Add end-to-end MXFP4 Metal support

src/engine/metal/glm.rs

@@ -434,13 +434,17 @@ impl GlmExecutor {
             .as_ref()
             .map(|plan| glm_streaming_model_spans(&model, &weights, plan))
             .transpose()?;
+        let context_spans = model_spans
+            .as_ref()
+            .map(|(spans, max_tensor_bytes)| (spans.as_slice(), *max_tensor_bytes));
         let context_handle = Context::open(
             &model,
             quality,
             effective_ssd.enabled,
             admission,
-            model_spans.as_deref(),
+            context_spans,
         )?;
+        let model_spans = model_spans.map(|(spans, _)| spans);
         configure_streaming(&model, &weights, streaming.as_ref())?;
         let scratch = GlmScratch::allocate(&model, context)?;
         let caches = (0..weights.layers.len())
@@ -2851,7 +2855,7 @@ fn glm_streaming_model_spans(
     model: &Model,
     weights: &GlmWeights,
     plan: &GlmStreamingPlan,
-) -> Result<Vec<(u64, u64)>, String> {
+) -> Result<(Vec<(u64, u64)>, u64), String> {
     let full_before = model
         .shape
         .leading_dense
@@ -2890,6 +2894,11 @@ fn glm_streaming_model_spans(
         })
         .map(|(_, tensor)| (tensor.offset, tensor.bytes))
         .collect::<Vec<_>>();
+    let max_tensor_bytes = spans
+        .iter()
+        .map(|(_, bytes)| *bytes)
+        .max()
+        .ok_or("GLM SSD streaming found no resident model tensors")?;
     spans.sort_unstable_by_key(|span| span.0);
     let mut merged: Vec<(u64, u64)> = Vec::new();
     for (offset, bytes) in spans {
@@ -2903,10 +2912,7 @@ fn glm_streaming_model_spans(
         }
         merged.push((offset, bytes));
     }
-    if merged.is_empty() {
-        return Err("GLM SSD streaming found no resident model tensors".into());
-    }
-    Ok(merged)
+    Ok((merged, max_tensor_bytes))
 }
 
 fn glm_layer_model_spans(model: &Model, layer: u32) -> Result<Vec<(u64, u64)>, String> {