Merge branch 'master' into concedo_experimental

# Conflicts:
#	.devops/nix/scope.nix
#	.github/workflows/nix-ci-aarch64.yml
#	.github/workflows/nix-ci.yml
#	README.md
#	scripts/sync-ggml.last

convert-hf-to-gguf.py

@@ -218,6 +218,8 @@ def from_model_architecture(model_architecture):
             return BertModel
         if model_architecture == "NomicBertModel":
             return NomicBertModel
+        if model_architecture == "GemmaForCausalLM":
+            return GemmaModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -277,6 +279,8 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
             return gguf.MODEL_ARCH.BERT
         if arch == "NomicBertModel":
             return gguf.MODEL_ARCH.NOMIC_BERT
+        if arch == "GemmaForCausalLM":
+            return gguf.MODEL_ARCH.GEMMA
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -618,11 +622,6 @@ def write_tensors(self):
 
             self.gguf_writer.add_tensor(new_name, data)
 
-            # note: MPT output is tied to (same as) wte in original model;
-            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
-            if new_name == "token_embd.weight":
-                self.gguf_writer.add_tensor("output.weight", data)
-
 
 class OrionModel(Model):
     def set_vocab(self):
@@ -655,6 +654,8 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
+        # note: config provides rms norm but it is actually layer norm
+        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -1031,7 +1032,6 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1785,6 +1785,63 @@ def get_tensors(self):
             yield name, data
 
 
+class GemmaModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_key_length(hparams["head_dim"])
+        self.gguf_writer.add_value_length(hparams["head_dim"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in self.get_tensors():
+            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
+            if name.endswith("norm.weight"):
+                data_torch = data_torch + 1
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 ###### CONVERSION LOGIC ######
 
 

examples/server/README.md

@@ -151,7 +151,7 @@ node index.js
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
 
     `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -209,7 +209,7 @@ node index.js
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
@@ -242,7 +242,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
 - `model`: The path to the model loaded with `-m`
 - `prompt`: The provided `prompt`
 - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token

examples/server/server.cpp

@@ -401,6 +401,16 @@ struct llama_server_context
         return true;
     }
 
+    void validate_model_chat_template(server_params & sparams) {
+        llama_chat_message chat[] = {{"user", "test"}};
+        std::vector<char> buf(1);
+        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
+        if (res < 0) {
+            LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
+            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
+        }
+    }
+
     void initialize() {
         // create slots
         all_slots_are_idle = true;
@@ -1939,6 +1949,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
+    printf("  -ctk TYPE, --cache-type-k TYPE\n");
+    printf("                            KV cache data type for K (default: f16)\n");
+    printf("  -ctv TYPE, --cache-type-v TYPE\n");
+    printf("                            KV cache data type for V (default: f16)\n");
     printf("  --mmproj MMPROJ_FILE      path to a multimodal projector file for LLaVA.\n");
     printf("  --log-disable             disables logging to a file.\n");
     printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
@@ -2377,6 +2391,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             );
             llama.process_system_prompt_data(json::parse(systm_content));
         }
+        else if (arg == "-ctk" || arg == "--cache-type-k") {
+            params.cache_type_k = argv[++i];
+        }
+        else if (arg == "-ctv" || arg == "--cache-type-v") {
+            params.cache_type_v = argv[++i];
+        }
         else if(arg == "--mmproj")
         {
             if (++i >= argc)
@@ -2753,6 +2773,11 @@ int main(int argc, char **argv)
         LOG_INFO("model loaded", {});
     }
 
+    if (sparams.chat_template.empty()) { // custom chat template is not supplied
+        // check if the template comes with the model is supported by us
+        llama.validate_model_chat_template(sparams);
+    }
+
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation

ggml-cuda.cu

@@ -1,3 +1,7 @@
+#include "ggml-cuda.h"
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
 #include <algorithm>
 #include <assert.h>
 #include <atomic>
@@ -121,11 +125,6 @@
 
 #endif // defined(GGML_USE_HIPBLAS)
 
-// ggml-cuda need half type so keep ggml headers include at last
-#include "ggml-cuda.h"
-#include "ggml.h"
-#include "ggml-backend-impl.h"
-
 #define CUDART_HMAX     11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 
 #define CC_PASCAL     600

ggml-impl.h

@@ -53,11 +53,23 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    __fp16 tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
 
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    __fp16 tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
 
 #else
 
@@ -214,17 +226,18 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
-
+#if !defined(GGML_FP16_TO_FP32)
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
     return ggml_table_f32_f16[s];
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
 
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 #endif
 
 #define GGML_HASHTABLE_FULL ((size_t)-1)