Disable chunk prefill and prefix caching when model is attention free

Signed-off-by: Christian Pinto <christian.pinto@ibm.com>

Author: christian-pinto

vllm/config.py

@@ -4710,6 +4710,15 @@ def __post_init__(self):
                     "Only \"last\" pooling supports chunked "
                     "prefill and prefix caching; disabling both.")
 
+        if self.model_config.is_attention_free:
+            # If the model is not of pooling type and it is attention free,
+            # we make sure chunked prefill and prefix_caching are
+            # disabled so that the correct KVCacheCoordinator
+            # is loaded.
+            disable_chunked_prefill_reasons.append(
+                "This is an attention free model, "
+                "disabling chunked prefill and prefix caching.")
+
         if disable_chunked_prefill_reasons:
             for reason in disable_chunked_prefill_reasons:
                 logger.info(reason)

vllm/v1/core/kv_cache_coordinator.py

@@ -388,7 +388,7 @@ def get_kv_cache_coordinator(
         kv_cache_config: KVCacheConfig, max_model_len: int, use_eagle: bool,
         enable_caching: bool, caching_hash_fn: Callable,
         enable_kv_cache_events: bool) -> KVCacheCoordinator:
-    if not enable_caching or len(kv_cache_config.kv_cache_groups) == 0:
+    if not enable_caching:
         # We instantiate this coordinator also for attention free models that
         # have 0 kv_cache_groups
         return KVCacheCoordinatorNoPrefixCache(kv_cache_config, max_model_len,

vllm/v1/core/kv_cache_manager.py

@@ -89,7 +89,7 @@ def __init__(
         self.prefix_cache_stats = PrefixCacheStats() if log_stats else None
 
         self.block_size: Optional[int] = None
-        if self.enable_caching and len(kv_cache_config.kv_cache_groups) > 0:
+        if self.enable_caching:
             assert len(
                 set(g.kv_cache_spec.block_size
                     for g in kv_cache_config.kv_cache_groups)