quantize: be able to specify the token embedding tensor type

* metal : require ne00 >= 128 for mat-mat kernels

ggml-ci

* llama : pad n_ctx by 32

ggml-ci

.github/workflows/build.yml

@@ -41,7 +41,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -97,7 +97,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test

README-sycl.md

@@ -29,6 +29,7 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 ## News
 
 - 2024.3
+  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
   - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
   - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
   - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.

common/common.cpp

@@ -101,7 +101,7 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-void process_escapes(std::string& input) {
+void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
 

examples/batched/batched.cpp

@@ -48,6 +48,8 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    process_escapes(params.prompt);
+
     // init LLM
 
     llama_backend_init();
@@ -78,7 +80,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx = n_kv_req;
+    ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
     ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;

examples/quantize/quantize.cpp

@@ -189,6 +189,18 @@ static void prepare_imatrix(const std::string& imatrix_file,
     }
 }
 
+static ggml_type parse_ggml_type(const char * arg) {
+    ggml_type result = GGML_TYPE_COUNT;
+    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
+        auto type = ggml_type(j);
+        const auto * name = ggml_type_name(type);
+        if (name && strcmp(arg, name) == 0) {
+            result = type; break;
+        }
+    }
+    return result;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -203,6 +215,18 @@ int main(int argc, char ** argv) {
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
             params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {

llama.cpp

@@ -11971,27 +11971,34 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
     // with the quantization of the output tensor
     if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
-        int nx = tensor->ne[0];
-        if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
-            new_type = GGML_TYPE_Q8_0;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
-                 ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
-            new_type = GGML_TYPE_Q5_K;
-        }
-        else if (new_type != GGML_TYPE_Q8_0) {
-            new_type = GGML_TYPE_Q6_K;
+        if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->output_tensor_type;
+        } else {
+            int nx = tensor->ne[0];
+            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_Q5_K;
+            }
+            else if (new_type != GGML_TYPE_Q8_0) {
+                new_type = GGML_TYPE_Q6_K;
+            }
         }
     } else if (name == "token_embd.weight") {
-        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
-            ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
-            new_type = GGML_TYPE_Q2_K;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
-            new_type = GGML_TYPE_IQ3_S;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
-            new_type = GGML_TYPE_IQ3_S;
+        if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->token_embedding_type;
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+                new_type = GGML_TYPE_Q2_K;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
         }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
                ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
@@ -12887,6 +12894,8 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
     struct llama_model_quantize_params result = {
         /*.nthread                     =*/ 0,
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
+        /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
+        /*.token_embedding_type        =*/ GGML_TYPE_COUNT,
         /*.allow_requantize            =*/ false,
         /*.quantize_output_tensor      =*/ true,
         /*.only_copy                   =*/ false,
@@ -13044,6 +13053,9 @@ struct llama_context * llama_new_context_with_model(
     cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
     cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
 
+    // this is necessary due to kv_self.n being padded later during inference
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 32);
+
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
     cparams.n_ubatch         = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);

llama.h

@@ -275,13 +275,15 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;      // quantize to this llama_ftype
-        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor; // quantize output.weight
-        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                   // quantize all tensors to the default type
-        void * imatrix;              // pointer to importance matrix data
+        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;              // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;   // output tensor type
+        enum ggml_type token_embedding_type; // itoken embeddings tensor type
+        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;         // quantize output.weight
+        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                           // quantize all tensors to the default type
+        void * imatrix;                      // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types

tests/test-backend-ops.cpp

@@ -2091,6 +2091,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         }
     }
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,  128, { 8,  1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,  128, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,   64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,   64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+
     for (ggml_type type_a : all_types) {
         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
             for (int n_mats : {2, 4, 8}) {