Skip to content

GitLab

Commit e019635f authored by xuxzh1's avatar xuxzh1 🎱
Browse files

update

parent 64def8e2
Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 7279 deletions
load_tests/starcoder_load.js deleted 100644 → 0
View file @ 64def8e2
import {check} from 'k6';
import http from 'k6/http';
import {Trend} from 'k6/metrics';
const host = __ENV.HOST || '127.0.0.1:3000';
const totalTime = new Trend('total_time', true);
const validationTime = new Trend('validation_time', true);
const queueTime = new Trend('queue_time', true);
const inferenceTime = new Trend('inference_time', true);
const timePerToken = new Trend('time_per_token', true);
const example = {
payload: JSON.stringify({
inputs: '# This is a fibonacci function written in the Python programming language.' +
'def fibonacci',
parameters: {
details: true,
max_new_tokens: 60,
temperature: 0.2,
top_p: 0.95,
seed: 0,
},
}),
generated_tokens: 60
};
export const options = {
thresholds: {
http_req_failed: ['rate==0'],
time_per_token: ['p(95)<90'],
queue_time: ['p(95)<1500'],
},
scenarios: {
load_test: {
executor: 'constant-arrival-rate',
duration: '60s',
preAllocatedVUs: 100,
rate: 10,
timeUnit: '1s',
},
},
};
export default function () {
const headers = {'Content-Type': 'application/json'};
const res = http.post(`http://${host}/generate`, example.payload, {
headers,
});
check(res, {
'Post status is 200': (r) => res.status === 200,
'Post response generated tokens': (r) => res.status === 200 && res.json().details.generated_tokens === example.generated_tokens,
});
if (res.status === 200) {
totalTime.add(res.headers["X-Total-Time"]);
validationTime.add(res.headers["X-Validation-Time"]);
queueTime.add(res.headers["X-Queue-Time"]);
inferenceTime.add(res.headers["X-Inference-Time"]);
timePerToken.add(res.headers["X-Time-Per-Token"]);
}
}
router/client/src/client.rs deleted 100644 → 0
View file @ 64def8e2
/// Single shard Client
use crate::pb::generate::v2::text_generation_service_client::TextGenerationServiceClient;
use crate::pb::generate::v2::*;
use crate::Result;
use grpc_metadata::InjectTelemetryContext;
use std::cmp::min;
use std::time::Duration;
use tonic::transport::{Channel, Uri};
use tracing::instrument;
/// Text Generation Inference gRPC client
#[derive(Debug, Clone)]
pub struct Client {
stub: TextGenerationServiceClient<Channel>,
}
impl Client {
/// Returns a client connected to the given url
pub async fn connect(uri: Uri) -> Result<Self> {
let channel = Channel::builder(uri).connect().await?;
Ok(Self {
stub: TextGenerationServiceClient::new(channel),
})
}
/// Returns a client connected to the given unix socket
pub async fn connect_uds(path: String) -> Result<Self> {
let channel = Channel::from_shared("http://[::]:50051".to_string())
.unwrap()
.connect_with_connector(tower::service_fn(move |_: Uri| {
tokio::net::UnixStream::connect(path.clone())
}))
.await?;
Ok(Self {
stub: TextGenerationServiceClient::new(channel),
})
}
/// Returns a list of uris or unix sockets of all shards
#[instrument(skip(self))]
pub async fn service_discovery(&mut self) -> Result<Vec<String>> {
let request = tonic::Request::new(ServiceDiscoveryRequest {}).inject_context();
let response = self.stub.service_discovery(request).await?;
let urls = response
.into_inner()
.urls
.into_iter()
// Remove unix socket prefix
.map(|url| match url.strip_prefix("unix://") {
None => url,
Some(stripped_url) => stripped_url.to_string(),
})
.collect();
Ok(urls)
}
/// Get model info
#[instrument(skip(self))]
pub async fn info(&mut self) -> Result<InfoResponse> {
let request = tonic::Request::new(InfoRequest {}).inject_context();
let response = self.stub.info(request).await?.into_inner();
Ok(response)
}
/// Get model health
#[instrument(skip(self))]
pub async fn health(&mut self) -> Result<HealthResponse> {
let request = tonic::Request::new(HealthRequest {}).inject_context();
let response = self.stub.health(request).await?.into_inner();
Ok(response)
}
/// Clear the past generations cache
#[instrument(skip(self))]
pub async fn clear_cache(&mut self, batch_id: Option<u64>) -> Result<()> {
let request = tonic::Request::new(ClearCacheRequest { id: batch_id }).inject_context();
self.stub.clear_cache(request).await?;
Ok(())
}
/// Filter a cached batch
#[instrument(skip(self))]
pub async fn filter_batch(
&mut self,
batch_id: u64,
request_ids: Vec<u64>,
) -> Result<Option<CachedBatch>> {
let request = tonic::Request::new(FilterBatchRequest {
batch_id,
request_ids,
})
.inject_context();
let filtered_batch = self.stub.filter_batch(request).await?.into_inner();
Ok(filtered_batch.batch)
}
/// Warmup on a max size batch
///
/// Returns the maximum amount of tokens supported by the hardware
#[instrument(skip_all)]
pub async fn warmup(
&mut self,
max_input_length: u32,
max_prefill_tokens: u32,
max_total_tokens: u32,
max_batch_size: Option<usize>,
) -> Result<Option<u32>> {
let mut n_tokens = 0;
let mut requests = Vec::new();
// Create requests
while n_tokens < max_prefill_tokens {
let truncate = min(max_input_length, max_prefill_tokens - n_tokens);
let mut inputs = String::new();
inputs.push_str(&"_test ".to_string().repeat(max_input_length as usize));
if n_tokens == 0 {
// 1 request is enough to test vision heads.
// Sending images on other queries messes up easily with truncation.
inputs.push_str("![](data:image/jpeg;base64,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)");
}
requests.push(Request {
id: 0,
// We truncate the input on the server side to be sure that it has the correct size
inputs,
truncate,
// Set sampling parameters to also take these ops into account in the max memory
parameters: Some(NextTokenChooserParameters {
temperature: 0.9,
top_k: 10,
top_p: 0.9,
typical_p: 0.9,
do_sample: false,
seed: 0,
repetition_penalty: 1.2,
frequency_penalty: 0.1,
watermark: true,
grammar: String::new(),
grammar_type: GrammarType::None as i32,
}),
stopping_parameters: Some(StoppingCriteriaParameters {
max_new_tokens: max_total_tokens - truncate,
stop_sequences: vec![],
ignore_eos_token: true,
}),
prefill_logprobs: true,
top_n_tokens: 20,
});
n_tokens += max_input_length;
// Check max_batch_size
if Some(requests.len()) == max_batch_size {
break;
}
}
let batch = Batch {
id: 0,
size: requests.len() as u32,
requests,
max_tokens: 0,
};
let request = tonic::Request::new(WarmupRequest {
batch: Some(batch),
max_input_length,
max_prefill_tokens,
max_total_tokens,
})
.inject_context();
let response = self.stub.warmup(request).await?.into_inner();
Ok(response.max_supported_total_tokens)
}
/// Generate one token for each request in the given batch
///
/// Returns Generation for each request in batch
/// and the next cached batch
#[instrument(skip_all, fields(id = &batch.id, size = &batch.size))]
pub async fn prefill(
&mut self,
batch: Batch,
) -> Result<(Vec<Generation>, Option<CachedBatch>, PrefillTimings)> {
let request = tonic::Request::new(PrefillRequest { batch: Some(batch) }).inject_context();
let response = self.stub.prefill(request).await?.into_inner();
Ok((
response.generations,
response.batch,
PrefillTimings::new(response.forward_ns, response.decode_ns, response.total_ns),
))
}
/// Generate one token for each request in the given cached batches
///
/// Returns Generation for each request in batches
/// and the next cached batch
#[instrument(skip_all, fields(size = batches.iter().map(|batch|{batch.size}).sum::<u32>()))]
pub async fn decode(
&mut self,
batches: Vec<CachedBatch>,
) -> Result<(Vec<Generation>, Option<CachedBatch>, DecodeTimings)> {
let request = tonic::Request::new(DecodeRequest { batches }).inject_context();
let response = self.stub.decode(request).await?.into_inner();
Ok((
response.generations,
response.batch,
DecodeTimings::new(
response.concat_ns,
response.forward_ns,
response.decode_ns,
response.total_ns,
),
))
}
}
pub struct PrefillTimings {
pub forward: Duration,
pub decode: Duration,
pub total: Duration,
}
impl PrefillTimings {
fn new(forward_ns: u64, decode_ns: u64, total_ns: u64) -> Self {
Self {
forward: Duration::from_nanos(forward_ns),
decode: Duration::from_nanos(decode_ns),
total: Duration::from_nanos(total_ns),
}
}
}
pub struct DecodeTimings {
pub concat: Option<Duration>,
pub forward: Duration,
pub decode: Duration,
pub total: Duration,
}
impl DecodeTimings {
fn new(concat_ns: Option<u64>, forward_ns: u64, decode_ns: u64, total_ns: u64) -> Self {
Self {
concat: concat_ns.map(Duration::from_nanos),
forward: Duration::from_nanos(forward_ns),
decode: Duration::from_nanos(decode_ns),
total: Duration::from_nanos(total_ns),
}
}
}
router/client/src/sharded_client.rs deleted 100644 → 0
View file @ 64def8e2
use crate::client::{DecodeTimings, PrefillTimings};
/// Multi shard Client
use crate::{Batch, CachedBatch, Client, Generation, HealthResponse, ShardInfo};
use crate::{ClientError, Result};
use futures::future::join_all;
use tonic::transport::Uri;
use tracing::instrument;
#[derive(Debug, Clone)]
/// Text Generation Inference gRPC multi client
pub struct ShardedClient {
clients: Vec<Client>,
}
impl ShardedClient {
fn new(clients: Vec<Client>) -> Self {
Self { clients }
}
/// Create a new ShardedClient from a master client. The master client will communicate with
/// the other shards and returns all uris/unix sockets with the `service_discovery` gRPC method.
async fn from_master_client(mut master_client: Client) -> Result<Self> {
// Get all uris/unix sockets from the master client
let uris = master_client.service_discovery().await?;
let futures = uris.into_iter().map(Client::connect_uds);
let clients: Result<Vec<Client>> = join_all(futures).await.into_iter().collect();
Ok(Self::new(clients?))
}
/// Returns a client connected to the given uri
pub async fn connect(uri: Uri) -> Result<Self> {
let master_client = Client::connect(uri).await?;
Self::from_master_client(master_client).await
}
/// Returns a client connected to the given unix socket
pub async fn connect_uds(path: String) -> Result<Self> {
let master_client = Client::connect_uds(path).await?;
Self::from_master_client(master_client).await
}
/// Get the model info
#[instrument(skip(self))]
pub async fn info(&mut self) -> Result<ShardInfo> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| client.info())
.collect();
join_all(futures).await.pop().unwrap()
}
/// GRPC health check
#[instrument(skip(self))]
pub async fn health(&mut self) -> Result<HealthResponse> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| client.health())
.collect();
join_all(futures).await.pop().unwrap()
}
/// Clear the past generations cache
#[instrument(skip(self))]
pub async fn clear_cache(&mut self, batch_id: Option<u64>) -> Result<()> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| client.clear_cache(batch_id))
.collect();
join_all(futures).await.into_iter().collect()
}
/// Filter a cached batch
#[instrument(skip(self))]
pub async fn filter_batch(
&mut self,
batch_id: u64,
request_ids: Vec<u64>,
) -> Result<Option<CachedBatch>> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| Box::pin(client.filter_batch(batch_id, request_ids.clone())))
.collect();
// all shards return the same message
join_all(futures).await.pop().unwrap()
}
/// Warmup on a max size batch
///
/// Returns the maximum amount of tokens supported by the hardware
#[instrument(skip(self))]
pub async fn warmup(
&mut self,
max_input_length: u32,
max_prefill_tokens: u32,
max_total_tokens: u32,
max_batch_size: Option<usize>,
) -> Result<Option<u32>> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| {
Box::pin(client.warmup(
max_input_length,
max_prefill_tokens,
max_total_tokens,
max_batch_size,
))
})
.collect();
// Take the minimum value
let results = join_all(futures)
.await
.into_iter()
.collect::<Result<Vec<Option<u32>>>>()?;
Ok(results.into_iter().flatten().min())
}
/// Generate one token for each request in the given batch
///
/// Returns Generation for each request in batch
/// and the next cached batch
#[instrument(skip_all, fields(id = & batch.id, size = & batch.size))]
pub async fn prefill(
&mut self,
batch: Batch,
) -> Result<(Vec<Generation>, Option<CachedBatch>, PrefillTimings)> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| Box::pin(client.prefill(batch.clone())))
.collect();
#[allow(clippy::type_complexity)]
let results: Result<Vec<(Vec<Generation>, Option<CachedBatch>, PrefillTimings)>> =
join_all(futures).await.into_iter().collect();
let mut results = results?;
let (mut generations, next_batch, mut timings) =
results.pop().ok_or(ClientError::EmptyResults)?;
// Merge generations from different model shards
for (mut shard_generations, _, shard_timings) in results.into_iter() {
generations.append(&mut shard_generations);
// Return the timings of the slowest shard
if shard_timings.total > timings.total {
timings = shard_timings;
}
}
Ok((generations, next_batch, timings))
}
/// Generate one token for each request in the given cached batches
///
/// Returns Generation for each request in batches
/// and the next cached batch
#[instrument(skip_all, fields(size = batches.iter().map(| batch | {batch.size}).sum::< u32 > ()))]
pub async fn decode(
&mut self,
batches: Vec<CachedBatch>,
) -> Result<(Vec<Generation>, Option<CachedBatch>, DecodeTimings)> {
let futures: Vec<_> = self
.clients
.iter_mut()
.map(|client| Box::pin(client.decode(batches.clone())))
.collect();
#[allow(clippy::type_complexity)]
let results: Result<Vec<(Vec<Generation>, Option<CachedBatch>, DecodeTimings)>> =
join_all(futures).await.into_iter().collect();
let mut results = results?;
let (mut generations, next_batch, mut timings) =
results.pop().ok_or(ClientError::EmptyResults)?;
// Merge generations from different model shards
for (mut shard_generations, _, shard_timings) in results.into_iter() {
generations.append(&mut shard_generations);
// Return the timings of the slowest shard
if shard_timings.total > timings.total {
timings = shard_timings;
}
}
Ok((generations, next_batch, timings))
}
}
router/src/health.rs deleted 100644 → 0
View file @ 64def8e2
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;
use text_generation_client::GrammarType as ProtoGrammarType;
use text_generation_client::{
Batch, NextTokenChooserParameters, Request, ShardedClient, StoppingCriteriaParameters,
};
// Note: Request ids and batch ids cannot collide.
const LIVENESS_ID: u64 = u64::MAX;
const BATCH_ID: u64 = u64::MAX;
#[derive(Clone, Debug)]
pub(crate) struct Health {
client: ShardedClient,
generation_health: Arc<AtomicBool>,
}
impl Health {
pub(crate) fn new(client: ShardedClient, generation_health: Arc<AtomicBool>) -> Self {
Self {
client,
generation_health,
}
}
pub(crate) async fn check(&mut self) -> bool {
if self.generation_health.load(Ordering::SeqCst) {
// Generation is healthy, we only check that the shards are answering gRPC calls
self.client.health().await.is_ok()
} else {
// Generation is unhealthy or have not sent any generation request yet
// Dummy batch of 1 token and 1 generated token
let liveness_request = Request {
id: LIVENESS_ID,
inputs: "liveness".to_string(),
truncate: 10,
prefill_logprobs: false,
parameters: Some(NextTokenChooserParameters {
temperature: 1.0,
top_k: 0,
top_p: 1.0,
typical_p: 1.0,
do_sample: false,
seed: 0,
repetition_penalty: 1.0,
frequency_penalty: 0.0,
watermark: false,
grammar: String::new(),
grammar_type: ProtoGrammarType::None as i32,
}),
stopping_parameters: Some(StoppingCriteriaParameters {
max_new_tokens: 1,
stop_sequences: vec![],
ignore_eos_token: false,
}),
top_n_tokens: 0,
};
let batch = Batch {
id: BATCH_ID,
requests: vec![liveness_request],
size: 1,
max_tokens: 2,
};
// Skips the queue
let value = self.client.prefill(batch).await.is_ok();
// Update generation health
self.generation_health.store(value, Ordering::SeqCst);
value
}
}
}
router/src/infer.rs deleted 100644 → 0
View file @ 64def8e2
/// Batching and inference logic
use crate::validation::{Validation, ValidationError};
use crate::{
ChatTemplateInputs, ChatTemplateVersions, Entry, GenerateRequest, GenerateStreamResponse,
HubTokenizerConfig, Message, PrefillToken, Queue, Token,
};
use crate::{FunctionRef, FunctionsMap, GrammarType, Properties, Tool, ToolType, Tools};
use futures::future::try_join_all;
use minijinja::{Environment, ErrorKind, Template};
use nohash_hasher::IntMap;
use serde_json::{json, Map, Value};
use std::collections::HashMap;
use std::sync::{
atomic::{AtomicBool, Ordering},
Arc,
};
use text_generation_client::{
Batch, CachedBatch, ClientError, GeneratedText, Generation, ShardedClient, Tokens,
};
use thiserror::Error;
use tokio::sync::mpsc::error::SendError;
use tokio::sync::{mpsc, Notify, Semaphore, TryAcquireError};
use tokio::time::Instant;
use tokio_stream::wrappers::UnboundedReceiverStream;
use tokio_stream::StreamExt;
use tracing::{info_span, instrument, Instrument, Span};
/// Inference struct
#[derive(Clone)]
pub struct Infer {
/// Validation
validation: Validation,
/// Request queue
queue: Queue,
/// Shared state
shared: Arc<Shared>,
/// Chat template
chat_template: Option<ChatTemplate>,
/// Inference limit
limit_concurrent_requests: Arc<Semaphore>,
}
/// Infer shared state
struct Shared {
/// Batching background Tokio task notifier
batching_task: Notify,
}
/// Raise a exception (custom function) used in the chat templates
fn raise_exception(err_text: String) -> Result<String, minijinja::Error> {
Err(minijinja::Error::new(ErrorKind::SyntaxError, err_text))
}
impl Infer {
#[allow(clippy::too_many_arguments)]
pub(crate) fn new(
client: ShardedClient,
validation: Validation,
waiting_served_ratio: f32,
max_batch_prefill_tokens: u32,
max_batch_total_tokens: u32,
max_waiting_tokens: usize,
max_batch_size: Option<usize>,
max_concurrent_requests: usize,
requires_padding: bool,
window_size: Option<u32>,
speculate: u32,
generation_health: Arc<AtomicBool>,
tokenizer_config: HubTokenizerConfig,
) -> Self {
// Infer shared state
let queue = Queue::new(requires_padding, 16, window_size, speculate);
let shared = Arc::new(Shared {
batching_task: Notify::new(),
});
// Spawn batching background task that contains all the inference logic
tokio::spawn(batching_task(
client,
waiting_served_ratio,
max_batch_prefill_tokens,
max_batch_total_tokens,
max_waiting_tokens,
max_batch_size,
queue.clone(),
shared.clone(),
generation_health,
));
let chat_template = tokenizer_config
.chat_template
.and_then(|t| match t {
ChatTemplateVersions::Single(template) => Some(template),
ChatTemplateVersions::Multiple(templates) => templates
.into_iter()
.find(|t| t.name == "default")
.map(|t| t.template),
})
.map(|t| {
// .strip() is not supported in minijinja
let t = t.replace(".strip()", " | trim");
ChatTemplate::new(t, tokenizer_config.bos_token, tokenizer_config.eos_token)
});
// Inference limit with a semaphore
let semaphore = Arc::new(Semaphore::new(max_concurrent_requests));
Self {
validation,
queue,
shared,
chat_template,
limit_concurrent_requests: semaphore,
}
}
/// Add a new request to the queue and return a stream of InferStreamResponse
#[instrument(skip_all)]
pub(crate) async fn generate_stream(
&self,
request: GenerateRequest,
) -> Result<GenerateStreamResponse, InferError> {
// Limit concurrent requests by acquiring a permit from the semaphore
let permit = self
.clone()
.limit_concurrent_requests
.try_acquire_owned()
.map_err(|err| {
metrics::increment_counter!("tgi_request_failure", "err" => "overloaded");
tracing::error!("{err}");
err
})?;
// Validate request
let valid_request = self.validation.validate(request).await.map_err(|err| {
metrics::increment_counter!("tgi_request_failure", "err" => "validation");
tracing::error!("{err}");
err
})?;
// MPSC channel to communicate with the background batching task
let (response_tx, response_rx) = mpsc::unbounded_channel();
let input_length = valid_request.input_length;
// Append the request to the queue
self.queue.append(Entry {
request: valid_request,
response_tx,
span: Span::current(),
temp_span: None,
queue_time: Instant::now(),
batch_time: None,
});
// Notify the background task that we have a new entry in the queue that needs
// to be batched
self.shared.batching_task.notify_one();
// Return stream
Ok((
permit,
input_length,
UnboundedReceiverStream::new(response_rx),
))
}
/// Tokenizer the input
#[instrument(skip_all)]
pub(crate) async fn tokenize(
&self,
request: GenerateRequest,
) -> Result<Option<tokenizers::Encoding>, InferError> {
// Tokenize request
let inputs = request.inputs;
let truncate = request.parameters.truncate;
let encoding = self
.validation
.tokenize(inputs, truncate)
.await
.map_err(|err| {
tracing::error!("Tokenization {err}");
err
})?;
// Return Encoding
Ok(encoding.map(|(encoding, _)| encoding))
}
/// Apply the chat template to the chat request
#[instrument(skip_all)]
pub(crate) fn apply_chat_template(
&self,
messages: Vec<Message>,
grammar_with_prompt: Option<(GrammarType, String)>,
) -> Result<String, InferError> {
self.chat_template
.as_ref()
.ok_or_else(|| InferError::TemplateError(ErrorKind::TemplateNotFound.into()))?
.apply(messages, grammar_with_prompt)
.map_err(|e| {
metrics::increment_counter!("tgi_request_failure", "err" => "template");
tracing::error!("{e}");
e
})
}
/// Add a new request to the queue and return a InferResponse
#[instrument(skip_all)]
pub(crate) async fn generate(
&self,
request: GenerateRequest,
) -> Result<InferResponse, InferError> {
let use_top_tokens = request.parameters.top_n_tokens.is_some_and(|x| x > 0);
// Create stream and keep semaphore permit as long as generate lives
let (_permit, _input_length, mut stream) = self.generate_stream(request).await?;
// Return values
let mut result_prefill = Vec::new();
let mut result_tokens = Vec::new();
let mut result_top_tokens = Vec::new();
let mut result_generated_text = None;
let mut result_start = None;
let mut result_queued = None;
// Iterate on stream
while let Some(response) = stream.next().await {
match response? {
// Add prefill tokens
InferStreamResponse::Prefill(tokens) => {
// Create Token objects
// We do that here instead of in the Python code as Rust for loops are faster
result_prefill = tokens
.ids
.into_iter()
.zip(tokens.logprobs.into_iter())
.zip(tokens.texts.into_iter())
.map(|((id, logprob), text)| PrefillToken { id, text, logprob })
.collect();
}
// Push last token
InferStreamResponse::Intermediate { token, top_tokens } => {
result_tokens.push(token);
result_top_tokens.push(top_tokens);
}
// Final message
// Set return values
InferStreamResponse::End {
token,
generated_text,
start,
queued,
top_tokens,
} => {
result_tokens.push(token);
result_top_tokens.push(top_tokens);
result_generated_text = Some(generated_text);
result_start = Some(start);
result_queued = Some(queued)
}
}
}
// Check that we received a `InferStreamResponse::End` message
if let (Some(generated_text), Some(queued), Some(start)) =
(result_generated_text, result_queued, result_start)
{
Ok(InferResponse {
prefill: result_prefill,
_input_length,
tokens: result_tokens,
generated_text,
queued,
start,
top_tokens: if use_top_tokens {
result_top_tokens
} else {
Vec::new()
},
})
} else {
let err = InferError::IncompleteGeneration;
metrics::increment_counter!("tgi_request_failure", "err" => "incomplete");
tracing::error!("{err}");
Err(err)
}
}
/// Add best_of new requests to the queue and return a InferResponse of the sequence with
/// the highest log probability per token
#[instrument(skip(self, request))]
pub(crate) async fn generate_best_of(
&self,
request: GenerateRequest,
best_of: usize,
) -> Result<(InferResponse, Vec<InferResponse>), InferError> {
// validate best_of parameter separately
let best_of = self.validation.validate_best_of(best_of)?;
// create multiple generate requests
let mut infer_responses: Vec<InferResponse> =
try_join_all((0..best_of).map(|_| self.generate(request.clone()))).await?;
// get the sequence with the highest log probability per token
let mut max_index = 0;
let mut max_logprob: f32 = f32::MIN;
for (i, response) in infer_responses.iter().enumerate() {
// mean logprobs of the generated tokens
let sequence_logprob = response
.tokens
.iter()
.map(|token| token.logprob)
.sum::<f32>()
/ response.tokens.len() as f32;
// set best sequence
if sequence_logprob > max_logprob {
max_index = i;
max_logprob = sequence_logprob;
}
}
let best_response = infer_responses.remove(max_index);
Ok((best_response, infer_responses))
}
}
#[derive(Clone)]
struct ChatTemplate {
template: Template<'static, 'static>,
bos_token: Option<String>,
eos_token: Option<String>,
use_default_tool_template: bool,
}
impl ChatTemplate {
fn new(template: String, bos_token: Option<String>, eos_token: Option<String>) -> Self {
let mut env = Box::new(Environment::new());
let template_str = template.into_boxed_str();
env.add_function("raise_exception", raise_exception);
// check if contains the tools variable within the template
let use_default_tool_template =
!template_str.as_ref().replace(' ', "").contains("{{tools}}");
// leaking env and template_str as read-only, static resources for performance.
let template = Box::leak(env)
.template_from_str(Box::leak(template_str))
.unwrap();
Self {
template,
bos_token,
eos_token,
use_default_tool_template,
}
}
fn apply(
&self,
mut messages: Vec<Message>,
grammar_with_prompt: Option<(GrammarType, String)>,
) -> Result<String, InferError> {
if self.use_default_tool_template {
if let Some(last_message) = messages.last_mut() {
if let Some((GrammarType::Json(tools), tool_prompt)) = grammar_with_prompt {
last_message.content = Some(format!(
"{}\n---\n{}\n{}",
last_message.content.as_deref().unwrap_or_default(),
tool_prompt,
tools
));
}
}
}
self.template
.render(ChatTemplateInputs {
messages,
bos_token: self.bos_token.as_deref(),
eos_token: self.eos_token.as_deref(),
add_generation_prompt: true,
tools: None,
tools_prompt: None,
})
.map_err(InferError::TemplateError)
}
}
pub struct ToolGrammar {}
impl ToolGrammar {
pub fn apply(
tools: Option<Vec<Tool>>,
tool_choice: Option<ToolType>,
) -> Result<Option<Tools>, InferError> {
if let Some((req_tools, tool_choice)) = tools.zip(tool_choice) {
// let tool_prompt = tool_prompt.unwrap_or_default();
let tools_to_use = match tool_choice {
ToolType::FunctionName(name) => {
vec![req_tools
.iter()
.find(|tool| tool.function.name == *name)
.unwrap_or_else(|| panic!("Tool with name {} not found", name))
.clone()]
}
ToolType::OneOf => req_tools.to_owned(),
};
// adds the error notification function for LLM feedback if required
let mut text_response_properties = Map::new();
text_response_properties.insert(
"error".to_string(),
serde_json::json!({
"type": "string",
"description": "The error or issue to notify"
}),
);
text_response_properties.insert(
"_name".to_string(),
serde_json::json!({
"type": "string",
"const": "notify_error"
}),
);
let functions: HashMap<String, serde_json::Value> = tools_to_use
.iter()
.map(|tool| {
let func = tool.function.clone();
// Clone the existing parameters, which are expected to be a JSON object
let mut params = if let Value::Object(params) = &func.arguments {
params.clone()
} else {
Map::new()
};
// Insert the function's description at the top level, outside of properties
params.insert(
"description".to_string(),
Value::String(func.description.clone().unwrap_or_default()),
);
// Ensure 'properties' exists and is an object
let properties = params
.entry("properties".to_string())
.or_insert_with(|| json!({}))
.as_object_mut()
.unwrap();
// Insert the constant for the function name inside 'properties'
properties.insert(
"_name".to_string(),
json!({
"type": "string",
"const": func.name.clone(),
// "description": "The name of the function"
}),
);
// Check if 'required' exists, and it is an array. If not, create an empty array.
let required = params
.entry("required".to_string())
.or_insert_with(|| json!([]))
.as_array_mut()
.unwrap();
// Add 'name' to the 'required' array if it is not already present
if !required.iter().any(|r| r == "_name") {
required.push(json!("_name"));
}
(func.name, Value::Object(params))
})
.chain([(
"notify_error".to_string(),
serde_json::json!({
"properties": text_response_properties,
"required": ["error", "_name"],
"type": "object"
}),
)])
.collect();
let tools = Tools {
functions_map: FunctionsMap { functions },
properties: Properties {
function: tools_to_use
.iter()
.map(|tool| FunctionRef {
ref_path: format!("#/$functions/{}", tool.function.name.clone()),
})
.chain(std::iter::once(FunctionRef {
ref_path: "#/$functions/notify_error".to_string(),
}))
.collect(),
},
};
return Ok(Some(tools));
}
// Err(InferError::ToolError("No tools provided".to_string()))
Ok(None)
}
}
/// Batching logic
/// Will be launched in a background Tokio task
///
/// Batches requests and sends them to the inference server
#[allow(clippy::too_many_arguments)]
async fn batching_task(
mut client: ShardedClient,
waiting_served_ratio: f32,
max_batch_prefill_tokens: u32,
max_batch_total_tokens: u32,
max_waiting_tokens: usize,
max_batch_size: Option<usize>,
queue: Queue,
shared: Arc<Shared>,
generation_health: Arc<AtomicBool>,
) {
// Infinite loop
loop {
// Wait for a notification from the Infer struct
shared.batching_task.notified().await;
// Get the next batch from the queue
// This batch might be smaller than the maximum batch size if there are not enough requests
// waiting in the queue
while let Some((mut entries, batch, span)) = queue
.next_batch(
None,
max_batch_size,
max_batch_prefill_tokens,
max_batch_total_tokens,
)
.await
{
let mut cached_batch = prefill(&mut client, batch, &mut entries, &generation_health)
.instrument(span)
.await;
let mut waiting_tokens = 1;
// We loop until we do not receive any cached batch from the inference server (== until
// all requests have met their stopping criteria)
while let Some(batch) = cached_batch {
// Get current batch info
let batch_size = batch.size;
let batch_max_tokens = batch.max_tokens;
let mut batches = vec![batch];
metrics::gauge!("tgi_batch_current_size", batch_size as f64);
metrics::gauge!("tgi_batch_current_max_tokens", batch_max_tokens as f64);
let min_size = if waiting_tokens >= max_waiting_tokens {
// If we didn't onboard any new requests since >= max_waiting_tokens, we try
// to add a new batch even though its size might be small
None
} else {
// Minimum batch size
Some((batch_size as f32 * waiting_served_ratio).floor() as usize)
};
let token_budget = max_batch_total_tokens.saturating_sub(batch_max_tokens);
let max_size = max_batch_size.map(|max_size| max_size - batch_size as usize);
// Try to get a new batch
if let Some((mut new_entries, new_batch, span)) = queue
.next_batch(min_size, max_size, max_batch_prefill_tokens, token_budget)
.await
{
// Tracking metrics
if min_size.is_some() {
metrics::increment_counter!("tgi_batch_concat", "reason" => "backpressure");
} else {
metrics::increment_counter!("tgi_batch_concat", "reason" => "wait_exceeded");
}
entries.iter_mut().for_each(|(_, entry)| {
// Create a new span to add the info that this entry is waiting
// because a new batch is being computed
let entry_waiting_span = info_span!(parent: &entry.span, "waiting");
// Add relationships
span.follows_from(&entry_waiting_span);
entry_waiting_span.follows_from(&span);
// Update entry
entry.temp_span = Some(entry_waiting_span);
});
// Generate one token for this new batch to have the attention past in cache
let new_cached_batch =
prefill(&mut client, new_batch, &mut new_entries, &generation_health)
.instrument(span)
.await;
// Reset waiting counter
waiting_tokens = 1;
// Extend current batch with the new batch
if let Some(new_cached_batch) = new_cached_batch {
entries.extend(new_entries);
batches.push(new_cached_batch);
}
}
// Create span for this batch to add context to inference calls
let next_batch_size = entries.len();
let next_batch_span =
info_span!(parent: None, "batch", batch_size = next_batch_size);
entries.iter_mut().for_each(|(_, entry)| {
// Create a new span to link the batch back to this entry
let entry_batch_span = info_span!(parent: &entry.span, "infer");
// Add relationships
next_batch_span.follows_from(&entry_batch_span);
entry_batch_span.follows_from(&next_batch_span);
// Update entry
entry.temp_span = Some(entry_batch_span);
});
cached_batch = decode(&mut client, batches, &mut entries, &generation_health)
.instrument(next_batch_span)
.await;
waiting_tokens += 1;
}
metrics::gauge!("tgi_batch_current_size", 0.0);
metrics::gauge!("tgi_batch_current_max_tokens", 0.0);
}
}
}
#[instrument(skip_all)]
async fn prefill(
client: &mut ShardedClient,
batch: Batch,
entries: &mut IntMap<u64, Entry>,
generation_health: &Arc<AtomicBool>,
) -> Option<CachedBatch> {
let start_time = Instant::now();
let batch_id = batch.id;
metrics::increment_counter!("tgi_batch_inference_count", "method" => "prefill");
match client.prefill(batch).await {
Ok((generations, next_batch, timings)) => {
// Update health
generation_health.store(true, Ordering::SeqCst);
let start_filtering_time = Instant::now();
// Send generated tokens and filter stopped entries
filter_send_generations(generations, entries);
// Filter next batch and remove requests that were stopped
let next_batch = filter_batch(client, next_batch, entries).await;
metrics::histogram!("tgi_batch_forward_duration", timings.forward.as_secs_f64(), "method" => "prefill");
metrics::histogram!("tgi_batch_decode_duration", timings.decode.as_secs_f64(), "method" => "prefill");
metrics::histogram!("tgi_batch_filter_duration", start_filtering_time.elapsed().as_secs_f64(), "method" => "prefill");
metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "prefill");
metrics::increment_counter!("tgi_batch_inference_success", "method" => "prefill");
next_batch
}
// If we have an error, we discard the whole batch
Err(err) => {
// Update health
generation_health.store(false, Ordering::SeqCst);
let _ = client.clear_cache(Some(batch_id)).await;
send_errors(err, entries);
metrics::increment_counter!("tgi_batch_inference_failure", "method" => "prefill");
None
}
}
}
#[instrument(skip_all)]
async fn decode(
client: &mut ShardedClient,
batches: Vec<CachedBatch>,
entries: &mut IntMap<u64, Entry>,
generation_health: &Arc<AtomicBool>,
) -> Option<CachedBatch> {
let start_time = Instant::now();
let batch_ids: Vec<u64> = batches.iter().map(|b| b.id).collect();
metrics::increment_counter!("tgi_batch_inference_count", "method" => "decode");
match client.decode(batches).await {
Ok((generations, next_batch, timings)) => {
// Update health
generation_health.store(true, Ordering::SeqCst);
let start_filtering_time = Instant::now();
// Send generated tokens and filter stopped entries
filter_send_generations(generations, entries);
// Filter next batch and remove requests that were stopped
let next_batch = filter_batch(client, next_batch, entries).await;
if let Some(concat_duration) = timings.concat {
metrics::histogram!("tgi_batch_concat_duration", concat_duration.as_secs_f64(), "method" => "decode");
}
metrics::histogram!("tgi_batch_forward_duration", timings.forward.as_secs_f64(), "method" => "decode");
metrics::histogram!("tgi_batch_decode_duration", timings.decode.as_secs_f64(), "method" => "decode");
metrics::histogram!("tgi_batch_filter_duration", start_filtering_time.elapsed().as_secs_f64(), "method" => "decode");
metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "decode");
metrics::increment_counter!("tgi_batch_inference_success", "method" => "decode");
next_batch
}
// If we have an error, we discard the whole batch
Err(err) => {
generation_health.store(false, Ordering::SeqCst);
for id in batch_ids {
let _ = client.clear_cache(Some(id)).await;
}
send_errors(err, entries);
metrics::increment_counter!("tgi_batch_inference_failure", "method" => "decode");
None
}
}
}
/// Filter a `batch` and remove all requests not present in `entries`
#[instrument(skip_all)]
async fn filter_batch(
client: &mut ShardedClient,
next_batch: Option<CachedBatch>,
entries: &IntMap<u64, Entry>,
) -> Option<CachedBatch> {
let mut batch = next_batch?;
// No need to filter
if batch.size as usize == entries.len() {
return Some(batch);
}
let id = batch.id;
// Retain only requests that are still in entries
batch.request_ids.retain(|id| entries.contains_key(id));
if batch.request_ids.is_empty() {
// All requests have been filtered out
// Next batch is now empty
// Clear it from the Python shards cache
// We unwrap here as we need to panic since we cannot recover if this method fails
client.clear_cache(Some(id)).await.unwrap();
None
} else {
// Filter Python shard cache
// We unwrap here as we need to panic since we cannot recover if this method fails
client.filter_batch(id, batch.request_ids).await.unwrap()
}
}
/// Send one or multiple `InferStreamResponse` to Infer for all `entries`
/// and filter entries
#[instrument(skip_all)]
fn filter_send_generations(generations: Vec<Generation>, entries: &mut IntMap<u64, Entry>) {
generations.into_iter().for_each(|generation| {
let id = generation.request_id;
// Get entry
// We can `expect` here as the request id should always be in the entries
let entry = entries
.get(&id)
.expect("ID not found in entries. This is a bug.");
// Create and enter a span to link this function back to the entry
let _span = info_span!(parent: entry.temp_span.as_ref().expect("batch_span is None. This is a bug."), "send_generation", generation = ?generation).entered();
// Send generation responses back to the infer task
// If the receive an error from the Flume channel, it means that the client dropped the
// request and we need to stop generating hence why we unwrap_or(true)
let stopped = send_responses(generation, entry).map_err(|err| {
tracing::error!("Entry response channel error.");
metrics::increment_counter!("tgi_request_failure", "err" => "dropped");
err
}).unwrap_or(true);
if stopped {
entries.remove(&id).expect("ID not found in entries. This is a bug.");
}
});
}
/// Send responses through the `entry` response channel
fn send_responses(
generation: Generation,
entry: &Entry,
) -> Result<bool, Box<SendError<Result<InferStreamResponse, InferError>>>> {
// Return directly if the channel is disconnected
if entry.response_tx.is_closed() {
metrics::increment_counter!("tgi_request_failure", "err" => "dropped");
return Ok(true);
}
let mut stopped = false;
if let Some(prefill_tokens) = generation.prefill_tokens {
// Send message
entry
.response_tx
.send(Ok(InferStreamResponse::Prefill(prefill_tokens)))?;
}
// Create last Token
let tokens_ = generation.tokens.expect("Non empty tokens in generation");
let n = tokens_.ids.len();
metrics::histogram!("tgi_request_skipped_tokens", (n - 1) as f64);
let mut iterator = tokens_
.ids
.into_iter()
.zip(tokens_.logprobs)
.zip(tokens_.texts)
.zip(tokens_.is_special)
.enumerate()
.peekable();
while let Some((i, (((id, logprob), text), special))) = iterator.next() {
let token = Token {
id,
text,
logprob,
special,
};
let top_tokens = if let Some(top_tokens_) = generation.top_tokens.get(i) {
top_tokens_
.ids
.iter()
.zip(top_tokens_.logprobs.iter())
.zip(top_tokens_.texts.iter())
.zip(top_tokens_.is_special.iter())
.map(|(((&id, &logprob), text), &special)| Token {
id,
text: text.to_string(),
logprob,
special,
})
.collect()
} else {
vec![]
};
match (&generation.generated_text, iterator.peek()) {
(Some(generated_text), None) => {
// Generation has ended
stopped = true;
// Send message
entry.response_tx.send(Ok(InferStreamResponse::End {
token,
top_tokens,
generated_text: generated_text.clone(),
queued: entry.queue_time,
start: entry.batch_time.unwrap(),
}))?;
}
_ => {
// Send message
entry
.response_tx
.send(Ok(InferStreamResponse::Intermediate { token, top_tokens }))?;
}
}
}
Ok(stopped)
}
/// Send errors to Infer for all `entries`
#[instrument(skip_all)]
fn send_errors(error: ClientError, entries: &mut IntMap<u64, Entry>) {
entries.drain().for_each(|(_, entry)| {
// Create and enter a span to link this function back to the entry
let _send_error_span = info_span!(parent: entry.temp_span.as_ref().expect("batch_span is None. This is a bug."), "send_error").entered();
let err = InferError::GenerationError(error.to_string());
metrics::increment_counter!("tgi_request_failure", "err" => "generation");
tracing::error!("{err}");
// unwrap_or is valid here as we don't care if the receiver is gone.
entry
.response_tx
.send(Err(err))
.unwrap_or(());
});
}
#[derive(Debug)]
pub(crate) enum InferStreamResponse {
// Optional first message
Prefill(Tokens),
// Intermediate messages
Intermediate {
token: Token,
top_tokens: Vec<Token>,
},
// Last message
End {
token: Token,
top_tokens: Vec<Token>,
generated_text: GeneratedText,
start: Instant,
queued: Instant,
},
}
#[derive(Debug)]
pub(crate) struct InferResponse {
/// input_length is the input as perceived by the rust tokenizer in the
/// validation pathway. It is redundant with prefill.len() but prefill
/// has data only if the user asked for it. This will always be filled.
pub(crate) _input_length: u32,
pub(crate) prefill: Vec<PrefillToken>,
pub(crate) tokens: Vec<Token>,
pub(crate) generated_text: GeneratedText,
pub(crate) queued: Instant,
pub(crate) start: Instant,
pub(crate) top_tokens: Vec<Vec<Token>>,
}
#[derive(Debug, Error)]
pub enum InferError {
#[error("Request failed during generation: {0}")]
GenerationError(String),
#[error("Model is overloaded")]
Overloaded(#[from] TryAcquireError),
#[error("Input validation error: {0}")]
ValidationError(#[from] ValidationError),
#[error("Incomplete generation")]
IncompleteGeneration,
#[error("Template error: {0}")]
TemplateError(#[from] minijinja::Error),
#[error("Tool error: {0}")]
ToolError(String),
}
impl InferError {
pub(crate) fn error_type(&self) -> &str {
match self {
InferError::GenerationError(_) => "generation",
InferError::Overloaded(_) => "overloaded",
InferError::ValidationError(_) => "validation",
InferError::IncompleteGeneration => "incomplete_generation",
InferError::TemplateError(_) => "template_error",
InferError::ToolError(_) => "tool_error",
}
}
}
// tests
#[cfg(test)]
mod tests {
use crate::infer::raise_exception;
use crate::ChatTemplateInputs;
use crate::Message;
use minijinja::Environment;
#[test]
fn test_chat_template() {
let env = Environment::new();
let source = r#"
{% for message in messages %}
{% if message['role'] == 'system' %}
{% if message['content']%}
{{'### System:\n' + message['content']+'\n\n'}}
{% endif %}
{% elif message['role'] == 'user' %}
{{'### User:\n' + message['content']+'\n\n'}}
{% elif message['role'] == 'assistant' %}
{{'### Assistant:\n' + message['content']}}
{% endif %}
{% if loop.last and add_generation_prompt %}
{{ '### Assistant:\n' }}
{% endif %}
{% endfor %}"#;
// trim all the whitespace
let source = source
.lines()
.map(|line| line.trim())
.collect::<Vec<&str>>()
.join("");
let tmpl = env.template_from_str(&source);
let chat_template_inputs = ChatTemplateInputs {
messages: vec![
Message {
role: "user".to_string(),
content: Some("Hi!".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("Hello how can I help?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("What is Deep Learning?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("magic!".to_string()),
name: None,
tool_calls: None,
},
],
bos_token: Some("[BOS]"),
eos_token: Some("[EOS]"),
add_generation_prompt: true,
..Default::default()
};
let result = tmpl.unwrap().render(chat_template_inputs).unwrap();
assert_eq!(
result,
"### User:\nHi!\n\n### Assistant:\nHello how can I help?### User:\nWhat is Deep Learning?\n\n### Assistant:\nmagic!### Assistant:\n"
);
}
#[test]
fn test_chat_template_invalid_with_raise() {
let mut env = Environment::new();
env.add_function("raise_exception", raise_exception);
let source = r#"
{{ bos_token }}
{% for message in messages %}
{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}
{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}
{% endif %}
{% if message['role'] == 'user' %}
{{ '[INST] ' + message['content'] + ' [/INST]' }}
{% elif message['role'] == 'assistant' %}
{{ message['content'] + eos_token}}
{% else %}
{{ raise_exception('Only user and assistant roles are supported!') }}
{% endif %}
{% endfor %}"#;
// trim all the whitespace
let source = source
.lines()
.map(|line| line.trim())
.collect::<Vec<&str>>()
.join("");
let tmpl = env.template_from_str(&source);
let chat_template_inputs = ChatTemplateInputs {
messages: vec![
Message {
role: "user".to_string(),
content: Some("Hi!".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("Hi again!".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("Hello how can I help?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("What is Deep Learning?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("magic!".to_string()),
name: None,
tool_calls: None,
},
],
bos_token: Some("[BOS]"),
eos_token: Some("[EOS]"),
add_generation_prompt: true,
..Default::default()
};
let result = tmpl.unwrap().render(chat_template_inputs); //.err().unwrap();
match result {
Ok(_) => panic!("Should have failed"),
Err(e) => {
assert_eq!(
e.detail().unwrap(),
"Conversation roles must alternate user/assistant/user/assistant/..."
);
}
}
}
#[test]
fn test_chat_template_valid_with_raise() {
let mut env = Environment::new();
env.add_function("raise_exception", raise_exception);
let source = r#"
{{ bos_token }}
{% for message in messages %}
{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}
{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}
{% endif %}
{% if message['role'] == 'user' %}
{{ '[INST] ' + message['content'] + ' [/INST]' }}
{% elif message['role'] == 'assistant' %}
{{ message['content'] + eos_token}}
{% else %}
{{ raise_exception('Only user and assistant roles are supported!') }}
{% endif %}
{% endfor %}"#;
// trim all the whitespace
let source = source
.lines()
.map(|line| line.trim())
.collect::<Vec<&str>>()
.join("");
let tmpl = env.template_from_str(&source);
let chat_template_inputs = ChatTemplateInputs {
messages: vec![
Message {
role: "user".to_string(),
content: Some("Hi!".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("Hello how can I help?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("What is Deep Learning?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("magic!".to_string()),
name: None,
tool_calls: None,
},
],
bos_token: Some("[BOS]"),
eos_token: Some("[EOS]"),
add_generation_prompt: true,
..Default::default()
};
let result = tmpl.unwrap().render(chat_template_inputs).unwrap();
assert_eq!(result, "[BOS][INST] Hi! [/INST]Hello how can I help?[EOS][INST] What is Deep Learning? [/INST]magic![EOS]");
}
#[test]
fn test_chat_template_valid_with_add_generation_prompt() {
let mut env = Environment::new();
env.add_function("raise_exception", raise_exception);
let source = r#"
{% for message in messages %}
{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
{% endfor %}
{% if add_generation_prompt %}
{{ '<|im_start|>assistant\n' }}
{% endif %}"#;
// trim all the whitespace
let source = source
.lines()
.map(|line| line.trim())
.collect::<Vec<&str>>()
.join("");
let tmpl = env.template_from_str(&source);
let chat_template_inputs = ChatTemplateInputs {
messages: vec![
Message {
role: "user".to_string(),
content: Some("Hi!".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("Hello how can I help?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("What is Deep Learning?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("magic!".to_string()),
name: None,
tool_calls: None,
},
],
bos_token: Some("[BOS]"),
eos_token: Some("[EOS]"),
add_generation_prompt: true,
..Default::default()
};
let result = tmpl.unwrap().render(chat_template_inputs).unwrap();
assert_eq!(result, "<|im_start|>user\nHi!<|im_end|>\n<|im_start|>assistant\nHello how can I help?<|im_end|>\n<|im_start|>user\nWhat is Deep Learning?<|im_end|>\n<|im_start|>assistant\nmagic!<|im_end|>\n<|im_start|>assistant\n");
}
struct ChatTemplateTestItem {
name: &'static str,
chat_template: &'static str,
input: ChatTemplateInputs<'static>,
target: &'static str,
}
#[test]
fn test_many_chat_templates() {
let example_chat = vec![
Message {
role: "user".to_string(),
content: Some("Hello, how are you?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "assistant".to_string(),
content: Some("I'm doing great. How can I help you today?".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("I'd like to show off how chat templating works!".to_string()),
name: None,
tool_calls: None,
},
];
let example_chat_with_system = vec![Message {
role: "system".to_string(),
content: Some(
"You are a friendly chatbot who always responds in the style of a pirate"
.to_string(),
),
name: None,
tool_calls: None,
}]
.iter()
.chain(&example_chat)
.cloned()
.collect::<Vec<_>>();
let test_default_templates = vec![
ChatTemplateTestItem {
name: "_base",
chat_template: "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some(""),
..Default::default()
},
target: "<|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing great. How can I help you today?<|im_end|>\n<|im_start|>user\nI'd like to show off how chat templating works!<|im_end|>\n",
},
ChatTemplateTestItem {
name: "blenderbot",
chat_template: "{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("</s>"),
..Default::default()
},
target: " Hello, how are you? I'm doing great. How can I help you today? I'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "blenderbot_small",
chat_template: "{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("</s>"),
..Default::default()
},
target: " Hello, how are you? I'm doing great. How can I help you today? I'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "bloom",
chat_template: "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("</s>"),
..Default::default()
},
target: "Hello, how are you?</s>I'm doing great. How can I help you today?</s>I'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "gpt_neox",
chat_template: "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("<|endoftext|>"),
..Default::default()
},
target: "Hello, how are you?<|endoftext|>I'm doing great. How can I help you today?<|endoftext|>I'd like to show off how chat templating works!<|endoftext|>",
},
ChatTemplateTestItem {
name: "gpt2",
chat_template: "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("<|endoftext|>"),
..Default::default()
},
target: "Hello, how are you?<|endoftext|>I'm doing great. How can I help you today?<|endoftext|>I'd like to show off how chat templating works!<|endoftext|>",
},
ChatTemplateTestItem {
name: "llama",
// NOTE: the `.strip()` has been replaced with `| trim` in the following template
chat_template: "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif USE_DEFAULT_PROMPT == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\n' + system_message + '\\n<</SYS>>\\n\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token +'[INST] ' + content | trim + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\n' + content | trim + '\\n<</SYS>>\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content | trim + ' ' + eos_token }}{% endif %}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat_with_system.clone(),
add_generation_prompt: true,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s>[INST] <<SYS>>\nYou are a friendly chatbot who always responds in the style of a pirate\n<</SYS>>\n\nHello, how are you? [/INST] I'm doing great. How can I help you today? </s><s>[INST] I'd like to show off how chat templating works! [/INST]",
},
ChatTemplateTestItem {
name: "whisper",
chat_template: "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: true,
bos_token: Some(""),
eos_token: Some("<|endoftext|>"),
..Default::default()
},
target: "Hello, how are you?<|endoftext|>I'm doing great. How can I help you today?<|endoftext|>I'd like to show off how chat templating works!<|endoftext|>",
},
];
#[allow(unused_variables)] // name is unused
for ChatTemplateTestItem {
name,
chat_template,
input,
target,
} in test_default_templates
{
let mut env = Environment::new();
env.add_function("raise_exception", raise_exception);
let tmpl = env.template_from_str(&chat_template);
let result = tmpl.unwrap().render(input).unwrap();
assert_eq!(result, target);
}
let test_custom_templates = vec![
ChatTemplateTestItem {
name: "HuggingFaceH4/zephyr-7b-beta (add_generation_prompt=false)",
chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat_with_system.clone(),
add_generation_prompt: false,
bos_token: Some(""),
eos_token: Some("</s>"),
..Default::default()
},
target: "<|system|>\nYou are a friendly chatbot who always responds in the style of a pirate</s><|user|>\nHello, how are you?</s><|assistant|>\nI'm doing great. How can I help you today?</s><|user|>\nI'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "HuggingFaceH4/zephyr-7b-beta (add_generation_prompt=true)",
chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}",
input: ChatTemplateInputs {
messages: vec![
Message {
role: "system".to_string(),
content: Some("You are a friendly chatbot who always responds in the style of a pirate".to_string()),
name: None,
tool_calls: None,
},
Message {
role: "user".to_string(),
content: Some("How many helicopters can a human eat in one sitting?".to_string()),
name: None,
tool_calls: None,
},
],
add_generation_prompt: true,
bos_token: Some(""),
eos_token: Some("</s>"),
..Default::default()
},
target: "<|system|>\nYou are a friendly chatbot who always responds in the style of a pirate</s><|user|>\nHow many helicopters can a human eat in one sitting?</s><|assistant|>",
},
ChatTemplateTestItem {
name: "HuggingFaceH4/zephyr-7b-gemma-v0.1",
chat_template: "{% if messages[0]['role'] == 'user' or messages[0]['role'] == 'system' %}{{ bos_token }}{% endif %}{% for message in messages %}{{ '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% elif messages[-1]['role'] == 'assistant' %}{{ eos_token }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<bos>"),
eos_token: Some("<eos>"),
..Default::default()
},
target: "<bos><|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing great. How can I help you today?<|im_end|>\n<|im_start|>user\nI'd like to show off how chat templating works!<|im_end|>\n",
},
ChatTemplateTestItem {
name: "mistralai/Mistral-7B-Instruct-v0.1",
chat_template: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token + ' ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s>[INST] Hello, how are you? [/INST]I'm doing great. How can I help you today?</s> [INST] I'd like to show off how chat templating works! [/INST]",
},
ChatTemplateTestItem {
name: "mistralai/Mixtral-8x7B-Instruct-v0.1",
chat_template: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s>[INST] Hello, how are you? [/INST]I'm doing great. How can I help you today?</s>[INST] I'd like to show off how chat templating works! [/INST]",
},
ChatTemplateTestItem {
name: "cognitivecomputations/dolphin-2.5-mixtral-8x7b",
chat_template: "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing great. How can I help you today?<|im_end|>\n<|im_start|>user\nI'd like to show off how chat templating works!<|im_end|>\n",
},
ChatTemplateTestItem {
name: "openchat/openchat-3.5-0106",
// `.title()` has been replaced with `| upper` in the following template
chat_template: "{{ bos_token }}{% for message in messages %}{{ 'GPT4 Correct ' + (message['role'] | title) + ': ' + message['content'] + '<|end_of_turn|>'}}{% endfor %}{% if add_generation_prompt %}{{ 'GPT4 Correct Assistant:' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s>GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT4 Correct User: I'd like to show off how chat templating works!<|end_of_turn|>",
},
ChatTemplateTestItem {
name: "upstage/SOLAR-10.7B-Instruct-v1.0",
chat_template: "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "Hello, how are you?</s>I'm doing great. How can I help you today?</s>I'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "codellama/CodeLlama-70b-Instruct-hf",
// NOTE: `.strip()` has been replaced with `| trim` in the following template
chat_template: "{% if messages[0]['role'] == 'system' %}{% set user_index = 1 %}{% else %}{% set user_index = 0 %}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != ((loop.index0 + user_index) % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 %}{{ '<s>' }}{% endif %}{% set content = 'Source: ' + message['role'] + '\\n\\n ' + message['content'] | trim %}{{ content + ' <step> ' }}{% endfor %}{{'Source: assistant\\nDestination: user\\n\\n '}}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s>Source: user\n\n Hello, how are you? <step> Source: assistant\n\n I'm doing great. How can I help you today? <step> Source: user\n\n I'd like to show off how chat templating works! <step> Source: assistant\nDestination: user\n\n ",
},
ChatTemplateTestItem {
name: "Deci/DeciLM-7B-instruct",
chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '### User:\\n' + message['content'] }}\n{% elif message['role'] == 'system' %}\n{{ '### System:\\n' + message['content'] }}\n{% elif message['role'] == 'assistant' %}\n{{ '### Assistant:\\n' + message['content'] }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '### Assistant:' }}\n{% endif %}\n{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "### User:\nHello, how are you?### Assistant:\nI'm doing great. How can I help you today?### User:\nI'd like to show off how chat templating works!",
},
ChatTemplateTestItem {
name: "Qwen/Qwen1.5-72B-Chat",
chat_template: "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system\\nYou are a helpful assistant<|im_end|>\\n' }}{% endif %}{{'<|im_start|>' + message['role'] + '\\n' + message['content']}}{% if (loop.last and add_generation_prompt) or not loop.last %}{{ '<|im_end|>' + '\\n'}}{% endif %}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<|im_start|>system\nYou are a helpful assistant<|im_end|>\n<|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing great. How can I help you today?<|im_end|>\n<|im_start|>user\nI'd like to show off how chat templating works!",
},
ChatTemplateTestItem {
name: "deepseek-ai/deepseek-llm-7b-chat",
chat_template: "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ 'User: ' + message['content'] + '\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'] + '\\n\\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<｜begin▁of▁sentence｜>"),
eos_token: Some("<｜end▁of▁sentence｜>"),
..Default::default()
},
target: "<｜begin▁of▁sentence｜>User: Hello, how are you?\n\nAssistant: I'm doing great. How can I help you today?<｜end▁of▁sentence｜>User: I'd like to show off how chat templating works!\n\n",
},
ChatTemplateTestItem {
name: "h2oai/h2o-danube-1.8b-chat",
chat_template: "{% for message in messages %}{% if message['role'] == 'user' %}{{ '<|prompt|>' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ '<|system|>' + message['content'] + eos_token }}{% elif message['role'] == 'assistant' %}{{ '<|answer|>' + message['content'] + eos_token }}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|answer|>' }}{% endif %}{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<|prompt|>Hello, how are you?</s><|answer|>I'm doing great. How can I help you today?</s><|prompt|>I'd like to show off how chat templating works!</s>",
},
ChatTemplateTestItem {
name: "internlm/internlm2-chat-7b",
chat_template: "{% if messages[0]['role'] == 'user' or messages[0]['role'] == 'system' %}{{ bos_token }}{% endif %}{% for message in messages %}{{ '<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% elif messages[-1]['role'] == 'assistant' %}{{ eos_token }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "<s><|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\nI'm doing great. How can I help you today?<|im_end|>\n<|im_start|>user\nI'd like to show off how chat templating works!<|im_end|>\n",
},
ChatTemplateTestItem {
name: "TheBloke/deepseek-coder-33B-instruct-AWQ",
chat_template: "{%- set found_item = false -%}\n{%- for message in messages -%}\n {%- if message['role'] == 'system' -%}\n {%- set found_item = true -%}\n {%- endif -%}\n{%- endfor -%}\n{%- if not found_item -%}\n{{'You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.\\n'}}\n{%- endif %}\n{%- for message in messages %}\n {%- if message['role'] == 'system' %}\n{{ message['content'] }}\n {%- else %}\n {%- if message['role'] == 'user' %}\n{{'### Instruction:\\n' + message['content'] + '\\n'}}\n {%- else %}\n{{'### Response:\\n' + message['content'] + '\\n<|EOT|>\\n'}}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{{'### Response:\\n'}}\n",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<｜begin▁of▁sentence｜>"),
eos_token: Some("<|EOT|>"),
..Default::default()
},
target: "You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.\n### Instruction:\nHello, how are you?\n### Response:\nI'm doing great. How can I help you today?\n<|EOT|>\n### Instruction:\nI'd like to show off how chat templating works!\n### Response:\n",
},
ChatTemplateTestItem {
name: "ericzzz/falcon-rw-1b-chat",
// `.strip()` has been replaced with `| trim` in the following template
chat_template: "{% for message in messages %}{% if loop.index > 1 and loop.previtem['role'] != 'assistant' %}{{ ' ' }}{% endif %}{% if message['role'] == 'system' %}{{ '[SYS] ' + message['content'] | trim }}{% elif message['role'] == 'user' %}{{ '[INST] ' + message['content'] | trim }}{% elif message['role'] == 'assistant' %}{{ '[RESP] ' + message['content'] + eos_token }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ ' [RESP] ' }}{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<|endoftext|>"),
eos_token: Some("<|endoftext|>"),
..Default::default()
},
target: "[INST] Hello, how are you? [RESP] I'm doing great. How can I help you today?<|endoftext|>[INST] I'd like to show off how chat templating works!",
},
ChatTemplateTestItem {
name: "abacusai/Smaug-34B-v0.1",
chat_template: "{%- for idx in range(0, messages|length) -%}\n{%- if messages[idx]['role'] == 'user' -%}\n{%- if idx > 1 -%}\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\n{%- else -%}\n{{- messages[idx]['content'] + ' [/INST]' -}}\n{%- endif -%}\n{% elif messages[idx]['role'] == 'system' %}\n{{- '[INST] <<SYS>>\\n' + messages[idx]['content'] + '\\n<</SYS>>\\n\\n' -}}\n{%- elif messages[idx]['role'] == 'assistant' -%}\n{{- ' ' + messages[idx]['content'] + ' ' + eos_token -}}\n{% endif %}\n{% endfor %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "Hello, how are you? [/INST] I'm doing great. How can I help you today? </s><s>[INST] I'd like to show off how chat templating works! [/INST]",
},
ChatTemplateTestItem {
name: "maywell/Synatra-Mixtral-8x7B",
chat_template: "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n{% for message in messages %}{% if message['role'] == 'user' %}### Instruction:\n{{ message['content']|trim -}}{% if not loop.last %}{% endif %}\n{% elif message['role'] == 'assistant' %}### Response:\n{{ message['content']|trim -}}{% if not loop.last %}{% endif %}\n{% elif message['role'] == 'system' %}{{ message['content']|trim -}}{% if not loop.last %}{% endif %}\n{% endif %}\n{% endfor %}\n{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}\n### Response:\n{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "Below is an instruction that describes a task. Write a response that appropriately completes the request.### Instruction:Hello, how are you?### Response:I'm doing great. How can I help you today?### Instruction:I'd like to show off how chat templating works!",
},
ChatTemplateTestItem {
name: "deepseek-ai/deepseek-coder-33b-instruct",
chat_template: "{% if not add_generation_prompt is defined %}\n{% set add_generation_prompt = false %}\n{% endif %}\n{%- set ns = namespace(found=false) -%}\n{%- for message in messages -%}\n {%- if message['role'] == 'system' -%}\n {%- set ns.found = true -%}\n {%- endif -%}\n{%- endfor -%}\n{{bos_token}}{%- if not ns.found -%}\n{{'You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer\\n'}}\n{%- endif %}\n{%- for message in messages %}\n {%- if message['role'] == 'system' %}\n{{ message['content'] }}\n {%- else %}\n {%- if message['role'] == 'user' %}\n{{'### Instruction:\\n' + message['content'] + '\\n'}}\n {%- else %}\n{{'### Response:\\n' + message['content'] + '\\n<|EOT|>\\n'}}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{% if add_generation_prompt %}\n{{'### Response:'}}\n{% endif %}",
input: ChatTemplateInputs {
messages: example_chat.clone(),
add_generation_prompt: false,
bos_token: Some("<｜begin▁of▁sentence｜>"),
eos_token: Some("</EOT>"),
..Default::default()
},
target: "<｜begin▁of▁sentence｜>You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer\n### Instruction:\nHello, how are you?\n### Response:\nI'm doing great. How can I help you today?\n<|EOT|>\n### Instruction:\nI'd like to show off how chat templating works!\n",
},
// NOT INCLUDED
// - meetkai/functionary-medium-v2.2
// - fireworks-ai/firefunction-v1
// https://github
ChatTemplateTestItem {
name: "maywell/PiVoT-MoE",
chat_template: "{{ (messages|selectattr('role', 'equalto', 'system')|list|last).content|trim if (messages|selectattr('role', 'equalto', 'system')|list) else '' }}{% for message in messages %}{% if message['role'] == 'system' %}{{ message['content']|trim }}{% elif message['role'] == 'user' %}### Instruction: {{ message['content']|trim }}{% elif message['role'] == 'assistant' %}### Response: {{ message['content']|trim }}{% elif message['role'] == 'user_context' %}### Input: {{ message['content']|trim }}{% endif %}{% if not loop.last %}\n{% endif %}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}### Response:{% endif %}",
input: ChatTemplateInputs {
messages: example_chat_with_system.clone(),
add_generation_prompt: false,
bos_token: Some("<s>"),
eos_token: Some("</s>"),
..Default::default()
},
target: "You are a friendly chatbot who always responds in the style of a pirateYou are a friendly chatbot who always responds in the style of a pirate### Instruction: Hello, how are you?### Response: I'm doing great. How can I help you today?### Instruction: I'd like to show off how chat templating works!",
},
];
#[allow(unused_variables)] // name is unused
for ChatTemplateTestItem {
name,
chat_template,
input,
target,
} in test_custom_templates
{
let mut env = Environment::new();
env.add_function("raise_exception", raise_exception);
// trim all the whitespace
let chat_template = chat_template
.lines()
.map(|line| line.trim())
.collect::<Vec<&str>>()
.join("");
let tmpl = env.template_from_str(&chat_template);
let result = tmpl.unwrap().render(input).unwrap();
assert_eq!(result, target);
}
}
}
router/src/queue.rs deleted 100644 → 0
View file @ 64def8e2
use crate::infer::InferError;
use crate::infer::InferStreamResponse;
use crate::validation::ValidGenerateRequest;
use nohash_hasher::{BuildNoHashHasher, IntMap};
use std::cmp::min;
use std::collections::VecDeque;
use text_generation_client::{Batch, Request};
use tokio::sync::{mpsc, oneshot};
use tokio::time::Instant;
use tracing::{info_span, instrument, Span};
/// Queue entry
#[derive(Debug)]
pub(crate) struct Entry {
/// Request
pub request: ValidGenerateRequest,
/// Response sender to communicate between the Infer struct and the batching_task
pub response_tx: mpsc::UnboundedSender<Result<InferStreamResponse, InferError>>,
/// Span that will live as long as entry
pub span: Span,
/// Temporary span used as a guard when logging inference, wait times...
pub temp_span: Option<Span>,
/// Instant when this entry was queued
pub queue_time: Instant,
/// Instant when this entry was added to a batch
pub batch_time: Option<Instant>,
}
/// Request Queue
#[derive(Debug, Clone)]
pub(crate) struct Queue {
/// Channel to communicate with the background queue task
queue_sender: mpsc::UnboundedSender<QueueCommand>,
}
impl Queue {
pub(crate) fn new(
requires_padding: bool,
block_size: u32,
window_size: Option<u32>,
speculate: u32,
) -> Self {
// Create channel
let (queue_sender, queue_receiver) = mpsc::unbounded_channel();
// Launch background queue task
tokio::spawn(queue_task(
requires_padding,
block_size,
window_size,
speculate,
queue_receiver,
));
Self { queue_sender }
}
/// Append an entry to the queue
#[instrument(skip_all)]
pub(crate) fn append(&self, entry: Entry) {
// Send append command to the background task managing the state
// Unwrap is safe here
self.queue_sender
.send(QueueCommand::Append(Box::new(entry), Span::current()))
.unwrap();
}
// Get the next batch
#[instrument(skip(self))]
pub(crate) async fn next_batch(
&self,
min_size: Option<usize>,
max_size: Option<usize>,
prefill_token_budget: u32,
token_budget: u32,
) -> Option<NextBatch> {
// Create response channel
let (response_sender, response_receiver) = oneshot::channel();
// Send next batch command to the background task managing the state
// Unwrap is safe here
self.queue_sender
.send(QueueCommand::NextBatch {
min_size,
max_size,
prefill_token_budget,
token_budget,
response_sender,
span: Span::current(),
})
.unwrap();
// Await on response channel
// Unwrap is safe here
response_receiver.await.unwrap()
}
}
// Background task responsible of the queue state
async fn queue_task(
requires_padding: bool,
block_size: u32,
window_size: Option<u32>,
speculate: u32,
mut receiver: mpsc::UnboundedReceiver<QueueCommand>,
) {
let mut state = State::new(requires_padding, block_size, window_size, speculate);
while let Some(cmd) = receiver.recv().await {
match cmd {
QueueCommand::Append(entry, span) => {
span.in_scope(|| state.append(*entry));
metrics::increment_gauge!("tgi_queue_size", 1.0);
}
QueueCommand::NextBatch {
min_size,
max_size,
prefill_token_budget,
token_budget,
response_sender,
span,
} => span.in_scope(|| {
let next_batch =
state.next_batch(min_size, max_size, prefill_token_budget, token_budget);
response_sender.send(next_batch).unwrap();
metrics::gauge!("tgi_queue_size", state.entries.len() as f64);
}),
}
}
}
/// Queue State
#[derive(Debug)]
struct State {
/// Queue entries organized in a Vec
entries: VecDeque<(u64, Entry)>,
/// Id of the next entry
next_id: u64,
/// Id of the next batch
next_batch_id: u64,
/// Whether the model is using padding
requires_padding: bool,
/// Paged Attention block size
block_size: u32,
/// Sliding window
window_size: Option<u32>,
/// Speculation amount
speculate: u32,
}
impl State {
fn new(
requires_padding: bool,
block_size: u32,
window_size: Option<u32>,
speculate: u32,
) -> Self {
Self {
entries: VecDeque::with_capacity(128),
next_id: 0,
next_batch_id: 0,
requires_padding,
block_size,
window_size,
speculate,
}
}
/// Append an entry to the queue
fn append(&mut self, mut entry: Entry) {
// Create a span that will live as long as the entry is in the queue waiting to be batched
let queue_span = info_span!(parent: &entry.span, "queued");
entry.temp_span = Some(queue_span);
// Push entry in the queue
self.entries.push_back((self.next_id, entry));
self.next_id += 1;
}
// Get the next batch
fn next_batch(
&mut self,
min_size: Option<usize>,
max_size: Option<usize>,
prefill_token_budget: u32,
token_budget: u32,
) -> Option<NextBatch> {
if self.entries.is_empty() {
tracing::debug!("No queue");
return None;
}
// Check if we have enough entries
if let Some(min_size) = min_size {
if self.entries.len() < min_size {
tracing::debug!("Not enough entries");
return None;
}
}
// Pad prefill_token_budget to be a multiple of block size
let prefill_token_budget =
((prefill_token_budget + self.block_size - 1) / self.block_size) * self.block_size;
// Create span for this batch to add context to inference calls
let next_batch_span = info_span!(parent: None, "batch", batch_size = tracing::field::Empty);
next_batch_span.follows_from(&Span::current());
let mut batch_requests = Vec::with_capacity(self.entries.len());
let mut batch_entries =
IntMap::with_capacity_and_hasher(self.entries.len(), BuildNoHashHasher::default());
let mut max_input_length = 0;
let mut prefill_tokens: u32 = 0;
let mut decode_tokens: u32 = 0;
// Pop entries starting from the front of the queue
while let Some((id, mut entry)) = self.entries.pop_front() {
// Filter entries where the response receiver was dropped (== entries where the request
// was dropped by the client)
if entry.response_tx.is_closed() {
metrics::increment_counter!("tgi_request_failure", "err" => "dropped");
tracing::debug!("Dropping entry");
continue;
}
if self.requires_padding {
// We pad to max input length in the Python shards
// We need to take these padding tokens into the equation
max_input_length = max_input_length.max(entry.request.input_length);
prefill_tokens = (batch_requests.len() + 1) as u32 * max_input_length
} else {
// pad to block size
prefill_tokens += ((entry.request.input_length + self.block_size - 1)
/ self.block_size)
* self.block_size;
}
if self.requires_padding {
decode_tokens += entry.request.stopping_parameters.max_new_tokens;
} else {
let max_new_tokens = match self.window_size {
None => entry.request.stopping_parameters.max_new_tokens,
Some(window_size) => min(
window_size.saturating_sub(entry.request.input_length),
entry.request.stopping_parameters.max_new_tokens,
),
};
// pad to block size
decode_tokens +=
((max_new_tokens + self.block_size - 1) / self.block_size) * self.block_size;
}
if prefill_tokens > prefill_token_budget
|| (prefill_tokens + decode_tokens + self.speculate) > token_budget
{
// Entry is over budget
// Add it back to the front
tracing::debug!("Over budget: prefill_tokens={prefill_tokens} > {prefill_token_budget} || {prefill_tokens} + {decode_tokens} + {} > {token_budget}", self.speculate);
self.entries.push_front((id, entry));
break;
}
tracing::debug!("Accepting entry");
// Create a new span to link the batch back to this entry
let entry_batch_span = info_span!(parent: &entry.span, "infer");
// Add relationships
next_batch_span.follows_from(&entry_batch_span);
entry_batch_span.follows_from(&next_batch_span);
// Update entry
entry.temp_span = Some(entry_batch_span);
batch_requests.push(Request {
id,
prefill_logprobs: entry.request.decoder_input_details,
inputs: entry.request.inputs.clone(),
truncate: entry.request.truncate,
parameters: Some(entry.request.parameters.clone()),
stopping_parameters: Some(entry.request.stopping_parameters.clone()),
top_n_tokens: entry.request.top_n_tokens,
});
// Set batch_time
entry.batch_time = Some(Instant::now());
// Insert in batch_entries IntMap
batch_entries.insert(id, entry);
// Check if max_size
if Some(batch_requests.len()) == max_size {
break;
}
}
// Empty batch
if batch_requests.is_empty() {
tracing::debug!("Filterered out all entries");
return None;
}
// Check if our batch is big enough
if let Some(min_size) = min_size {
// Batch is too small
if batch_requests.len() < min_size {
// Add back entries to the queue in the correct order
for r in batch_requests.into_iter().rev() {
let id = r.id;
let entry = batch_entries.remove(&id).unwrap();
self.entries.push_front((id, entry));
}
return None;
}
}
// Final batch size
let size = batch_requests.len() as u32;
next_batch_span.record("batch_size", size);
let batch = Batch {
id: self.next_batch_id,
requests: batch_requests,
size,
max_tokens: (prefill_tokens + decode_tokens),
};
// Increment batch id
self.next_batch_id += 1;
metrics::histogram!("tgi_batch_next_size", batch.size as f64);
Some((batch_entries, batch, next_batch_span))
}
}
type NextBatch = (IntMap<u64, Entry>, Batch, Span);
#[derive(Debug)]
enum QueueCommand {
Append(Box<Entry>, Span),
NextBatch {
min_size: Option<usize>,
max_size: Option<usize>,
prefill_token_budget: u32,
token_budget: u32,
response_sender: oneshot::Sender<Option<NextBatch>>,
span: Span,
},
}
#[cfg(test)]
mod tests {
use super::*;
use text_generation_client::{
GrammarType as ProtoGrammarType, NextTokenChooserParameters, StoppingCriteriaParameters,
};
use tracing::info_span;
fn default_entry() -> (
Entry,
mpsc::UnboundedReceiver<Result<InferStreamResponse, InferError>>,
) {
let (response_tx, receiver_tx) = mpsc::unbounded_channel();
let entry = Entry {
request: ValidGenerateRequest {
inputs: String::new(),
input_length: 0,
truncate: 0,
decoder_input_details: false,
parameters: NextTokenChooserParameters {
temperature: 0.0,
top_k: 0,
top_p: 0.0,
typical_p: 0.0,
do_sample: false,
seed: 0,
repetition_penalty: 0.0,
frequency_penalty: 0.0,
watermark: false,
grammar: String::new(),
grammar_type: ProtoGrammarType::None as i32,
},
stopping_parameters: StoppingCriteriaParameters {
ignore_eos_token: false,
max_new_tokens: 1,
stop_sequences: vec![],
},
top_n_tokens: 0,
},
response_tx,
span: info_span!("entry"),
temp_span: None,
queue_time: Instant::now(),
batch_time: None,
};
(entry, receiver_tx)
}
#[test]
fn test_append() {
let mut state = State::new(false, 1, None, 0);
let (entry, _guard) = default_entry();
assert_eq!(state.next_id, 0);
assert_eq!(state.entries.len(), 0);
state.append(entry);
assert_eq!(state.next_id, 1);
assert_eq!(state.entries.len(), 1);
let (id, _) = state.entries.remove(0).unwrap();
assert_eq!(id, 0);
}
#[test]
fn test_next_batch_empty() {
let mut state = State::new(false, 1, None, 0);
assert!(state.next_batch(None, None, 1, 1).is_none());
assert!(state.next_batch(Some(1), None, 1, 1).is_none());
}
#[test]
fn test_next_batch_min_size() {
let mut state = State::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
state.append(entry1);
state.append(entry2);
let (entries, batch, _) = state.next_batch(None, None, 2, 2).unwrap();
assert_eq!(entries.len(), 2);
assert!(entries.contains_key(&0));
assert!(entries.contains_key(&1));
assert!(entries.get(&0).unwrap().batch_time.is_some());
assert!(entries.get(&1).unwrap().batch_time.is_some());
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 2);
assert_eq!(state.next_id, 2);
assert_eq!(state.entries.len(), 0);
assert_eq!(state.next_batch_id, 1);
let (entry3, _guard3) = default_entry();
state.append(entry3);
assert!(state.next_batch(Some(2), None, 2, 2).is_none());
assert_eq!(state.next_id, 3);
assert_eq!(state.entries.len(), 1);
let (id, _) = state.entries.remove(0).unwrap();
assert_eq!(id, 2);
}
#[test]
fn test_next_batch_max_size() {
let mut state = State::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
state.append(entry1);
state.append(entry2);
let (entries, batch, _) = state.next_batch(None, Some(1), 2, 2).unwrap();
assert_eq!(entries.len(), 1);
assert!(entries.contains_key(&0));
assert!(entries.get(&0).unwrap().batch_time.is_some());
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 1);
assert_eq!(state.next_id, 2);
assert_eq!(state.entries.len(), 1);
assert_eq!(state.next_batch_id, 1);
}
#[test]
fn test_next_batch_token_budget() {
let mut state = State::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
state.append(entry1);
state.append(entry2);
let (entries, batch, _) = state.next_batch(None, None, 1, 1).unwrap();
assert_eq!(entries.len(), 1);
assert!(entries.contains_key(&0));
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 1);
assert_eq!(state.next_id, 2);
assert_eq!(state.entries.len(), 1);
assert_eq!(state.next_batch_id, 1);
let (entry3, _guard3) = default_entry();
state.append(entry3);
let (entries, batch, _) = state.next_batch(None, None, 3, 3).unwrap();
assert_eq!(entries.len(), 2);
assert!(entries.contains_key(&1));
assert!(entries.contains_key(&2));
assert_eq!(batch.id, 1);
assert_eq!(batch.size, 2);
assert_eq!(state.next_id, 3);
assert_eq!(state.entries.len(), 0);
assert_eq!(state.next_batch_id, 2);
}
#[tokio::test]
async fn test_queue_append() {
let queue = Queue::new(false, 1, None, 0);
let (entry, _guard) = default_entry();
queue.append(entry);
}
#[tokio::test]
async fn test_queue_next_batch_empty() {
let queue = Queue::new(false, 1, None, 0);
assert!(queue.next_batch(None, None, 1, 1).await.is_none());
assert!(queue.next_batch(Some(1), None, 1, 1).await.is_none());
}
#[tokio::test]
async fn test_queue_next_batch_min_size() {
let queue = Queue::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
queue.append(entry1);
queue.append(entry2);
let (entries, batch, _) = queue.next_batch(None, None, 2, 2).await.unwrap();
assert_eq!(entries.len(), 2);
assert!(entries.contains_key(&0));
assert!(entries.contains_key(&1));
assert!(entries.get(&0).unwrap().batch_time.is_some());
assert!(entries.get(&1).unwrap().batch_time.is_some());
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 2);
let (entry3, _guard3) = default_entry();
queue.append(entry3);
// Not enough requests pending
assert!(queue.next_batch(Some(2), None, 2, 2).await.is_none());
// Not enough token budget
assert!(queue.next_batch(Some(1), None, 0, 0).await.is_none());
// Ok
let (entries2, batch2, _) = queue.next_batch(Some(1), None, 2, 2).await.unwrap();
assert_eq!(entries2.len(), 1);
assert!(entries2.contains_key(&2));
assert!(entries2.get(&2).unwrap().batch_time.is_some());
assert_eq!(batch2.id, 1);
assert_eq!(batch2.size, 1);
}
#[tokio::test]
async fn test_queue_next_batch_max_size() {
let queue = Queue::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
queue.append(entry1);
queue.append(entry2);
let (entries, batch, _) = queue.next_batch(None, Some(1), 2, 2).await.unwrap();
assert_eq!(entries.len(), 1);
assert!(entries.contains_key(&0));
assert!(entries.get(&0).unwrap().batch_time.is_some());
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 1);
}
#[tokio::test]
async fn test_queue_next_batch_token_budget() {
let queue = Queue::new(false, 1, None, 0);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
queue.append(entry1);
queue.append(entry2);
let (entries, batch, _) = queue.next_batch(None, None, 1, 1).await.unwrap();
assert_eq!(entries.len(), 1);
assert!(entries.contains_key(&0));
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 1);
let (entry3, _guard3) = default_entry();
queue.append(entry3);
let (entries, batch, _) = queue.next_batch(None, None, 3, 3).await.unwrap();
assert_eq!(entries.len(), 2);
assert!(entries.contains_key(&1));
assert!(entries.contains_key(&2));
assert_eq!(batch.id, 1);
assert_eq!(batch.size, 2);
}
#[tokio::test]
async fn test_queue_next_batch_token_speculate() {
let queue = Queue::new(false, 1, None, 2);
let (entry1, _guard1) = default_entry();
let (entry2, _guard2) = default_entry();
queue.append(entry1);
queue.append(entry2);
// Budget of 1 is not enough
assert!(queue.next_batch(None, None, 1, 1).await.is_none());
let (entries, batch, _) = queue.next_batch(None, None, 6, 6).await.unwrap();
assert_eq!(entries.len(), 2);
assert!(entries.contains_key(&0));
assert!(entries.contains_key(&1));
assert_eq!(batch.id, 0);
assert_eq!(batch.size, 2);
}
#[tokio::test]
async fn test_queue_next_batch_dropped_receiver() {
let queue = Queue::new(false, 1, None, 0);
let (entry, _) = default_entry();
queue.append(entry);
assert!(queue.next_batch(None, None, 1, 1).await.is_none());
}
}
server/text_generation_server/models/cache_manager.py deleted 100644 → 0
View file @ 64def8e2
import math
import torch
from typing import Optional, List, Tuple
from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
BLOCK_SIZE: int = 16
# Will be set in warmup
CACHE_MANAGER: Optional["CacheManager"] = None
class CacheManager:
def __init__(
self,
num_blocks: int,
num_layers: int,
num_heads: int,
head_size: int,
repeat_slots: bool,
dtype: torch.dtype,
device: torch.device,
):
self.block_size = BLOCK_SIZE
self.num_blocks = num_blocks
self.repeat_slots = repeat_slots
element_size = torch.tensor([], dtype=dtype).element_size()
if IS_XPU_SYSTEM:
x = 1
else:
x = self.block_size // element_size
self.kv_cache = [
(
torch.empty(
(num_blocks, num_heads, head_size // x, self.block_size, x),
dtype=dtype,
device=device,
),
torch.empty(
(num_blocks, num_heads, head_size, self.block_size),
dtype=dtype,
device=device,
),
)
for _ in range(num_layers)
]
self.free_block_mask = torch.ones(num_blocks, dtype=torch.int32, device="cpu")
self.slots = torch.arange(
0, num_blocks * self.block_size, dtype=torch.int64
).view(num_blocks, self.block_size)
def allocate(
self,
needed_blocks_slots: List[Tuple[int, int]],
blocks: int,
max_blocks: int,
device: torch.device,
):
# Get free blocks indices by finding values in mask that are not set to 0
free_block_indices = self.free_block_mask.nonzero()
if blocks > len(free_block_indices):
raise RuntimeError(
f"Out of available cache blocks: asked {blocks}, only {len(free_block_indices)} free blocks"
)
# Slice by the number of required blocks
block_indices = free_block_indices[:blocks]
block_indices = block_indices.flatten()
# Padded block tables
block_tables_tensor = torch.zeros(
(len(needed_blocks_slots), max_blocks), dtype=torch.int32
)
# Allocate paged attention blocks
cumulative_blocks = 0
slots = []
block_tables = []
for i, (needed_blocks, needed_slots) in enumerate(needed_blocks_slots):
# Get allocated blocks for this sequence
allocated_blocks = block_indices[
cumulative_blocks : cumulative_blocks + needed_blocks
]
# Get slots for the allocated blocks
all_slots = self.slots[allocated_blocks].flatten()
# Repeat slots in the case of context sliding window
if needed_slots > len(all_slots) and self.repeat_slots:
repeats = math.ceil(needed_slots / len(all_slots))
all_slots = all_slots.repeat(repeats)
allocated_slots = all_slots[:needed_slots]
slots.append(allocated_slots)
block_tables.append(allocated_blocks.tolist())
block_tables_tensor[i, :needed_blocks] = allocated_blocks
cumulative_blocks += needed_blocks
block_tables = block_tables
block_tables_tensor = block_tables_tensor.to(device)
slots = torch.concat(slots).to(device)
# Allocate the required number of blocks by setting the mask to 0
self.free_block_mask[block_indices] = 0
return block_tables, block_tables_tensor, slots
def free(self, block_indices: Optional[List[int]]):
if block_indices is not None and block_indices:
# Reset mask
self.free_block_mask[block_indices] = 1
def set_cache_manager(
num_blocks: int,
num_layers: int,
num_heads: int,
head_size: int,
repeat_slots: bool,
dtype: torch.dtype,
device: torch.device,
) -> CacheManager:
global CACHE_MANAGER
if CACHE_MANAGER is not None:
del CACHE_MANAGER
torch.cuda.empty_cache()
CACHE_MANAGER = CacheManager(
num_blocks, num_layers, num_heads, head_size, repeat_slots, dtype, device
)
return CACHE_MANAGER
def get_cache_manager() -> CacheManager:
global CACHE_MANAGER
if CACHE_MANAGER is None:
raise RuntimeError("cache manager was not initialized")
return CACHE_MANAGER
server/text_generation_server/utils/awq/conversion_utils.py deleted 100644 → 0
View file @ 64def8e2
import torch
from typing import List
AWQ_PACK_ORDER = [0, 2, 4, 6, 1, 3, 5, 7]
REVERSE_AWQ_PACK_ORDER = [0, 4, 1, 5, 2, 6, 3, 7]
def pack(imatrix: torch.Tensor, direction: str = "column"):
"""
Packs a 4-bit integer matrix into a packed 32-bit integer matrix.
Args:
imatrix (torch.Tensor): matrix of integers
direction (str): direction of packing, either "column" or "row"
Returns:
qmatrix (torch.Tensor): packed matrix of integers
"""
shifts = torch.arange(0, 32, 4, dtype=torch.int32, device=imatrix.device)
imatrix = imatrix.to(torch.int8) & 0x0F # eventually correct overflow
if direction == "column":
imatrix = imatrix.view(-1, imatrix.shape[1] // (32 // 4), (32 // 4))
qmatrix = torch.bitwise_left_shift(imatrix, shifts[None, None, :]).sum(dim=-1)
elif direction == "row":
imatrix = imatrix.view(imatrix.shape[0] // (32 // 4), (32 // 4), -1)
qmatrix = torch.bitwise_left_shift(imatrix, shifts[None, :, None]).sum(dim=1)
qmatrix = qmatrix.to(torch.int32)
return qmatrix
def unpack(qmatrix: torch.Tensor, direction: str = "column"):
"""
Unpacks a 32-bit packed integer matrix into a 4-bit integer matrix.
Args:
qmatrix (torch.Tensor): matrix of packed integers
direction (str): direction of unpacking, either "column" or "row"
Returns:
imatrix (torch.Tensor): matrix of integers
"""
shifts = torch.arange(0, 32, 4, device=qmatrix.device)
if direction == "column":
imatrix = torch.bitwise_right_shift(
qmatrix[:, :, None], shifts[None, None, :]
).view(qmatrix.shape[0], -1)
elif direction == "row":
imatrix = torch.bitwise_right_shift(
qmatrix[:, None, :], shifts[None, :, None]
).view(-1, qmatrix.shape[-1])
imatrix = imatrix.to(torch.int8) & 0x0F # eventually correct overflow
return imatrix
def apply_order(
imatrix: torch.Tensor,
direction: str = "column",
order: List[int] = AWQ_PACK_ORDER,
):
"""
Applies the order to a 4-bit integer matrix.
Args:
imatrix (torch.Tensor): matrix of integers
direction (str): direction of applying order, either "column" or "row"
order (List[int]): order to apply, default is AWQ_PACK_ORDER
Returns:
imatrix (torch.Tensor): matrix of integers
"""
if direction == "column":
imatrix = imatrix.view(-1, (32 // 4))[:, order].view(imatrix.shape)
elif direction == "row":
imatrix = imatrix.view((32 // 4), -1)[order, :].view(imatrix.shape)
return imatrix
def fast_awq_to_gptq(qweight, qzeros):
# awq uses column packing for both weights and zeros
izeros = unpack(qzeros, direction="column")
iweights = unpack(qweight, direction="column")
# Reverse the order of the iweight and izeros tensors
izeros = apply_order(izeros, direction="column", order=REVERSE_AWQ_PACK_ORDER)
iweights = apply_order(iweights, direction="column", order=REVERSE_AWQ_PACK_ORDER)
# Subtract 1 from the izeros tensor (gptq adds 1 to the zeros)
izeros = izeros - 1
# exllama uses row packing for weights and column packing for zeros
qzeros = pack(izeros, direction="column")
qweight = pack(iweights, direction="row")
return qweight, qzeros
server/text_generation_server/utils/awq/quantize/qmodule.py deleted 100644 → 0
View file @ 64def8e2
# Copied logic from https://github.com/mit-han-lab/llm-awq/blob/f084f40bd996f3cf3a0633c1ad7d9d476c318aaa/awq/quantize/qmodule.py
import math
import torch
import torch.nn as nn
import awq_inference_engine # with CUDA kernels
# class ScaledActivation(nn.Module):
# def __init__(self, module, scales):
# super().__init__()
# self.act = module
# self.scales = nn.Parameter(scales.data)
#
# def forward(self, x):
# return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
class WQLinear(nn.Module):
def __init__(self, w_bit, group_size, qweight, qzeros, scales, bias):
super().__init__()
if w_bit not in [4]:
raise NotImplementedError("Only 4-bit are supported for now.")
self.in_features = qweight.shape[0]
self.out_features = qweight.shape[1] * 32 // w_bit
self.w_bit = w_bit
self.group_size = group_size if group_size != -1 else self.in_features
# quick sanity check (make sure aligment)
assert self.in_features % self.group_size == 0
assert self.out_features % (32 // self.w_bit) == 0
self.qweight = qweight
self.qzeros = qzeros
self.scales = scales
if bias:
self.bias = bias
else:
self.bias = None
@torch.no_grad()
def forward(self, x):
out_shape = x.shape[:-1] + (self.out_features,)
out = awq_inference_engine.gemm_forward_cuda(
x.reshape(-1, x.shape[-1]), self.qweight, self.scales, self.qzeros, 8
)
out = out + self.bias if self.bias is not None else out
return out.reshape(out_shape)
server/text_generation_server/utils/flash_attn.py deleted 100644 → 0
View file @ 64def8e2
import os
import torch
from loguru import logger
import math
from text_generation_server.utils.import_utils import (
IS_CUDA_SYSTEM,
IS_ROCM_SYSTEM,
IS_XPU_SYSTEM,
)
if os.getenv("USE_FLASH_ATTENTION", "").lower() == "false":
raise ImportError("`USE_FLASH_ATTENTION` is false.")
HAS_FLASH_ATTN = True
HAS_FLASH_ATTN_V2_CUDA = False
HAS_FLASH_ATTN_V2_ROCM = False
if IS_XPU_SYSTEM:
import intel_extension_for_pytorch as ipex
if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:
if not torch.cuda.is_available():
raise ImportError("CUDA is not available")
major, minor = torch.cuda.get_device_capability()
is_sm75 = major == 7 and minor == 5
is_sm8x = major == 8 and minor >= 0
is_sm90 = major == 9 and minor == 0
HAS_FLASH_ATTN = False
HAS_FLASH_ATTN_V2_CUDA = False
HAS_FLASH_ATTN_V2_ROCM = False
try:
try:
import flash_attn_2_cuda
except ImportError:
architecture_suffix = ""
if IS_CUDA_SYSTEM:
architecture_suffix = "-cuda"
elif IS_ROCM_SYSTEM:
architecture_suffix = "-rocm"
raise ImportError(
"Flash Attention V2 is not installed.\n"
"Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) "
f"or install flash attention v2 with `cd server && make install install-flash-attention-v2{architecture_suffix}`"
)
if not (is_sm8x or is_sm90) and IS_CUDA_SYSTEM:
raise ImportError(
f"GPU with CUDA capability {major} {minor} is not supported for "
"Flash Attention V2"
)
HAS_FLASH_ATTN_V2_CUDA = IS_CUDA_SYSTEM
HAS_FLASH_ATTN_V2_ROCM = IS_ROCM_SYSTEM
except ImportError as e:
try:
import flash_attn_cuda
except ImportError:
raise ImportError(
"Flash Attention is not installed.\n"
"Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) "
"or install flash attention with `cd server && make install install-flash-attention`"
) from e
if IS_CUDA_SYSTEM and not (is_sm75 or is_sm8x or is_sm90):
raise ImportError(
f"GPU with CUDA capability {major} {minor} is not supported"
) from e
elif IS_ROCM_SYSTEM:
for idx in range(torch.cuda.device_count()):
if "MI210" not in torch.cuda.get_device_name(
idx
) and "MI250" not in torch.cuda.get_device_name(idx):
raise ImportError(
f"AMD GPU {torch.cuda.get_device_name(idx)} does not support flash-attention"
)
logger.warning(f"Unable to use Flash Attention V2: {e}")
HAS_FLASH_ATTN = True
def attention(
q,
k,
v,
out,
cu_seqlens,
max_s,
softmax_scale,
window_size_left=-1,
):
if window_size_left <= 0 and window_size_left != -1:
raise ValueError("`window_size_left` must be > 0 or -1")
if IS_XPU_SYSTEM:
if window_size_left != -1:
raise ValueError(
f"XPU version of Flash Attention does not support window attention (window_size_left != -1, got window_size_left={window_size_left})."
)
return ipex.llm.functional.varlen_attention(
q,
k,
v,
out,
cu_seqlens,
cu_seqlens,
max_s,
max_s,
0.0,
softmax_scale,
False,
True,
False,
None,
)
if HAS_FLASH_ATTN_V2_CUDA:
return flash_attn_2_cuda.varlen_fwd(
q,
k,
v,
out,
cu_seqlens,
cu_seqlens,
None,
None,
None,
max_s,
max_s,
0.0,
softmax_scale,
False,
True,
window_size_left,
0,
False,
None,
)
elif HAS_FLASH_ATTN_V2_ROCM:
if window_size_left != -1:
raise ValueError(
f"RoCm version of Flash Attention v2 does not support window attention (window_size_left != -1, got window_size_left={window_size_left})."
)
# RoCm flash API does not take the window_size_left and window_size_right arguments.
return flash_attn_2_cuda.varlen_fwd(
q,
k,
v,
out,
cu_seqlens,
cu_seqlens,
max_s,
max_s,
0.0,
softmax_scale,
False,
True,
False,
None,
)
elif HAS_FLASH_ATTN:
if window_size_left != -1:
raise NotImplementedError(
"window_size_left is only available with flash attn v2"
)
# Flash attention v1 requires q, k and v to have the same number of heads
if k.shape[1] != q.shape[1]:
# MQA expand
if k.shape[1] == 1:
k = k.expand(-1, q.shape[1], -1)
# Grouped attention reshape
else:
original_shape = k.shape
k = (
k.unsqueeze(2)
.expand(-1, -1, q.shape[1] // k.shape[1], -1)
.reshape(original_shape[0], -1, original_shape[2])
)
if v.shape[1] != q.shape[1]:
# MQA expand
if v.shape[1] == 1:
v = v.expand(-1, q.shape[1], -1)
# Grouped attention reshape
else:
original_shape = v.shape
v = (
v.unsqueeze(2)
.expand(-1, -1, q.shape[1] // v.shape[1], -1)
.reshape(original_shape[0], -1, original_shape[2])
)
return flash_attn_cuda.fwd(
q,
k,
v,
out,
cu_seqlens,
cu_seqlens,
max_s,
max_s,
0.0,
softmax_scale,
False,
True,
False,
0,
None,
)
raise NotImplementedError("flash attention is not installed")
server/text_generation_server/utils/gptq/custom_autotune.py deleted 100644 → 0
View file @ 64def8e2
# https://github.com/fpgaminer/GPTQ-triton
"""
Mostly the same as the autotuner in Triton, but with a few changes like using 40 runs instead of 100.
"""
import builtins
import math
import time
from typing import Dict
import triton
class Autotuner(triton.KernelInterface):
def __init__(
self,
fn,
arg_names,
configs,
key,
reset_to_zero,
prune_configs_by: Dict = None,
nearest_power_of_two: bool = False,
):
"""
:param prune_configs_by: a dict of functions that are used to prune configs, fields:
'perf_model': performance model used to predicate running time with different configs, returns running time
'top_k': number of configs to bench
'prune_num_stages_by'(optional): a function used to prune num_stages. It take configs:List[Config] as its input, and returns pruned configs.
'nearest_power_of_two'(optional): whether to round key arguments to the nearest power of two when caching tuning results
"""
if not configs:
self.configs = [triton.Config({}, num_warps=4, num_stages=2)]
else:
self.configs = configs
self.key_idx = [arg_names.index(k) for k in key]
self.nearest_power_of_two = nearest_power_of_two
self.cache = {}
# hook to reset all required tensor to zeros before relaunching a kernel
self.hook = lambda args: 0
if reset_to_zero is not None:
self.reset_idx = [arg_names.index(k) for k in reset_to_zero]
def _hook(args):
for i in self.reset_idx:
args[i].zero_()
self.hook = _hook
self.arg_names = arg_names
# prune configs
if prune_configs_by:
perf_model, top_k = (
prune_configs_by["perf_model"],
prune_configs_by["top_k"],
)
if "early_config_prune" in prune_configs_by:
early_config_prune = prune_configs_by["early_config_prune"]
else:
perf_model, top_k, early_config_prune = None, None, None
self.perf_model, self.configs_top_k = perf_model, top_k
self.early_config_prune = early_config_prune
self.fn = fn
def _bench(self, *args, config, **meta):
# check for conflicts, i.e. meta-parameters both provided
# as kwargs and by the autotuner
conflicts = meta.keys() & config.kwargs.keys()
if conflicts:
raise ValueError(
f"Conflicting meta-parameters: {', '.join(conflicts)}."
" Make sure that you don't re-define auto-tuned symbols."
)
# augment meta-parameters with tunable ones
current = dict(meta, **config.kwargs)
def kernel_call():
if config.pre_hook:
config.pre_hook(self.nargs)
self.hook(args)
self.fn.run(
*args,
num_warps=config.num_warps,
num_stages=config.num_stages,
**current,
)
try:
# In testings using only 40 reps seems to be close enough and it appears to be what PyTorch uses
# PyTorch also sets fast_flush to True, but I didn't see any speedup so I'll leave the default
return triton.testing.do_bench(
kernel_call, quantiles=(0.5, 0.2, 0.8), rep=40
)
except triton.OutOfResources:
return (float("inf"), float("inf"), float("inf"))
def run(self, *args, **kwargs):
self.nargs = dict(zip(self.arg_names, args))
if len(self.configs) > 1:
key = tuple(args[i] for i in self.key_idx)
# This reduces the amount of autotuning by rounding the keys to the nearest power of two
# In my testing this gives decent results, and greatly reduces the amount of tuning required
if self.nearest_power_of_two:
key = tuple([2 ** int(math.log2(x) + 0.5) for x in key])
if key not in self.cache:
# prune configs
pruned_configs = self.prune_configs(kwargs)
bench_start = time.time()
timings = {
config: self._bench(*args, config=config, **kwargs)
for config in pruned_configs
}
bench_end = time.time()
self.bench_time = bench_end - bench_start
self.cache[key] = builtins.min(timings, key=timings.get)
self.hook(args)
self.configs_timings = timings
config = self.cache[key]
else:
config = self.configs[0]
self.best_config = config
if config.pre_hook is not None:
config.pre_hook(self.nargs)
return self.fn.run(
*args,
num_warps=config.num_warps,
num_stages=config.num_stages,
**kwargs,
**config.kwargs,
)
def prune_configs(self, kwargs):
pruned_configs = self.configs
if self.early_config_prune:
pruned_configs = self.early_config_prune(self.configs, self.nargs)
if self.perf_model:
top_k = self.configs_top_k
if isinstance(top_k, float) and top_k <= 1.0:
top_k = int(len(self.configs) * top_k)
if len(pruned_configs) > top_k:
est_timing = {
config: self.perf_model(
**self.nargs,
**kwargs,
**config.kwargs,
num_stages=config.num_stages,
num_warps=config.num_warps,
)
for config in pruned_configs
}
pruned_configs = sorted(est_timing.keys(), key=lambda x: est_timing[x])[
:top_k
]
return pruned_configs
def warmup(self, *args, **kwargs):
self.nargs = dict(zip(self.arg_names, args))
for config in self.prune_configs(kwargs):
self.fn.warmup(
*args,
num_warps=config.num_warps,
num_stages=config.num_stages,
**kwargs,
**config.kwargs,
)
self.nargs = None
def autotune(
configs, key, prune_configs_by=None, reset_to_zero=None, nearest_power_of_two=False
):
"""
Decorator for auto-tuning a :code:`triton.jit`'d function.
.. highlight:: python
.. code-block:: python
@triton.autotune(configs=[
triton.Config(meta={'BLOCK_SIZE': 128}, num_warps=4),
triton.Config(meta={'BLOCK_SIZE': 1024}, num_warps=8),
],
key=['x_size'] # the two above configs will be evaluated anytime
# the value of x_size changes
)
@triton.jit
def kernel(x_ptr, x_size, **META):
BLOCK_SIZE = META['BLOCK_SIZE']
:note: When all the configurations are evaluated, the kernel will run multiple time.
This means that whatever value the kernel updates will be updated multiple times.
To avoid this undesired behavior, you can use the `reset_to_zero` argument, which
reset the value of the provided tensor to `zero` before running any configuration.
:param configs: a list of :code:`triton.Config` objects
:type configs: list[triton.Config]
:param key: a list of argument names whose change in value will trigger the evaluation of all provided configs.
:type key: list[str]
:param prune_configs_by: a dict of functions that are used to prune configs, fields:
'perf_model': performance model used to predicate running time with different configs, returns running time
'top_k': number of configs to bench
'early_config_prune'(optional): a function used to do early prune (eg, num_stages). It take configs:List[Config] as its input, and returns pruned configs.
:param reset_to_zero: a list of argument names whose value will be reset to zero before evaluating any configs.
:type reset_to_zero: list[str]
"""
def decorator(fn):
return Autotuner(
fn,
fn.arg_names,
configs,
key,
reset_to_zero,
prune_configs_by,
nearest_power_of_two,
)
return decorator
def matmul248_kernel_config_pruner(configs, nargs):
"""
The main purpose of this function is to shrink BLOCK_SIZE_* when the corresponding dimension is smaller.
"""
m = max(2 ** int(math.ceil(math.log2(nargs["M"]))), 16)
n = max(2 ** int(math.ceil(math.log2(nargs["N"]))), 16)
k = max(2 ** int(math.ceil(math.log2(nargs["K"]))), 16)
used = set()
for config in configs:
block_size_m = min(m, config.kwargs["BLOCK_SIZE_M"])
block_size_n = min(n, config.kwargs["BLOCK_SIZE_N"])
block_size_k = min(k, config.kwargs["BLOCK_SIZE_K"])
group_size_m = config.kwargs["GROUP_SIZE_M"]
if (
block_size_m,
block_size_n,
block_size_k,
group_size_m,
config.num_stages,
config.num_warps,
) in used:
continue
used.add(
(
block_size_m,
block_size_n,
block_size_k,
group_size_m,
config.num_stages,
config.num_warps,
)
)
yield triton.Config(
{
"BLOCK_SIZE_M": block_size_m,
"BLOCK_SIZE_N": block_size_n,
"BLOCK_SIZE_K": block_size_k,
"GROUP_SIZE_M": group_size_m,
},
num_stages=config.num_stages,
num_warps=config.num_warps,
)
server/text_generation_server/utils/gptq/exllama.py deleted 100644 → 0
View file @ 64def8e2
import torch
from exllama_kernels import make_q4, q4_matmul, prepare_buffers, set_tuning_params
# Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
none_tensor = torch.empty((1, 1), device="meta")
def ext_make_q4(qweight, qzeros, scales, g_idx, device):
"""Construct Q4Matrix, return handle"""
return make_q4(
qweight, qzeros, scales, g_idx if g_idx is not None else none_tensor, device
)
def ext_q4_matmul(x, q4, q4_width):
"""Matrix multiplication, returns x @ q4"""
outshape = x.shape[:-1] + (q4_width,)
x = x.view(-1, x.shape[-1])
output = torch.empty((x.shape[0], q4_width), dtype=torch.float16, device=x.device)
q4_matmul(x, q4, output)
return output.view(outshape)
MAX_DQ = 1
MAX_INNER = 1
ACT_ORDER = False
DEVICE = None
TEMP_STATE = None
TEMP_DQ = None
def set_device(device):
global DEVICE
DEVICE = device
def create_exllama_buffers(max_total_tokens: int):
global MAX_DQ, MAX_INNER, ACT_ORDER, DEVICE, TEMP_STATE, TEMP_DQ
assert DEVICE is not None, "call set_device first"
if not ACT_ORDER:
max_total_tokens = 1
# This temp_state buffer is required to reorder X in the act-order case.
temp_state = torch.zeros(
(max_total_tokens, MAX_INNER), dtype=torch.float16, device=DEVICE
)
temp_dq = torch.zeros((1, MAX_DQ), dtype=torch.float16, device=DEVICE)
# This temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill.
prepare_buffers(DEVICE, temp_state, temp_dq)
matmul_recons_thd = 8
matmul_fused_remap = False
matmul_no_half2 = False
set_tuning_params(matmul_recons_thd, matmul_fused_remap, matmul_no_half2)
TEMP_STATE, TEMP_DQ = temp_state, temp_dq
class Ex4bitLinear(torch.nn.Module):
"""Linear layer implementation with per-group 4-bit quantization of the weights"""
def __init__(self, qweight, qzeros, scales, g_idx, bias, bits, groupsize):
super().__init__()
global MAX_DQ, MAX_INNER, ACT_ORDER, DEVICE
assert bits == 4
self.device = qweight.device
self.qweight = qweight
self.qzeros = qzeros
self.scales = scales
self.g_idx = g_idx.cpu() if g_idx is not None else None
self.bias = bias if bias is not None else None
if self.g_idx is not None and (
(self.g_idx == 0).all()
or torch.equal(
g_idx.cpu(),
torch.tensor(
[i // groupsize for i in range(g_idx.shape[0])], dtype=torch.int32
),
)
):
self.empty_g_idx = True
self.g_idx = None
assert self.device.type == "cuda"
assert self.device.index is not None
self.q4 = ext_make_q4(
self.qweight, self.qzeros, self.scales, self.g_idx, self.device.index
)
self.height = qweight.shape[0] * 8
self.width = qweight.shape[1]
# Infer groupsize from height of qzeros
self.groupsize = None
if self.qzeros.shape[0] > 1:
self.groupsize = (self.qweight.shape[0] * 8) // (self.qzeros.shape[0])
if self.groupsize is not None:
assert groupsize == self.groupsize
# Handle act-order matrix
if self.g_idx is not None:
if self.groupsize is None:
raise ValueError("Found group index but no groupsize. What do?")
self.act_order = True
else:
self.act_order = False
DEVICE = self.qweight.device
MAX_DQ = max(MAX_DQ, self.qweight.numel() * 8)
if self.act_order:
MAX_INNER = max(MAX_INNER, self.height, self.width)
ACT_ORDER = True
def forward(self, x):
out = ext_q4_matmul(x, self.q4, self.width)
if self.bias is not None:
out.add_(self.bias)
return out
server/text_generation_server/utils/gptq/exllamav2.py deleted 100644 → 0
View file @ 64def8e2
# Adapted from turboderp exllama: https://github.com/turboderp/exllamav2
import torch
import torch.nn as nn
from loguru import logger
try:
from exllamav2_kernels import make_q_matrix, gemm_half_q_half
except ImportError:
logger.error("exllamav2_kernels not installed.")
raise
# Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
none_tensor = torch.empty((1, 1), device="meta")
def ext_gemm_half_q_half(x, q_handle, q4_width, force_cuda):
"""Matrix multiplication, returns x @ q4"""
output_shape = x.shape[:-1] + (q4_width,)
x = x.view(-1, x.shape[-1])
output = torch.empty((x.shape[0], q4_width), dtype=torch.half, device=x.device)
gemm_half_q_half(x, q_handle, output, force_cuda)
return output.view(output_shape)
# Group map needed for irregular group sizes
def make_group_map(q_groups, num_qrows):
gr = q_groups.tolist()
group_map = []
num_groups = len(gr) // 2
for i in range(num_groups):
bits = gr[i * 2]
if i < num_groups - 1:
qrows = gr[i * 2 + 3] - gr[i * 2 + 1]
else:
qrows = num_qrows - gr[i * 2 + 1]
rows = qrows * 32 // bits
for j in range(rows):
group_map += [i]
group_map += [rows - j]
return torch.tensor(group_map, dtype=torch.short, device=q_groups.device)
# Create Q matrix
def ext_make_q_matrix(w: dict, temp_dq, key: str = None):
"""
Create Q matrix
"""
# EXL2
# won't work as the moment because the tensors are not the same.
if "q_weight" in w:
w["q_scale_max"] /= 256
w["q_perm"] = w["q_perm"].short()
w["q_invperm"] = w["q_invperm"].short()
if "q_group_map" not in w:
w["q_group_map"] = make_group_map(w["q_groups"], w["q_weight"].shape[0])
return make_q_matrix(
w["q_weight"],
w["q_perm"],
w["q_invperm"],
w["q_scale"],
w["q_scale_max"],
w["q_groups"],
w["q_group_map"],
none_tensor,
none_tensor,
none_tensor,
temp_dq,
)
# GPTQ
elif "qweight" in w:
if w["scales"].dtype == torch.float:
w["scales"] = w["scales"].half()
# GPTQ with g_idx (act_order)
if w.get("g_idx", None) is not None and not (w["g_idx"] == 0).all().item():
w["q_perm"] = torch.empty(
(w["qweight"].shape[0] * 8,),
dtype=torch.short,
device=w["qweight"].device,
)
w["q_invperm"] = torch.empty_like(w["q_perm"])
# make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx.
return make_q_matrix(
w["qweight"],
w["q_perm"],
w["q_invperm"],
none_tensor,
none_tensor,
none_tensor,
none_tensor,
w["qzeros"],
w["scales"],
w["g_idx"].cpu(),
temp_dq,
)
# GPTQ without g_idx
else:
return make_q_matrix(
w["qweight"],
none_tensor,
none_tensor,
none_tensor,
none_tensor,
none_tensor,
none_tensor,
w["qzeros"],
w["scales"],
none_tensor,
temp_dq,
)
DEVICE = None
FIXED_BYTES = 0
LAYERS = []
def set_device(device):
global DEVICE
DEVICE = device
def create_exllama_buffers(max_total_tokens: int):
global FIXED_BYTES, LAYERS, DEVICE
temp_dq = ExLlamaV2DeviceTensors(DEVICE, FIXED_BYTES)
for layer in LAYERS:
layer.post_init(temp_dq)
class QuantLinear(nn.Module):
QUANT_TYPE = "exllamav2"
"""Linear layer implementation with per-group 4-bit quantization of the weights"""
# def __init__(self, bits, group_size, infeatures, outfeatures, bias, trainable=False, **kwargs):
def __init__(self, qweight, qzeros, scales, g_idx, bias, bits, groupsize):
super().__init__()
if bits != 4:
raise ValueError(
f"Exllamav2 kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization."
)
self.q_handle = None
self.q_tensors = None
self.bits = bits
self.maxq = 2**self.bits - 1
self.infeatures = qweight.shape[0] // self.bits * 32
self.outfeatures = qweight.shape[1]
self.padding = -self.outfeatures % 32
self.outfeatures = self.outfeatures + self.padding
self.device = qweight.device
self.qweight = qweight
self.qzeros = qzeros
self.scales = scales
self.g_idx = g_idx
self.bias = bias if bias is not None else None
self.group_size = groupsize
global FIXED_BYTES, LAYERS
FIXED_BYTES = max(FIXED_BYTES, self.scratch_space_fixed())
LAYERS.append(self)
def post_init(self, temp_dq):
assert self.qweight.device.type == "cuda"
assert self.qweight.device.index is not None
self.q_tensors = {
"qweight": self.qweight,
"qzeros": self.qzeros,
"scales": self.scales,
"g_idx": self.g_idx,
}
temp_dq = temp_dq.get_scratch_slice(self.temp_dq_size())
# We NEED to keep a pointer on Python side, otherwise the garbage collector will mess with us,
# and `Memory access fault by GPU node-2` will EAT you.
self.temp_dq = temp_dq
self.q_handle = ext_make_q_matrix(self.q_tensors, temp_dq)
def forward(self, x, force_cuda=False):
output = ext_gemm_half_q_half(x, self.q_handle, self.outfeatures, force_cuda)
if self.bias is not None:
output.add_(self.bias)
return output
def temp_dq_size(self):
return self.infeatures * self.outfeatures * 2 + 128
def temp_fwd_size(self, max_input_len, max_batch_size):
return self.outfeatures * max_input_len * max_batch_size * 4 + 128
def scratch_space_fixed(self, max_input_len=4096, max_batch_size=16):
return self.temp_dq_size() + self.temp_fwd_size(max_input_len, max_batch_size)
class ExLlamaV2DeviceTensors:
device_idx: int
scratch_bytes: int
scratch_idx: int
scratch: torch.tensor = None
def __init__(self, device, scratch_bytes):
self.device = device
self.scratch_bytes = scratch_bytes
def prepare(self):
self.scratch = torch.empty(
(self.scratch_bytes // 2,), dtype=torch.half, device=self.device
)
def get_scratch_slice(self, size_bytes):
if self.scratch is None:
self.prepare()
size_bytes = ((size_bytes + 127) // 128) * 128
size_half = size_bytes // 2
scratch_slice = self.scratch.narrow(0, 0, size_half)
return scratch_slice
server/text_generation_server/utils/gptq/quant_linear.py deleted 100644 → 0
View file @ 64def8e2
import math
import numpy as np
import torch
import torch.nn as nn
from torch.cuda.amp import custom_bwd, custom_fwd
try:
import triton
import triton.language as tl
from . import custom_autotune
# code based https://github.com/fpgaminer/GPTQ-triton
@custom_autotune.autotune(
configs=[
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=4,
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=4,
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=4,
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=4,
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=4,
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
},
num_stages=2,
num_warps=8,
),
triton.Config(
{
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
},
num_stages=3,
num_warps=8,
),
triton.Config(
{
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
},
num_stages=2,
num_warps=4,
),
],
key=["M", "N", "K"],
nearest_power_of_two=True,
prune_configs_by={
"early_config_prune": custom_autotune.matmul248_kernel_config_pruner,
"perf_model": None,
"top_k": None,
},
)
@triton.jit
def matmul_248_kernel(
a_ptr,
b_ptr,
c_ptr,
scales_ptr,
zeros_ptr,
g_ptr,
M,
N,
K,
bits,
maxq,
stride_am,
stride_ak,
stride_bk,
stride_bn,
stride_cm,
stride_cn,
stride_scales,
stride_zeros,
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
):
"""
Compute the matrix multiplication C = A x B.
A is of shape (M, K) float16
B is of shape (K//8, N) int32
C is of shape (M, N) float16
scales is of shape (G, N) float16
zeros is of shape (G, N) float16
g_ptr is of shape (K) int32
"""
infearure_per_bits = 32 // bits
pid = tl.program_id(axis=0)
num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
num_pid_k = tl.cdiv(K, BLOCK_SIZE_K)
num_pid_in_group = GROUP_SIZE_M * num_pid_n
group_id = pid // num_pid_in_group
first_pid_m = group_id * GROUP_SIZE_M
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
pid_m = first_pid_m + (pid % group_size_m)
pid_n = (pid % num_pid_in_group) // group_size_m
offs_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
offs_k = tl.arange(0, BLOCK_SIZE_K)
a_ptrs = a_ptr + (
offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak
) # (BLOCK_SIZE_M, BLOCK_SIZE_K)
a_mask = offs_am[:, None] < M
# b_ptrs is set up such that it repeats elements along the K axis 8 times
b_ptrs = b_ptr + (
(offs_k[:, None] // infearure_per_bits) * stride_bk
+ offs_bn[None, :] * stride_bn
) # (BLOCK_SIZE_K, BLOCK_SIZE_N)
g_ptrs = g_ptr + offs_k
# shifter is used to extract the N bits of each element in the 32-bit word from B
scales_ptrs = scales_ptr + offs_bn[None, :]
zeros_ptrs = zeros_ptr + (offs_bn[None, :] // infearure_per_bits)
shifter = (offs_k % infearure_per_bits) * bits
zeros_shifter = (offs_bn % infearure_per_bits) * bits
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, num_pid_k):
g_idx = tl.load(g_ptrs)
# Fetch scales and zeros; these are per-outfeature and thus reused in the inner loop
scales = tl.load(
scales_ptrs + g_idx[:, None] * stride_scales
) # (BLOCK_SIZE_K, BLOCK_SIZE_N,)
zeros = tl.load(
zeros_ptrs + g_idx[:, None] * stride_zeros
) # (BLOCK_SIZE_K, BLOCK_SIZE_N,)
zeros = (zeros >> zeros_shifter[None, :]) & maxq
zeros = (zeros + 1) & maxq # eventually avoid overflow
a = tl.load(a_ptrs, mask=a_mask, other=0.0) # (BLOCK_SIZE_M, BLOCK_SIZE_K)
b = tl.load(b_ptrs) # (BLOCK_SIZE_K, BLOCK_SIZE_N), but repeated
# Now we need to unpack b (which is N-bit values) into 32-bit values
b = (b >> shifter[:, None]) & maxq # Extract the N-bit values
b = (b - zeros) * scales # Scale and shift
accumulator += tl.dot(a, b)
a_ptrs += BLOCK_SIZE_K
b_ptrs += (BLOCK_SIZE_K // infearure_per_bits) * stride_bk
g_ptrs += BLOCK_SIZE_K
c_ptrs = c_ptr + stride_cm * offs_am[:, None] + stride_cn * offs_bn[None, :]
c_mask = (offs_am[:, None] < M) & (offs_bn[None, :] < N)
tl.store(c_ptrs, accumulator, mask=c_mask)
except:
print("triton not installed.")
def matmul248(input, qweight, scales, qzeros, g_idx, bits, maxq):
with torch.cuda.device(input.device):
output = torch.empty(
(input.shape[0], qweight.shape[1]), device=input.device, dtype=torch.float16
)
grid = lambda META: (
triton.cdiv(input.shape[0], META["BLOCK_SIZE_M"])
* triton.cdiv(qweight.shape[1], META["BLOCK_SIZE_N"]),
)
matmul_248_kernel[grid](
input,
qweight,
output,
scales,
qzeros,
g_idx,
input.shape[0],
qweight.shape[1],
input.shape[1],
bits,
maxq,
input.stride(0),
input.stride(1),
qweight.stride(0),
qweight.stride(1),
output.stride(0),
output.stride(1),
scales.stride(0),
qzeros.stride(0),
)
return output
class QuantLinearFunction(torch.autograd.Function):
@staticmethod
@custom_fwd(cast_inputs=torch.float16)
def forward(ctx, input, qweight, scales, qzeros, g_idx, bits, maxq):
output = matmul248(input, qweight, scales, qzeros, g_idx, bits, maxq)
return output
class QuantLinear(nn.Module):
def __init__(self, qweight, qzeros, scales, g_idx, bias, bits, groupsize):
super().__init__()
self.register_buffer("qweight", qweight)
self.register_buffer("qzeros", qzeros)
self.register_buffer("scales", scales)
self.register_buffer("g_idx", g_idx)
if bias is not None:
self.register_buffer("bias", bias)
else:
self.bias = None
if bits not in [2, 4, 8]:
raise NotImplementedError("Only 2,4,8 bits are supported.")
self.bits = bits
self.maxq = 2**self.bits - 1
self.groupsize = groupsize
self.outfeatures = qweight.shape[1]
self.infeatures = qweight.shape[0] * 32 // bits
@classmethod
def new(cls, bits, groupsize, infeatures, outfeatures, bias):
if bits not in [2, 4, 8]:
raise NotImplementedError("Only 2,4,8 bits are supported.")
qweight = torch.zeros((infeatures // 32 * bits, outfeatures), dtype=torch.int32)
qzeros = torch.zeros(
(math.ceil(infeatures / groupsize), outfeatures // 32 * bits),
dtype=torch.int32,
)
scales = torch.zeros(
(math.ceil(infeatures / groupsize), outfeatures), dtype=torch.float16
)
g_idx = torch.tensor(
[i // groupsize for i in range(infeatures)], dtype=torch.int32
)
if bias:
bias = torch.zeros((outfeatures), dtype=torch.float16)
else:
bias = None
return cls(qweight, qzeros, scales, g_idx, bias, bits, groupsize)
def pack(self, linear, scales, zeros, g_idx=None):
self.g_idx = g_idx.clone() if g_idx is not None else self.g_idx
scales = scales.t().contiguous()
zeros = zeros.t().contiguous()
scale_zeros = zeros * scales
self.scales = scales.clone().half()
if linear.bias is not None:
self.bias = linear.bias.clone().half()
intweight = []
for idx in range(self.infeatures):
intweight.append(
torch.round(
(linear.weight.data[:, idx] + scale_zeros[self.g_idx[idx]])
/ self.scales[self.g_idx[idx]]
).to(torch.int)[:, None]
)
intweight = torch.cat(intweight, dim=1)
intweight = intweight.t().contiguous()
intweight = intweight.numpy().astype(np.uint32)
qweight = np.zeros(
(intweight.shape[0] // 32 * self.bits, intweight.shape[1]), dtype=np.uint32
)
i = 0
row = 0
while row < qweight.shape[0]:
if self.bits in [2, 4, 8]:
for j in range(i, i + (32 // self.bits)):
qweight[row] |= intweight[j] << (self.bits * (j - i))
i += 32 // self.bits
row += 1
else:
raise NotImplementedError("Only 2,4,8 bits are supported.")
qweight = qweight.astype(np.int32)
self.qweight = torch.from_numpy(qweight)
zeros -= 1
zeros = zeros.numpy().astype(np.uint32)
qzeros = np.zeros(
(zeros.shape[0], zeros.shape[1] // 32 * self.bits), dtype=np.uint32
)
i = 0
col = 0
while col < qzeros.shape[1]:
if self.bits in [2, 4, 8]:
for j in range(i, i + (32 // self.bits)):
qzeros[:, col] |= zeros[:, j] << (self.bits * (j - i))
i += 32 // self.bits
col += 1
else:
raise NotImplementedError("Only 2,4,8 bits are supported.")
qzeros = qzeros.astype(np.int32)
self.qzeros = torch.from_numpy(qzeros)
def forward(self, x):
out_shape = x.shape[:-1] + (self.outfeatures,)
out = QuantLinearFunction.apply(
x.reshape(-1, x.shape[-1]),
self.qweight,
self.scales,
self.qzeros,
self.g_idx,
self.bits,
self.maxq,
)
out = out + self.bias if self.bias is not None else out
return out.reshape(out_shape)
server/text_generation_server/utils/gptq/quantize.py deleted 100644 → 0
View file @ 64def8e2
import time
import torch.nn as nn
import math
import json
import os
import torch
import transformers
from texttable import Texttable
from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer
from huggingface_hub import HfApi
from accelerate import init_empty_weights
from text_generation_server.utils import initialize_torch_distributed, Weights
from text_generation_server.utils.hub import weight_files
from text_generation_server.utils.gptq.quant_linear import QuantLinear
from loguru import logger
from typing import Optional
DEV = torch.device("cuda:0")
class Quantizer(nn.Module):
def __init__(self, shape=1):
super(Quantizer, self).__init__()
self.register_buffer("maxq", torch.tensor(0))
self.register_buffer("scale", torch.zeros(shape))
self.register_buffer("zero", torch.zeros(shape))
def configure(
self,
bits,
perchannel=False,
sym=True,
mse=False,
norm=2.4,
grid=100,
maxshrink=0.8,
trits=False,
):
self.maxq = torch.tensor(2**bits - 1)
self.perchannel = perchannel
self.sym = sym
self.mse = mse
self.norm = norm
self.grid = grid
self.maxshrink = maxshrink
if trits:
self.maxq = torch.tensor(-1)
self.scale = torch.zeros_like(self.scale)
def _quantize(self, x, scale, zero, maxq):
if maxq < 0:
return (x > scale / 2).float() * scale + (x < zero / 2).float() * zero
q = torch.clamp(torch.round(x / scale) + zero, 0, maxq)
return scale * (q - zero)
def find_params(self, x, weight=False):
dev = x.device
self.maxq = self.maxq.to(dev)
shape = x.shape
if self.perchannel:
if weight:
x = x.flatten(1)
else:
if len(shape) == 4:
x = x.permute([1, 0, 2, 3])
x = x.flatten(1)
if len(shape) == 3:
x = x.reshape((-1, shape[-1])).t()
if len(shape) == 2:
x = x.t()
else:
x = x.flatten().unsqueeze(0)
tmp = torch.zeros(x.shape[0], device=dev)
xmin = torch.minimum(x.min(1)[0], tmp)
xmax = torch.maximum(x.max(1)[0], tmp)
if self.sym:
xmax = torch.maximum(torch.abs(xmin), xmax)
tmp = xmin < 0
if torch.any(tmp):
xmin[tmp] = -xmax[tmp]
tmp = (xmin == 0) & (xmax == 0)
xmin[tmp] = -1
xmax[tmp] = +1
if self.maxq < 0:
self.scale = xmax
self.zero = xmin
else:
self.scale = (xmax - xmin) / self.maxq
if self.sym:
self.zero = torch.full_like(self.scale, (self.maxq + 1) / 2)
else:
self.zero = torch.round(-xmin / self.scale)
if self.mse:
best = torch.full([x.shape[0]], float("inf"), device=dev)
for i in range(int(self.maxshrink * self.grid)):
p = 1 - i / self.grid
xmin1 = p * xmin
xmax1 = p * xmax
scale1 = (xmax1 - xmin1) / self.maxq
zero1 = torch.round(-xmin1 / scale1) if not self.sym else self.zero
q = self._quantize(
x, scale1.unsqueeze(1), zero1.unsqueeze(1), self.maxq
)
q -= x
q.abs_()
q.pow_(self.norm)
err = torch.sum(q, 1)
tmp = err < best
if torch.any(tmp):
best[tmp] = err[tmp]
self.scale[tmp] = scale1[tmp]
self.zero[tmp] = zero1[tmp]
if not self.perchannel:
if weight:
tmp = shape[0]
else:
tmp = shape[1] if len(shape) != 3 else shape[2]
self.scale = self.scale.repeat(tmp)
self.zero = self.zero.repeat(tmp)
if weight:
shape = [-1] + [1] * (len(shape) - 1)
self.scale = self.scale.reshape(shape)
self.zero = self.zero.reshape(shape)
return
if len(shape) == 4:
self.scale = self.scale.reshape((1, -1, 1, 1))
self.zero = self.zero.reshape((1, -1, 1, 1))
if len(shape) == 3:
self.scale = self.scale.reshape((1, 1, -1))
self.zero = self.zero.reshape((1, 1, -1))
if len(shape) == 2:
self.scale = self.scale.unsqueeze(0)
self.zero = self.zero.unsqueeze(0)
def quantize(self, x):
if self.ready():
return self._quantize(x, self.scale, self.zero, self.maxq)
return x
def enabled(self):
return self.maxq > 0
def ready(self):
return torch.all(self.scale != 0)
class GPTQ:
def __init__(self, layer, observe=False):
self.layer = layer
self.dev = self.layer.weight.device
W = layer.weight.data.clone()
if isinstance(self.layer, nn.Conv2d):
W = W.flatten(1)
if isinstance(self.layer, transformers.Conv1D):
W = W.t()
self.rows = W.shape[0]
self.columns = W.shape[1]
self.H = torch.zeros((self.columns, self.columns), device=self.dev)
self.nsamples = 0
self.quantizer = Quantizer()
self.observe = observe
def add_batch(self, inp, out):
# Hessian H = 2 X XT + λ I
if self.observe:
self.inp1 = inp
self.out1 = out
else:
self.inp1 = None
self.out1 = None
if len(inp.shape) == 2:
inp = inp.unsqueeze(0)
tmp = inp.shape[0]
if isinstance(self.layer, nn.Linear) or isinstance(
self.layer, transformers.Conv1D
):
if len(inp.shape) == 3:
inp = inp.reshape((-1, inp.shape[-1]))
inp = inp.t()
if isinstance(self.layer, nn.Conv2d):
unfold = nn.Unfold(
self.layer.kernel_size,
dilation=self.layer.dilation,
padding=self.layer.padding,
stride=self.layer.stride,
)
inp = unfold(inp)
inp = inp.permute([1, 0, 2])
inp = inp.flatten(1)
self.H *= self.nsamples / (self.nsamples + tmp)
self.nsamples += tmp
# inp = inp.float()
inp = math.sqrt(2 / self.nsamples) * inp.float()
# self.H += 2 / self.nsamples * inp.matmul(inp.t())
self.H += inp.matmul(inp.t())
def print_loss(self, name, q_weight, weight_error, timecost):
table = Texttable()
length = 28
name = (
(name + " " * (length - len(name)))
if len(name) <= length
else name[:length]
)
table.header(["name", "weight_error", "fp_inp_SNR", "q_inp_SNR", "time"])
# assign weight
self.layer.weight.data = q_weight.reshape(self.layer.weight.shape).to(
self.layer.weight.data.dtype
)
if self.inp1 is not None:
# quantize input to int8
quantizer = Quantizer()
quantizer.configure(8, perchannel=False, sym=True, mse=False)
quantizer.find_params(self.inp1)
q_in = quantizer.quantize(self.inp1).type(torch.float16)
q_out = self.layer(q_in)
# get kinds of SNR
q_SNR = torch_snr_error(q_out, self.out1).item()
fp_SNR = torch_snr_error(self.layer(self.inp1), self.out1).item()
else:
q_SNR = "-"
fp_SNR = "-"
table.add_row([name, weight_error, fp_SNR, q_SNR, timecost])
print(table.draw().split("\n")[-2])
def fasterquant(
self, blocksize=128, percdamp=0.01, groupsize=-1, act_order=False, name=""
):
self.layer.to(self.dev)
W = self.layer.weight.data.clone()
if isinstance(self.layer, nn.Conv2d):
W = W.flatten(1)
if isinstance(self.layer, transformers.Conv1D):
W = W.t()
W = W.float()
tick = time.time()
if not self.quantizer.ready():
self.quantizer.find_params(W, weight=True)
H = self.H
if not self.observe:
del self.H
dead = torch.diag(H) == 0
H[dead, dead] = 1
W[:, dead] = 0
if act_order:
perm = torch.argsort(torch.diag(H), descending=True)
W = W[:, perm]
H = H[perm][:, perm]
Losses = torch.zeros_like(W)
Q = torch.zeros_like(W)
damp = percdamp * torch.mean(torch.diag(H))
diag = torch.arange(self.columns, device=self.dev)
H[diag, diag] += damp
H = torch.linalg.cholesky(H)
H = torch.cholesky_inverse(H)
try:
H = torch.linalg.cholesky(H, upper=True)
except Exception:
# Addition because Falcon fails on h_to_4h
H = torch.linalg.cholesky(
H + 1e-5 * torch.eye(H.shape[0]).to(H.device), upper=True
)
Hinv = H
g_idx = []
scale = []
zero = []
now_idx = 1
for i1 in range(0, self.columns, blocksize):
i2 = min(i1 + blocksize, self.columns)
count = i2 - i1
W1 = W[:, i1:i2].clone()
Q1 = torch.zeros_like(W1)
Err1 = torch.zeros_like(W1)
Losses1 = torch.zeros_like(W1)
Hinv1 = Hinv[i1:i2, i1:i2]
for i in range(count):
w = W1[:, i]
d = Hinv1[i, i]
if groupsize != -1:
if (i1 + i) % groupsize == 0:
self.quantizer.find_params(
W[:, (i1 + i) : (i1 + i + groupsize)], weight=True
)
if ((i1 + i) // groupsize) - now_idx == -1:
scale.append(self.quantizer.scale)
zero.append(self.quantizer.zero)
now_idx += 1
q = self.quantizer.quantize(w.unsqueeze(1)).flatten()
Q1[:, i] = q
Losses1[:, i] = (w - q) ** 2 / d**2
err1 = (w - q) / d
W1[:, i:] -= err1.unsqueeze(1).matmul(Hinv1[i, i:].unsqueeze(0))
Err1[:, i] = err1
Q[:, i1:i2] = Q1
Losses[:, i1:i2] = Losses1 / 2
W[:, i2:] -= Err1.matmul(Hinv[i1:i2, i2:])
torch.cuda.synchronize()
error = torch.sum(Losses).item()
groupsize = groupsize if groupsize != -1 else self.columns
g_idx = [i // groupsize for i in range(self.columns)]
g_idx = torch.tensor(g_idx, dtype=torch.int32, device=Q.device)
if act_order:
invperm = torch.argsort(perm)
Q = Q[:, invperm]
g_idx = g_idx[invperm]
if isinstance(self.layer, transformers.Conv1D):
Q = Q.t()
self.print_loss(
name=name, q_weight=Q, weight_error=error, timecost=(time.time() - tick)
)
if scale == []:
scale.append(self.quantizer.scale)
zero.append(self.quantizer.zero)
scale = torch.cat(scale, dim=1)
zero = torch.cat(zero, dim=1)
return scale, zero, g_idx, error
def free(self):
self.inp1 = None
self.out1 = None
self.H = None
self.Losses = None
self.Trace = None
torch.cuda.empty_cache()
def get_wikitext2(nsamples, seed, seqlen, model_id, trust_remote_code):
from datasets import load_dataset
traindata = load_dataset("wikitext", "wikitext-2-raw-v1", split="train")
testdata = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=False, trust_remote_code=trust_remote_code
)
except:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=True, trust_remote_code=trust_remote_code
)
trainenc = tokenizer("\n\n".join(traindata["text"]), return_tensors="pt")
testenc = tokenizer("\n\n".join(testdata["text"]), return_tensors="pt")
import random
random.seed(seed)
trainloader = []
for _ in range(nsamples):
i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
inp = trainenc.input_ids[:, i:j]
tar = inp.clone()
tar[:, :-1] = -100
trainloader.append((inp, tar))
return trainloader, testenc
def get_ptb(nsamples, seed, seqlen, model_id, trust_remote_code):
from datasets import load_dataset
traindata = load_dataset("ptb_text_only", "penn_treebank", split="train")
valdata = load_dataset("ptb_text_only", "penn_treebank", split="validation")
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=False, trust_remote_code=trust_remote_code
)
except:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=True, trust_remote_code=trust_remote_code
)
trainenc = tokenizer("\n\n".join(traindata["sentence"]), return_tensors="pt")
testenc = tokenizer("\n\n".join(valdata["sentence"]), return_tensors="pt")
import random
random.seed(seed)
trainloader = []
for _ in range(nsamples):
i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
inp = trainenc.input_ids[:, i:j]
tar = inp.clone()
tar[:, :-1] = -100
trainloader.append((inp, tar))
return trainloader, testenc
def get_c4(nsamples, seed, seqlen, model_id, trust_remote_code):
from datasets import load_dataset
traindata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"train": "en/c4-train.00000-of-01024.json.gz"},
split="train",
use_auth_token=False,
)
valdata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"validation": "en/c4-validation.00000-of-00008.json.gz"},
split="validation",
use_auth_token=False,
)
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=False, trust_remote_code=trust_remote_code
)
except:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=True, trust_remote_code=trust_remote_code
)
import random
random.seed(seed)
trainloader = []
for _ in range(nsamples):
while True:
i = random.randint(0, len(traindata) - 1)
trainenc = tokenizer(traindata[i]["text"], return_tensors="pt")
if trainenc.input_ids.shape[1] >= seqlen:
break
i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
inp = trainenc.input_ids[:, i:j]
tar = inp.clone()
tar[:, :-1] = -100
trainloader.append((inp, tar))
import random
random.seed(0)
valenc = []
for _ in range(256):
while True:
i = random.randint(0, len(valdata) - 1)
tmp = tokenizer(valdata[i]["text"], return_tensors="pt")
if tmp.input_ids.shape[1] >= seqlen:
break
i = random.randint(0, tmp.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
valenc.append(tmp.input_ids[:, i:j])
valenc = torch.hstack(valenc)
class TokenizerWrapper:
def __init__(self, input_ids):
self.input_ids = input_ids
valenc = TokenizerWrapper(valenc)
return trainloader, valenc
def get_ptb_new(nsamples, seed, seqlen, model_id, trust_remote_code):
from datasets import load_dataset
traindata = load_dataset("ptb_text_only", "penn_treebank", split="train")
testdata = load_dataset("ptb_text_only", "penn_treebank", split="test")
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=False, trust_remote_code=trust_remote_code
)
except:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=True, trust_remote_code=trust_remote_code
)
trainenc = tokenizer(" ".join(traindata["sentence"]), return_tensors="pt")
testenc = tokenizer(" ".join(testdata["sentence"]), return_tensors="pt")
import random
random.seed(seed)
trainloader = []
for _ in range(nsamples):
i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
inp = trainenc.input_ids[:, i:j]
tar = inp.clone()
tar[:, :-1] = -100
trainloader.append((inp, tar))
return trainloader, testenc
def get_c4_new(nsamples, seed, seqlen, model_id, trust_remote_code):
from datasets import load_dataset
traindata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"train": "en/c4-train.00000-of-01024.json.gz"},
split="train",
)
valdata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"validation": "en/c4-validation.00000-of-00008.json.gz"},
split="validation",
)
try:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=False, trust_remote_code=trust_remote_code
)
except:
tokenizer = AutoTokenizer.from_pretrained(
model_id, use_fast=True, trust_remote_code=trust_remote_code
)
import random
random.seed(seed)
trainloader = []
for _ in range(nsamples):
while True:
i = random.randint(0, len(traindata) - 1)
trainenc = tokenizer(traindata[i]["text"], return_tensors="pt")
if trainenc.input_ids.shape[1] >= seqlen:
break
i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
j = i + seqlen
inp = trainenc.input_ids[:, i:j]
tar = inp.clone()
tar[:, :-1] = -100
trainloader.append((inp, tar))
valenc = tokenizer(" ".join(valdata[:1100]["text"]), return_tensors="pt")
valenc = valenc.input_ids[:, : (256 * seqlen)]
class TokenizerWrapper:
def __init__(self, input_ids):
self.input_ids = input_ids
valenc = TokenizerWrapper(valenc)
return trainloader, valenc
def get_loaders(
name, nsamples=128, seed=0, seqlen=2048, model_id="", trust_remote_code=False
):
if "wikitext2" in name:
return get_wikitext2(nsamples, seed, seqlen, model_id, trust_remote_code)
if "ptb" in name:
if "new" in name:
return get_ptb_new(nsamples, seed, seqlen, model_id, trust_remote_code)
return get_ptb(nsamples, seed, seqlen, model_id, trust_remote_code)
if "c4" in name:
if "new" in name:
return get_c4_new(nsamples, seed, seqlen, model_id, trust_remote_code)
return get_c4(nsamples, seed, seqlen, model_id, trust_remote_code)
def find_layers(module, layers=(nn.Conv2d, nn.Linear), name=""):
# Skip last lm_head linear
# Need isintance Falcon is inheriting Linear.
if isinstance(module, layers) and "lm_head" not in name:
return {name: module}
res = {}
for name1, child in module.named_children():
res.update(
find_layers(
child, layers=layers, name=name + "." + name1 if name != "" else name1
)
)
return res
@torch.no_grad()
def sequential(
model,
dataloader,
dev,
nsamples,
bits,
groupsize,
*,
hooks,
percdamp=0.01,
sym: bool = False,
act_order: bool = False,
):
print("Starting ...")
use_cache = model.config.use_cache
model.config.use_cache = False
try:
layers = model.model.layers
prefix = "model.layers"
except Exception:
layers = model.transformer.h
prefix = "transformer.h"
dtype = next(iter(model.parameters())).dtype
inps = torch.zeros(
(nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev
)
cache = {"i": 0}
extra = {}
class Catcher(nn.Module):
def __init__(self, module):
super().__init__()
self.module = module
def forward(self, inp, **kwargs):
inps[cache["i"]] = inp
cache["i"] += 1
extra.update(kwargs.copy())
raise ValueError
layers[0] = Catcher(layers[0])
for batch in dataloader:
try:
model(batch[0].cuda())
except ValueError:
pass
layers[0] = layers[0].module
# layers[0] = layers[0].cpu()
# model.model.embed_tokens = model.model.embed_tokens.cpu()
# model.model.norm = model.model.norm.cpu()
torch.cuda.empty_cache()
for hook in hooks:
hook.remove()
outs = torch.zeros_like(inps)
extra = {
k: v.to(dev) if isinstance(v, torch.Tensor) else v for k, v in extra.items()
}
print("Ready.")
quantizers = {}
for i in range(len(layers)):
print(f"Quantizing layer {i+1}/{len(layers)}..")
print("+------------------+--------------+------------+-----------+-------+")
print("| name | weight_error | fp_inp_SNR | q_inp_SNR | time |")
print("+==================+==============+============+===========+=======+")
layer = layers[i]
layer.load()
full = find_layers(layer)
sequential = [list(full.keys())]
for names in sequential:
subset = {n: full[n] for n in names}
gptq = {}
for name in subset:
gptq[name] = GPTQ(subset[name])
gptq[name].quantizer.configure(
bits, perchannel=True, sym=sym, mse=False
)
pass
def add_batch(name):
def tmp(_, inp, out):
gptq[name].add_batch(inp[0].data, out.data)
return tmp
handles = []
for name in subset:
handles.append(subset[name].register_forward_hook(add_batch(name)))
for j in range(nsamples):
outs[j] = layer(inps[j].unsqueeze(0), **extra)[0]
for h in handles:
h.remove()
for name in subset:
scale, zero, g_idx, error = gptq[name].fasterquant(
percdamp=percdamp,
groupsize=groupsize,
act_order=act_order,
name=name,
)
quantizers[f"{prefix}.{i}.{name}"] = (
gptq[name].quantizer.cpu(),
scale.cpu(),
zero.cpu(),
g_idx.cpu(),
bits,
groupsize,
)
gptq[name].free()
for j in range(nsamples):
outs[j] = layer(inps[j].unsqueeze(0), **extra)[0]
layer.unload()
del layer
del gptq
torch.cuda.empty_cache()
inps, outs = outs, inps
print("+------------------+--------------+------------+-----------+-------+")
print("\n")
model.config.use_cache = use_cache
return quantizers
def make_quant_linear(module, names, bits, groupsize, name=""):
if isinstance(module, QuantLinear):
return
for attr in dir(module):
tmp = getattr(module, attr)
name1 = name + "." + attr if name != "" else attr
if name1 in names:
delattr(module, attr)
setattr(
module,
attr,
QuantLinear.new(
bits,
groupsize,
tmp.in_features,
tmp.out_features,
tmp.bias is not None,
),
)
for name1, child in module.named_children():
make_quant_linear(
child, names, bits, groupsize, name + "." + name1 if name != "" else name1
)
# TODO: perform packing on GPU
def pack(model, quantizers, bits, groupsize):
layers = find_layers(model)
layers = {n: layers[n] for n in quantizers}
make_quant_linear(model, quantizers, bits, groupsize)
qlayers = find_layers(model, (QuantLinear,))
print("Packing ...")
for name in qlayers:
print(name)
quantizers[name], scale, zero, g_idx, _, _ = quantizers[name]
qlayers[name].pack(layers[name], scale, zero, g_idx)
print("Done.")
return model
def setdeepattr(module, full_name, tensor):
current = module
tokens = full_name.split(".")
for token in tokens[:-1]:
current = getattr(current, token)
setattr(current, tokens[-1], tensor)
def getdeepattr(module, full_name):
current = module
tokens = full_name.split(".")
for token in tokens:
current = getattr(current, token)
return current
def load_weights_pre_hook(module_name, weights, recursive=False):
def inner(module, args):
print(f"Pre hook {module_name}")
local_params = {}
for k, v in module.named_parameters():
if not recursive and k.count(".") != 1:
continue
local_params[k] = v
for k, v in module.named_buffers():
if not recursive and k.count(".") != 1:
continue
local_params[k] = v
for local_param in local_params:
current_tensor = getdeepattr(module, local_param)
if current_tensor.device == torch.device("meta"):
# print(f"Loading {local_param}")
if module_name:
tensor_name = f"{module_name}.{local_param}"
else:
tensor_name = local_param
tensor = weights.get_tensor(tensor_name)
setdeepattr(module, local_param, nn.Parameter(tensor))
else:
tensor = current_tensor.to(device=torch.device("cuda:0"))
if current_tensor.requires_grad:
tensor = nn.Parameter(tensor)
setdeepattr(module, local_param, tensor)
return inner
def load_weights_post_hook(module_name, weights, recursive=False):
def inner(module, args, output):
print(f"Post hook {module_name}")
local_params = {}
for k, v in module.named_parameters():
if not recursive and k.count(".") != 1:
continue
local_params[k] = v
for k, v in module.named_buffers():
if not recursive and k.count(".") != 1:
continue
local_params[k] = v
for local_param in local_params:
# print(f"Unloading {local_param}")
current_tensor = getdeepattr(module, local_param)
setdeepattr(
module,
local_param,
nn.Parameter(current_tensor.to(device=torch.device("cpu"))),
)
return output
return inner
def quantize(
model_id: str,
bits: int,
groupsize: int,
output_dir: str,
revision: str,
trust_remote_code: bool,
upload_to_model_id: Optional[str],
percdamp: float,
act_order: bool,
):
print("loading model")
config = AutoConfig.from_pretrained(
model_id,
trust_remote_code=trust_remote_code,
)
with init_empty_weights():
model = AutoModelForCausalLM.from_config(
config, torch_dtype=torch.float16, trust_remote_code=trust_remote_code
)
model = model.eval()
print("LOADED model")
files = weight_files(model_id, revision, extension=".safetensors")
process_group, _, _ = initialize_torch_distributed()
weights = Weights(
files,
device=torch.device("cuda:0"),
dtype=torch.float16,
process_group=process_group,
aliases={"embed_tokens.weight": ["lm_head.weight"]},
)
hooks = []
for name, module in model.named_modules():
def load(module, name):
def _load():
load_weights_pre_hook(name, weights, recursive=True)(module, None)
return _load
def unload(module, name):
def _unload():
load_weights_post_hook(name, weights, recursive=True)(
module, None, None
)
return _unload
module.load = load(module, name)
module.unload = unload(module, name)
hooks.append(
module.register_forward_pre_hook(load_weights_pre_hook(name, weights))
)
hooks.append(
module.register_forward_hook(load_weights_post_hook(name, weights))
)
model.seqlen = 2048
dataset = "wikitext2"
nsamples = 128
seed = None
dataloader, testloader = get_loaders(
dataset,
nsamples=nsamples,
seed=seed,
model_id=model_id,
seqlen=model.seqlen,
trust_remote_code=trust_remote_code,
)
tick = time.time()
quantizers = sequential(
model,
dataloader,
DEV,
nsamples,
bits,
groupsize,
percdamp=percdamp,
act_order=act_order,
hooks=hooks,
)
print(time.time() - tick)
pack(model, quantizers, bits, groupsize)
from safetensors.torch import save_file
from transformers.modeling_utils import shard_checkpoint
state_dict = model.state_dict()
state_dict = {k: v.cpu().contiguous() for k, v in state_dict.items()}
state_dict["gptq_bits"] = torch.LongTensor([bits])
state_dict["gptq_groupsize"] = torch.LongTensor([groupsize])
max_shard_size = "10GB"
shards, index = shard_checkpoint(
state_dict, max_shard_size=max_shard_size, weights_name="model.safetensors"
)
os.makedirs(output_dir, exist_ok=True)
for shard_file, shard in shards.items():
save_file(
shard,
os.path.join(output_dir, shard_file),
metadata={
"format": "pt",
"quantized": "gptq",
"origin": "text-generation-inference",
},
)
if index is None:
path_to_weights = os.path.join(output_dir, "model.safetensors")
logger.info(f"Model weights saved in {path_to_weights}")
else:
save_index_file = "model.safetensors.index.json"
save_index_file = os.path.join(output_dir, save_index_file)
with open(save_index_file, "w", encoding="utf-8") as f:
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
f.write(content)
logger.info(
f"The model is bigger than the maximum size per checkpoint ({max_shard_size}) and is going to be "
f"split in {len(shards)} checkpoint shards. You can find where each parameters has been saved in the "
f"index located at {save_index_file}."
)
config = AutoConfig.from_pretrained(model_id, trust_remote_code=trust_remote_code)
config.save_pretrained(output_dir)
logger.info("Saved config")
logger.info("Saving tokenizer")
tokenizer = AutoTokenizer.from_pretrained(
model_id, trust_remote_code=trust_remote_code
)
tokenizer.save_pretrained(output_dir)
logger.info("Saved tokenizer")
if upload_to_model_id:
api = HfApi()
api.upload_folder(
folder_path=output_dir, repo_id=upload_to_model_id, repo_type="model"
)
server/text_generation_server/utils/layers.py deleted 100644 → 0
View file @ 64def8e2
import os
import torch
import torch.distributed
from torch import nn
from torch.nn import functional as F
from typing import List, Tuple, Optional
from loguru import logger
from functools import lru_cache
from text_generation_server.utils.speculate import get_speculate
HAS_BITS_AND_BYTES = True
try:
import bitsandbytes as bnb
from bitsandbytes.nn import Int8Params, Params4bit
except ImportError:
HAS_BITS_AND_BYTES = False
from accelerate import init_empty_weights
from text_generation_server.utils.gptq.quant_linear import QuantLinear
from text_generation_server.utils.import_utils import (
IS_CUDA_SYSTEM,
IS_ROCM_SYSTEM,
IS_XPU_SYSTEM,
)
if IS_XPU_SYSTEM:
import intel_extension_for_pytorch as ipex
HAS_AWQ = True
try:
from text_generation_server.utils.awq.quantize.qmodule import WQLinear
except ImportError:
HAS_AWQ = False
try:
major, _minor = torch.cuda.get_device_capability()
except Exception:
major = 1
HAS_EXLLAMA = False
CAN_EXLLAMA = major >= 8 or IS_ROCM_SYSTEM
V2 = os.getenv("EXLLAMA_VERSION", "2") == "2"
if os.getenv("DISABLE_EXLLAMA") == "True":
HAS_EXLLAMA = False
elif CAN_EXLLAMA:
try:
if V2:
from text_generation_server.utils.gptq.exllamav2 import (
QuantLinear as ExllamaQuantLinear,
create_exllama_buffers,
set_device,
)
HAS_EXLLAMA = "2"
else:
from text_generation_server.utils.gptq.exllama import (
Ex4bitLinear as ExllamaQuantLinear,
create_exllama_buffers,
set_device,
)
HAS_EXLLAMA = "1"
except ImportError:
pass
HAS_EETQ = False
try:
from EETQ import quant_weights, w8_a16_gemm
HAS_EETQ = True
except ImportError:
pass
# Monkey patching
@classmethod
def load_layer_norm(cls, prefix, weights, eps):
weight = weights.get_tensor(f"{prefix}.weight")
bias = weights.get_tensor(f"{prefix}.bias")
with init_empty_weights():
ln = cls(weight.shape, eps=eps)
ln.weight = nn.Parameter(weight)
ln.bias = nn.Parameter(bias)
return ln
@classmethod
def load_layer_norm_no_bias(cls, prefix, weights, eps):
weight = weights.get_tensor(f"{prefix}.weight")
with init_empty_weights():
ln = cls(weight.shape, eps=eps)
ln.weight = nn.Parameter(weight)
ln.bias = None
return ln
@classmethod
def load_conv2d(cls, prefix, weights, in_channels, out_channels, kernel_size, stride):
weight = weights.get_tensor(f"{prefix}.weight")
bias = weights.get_tensor(f"{prefix}.bias")
with init_empty_weights():
conv2d = cls(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
)
conv2d.weight = nn.Parameter(weight)
conv2d.bias = nn.Parameter(bias)
return conv2d
@classmethod
def load_conv2d_no_bias(
cls, prefix, weights, in_channels, out_channels, kernel_size, stride
):
weight = weights.get_tensor(f"{prefix}.weight")
with init_empty_weights():
conv2d = cls(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
)
conv2d.weight = nn.Parameter(weight)
conv2d.bias = None
return conv2d
torch.nn.Conv2d.load = load_conv2d
torch.nn.Conv2d.load_no_bias = load_conv2d_no_bias
torch.nn.LayerNorm.load = load_layer_norm
torch.nn.LayerNorm.load_no_bias = load_layer_norm_no_bias
class FastLinear(nn.Module):
def __init__(
self,
weight,
bias,
) -> None:
super().__init__()
self.weight = nn.Parameter(weight)
if bias is not None:
self.bias = nn.Parameter(bias)
else:
self.bias = None
@classmethod
def load(cls, config, prefix: str, weights, bias: bool):
weight = weights.get_tensor(f"{prefix}.weight")
if bias:
bias = weights.get_tensor(f"{prefix}.bias")
else:
bias = None
return cls(weight, bias)
def forward(self, input: torch.Tensor) -> torch.Tensor:
return F.linear(input, self.weight, self.bias)
class EETQLinear(nn.Module):
def __init__(
self,
weight,
bias,
) -> None:
super().__init__()
device = weight.device
if weight.dtype != torch.float16:
weight = weight.to(dtype=torch.float16)
weight = torch.t(weight).contiguous().cpu()
weight, scale = quant_weights(weight, torch.int8, False)
self.weight = weight.cuda(device)
self.scale = scale.cuda(device)
self.bias = bias.cuda(device) if bias is not None else None
def forward(self, input: torch.Tensor) -> torch.Tensor:
output = w8_a16_gemm(input, self.weight, self.scale)
output = output + self.bias if self.bias is not None else output
return output
def fp8_quantize(weight, qdtype=torch.float8_e4m3fn):
device = weight.device
# weight, scale = quant_weights(weight, torch.int8, False)
finfo = torch.finfo(qdtype)
# Calculate the scale as dtype max divided by absmax
scale = finfo.max / weight.abs().max().clamp(min=1e-12)
# scale and clamp the tensor to bring it to
# the representative range of float8 data type
# (as default cast is unsaturated)
qweight = (weight * scale).clamp(min=finfo.min, max=finfo.max)
# Return both float8 data and the inverse scale (as float),
# as both required as inputs to torch._scaled_mm
qweight = qweight.to(qdtype)
scale = scale.float().reciprocal()
return qweight, scale
class Fp8Linear(nn.Module):
def __init__(
self,
weight,
bias,
) -> None:
super().__init__()
self.dtype = weight.dtype
self.qweight, self.scale = fp8_quantize(weight)
self.bias = bias if bias is not None else None
def forward(self, input: torch.Tensor) -> torch.Tensor:
qinput, scale = fp8_quantize(input)
output, _ = torch._scaled_mm(
qinput,
self.qweight.t(),
out_dtype=self.dtype,
scale_a=scale,
scale_b=self.scale,
bias=self.bias,
)
return output
class Linear8bitLt(nn.Module):
def __init__(
self,
weight,
bias,
has_fp16_weights=True,
memory_efficient_backward=False,
threshold=0.0,
index=None,
):
super().__init__()
assert (
not memory_efficient_backward
), "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
self.state = bnb.MatmulLtState()
self.index = index
# Necessary for stacked layers
self.state.threshold = threshold
self.state.has_fp16_weights = has_fp16_weights
self.state.memory_efficient_backward = memory_efficient_backward
if threshold > 0.0 and not has_fp16_weights:
self.state.use_pool = True
self.weight = Int8Params(
weight.data,
has_fp16_weights=has_fp16_weights,
requires_grad=has_fp16_weights,
)
self.weight.cuda(weight.device)
self.bias = bias
def init_8bit_state(self):
self.state.CB = self.weight.CB
self.state.SCB = self.weight.SCB
self.weight.CB = None
self.weight.SCB = None
def forward(self, x: torch.Tensor):
self.state.is_training = self.training
if self.weight.CB is not None:
self.init_8bit_state()
# weights are cast automatically as Int8Params, but the bias has to be cast manually
if self.bias is not None and self.bias.dtype != x.dtype:
self.bias.data = self.bias.data.to(x.dtype)
out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
if not self.state.has_fp16_weights:
if self.state.CB is not None and self.state.CxB is not None:
# we converted 8-bit row major to turing/ampere format in the first inference pass
# we no longer need the row-major weight
del self.state.CB
self.weight.data = self.state.CxB
return out
class Linear4bit(nn.Module):
def __init__(self, weight, bias, quant_type):
super().__init__()
self.weight = Params4bit(
weight.data,
requires_grad=False,
compress_statistics=True,
quant_type=quant_type,
)
self.compute_dtype = None
self.weight.cuda(weight.device)
self.bias = bias
def forward(self, x: torch.Tensor):
# weights are cast automatically as Int8Params, but the bias has to be cast manually
if self.bias is not None and self.bias.dtype != x.dtype:
self.bias.data = self.bias.data.to(x.dtype)
if getattr(self.weight, "quant_state", None) is None:
print(
"FP4 quantization state not initialized. Please call .cuda() or .to(device) on the LinearFP4 layer first."
)
inp_dtype = x.dtype
if self.compute_dtype is not None:
x = x.to(self.compute_dtype)
bias = None if self.bias is None else self.bias.to(self.compute_dtype)
out = bnb.matmul_4bit(
x, self.weight.t(), bias=bias, quant_state=self.weight.quant_state
)
out = out.to(inp_dtype)
return out
@lru_cache(1)
def warn_deprecate_bnb():
logger.warning(
"Bitsandbytes 8bit is deprecated, using `eetq` is a drop-in replacement, and has much better performnce"
)
def get_linear(weight, bias, quantize):
if quantize is None:
linear = FastLinear(weight, bias)
elif quantize == "eetq":
if HAS_EETQ:
linear = EETQLinear(weight, bias)
else:
raise ImportError(
"Please install EETQ from https://github.com/NetEase-FuXi/EETQ"
)
elif quantize == "fp8":
linear = Fp8Linear(weight, bias)
elif quantize == "bitsandbytes":
warn_deprecate_bnb()
linear = Linear8bitLt(
weight,
bias,
has_fp16_weights=False,
threshold=6.0,
)
if bias is not None:
linear.bias = nn.Parameter(bias)
elif quantize == "bitsandbytes-fp4":
linear = Linear4bit(
weight,
bias,
quant_type="fp4",
)
elif quantize == "bitsandbytes-nf4":
linear = Linear4bit(
weight,
bias,
quant_type="nf4",
)
elif quantize == "gptq":
try:
qweight, qzeros, scales, g_idx, bits, groupsize, use_exllama = weight
except Exception:
raise NotImplementedError(
f"The passed weight is not `gptq` compatible, loader needs to be updated."
)
if use_exllama:
linear = ExllamaQuantLinear(
qweight, qzeros, scales, g_idx, bias, bits, groupsize
)
else:
linear = QuantLinear(
qweight,
qzeros,
scales,
g_idx,
bias,
bits,
groupsize,
)
elif quantize == "awq":
try:
qweight, qzeros, scales, _, bits, groupsize, _ = weight
except Exception:
raise NotImplementedError(
f"The passed weight is not `awq` compatible, loader needs to be updated."
)
if IS_ROCM_SYSTEM:
raise NotImplementedError(
"AWQ GEMM kernel can't be used on ROCm systems, please use `--quantize gptq` instead "
"to use Exllama/GPTQ kernels for AWQ inference."
)
if not HAS_AWQ:
raise NotImplementedError(
"You do not seem to have awq installed, either install it (cd server && make install-awq), or try using GPTQ `---quantize gptq` a conversion AWQ->GPTQ will happen on the fly"
)
linear = WQLinear(
w_bit=bits,
group_size=groupsize,
qweight=qweight,
qzeros=qzeros,
scales=scales,
bias=bias is not None,
)
else:
raise NotImplementedError(f"Quantization `{quantize}` is not implemented yet.")
return linear
class SuperLayer(nn.Module):
def __init__(self, linear):
super().__init__()
self.linear = linear
def forward(self, x):
return self.linear.forward(x)
class ResBlock(torch.nn.Module):
def __init__(self, config, prefix, weights):
super().__init__()
self.linear = FastLinear.load(
config, prefix=f"{prefix}.linear", weights=weights, bias=True
)
self.act = torch.nn.SiLU()
def forward(self, x):
return x + self.act(self.linear(x))
class MedusaModel(torch.nn.Module):
def __init__(self, config, medusa_config, weights):
super().__init__()
self.heads = torch.nn.ModuleList(
[
MedusaHead(config, medusa_config, prefix=f"{i}", weights=weights)
for i in range(get_speculate())
]
)
def forward(self, x):
speculative_logits = torch.stack([head(x) for head in self.heads], dim=1)
return speculative_logits
class MedusaHead(torch.nn.Module):
def __init__(self, config, medusa_config, prefix, weights):
super().__init__()
self.blocks = torch.nn.ModuleList(
[
ResBlock(config, prefix=f"{prefix}.{i}", weights=weights)
for i in range(medusa_config["medusa_num_layers"])
]
)
n = len(self.blocks)
self.out = FastLinear.load(
config, prefix=f"{prefix}.{n}", weights=weights, bias=False
)
def forward(self, x):
for block in self.blocks:
x = block(x)
x = self.out(x)
return x
class MedusaHeadV1(nn.Module):
def __init__(self, lm_head, medusa):
super().__init__()
self.lm_head = lm_head
self.medusa = medusa
@staticmethod
def load(config, prefix: str, weights):
from pathlib import Path
from safetensors import safe_open
import json
use_medusa = config.use_medusa
medusa_config = str(Path(use_medusa) / "config.json")
filename = str(Path(use_medusa) / "medusa_lm_head.safetensors")
with open(medusa_config, "r") as f:
medusa_config = json.load(f)
routing = weights.routing
with safe_open(filename, framework="pytorch") as f:
for k in f.keys():
if k in routing and routing[k] != filename:
raise RuntimeError(
f"Key {k} was found in multiple files: {filename} and {routing[k]}"
)
routing[k] = filename
medusa = MedusaModel(config, medusa_config, weights)
lm_head = TensorParallelHead.load(config, prefix, weights)
return MedusaHeadV1(lm_head, medusa)
def forward(
self, input: torch.Tensor
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
logits = self.lm_head(input)
# If we have too many tokens, we skip speculative logits
if input.shape[0] > 128:
return logits, None
speculative_logits = self.medusa(input)
return logits, speculative_logits
class MedusaHeadV2(nn.Module):
def __init__(self, config, prefix, weights):
super().__init__()
from pathlib import Path
from safetensors import safe_open
import json
use_medusa = config.use_medusa
medusa_config = str(Path(use_medusa) / "config.json")
filename = str(Path(use_medusa) / "medusa_lm_head.safetensors")
with open(medusa_config, "r") as f:
medusa_config = json.load(f)
routing = weights.routing
with safe_open(filename, framework="pytorch") as f:
for k in f.keys():
if k in routing and routing[k] != filename:
raise RuntimeError(
f"Key {k} was found in multiple files: {filename} and {routing[k]}"
)
routing[k] = filename
self.n_medusa_heads = get_speculate()
assert medusa_config["medusa_num_layers"] == 1
self.linear = TensorParallelColumnLinear.load_multi(
config,
prefixes=[f"{i}.0.linear" for i in range(self.n_medusa_heads)],
dim=0,
weights=weights,
bias=True,
)
self.process_group = weights.process_group
self.world_size = self.process_group.size()
self.rank = self.process_group.rank()
self.act = torch.nn.SiLU()
self.lm_head = TensorParallelHead.load(config, prefix, weights)
def forward(self, x):
# If we have too many tokens, we skip speculative logits
if x.shape[0] > 128:
logits = self.lm_head(x)
return logits, None
size = x.shape[-1]
block_size = (size + self.world_size - 1) // self.world_size
start = self.rank * block_size
stop = (self.rank + 1) * block_size
x_block = x[:, start:stop]
# Compute all medusa heads at the same time, then reshape and move the n_medusa_heads dim to dim 1
medusa_res = self.act(self.linear(x)).reshape(
*x_block.shape[:-1], self.n_medusa_heads, x_block.shape[-1]
)
# Apply all residual medusa heads
output = x[:, start:stop].unsqueeze(-2) + medusa_res
# Gather medusa heads
world_output = [
torch.empty_like(output) for _ in range(self.process_group.size())
]
torch.distributed.all_gather(world_output, output, group=self.process_group)
world_output = torch.cat(world_output, dim=-1)
# Stack x and medusa residual x
stacked_x = torch.cat([x.unsqueeze(-2), world_output], dim=-2)
# Compute lm head on x + medusa residual x
logits = self.lm_head(stacked_x)
# Finally, split logits from speculative logits
logits, speculative_logits = torch.split(
logits, [1, self.n_medusa_heads], dim=-2
)
# Squeeze added dimension
logits = logits.squeeze(-2)
return logits, speculative_logits
class SpeculativeHead(nn.Module):
def __init__(self, lm_head, medusa):
super().__init__()
self.head = lm_head
self.medusa = medusa
@staticmethod
def load(config, prefix: str, weights):
use_medusa = config.use_medusa
if use_medusa:
lm_head = None
try:
medusa = MedusaHeadV1.load(config, prefix, weights)
except:
medusa = MedusaHeadV2(config, prefix, weights)
else:
lm_head = TensorParallelHead.load(config, prefix, weights)
medusa = None
return SpeculativeHead(lm_head, medusa)
def forward(
self, input: torch.Tensor
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
if self.medusa is not None:
return self.medusa(input)
assert self.head is not None
logits = self.head(input)
return logits, None
class TensorParallelHead(SuperLayer):
def __init__(self, linear, process_group, should_gather: bool):
super().__init__(linear)
self.process_group = process_group
self.should_gather = should_gather
@staticmethod
def load(config, prefix: str, weights):
if weights.process_group.size() > 1:
try:
weight = weights.get_sharded(f"{prefix}.weight", dim=0)
should_gather = True
except AssertionError:
# If the vocab size is not divisible by number of shards
# just load the entire thing.
weight = weights.get_tensor(f"{prefix}.weight")
should_gather = False
else:
weight = weights.get_tensor(f"{prefix}.weight")
should_gather = False
# GPTQ,AWQ,EETQ don't quantize heads (nor embeddings)
if config.quantize in ["gptq", "awq", "eetq"]:
quantize = None
else:
quantize = config.quantize
return TensorParallelHead(
get_linear(weight, bias=None, quantize=quantize),
process_group=weights.process_group,
should_gather=should_gather,
)
def forward(self, input: torch.Tensor) -> torch.Tensor:
if not self.should_gather:
return super().forward(input)
world_size = self.process_group.size()
if len(input.shape) == 2 and isinstance(self.linear, FastLinear):
out_dim = self.linear.weight.shape[0]
if input.shape[0] == 1:
world_out = input.new_empty(1, out_dim * world_size)
local_out = input.new_empty(1, out_dim)
gather_input = local_out
else:
world_out = input.new_empty(out_dim * world_size, input.shape[0])
gather_input = input.new_empty(out_dim, input.shape[0])
local_out = gather_input.T
torch.mm(input, self.linear.weight.T, out=local_out)
torch.distributed.all_gather_into_tensor(
world_out, gather_input, group=self.process_group
)
if input.shape[0] == 1:
return world_out
return world_out.T
output = super().forward(input)
world_output = [
torch.empty_like(output) for _ in range(self.process_group.size())
]
torch.distributed.all_gather(world_output, output, group=self.process_group)
world_output = torch.cat(world_output, dim=-1)
return world_output
class TensorParallelColumnLinear(SuperLayer):
@classmethod
def load_gate_up(cls, config, prefix: str, weights, bias: bool):
"""Specific method when the QKV was joined after the fact"""
weight = weights.get_weights_col_packed_gate_up(
prefix, quantize=config.quantize
)
if bias:
raise NotImplementedError("packed_gate_up only implemented without bias")
else:
bias = None
linear = get_linear(weight, bias, config.quantize)
return cls(linear)
@classmethod
def load_qkv(cls, config, prefix: str, weights, bias: bool):
"""Specific method when the QKV was joined after the fact"""
weight = weights.get_weights_col_packed_qkv(prefix, quantize=config.quantize)
if bias:
raise NotImplementedError("packed_qkv only implemented for baichuan")
else:
bias = None
linear = get_linear(weight, bias, config.quantize)
return cls(linear)
@classmethod
def load(cls, config, prefix: str, weights, bias: bool):
return cls.load_multi(config, [prefix], weights, bias, dim=0)
@classmethod
def load_multi(cls, config, prefixes: List[str], weights, bias: bool, dim: int):
weight = weights.get_multi_weights_col(
prefixes, quantize=config.quantize, dim=dim
)
if bias:
b = [weights.get_sharded(f"{p}.bias", dim=0) for p in prefixes]
bias = torch.cat(b, dim=dim)
else:
bias = None
linear = get_linear(weight, bias, config.quantize)
return cls(linear)
class TensorParallelRowLinear(SuperLayer):
def __init__(self, linear, process_group):
super().__init__(linear)
self.process_group = process_group
@classmethod
def load(cls, config, prefix: str, weights, bias: bool):
weight = weights.get_multi_weights_row(prefix, quantize=config.quantize)
if bias and weights.process_group.rank() == 0:
# Rank is only on the first rank process
bias = weights.get_tensor(f"{prefix}.bias")
else:
bias = None
return cls(
get_linear(weight, bias, config.quantize),
process_group=weights.process_group,
)
def forward(self, input: torch.Tensor, reduce: bool = True) -> torch.Tensor:
out = super().forward(input)
if self.process_group.size() > 1 and reduce:
torch.distributed.all_reduce(out, group=self.process_group)
return out
class TensorParallelEmbedding(nn.Module):
def __init__(self, prefix: str, weights, reduce=True):
super().__init__()
weight = weights.get_partial_sharded(f"{prefix}.weight", dim=0)
num_embeddings = weights.get_shape(f"{prefix}.weight")[0]
process_group = weights.process_group
world_size = process_group.size()
rank = process_group.rank()
block_size = (num_embeddings + world_size - 1) // world_size
self.min_id = rank * block_size
self.max_id = min(num_embeddings, (rank + 1) * block_size)
self.null_idx = weight.shape[
0
] # Usually block_size, might be less in non even vocab_size.
self.process_group = weights.process_group
self.reduce = reduce
"""Additional 0 entry used for masking"""
self.weight = nn.Parameter(F.pad(weight, (0, 0, 0, 1)))
def forward(self, input: torch.Tensor) -> torch.Tensor:
# default all out of bounds values to `self.null_idx` that will then be mapped to 0
# translate for [0, self.max_id - self.min_id[
input = torch.where(
(self.min_id > input) | (input >= self.max_id),
self.null_idx,
input - self.min_id,
)
out = torch.nn.functional.embedding(input, self.weight)
if self.reduce and self.process_group.size() > 1:
torch.distributed.all_reduce(out, group=self.process_group)
return out
try:
if IS_CUDA_SYSTEM:
import dropout_layer_norm
elif IS_ROCM_SYSTEM:
from vllm import layernorm_ops
else:
dropout_layer_norm = None
class FastLayerNorm(nn.LayerNorm):
def forward(self, hidden_states, residual=None):
if IS_XPU_SYSTEM:
res_out = hidden_states
out = ipex.llm.functional.add_layer_norm(
residual, hidden_states, self.weight, self.bias, self.eps, True
)
if residual is not None:
res_out = residual
return out, res_out
elif hidden_states.shape[-1] > 8192 or IS_ROCM_SYSTEM:
if residual is not None:
hidden_states += residual
residual = hidden_states
return super(FastLayerNorm, self).forward(hidden_states), residual
else:
(
normed_hidden_states,
residual,
*rest,
) = dropout_layer_norm.dropout_add_ln_fwd(
hidden_states,
residual,
self.weight,
self.bias,
None,
None,
None,
None,
0.0,
self.eps,
1.0,
0,
None,
False,
False,
)
if residual is None:
residual = hidden_states
return normed_hidden_states, residual
class FastRMSNorm(nn.Module):
def __init__(self, weight: torch.Tensor, eps: float):
super().__init__()
self.weight = nn.Parameter(weight)
self.variance_epsilon = eps
@classmethod
def load(cls, prefix, weights, eps=1e-6):
weight = weights.get_tensor(f"{prefix}.weight")
return cls(weight, eps)
def forward(self, hidden_states, residual=None):
if IS_XPU_SYSTEM:
residual_out = hidden_states
out = ipex.llm.functional.add_rms_norm(
residual,
hidden_states,
self.weight,
None,
self.variance_epsilon,
True,
)
if residual is not None:
residual_out = residual
return out, residual_out
elif hidden_states.shape[-1] > 8192:
if residual is not None:
hidden_states += residual
residual = hidden_states
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(
variance + self.variance_epsilon
)
# convert into half-precision if necessary
if self.weight.dtype in [torch.float16, torch.bfloat16]:
hidden_states = hidden_states.to(self.weight.dtype)
return self.weight * hidden_states, residual
elif IS_CUDA_SYSTEM:
# faster post attention rms norm
(
normed_hidden_states,
res,
*rest,
) = dropout_layer_norm.dropout_add_ln_fwd(
hidden_states,
residual,
self.weight,
None,
None,
None,
None,
None,
0.0,
self.variance_epsilon,
1.0,
0,
None,
False,
True, # Activate RMSNorm
)
if res is None:
res = hidden_states
return normed_hidden_states, res
elif IS_ROCM_SYSTEM:
# We use VLLM RMSNorm kernel that can be compiled for RoCm, instead of Flash Attention ones that can not.
if residual is not None:
hidden_states += residual
residual = hidden_states
out = torch.empty_like(hidden_states)
layernorm_ops.rms_norm(
out,
hidden_states,
self.weight.data,
self.variance_epsilon,
)
return out, residual
else:
raise ValueError(
"Your system seem to be not supported. Please check your install or open an issue at https://github.com/huggingface/text-generation-inference/issues with a clear reproduction."
)
except ImportError:
pass
try:
if IS_CUDA_SYSTEM:
from flash_attn.layers.rotary import RotaryEmbedding
import rotary_emb
elif IS_ROCM_SYSTEM:
from vllm import pos_encoding_ops
def _create_inv_freq(dim, base, device):
inv_freq = 1.0 / (
base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim)
)
return inv_freq
def _get_rope_config(config):
if os.getenv("ROPE_SCALING", None) is not None:
rope_scaling = {
"type": os.environ["ROPE_SCALING"],
"factor": float(os.environ["ROPE_FACTOR"]),
}
return rope_scaling
return getattr(config, "rope_scaling", None)
class PositionRotaryEmbedding(nn.Module):
def __init__(self, inv_freq, scaling_factor):
super().__init__()
self.inv_freq = inv_freq
self._seq_len_cached = 0
self._cos_cached = None
self._sin_cached = None
self._cos_k_cached = None
self._sin_k_cached = None
self.scaling_factor = scaling_factor
self.dynamic_args = None
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
):
# Such controlflows may add some overhead.
if IS_CUDA_SYSTEM:
rotary_dim = cos.shape[-1]
q1 = query[..., :rotary_dim]
q2 = query[..., rotary_dim : 2 * rotary_dim]
rotary_emb.apply_rotary(q1, q2, cos, sin, q1, q2, False)
k1 = key[..., :rotary_dim]
k2 = key[..., rotary_dim : 2 * rotary_dim]
rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
elif IS_ROCM_SYSTEM:
# NOTE: On RoCm systems, we use a ROPE implementatation adapted from VLLM which launches a single kernel for both query/key, contrary to flash-attn implementation used on NVIDIA systems.
# Compiling flash-attn rotary on RoCm, it appears hipcc is unable to unroll loops, resulting in an even slower inference compared to eager: https://github.com/pytorch/pytorch/issues/113773
head_size = query.shape[-1]
# Inplace operation, updating query and key.
pos_encoding_ops.rotary_embedding(query, key, head_size, cos, sin, True)
elif IS_XPU_SYSTEM:
ipex.llm.functional.rotary_embedding(
query, key, sin, cos, query.size(-1), True
)
else:
raise ValueError(
"Your system seem to be not supported. Please check your install or open an issue at https://github.com/huggingface/text-generation-inference/issues with a clear reproduction."
)
@classmethod
def static(cls, config, dim, base, device):
inv_freq = _create_inv_freq(dim, base, device)
scaling_factor = None
rope_scaling = _get_rope_config(config)
if rope_scaling is not None:
scaling_factor = rope_scaling["factor"]
if rope_scaling["type"] == "linear":
pass
elif rope_scaling["type"] == "dynamic":
return DynamicPositionRotaryEmbedding(
dim=dim,
max_position_embeddings=config.max_position_embeddings,
base=base,
device=inv_freq.device,
scaling_factor=scaling_factor,
)
elif rope_scaling["type"] == "yarn":
return YarnPositionRotaryEmbedding(
dim=2 * inv_freq.shape[0],
max_position_embeddings=rope_scaling[
"original_max_position_embeddings"
],
base=10000.0,
device=inv_freq.device,
scaling_factor=scaling_factor,
extrapolation_factor=1,
attn_factor=1,
beta_fast=32,
beta_slow=1,
)
else:
raise NotImplementedError(
f"rope scaling type {rope_scaling['type']} is not implemented or invalid"
)
return cls(inv_freq, scaling_factor)
@classmethod
def load(cls, config, prefix, weights):
# XXX: Always load this in float32 !
dtype = weights.dtype
weights.dtype = torch.float32
inv_freq = weights.get_tensor(f"{prefix}.inv_freq")
weights.dtype = dtype
scaling_factor = None
rope_scaling = _get_rope_config(config)
if rope_scaling is not None:
scaling_factor = rope_scaling["factor"]
if rope_scaling["type"] == "linear":
pass
elif rope_scaling["type"] == "dynamic":
return DynamicPositionRotaryEmbedding(
dim=2 * inv_freq.shape[0],
max_position_embeddings=config.max_position_embeddings,
base=10000.0,
device=inv_freq.device,
scaling_factor=scaling_factor,
)
elif rope_scaling["type"] == "yarn":
return YarnPositionRotaryEmbedding(
dim=2 * inv_freq.shape[0],
max_position_embeddings=rope_scaling[
"original_max_position_embeddings"
],
base=10000.0,
device=inv_freq.device,
scaling_factor=scaling_factor,
extrapolation_factor=1,
attn_factor=1,
beta_fast=32,
beta_slow=1,
)
else:
raise NotImplementedError(
f"rope scaling type {rope_scaling['type']} is not implemented or invalid"
)
return cls(inv_freq, scaling_factor)
def _update_cos_sin_cache(self, dtype, device, seqlen):
# Reset the tables if the sequence length has changed,
# or if we're on a new device (possibly due to tracing for instance)
if (
seqlen > self._seq_len_cached
or self._cos_cached.device != device
or self._cos_cached.dtype != dtype
):
self._seq_len_cached = seqlen
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
if self.scaling_factor is not None:
t /= self.scaling_factor
# Don't do einsum, it converts fp32 to fp16
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)
freqs = torch.outer(t, self.inv_freq.to(device=t.device))
self._cos_cached = torch.cos(freqs).to(dtype)
self._sin_cached = torch.sin(freqs).to(dtype)
def get_cos_sin(
self, position_ids: torch.Tensor, max_s: int, dtype: torch.dtype
):
"""
Return cos and sin for the asked position ids
"""
if IS_ROCM_SYSTEM:
# For RoCm, we always use float cos/sin to avoid a cast.
# For NVIDIA, for some reason, the flash-attn rotary kernel requires cos/sin and query/key to be of same dtype: https://github.com/Dao-AILab/flash-attention/blob/017716451d446e464dde9aca3a3c1ed2209caaa9/csrc/rotary/rotary.cpp#L26
# But later on goes and cast cos/sin to float anyway: https://github.com/Dao-AILab/flash-attention/blob/017716451d446e464dde9aca3a3c1ed2209caaa9/csrc/rotary/rotary_cuda.cu#L29, which looks suboptimal.
dtype = torch.float32
self._update_cos_sin_cache(dtype, position_ids.device, max_s)
cos = torch.index_select(self._cos_cached, 0, position_ids)
sin = torch.index_select(self._sin_cached, 0, position_ids)
# Note: this unsqueeze is not necessary on RoCm + VLLM ROPE implementation, but we leave it as is to avoid yet an other controlflow.
return cos.unsqueeze(1), sin.unsqueeze(1)
class DynamicPositionRotaryEmbedding(PositionRotaryEmbedding):
def __init__(self, dim, max_position_embeddings, base, device, scaling_factor):
inv_freq = _create_inv_freq(dim, base, device)
super().__init__(inv_freq, scaling_factor)
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
def _update_cos_sin_cache(self, dtype, device, seqlen):
# Reset the tables if the sequence length has changed,
# or if we're on a new device (possibly due to tracing for instance)
if (
seqlen > self._seq_len_cached
or self._cos_cached.device != device
or self._cos_cached.dtype != dtype
):
if seqlen > self.max_position_embeddings:
newbase = self.base * (
(self.scaling_factor * seqlen / self.max_position_embeddings)
- (self.scaling_factor - 1)
) ** (self.dim / (self.dim - 2))
self.inv_freq = _create_inv_freq(
self.dim, newbase, self.inv_freq.device
)
self._seq_len_cached = seqlen
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
# Don't do einsum, it converts fp32 to fp16
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)
freqs = torch.outer(t, self.inv_freq.to(device=t.device))
self._cos_cached = torch.cos(freqs).to(dtype)
self._sin_cached = torch.sin(freqs).to(dtype)
# Inverse dim formula to find dim based on number of rotations
import math
def find_correction_dim(
num_rotations, dim, base=10000, max_position_embeddings=2048
):
return (
dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))
) / (2 * math.log(base))
# Find dim range bounds based on rotations
def find_correction_range(
low_rot, high_rot, dim, base=10000, max_position_embeddings=2048
):
low = math.floor(
find_correction_dim(low_rot, dim, base, max_position_embeddings)
)
high = math.ceil(
find_correction_dim(high_rot, dim, base, max_position_embeddings)
)
return max(low, 0), min(high, dim - 1) # Clamp values just in case
def linear_ramp_mask(min, max, dim):
if min == max:
max += 0.001 # Prevent singularity
linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
ramp_func = torch.clamp(linear_func, 0, 1)
return ramp_func
def get_mscale(scale=1):
if scale <= 1:
return 1.0
return 0.1 * math.log(scale) + 1.0
class YarnPositionRotaryEmbedding(PositionRotaryEmbedding):
def __init__(
self,
dim,
max_position_embeddings,
base,
device,
scaling_factor,
*,
extrapolation_factor,
attn_factor,
beta_fast,
beta_slow,
):
inv_freq = _create_inv_freq(dim, base, device)
super().__init__(inv_freq, scaling_factor)
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
self.extrapolation_factor = extrapolation_factor
self.attn_factor = attn_factor
self.beta_fast = beta_fast
self.beta_slow = beta_slow
self.mscale = float(
get_mscale(self.scaling_factor) * self.attn_factor
) # Get n-d magnitude scaling corrected for interpolation
def _update_cos_sin_cache(self, dtype, device, seqlen):
# Reset the tables if the sequence length has changed,
# or if we're on a new device (possibly due to tracing for instance)
if (
seqlen > self._seq_len_cached
or self._cos_cached.device != device
or self._cos_cached.dtype != dtype
):
if seqlen > self.max_position_embeddings:
inv_freq_extrapolation = _create_inv_freq(
self.dim, self.base, self.inv_freq.device
)
freqs = 1.0 / inv_freq_extrapolation
inv_freq_interpolation = 1.0 / (self.scaling_factor * freqs)
low, high = find_correction_range(
self.beta_fast,
self.beta_slow,
self.dim,
self.base,
self.max_position_embeddings,
)
inv_freq_mask = (
1
- linear_ramp_mask(low, high, self.dim // 2).float().to(device)
) * self.extrapolation_factor # Get n-d rotational scaling corrected for extrapolation
inv_freq = (
inv_freq_interpolation * (1 - inv_freq_mask)
+ inv_freq_extrapolation * inv_freq_mask
)
self.inv_freq = inv_freq
self.mscale = float(
get_mscale(self.scaling_factor) * self.attn_factor
) # Get n-d magnitude scaling corrected for interpolation
self._seq_len_cached = seqlen
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
# Don't do einsum, it converts fp32 to fp16
# freqs = torch.einsum("i,j->ij", t, self.inv_freq)
freqs = torch.outer(t, self.inv_freq.to(device=t.device))
self._cos_cached = (torch.cos(freqs) * self.mscale).to(dtype)
self._sin_cached = (torch.sin(freqs) * self.mscale).to(dtype)
except ImportError:
pass
server/text_generation_server/utils/paged_attention.py deleted 100644 → 0
View file @ 64def8e2
import torch
from text_generation_server.utils.import_utils import (
IS_CUDA_SYSTEM,
IS_ROCM_SYSTEM,
IS_XPU_SYSTEM,
)
_PARTITION_SIZE = 512
if IS_XPU_SYSTEM:
import intel_extension_for_pytorch as ipex
def reshape_and_cache(
key: torch.Tensor,
value: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
slots: torch.Tensor,
):
if IS_CUDA_SYSTEM:
from vllm._C import cache_ops
cache_ops.reshape_and_cache(
key, value, key_cache, value_cache, slots, "auto", 1.0
)
elif IS_ROCM_SYSTEM:
from vllm import cache_ops
# cache_ops.reshape_and_cache(key, value, key_cache, value_cache, slots)
cache_ops.reshape_and_cache(key, value, key_cache, value_cache, slots.int())
elif IS_XPU_SYSTEM:
ipex.llm.modules.PagedAttention.reshape_and_cache(
key, value, key_cache, value_cache, slots
)
else:
raise ValueError("vllm is not supported on your system")
def attention(
out: torch.Tensor,
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
kv_head_mapping: torch.Tensor,
softmax_scale: float,
block_tables: torch.Tensor,
input_lengths: torch.Tensor,
max_s: int,
):
# Adapted from: https://github.com/vllm-project/vllm/blob/f8a1e39fae05ca610be8d5a78be9d40f5274e5fc/vllm/model_executor/layers/attention.py
# Copyright 2023 The vLLM team. All rights
# reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# value_cache => [num_blocks, num_heads, head_size, block_size]
block_size = value_cache.shape[3]
num_seqs, num_heads, head_size = query.shape
max_num_partitions = (max_s + _PARTITION_SIZE - 1) // _PARTITION_SIZE
if IS_XPU_SYSTEM:
query = query.contiguous()
return ipex.llm.modules.PagedAttention.single_query_cached_kv_attention(
out,
query,
key_cache,
value_cache,
kv_head_mapping,
softmax_scale,
block_tables,
input_lengths,
block_size,
max_s,
None,
)
# NOTE(woosuk): We use a simple heuristic to decide whether to use
# PagedAttention V1 or V2. If the number of partitions is 1, we use
# V1 to avoid the overhead of reduction. Also, if the number of
# sequences or heads is large, we use V1 since there is enough work
# to parallelize.
use_v1 = max_s <= 8192 and (max_num_partitions == 1 or num_seqs * num_heads > 512)
if use_v1:
if IS_CUDA_SYSTEM:
from vllm._C import ops
ops.paged_attention_v1(
out,
query,
key_cache,
value_cache,
kv_head_mapping,
softmax_scale,
block_tables,
input_lengths,
block_size,
max_s,
None,
"auto",
1.0,
)
elif IS_ROCM_SYSTEM:
from vllm import attention_ops
attention_ops.paged_attention_v1(
out,
query,
key_cache,
value_cache,
kv_head_mapping,
softmax_scale,
block_tables,
input_lengths,
block_size,
max_s,
None,
)
else:
raise ValueError("vllm is not supported on your system")
else:
# Run PagedAttention V2.
assert _PARTITION_SIZE % block_size == 0
tmp_output = torch.empty(
size=(num_seqs, num_heads, max_num_partitions, head_size),
dtype=out.dtype,
device=out.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=out.device,
)
max_logits = torch.empty_like(exp_sums)
if IS_CUDA_SYSTEM:
from vllm._C import ops
ops.paged_attention_v2(
out,
exp_sums,
max_logits,
tmp_output,
query,
key_cache,
value_cache,
kv_head_mapping,
softmax_scale,
block_tables,
input_lengths,
block_size,
max_s,
None,
"auto",
1.0,
)
elif IS_ROCM_SYSTEM:
from vllm import attention_ops
attention_ops.paged_attention_v2(
out,
exp_sums,
max_logits,
tmp_output,
query,
key_cache,
value_cache,
kv_head_mapping,
softmax_scale,
block_tables,
input_lengths,
block_size,
max_s,
None,
)
else:
raise ValueError("vllm is not supported on your system")
server/vllm/.pylintrc deleted 100644 → 0
View file @ 64def8e2
# This Pylint rcfile contains a best-effort configuration to uphold the
# best-practices and style described in the Google Python style guide:
# https://google.github.io/styleguide/pyguide.html
#
# Its canonical open-source location is:
# https://google.github.io/styleguide/pylintrc
[MASTER]
# Files or directories to be skipped. They should be base names, not paths.
ignore=docs
# Files or directories matching the regex patterns are skipped. The regex
# matches against base names, not paths.
ignore-patterns=
# Pickle collected data for later comparisons.
persistent=no
# List of plugins (as comma separated values of python modules names) to load,
# usually to register additional checkers.
load-plugins=
# Use multiple processes to speed up Pylint.
jobs=4
# Allow loading of arbitrary C extensions. Extensions are imported into the
# active Python interpreter and may run arbitrary code.
unsafe-load-any-extension=no
[MESSAGES CONTROL]
# Only show warnings with the listed confidence levels. Leave empty to show
# all. Valid levels: HIGH, INFERENCE, INFERENCE_FAILURE, UNDEFINED
confidence=
# Enable the message, report, category or checker with the given id(s). You can
# either give multiple identifier separated by comma (,) or put this option
# multiple time (only on the command line, not in the configuration file where
# it should appear only once). See also the "--disable" option for examples.
#enable=
# Disable the message, report, category or checker with the given id(s). You
# can either give multiple identifiers separated by comma (,) or put this
# option multiple times (only on the command line, not in the configuration
# file where it should appear only once).You can also use "--disable=all" to
# disable everything first and then reenable specific checks. For example, if
# you want to run only the similarities checker, you can use "--disable=all
# --enable=similarities". If you want to run only the classes checker, but have
# no Warning level messages displayed, use"--disable=all --enable=classes
# --disable=W"
disable=abstract-method,
apply-builtin,
arguments-differ,
attribute-defined-outside-init,
backtick,
bad-option-value,
basestring-builtin,
buffer-builtin,
c-extension-no-member,
consider-using-enumerate,
cmp-builtin,
cmp-method,
coerce-builtin,
coerce-method,
delslice-method,
div-method,
duplicate-code,
eq-without-hash,
execfile-builtin,
file-builtin,
filter-builtin-not-iterating,
fixme,
getslice-method,
global-statement,
hex-method,
idiv-method,
implicit-str-concat-in-sequence,
import-error,
import-self,
import-star-module-level,
inconsistent-return-statements,
input-builtin,
intern-builtin,
invalid-str-codec,
locally-disabled,
logging-fstring-interpolation, # added by vLLM
logging-not-lazy, # added by vLLM
long-builtin,
long-suffix,
map-builtin-not-iterating,
misplaced-comparison-constant,
missing-class-docstring, # TODO (vLLM): enable
missing-function-docstring,
missing-module-docstring, # TODO (vLLM): enable
metaclass-assignment,
next-method-called,
next-method-defined,
no-absolute-import,
no-else-break,
no-else-continue,
no-else-raise,
no-else-return,
no-init, # added
no-member,
no-name-in-module,
no-self-use,
nonzero-method,
oct-method,
old-division,
old-ne-operator,
old-octal-literal,
old-raise-syntax,
parameter-unpacking,
print-statement,
raising-string,
range-builtin-not-iterating,
raw_input-builtin,
rdiv-method,
reduce-builtin,
relative-import,
reload-builtin,
round-builtin,
setslice-method,
signature-differs,
standarderror-builtin,
suppressed-message,
sys-max-int,
too-few-public-methods,
too-many-ancestors,
too-many-arguments,
too-many-boolean-expressions,
too-many-branches,
too-many-instance-attributes,
too-many-locals,
too-many-nested-blocks,
too-many-public-methods,
too-many-return-statements,
too-many-statements,
trailing-newlines,
unichr-builtin,
unicode-builtin,
unnecessary-pass,
unpacking-in-except,
unspecified-encoding,
useless-else-on-loop,
useless-object-inheritance,
useless-suppression,
using-cmp-argument,
wrong-import-order,
xrange-builtin,
zip-builtin-not-iterating,
[REPORTS]
# Set the output format. Available formats are text, parseable, colorized, msvs
# (visual studio) and html. You can also give a reporter class, eg
# mypackage.mymodule.MyReporterClass.
output-format=text
# Tells whether to display a full report or only the messages
reports=no
# Python expression which should return a note less than 10 (10 is the highest
# note). You have access to the variables errors warning, statement which
# respectively contain the number of errors / warnings messages and the total
# number of statements analyzed. This is used by the global evaluation report
# (RP0004).
evaluation=10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10)
# Template used to display messages. This is a python new-style format string
# used to format the message information. See doc for all details
#msg-template=
[BASIC]
# Good variable names which should always be accepted, separated by a comma
good-names=main,_
# Bad variable names which should always be refused, separated by a comma
bad-names=
# Colon-delimited sets of names that determine each other's naming style when
# the name regexes allow several styles.
name-group=
# Include a hint for the correct naming format with invalid-name
include-naming-hint=no
# List of decorators that produce properties, such as abc.abstractproperty. Add
# to this list to register other decorators that produce valid properties.
property-classes=abc.abstractproperty,cached_property.cached_property,cached_property.threaded_cached_property,cached_property.cached_property_with_ttl,cached_property.threaded_cached_property_with_ttl
# Regular expression matching correct function names
function-rgx=^(?:(?P<exempt>setUp|tearDown|setUpModule|tearDownModule)|(?P<camel_case>_?[A-Z][a-zA-Z0-9]*)|(?P<snake_case>_?[a-z][a-z0-9_]*))$
# Regular expression matching correct variable names
variable-rgx=^[a-z][a-z0-9_]*$
# Regular expression matching correct constant names
const-rgx=^(_?[A-Z][A-Z0-9_]*|__[a-z0-9_]+__|_?[a-z][a-z0-9_]*)$
# Regular expression matching correct attribute names
attr-rgx=^_{0,2}[a-z][a-z0-9_]*$
# Regular expression matching correct argument names
argument-rgx=^[a-z][a-z0-9_]*$
# Regular expression matching correct class attribute names
class-attribute-rgx=^(_?[A-Z][A-Z0-9_]*|__[a-z0-9_]+__|_?[a-z][a-z0-9_]*)$
# Regular expression matching correct inline iteration names
inlinevar-rgx=^[a-z][a-z0-9_]*$
# Regular expression matching correct class names
class-rgx=^_?[A-Z][a-zA-Z0-9]*$
# Regular expression matching correct module names
module-rgx=^(_?[a-z][a-z0-9_]*|__init__)$
# Regular expression matching correct method names
method-rgx=(?x)^(?:(?P<exempt>_[a-z0-9_]+__|runTest|setUp|tearDown|setUpTestCase|tearDownTestCase|setupSelf|tearDownClass|setUpClass|(test|assert)_*[A-Z0-9][a-zA-Z0-9_]*|next)|(?P<camel_case>_{0,2}[A-Z][a-zA-Z0-9_]*)|(?P<snake_case>_{0,2}[a-z][a-z0-9_]*))$
# Regular expression which should only match function or class names that do
# not require a docstring.
no-docstring-rgx=(__.*__|main|test.*|.*test|.*Test)$
# Minimum line length for functions/classes that require docstrings, shorter
# ones are exempt.
docstring-min-length=10
[TYPECHECK]
# List of decorators that produce context managers, such as
# contextlib.contextmanager. Add to this list to register other decorators that
# produce valid context managers.
contextmanager-decorators=contextlib.contextmanager,contextlib2.contextmanager
# Tells whether missing members accessed in mixin class should be ignored. A
# mixin class is detected if its name ends with "mixin" (case insensitive).
ignore-mixin-members=yes
# List of module names for which member attributes should not be checked
# (useful for modules/projects where namespaces are manipulated during runtime
# and thus existing member attributes cannot be deduced by static analysis. It
# supports qualified module names, as well as Unix pattern matching.
ignored-modules=
# List of class names for which member attributes should not be checked (useful
# for classes with dynamically set attributes). This supports the use of
# qualified names.
ignored-classes=optparse.Values,thread._local,_thread._local
# List of members which are set dynamically and missed by pylint inference
# system, and so shouldn't trigger E1101 when accessed. Python regular
# expressions are accepted.
generated-members=
[FORMAT]
# Maximum number of characters on a single line.
max-line-length=80
# TODO(https://github.com/PyCQA/pylint/issues/3352): Direct pylint to exempt
# lines made too long by directives to pytype.
# Regexp for a line that is allowed to be longer than the limit.
ignore-long-lines=(?x)(
^\s*(\#\ )?<?https?://\S+>?$|
^\s*(from\s+\S+\s+)?import\s+.+$)
# Allow the body of an if to be on the same line as the test if there is no
# else.
single-line-if-stmt=yes
# Maximum number of lines in a module
max-module-lines=99999
# String used as indentation unit. The internal Google style guide mandates 2
# spaces. Google's externaly-published style guide says 4, consistent with
# PEP 8. Here, we use 2 spaces, for conformity with many open-sourced Google
# projects (like TensorFlow).
indent-string=' '
# Number of spaces of indent required inside a hanging or continued line.
indent-after-paren=4
# Expected format of line ending, e.g. empty (any line ending), LF or CRLF.
expected-line-ending-format=
[MISCELLANEOUS]
# List of note tags to take in consideration, separated by a comma.
notes=TODO
[STRING]
# This flag controls whether inconsistent-quotes generates a warning when the
# character used as a quote delimiter is used inconsistently within a module.
check-quote-consistency=yes
[VARIABLES]
# Tells whether we should check for unused import in __init__ files.
init-import=no
# A regular expression matching the name of dummy variables (i.e. expectedly
# not used).
dummy-variables-rgx=^\*{0,2}(_$|unused_|dummy_)
# List of additional names supposed to be defined in builtins. Remember that
# you should avoid to define new builtins when possible.
additional-builtins=
# List of strings which can identify a callback function by name. A callback
# name must start or end with one of those strings.
callbacks=cb_,_cb
# List of qualified module names which can have objects that can redefine
# builtins.
redefining-builtins-modules=six,six.moves,past.builtins,future.builtins,functools
[LOGGING]
# Logging modules to check that the string format arguments are in logging
# function parameter format
logging-modules=logging,absl.logging,tensorflow.io.logging
[SIMILARITIES]
# Minimum lines number of a similarity.
min-similarity-lines=4
# Ignore comments when computing similarities.
ignore-comments=yes
# Ignore docstrings when computing similarities.
ignore-docstrings=yes
# Ignore imports when computing similarities.
ignore-imports=no
[SPELLING]
# Spelling dictionary name. Available dictionaries: none. To make it working
# install python-enchant package.
spelling-dict=
# List of comma separated words that should not be checked.
spelling-ignore-words=
# A path to a file that contains private dictionary; one word per line.
spelling-private-dict-file=
# Tells whether to store unknown words to indicated private dictionary in
# --spelling-private-dict-file option instead of raising a message.
spelling-store-unknown-words=no
[IMPORTS]
# Deprecated modules which should not be used, separated by a comma
deprecated-modules=regsub,
TERMIOS,
Bastion,
rexec,
sets
# Create a graph of every (i.e. internal and external) dependencies in the
# given file (report RP0402 must not be disabled)
import-graph=
# Create a graph of external dependencies in the given file (report RP0402 must
# not be disabled)
ext-import-graph=
# Create a graph of internal dependencies in the given file (report RP0402 must
# not be disabled)
int-import-graph=
# Force import order to recognize a module as part of the standard
# compatibility libraries.
known-standard-library=
# Force import order to recognize a module as part of a third party library.
known-third-party=enchant, absl
# Analyse import fallback blocks. This can be used to support both Python 2 and
# 3 compatible code, which means that the block might have code that exists
# only in one or another interpreter, leading to false positives when analysed.
analyse-fallback-blocks=no
[CLASSES]
# List of method names used to declare (i.e. assign) instance attributes.
defining-attr-methods=__init__,
__new__,
setUp
# List of member names, which should be excluded from the protected access
# warning.
exclude-protected=_asdict,
_fields,
_replace,
_source,
_make
# List of valid names for the first argument in a class method.
valid-classmethod-first-arg=cls,
class_
# List of valid names for the first argument in a metaclass class method.
valid-metaclass-classmethod-first-arg=mcs
[EXCEPTIONS]
# Exceptions that will emit a warning when being caught. Defaults to
# "Exception"
overgeneral-exceptions=StandardError,
Exception,
BaseException
server/vllm/.readthedocs.yaml deleted 100644 → 0
View file @ 64def8e2
# Read the Docs configuration file
# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
version: 2
build:
os: ubuntu-22.04
tools:
python: "3.8"
sphinx:
configuration: docs/source/conf.py
# If using Sphinx, optionally build your docs in additional formats such as PDF
formats:
- pdf
# Optionally declare the Python requirements required to build your docs
python:
install:
- requirements: docs/requirements-docs.txt
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment