Merge pull request #855 from n-k/main

Add support for adding tensor buffer type overrides

Author: MarcusDunn

examples/simple/src/main.rs

@@ -48,6 +48,9 @@ struct Args {
     #[cfg(any(feature = "cuda", feature = "vulkan"))]
     #[clap(long)]
     disable_gpu: bool,
+    #[cfg(any(feature = "cuda", feature = "vulkan"))]
+    #[arg(long, help = "Keep MoE layers on CPU")]
+    cmoe: bool,
     #[arg(short = 's', long, help = "RNG seed (default: 1234)")]
     seed: Option<u32>,
     #[arg(
@@ -129,6 +132,8 @@ fn main() -> Result<()> {
         file,
         #[cfg(any(feature = "cuda", feature = "vulkan"))]
         disable_gpu,
+        #[cfg(any(feature = "cuda", feature = "vulkan"))]
+        cmoe,
         key_value_overrides,
         seed,
         threads,
@@ -176,6 +181,13 @@ fn main() -> Result<()> {
         model_params.as_mut().append_kv_override(k.as_c_str(), *v);
     }
 
+    #[cfg(any(feature = "cuda", feature = "vulkan"))]
+    {
+        if !disable_gpu && cmoe {
+            model_params.as_mut().add_cpu_moe_override();
+        }
+    }
+
     let model_path = model
         .get_or_load()
         .with_context(|| "failed to get model from args")?;

llama-cpp-2/src/model/params.rs

@@ -13,6 +13,7 @@ pub mod kv_overrides;
 pub struct LlamaModelParams {
     pub(crate) params: llama_cpp_sys_2::llama_model_params,
     kv_overrides: Vec<llama_cpp_sys_2::llama_model_kv_override>,
+    buft_overrides: Vec<llama_cpp_sys_2::llama_model_tensor_buft_override>,
 }
 
 impl Debug for LlamaModelParams {
@@ -107,6 +108,48 @@ impl LlamaModelParams {
     }
 }
 
+impl LlamaModelParams {
+    /// Adds buffer type overides to move all mixture-of-experts layers to CPU.
+    pub fn add_cpu_moe_override(self: Pin<&mut Self>) {
+        self.add_cpu_buft_override(c"\\.ffn_(up|down|gate)_(ch|)exps");
+    }
+
+    /// Appends a buffer type override to the model parameters, to move layers matching pattern to CPU.
+    /// It must be pinned as this creates a self-referential struct.
+    pub fn add_cpu_buft_override(mut self: Pin<&mut Self>, key: &CStr) {
+        let buft_override = self
+            .buft_overrides
+            .get_mut(0)
+            .expect("buft_overrides did not have a next allocated");
+
+        assert!(
+            buft_override.pattern.is_null(),
+            "last buft_override was not empty"
+        );
+
+        // There should be some way to do this without iterating over everything.
+        for (_i, &c) in key.to_bytes_with_nul().iter().enumerate() {
+            c_char::try_from(c).expect("invalid character in key");
+        }
+
+        buft_override.pattern = key.as_ptr();
+        buft_override.buft = unsafe { llama_cpp_sys_2::ggml_backend_cpu_buffer_type() };
+
+        // set to null pointer for panic safety (as push may move the vector, invalidating the pointer)
+        self.params.tensor_buft_overrides = null();
+
+        // push the next one to ensure we maintain the iterator invariant of ending with a 0
+        self.buft_overrides
+            .push(llama_cpp_sys_2::llama_model_tensor_buft_override {
+                pattern: std::ptr::null(),
+                buft: std::ptr::null_mut(),
+            });
+
+        // set the pointer to the (potentially) new vector
+        self.params.tensor_buft_overrides = self.buft_overrides.as_ptr();
+    }
+}
+
 impl LlamaModelParams {
     /// Get the number of layers to offload to the GPU.
     #[must_use]
@@ -199,6 +242,10 @@ impl Default for LlamaModelParams {
                     val_i64: 0,
                 },
             }],
+            buft_overrides: vec![llama_cpp_sys_2::llama_model_tensor_buft_override {
+                pattern: std::ptr::null(),
+                buft: std::ptr::null_mut(),
+            }],
         }
     }
 }