build(llama-infini): 更新算子库和 infini-toolkit

Signed-off-by: YdrMaster <ydrml@hotmail.com>

Author: YdrMaster

Cargo.toml

@@ -30,7 +30,7 @@ regex = "1.11"
 env_logger = "0.11"
 build-script-cfg = "0.0"
 
-operators = { git = "https://github.com/YdrMaster/operators-rs", rev = "b9e6fdd", default-features = false }
+operators = { git = "https://github.com/YdrMaster/operators-rs", rev = "7aa1051", default-features = false }
 search-cl-tools = { git = "https://github.com/InfiniTensor/clrt", rev = "6846d52" }
-search-infini-tools = { git = "https://github.com/InfiniTensor/infini-rt", rev = "136c30b" }
+search-infini-tools = { git = "https://github.com/InfiniTensor/infini-toolkit", rev = "9943c78" }
 search-cuda-tools = { git = "https://github.com/YdrMaster/cuda-driver", rev = "e2ec203" }

models/llama/infini/src/infer.rs

@@ -2,7 +2,8 @@
 use gguf::GGufModel;
 use llama::{ext::ggml_quants::f16, LlamaRequest, LlamaStorage, LlamaWorker, Tensor};
 use operators::{
-    infini_rt::{self, Device, DeviceType::DEVICE_CPU},
+    infini::Device,
+    infini_rt::{self, DeviceType::DEVICE_CPU},
     random_sample::{KVPair, SampleArgs},
     TopoNode,
 };
@@ -59,12 +60,7 @@ fn test_infer() {
     println!("distribution: {devices:?}");
 
     infini_rt::init(DEVICE_CPU);
-    let (seeds, senders) = WorkerSeed::new(
-        devices
-            .into_iter()
-            .map(|id| Device { ty: DEVICE_CPU, id })
-            .collect(),
-    );
+    let (seeds, senders) = WorkerSeed::new(devices.into_iter().map(|_| Device::cpu()).collect());
     thread::scope(|s| {
         let _workers = zip(lens, seeds)
             .enumerate()

models/llama/infini/src/lib.rs

@@ -3,7 +3,8 @@
 use llama::{BlkWeight, Contiguous, LlamaBlkStorage, LlamaStorage, Tensor, WeightLoader};
 use operators::{
     all_reduce::{AllReduce, NonAllReduce},
-    infini_rt::{DevBlob, DevByte, Device, Event, HostBlob, Stream},
+    infini::Device,
+    infini_rt::{DevBlob, DevByte, Event, HostBlob, Stream},
     random_sample::infini::Operator as RandomSampleNpu,
     rearrange::infini::Operator as Rearrange,
     ByteOf, QueueOf, TopoNode,