Merge branch 'master' of github.com:ggerganov/llama.cpp into mfalcon_mamba_cuda

.devops/llama-server-cuda.Dockerfile

@@ -24,6 +24,8 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 ENV GGML_CUDA=1
 # Enable cURL
 ENV LLAMA_CURL=1
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 RUN make -j$(nproc) llama-server
 

.devops/llama-server-intel.Dockerfile

@@ -26,6 +26,8 @@ RUN apt-get update && \
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-rocm.Dockerfile

@@ -39,6 +39,8 @@ ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 # Enable cURL
 ENV LLAMA_CURL=1

.devops/llama-server-vulkan.Dockerfile

@@ -23,6 +23,8 @@ RUN cp /app/build/bin/llama-server /llama-server && \
     rm -rf /app
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server.Dockerfile

@@ -21,6 +21,8 @@ RUN apt-get update && \
 COPY --from=build /app/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.ecrc

@@ -1,5 +1,5 @@
 {
-  "Exclude": ["^\\.gitmodules$"],
+  "Exclude": ["^\\.gitmodules$", "stb_image\\.h"],
   "Disable": {
     "IndentSize": true
   }

common/common.cpp

@@ -327,6 +327,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 void gpt_params_parse_from_env(gpt_params & params) {
     // we only care about server-related params for now
     get_env("LLAMA_ARG_MODEL",            params.model);
+    get_env("LLAMA_ARG_MODEL_URL",        params.model_url);
+    get_env("LLAMA_ARG_MODEL_ALIAS",      params.model_alias);
+    get_env("LLAMA_ARG_HF_REPO",          params.hf_repo);
+    get_env("LLAMA_ARG_HF_FILE",          params.hf_file);
     get_env("LLAMA_ARG_THREADS",          params.n_threads);
     get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
     get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
@@ -341,6 +345,9 @@ void gpt_params_parse_from_env(gpt_params & params) {
     get_env("LLAMA_ARG_EMBEDDINGS",       params.embedding);
     get_env("LLAMA_ARG_FLASH_ATTN",       params.flash_attn);
     get_env("LLAMA_ARG_DEFRAG_THOLD",     params.defrag_thold);
+    get_env("LLAMA_ARG_CONT_BATCHING",    params.cont_batching);
+    get_env("LLAMA_ARG_HOST",             params.hostname);
+    get_env("LLAMA_ARG_PORT",             params.port);
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {

convert_hf_to_gguf.py

@@ -1572,7 +1572,7 @@ class LlamaModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -3820,7 +3820,7 @@ class ExaoneModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)

examples/server/README.md

@@ -249,23 +249,49 @@ logging:
 
 Available environment variables (if specified, these variables will override parameters specified in arguments):
 
-- `LLAMA_CACHE` (cache directory, used by `--hf-repo`)
+- `LLAMA_CACHE`: cache directory, used by `--hf-repo`
-- `HF_TOKEN` (Hugging Face access token, used when accessing a gated model with `--hf-repo`)
+- `HF_TOKEN`: Hugging Face access token, used when accessing a gated model with `--hf-repo`
-- `LLAMA_ARG_MODEL`
+- `LLAMA_ARG_MODEL`: equivalent to `-m`
-- `LLAMA_ARG_THREADS`
+- `LLAMA_ARG_MODEL_URL`: equivalent to `-mu`
-- `LLAMA_ARG_CTX_SIZE`
+- `LLAMA_ARG_MODEL_ALIAS`: equivalent to `-a`
-- `LLAMA_ARG_N_PARALLEL`
+- `LLAMA_ARG_HF_REPO`: equivalent to `--hf-repo`
-- `LLAMA_ARG_BATCH`
+- `LLAMA_ARG_HF_FILE`: equivalent to `--hf-file`
-- `LLAMA_ARG_UBATCH`
+- `LLAMA_ARG_THREADS`: equivalent to `-t`
-- `LLAMA_ARG_N_GPU_LAYERS`
+- `LLAMA_ARG_CTX_SIZE`: equivalent to `-c`
-- `LLAMA_ARG_THREADS_HTTP`
+- `LLAMA_ARG_N_PARALLEL`: equivalent to `-np`
-- `LLAMA_ARG_CHAT_TEMPLATE`
+- `LLAMA_ARG_BATCH`: equivalent to `-b`
-- `LLAMA_ARG_N_PREDICT`
+- `LLAMA_ARG_UBATCH`: equivalent to `-ub`
-- `LLAMA_ARG_ENDPOINT_METRICS`
+- `LLAMA_ARG_N_GPU_LAYERS`: equivalent to `-ngl`
-- `LLAMA_ARG_ENDPOINT_SLOTS`
+- `LLAMA_ARG_THREADS_HTTP`: equivalent to `--threads-http`
-- `LLAMA_ARG_EMBEDDINGS`
+- `LLAMA_ARG_CHAT_TEMPLATE`: equivalent to `--chat-template`
-- `LLAMA_ARG_FLASH_ATTN`
+- `LLAMA_ARG_N_PREDICT`: equivalent to `-n`
-- `LLAMA_ARG_DEFRAG_THOLD`
+- `LLAMA_ARG_ENDPOINT_METRICS`: if set to `1`, it will enable metrics endpoint (equivalent to `--metrics`)
+- `LLAMA_ARG_ENDPOINT_SLOTS`: if set to `0`, it will **disable** slots endpoint (equivalent to `--no-slots`). This feature is enabled by default.
+- `LLAMA_ARG_EMBEDDINGS`: if set to `1`, it will enable embeddings endpoint (equivalent to `--embeddings`)
+- `LLAMA_ARG_FLASH_ATTN`: if set to `1`, it will enable flash attention (equivalent to `-fa`)
+- `LLAMA_ARG_CONT_BATCHING`: if set to `0`, it will **disable** continuous batching (equivalent to `--no-cont-batching`). This feature is enabled by default.
+- `LLAMA_ARG_DEFRAG_THOLD`: equivalent to `-dt`
+- `LLAMA_ARG_HOST`: equivalent to `--host`
+- `LLAMA_ARG_PORT`: equivalent to `--port`
+
+Example usage of docker compose with environment variables:
+
+```yml
+services:
+  llamacpp-server:
+    image: ghcr.io/ggerganov/llama.cpp:server
+    ports:
+      - 8080:8080
+    volumes:
+      - ./models:/models
+    environment:
+      # alternatively, you can use "LLAMA_ARG_MODEL_URL" to download the model
+      LLAMA_ARG_MODEL: /models/my_model.gguf
+      LLAMA_ARG_CTX_SIZE: 4096
+      LLAMA_ARG_N_PARALLEL: 2
+      LLAMA_ARG_ENDPOINT_METRICS: 1  # to disable, either remove or set to 0
+      LLAMA_ARG_PORT: 8080
+```
 
 ## Build
 

src/llama.cpp

@@ -6605,6 +6605,7 @@ static bool llm_load_tensors(
         const int64_t n_embd_gqa    = n_embd_v_gqa;
         const int64_t n_vocab       = hparams.n_vocab;
         const int64_t n_vocab_type  = hparams.n_vocab_type;
+        const int64_t n_rot         = hparams.n_rot;
         const int64_t n_expert      = hparams.n_expert;
         const int64_t n_expert_used = hparams.n_expert_used;
         const int64_t n_ctx_train   = hparams.n_ctx_train;
@@ -6662,7 +6663,7 @@ static bool llm_load_tensors(
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
 
-                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
 
                         if (n_expert == 0) {
                             layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
@@ -8115,8 +8116,8 @@ static bool llm_load_tensors(
 
                         layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + (hparams.n_embd_head_k << 2)});
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + (hparams.n_embd_head_k << 2)});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
 
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
 
@@ -8193,7 +8194,7 @@ static bool llm_load_tensors(
                         layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                         layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});