Add GLM 5.2 Metal execution

src/engine/metal.rs

@@ -1,6 +1,8 @@
 mod checkpoint;
+mod glm;
 mod gpu;
 
+use glm::GlmExecutor;
 use gpu::*;
 
 use super::gguf::{F16, Gguf, Q8_0, Tensor as GgufTensor};
@@ -563,7 +565,7 @@ impl Session {
 // and `_context` must drop before `model` unmaps memory wrapped without copying
 // by native/metal/ds4_metal.m:10329. This intentionally differs from
 // ../ds4/ds4.c:56287-56288; do not reorder these fields to match it.
-pub(super) struct Executor {
+pub(super) struct DeepSeekExecutor {
     weights: Weights,
     session: Session,
     logits: Vec<f32>,
@@ -576,7 +578,7 @@ pub(super) struct Executor {
     model: Model,
 }
 
-impl Executor {
+impl DeepSeekExecutor {
     pub(super) fn open(
         model: Model,
         context: u32,
@@ -584,7 +586,7 @@ impl Executor {
         prefill_chunk: u32,
     ) -> Result<Self, String> {
         let weights = Weights::bind(&model)?;
-        let context_handle = Context::open(&model, quality)?;
+        let context_handle = Context::open(&model, quality, false, 0)?;
         let session = Session::new(
             &model,
             context,
@@ -849,6 +851,136 @@ impl Executor {
     }
 }
 
+/// Model-family dispatch over the two Rust-owned Metal graphs.
+pub(super) enum Executor {
+    DeepSeek(Box<DeepSeekExecutor>),
+    Glm(Box<GlmExecutor>),
+}
+
+impl Executor {
+    #[allow(dead_code)]
+    pub(super) fn open(
+        model: Model,
+        context: u32,
+        quality: bool,
+        prefill_chunk: u32,
+    ) -> Result<Self, String> {
+        Self::open_configured(
+            model,
+            context,
+            quality,
+            prefill_chunk,
+            crate::settings::EngineSsdSettings {
+                enabled: false,
+                cold: false,
+                cache_experts: 0,
+                cache_bytes: 0,
+                full_layers: 0,
+                full_layers_set: false,
+                preload_experts: 0,
+            },
+        )
+    }
+
+    pub(super) fn open_configured(
+        model: Model,
+        context: u32,
+        quality: bool,
+        prefill_chunk: u32,
+        ssd: crate::settings::EngineSsdSettings,
+    ) -> Result<Self, String> {
+        match model.shape.family {
+            ModelFamily::DeepSeek => DeepSeekExecutor::open(model, context, quality, prefill_chunk)
+                .map(Box::new)
+                .map(Self::DeepSeek),
+            ModelFamily::Glm => GlmExecutor::open(model, context, quality, ssd)
+                .map(Box::new)
+                .map(Self::Glm),
+        }
+    }
+
+    pub(super) fn eval(&mut self, token: i32) -> Result<(), String> {
+        match self {
+            Self::DeepSeek(executor) => executor.eval(token),
+            Self::Glm(executor) => executor.eval(token),
+        }
+    }
+
+    pub(super) fn prefill(
+        &mut self,
+        tokens: &[i32],
+        progress: impl FnMut(u32) -> bool,
+    ) -> Result<usize, String> {
+        match self {
+            Self::DeepSeek(executor) => executor.prefill(tokens, progress),
+            Self::Glm(executor) => executor.prefill(tokens, progress),
+        }
+    }
+
+    pub(super) fn logits(&self) -> &[f32] {
+        match self {
+            Self::DeepSeek(executor) => executor.logits(),
+            Self::Glm(executor) => executor.logits(),
+        }
+    }
+
+    pub(super) fn model(&self) -> &Model {
+        match self {
+            Self::DeepSeek(executor) => executor.model(),
+            Self::Glm(executor) => executor.model(),
+        }
+    }
+
+    pub(super) fn context(&self) -> u32 {
+        match self {
+            Self::DeepSeek(executor) => executor.context(),
+            Self::Glm(executor) => executor.context(),
+        }
+    }
+
+    pub(super) fn position(&self) -> u32 {
+        match self {
+            Self::DeepSeek(executor) => executor.position(),
+            Self::Glm(executor) => executor.position(),
+        }
+    }
+
+    pub(super) fn reset(&mut self) -> Result<(), String> {
+        match self {
+            Self::DeepSeek(executor) => executor.reset(),
+            Self::Glm(executor) => executor.reset(),
+        }
+    }
+
+    pub(super) fn align_prompt(&mut self, tokens: &[i32]) -> Result<usize, String> {
+        match self {
+            Self::DeepSeek(executor) => executor.align_prompt(tokens),
+            Self::Glm(executor) => executor.align_prompt(tokens),
+        }
+    }
+
+    pub(super) fn tokens(&self) -> &[i32] {
+        match self {
+            Self::DeepSeek(executor) => executor.tokens(),
+            Self::Glm(executor) => executor.tokens(),
+        }
+    }
+
+    pub(super) fn checkpoint_tag(&self) -> [u8; 32] {
+        match self {
+            Self::DeepSeek(executor) => executor.checkpoint_tag(),
+            Self::Glm(executor) => executor.checkpoint_tag(),
+        }
+    }
+
+    pub(super) fn note_checkpoint_tag(&mut self, tag: [u8; 32]) {
+        match self {
+            Self::DeepSeek(executor) => executor.note_checkpoint_tag(tag),
+            Self::Glm(executor) => executor.note_checkpoint_tag(tag),
+        }
+    }
+}
+
 #[allow(clippy::too_many_arguments)]
 fn refresh_ratio4_compressor_state(
     s: &BatchScratch,