Implement Qwen text core

src/engine/metal/checkpoint.rs

@@ -320,6 +320,7 @@ impl Executor {
         match self {
             Self::DeepSeek(executor) => executor.save_checkpoint(path, tag, progress),
             Self::Glm(executor) => executor.save_checkpoint(path, tag, progress),
+            Self::Qwen(executor) => executor.save_checkpoint(path, tag, progress),
         }
     }
 
@@ -331,6 +332,7 @@ impl Executor {
         match self {
             Self::DeepSeek(executor) => executor.load_checkpoint(path, progress),
             Self::Glm(executor) => executor.load_checkpoint(path, progress),
+            Self::Qwen(executor) => executor.load_checkpoint(path, progress),
         }
     }
 }

src/engine/metal/gpu.rs

@@ -5,6 +5,22 @@ pub(super) struct GpuTensor {
     _private: [u8; 0],
 }
 
+#[derive(Clone, Copy)]
+#[repr(C)]
+pub(super) struct QwenWeightView {
+    pub(super) map: *const c_void,
+    pub(super) size: u64,
+    pub(super) offset: u64,
+    pub(super) bytes: u64,
+}
+
+#[derive(Clone, Copy, Default)]
+#[repr(C)]
+pub(super) struct QwenKernelArgs {
+    pub(super) u: [u32; 16],
+    pub(super) f: [f32; 8],
+}
+
 #[derive(Clone, Copy, Default)]
 #[repr(C)]
 pub(super) struct Glm53VisionLayerWeights {
@@ -81,6 +97,18 @@ unsafe extern "C" {
         map_size: u64,
         max_tensor_bytes: u64,
     ) -> i32;
+    pub(super) fn ds4_gpu_qwen_dispatch(
+        kernel: *const c_char,
+        out: *mut GpuTensor,
+        a: *const GpuTensor,
+        b: *const GpuTensor,
+        c: *const GpuTensor,
+        weights: *const QwenWeightView,
+        weight_count: u32,
+        args: *const QwenKernelArgs,
+        grid_x: u32,
+        grid_y: u32,
+    ) -> i32;
     pub(super) fn ds4_gpu_set_transient_model_map_range(
         model_map: *const c_void,
         model_size: u64,
@@ -1804,7 +1832,7 @@ unsafe extern "C" {
 }
 
 pub(super) struct Context {
-    _model_file: File,
+    _model_file: Option<File>,
 }
 
 impl Context {
@@ -1908,9 +1936,29 @@ impl Context {
             }
         }
         Ok(Self {
-            _model_file: model_file,
+            _model_file: Some(model_file),
         })
     }
+
+    pub(super) fn open_qwen(admission_bytes: u64) -> Result<Self, String> {
+        check(unsafe { ds4_gpu_init() }, "Metal initialization")?;
+        unsafe {
+            ds4_gpu_set_glm_model(false);
+            ds4_gpu_set_ssd_streaming(false);
+            ds4_gpu_set_decode_pipeline_fast_lookup(0);
+        }
+        let recommended = unsafe { ds4_gpu_recommended_working_set_size() };
+        if admission_bytes != 0 && recommended != 0 && admission_bytes > recommended {
+            unsafe { ds4_gpu_cleanup() };
+            return Err(format!(
+                "Qwen model load needs {:.1} GiB including context and scratch, but Metal recommends at most {:.1} GiB",
+                admission_bytes as f64 / 1_073_741_824.0,
+                recommended as f64 / 1_073_741_824.0,
+            ));
+        }
+        unsafe { ds4_gpu_set_quality(true) };
+        Ok(Self { _model_file: None })
+    }
 }
 
 impl Drop for Context {