📊 feat: distinguish compute density and communication density in DensityObserver

- Add record_comm_density() call in select_blocks to track CPU block selection
- Add get_per_layer_comm_density() method for detailed analysis
- Update print_summary() to show both densities and H2D savings ratio
- Set DensityObserver mode (offload/gpu_only) in test_ruler.py
- Update get_summary() to return both density types

Key insight: Comm density can be 100% even when compute density is ~37%
because sparse BSA blocks are distributed across all CPU blocks.
Since CPU block granularity is 32x coarser (4096 vs 128 tokens),
any() aggregation across heads/Q-blocks results in all CPU blocks being needed.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

nanovllm/kvcache/sparse/xattn_bsa.py

@@ -905,6 +905,15 @@ class XAttentionBSAPolicy(SparsePolicy):
             self._stats_total_selected_blocks += len(selected_block_ids)
             self._stats_num_chunks += 1
 
+            # Record communication density to DensityObserver
+            # Comm density = selected_cpu_blocks / available_cpu_blocks
+            # This is different from compute density (BSA block granularity)
+            DensityObserver.record_comm_density(
+                layer_id=layer_id,
+                selected_cpu_blocks=len(selected_block_ids),
+                total_cpu_blocks=len(available_blocks),
+            )
+
             # Log per-chunk density
             chunk_density = len(selected_block_ids) / len(available_blocks)
             logger.debug(f"[XAttn] chunk={ctx.query_chunk_idx}, available={len(available_blocks)}, "

nanovllm/utils/density_observer.py

@@ -266,14 +266,31 @@ class DensityObserver(Observer):
             return 0.0
         return sum(all_densities) / len(all_densities)
 
+    @classmethod
+    def get_per_layer_comm_density(cls) -> Dict[int, float]:
+        """
+        获取每层的 communication density (CPU block 粒度)。
+
+        Returns:
+            Dict[layer_id, avg_comm_density]
+        """
+        result = {}
+        for layer_id, densities in cls._layer_comm_densities.items():
+            if densities:
+                result[layer_id] = sum(densities) / len(densities)
+        return result
+
     @classmethod
     def get_summary(cls) -> dict:
         """返回统计摘要"""
         per_layer = cls.get_per_layer_density()
+        per_layer_comm = cls.get_per_layer_comm_density()
         return {
             "mode": cls._mode,
-            "overall_density": cls.get_overall_density(),
+            "overall_compute_density": cls.get_overall_density(),
-            "per_layer_density": per_layer,
+            "overall_comm_density": cls.get_overall_comm_density(),
+            "per_layer_compute_density": per_layer,
+            "per_layer_comm_density": per_layer_comm,
             "num_layers": len(per_layer),
             "last_mask_shape": {
                 "q_blocks": cls._last_q_blocks,
@@ -301,7 +318,9 @@ class DensityObserver(Observer):
         print(f"[DensityObserver] Mode: {cls._mode}")
         print(f"  Compute density: {overall:.4f} (min: {min_density:.4f} @ layer {min_layer})")
         if overall_comm > 0:
-            print(f"  Comm density: {overall_comm:.4f}")
+            # Offload mode: show both densities with explanation
+            print(f"  Comm density:    {overall_comm:.4f} (CPU block granularity)")
+            print(f"  Savings ratio:   {1 - overall_comm:.1%} H2D transfer reduction")
         print(f"  Num layers: {len(per_layer)}")
         # 输出 layer 0 的 density 用于对比
         if 0 in per_layer:

tests/test_ruler.py

@@ -386,8 +386,11 @@ def run_ruler_benchmark(
     if sparse_policy and sparse_policy.upper() == "XATTN_BSA":
         DensityObserver.enable()
         DensityObserver.complete_reset()
+        # Set mode for correct density interpretation
+        DensityObserver.set_mode("offload" if enable_cpu_offload else "gpu_only")
         if not json_output:
-            print("[DensityObserver] Enabled for XAttention BSA")
+            mode_str = "offload" if enable_cpu_offload else "gpu_only"
+            print(f"[DensityObserver] Enabled for XAttention BSA (mode: {mode_str})")
 
     # LLM initialization kwargs
     llm_kwargs = {