Merge branch 'upstream' into concedo_experimental

# Conflicts: # README.md # examples/server/README.md # examples/speculative/speculative.cpp # flake.lock # ggml/src/CMakeLists.txt # scripts/sync-ggml.last # tests/test-backend-ops.cpp
LostRuins · Nov 14, 2024 · df080b0 · df080b0
2 parents bfa118e + 2a82891
commit df080b0
Show file tree

Hide file tree

Showing 41 changed files with 147,491 additions and 145,726 deletions.
diff --git a/common/common.h b/common/common.h
@@ -174,7 +174,7 @@ struct common_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
+    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;

diff --git a/examples/chat-persistent.sh b/examples/chat-persistent.sh
@@ -23,8 +23,9 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
 NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
 NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
 
-SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
-SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
+SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
+'|'\
+'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
 CTX_SIZE=2048
@@ -129,15 +130,12 @@ while read -e line; do
 
     printf ' '
 
-    # HACK get num tokens from debug message
-    # TODO get both messages in one go
-    if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
-        ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+    if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
         echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
         exit 1
     fi
 
-    n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
+    n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
 
     if ((n_tokens > CTX_ROTATE_POINT)); then
         tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"

diff --git a/examples/convert_legacy_llama.py b/examples/convert_legacy_llama.py
@@ -840,6 +840,8 @@ def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None
                         self.gguf.add_base_model_version(key, base_model_entry["version"])
                     if "organization" in base_model_entry:
                         self.gguf.add_base_model_organization(key, base_model_entry["organization"])
+                    if "description" in base_model_entry:
+                        self.gguf.add_base_model_description(key, base_model_entry["description"])
                     if "url" in base_model_entry:
                         self.gguf.add_base_model_url(key, base_model_entry["url"])
                     if "doi" in base_model_entry:
@@ -849,12 +851,32 @@ def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None
                     if "repo_url" in base_model_entry:
                         self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
 
+            if metadata.datasets is not None:
+                self.gguf.add_dataset_count(len(metadata.datasets))
+                for key, dataset_entry in enumerate(metadata.datasets):
+                    if "name" in dataset_entry:
+                        self.gguf.add_dataset_name(key, dataset_entry["name"])
+                    if "author" in dataset_entry:
+                        self.gguf.add_dataset_author(key, dataset_entry["author"])
+                    if "version" in dataset_entry:
+                        self.gguf.add_dataset_version(key, dataset_entry["version"])
+                    if "organization" in dataset_entry:
+                        self.gguf.add_dataset_organization(key, dataset_entry["organization"])
+                    if "description" in dataset_entry:
+                        self.gguf.add_dataset_description(key, dataset_entry["description"])
+                    if "url" in dataset_entry:
+                        self.gguf.add_dataset_url(key, dataset_entry["url"])
+                    if "doi" in dataset_entry:
+                        self.gguf.add_dataset_doi(key, dataset_entry["doi"])
+                    if "uuid" in dataset_entry:
+                        self.gguf.add_dataset_uuid(key, dataset_entry["uuid"])
+                    if "repo_url" in dataset_entry:
+                        self.gguf.add_dataset_repo_url(key, dataset_entry["repo_url"])
+
             if metadata.tags is not None:
                 self.gguf.add_tags(metadata.tags)
             if metadata.languages is not None:
                 self.gguf.add_languages(metadata.languages)
-            if metadata.datasets is not None:
-                self.gguf.add_datasets(metadata.datasets)
 
     def add_meta_arch(self, params: Params) -> None:
         # Metadata About The Neural Architecture Itself

diff --git a/examples/server/public/index.html b/examples/server/public/index.html
@@ -200,23 +200,38 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
             <div class="label">System Message</div>
             <textarea class="textarea textarea-bordered h-24" :placeholder="'Default: ' + configDefault.systemMessage" v-model="config.systemMessage"></textarea>
           </label>
-          <template v-for="key in ['temperature', 'top_k', 'top_p', 'min_p', 'max_tokens']">
-            <label class="input input-bordered flex items-center gap-2 mb-2">
-              <b>{{ key }}</b>
-              <input type="text" class="grow" :placeholder="'Default: ' + (configDefault[key] || 'none')" v-model="config[key]" />
-            </label>
+          <template v-for="configKey in ['temperature', 'top_k', 'top_p', 'min_p', 'max_tokens']">
+            <settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
           </template>
           <!-- TODO: add more sampling-related configs, please regroup them into different "collapse" sections -->
-          <div class="collapse collapse-arrow bg-base-200 mb-2">
-            <input type="checkbox" />
-            <div class="collapse-title font-bold">Advanced config</div>
+          <!-- Section: Other sampler settings -->
+          <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
+            <summary class="collapse-title font-bold">Other sampler settings</summary>
+            <div class="collapse-content">
+              <template v-for="configKey in ['dynatemp_range', 'dynatemp_exponent', 'typical_p', 'xtc_probability', 'xtc_threshold']">
+                <settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
+              </template>
+            </div>
+          </details>
+          <!-- Section: Penalties settings -->
+          <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
+            <summary class="collapse-title font-bold">Penalties settings</summary>
+            <div class="collapse-content">
+              <template v-for="configKey in ['repeat_last_n', 'repeat_penalty', 'presence_penalty', 'frequency_penalty', 'dry_multiplier', 'dry_base', 'dry_allowed_length', 'dry_penalty_last_n']">
+                <settings-modal-numeric-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
+              </template>
+            </div>
+          </details>
+          <!-- Section: Advanced config -->
+          <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
+            <summary class="collapse-title font-bold">Advanced config</summary>
             <div class="collapse-content">
               <label class="form-control mb-2">
                 <div class="label inline">Custom JSON config (For more info, refer to <a class="underline" href="https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md" target="_blank" rel="noopener noreferrer">server documentation</a>)</div>
                 <textarea class="textarea textarea-bordered h-24" placeholder="Example: { &quot;mirostat&quot;: 1, &quot;min_p&quot;: 0.1 }" v-model="config.custom"></textarea>
               </label>
             </div>
-          </div>
+          </details>
         </div>
 
         <!-- action buttons -->
@@ -229,6 +244,21 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
     </dialog>
   </div>
 
+  <!-- Template to be used by settings modal -->
+  <template id="settings-modal-numeric-input">
+    <label class="input input-bordered join-item grow flex items-center gap-2 mb-2">
+      <!-- Show help message on hovering on the input label -->
+      <div class="dropdown dropdown-hover">
+        <div tabindex="0" role="button" class="font-bold">{{ configKey }}</div>
+        <div class="dropdown-content menu bg-base-100 rounded-box z-10 w-64 p-2 shadow mt-4">
+          {{ configInfo[configKey] || '(no help message available)' }}
+        </div>
+      </div>
+      <!-- Here we forward v-model from parent to child component, see: https://stackoverflow.com/questions/47311936/v-model-and-child-components -->
+      <input type="text" class="grow" :placeholder="'Default: ' + (configDefault[configKey] || 'none')" :value="modelValue" @input="$emit('update:modelValue', $event.target.value)" />
+    </label>
+  </template>
+
   <script src="./deps_markdown-it.js"></script>
   <script type="module">
     import { createApp, defineComponent, shallowRef, computed, h } from './deps_vue.esm-browser.js';
@@ -245,12 +275,48 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
       systemMessage: 'You are a helpful assistant.',
       // make sure these default values are in sync with `common.h`
       temperature: 0.8,
+      dynatemp_range: 0.0,
+      dynatemp_exponent: 1.0,
       top_k: 40,
       top_p: 0.95,
       min_p: 0.05,
+      xtc_probability: 0.0,
+      xtc_threshold: 0.1,
+      typical_p: 1.0,
+      repeat_last_n: 64,
+      repeat_penalty: 1.0,
+      presence_penalty: 0.0,
+      frequency_penalty: 0.0,
+      dry_multiplier: 0.0,
+      dry_base: 1.75,
+      dry_allowed_length: 2,
+      dry_penalty_last_n: -1,
       max_tokens: -1,
       custom: '', // custom json-stringified object
     };
+    const CONFIG_INFO = {
+      apiKey: '',
+      systemMessage: 'The starting message that defines how model should behave.',
+      temperature: 'Controls the randomness of the generated text by affecting the probability distribution of the output tokens. Higher = more random, lower = more focused.',
+      dynatemp_range: 'Addon for the temperature sampler. The added value to the range of dynamic temperature, which adjusts probabilities by entropy of tokens.',
+      dynatemp_exponent: 'Addon for the temperature sampler. Smoothes out the probability redistribution based on the most probable token.',
+      top_k: 'Keeps only k top tokens.',
+      top_p: 'Limits tokens to those that together have a cumulative probability of at least p',
+      min_p: 'Limits tokens based on the minimum probability for a token to be considered, relative to the probability of the most likely token.',
+      xtc_probability: 'XTC sampler cuts out top tokens; this parameter controls the chance of cutting tokens at all. 0 disables XTC.',
+      xtc_threshold: 'XTC sampler cuts out top tokens; this parameter controls the token probability that is required to cut that token.',
+      typical_p: 'Sorts and limits tokens based on the difference between log-probability and entropy.',
+      repeat_last_n: 'Last n tokens to consider for penalizing repetition',
+      repeat_penalty: 'Controls the repetition of token sequences in the generated text',
+      presence_penalty: 'Limits tokens based on whether they appear in the output or not.',
+      frequency_penalty: 'Limits tokens based on how often they appear in the output.',
+      dry_multiplier: 'DRY sampling reduces repetition in generated text even across long contexts. This parameter sets the DRY sampling multiplier.',
+      dry_base: 'DRY sampling reduces repetition in generated text even across long contexts. This parameter sets the DRY sampling base value.',
+      dry_allowed_length: 'DRY sampling reduces repetition in generated text even across long contexts. This parameter sets the allowed length for DRY sampling.',
+      dry_penalty_last_n: 'DRY sampling reduces repetition in generated text even across long contexts. This parameter sets DRY penalty for the last n tokens.',
+      max_tokens: 'The maximum number of token per output.',
+      custom: '', // custom json-stringified object
+    };
     // config keys having numeric value (i.e. temperature, top_k, top_p, etc)
     const CONFIG_NUMERIC_KEYS = Object.entries(CONFIG_DEFAULT).filter(e => isNumeric(e[1])).map(e => e[0]);
     // list of themes supported by daisyui
@@ -269,6 +335,12 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
       { props: ["source", "options", "plugins"] }
     );
 
+    // inout field to be used by settings modal
+    const SettingsModalNumericInput = defineComponent({
+      template: document.getElementById('settings-modal-numeric-input').innerHTML,
+      props: ['configKey', 'configDefault', 'configInfo', 'modelValue'],
+    });
+
     // coversations is stored in localStorage
     // format: { [convId]: { id: string, lastModified: number, messages: [...] } }
     // convId is a string prefixed with 'conv-'
@@ -359,6 +431,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
     const mainApp = createApp({
       components: {
         VueMarkdown,
+        SettingsModalNumericInput,
       },
       data() {
         return {
@@ -376,6 +449,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
           // const
           themes: THEMES,
           configDefault: {...CONFIG_DEFAULT},
+          configInfo: {...CONFIG_INFO},
         }
       },
       computed: {},
@@ -452,8 +526,22 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
               stream: true,
               cache_prompt: true,
               temperature: this.config.temperature,
+              dynatemp_range: this.config.dynatemp_range,
+              dynatemp_exponent: this.config.dynatemp_exponent,
               top_k: this.config.top_k,
               top_p: this.config.top_p,
+              min_p: this.config.min_p,
+              typical_p: this.config.typical_p,
+              xtc_probability: this.config.xtc_probability,
+              xtc_threshold: this.config.xtc_threshold,
+              repeat_last_n: this.config.repeat_last_n,
+              repeat_penalty: this.config.repeat_penalty,
+              presence_penalty: this.config.presence_penalty,
+              frequency_penalty: this.config.frequency_penalty,
+              dry_multiplier: this.config.dry_multiplier,
+              dry_base: this.config.dry_base,
+              dry_allowed_length: this.config.dry_allowed_length,
+              dry_penalty_last_n: this.config.dry_penalty_last_n,
               max_tokens: this.config.max_tokens,
               ...(this.config.custom.length ? JSON.parse(this.config.custom) : {}),
               ...(this.config.apiKey ? { api_key: this.config.apiKey } : {}),

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
@@ -656,11 +656,16 @@ struct server_context {
     }
 
     bool validate_model_chat_template() const {
-        llama_chat_message chat[] = {{"user", "test"}};
-
-        const int res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0);
-
-        return res > 0;
+        std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
+        std::string template_key = "tokenizer.chat_template";
+        int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
+        if (res >= 0) {
+            llama_chat_message chat[] = {{"user", "test"}};
+            std::string tmpl = std::string(model_template.data(), model_template.size());
+            int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
+            return chat_res > 0;
+        }
+        return false;
     }
 
     void init() {

diff --git a/ggml/src/ggml-cuda/count-equal.cu b/ggml/src/ggml-cuda/count-equal.cu
@@ -44,7 +44,7 @@ void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int");
-    const int64_t dne = GGML_PAD(ne / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
+    const int64_t dne = GGML_PAD((ne + 4*nsm - 1) / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
 
     CUDA_CHECK(cudaMemsetAsync(dst_d, 0, ggml_nbytes(dst), stream));
 

diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
@@ -596,17 +596,12 @@ @implementation GGMLMetalClass
             ctx->kernels[i].pipeline = nil;
         }
 
-        /*
-            GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
-                    (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
-                    (int) kernel->pipeline.threadExecutionWidth); \
-        */
 #define GGML_METAL_ADD_KERNEL(e, name, supported) \
         if (supported) { \
             struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
             id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
             kernel->pipeline = [device newComputePipelineStateWithFunction:metal_function error:&error]; \
-            GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+            GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
                     (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
                     (int) kernel->pipeline.threadExecutionWidth); \
             [metal_function release]; \
@@ -3046,6 +3041,8 @@ static void ggml_metal_encode_node(
 
                 bool use_vec_kernel = false;
 
+                // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
+                //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 if (ne01 >= 4 || (ne00%128 != 0)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16: