Support for attention free models

Forces 0 KV Cache groups to disable KV Cache in attention free models

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

Author: christian-pinto

vllm/v1/core/kv_cache_coordinator.py

@@ -221,7 +221,9 @@ def __init__(self, kv_cache_config: KVCacheConfig, max_model_len: int,
         super().__init__(kv_cache_config, max_model_len, use_eagle,
                          enable_caching, caching_hash_fn,
                          enable_kv_cache_events)
-        self.verify_and_split_kv_cache_groups()
+        # attention free models are initialized with 0 kv_cache_groups
+        if len(self.kv_cache_config.kv_cache_groups) > 0:
+            self.verify_and_split_kv_cache_groups()
 
     def verify_and_split_kv_cache_groups(self) -> None:
         """

vllm/v1/core/kv_cache_manager.py

@@ -86,7 +86,7 @@ def __init__(
         self.prefix_cache_stats = PrefixCacheStats() if log_stats else None
 
         self.block_size: Optional[int] = None
-        if self.enable_caching:
+        if self.enable_caching and len(self.kv_cache_config.kv_cache_groups) > 0:
             assert len(
                 set(g.kv_cache_spec.block_size
                     for g in kv_cache_config.kv_cache_groups)

vllm/v1/core/kv_cache_utils.py

@@ -551,6 +551,10 @@ def check_enough_kv_cache_memory(vllm_config: VllmConfig,
         ValueError: If there is not enough memory available for the KV cache.
     """
 
+    # No need to check for available memory if the model is attention free
+    if vllm_config.model_config.is_attention_free:
+        return
+
     if available_memory <= 0:
         raise ValueError("No available memory for the cache blocks. "
                          "Try increasing `gpu_memory_utilization` when "
@@ -736,6 +740,12 @@ def is_kv_cache_page_size_uniform(
     page_sizes = {layer.page_size_bytes for layer in kv_cache_spec.values()}
     return len(page_sizes) == 1
 
+def is_kv_cache_type_attention_free(
+        kv_cache_spec: dict[str, KVCacheSpec]) -> bool:
+
+    # kv_cache_spec is an empty dict for attention free models
+    if not kv_cache_spec:
+        return True
 
 def _get_kv_cache_config_uniform_page_size(
         vllm_config: VllmConfig, kv_cache_spec: dict[str, KVCacheSpec],
@@ -879,6 +889,11 @@ def _get_kv_cache_config_uniform_page_size(
     return kv_cache_config
 
 
+def _get_kv_cache_config_attention_free() -> KVCacheConfig:
+    return KVCacheConfig(num_blocks=1,
+                         kv_cache_tensors=[],
+                         kv_cache_groups=[])
+
 def unify_hybrid_kv_cache_specs(kv_cache_spec: dict[str, KVCacheSpec]):
     """
     This function tries to convert the KV cache specs to one type if the model
@@ -945,7 +960,11 @@ def get_kv_cache_config(
     if vllm_config.scheduler_config.disable_hybrid_kv_cache_manager:
         unify_hybrid_kv_cache_specs(kv_cache_spec)
 
-    if is_kv_cache_type_uniform(kv_cache_spec):
+    if is_kv_cache_type_attention_free(kv_cache_spec):
+        # This returns a kv_cahce config with 0 kv cache groups and 1 block
+        # to allow for the KVCache manager to handle attention fre models.
+        return _get_kv_cache_config_attention_free()
+    elif is_kv_cache_type_uniform(kv_cache_spec):
         # KV cache of all layers are the same, which is true for
         # most models. Allocate the same amount of memory for
         # each layer.

vllm/v1/executor/abstract.py

@@ -73,10 +73,16 @@ def register_failure_callback(self, callback: FailureCallback):
         pass
 
     def determine_available_memory(self) -> list[int]:  # in bytes
+        if self.vllm_config.model_config.is_attention_free:
+            return [0]
+
         output = self.collective_rpc("determine_available_memory")
         return output
 
     def get_kv_cache_specs(self) -> list[dict[str, KVCacheSpec]]:
+        if self.vllm_config.model_config.is_attention_free:
+            return [{}]
+
         output = self.collective_rpc("get_kv_cache_spec")
         return output