chore: upgrade nixl to 0.6.0 (#3550)

Signed-off-by: richardhuo-nv <rihuo@nvidia.com>

Cargo.lock

@@ -5103,9 +5103,9 @@ dependencies = [
 
 [[package]]
 name = "nixl-sys"
-version = "0.4.1"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "743ed1038b386b75451f9e0bba37cb2e3eea75873635268337d6531be99c9303"
+checksum = "06b070d86389c00f5144bc0f7a456913212c25fa4e4713c482f2f0bfa0839ee2"
 dependencies = [
  "bindgen 0.71.1",
  "cc",

container/Dockerfile

@@ -39,7 +39,7 @@ ARG SCCACHE_REGION=""
 
 # NIXL configuration
 ARG NIXL_UCX_REF=v1.19.0
-ARG NIXL_REF=0.4.1
+ARG NIXL_REF=0.6.0
 
 # Python configuration
 ARG PYTHON_VERSION=3.12

container/build.sh

@@ -115,7 +115,7 @@ NONE_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
 SGLANG_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
 SGLANG_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
 
-NIXL_REF=0.4.1
+NIXL_REF=0.6.0
 NIXL_UCX_REF=v1.19.0
 NIXL_UCX_EFA_REF=9d2b88a1f67faf9876f267658bd077b379b8bb76
 

docs/support_matrix.md

@@ -58,14 +58,14 @@ If you are using a **GPU**, the following GPU models and architectures are suppo
 | :----------------- | :---------- | :------------------------------------ | :----------- |
 | ai-dynamo          | 0.5.1       | >=2.28                                |              |
 | ai-dynamo-runtime  | 0.5.1       | >=2.28 (Python 3.12 has known issues) |              |
-| NIXL               | 0.4.1       | >=2.27                                | >=11.8       |
+| NIXL               | 0.6.0       | >=2.27                                | >=11.8       |
 
 ### Build Dependency
 
 | **Build Dependency** | **Version**                                                                      |
 | :------------------- | :------------------------------------------------------------------------------- |
 | **TensorRT-LLM**     | 1.1.0rc5                                                                         |
-| **NIXL**             | 0.4.1                                                                            |
+| **NIXL**             | 0.6.0                                                                            |
 | **vLLM**             | 0.10.1.1                                                                         |
 | **SGLang**           | 0.5.3rc0                                                                         |
 

lib/bindings/python/Cargo.lock

@@ -3638,9 +3638,9 @@ dependencies = [
 
 [[package]]
 name = "nixl-sys"
-version = "0.4.1"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "743ed1038b386b75451f9e0bba37cb2e3eea75873635268337d6531be99c9303"
+checksum = "06b070d86389c00f5144bc0f7a456913212c25fa4e4713c482f2f0bfa0839ee2"
 dependencies = [
  "bindgen 0.71.1",
  "cc",

lib/llm/Cargo.toml

@@ -93,7 +93,7 @@ dialoguer = { version = "0.11", default-features = false, features = [
 ] }
 
 # block_manager
-nixl-sys = { version = "0.4.1", optional = true }
+nixl-sys = { version = "=0.6.0", optional = true }
 cudarc = { version = "0.17.1", features = ["cuda-12020"], optional = true }
 ndarray = { version = "0.16", optional = true }
 nix = { version = "0.26", optional = true }

lib/llm/src/block_manager/block/transfer/nixl.rs

@@ -4,7 +4,7 @@
 use super::*;
 
 use anyhow::Result;
-use nixl_sys::{MemoryRegion, NixlDescriptor, XferDescList};
+use nixl_sys::{MemoryRegion, NixlDescriptor, XferDescList, XferStatus};
 use std::future::Future;
 
 fn append_xfer_request<Source, Destination>(
@@ -109,15 +109,13 @@ where
         .storage_type()
         .nixl_mem_type();
 
-    let mut src_dl = XferDescList::new(src_mem_type, false)?;
-    let mut dst_dl = XferDescList::new(dst_mem_type, false)?;
+    let mut src_dl = XferDescList::new(src_mem_type)?;
+    let mut dst_dl = XferDescList::new(dst_mem_type)?;
 
     for (src, dst) in src.iter().zip(dst.iter_mut()) {
         append_xfer_request(src, dst, &mut src_dl, &mut dst_dl)?;
     }
 
-    debug_assert!(!src_dl.has_overlaps()? && !dst_dl.has_overlaps()?);
-
     let xfer_req = nixl_agent.create_xfer_req(
         transfer_type.as_xfer_op(),
         &src_dl,
@@ -137,8 +135,10 @@ where
 
             loop {
                 match nixl_agent.get_xfer_status(&xfer_req) {
-                    Ok(false) => break, // Transfer is complete.
-                    Ok(true) => tokio::time::sleep(std::time::Duration::from_millis(5)).await, // Transfer is still in progress.
+                    Ok(XferStatus::Success) => break, // Transfer is complete.
+                    Ok(XferStatus::InProgress) => {
+                        tokio::time::sleep(std::time::Duration::from_millis(5)).await
+                    } // Transfer is still in progress.
                     Err(e) => {
                         tracing::error!("Error getting transfer status: {}", e);
                         break;

pyproject.toml

@@ -53,13 +53,13 @@ trtllm =[
 
 vllm = [
     "uvloop",
-    "nixl<=0.4.1",
+    "nixl<=0.6.0",
     "vllm[flashinfer]==0.10.2",
 ]
 
 sglang = [
     "uvloop",
-    "nixl<=0.4.1",
+    "nixl<=0.6.0",
     "sglang[all]==0.5.3",
 ]