Merge branch 'master' into pr/8836

Author: Nexesenex

README.md

@@ -106,6 +106,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
 - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
+- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 

common/common.cpp

@@ -77,6 +77,41 @@
 
 using json = nlohmann::ordered_json;
 
+//
+// Environment variable utils
+//
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, std::string>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::string(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<!std::is_same<T, bool>::value && std::is_integral<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stoi(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_floating_point<T>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    target = value ? std::stof(value) : target;
+}
+
+template<typename T>
+static typename std::enable_if<std::is_same<T, bool>::value, void>::type
+get_env(std::string name, T & target) {
+    char * value = std::getenv(name.c_str());
+    if (value) {
+        std::string val(value);
+        target = val == "1" || val == "true";
+    }
+}
+
 //
 // CPU utils
 //
@@ -220,12 +255,6 @@ int32_t cpu_get_num_math() {
 // CLI argument parsing
 //
 
-void gpt_params_handle_hf_token(gpt_params & params) {
-    if (params.hf_token.empty() && std::getenv("HF_TOKEN")) {
-        params.hf_token = std::getenv("HF_TOKEN");
-    }
-}
-
 void gpt_params_handle_model_default(gpt_params & params) {
     if (!params.hf_repo.empty()) {
         // short-hand to avoid specifying --hf-file -> default it to --model
@@ -273,7 +302,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 
     gpt_params_handle_model_default(params);
 
-    gpt_params_handle_hf_token(params);
+    if (params.hf_token.empty()) {
+        get_env("HF_TOKEN", params.hf_token);
+    }
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -293,6 +324,25 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     return true;
 }
 
+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_THREADS",          params.n_threads);
+    get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
+    get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
+    get_env("LLAMA_ARG_BATCH",            params.n_batch);
+    get_env("LLAMA_ARG_UBATCH",           params.n_ubatch);
+    get_env("LLAMA_ARG_N_GPU_LAYERS",     params.n_gpu_layers);
+    get_env("LLAMA_ARG_THREADS_HTTP",     params.n_threads_http);
+    get_env("LLAMA_ARG_CHAT_TEMPLATE",    params.chat_template);
+    get_env("LLAMA_ARG_N_PREDICT",        params.n_predict);
+    get_env("LLAMA_ARG_ENDPOINT_METRICS", params.endpoint_metrics);
+    get_env("LLAMA_ARG_ENDPOINT_SLOTS",   params.endpoint_slots);
+    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);
+}
+
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     const auto params_org = params; // the example can modify the default params
 

common/common.h

@@ -267,7 +267,7 @@ struct gpt_params {
     std::string lora_outfile = "ggml-lora-merged-f16.gguf";
 };
 
-void gpt_params_handle_hf_token(gpt_params & params);
+void gpt_params_parse_from_env(gpt_params & params);
 void gpt_params_handle_model_default(gpt_params & params);
 
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);

convert_hf_to_gguf.py

@@ -295,6 +295,7 @@ def prepare_tensors(self):
                             gguf.MODEL_TENSOR.FFN_GATE_INP,
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
+                            gguf.MODEL_TENSOR.SSM_CONV1D,
                         )
                     )
                     or not name.endswith(".weight")
@@ -2711,7 +2712,7 @@ class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
 
 
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
+@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
@@ -2742,20 +2743,24 @@ def set_gguf_parameters(self):
         # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
         dt_rank      = self.find_hparam(["time_step_rank",     "dt_rank"],      optional=True) or -(d_model // -16)
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
+        use_dt_b_c_norm = False
+        # For falconmamba we do apply RMS norm on B / DT and C layers
+        if self.find_hparam(["model_type"], optional=True) in ("falcon_mamba",):
+            use_dt_b_c_norm = True
         # Fail early for models which don't have a block expansion factor of 2
         assert d_inner == 2 * d_model
 
         self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
         self.gguf_writer.add_embedding_length(d_model)
         self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
         self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_ssm_conv_kernel(d_conv)
         self.gguf_writer.add_ssm_inner_size(d_inner)
         self.gguf_writer.add_ssm_state_size(d_state)
         self.gguf_writer.add_ssm_time_step_rank(dt_rank)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_ssm_dt_b_c_rms(use_dt_b_c_norm) # For classic Mamba we don't apply rms norm on B / DT layers
         self.gguf_writer.add_file_type(self.ftype)
 
     _tok_embd = None
@@ -2782,23 +2787,6 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
 
         return [(new_name, data_torch)]
 
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        if bid is not None and new_name in (
-            self.format_tensor_name(
-                n, bid, ".weight" if name.endswith(".weight") else ""
-            )
-            for n in [
-                gguf.MODEL_TENSOR.SSM_CONV1D,
-                gguf.MODEL_TENSOR.SSM_X,
-                gguf.MODEL_TENSOR.SSM_DT,
-                gguf.MODEL_TENSOR.SSM_A,
-                gguf.MODEL_TENSOR.SSM_D,
-            ]
-        ):
-            return gguf.GGMLQuantizationType.F32
-
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
 
 @Model.register("CohereForCausalLM")
 class CommandR2Model(Model):
@@ -3792,7 +3780,7 @@ class ExaoneModel(Model):
     def set_gguf_parameters(self):
         hparams = self.hparams
 
-        assert(hparams["activation_function"] == "silu")
+        assert (hparams["activation_function"] == "silu")
 
         max_position_embeddings = hparams["max_position_embeddings"]
         embed_dim = hparams["hidden_size"]
@@ -3855,8 +3843,8 @@ def prepare_tensors(self):
 
         super().prepare_tensors()
 
-###### CONVERSION LOGIC ######
 
+###### CONVERSION LOGIC ######
 
 # tree of lazy tensors
 class LazyTorchTensor(gguf.LazyBase):

examples/llava/clip.cpp

@@ -20,6 +20,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -1108,7 +1112,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx;
+    clip_ctx * new_clip = new clip_ctx{};
 
     // update projector type
     {
@@ -1142,6 +1146,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     LOG_TEE("%s: CLIP using CANN backend\n", __func__);
 #endif
 
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_TEE("%s: CLIP using Vulkan backend\n", __func__);
+#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();

examples/server/README.md

@@ -247,6 +247,25 @@ logging:
          --log-append             Don't truncate the old log file.
 ```
 
+Available environment variables (if specified, these variables will override parameters specified in arguments):
+
+- `LLAMA_CACHE` (cache directory, used by `--hf-repo`)
+- `HF_TOKEN` (Hugging Face access token, used when accessing a gated model with `--hf-repo`)
+- `LLAMA_ARG_MODEL`
+- `LLAMA_ARG_THREADS`
+- `LLAMA_ARG_CTX_SIZE`
+- `LLAMA_ARG_N_PARALLEL`
+- `LLAMA_ARG_BATCH`
+- `LLAMA_ARG_UBATCH`
+- `LLAMA_ARG_N_GPU_LAYERS`
+- `LLAMA_ARG_THREADS_HTTP`
+- `LLAMA_ARG_CHAT_TEMPLATE`
+- `LLAMA_ARG_N_PREDICT`
+- `LLAMA_ARG_ENDPOINT_METRICS`
+- `LLAMA_ARG_ENDPOINT_SLOTS`
+- `LLAMA_ARG_EMBEDDINGS`
+- `LLAMA_ARG_FLASH_ATTN`
+- `LLAMA_ARG_DEFRAG_THOLD`
 
 ## Build
 

examples/server/server.cpp

@@ -2507,6 +2507,9 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    // parse arguments from environment variables
+    gpt_params_parse_from_env(params);
+
     // TODO: not great to use extern vars
     server_log_json = params.log_json;
     server_verbose = params.verbosity > 0;

ggml/src/ggml-sycl.cpp

@@ -893,43 +893,6 @@ static void clamp_f32(const float * x, float * dst, const float min, const float
     dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
 }
 
-template <typename T>
-static void im2col_kernel(const float *x, T *dst, int offset_delta,
-                           int IW, int IH, int OW, int KW, int KH,
-                           int pelements, int CHW, int s0, int s1, int p0,
-                           int p1, int d0, int d1,
-                           const sycl::nd_item<3> &item_ct1) {
-    const int i = item_ct1.get_local_id(2) +
-                  item_ct1.get_group(2) * item_ct1.get_local_range(2);
-    if (i >= pelements) {
-        return;
-    }
-
-    const int ksize = OW * (KH > 1 ? KW : 1);
-    const int kx = i / ksize;
-    const int kd = kx * ksize;
-    const int ky = (i - kd) / OW;
-    const int ix = i % OW;
-
-    const int64_t iiw = ix * s0 + kx * d0 - p0;
-    const int64_t iih = item_ct1.get_group(1) * s1 + ky * d1 - p1;
-
-    const int64_t offset_dst =
-        (item_ct1.get_group(1) * OW + ix) * CHW +
-        (item_ct1.get_group(0) * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] =
-            sycl::vec<float, 1>(0.0f)
-                .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
-    } else {
-        const int64_t offset_src = item_ct1.get_group(0) * offset_delta;
-        dst[offset_dst] =
-            sycl::vec<float, 1>(x[offset_src + iih * IW + iiw])
-                .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
-    }
-}
-
 template <typename Ti, typename To>
 static  void pool2d_nchw_kernel(
         const int ih, const int iw, const int oh, const int ow,
@@ -1742,32 +1705,6 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst,
                          });
 }
 
-template <typename T>
-static void im2col_sycl(const float *x, T *dst, int IW, int IH,
-                                int OW, int OH, int KW, int KH, int IC,
-                                int offset_delta, int s0, int s1, int p0,
-                                int p1, int d0, int d1,
-                                queue_ptr stream) {
-    const int parallel_elements = OW * KW * KH;
-    const int num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
-    sycl::range<3> block_nums(IC, OH, num_blocks);
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums *
-                                  sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                im2col_kernel(x, dst, offset_delta, IW, IH, OW, KW, KH,
-                               parallel_elements, (IC * KH * KW), s0, s1, p0,
-                               p1, d0, d1, item_ct1);
-            });
-    }
-}
-
-
 static bool g_sycl_loaded = false;
 
 bool ggml_sycl_loaded(void) {
@@ -2636,47 +2573,6 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
     (void) src1_dd;
 }
 
-inline void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                const ggml_tensor *src1, ggml_tensor *dst,
-                                const float *src0_dd, const float *src1_dd,
-                                float *dst_dd,
-                                const queue_ptr &main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
-
-    const int64_t IC = src1->ne[is_2D ? 2 : 1];
-    const int64_t IH = is_2D ? src1->ne[1] : 1;
-    const int64_t IW =         src1->ne[0];
-
-    const int64_t KH = is_2D ? src0->ne[1] : 1;
-    const int64_t KW =         src0->ne[0];
-
-    const int64_t OH = is_2D ? dst->ne[2] : 1;
-    const int64_t OW =         dst->ne[1];
-
-    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-
-    if (dst->type == GGML_TYPE_F16) {
-        im2col_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
-    } else {
-        im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
-    }
-
-    (void) src0;
-    (void) src0_dd;
-}
-
 inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                   const ggml_tensor *src1, ggml_tensor *dst,
                                   const float *src0_dd, const float *src1_dd,
@@ -3581,7 +3477,8 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     bool use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE
+        && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda || src1->ne[1] > MMVQ_MIN_BATCH_SIZE);
 
     bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;