[v1][core] Support for attention free models

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

@@ -389,6 +389,8 @@ def get_kv_cache_coordinator(
         enable_caching: bool, caching_hash_fn: Callable,
         enable_kv_cache_events: bool) -> KVCacheCoordinator:
     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,
                                                use_eagle, caching_hash_fn,
                                                enable_kv_cache_events)

vllm/v1/core/kv_cache_manager.py

@@ -78,7 +78,9 @@ def __init__(
     ) -> None:
         self.max_model_len = max_model_len
 
-        self.enable_caching = enable_caching
+        self.enable_caching = (enable_caching
+                               if len(kv_cache_config.kv_cache_groups) > 0
+                               else False)
         self.caching_hash_fn = (
             sha256_cbor_64bit if caching_hash_algo == "sha256_cbor_64bit" else
             sha256 if caching_hash_algo == "sha256" else hash)
@@ -101,7 +103,7 @@ def __init__(
             kv_cache_config=kv_cache_config,
             max_model_len=self.max_model_len,
             use_eagle=self.use_eagle,
-            enable_caching=enable_caching,
+            enable_caching=self.enable_caching,
             caching_hash_fn=self.caching_hash_fn,
             enable_kv_cache_events=enable_kv_cache_events,
         )

vllm/v1/core/kv_cache_utils.py

@@ -563,6 +563,10 @@ 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 kv_cache_spec is empty
+    if not kv_cache_spec:
+        return
+
     if available_memory <= 0:
         raise ValueError("No available memory for the cache blocks. "
                          "Try increasing `gpu_memory_utilization` when "
@@ -749,6 +753,13 @@ def is_kv_cache_page_size_uniform(
     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
+    return not kv_cache_spec
+
+
 def _get_kv_cache_config_uniform_page_size(
         vllm_config: VllmConfig, kv_cache_spec: dict[str, KVCacheSpec],
         available_memory: int) -> KVCacheConfig:
@@ -891,6 +902,10 @@ 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
@@ -957,7 +972,11 @@ 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_cache config with 0 kv_cache groups and 1 block
+        # to allow for the KVCache manager to handle attention free 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/engine/core.py

@@ -139,7 +139,11 @@ def _initialize_kv_caches(
 
         # Profiles the peak memory usage of the model to determine how much
         # memory can be allocated for kv cache.
-        available_gpu_memory = self.model_executor.determine_available_memory()
+        check_available_memory = not(len(kv_cache_specs) == 1 and not kv_cache_specs[0])
+        available_gpu_memory = [0]
+        if check_available_memory:
+            available_gpu_memory = (
+                self.model_executor.determine_available_memory())
 
         assert len(kv_cache_specs) == len(available_gpu_memory)
         # Get the kv cache tensor size

vllm/v1/worker/gpu_worker.py

@@ -209,6 +209,7 @@ def determine_available_memory(self) -> int:
             You may limit the usage of GPU memory
             by adjusting the `gpu_memory_utilization` parameter.
         """
+
         torch.cuda.empty_cache()
         torch.cuda.reset_peak_memory_stats()
         GiB = lambda b: b / GiB_bytes