Support DeepSeek V4 Flash 0731

src/engine/metal/glm.rs

@@ -401,6 +401,7 @@ impl GlmExecutor {
                 dspark_confidence_threshold: 0.9,
                 dspark_confidence_threshold_set: false,
                 dspark_strict: false,
+                dspark_exact_sampling: false,
             },
             None,
         )
@@ -3412,6 +3413,7 @@ mod tests {
                     dspark_confidence_threshold: 0.9,
                     dspark_confidence_threshold_set: false,
                     dspark_strict: false,
+                    dspark_exact_sampling: false,
                 },
                 None,
             )

src/engine/metal/gpu.rs

@@ -1283,18 +1283,6 @@ unsafe extern "C" {
         hc: u32,
         eps: f32,
     ) -> i32;
-    pub(super) fn ds4_gpu_hc_weighted_sum_norm_tensor(
-        out: *mut GpuTensor,
-        norm: *mut GpuTensor,
-        residual: *const GpuTensor,
-        weights: *const GpuTensor,
-        map: *const c_void,
-        size: u64,
-        norm_weight: u64,
-        embd: u32,
-        hc: u32,
-        eps: f32,
-    ) -> i32;
     pub(super) fn ds4_gpu_hc_weighted_sum_tensor(
         out: *mut GpuTensor,
         residual: *const GpuTensor,
@@ -1443,9 +1431,9 @@ impl Buffer {
     }
 
     pub(super) fn view(&self, offset: u64, bytes: u64) -> Result<Self, String> {
-        // SAFETY: native/metal/ds4_metal.m:7916-7940 bounds-checks the view,
-        // ARC-retains base_obj.buffer at :7926, and marks the view non-owning at
-        // :7929, so it may outlive and be freed independently of this wrapper.
+        // SAFETY: `ds4_gpu_tensor_view` in native/metal/ds4_metal.m bounds-checks
+        // the view, ARC-retains base_obj.buffer, and marks the view non-owning, so
+        // it may outlive and be freed independently of this wrapper.
         // Recheck those guarantees whenever the vendored Metal file is re-synced.
         NonNull::new(unsafe { ds4_gpu_tensor_view(self.raw(), offset, bytes) })
             .map(Self)