto * it : tensors) {
+        uint16_t i_split = params->keep_split ? it->idx : 0;
+        struct ggml_tensor * tensor = it->tensor;
         if (ctx_outs[i_split] == NULL) {
             ctx_outs[i_split] = gguf_init_empty();
         }
@@ -18726,8 +18725,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     const auto tn = LLM_TN(model.arch);
     new_ofstream(0);
-    for (const auto & it : weights_map) {
-        const auto & weight = it.second;
+    for (const auto * it : tensors) {
+        const auto & weight = *it;
         struct ggml_tensor * tensor = weight.tensor;
         if (weight.idx != cur_split && params->keep_split) {
             close_ofstream();