diff --git a/Cargo.toml b/Cargo.toml
index e2baa7e..ed476c1 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,4 +1,4 @@
-workspace = { members = ["llama"] }
+workspace = { members = [ "api", "common","llama"] }
 [package]
 name = "openllama"
 version = "0.1.0"
diff --git a/api/Cargo.toml b/api/Cargo.toml
new file mode 100644
index 0000000..d968b7b
--- /dev/null
+++ b/api/Cargo.toml
@@ -0,0 +1,8 @@
+[package]
+name = "api"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
diff --git a/api/src/ggml.rs b/api/src/ggml.rs
new file mode 100644
index 0000000..005d792
--- /dev/null
+++ b/api/src/ggml.rs
@@ -0,0 +1,3 @@
+macro_rules! ggml_file_magic {
+    () => {0x67676d6c};
+}
\ No newline at end of file
diff --git a/api/src/ggml_backend.rs b/api/src/ggml_backend.rs
new file mode 100644
index 0000000..ec8e960
--- /dev/null
+++ b/api/src/ggml_backend.rs
@@ -0,0 +1 @@
+type GgmlBackendSchedEvalCallback = fn(struct ggml_tensor * t, bool ask, void * user_data) -> bool;
\ No newline at end of file
diff --git a/api/src/lib.rs b/api/src/lib.rs
new file mode 100644
index 0000000..b2d15b8
--- /dev/null
+++ b/api/src/lib.rs
@@ -0,0 +1,18 @@
+pub mod llama;
+mod ggml_backend;
+mod ggml;
+
+pub fn add(left: usize, right: usize) -> usize {
+    left + right
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn it_works() {
+        let result = add(2, 2);
+        assert_eq!(result, 4);
+    }
+}
diff --git a/api/src/llama.rs b/api/src/llama.rs
new file mode 100644
index 0000000..6b0631b
--- /dev/null
+++ b/api/src/llama.rs
@@ -0,0 +1,254 @@
+use crate::llama;
+
+pub const LLAMA_DEFAULT_SEED: i32 = 0xFFFFFFFF;
+
+pub type LlamaPos = i32;
+pub type LlamaToken = i32;
+pub type LlamaSeqId = i32;
+
+
+pub enum LlamaVocabType {
+  LlamaVocabTypeNone = 0, // For models without vocab
+  LlamaVocabTypeSpm = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
+  LlamaVocabTypeBpe = 2, // GPT-2 tokenizer based on byte-level BPE
+  LlamaVocabTypeWpm = 3, // BERT tokenizer based on WordPiece
+}
+
+// pre-tokenization types
+pub enum LlamaVocabPreType {
+  LlamaVocabPreTypeDefault = 0,
+  LlamaVocabPreTypeLlama3 = 1,
+  LlamaVocabPreTypeDeepSeekLlm = 2,
+  LlamaVocabPreTypeDeepSeekCoder = 3,
+  LlamaVocabPreTypeFalcon = 4,
+  LlamaVocabPreTypeMpt = 5,
+  LlamaVocabPreTypeStarCoder = 6,
+  LlamaVocabPreTypeGpt2 = 7,
+  LlamaVocabPreTypeRefactor = 8,
+  LlamaVocabPreTypeCommandR = 9,
+  LlamaVocabPreTypeStableLm2 = 10,
+  LlamaVocabPreTypeQwen2 = 11,
+  LlamaVocabPreTypeOlmo = 12,
+  LlamaVocabPreTypeDbrx = 13,
+}
+
+// note: these values should be synchronized with ggml_rope
+// TODO: maybe move this enum to ggml.h (ggml_rope_type)
+pub enum LlamaRopeType {
+  LlamaRopeTypeNone = -1,
+  LlamaRopeTypeNorm = 0,
+  LlamaRopeTypeNeox = 2,
+  LlamaRopeTypeGlm = 4,
+}
+
+pub enum LlamaTokenType {
+  LlamaTokenTypeUndefined = 0,
+  LlamaTokenTypeNormal = 1,
+  LlamaTokenTypeUnknown = 2,
+  LlamaTokenTypeControl = 3,
+  LlamaTokenTypeUserDefined = 4,
+  LlamaTokenTypeUnused = 5,
+  LlamaTokenTypeByte = 6,
+}
+
+// model file types
+pub enum LlamaFtype {
+  LlamaFtypeAllF32 = 0,
+  LlamaFtypeMostlyF16 = 1,  // except 1d tensors
+  LlamaFtypeMostlyQ40 = 2,  // except 1d tensors
+  LlamaFtypeMostlyQ41 = 3,  // except 1d tensors
+  LlamaFtypeMostlyQ41SomeF16 = 4,  // tok_embeddings.weight and output.weight are F16
+  // LlamaFtypeMostlyQ42 = 5,  // support has been removed
+  // LlamaFtypeMostlyQ43 = 6,  // support has been removed
+  LlamaFtypeMostlyQ80 = 7,  // except 1d tensors
+  LlamaFtypeMostlyQ50 = 8,  // except 1d tensors
+  LlamaFtypeMostlyQ51 = 9,  // except 1d tensors
+  LlamaFtypeMostlyQ2K = 10, // except 1d tensors
+  LlamaFtypeMostlyQ3KS = 11, // except 1d tensors
+  LlamaFtypeMostlyQ3KM = 12, // except 1d tensors
+  LlamaFtypeMostlyQ3KL = 13, // except 1d tensors
+  LlamaFtypeMostlyQ4KS = 14, // except 1d tensors
+  LlamaFtypeMostlyQ4KM = 15, // except 1d tensors
+  LlamaFtypeMostlyQ5KS = 16, // except 1d tensors
+  LlamaFtypeMostlyQ5KM = 17, // except 1d tensors
+  LlamaFtypeMostlyQ6K = 18, // except 1d tensors
+  LlamaFtypeMostlyIq2Xxs = 19, // except 1d tensors
+  LlamaFtypeMostlyIq2Xs = 20, // except 1d tensors
+  LlamaFtypeMostlyQ2KS = 21, // except 1d tensors
+  LlamaFtypeMostlyIq3Xs = 22, // except 1d tensors
+  LlamaFtypeMostlyIq3Xxs = 23, // except 1d tensors
+  LlamaFtypeMostlyIq1S = 24, // except 1d tensors
+  LlamaFtypeMostlyIq4Nl = 25, // except 1d tensors
+  LlamaFtypeMostlyIq3S = 26, // except 1d tensors
+  LlamaFtypeMostlyIq3M = 27, // except 1d tensors
+  LlamaFtypeMostlyIq2S = 28, // except 1d tensors
+  LlamaFtypeMostlyIq2M = 29, // except 1d tensors
+  LlamaFtypeMostlyIq4Xs = 30, // except 1d tensors
+  LlamaFtypeMostlyIq1M = 31, // except 1d tensors
+  LlamaFtypeMostlyBf16 = 32, // except 1d tensors
+
+  LlamaFtypeGuessed = 1024, // not specified in the model file
+}
+
+pub enum LlamaRopeScalingType {
+  LlamaRopeScalingTypeUnspecified = -1,
+  LlamaRopeScalingTypeNone = 0,
+  LlamaRopeScalingTypeLinear = 1,
+  LlamaRopeScalingTypeYarn = 2,
+  LlamaRopeScalingTypeMaxValue = LlamaRopeScalingType::LlamaRopeScalingTypeYarn::into(),
+}
+
+pub enum LlamaPoolingType {
+  LlamaPoolingTypeUnspecified = -1,
+  LlamaPoolingTypeNone = 0,
+  LlamaPoolingTypeMean = 1,
+  LlamaPoolingTypeCls = 2,
+}
+
+pub enum LlamaSplitMode {
+  LlamaSplitModeNone = 0, // single GPU
+  LlamaSplitModeLayer = 1, // split layers and KV across GPUs
+  LlamaSplitModeRow = 2, // split rows across GPUs
+}
+
+pub struct LlamaTokenData {
+  pub id: llama::LlamaToken, // token id
+  pub logit: f64,    // log-odds of the token
+  pub p: f64,        // probability of the token
+}
+
+pub struct LlamaTokenDataArray<'a> {
+  pub data: &'a LlamaTokenData,
+  pub size: usize,
+  pub sorted: bool,
+}
+
+pub type LlamaProgressCallback = fn(progress: f64, data: &[i8]) -> bool;
+
+
+// Input data for llama_decode
+// A LlamaBatch object can contain input about one or many sequences
+// The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+//
+// - token  : the token ids of the input (used when embd is NULL)
+// - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+// - pos    : the positions of the respective token in the sequence
+// - seq_id : the sequence to which the respective token belongs
+// - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+//
+pub struct LlamaBatch {
+  n_tokens: i32,
+
+  token: llama::LlamaToken,
+  embd: f32,
+  pos: llama::LlamaPos,
+  n_seq_id: i32,
+  seq_id: llama::LlamaSeqId,
+  logits: i8, // TODO: rename this to "output"
+
+  // NOTE: helpers for smooth API transition - can be deprecated in the future
+  //       for future-proof code, use the above fields instead and ignore everything below
+  //
+  // pos[i] = all_pos_0 + i*all_pos_1
+  //
+  all_pos_0: llama::LlamaPos,  // used if pos == NULL
+  all_pos_1: llama::LlamaPos,  // used if pos == NULL
+  all_seq_id: llama::LlamaSeqId, // used if seq_id == NULL
+}
+
+pub enum LlamaModelKvOverrideType {
+  LlamaKvOverrideTypeInt,
+  LlamaKvOverrideTypeFloat,
+  LlamaKvOverrideTypeBool,
+  LlamaKvOverrideTypeStr,
+}
+
+
+pub union LlamaModelValueOverride {
+  val_i64: i64,
+  val_f64: f64,
+  val_bool: bool,
+  val_str: [char; 128],
+}
+
+pub struct LlamaModelKvOverride {
+  tag: LlamaModelKvOverrideType,
+  key: [char; 128],
+  value: LlamaModelValueOverride,
+}
+
+pub struct LlamaModelParams {
+  n_gpu_layers: i32, // number of layers to store in VRAM
+  split_mode: LlamaSplitMode, // how to split the model across multiple GPUs
+
+  // main_gpu interpretation depends on split_mode:
+  // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+  // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+  // LLAMA_SPLIT_LAYER: ignored
+  main_gpu: i32,
+
+  // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
+  tensor_split: Option<f64>,
+
+  // comma separated list of RPC servers to use for offloading
+  rpc_servers: Option<char>,
+
+  // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+  // If the provided progress_callback returns true, model loading continues.
+  // If it returns false, model loading is immediately aborted.
+  progress_callback: LlamaProgressCallback,
+
+  // context pointer passed to the progress callback
+  progress_callback_user_data: Option<[i8]>,
+
+  // override key-value pairs of the model meta data
+  kv_overrides: Option<LlamaModelKvOverride>,
+  // Keep the booleans together to avoid misalignment during copy-by-value.
+  vocab_only: bool,    // only load the vocabulary, no weights
+  use_mmap: bool,      // use mmap if possible
+  use_mlock: bool,     // force system to keep model in RAM
+  check_tensors: bool, // validate model tensor data
+}
+
+
+pub struct LlamaContextParams {
+  pub seed: u32,              // RNG seed, -1 for random
+  pub n_ctx: u32,             // text context, 0 = from model
+  pub n_batch: u32,           // logical maximum batch size that can be submitted to llama_decode
+  pub n_ubatch: u32,          // physical maximum batch size
+  pub n_seq_max: u32,         // max number of sequences (i.e. distinct states for recurrent models)
+  pub n_threads: u32,         // number of threads to use for generation
+  pub n_threads_batch: u32,   // number of threads to use for batch processing
+
+  rope_scaling_type: LlamaRopeScalingType, // RoPE scaling type, from `enum llama_rope_scaling_type`
+  pooling_type: LlamaPoolingType,      // whether to pool (sum) embedding results by sequence id
+  // (ignored if no pooling layer)
+
+  // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+  pub rope_freq_base: f64,   // RoPE base frequency, 0 = from model
+  pub rope_freq_scale: f64,  // RoPE frequency scaling factor, 0 = from model
+  pub yarn_ext_factor: f64,  // YaRN extrapolation mix factor, negative = from model
+  pub yarn_attn_factor: f64, // YaRN magnitude scaling factor
+  pub yarn_beta_fast: f64,   // YaRN low correction dim
+  pub yarn_beta_slow: f64,   // YaRN high correction dim
+  pub yarn_orig_ctx: u32,    // YaRN original context size
+  pub defrag_thold: f64,     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+
+   cb_eval:,
+  void * cb_eval_user_data;
+
+  enum ggml_type type_k; // data type for K cache
+  enum ggml_type type_v; // data type for V cache
+
+  // Keep the booleans together to avoid misalignment during copy-by-value.
+  pub logits_all: bool,  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
+  pub embeddings: bool,  // if true, extract embeddings (together with logits)
+  pub offload_kqv: bool, // whether to offload the KQV ops (including the KV cache) to GPU
+  pub flash_attn: bool,  // whether to use flash attention
+
+  // Abort callback
+  // if it returns true, execution of llama_decode() will be aborted
+  // currently works only with CPU execution
+  ggml_abort_callback abort_callback;
+  void *              abort_callback_data;
+}
\ No newline at end of file
diff --git a/common/Cargo.toml b/common/Cargo.toml
new file mode 100644
index 0000000..e99aa0b
--- /dev/null
+++ b/common/Cargo.toml
@@ -0,0 +1,10 @@
+[package]
+name = "common"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+
+api = { path = "../api" }
\ No newline at end of file
diff --git a/common/src/common.rs b/common/src/common.rs
new file mode 100644
index 0000000..e77b9a5
--- /dev/null
+++ b/common/src/common.rs
@@ -0,0 +1,4 @@
+use api::llama;
+
+
+pub struct GPTParams {}
\ No newline at end of file
diff --git a/common/src/lib.rs b/common/src/lib.rs
new file mode 100644
index 0000000..b32ff94
--- /dev/null
+++ b/common/src/lib.rs
@@ -0,0 +1,16 @@
+mod common;
+
+pub fn add(left: usize, right: usize) -> usize {
+    left + right
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn it_works() {
+        let result = add(2, 2);
+        assert_eq!(result, 4);
+    }
+}
diff --git a/llama/src/lib.rs b/llama/src/lib.rs
index 8b13789..356099b 100644
--- a/llama/src/lib.rs
+++ b/llama/src/lib.rs
@@ -1 +1 @@
-
+mod llama;
\ No newline at end of file
diff --git a/llama/src/llama.rs b/llama/src/llama.rs
index 356099b..e69de29 100644
--- a/llama/src/llama.rs
+++ b/llama/src/llama.rs
@@ -1 +0,0 @@
-mod llama;
\ No newline at end of file