Tmp

backends/v3/src/backend.rs

@@ -35,20 +35,14 @@ impl BackendV3 {
         window_size: Option<u32>,
         speculate: u32,
     ) -> Self {
-        let prefix_caching = if let Ok(prefix_caching) = std::env::var("USE_PREFIX_CACHING") {
+        let prefix_caching =
-            matches!(prefix_caching.as_str(), "true" | "1")
+            std::env::var("USE_PREFIX_CACHING").expect("Expect prefix caching env var");
-        } else {
+        let prefix_caching = matches!(prefix_caching.as_str(), "true" | "1");
-            true
+        let attention: String = std::env::var("ATTENTION").expect("attention env var");
-        };
+
-        let attention = if let Ok(attention) = std::env::var("ATTENTION") {
+        let attention: Attention = attention
-            attention
             .parse()
-                .unwrap_or_else(|_| panic!("Invalid attention was specified :`{attention}`"))
+            .unwrap_or_else(|_| panic!("Invalid attention was specified :`{attention}`"));
-        } else if prefix_caching {
-            Attention::FlashInfer
-        } else {
-            Attention::FlashDecoding
-        };
         let block_size = if attention == Attention::FlashDecoding {
             256
         } else if attention == Attention::FlashInfer {

backends/v3/src/block_allocator.rs

@@ -1,4 +1,4 @@
-use std::{cmp::min, sync::Arc};
+use std::sync::Arc;
 use tokio::sync::{mpsc, oneshot};
 
 use crate::radix::RadixAllocator;
@@ -91,11 +91,7 @@ async fn block_allocator_task(
     window_size: Option<u32>,
     mut receiver: mpsc::UnboundedReceiver<BlockAllocatorCommand>,
 ) {
-    let mut allocator: Box<dyn Allocator + Send> = if prefix_caching {
+    let mut allocator = RadixAllocator::new(block_size, blocks, window_size, prefix_caching);
-        Box::new(RadixAllocator::new(block_size, blocks, window_size))
-    } else {
-        Box::new(SimpleAllocator::new(blocks, block_size, window_size))
-    };
     while let Some(cmd) = receiver.recv().await {
         match cmd {
             BlockAllocatorCommand::Free {
@@ -128,83 +124,12 @@ enum BlockAllocatorCommand {
     },
 }
 
-pub trait Allocator {
+// pub trait Allocator {
-    fn allocate(
+//     fn allocate(
-        &mut self,
+//         &mut self,
-        tokens: u32,
+//         tokens: u32,
-        prefill_tokens: Option<Arc<Vec<u32>>>,
+//         prefill_tokens: Option<Arc<Vec<u32>>>,
-    ) -> Option<BlockAllocation>;
+//     ) -> Option<BlockAllocation>;
-
+//
-    fn free(&mut self, blocks: Vec<u32>, allocation_id: u64);
+//     fn free(&mut self, blocks: Vec<u32>, allocation_id: u64);
-}
+// }
-
-pub struct SimpleAllocator {
-    free_blocks: Vec<u32>,
-    block_size: u32,
-    window_size: Option<u32>,
-}
-
-impl SimpleAllocator {
-    fn new(blocks: u32, block_size: u32, window_size: Option<u32>) -> Self {
-        SimpleAllocator {
-            block_size,
-            // Block 0 is reserved for health checks
-            free_blocks: (1..blocks).collect(),
-            window_size,
-        }
-    }
-}
-
-impl Allocator for SimpleAllocator {
-    fn allocate(
-        &mut self,
-        tokens: u32,
-        _prefill_tokens: Option<Arc<Vec<u32>>>,
-    ) -> Option<BlockAllocation> {
-        // Apply window size
-        let (required_blocks, repeats) = {
-            let (tokens, repeats) = match self.window_size {
-                None => (tokens, 1),
-                Some(window_size) => {
-                    let repeats = (tokens + window_size - 1) / window_size;
-                    let tokens = min(tokens, window_size);
-                    (tokens, repeats as usize)
-                }
-            };
-            // Pad to a multiple of block size
-            let required_blocks = (tokens + self.block_size - 1) / self.block_size;
-            (required_blocks, repeats)
-        };
-
-        let tokens = tokens as usize;
-        if required_blocks > self.free_blocks.len() as u32 {
-            None
-        } else {
-            let blocks = self
-                .free_blocks
-                .split_off(self.free_blocks.len() - required_blocks as usize);
-            let mut slots =
-                Vec::with_capacity((required_blocks * self.block_size * repeats as u32) as usize);
-
-            'slots: for block_id in blocks.repeat(repeats).iter() {
-                for s in (block_id * self.block_size)..((block_id + 1) * self.block_size) {
-                    slots.push(s);
-                    if slots.len() == tokens {
-                        break 'slots;
-                    }
-                }
-            }
-            Some(BlockAllocation {
-                allocation_id: 0,
-                blocks,
-                slots,
-                prefix_len: 0,
-                block_allocator: None,
-            })
-        }
-    }
-
-    fn free(&mut self, blocks: Vec<u32>, _allocation_id: u64) {
-        self.free_blocks.extend(blocks)
-    }
-}

backends/v3/src/radix.rs

@@ -5,7 +5,7 @@ use std::{
 
 use slotmap::{DefaultKey, SlotMap};
 
-use crate::block_allocator::{Allocator, BlockAllocation};
+use crate::block_allocator::BlockAllocation;
 
 pub struct RadixAllocator {
     allocation_id: u64,
@@ -21,10 +21,18 @@ pub struct RadixAllocator {
     // This isn't used because the prefix need to match without the windowing
     // mecanism. This at worst is overallocating, not necessarily being wrong.
     window_size: Option<u32>,
+
+    /// Wether to actual use the radix tree for searching or not.
+    prefix_caching: bool,
 }
 
 impl RadixAllocator {
-    pub fn new(block_size: u32, n_blocks: u32, window_size: Option<u32>) -> Self {
+    pub fn new(
+        block_size: u32,
+        n_blocks: u32,
+        window_size: Option<u32>,
+        prefix_caching: bool,
+    ) -> Self {
         assert_eq!(
             block_size, 1,
             "Radix tree allocator only works with block_size=1, was: {}",
@@ -42,6 +50,7 @@ impl RadixAllocator {
             // Block 0 is reserved for health checks.
             free_blocks: (1..n_blocks).collect(),
             window_size,
+            prefix_caching,
         }
     }
 
@@ -69,14 +78,16 @@ impl RadixAllocator {
     }
 }
 
-impl Allocator for RadixAllocator {
+// Allocator trait
-    fn allocate(
+impl RadixAllocator {
+    pub fn allocate(
         &mut self,
         tokens: u32,
         prefill_tokens: Option<Arc<Vec<u32>>>,
     ) -> Option<BlockAllocation> {
         let mut blocks = vec![];
-        let prefix_node = if let Some(prefill_tokens) = prefill_tokens.as_ref() {
+        let prefix_node = match (self.prefix_caching, prefill_tokens.as_ref()) {
+            (true, Some(prefill_tokens)) => {
                 let node_id = self
                     .cache_blocks
                     .find(prefill_tokens.as_slice(), &mut blocks);
@@ -84,8 +95,8 @@ impl Allocator for RadixAllocator {
                 // refcount to ensure that the prefix that was found is not evicted.
 
                 node_id
-        } else {
+            }
-            self.cache_blocks.root_id()
+            _ => self.cache_blocks.root_id(),
         };
 
         self.cache_blocks
@@ -126,7 +137,7 @@ impl Allocator for RadixAllocator {
         })
     }
 
-    fn free(&mut self, blocks: Vec<u32>, allocation_id: u64) {
+    pub fn free(&mut self, blocks: Vec<u32>, allocation_id: u64) {
         let allocation = match self.allocations.remove(&allocation_id) {
             Some(allocation) => allocation,
             None => unreachable!("Tried to free an unknown allocation."),
@@ -574,13 +585,11 @@ where
 mod tests {
     use std::sync::Arc;
 
-    use crate::block_allocator::Allocator;
-
     use super::RadixAllocator;
 
     #[test]
     fn allocator_reuses_prefixes() {
-        let mut cache = RadixAllocator::new(1, 12, None);
+        let mut cache = RadixAllocator::new(1, 12, None, true);
         let allocation = cache.allocate(8, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
         assert_eq!(allocation.blocks, vec![4, 5, 6, 7, 8, 9, 10, 11]);
         assert_eq!(allocation.slots, allocation.slots);
@@ -592,9 +601,23 @@ mod tests {
         assert_eq!(allocation.prefix_len, 4);
     }
 
+    #[test]
+    fn allocator_doesnt_reuses_prefixes() {
+        let mut cache = RadixAllocator::new(1, 12, None, false);
+        let allocation = cache.allocate(8, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
+        assert_eq!(allocation.blocks, vec![4, 5, 6, 7, 8, 9, 10, 11]);
+        assert_eq!(allocation.slots, allocation.slots);
+        assert_eq!(allocation.prefix_len, 0);
+        cache.free(allocation.blocks.clone(), allocation.allocation_id);
+
+        let allocation = cache.allocate(8, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
+        assert_eq!(allocation.blocks, vec![1, 2, 3, 8, 9, 10, 11, 7]);
+        assert_eq!(allocation.prefix_len, 0);
+    }
+
     #[test]
     fn allocator_collects_older_prefixes_first() {
-        let mut cache = RadixAllocator::new(1, 7, None);
+        let mut cache = RadixAllocator::new(1, 7, None, true);
         let allocation1 = cache.allocate(4, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
         assert_eq!(allocation1.blocks, vec![3, 4, 5, 6]);
         assert_eq!(allocation1.prefix_len, 0);
@@ -614,7 +637,7 @@ mod tests {
 
     #[test]
     fn allocator_frees_fully_overlapping_prefills() {
-        let mut cache = RadixAllocator::new(1, 10, None);
+        let mut cache = RadixAllocator::new(1, 10, None, true);
         let allocation1 = cache.allocate(4, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
         let allocation2 = cache.allocate(4, Some(Arc::new(vec![0, 1, 2, 3]))).unwrap();
 
@@ -630,7 +653,7 @@ mod tests {
 
     #[test]
     fn allocator_frees_partially_overlapping_prefills() {
-        let mut cache = RadixAllocator::new(1, 20, None);
+        let mut cache = RadixAllocator::new(1, 20, None, true);
         let allocation1 = cache.allocate(4, Some(Arc::new(vec![0, 1]))).unwrap();
         assert_eq!(allocation1.blocks, vec![16, 17, 18, 19]);
         assert_eq!(allocation1.prefix_len, 0);

launcher/src/main.rs

@@ -24,6 +24,44 @@ use tracing_subscriber::{filter::LevelFilter, EnvFilter};
 
 mod env_runtime;
 
+fn resolve_attention(config: &Config, lora_adapters: &Option<String>) -> (String, String) {
+    let mut prefix_caching: Option<String> = std::env::var("USE_PREFIX_CACHING").ok();
+    let mut attention: Option<String> = std::env::var("ATTENTION").ok();
+    match config.head_dim {
+        Some(h) if h == 64 || h == 128 || h == 256 => {
+            if lora_adapters.is_some() && prefix_caching.is_none() {
+                tracing::info!("Disabling prefix caching because of lora adapters");
+                prefix_caching = Some("0".to_string());
+            }
+            match config.model_type.as_deref() {
+                Some("gemma2") | Some("falcon") | Some("deepseek_v2") => {
+                    // Required because gemma2 needs bfloat16 which is not supported by
+                    // flashinfer ?
+                    if prefix_caching.is_none() {
+                        tracing::info!(
+                            "Forcing flash decoding because model {} requires it",
+                            config.model_type.as_ref().unwrap()
+                        );
+                        prefix_caching = Some("0".to_string());
+                        attention = Some("flashdecoding".to_string());
+                    }
+                }
+                _ => {}
+            }
+        }
+        _ => {
+            if prefix_caching.is_none() {
+                tracing::info!("Forcing flash decoding because head dim is not supported by flashinfer, also disabling prefix caching");
+                prefix_caching = Some("0".to_string());
+                attention = Some("flashdecoding".to_string());
+            }
+        }
+    }
+    let prefix_caching = prefix_caching.unwrap_or("true".to_string());
+    let attention = attention.unwrap_or("flashinfer".to_string());
+    (prefix_caching, attention)
+}
+
 #[derive(Deserialize)]
 struct RawConfig {
     max_position_embeddings: Option<usize>,
@@ -1496,28 +1534,10 @@ fn main() -> Result<(), LauncherError> {
             let config: RawConfig = serde_json::from_str(&content)?;
 
             let config: Config = config.into();
-            match config.head_dim {
+            let (prefix_caching, attention) = resolve_attention(&config, &args.lora_adapters);
-                Some(h) if h == 64 || h == 128 || h == 256 => {
+            tracing::info!("Using attention {attention} - Prefix caching {prefix_caching}");
-                    if args.lora_adapters.is_some() {
+            std::env::set_var("USE_PREFIX_CACHING", prefix_caching);
-                        tracing::info!("Disabling prefix caching because of lora adapters");
+            std::env::set_var("ATTENTION", attention);
-                        std::env::set_var("USE_PREFIX_CACHING", "0");
-                    }
-                    match config.model_type.as_deref() {
-                        Some("gemma2") | Some("falcon") | Some("deepseek_v2") => {
-                            // Required because gemma2 needs bfloat16 which is not supported by
-                            // flashinfer ?
-                            std::env::set_var("USE_PREFIX_CACHING", "0");
-                            std::env::set_var("ATTENTION", "flashdecoding");
-                        }
-                        _ => {}
-                    }
-                }
-                _ => {
-                    tracing::info!("Forcing flash decoding because head dim is not supported by flashinfer, also disabling prefix caching");
-                    std::env::set_var("USE_PREFIX_CACHING", "0");
-                    std::env::set_var("ATTENTION", "flashdecoding");
-                }
-            }
             let quantize = config.quantize;
 
             // Quantization usually means you're even more RAM constrained.

server/text_generation_server/models/flash_causal_lm.py

@@ -43,7 +43,6 @@ from text_generation_server.models.globals import (
     ATTENTION,
     BLOCK_SIZE,
     CUDA_GRAPHS,
-    PREFIX_CACHING,
     get_adapter_to_index,
 )
 from text_generation_server.layers.attention import Seqlen
@@ -266,7 +265,7 @@ class FlashCausalLMBatch(Batch):
 
             orig_input_length = len(tokenized_input)
 
-            if PREFIX_CACHING:
+            if ATTENTION == "flashinfer":
                 prefix_len = r.prefix_len
                 if prefix_len == orig_input_length:
                     assert prefix_len > 0
@@ -1452,7 +1451,7 @@ class FlashCausalLM(Model):
 
         if cu_seqlen_prefill is not None or cuda_graph is None:
             input_lengths = input_lengths + prefix_lens_tensor
-            if PREFIX_CACHING:
+            if ATTENTION == "flashinfer":
                 block_tables = block_tables_to_ragged(
                     block_tables=block_tables,
                     input_lengths=batch.input_lengths,

server/text_generation_server/models/globals.py

@@ -5,9 +5,9 @@ from typing import Dict, Optional
 
 from text_generation_server.utils.log import log_master
 
-PREFIX_CACHING = os.getenv("USE_PREFIX_CACHING", "1").lower() in {"1", "true"}
+PREFIX_CACHING = os.getenv("USE_PREFIX_CACHING").lower() in {"1", "true"}
 log_master(logger.info, f"Using prefix caching = {PREFIX_CACHING}")
-ATTENTION = os.getenv("ATTENTION", "flashinfer" if PREFIX_CACHING else "flashdecoding")
+ATTENTION = os.getenv("ATTENTION")
 _expected = {"paged", "flashdecoding", "flashinfer"}
 assert (
     ATTENTION in _expected