Simple attempt to fix the healthcheck block allocation.

This commit is contained in:
Nicolas Patry 2024-12-10 17:57:17 +01:00
parent 22dbc449cd
commit 45a86d5cf0
No known key found for this signature in database
GPG Key ID: D2920555C90F704C
4 changed files with 41 additions and 9 deletions

View File

@ -147,8 +147,8 @@ impl SimpleAllocator {
fn new(blocks: u32, block_size: u32, window_size: Option<u32>) -> Self {
SimpleAllocator {
block_size,
// Block 0 is reserved for health checks
free_blocks: (1..blocks).collect(),
// XXX: Block 0&1 is reserved for health checks
free_blocks: (2..blocks).collect(),
window_size,
}
}

View File

@ -37,6 +37,8 @@ pub enum ClientError {
Generation(String),
#[error("Sharded results are empty")]
EmptyResults,
#[error("Invalid attention {0}")]
InvalidAttention(String),
}
impl From<Status> for ClientError {

View File

@ -13,15 +13,38 @@ use futures::future::join_all;
use tonic::transport::Uri;
use tracing::instrument;
#[derive(Debug, Clone, Copy)]
pub enum Attn {
Flashdecoding,
Flashinfer,
Paged,
}
impl TryFrom<&str> for Attn {
type Error = ClientError;
fn try_from(value: &str) -> std::result::Result<Self, Self::Error> {
match value {
"flashdecoding" => Ok(Attn::Flashdecoding),
"flashinfer" => Ok(Attn::Flashinfer),
"paged" => Ok(Attn::Paged),
string => Err(ClientError::InvalidAttention(string.to_string())),
}
}
}
#[derive(Debug, Clone)]
/// Text Generation Inference gRPC multi client
pub struct ShardedClient {
clients: Vec<Client>,
attention_impl: Option<Attn>,
}
impl ShardedClient {
fn new(clients: Vec<Client>) -> Self {
Self { clients }
Self {
clients,
attention_impl: None,
}
}
/// Create a new ShardedClient from a master client. The master client will communicate with
@ -55,7 +78,9 @@ impl ShardedClient {
.iter_mut()
.map(|client| client.info())
.collect();
join_all(futures).await.pop().unwrap()
let info = join_all(futures).await.pop().unwrap()?;
self.attention_impl = Some((&*info.attention_impl).try_into()?);
Ok(info)
}
/// GRPC health check
@ -211,6 +236,12 @@ impl Health for ShardedClient {
async fn model_health(&self) -> Result<()> {
// Dummy batch of 1 token and 1 generated token
//
let (blocks, slots) = match self.attention_impl.expect("Attention to be set") {
Attn::Paged => (vec![0], (0..2).collect()),
Attn::Flashinfer => (vec![0, 1], (0..2).collect()),
Attn::Flashdecoding => (vec![0], (0..2).collect()),
};
let liveness_request = Request {
id: u64::MAX,
inputs: "liveness".to_string(),
@ -239,9 +270,8 @@ impl Health for ShardedClient {
ignore_eos_token: false,
}),
top_n_tokens: 0,
// Block 0 is reserved for health checks
blocks: vec![0],
slots: (0..16).collect(),
blocks,
slots,
cache_len: 0,
adapter_id: None,
chunk_len: None,

View File

@ -42,8 +42,8 @@ impl RadixAllocator {
allocations: HashMap::new(),
cache_blocks: RadixTrie::new(block_size as usize),
// Block 0 is reserved for health checks.
free_blocks: (1..n_blocks).collect(),
// XXX: Block 0,1 is reserved for health checks.
free_blocks: (2..n_blocks).collect(),
window_size,
block_size,
}