[Perf] Disable chunked local attention by default with llama4 (#21761)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

vllm/config.py

@@ -4769,12 +4769,23 @@ class VllmConfig:
                 # Hybrid KV cache manager is not compatible with KV events.
                 self.scheduler_config.disable_hybrid_kv_cache_manager = True
             if self.model_config is not None and \
-                self.model_config.attention_chunk_size is not None and \
-                self.speculative_config is not None and \
-                self.speculative_config.use_eagle():
-                # Hybrid KV cache manager is not yet supported with chunked
-                # local attention + eagle.
-                self.scheduler_config.disable_hybrid_kv_cache_manager = True
+                self.model_config.attention_chunk_size is not None:
+                if self.speculative_config is not None and \
+                    self.speculative_config.use_eagle():
+                    # Hybrid KV cache manager is not yet supported with chunked
+                    # local attention + eagle.
+                    self.scheduler_config.disable_hybrid_kv_cache_manager = True
+                elif \
+                    not envs.VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE:
+                    logger.warning(
+                        "There is a latency regression when using chunked local"
+                        " attention with the hybrid KV cache manager. Disabling"
+                        " it, by default. To enable it, set the environment "
+                        "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE=1."
+                    )
+                    # Hybrid KV cache manager is not yet supported with chunked
+                    # local attention.
+                    self.scheduler_config.disable_hybrid_kv_cache_manager = True
 
     def update_sizes_for_sequence_parallelism(self,
                                               possible_sizes: list) -> list:

vllm/envs.py

@@ -143,6 +143,7 @@ if TYPE_CHECKING:
     VLLM_USE_CUDNN_PREFILL: bool = False
     VLLM_ENABLE_CUDAGRAPH_GC: bool = False
     VLLM_LOOPBACK_IP: str = ""
+    VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
 
 
 def get_default_cache_root():
@@ -991,6 +992,17 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # The default value is "VLLM".
     "VLLM_PROCESS_NAME_PREFIX":
     lambda: os.getenv("VLLM_PROCESS_NAME_PREFIX", "VLLM"),
+
+    # Allow chunked local attention with hybrid kv cache manager.
+    # Currently using the Hybrid KV cache manager with chunked local attention
+    # in the Llama4 models (the only models currently using chunked local attn)
+    # causes a latency regression. For this reason, we disable it by default.
+    # This flag is used to allow users to enable it if they want to (to save on
+    # kv-cache memory usage and enable longer contexts)
+    # TODO(lucas): Remove this flag once latency regression is resolved.
+    "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE":
+    lambda: bool(int(os.getenv(\
+            "VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE", "0"))),
 }
 
 # --8<-- [end:env-vars-definition]