Using an enum for flash backens (paged/flashdecoding/flashinfer) (#2385)

* Using an enum for flash backens (paged/flashdecoding/flashinfer)

* Early exit on server too.

* Clippy.

* Fix clippy and fmt.

.gitignore

@@ -18,3 +18,4 @@ server/exllama_kernels/exllama_kernels/exllama_ext_hip.cpp
 
 data/
 load_tests/*.json
+server/fbgemmm

backends/v3/src/backend.rs

@@ -6,7 +6,7 @@ use nohash_hasher::IntMap;
 use std::sync::Arc;
 use text_generation_router::infer::{Backend, GeneratedText, InferError, InferStreamResponse};
 use text_generation_router::validation::ValidGenerateRequest;
-use text_generation_router::{FinishReason, PrefillToken, Token};
+use text_generation_router::{Attention, FinishReason, PrefillToken, Token};
 use tokio::sync::mpsc::error::SendError;
 use tokio::sync::{mpsc, Notify};
 use tokio::time::Instant;
@@ -35,12 +35,18 @@ impl BackendV3 {
         window_size: Option<u32>,
         speculate: u32,
     ) -> Self {
-        let flashdecoding = if let Ok(flashdecoding) = std::env::var("FLASH_DECODING") {
-            matches!(flashdecoding.to_lowercase().as_str(), "1" | "true")
+        let attention = if let Ok(attention) = std::env::var("ATTENTION") {
+            attention
+                .parse()
+                .unwrap_or_else(|_| panic!("Invalid attention was specified :`{attention}`"))
         } else {
-            false
+            Attention::Paged
+        };
+        let block_size = if attention == Attention::FlashDecoding {
+            256
+        } else {
+            16
         };
-        let block_size = if flashdecoding { 256 } else { 16 };
 
         let queue = Queue::new(
             requires_padding,

docs/openapi.json

@@ -2080,4 +2080,4 @@
       "description": "Hugging Face Text Generation Inference API"
     }
   ]
-}
+}

docs/source/conceptual/quantization.md

@@ -9,7 +9,7 @@ We recommend using the official quantization scripts for creating your quants:
 2. [GPTQ/ Marlin](https://github.com/AutoGPTQ/AutoGPTQ/blob/main/examples/quantization/basic_usage.py)
 3. [EXL2](https://github.com/turboderp/exllamav2/blob/master/doc/convert.md)
 
-For on-the-fly quantization you simply need to pass one of the supported quantization types and TGI takes care of the rest. 
+For on-the-fly quantization you simply need to pass one of the supported quantization types and TGI takes care of the rest.
 
 ## Quantization with bitsandbytes, EETQ & fp8
 
@@ -69,4 +69,4 @@ text-generation-launcher --model-id /data/falcon-40b-gptq/ --sharded true --num-
 You can learn more about the quantization options by running `text-generation-server quantize --help`.
 
 If you wish to do more with GPTQ models (e.g. train an adapter on top), you can read about transformers GPTQ integration [here](https://huggingface.co/blog/gptq-integration).
-You can learn more about GPTQ from the [paper](https://arxiv.org/pdf/2210.17323.pdf).
+You can learn more about GPTQ from the [paper](https://arxiv.org/pdf/2210.17323.pdf).

launcher/src/main.rs

@@ -1461,7 +1461,7 @@ fn main() -> Result<(), LauncherError> {
 
         if config.model_type == Some("gemma2".to_string()) {
             tracing::info!("Forcing flash decoding because of softcap usage");
-            std::env::set_var("FLASH_DECODING", "1");
+            std::env::set_var("ATTENTION", "flashdecoding");
         }
         let config: Config = config.into();
 

router/src/infer/v2/scheduler.rs

@@ -1,10 +1,10 @@
 /// Batching and inference logic
 use crate::infer::v2::queue::{Entry, Queue};
 use crate::infer::{
-    Backend, GenerateStreamResponse, GeneratedText, InferError, InferStreamResponse,
+    Attention, Backend, GenerateStreamResponse, GeneratedText, InferError, InferStreamResponse,
 };
 use crate::validation::ValidGenerateRequest;
-use crate::{FinishReason, PrefillToken, Token};
+use crate::{Attention, FinishReason, PrefillToken, Token};
 use nohash_hasher::IntMap;
 use std::sync::{
     atomic::{AtomicBool, Ordering},
@@ -40,12 +40,18 @@ impl BackendV2 {
         generation_health: Arc<AtomicBool>,
     ) -> Self {
         // Infer shared state
-        let flashdecoding = if let Ok(flashdecoding) = std::env::var("FLASH_DECODING") {
-            matches!(flashdecoding.to_lowercase().as_str(), "1" | "true")
+        let attention = if let Ok(attention) = std::env::var("ATTENTION") {
+            attention
+                .parse()
+                .expect(&format!("Invalid attention was specified :`{attention}`"))
         } else {
-            false
+            Attention::Paged
+        };
+        let block_size = if attention == Attention::FlashDecoding {
+            256
+        } else {
+            16
         };
-        let block_size = if flashdecoding { 256 } else { 16 };
         let queue = Queue::new(requires_padding, block_size, window_size, speculate);
         let batching_task_notifier = Arc::new(Notify::new());
 

router/src/lib.rs

@@ -15,6 +15,35 @@ use tracing::warn;
 use utoipa::ToSchema;
 use validation::Validation;
 
+#[derive(PartialEq)]
+pub enum Attention {
+    Paged,
+    FlashDecoding,
+    FlashInfer,
+}
+
+#[derive(Debug)]
+pub struct ParseError;
+
+impl std::fmt::Display for ParseError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "Cannot parse attention value")
+    }
+}
+impl std::error::Error for ParseError {}
+
+impl std::str::FromStr for Attention {
+    type Err = ParseError;
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "paged" => Ok(Attention::Paged),
+            "flashdecoding" => Ok(Attention::FlashDecoding),
+            "flashinfer" => Ok(Attention::FlashInfer),
+            _ => Err(ParseError),
+        }
+    }
+}
+
 #[derive(Clone, Deserialize, ToSchema)]
 pub(crate) struct VertexInstance {
     #[schema(example = "What is Deep Learning?")]

server/text_generation_server/layers/attention/common.py

@@ -1,10 +1,10 @@
 from dataclasses import dataclass
-from text_generation_server.models.globals import FLASH_DECODING, FLASH_INFER
+from text_generation_server.models.globals import ATTENTION
 import torch
 from typing import Optional
 
 
-if FLASH_DECODING or FLASH_INFER:
+if ATTENTION in {"flashinfer", "flashdecoding"}:
 
     @dataclass
     class Seqlen:

server/text_generation_server/layers/attention/cuda.py

@@ -1,9 +1,8 @@
 import torch
 from text_generation_server.utils.import_utils import SYSTEM
 from text_generation_server.models.globals import (
-    FLASH_DECODING,
+    ATTENTION,
     BLOCK_SIZE,
-    FLASH_INFER,
 )
 from text_generation_server.layers.attention import Seqlen
 from typing import Optional
@@ -27,7 +26,7 @@ def reshape_and_cache(
     value_cache: torch.Tensor,
     slots: torch.Tensor,
 ):
-    if FLASH_DECODING or FLASH_INFER:
+    if ATTENTION in {"flashdecoding", "flashinfer"}:
         shape = key_cache.shape
         key_cache.view(-1, shape[-2], shape[-1])[slots] = key
         value_cache.view(-1, shape[-2], shape[-1])[slots] = value
@@ -76,7 +75,7 @@ def paged_attention(
     # 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.
-    if FLASH_INFER:
+    if ATTENTION == "flashinfer":
         from text_generation_server.layers.attention.flash_infer import decode_state
 
         return decode_state.get().forward(
@@ -85,7 +84,7 @@ def paged_attention(
             logits_soft_cap=softcap,
             sm_scale=softmax_scale,
         )
-    elif FLASH_DECODING:
+    elif ATTENTION == "flashdecoding":
         max_q = 1
         max_k = max_s
         import flash_attn_2_cuda
@@ -219,7 +218,7 @@ def paged_attention(
 
 SUPPORTS_WINDOWING = V2
 
-if FLASH_INFER:
+if ATTENTION == "flashinfer":
 
     def attention(
         q,

server/text_generation_server/layers/attention/rocm.py

@@ -1,7 +1,7 @@
 import os
 import torch
 from text_generation_server.utils.import_utils import SYSTEM
-from text_generation_server.models.globals import FLASH_DECODING
+from text_generation_server.models.globals import ATTENTION
 from text_generation_server.layers.attention import Seqlen
 from text_generation_server.utils.log import log_master
 from loguru import logger
@@ -28,7 +28,7 @@ def reshape_and_cache(
     value_cache: torch.Tensor,
     slots: torch.Tensor,
 ):
-    if FLASH_DECODING:
+    if ATTENTION == "flashdecoding":
         shape = key_cache.shape
         key_cache.view(-1, shape[-2], shape[-1])[slots] = key
         value_cache.view(-1, shape[-2], shape[-1])[slots] = value

server/text_generation_server/models/flash_causal_lm.py

@@ -40,8 +40,7 @@
 from text_generation_server.pb import generate_pb2
 from text_generation_server.models.globals import (
     MEM_POOL,
-    FLASH_DECODING,
-    FLASH_INFER,
+    ATTENTION,
     BLOCK_SIZE,
     CUDA_GRAPHS,
     get_adapter_to_index,
@@ -938,7 +937,7 @@ def __init__(
         self.cuda_graphs = {}
         self.kv_cache = []
 
-        if FLASH_INFER:
+        if ATTENTION == "flashinfer":
             from text_generation_server.layers.attention.flash_infer import (
                 create_prefill_state,
                 create_decode_state,
@@ -990,7 +989,7 @@ def init_kv_cache(
         else:
             x = BLOCK_SIZE // element_size
 
-        if FLASH_DECODING or FLASH_INFER:
+        if ATTENTION in {"flashdecoding", "flashinfer"}:
             self.kv_cache = [
                 (
                     torch.empty(
@@ -1062,7 +1061,7 @@ def cuda_graph_warmup(self, bs: int, max_s: int, max_bt: int):
         graph = torch.cuda.CUDAGraph()
         self.cuda_graphs[bs]["graph"] = graph
 
-        if FLASH_INFER:
+        if ATTENTION == "flashinfer":
             from text_generation_server.layers.attention.flash_infer import (
                 create_decode_state_cuda_graphs,
             )
@@ -1766,7 +1765,7 @@ def _forward_context(
         input_lengths: torch.Tensor,
         state: Optional[Any] = None,
     ) -> ContextManager:
-        if not FLASH_INFER:
+        if ATTENTION != "flashinfer":
             return nullcontext()
 
         from text_generation_server.layers.attention.flash_infer import (

server/text_generation_server/models/globals.py

@@ -5,16 +5,16 @@
 
 from text_generation_server.utils.log import log_master
 
-FLASH_INFER = os.getenv("FLASH_INFER") in {"1", "true", "True"}
-if FLASH_INFER:
-    log_master(logger.info, "Using FLASH_INFER")
+ATTENTION = os.getenv("ATTENTION", "paged")
+_expected = {"paged", "flashdecoding", "flashinfer"}
+assert (
+    ATTENTION in _expected
+), f"Attention is not valid {ATTENTION}, expected {_expected}"
+log_master(logger.info, f"Using Attention = {ATTENTION}")
 
 MEM_POOL = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None
 # This is overridden by the cli
-FLASH_DECODING = os.getenv("FLASH_DECODING") in {"1", "true", "True"}
-BLOCK_SIZE: int = 256 if FLASH_DECODING else 16
-if FLASH_DECODING:
-    log_master(logger.info, "Using FLASH_DECODING")
+BLOCK_SIZE: int = 256 if ATTENTION == "flashdecoding" else 16
 
 
 cuda_graphs = os.getenv("CUDA_GRAPHS")