feat(multimodal): let vLLM workers expand Kimi-K3 and GLM-5.3-Flash image anchors

crates/multimodal/scripts/check_worker_anchor_parity.py

@@ -50,6 +50,17 @@
             "video": ("config_id", "video_token_index"),
         },
     },
+    "kimi_k3": {
+        "model_id": "moonshotai/Kimi-K3",
+        # Anchor = tokenizer token for the config's media_placeholder_token_id.
+        "anchors": {"image": ("config_id", "media_placeholder_token_id")},
+    },
+    "glm53_flash": {
+        "model_id": "zai-org/GLM-5.3-Flash",
+        # Anchor = tokenizer token for the config's image_token_id. Video is
+        # deliberately absent: vLLM targets a three-token block for it.
+        "anchors": {"image": ("config_id", "image_token_id")},
+    },
 }
 
 PREFIX_TEXT = "Describe what you see."

crates/multimodal/src/registry/glm53_flash.rs

@@ -119,6 +119,12 @@ impl ModelProcessorSpec for Glm53FlashSpec {
         }
     }
 
+    /// `<|image|>` is vLLM's single-token image target; video is not, its
+    /// target is the three-token `<|begin_of_video|>…<|end_of_video|>` block.
+    fn worker_expandable(&self, modality: Modality) -> bool {
+        matches!(modality, Modality::Image)
+    }
+
     fn modality_limits(
         &self,
         metadata: &ModelMetadata,
@@ -403,4 +409,12 @@ mod tests {
             Err(ModelRegistryError::InvalidPreprocessedField { .. })
         ));
     }
+
+    #[test]
+    fn only_the_image_anchor_can_be_expanded_by_the_worker() {
+        let spec = Glm53FlashSpec;
+        assert!(spec.worker_expandable(Modality::Image));
+        assert!(!spec.worker_expandable(Modality::Video));
+        assert!(!spec.worker_expandable(Modality::Audio));
+    }
 }

crates/multimodal/src/registry/kimi_k3.rs

@@ -82,6 +82,12 @@ impl ModelProcessorSpec for KimiK3VisionSpec {
         Self::pad_token_id(metadata)
     }
 
+    /// `<|media_pad|>` is the single token vLLM's `kimi_k3` prompt updates
+    /// replace with the dimensioned media block.
+    fn worker_expandable(&self, modality: Modality) -> bool {
+        matches!(modality, Modality::Image)
+    }
+
     fn modality_limits(
         &self,
         _metadata: &ModelMetadata,
@@ -279,6 +285,16 @@ mod tests {
         );
     }
 
+    #[test]
+    fn image_anchor_can_be_expanded_by_the_worker() {
+        use crate::registry::ModelProcessorSpec;
+
+        let spec = super::KimiK3VisionSpec;
+        assert!(spec.worker_expandable(Modality::Image));
+        assert!(!spec.worker_expandable(Modality::Video));
+        assert!(!spec.worker_expandable(Modality::Audio));
+    }
+
     #[test]
     fn kimi_k3_dimensions_are_per_image() {
         let tokenizer = k3_tokenizer();