refactor: 基于 causal-lm 重构 service/web-api/xtask

Signed-off-by: YdrMaster <[email protected]>

nvidia/common/src/lib.rs

@@ -247,3 +247,12 @@ where
         })
     }
 }
+
+pub fn synchronize() {
+    cuda::init();
+    for i in 0..cuda::Device::count() {
+        cuda::Device::new(i as _)
+            .retain_primary()
+            .apply(|ctx| ctx.synchronize());
+    }
+}

service/Cargo.toml

@@ -11,20 +11,9 @@ common = { path = "../common" }
 tensor = { path = "../tensor" }
 tokenizer = { path = "../tokenizer" }
 causal-lm = { path = "../causal-lm" }
-transformer = { path = "../transformer" }
-transformer-cpu = { path = "../transformer-cpu" }
-transformer-nv = { path = "../nvidia/transformer", optional = true }
-distributed = { path = "../nvidia/distributed", optional = true }
-half.workspace = true
 log.workspace = true
 tokio.workspace = true
 
-[build-dependencies]
-search-cuda-tools.workspace = true
-
 [dev-dependencies]
 colored = "2.1"
-tokio = { workspace = true, features = ["time"] }
-
-[features]
-nvidia = ["transformer-nv", "distributed"]
+transformer-cpu = { path = "../transformer-cpu" }

service/src/batcher.rs

@@ -1,18 +1,10 @@
-use crate::session::SessionContext;
-use common::utok;
-use std::sync::{
+use std::sync::{
     atomic::{
         AtomicBool,
         Ordering::{Acquire, Release},
     },
     Condvar, Mutex,
 };
-use tokio::sync::mpsc::UnboundedSender;
-
-pub struct Task<Cache> {
-    pub ctx: SessionContext<Cache>,
-    pub responsing: UnboundedSender<utok>,
-}
 
 pub struct Batcher<T> {
     queue: Mutex<Vec<T>>,