Making prefix/flashinfer the default and testing the full release tests.

backends/v3/src/backend.rs

@@ -38,7 +38,7 @@ impl BackendV3 {
         let prefix_caching = if let Ok(prefix_caching) = std::env::var("USE_PREFIX_CACHING") {
             matches!(prefix_caching.as_str(), "true" | "1")
         } else {
-            false
+            true
         };
         let attention = if let Ok(attention) = std::env::var("ATTENTION") {
             attention

server/text_generation_server/models/globals.py

@@ -5,7 +5,7 @@ from typing import Dict, Optional
 
 from text_generation_server.utils.log import log_master
 
-PREFIX_CACHING = os.getenv("USE_PREFIX_CACHING", "0").lower() in {"1", "true"}
+PREFIX_CACHING = os.getenv("USE_PREFIX_CACHING", "1").lower() in {"1", "true"}
 log_master(logger.info, f"Using prefix caching = {PREFIX_CACHING}")
 ATTENTION = os.getenv("ATTENTION", "flashinfer" if PREFIX_CACHING else "paged")
 _expected = {"paged", "flashdecoding", "flashinfer"}