load tested

backends/client/src/v3/client.rs

@@ -159,6 +159,7 @@ impl Client {
                 blocks: vec![],
                 slots: vec![],
                 prefix_len: 0,
+                postfix_len: truncate,
                 // Set sampling parameters to also take these ops into account in the max memory
                 parameters: Some(NextTokenChooserParameters {
                     temperature: 0.9,

backends/client/src/v3/sharded_client.rs

@@ -246,6 +246,7 @@ impl Health for ShardedClient {
             blocks: vec![0],
             slots: (0..16).collect(),
             prefix_len: 0,
+            postfix_len: 1,
             adapter_id: None,
         };
         let batch = Batch {

backends/v3/src/backend.rs

@@ -34,9 +34,13 @@ impl BackendV3 {
         requires_padding: bool,
         window_size: Option<u32>,
         speculate: u32,
+        support_chunking: bool,
     ) -> Self {
-        let prefix_caching =
-            std::env::var("USE_PREFIX_CACHING").unwrap_or("1".to_string());
+        if support_chunking {
+            tracing::warn!("Model supports prefill chunking. `waiting_served_ratio` and `max_waiting_tokens` will be ignored.");
+        }
+
+        let prefix_caching = std::env::var("USE_PREFIX_CACHING").unwrap_or("1".to_string());
         let prefix_caching = matches!(prefix_caching.as_str(), "true" | "1");
         let attention: String = std::env::var("ATTENTION").unwrap_or("flashinfer".to_string());
 
@@ -52,6 +56,7 @@ impl BackendV3 {
             window_size,
             speculate,
             max_batch_total_tokens,
+            support_chunking,
         );
         let batching_task_notifier = Arc::new(Notify::new());
 
@@ -63,6 +68,7 @@ impl BackendV3 {
             max_batch_total_tokens,
             max_waiting_tokens,
             max_batch_size,
+            support_chunking,
             queue.clone(),
             batching_task_notifier.clone(),
         ));
@@ -127,6 +133,7 @@ pub(crate) async fn batching_task(
     max_batch_total_tokens: u32,
     max_waiting_tokens: usize,
     max_batch_size: Option<usize>,
+    support_chunking: bool,
     queue: Queue,
     notifier: Arc<Notify>,
 ) {
@@ -158,28 +165,44 @@ pub(crate) async fn batching_task(
                 // Get current batch info
                 let batch_size = batch.size;
                 let batch_max_tokens = batch.max_tokens;
+                let current_tokens = batch.current_tokens;
                 let mut batches = vec![batch];
                 metrics::gauge!("tgi_batch_current_size").set(batch_size as f64);
                 metrics::gauge!("tgi_batch_current_max_tokens").set(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
-                    // TODO: temporarily disable to avoid incorrect deallocation +
-                    //       reallocation when using prefix caching.
-                    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.saturating_sub(batch_size as usize));
+
+                let (min_size, max_size, prefill_token_budget) = if support_chunking {
+                    // Since the next batch will be concatenated with the current batch,
+                    // the current batch tokens must be subtracted to the prefill budget
+                    // In the future, we could concatenate beforehand
+                    let prefill_token_budget = max_batch_prefill_tokens - current_tokens;
+                    // We can ignore min_size and max_size
+                    // Models than rely on max_size cannot support chunking
+                    // Regarding min_size, chunking allow us to consistently run at the compute
+                    // bound, making min_size useless.
+                    (None, None, prefill_token_budget)
+                } else {
+                    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
+                        // TODO: temporarily disable to avoid incorrect deallocation +
+                        //       reallocation when using prefix caching.
+                        Some((batch_size as f32 * waiting_served_ratio).floor() as usize)
+                    };
+
+                    let max_size =
+                        max_batch_size.map(|max_size| max_size.saturating_sub(batch_size as usize));
+
+                    (min_size, max_size, max_batch_prefill_tokens)
+                };
 
                 // 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)
+                    .next_batch(min_size, max_size, prefill_token_budget, token_budget)
                     .await
                 {
                     // Tracking metrics

backends/v3/src/client/grpc_client.rs

@@ -159,6 +159,7 @@ impl Client {
                 blocks: vec![],
                 slots: vec![],
                 prefix_len: 0,
+                postfix_len: truncate,
                 // Set sampling parameters to also take these ops into account in the max memory
                 parameters: Some(NextTokenChooserParameters {
                     temperature: 0.9,

backends/v3/src/client/mod.rs

@@ -29,15 +29,6 @@ pub trait Health {
     async fn model_health(&self) -> Result<()>;
 }
 
-#[derive(Debug)]
-pub struct ShardInfo {
-    pub requires_padding: bool,
-    pub dtype: String,
-    pub device_type: String,
-    pub window_size: Option<u32>,
-    pub speculate: u32,
-}
-
 #[derive(Error, Debug, Clone)]
 pub enum ClientError {
     #[error("Could not connect to Text Generation server: {0}")]

backends/v3/src/client/sharded_client.rs

@@ -1,6 +1,6 @@
-use crate::client::{ClientError, Result};
+use crate::client::Health;
 /// Multi shard Client
-use crate::client::{Health, ShardInfo};
+use crate::client::{ClientError, Result};
 
 use crate::client::grpc_client::{DecodeTimings, PrefillTimings};
 use crate::client::{
@@ -49,13 +49,13 @@ impl ShardedClient {
 
     /// Get the model info
     #[instrument(skip(self))]
-    pub async fn info(&mut self) -> Result<ShardInfo> {
+    pub async fn info(&mut self) -> Result<InfoResponse> {
         let futures: Vec<_> = self
             .clients
             .iter_mut()
             .map(|client| client.info())
             .collect();
-        join_all(futures).await.pop().unwrap().map(ShardInfo::from)
+        join_all(futures).await.pop().unwrap()
     }
 
     /// GRPC health check
@@ -194,18 +194,6 @@ impl ShardedClient {
     }
 }
 
-impl From<InfoResponse> for ShardInfo {
-    fn from(value: InfoResponse) -> Self {
-        Self {
-            requires_padding: value.requires_padding,
-            dtype: value.dtype,
-            device_type: value.device_type,
-            window_size: value.window_size,
-            speculate: value.speculate,
-        }
-    }
-}
-
 #[async_trait]
 impl Health for ShardedClient {
     async fn device_health(&self) -> Result<()> {
@@ -248,6 +236,7 @@ impl Health for ShardedClient {
             slots: (0..16).collect(),
             prefix_len: 0,
             adapter_id: None,
+            postfix_len: 1,
         };
         let batch = Batch {
             id: u64::MAX,

backends/v3/src/lib.rs

@@ -29,6 +29,8 @@ pub struct BackendInfo {
     pub max_waiting_tokens: usize,
     #[schema(nullable = true, example = "null")]
     pub max_batch_size: Option<usize>,
+    #[schema(example = "false")]
+    pub support_chunking: bool,
 }
 
 #[allow(clippy::too_many_arguments)]
@@ -110,6 +112,7 @@ pub async fn connect_backend(
         model_device_type: shard_info.device_type.clone(),
         model_dtype: shard_info.dtype.clone(),
         speculate: shard_info.speculate as usize,
+        support_chunking: shard_info.support_chunking,
     };
 
     let backend = BackendV3::new(
@@ -122,6 +125,7 @@ pub async fn connect_backend(
         shard_info.requires_padding,
         shard_info.window_size,
         shard_info.speculate,
+        shard_info.support_chunking,
     );
 
     tracing::info!("Using backend V3");

backends/v3/src/main.rs

@@ -131,25 +131,12 @@ async fn main() -> Result<(), RouterError> {
             "`max_input_tokens` must be < `max_total_tokens`".to_string(),
         ));
     }
-    if max_input_tokens as u32 > max_batch_prefill_tokens {
-        return Err(RouterError::ArgumentValidation(format!("`max_batch_prefill_tokens` must be >= `max_input_tokens`. Given: {max_batch_prefill_tokens} and {max_input_tokens}")));
-    }
 
     if validation_workers == 0 {
         return Err(RouterError::ArgumentValidation(
             "`validation_workers` must be > 0".to_string(),
         ));
     }
-
-    if let Some(ref max_batch_total_tokens) = max_batch_total_tokens {
-        if max_batch_prefill_tokens > *max_batch_total_tokens {
-            return Err(RouterError::ArgumentValidation(format!("`max_batch_prefill_tokens` must be <= `max_batch_total_tokens`. Given: {max_batch_prefill_tokens} and {max_batch_total_tokens}")));
-        }
-        if max_total_tokens as u32 > *max_batch_total_tokens {
-            return Err(RouterError::ArgumentValidation(format!("`max_total_tokens` must be <= `max_batch_total_tokens`. Given: {max_total_tokens} and {max_batch_total_tokens}")));
-        }
-    }
-
     if let Some(max_batch_size) = max_batch_size {
         if max_batch_size == 0 {
             return Err(RouterError::ArgumentValidation(
@@ -158,7 +145,7 @@ async fn main() -> Result<(), RouterError> {
         }
     }
 
-    let (backend, _backend_info) = connect_backend(
+    let (backend, backend_info) = connect_backend(
         max_input_tokens,
         max_total_tokens,
         master_shard_uds_path,
@@ -170,6 +157,19 @@ async fn main() -> Result<(), RouterError> {
     )
     .await?;
 
+    // Validate remaining args now that the backend is known
+    let support_chunking = backend_info.support_chunking;
+    let max_batch_total_tokens = backend_info.max_batch_total_tokens;
+    if max_input_tokens as u32 > max_batch_prefill_tokens && !support_chunking {
+        return Err(RouterError::ArgumentValidation(format!("`max_batch_prefill_tokens` must be >= `max_input_tokens`. Given: {max_batch_prefill_tokens} and {max_input_tokens}")));
+    }
+    if max_batch_prefill_tokens > max_batch_total_tokens {
+        return Err(RouterError::ArgumentValidation(format!("`max_batch_prefill_tokens` must be <= `max_batch_total_tokens`. Given: {max_batch_prefill_tokens} and {max_batch_total_tokens}")));
+    }
+    if max_total_tokens as u32 > max_batch_total_tokens {
+        return Err(RouterError::ArgumentValidation(format!("`max_total_tokens` must be <= `max_batch_total_tokens`. Given: {max_total_tokens} and {max_batch_total_tokens}")));
+    }
+
     // Run server
     server::run(
         backend,

backends/v3/src/queue.rs

@@ -4,7 +4,7 @@ use crate::client::{
     Batch, GrammarType, NextTokenChooserParameters, Request, StoppingCriteriaParameters,
 };
 use nohash_hasher::{BuildNoHashHasher, IntMap};
-use std::cmp::{max, min};
+use std::cmp::max;
 use std::collections::VecDeque;
 use text_generation_router::infer::InferError;
 use text_generation_router::infer::InferStreamResponse;
@@ -50,6 +50,7 @@ impl Queue {
         window_size: Option<u32>,
         speculate: u32,
         max_batch_total_tokens: u32,
+        support_chunking: bool,
     ) -> Self {
         // Create channel
         let (queue_sender, queue_receiver) = mpsc::unbounded_channel();
@@ -62,6 +63,7 @@ impl Queue {
             window_size,
             speculate,
             max_batch_total_tokens,
+            support_chunking,
             queue_receiver,
         ));
 
@@ -108,6 +110,7 @@ impl Queue {
 }
 
 // Background task responsible of the queue state
+#[allow(clippy::too_many_arguments)]
 async fn queue_task(
     requires_padding: bool,
     block_size: u32,
@@ -115,6 +118,7 @@ async fn queue_task(
     window_size: Option<u32>,
     speculate: u32,
     max_batch_total_tokens: u32,
+    support_chunking: bool,
     mut receiver: mpsc::UnboundedReceiver<QueueCommand>,
 ) {
     let mut state = State::new(
@@ -124,6 +128,7 @@ async fn queue_task(
         window_size,
         speculate,
         max_batch_total_tokens,
+        support_chunking,
     );
 
     while let Some(cmd) = receiver.recv().await {
@@ -166,12 +171,14 @@ struct State {
     /// Paged Attention block size
     block_size: u32,
 
-    /// Sliding window
-    window_size: Option<u32>,
-
     /// Speculation amount
     speculate: u32,
 
+    /// Whether the model allow the prefill chunking
+    /// If it does, the last request in the batch will be split to exactly match the prefill
+    /// token budget
+    support_chunking: bool,
+
     /// Paged Attention Block Allocation
     block_allocator: Option<BlockAllocator>,
 }
@@ -184,6 +191,7 @@ impl State {
         window_size: Option<u32>,
         speculate: u32,
         max_batch_total_tokens: u32,
+        support_chunking: bool,
     ) -> Self {
         let block_allocator = (!requires_padding).then(|| {
             BlockAllocator::new(
@@ -199,8 +207,8 @@ impl State {
             next_id: 0,
             next_batch_id: 0,
             block_size,
-            window_size,
             speculate,
+            support_chunking,
             block_allocator,
         }
     }
@@ -268,7 +276,7 @@ impl State {
                 continue;
             }
 
-            let block_allocation = match &self.block_allocator {
+            let (block_allocation, postfix_len) = match &self.block_allocator {
                 None => {
                     // We pad to max input length in the Python shards
                     // We need to take these padding tokens into the equation
@@ -285,34 +293,9 @@ impl State {
                         self.entries.push_front((id, entry));
                         break 'entry_loop;
                     }
-                    None
+                    (None, entry.request.input_length)
                 }
-                Some(_block_allocator) => {
-                    prefill_tokens += entry.request.input_length;
-                    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,
-                        ),
-                    };
-                    decode_tokens += max_new_tokens;
-
-                    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;
-                    }
-
-                    let tokens = entry.request.input_length
-                        + entry.request.stopping_parameters.max_new_tokens
-                        + self.speculate
-                        - 1;
-
+                Some(block_allocator) => {
                     // If users wants the prefill logprobs, we cannot reuse the cache.
                     // So no input_ids for the radix tree.
                     let input_ids = if entry.request.decoder_input_details {
@@ -321,10 +304,65 @@ impl State {
                         entry.request.input_ids.clone()
                     };
 
-                    Some((tokens, input_ids))
+                    let tokens = entry.request.input_length
+                        + entry.request.stopping_parameters.max_new_tokens
+                        + self.speculate
+                        - 1;
+                    tracing::debug!("Allocating {tokens} with {input_ids:?}");
+
+                    let block_allocation = match block_allocator.allocate(tokens, input_ids).await {
+                        None => {
+                            // Entry is over budget
+                            // Add it back to the front
+                            tracing::debug!("Over budget: not enough free blocks");
+                            self.entries.push_front((id, entry));
+                            break 'entry_loop;
+                        }
+                        Some(mut block_allocation) => {
+                            tracing::debug!("Allocation: {block_allocation:?}");
+                            max_blocks = max(max_blocks, block_allocation.blocks.len() as u32);
+
+                            if block_allocation.prefix_len == entry.request.input_length {
+                                // The whole request was found in the radix trie
+                                // However, for the transformer forward to work, we need to
+                                // have at least one token of postfix.
+                                block_allocation.prefix_len -= 1;
+                            }
+
+                            block_allocation
+                        }
+                    };
+
+                    let mut postfix_len = entry.request.input_length - block_allocation.prefix_len;
+
+                    // Check equality too as if we don't we might end up with a postfix_len = 0
+                    // in the next iteration of the loop
+                    if prefill_tokens + postfix_len >= prefill_token_budget {
+                        // Entry is over budget
+                        if self.support_chunking {
+                            // We support chunking, just set postfix_len to exactly match prefill_token_budget
+                            postfix_len = prefill_token_budget - prefill_tokens;
+                            // Push this entry inside the batch
+                            batch.push((id, entry, Some(block_allocation), postfix_len));
+                            break 'entry_loop;
+                        } else {
+                            // We don't support chunking, this entry needs to go back to the buffer
+                            // Add it back to the front
+                            tracing::debug!(
+                                "Over budget: prefill_tokens={} > {prefill_token_budget}",
+                                prefill_tokens + postfix_len
+                            );
+                            self.entries.push_front((id, entry));
+                            break 'entry_loop;
+                        }
+                    }
+
+                    prefill_tokens += postfix_len;
+
+                    (Some(block_allocation), postfix_len)
                 }
             };
-            batch.push((id, entry, block_allocation));
+            batch.push((id, entry, block_allocation, postfix_len));
             if Some(batch.len()) == max_size {
                 break;
             }
@@ -342,7 +380,7 @@ impl State {
             // Batch is too small
             if batch.len() < min_size {
                 // Add back entries to the queue in the correct order
-                for (id, entry, _) in batch.into_iter().rev() {
+                for (id, entry, _, _) in batch.into_iter().rev() {
                     self.entries.push_front((id, entry));
                 }
                 return None;
@@ -353,29 +391,7 @@ impl State {
         let mut batch_entries =
             IntMap::with_capacity_and_hasher(self.entries.len(), BuildNoHashHasher::default());
 
-        for (id, mut entry, block_allocation) in batch {
-            let block_allocation = if let (Some((tokens, input_ids)), Some(block_allocator)) =
-                (block_allocation, &self.block_allocator)
-            {
-                tracing::debug!("Allocating {tokens} with {input_ids:?}");
-                match block_allocator.allocate(tokens, input_ids).await {
-                    None => {
-                        // Entry is over budget
-                        // Add it back to the front
-                        tracing::debug!("Over budget: not enough free blocks");
-                        self.entries.push_front((id, entry));
-                        continue;
-                    }
-                    Some(block_allocation) => {
-                        tracing::debug!("Allocation: {block_allocation:?}");
-                        max_blocks = max(max_blocks, block_allocation.blocks.len() as u32);
-                        Some(block_allocation)
-                    }
-                }
-            } else {
-                None
-            };
-            tracing::debug!("Accepting entry");
+        for (id, mut entry, block_allocation, postfix_len) in batch {
             // Create a new span to link the batch back to this entry
             let entry_batch_span = info_span!(parent: &entry.span, "infer");
             // Add relationships
@@ -429,6 +445,7 @@ impl State {
                 slots,
                 prefix_len,
                 adapter_id: entry.request.adapter_id.clone(),
+                postfix_len,
             });
             // Set batch_time
             entry.batch_time = Some(Instant::now());
@@ -436,12 +453,6 @@ impl State {
             batch_entries.insert(id, entry);
         }
 
-        // Empty batch
-        if batch_requests.is_empty() {
-            tracing::debug!("Filterered out all entries");
-            return None;
-        }
-
         // Final batch size
         let size = batch_requests.len() as u32;
         next_batch_span.record("batch_size", size);

benchmark/src/generation.rs

@@ -159,6 +159,7 @@ async fn prefill(
             blocks: vec![],
             slots: vec![],
             prefix_len: 0,
+            postfix_len: sequence_length,
             adapter_id: None,
         })
         .collect();

proto/v3/generate.proto

@@ -34,6 +34,7 @@ message InfoResponse {
   string device_type = 3;
   optional uint32 window_size = 4;
   uint32 speculate = 5;
+  bool support_chunking = 6;
 }
 
 /// Empty request
@@ -139,6 +140,8 @@ message Request {
   uint32 prefix_len = 12;
   /// Context truncation
   bool add_special_tokens = 13;
+  /// Postfix length for prefill chunking
+  uint32 postfix_len = 14;
 }
 
 message Batch {
@@ -163,6 +166,8 @@ message CachedBatch {
   uint32 size = 3;
   /// Maximum number of tokens this batch will grow to
   uint32 max_tokens = 4;
+  /// Number of tokens in the next forward
+  uint32 current_tokens = 5;
 }
 
 enum FinishReason {

server/tests/conftest.py

@@ -1,7 +1,7 @@
 import pytest
-import os
 from text_generation_server.pb import generate_pb2
 
+
 @pytest.fixture
 def default_pb_parameters():
     return generate_pb2.NextTokenChooserParameters(

server/text_generation_server/models/causal_lm.py

@@ -76,6 +76,7 @@ class CausalLMBatch(Batch):
             request_ids=[r.id for r in self.requests],
             size=len(self),
             max_tokens=self.max_tokens,
+            current_tokens=len(self),
         )
 
     @classmethod

server/text_generation_server/models/flash_causal_lm.py

@@ -16,7 +16,17 @@ from transformers import (
     AutoTokenizer,
     GenerationConfig,
 )
-from typing import Any, ContextManager, Iterable, Optional, Tuple, List, Type, Dict, Union
+from typing import (
+    Any,
+    ContextManager,
+    Iterable,
+    Optional,
+    Tuple,
+    List,
+    Type,
+    Dict,
+    Union,
+)
 
 from text_generation_server.adapters import AdapterBatchData, AdapterBatchMetadata
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
@@ -24,6 +34,10 @@ from text_generation_server.utils.chunks import concat_text_chunks
 from text_generation_server.utils.import_utils import SYSTEM
 from text_generation_server.models import Model
 from text_generation_server.utils.log import log_master
+from text_generation_server.utils.prefill_chunking import (
+    get_support_chunking,
+    get_max_prefill_tokens,
+)
 from text_generation_server.utils.tokens import batch_top_tokens
 from text_generation_server.utils.speculate import get_speculate
 from text_generation_server.utils import (
@@ -60,12 +74,9 @@ from text_generation_server.utils.import_utils import (
 
 tracer = trace.get_tracer(__name__)
 
-
 # Will be set in init
 SLIDING_WINDOW: Optional[int] = None
 
-TOKEN_BUDGET = 8
-
 
 def set_sliding_window(sliding_window: int):
     global SLIDING_WINDOW
@@ -206,6 +217,11 @@ class FlashCausalLMBatch(Batch):
             request_ids=[r.id for r in self.requests],
             size=len(self),
             max_tokens=self.num_blocks * BLOCK_SIZE,
+            current_tokens=(
+                sum([len(i) for i in self.input_ids])
+                if isinstance(self.input_ids, list)
+                else len(self.input_ids)
+            ),
         )
 
     @classmethod
@@ -272,7 +288,7 @@ class FlashCausalLMBatch(Batch):
             prompt_lengths.append(prompt_length)
 
             prefix_length = r.prefix_len
-            postfix_length = prefix_length + 10
+            postfix_length = r.postfix_len
             assert (
                 prefix_length <= prompt_length
             ), f"Prefix {prefix_length} vs input {prompt_length}"
@@ -282,10 +298,13 @@ class FlashCausalLMBatch(Batch):
             if prefix_length + postfix_length < prompt_length:
                 # FIXME: speculate is not supported for context chunking at the moment
                 assert speculate == 0
+                assert get_support_chunking()
+                assert postfix_length > 0
 
             prefix_ids.append(tokenized_input[:prefix_length])
-            postfix_ids = tokenized_input[prefix_length : postfix_length]
-            # postfix_ids = tokenized_input[prefix_length:]
+            postfix_ids = tokenized_input[
+                prefix_length : prefix_length + postfix_length
+            ]
 
             postfix_length = len(postfix_ids)
             postfix_lengths.append(postfix_length)
@@ -371,7 +390,6 @@ class FlashCausalLMBatch(Batch):
             requests=pb.requests,
             requests_idx_mapping=requests_idx_mapping,
             input_ids=all_postfix_ids,
-
             block_tables=block_tables,
             block_tables_tensor=block_tables_tensor,
             prefix_lengths=prefix_lengths,
@@ -395,7 +413,6 @@ class FlashCausalLMBatch(Batch):
             max_blocks=max_blocks,
             speculative_ids=None,
             prompt_lengths_tensor=prompt_lengths_tensor,
-
             # These values will be set by `FlashCausalLMBatch.prepare_for_prefill`
             position_ids=None,
             cu_seqlen_prefill=None,
@@ -431,7 +448,7 @@ class FlashCausalLMBatch(Batch):
         if len(request_ids) == len(self):
             return self
 
-        device = self.input_ids.device
+        device = self.block_tables_tensor.device
 
         # New values after filtering
         requests_idx_mapping = {}
@@ -552,13 +569,13 @@ class FlashCausalLMBatch(Batch):
 
         if self.prefilling:
             # These values will be set by `FlashCausalLMBatch.prepare_for_prefill`
-            position_ids=None
-            start_slots=None
-            slot_indices=None
-            slots=None
-            prefix_lengths_tensor=None
-            postfix_lengths_tensor=None
-            adapter_meta=None
+            position_ids = None
+            start_slots = None
+            slot_indices = None
+            slots = None
+            prefix_lengths_tensor = None
+            postfix_lengths_tensor = None
+            adapter_meta = None
         else:
             # Index into tensors
             input_ids = self.input_ids[indices]
@@ -643,24 +660,24 @@ class FlashCausalLMBatch(Batch):
         max_current_length = 0
         for b in batches:
             total_batch_size += len(b)
-            total_slots += len(b.slots)
+            max_blocks = max(max_blocks, b.max_blocks)
+            # If `b` is prefilling and was just filtered, `b.slots` is None
+            # `total_slots` is not used if any of the batches is prefilling
+            total_slots += len(b.slots) if not b.prefilling else 0
             num_blocks += b.num_blocks
             speculative_length = (
                 b.speculative_ids.shape[1] if b.speculative_ids is not None else 0
             )
-            max_blocks = max(max_blocks, b.max_blocks)
             max_postfix_length = max(max_postfix_length, b.max_postfix_length)
             max_current_length = max(max_current_length, b.max_current_length)
             max_length = max(
                 max_length,
                 max(
-                    prefix_length
-                    + postfix_length
+                    prompt_length
                     + stopping_criteria.max_new_tokens
                     + speculative_length
-                    - stopping_criteria.current_tokens
-                    for prefix_length, postfix_length, stopping_criteria in zip(
-                        b.prefix_lengths, b.postfix_lengths, b.stopping_criterias
+                    for prompt_length, stopping_criteria in zip(
+                        b.prompt_lengths, b.stopping_criterias
                     )
                 ),
             )
@@ -669,14 +686,14 @@ class FlashCausalLMBatch(Batch):
         if prefilling:
             input_ids = []
             # These values will be set by `FlashCausalLMBatch.prepare_for_prefill`
-            position_ids=None
-            start_slots=None
-            slots=None
-            slot_indices=None
-            prefix_lengths_tensor=None
-            postfix_lengths_tensor=None
-            adapter_meta=None
-            adapter_segment_builder=None
+            position_ids = None
+            start_slots = None
+            slots = None
+            slot_indices = None
+            prefix_lengths_tensor = None
+            postfix_lengths_tensor = None
+            adapter_meta = None
+            adapter_segment_builder = None
         else:
             input_ids = batches[0].input_ids.new_empty(total_batch_size)
             position_ids = batches[0].position_ids.new_empty(total_batch_size)
@@ -746,8 +763,6 @@ class FlashCausalLMBatch(Batch):
 
             start_index = cumulative_batch_size
             end_index = cumulative_batch_size + len(batch)
-            slots_start_index = cumulative_slots
-            slots_end_index = cumulative_slots + len(batch.slots)
 
             # Copy tensors (GPU)
             top_n_tokens_tensor[start_index:end_index] = batch.top_n_tokens_tensor
@@ -761,10 +776,17 @@ class FlashCausalLMBatch(Batch):
             prompt_lengths_tensor[start_index:end_index] = batch.prompt_lengths_tensor
 
             if not prefilling:
+                slots_start_index = cumulative_slots
+                slots_end_index = cumulative_slots + len(batch.slots)
+
                 input_ids[start_index:end_index] = batch.input_ids
                 position_ids[start_index:end_index] = batch.position_ids
-                slot_indices[start_index:end_index] = batch.slot_indices + cumulative_slots
-                postfix_lengths_tensor[start_index:end_index] = batch.postfix_lengths_tensor
+                slot_indices[start_index:end_index] = (
+                    batch.slot_indices + cumulative_slots
+                )
+                postfix_lengths_tensor[start_index:end_index] = (
+                    batch.postfix_lengths_tensor
+                )
                 slots[slots_start_index:slots_end_index] = batch.slots
 
                 # Copy over adapter indices
@@ -779,11 +801,17 @@ class FlashCausalLMBatch(Batch):
                 cumulative_adapter_indices_size = adapter_end_index
                 adapter_set.update(batch.adapter_meta.adapter_set)
                 adapter_segment_builder.concat(
-                    batch.adapter_meta.adapter_segments, batch.adapter_meta.segment_indices
+                    batch.adapter_meta.adapter_segments,
+                    batch.adapter_meta.segment_indices,
+                )
+                prefix_lengths_tensor[start_index:end_index] = (
+                    batch.prefix_lengths_tensor
                 )
-                prefix_lengths_tensor[start_index:end_index] = batch.prefix_lengths_tensor
 
                 start_slots.append(batch.start_slots + cumulative_slots)
+
+                # Update
+                cumulative_slots += len(batch.slots)
             else:
                 if isinstance(batch.input_ids, torch.Tensor):
                     batch.input_ids = batch.input_ids.view(-1, 1).tolist()
@@ -810,7 +838,6 @@ class FlashCausalLMBatch(Batch):
 
             # Update
             cumulative_batch_size += len(batch)
-            cumulative_slots += len(batch.slots)
 
         if start_slots is not None:
             start_slots = torch.concat(start_slots)
@@ -915,7 +942,7 @@ class FlashCausalLMBatch(Batch):
             postfix_length,
             prompt_length,
             request_prefilling,
-            blocks
+            blocks,
         ) in enumerate(
             zip(
                 self.requests,
@@ -923,7 +950,7 @@ class FlashCausalLMBatch(Batch):
                 self.postfix_lengths,
                 self.prompt_lengths,
                 self.prefilling_mask,
-                self.block_tables
+                self.block_tables,
             )
         ):
             next_chunk_length = postfix_length
@@ -967,9 +994,7 @@ class FlashCausalLMBatch(Batch):
             no_prefill_logprobs = no_prefill_logprobs and not prefill_logprobs
 
             if prefill_logprobs:
-                prefill_head_indices.append(
-                    request_position_ids + cumulative_length
-                )
+                prefill_head_indices.append(request_position_ids + cumulative_length)
                 prefill_next_token_indices.append(
                     prefill_out_cumulative_length + postfix_length - 1
                 )
@@ -988,7 +1013,6 @@ class FlashCausalLMBatch(Batch):
                 prefill_cu_outlens.append(prefill_out_cumulative_length + 1)
                 prefill_out_cumulative_length += 1
 
-
             start_slots.append(cumulative_slot_tokens)
             slots.extend(request_slots)
             slot_indices.append(request_slot_indices)
@@ -998,9 +1022,7 @@ class FlashCausalLMBatch(Batch):
 
             ADAPTER_TO_INDEX = get_adapter_to_index()
             adapter_index = ADAPTER_TO_INDEX.get(r.adapter_id, 0)
-            adapter_indices_list.append(
-                torch.full((next_chunk_length,), adapter_index)
-            )
+            adapter_indices_list.append(torch.full((next_chunk_length,), adapter_index))
             adapter_set.add(adapter_index)
 
             # Update
@@ -1240,6 +1262,7 @@ class FlashCausalLM(Model):
             rank=rank,
             world_size=world_size,
             sliding_window=config.sliding_window,
+            support_chunking=True,
         )
 
     @property
@@ -1764,29 +1787,43 @@ class FlashCausalLM(Model):
         finished_prefilling = True
         next_chunk_lengths = []
         if prefill:
-            next_prefilling_mask = []
-            # Budget in tokens for the next batch
-            # We remove next input ids to always have enough space for at least a single decode
-            # for the remaining requests
-            batch_budget = TOKEN_BUDGET - len(batch)
-            for prefix_length, postfix_length, prompt_length in zip(
-                batch.prefix_lengths, batch.postfix_lengths, batch.prompt_lengths
-            ):
-                remaining_prefill_tokens = max(
-                    prompt_length - prefix_length - postfix_length, 0
-                )
-                if remaining_prefill_tokens > 0:
-                    next_chunk_length = max(
-                        min(remaining_prefill_tokens, batch_budget), 1
+            if get_support_chunking():
+                next_prefilling_mask = []
+                # Budget in tokens for the next batch
+                # We remove len(batch) to always have enough space for at least a single decode
+                # for the remaining requests
+                batch_budget = get_max_prefill_tokens() - len(batch)
+                # We reverse to prioritize older requests
+                # zip() is not reversible so reverse the underlying lists instead
+                for prefix_length, postfix_length, prompt_length in zip(
+                    reversed(batch.prefix_lengths),
+                    reversed(batch.postfix_lengths),
+                    reversed(batch.prompt_lengths),
+                ):
+                    remaining_prefill_tokens = max(
+                        prompt_length - prefix_length - postfix_length, 0
                     )
-                    batch_budget -= next_chunk_length
-                    finished_prefilling = False
-                    next_prefilling_mask.append(True)
-                else:
-                    # Since speculation will be turned off, this is always true
-                    next_chunk_length = 1
-                    next_prefilling_mask.append(False)
-                next_chunk_lengths.append(next_chunk_length)
+                    if remaining_prefill_tokens > 0:
+                        next_chunk_length = max(
+                            min(remaining_prefill_tokens, batch_budget), 1
+                        )
+                        batch_budget -= next_chunk_length
+                        finished_prefilling = False
+                        next_prefilling_mask.append(True)
+                    else:
+                        # Since speculation will be turned off, this is always true
+                        next_chunk_length = 1
+                        next_prefilling_mask.append(False)
+                    next_chunk_lengths.append(next_chunk_length)
+
+                # Reverse back the obtained values²
+                next_chunk_lengths.reverse()
+                next_prefilling_mask.reverse()
+            else:
+                # The model does not support chunking
+                # We know we only do a single prefill
+                finished_prefilling = True
+                next_prefilling_mask = [False] * len(batch)
 
             batch.prefilling = not finished_prefilling
             batch.prefilling_mask = next_prefilling_mask
@@ -2179,7 +2216,9 @@ class FlashCausalLM(Model):
                 # have more than one new token per request with speculative decoding
                 for next_token_id in _next_token_ids:
                     batch.next_token_chooser = (
-                        batch.next_token_chooser.advance_grammar_single(i, next_token_id)
+                        batch.next_token_chooser.advance_grammar_single(
+                            i, next_token_id
+                        )
                     )
 
             # Update values

server/text_generation_server/models/globals.py

@@ -18,7 +18,7 @@ if PREFIX_CACHING and ATTENTION not in {"flashinfer", "flashdecoding"}:
     raise RuntimeError("Prefix caching is only supported with flashinfer")
 
 MEM_POOL = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None
-TGI_WIGGLE_ROOM = float(os.getenv("TGI_WIGGLE_ROOM", "0.95"))
+TGI_WIGGLE_ROOM = float(os.getenv("TGI_WIGGLE_ROOM", "0.90"))
 assert TGI_WIGGLE_ROOM > 0
 assert TGI_WIGGLE_ROOM < 1
 

server/text_generation_server/models/idefics_causal_lm.py

@@ -83,6 +83,7 @@ class IdeficsCausalLMBatch(Batch):
             request_ids=[r.id for r in self.requests],
             size=len(self),
             max_tokens=self.max_tokens,
+            current_tokens=len(self),
         )
 
     @classmethod

server/text_generation_server/models/mamba.py

@@ -116,6 +116,7 @@ class MambaBatch(Batch):
             request_ids=[r.id for r in self.requests],
             size=len(self),
             max_tokens=self.max_tokens,
+            current_tokens=len(self),
         )
 
     @classmethod

server/text_generation_server/models/model.py

@@ -5,8 +5,11 @@ from abc import ABC, abstractmethod
 from typing import List, Tuple, Optional, TypeVar, Type, Dict
 from collections import defaultdict
 from transformers import PreTrainedTokenizerBase
+from loguru import logger
 
 from text_generation_server.models.types import Batch, Generation
+from text_generation_server.utils.log import log_master
+from text_generation_server.utils.prefill_chunking import set_support_chunking
 from text_generation_server.utils.speculate import get_speculate
 from text_generation_server.pb.generate_pb2 import InfoResponse
 from text_generation_server.adapters.weights import LayerAdapterWeights
@@ -31,6 +34,7 @@ class Model(ABC):
         sliding_window: Optional[int] = None,
         speculate: Optional[int] = None,
         adapter_id: str = BASE_MODEL_ADAPTER_ID,
+        support_chunking: bool = False,
     ):
         self.model_id = model_id
         self.model = model.eval()
@@ -60,6 +64,17 @@ class Model(ABC):
             speculate = get_speculate()
         self.speculate = speculate
 
+        if speculate != 0 and support_chunking:
+            log_master(
+                logger.warning,
+                "Prefill chunking does not support speculation yet. "
+                "Prefill chunking will be turned off",
+            )
+            support_chunking = False
+
+        self.support_chunking = support_chunking
+        set_support_chunking(support_chunking)
+
         self.has_position_ids = (
             inspect.signature(model.forward).parameters.get("position_ids", None)
             is not None
@@ -78,6 +93,7 @@ class Model(ABC):
             device_type=self.device.type,
             window_size=self.sliding_window,
             speculate=self.speculate,
+            support_chunking=self.support_chunking,
         )
 
     @property

server/text_generation_server/models/seq2seq_lm.py

@@ -80,6 +80,7 @@ class Seq2SeqLMBatch(Batch):
             request_ids=[r.id for r in self.requests],
             size=len(self),
             max_tokens=self.max_tokens,
+            current_tokens=len(self),
         )
 
     @classmethod

server/text_generation_server/models/vlm_causal_lm.py

@@ -357,7 +357,6 @@ class VlmCausalLM(FlashCausalLM):
         else:
             cuda_graph = None
         if cu_seqlen_prefill is not None or cuda_graph is None:
-            input_lengths = postfix_lengths + prefix_lengths_tensor
             if PREFIX_CACHING:
                 block_tables = block_tables_to_ragged(
                     block_tables=block_tables,
@@ -424,7 +423,7 @@ class VlmCausalLM(FlashCausalLM):
         cuda_graph["postfix_lengths"][: postfix_lengths.shape[0]] = postfix_lengths
         cuda_graph["prefix_lengths"].zero_()
         cuda_graph["prefix_lengths"][
-        : prefix_lengths_tensor.shape[0]
+            : prefix_lengths_tensor.shape[0]
         ] = prefix_lengths_tensor
 
         with self._forward_context(

server/text_generation_server/server.py

@@ -15,6 +15,7 @@ from text_generation_server.cache import Cache
 from text_generation_server.interceptor import ExceptionInterceptor
 from text_generation_server.models import Model, get_model_with_lora_adapters
 from text_generation_server.utils.adapter import AdapterInfo
+from text_generation_server.utils.prefill_chunking import set_max_prefill_tokens
 
 try:
     from text_generation_server.models.pali_gemma import PaliGemmaBatch
@@ -96,6 +97,8 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
         return generate_pb2.FilterBatchResponse(batch=filtered_batch.to_pb())
 
     async def Warmup(self, request, context):
+        set_max_prefill_tokens(request.max_prefill_tokens)
+
         if self.quantize in {"exl2", "gptq"}:
             try:
                 # When using GPTQ, Exllama kernels need some global kernels

server/text_generation_server/utils/prefill_chunking.py

@@ -0,0 +1,24 @@
+from typing import Optional
+
+SUPPORT_CHUNKING: Optional[bool] = None
+MAX_PREFILL_TOKENS: Optional[int] = None
+
+
+def set_support_chunking(support_chunking: bool):
+    global SUPPORT_CHUNKING
+    SUPPORT_CHUNKING = support_chunking
+
+
+def get_support_chunking() -> bool:
+    global SUPPORT_CHUNKING
+    return SUPPORT_CHUNKING
+
+
+def set_max_prefill_tokens(max_prefill_tokens: int):
+    global MAX_PREFILL_TOKENS
+    MAX_PREFILL_TOKENS = max_prefill_tokens
+
+
+def get_max_prefill_tokens() -> int:
+    global MAX_PREFILL_TOKENS
+    return MAX_PREFILL_TOKENS