create demo

Author: dreaming-panda

examples/verify_algo.py

@@ -140,14 +140,16 @@ def verify_algos(
         else:
             unique_result[item['query']] = max(item["score"], unique_result[item['query']])
 
-    
-    llm_cfg = AutoConfig.from_pretrained(model_name)
-    flow = vortex_torch.flow.build_vflow(vortex_module_name) 
-    memory_access_runtime = flow.run_indexer_virtual(
-        group_size=llm_cfg.num_attention_heads // llm_cfg.num_key_value_heads,
-        page_size=page_size,
-        head_dim=llm_cfg.head_dim,
-    )
+    if sparse_attention:
+        llm_cfg = AutoConfig.from_pretrained(model_name)
+        flow = vortex_torch.flow.build_vflow(vortex_module_name) 
+        memory_access_runtime = flow.run_indexer_virtual(
+            group_size=llm_cfg.num_attention_heads // llm_cfg.num_key_value_heads,
+            page_size=page_size,
+            head_dim=llm_cfg.head_dim,
+        )
+    else:
+        memory_access_runtime = 0.0
 
     global_summary = {
         f'mean@{trials}': total_accuracy / count if count > 0 else 0,
@@ -156,7 +158,7 @@ def verify_algos(
         "e2e_time": e2e_time,
         "total_tokens": total_tokens, 
         "throughput": total_tokens / e2e_time,
-        "memory_access_runtime (per page)": memory_access_runtime
+        "auxilary memory_access_runtime (bytes per page)": memory_access_runtime
     }
 
     return global_summary
@@ -229,3 +231,4 @@ def parse_args():
     )
     print(summary)
 
+    exit(0)

examples/verify_algo.sh

@@ -2,10 +2,7 @@
 set -e
 
 sparse_algos=(
-  "gqa_value_aware_sparse_attention"
-  # "gqa_block_sparse_attention"
-  # "gqa_quest_sparse_attention"
-  # "block_sparse_attention"
+
 )
 
 for algo in "${sparse_algos[@]}"; do
@@ -15,6 +12,6 @@ for algo in "${sparse_algos[@]}"; do
     --topk-val 30 \
     --vortex-module-name "${algo}" \
     --model-name Qwen/Qwen3-1.7B \
-    --mem 0.8
+    --mem 0.7
 done
 

vortex_torch/abs/context_base.py

@@ -113,6 +113,7 @@ def summary(self) -> None:
         Print fields; tensor fields show shape/dtype/device, and append memory totals incl. auxiliary.
         """
 
+        return 
         def _fmt_bytes(n: int) -> str:
             units = ("B", "KB", "MB", "GB", "TB", "PB")
             f = float(n)