vlm fixes

launcher/src/main.rs

@@ -58,7 +58,10 @@ fn get_config(
     };
 
     let content = std::fs::read_to_string(filename)?;
-    let config: RawConfig = serde_json::from_str(&content)?;
+    let mut config: RawConfig = serde_json::from_str(&content)?;
+    if let Some(text_config) = config.text_config {
+        config = *text_config;
+    }
 
     let config: Config = config.into();
     Ok(config)
@@ -79,7 +82,7 @@ fn resolve_attention(config: &Option<Config>, lora_adapters: &Option<String>) ->
                     prefix_caching = Some("0".to_string());
                 }
                 match config.model_type.as_deref() {
-                    Some("gemma2") | Some("falcon") | Some("deepseek_v2") => {
+                    Some("gemma") | Some("gemma2") | Some("falcon") | Some("deepseek_v2") => {
                         // Required because gemma2 needs bfloat16 which is not supported by
                         // flashinfer ?
                         if attention.is_none() {
@@ -96,7 +99,7 @@ fn resolve_attention(config: &Option<Config>, lora_adapters: &Option<String>) ->
             }
             _ => {
                 if attention.is_none() {
-                    tracing::info!("Forcing flash decoding because head dim is not supported by flashinfer, also disabling prefix caching");
+                    tracing::info!("Forcing flash decoding because head dim ({:?}) is not supported by flashinfer, also disabling prefix caching", config.head_dim);
                     attention = Some("flashdecoding".to_string());
                 }
                 if prefix_caching.is_none() {
@@ -122,6 +125,7 @@ struct RawConfig {
     num_attention_heads: Option<usize>,
     head_dim: Option<usize>,
     is_encoder_decoder: Option<bool>,
+    text_config: Option<Box<RawConfig>>,
 }
 
 #[derive(Deserialize)]

router/src/config.rs

@@ -7,6 +7,7 @@ pub struct LlavaNext {
     pub(crate) text_config: TextConfig,
     pub(crate) vision_config: VisionConfig,
     pub(crate) image_grid_pinpoints: Vec<(usize, usize)>,
+    pub(crate) image_token_index: u32,
 }
 
 fn get_anyres_image_grid_shape(
@@ -112,7 +113,9 @@ pub struct ClipVisionModel {
 
 #[derive(Clone, Debug, Serialize, Deserialize)]
 #[serde(rename_all = "snake_case")]
-pub struct Idefics2 {}
+pub struct Idefics2 {
+    pub(crate) image_token_id: u32,
+}
 
 impl Idefics2 {
     pub fn get_number_of_features(&self, _height: usize, _width: usize) -> usize {

router/src/validation.rs

@@ -1,5 +1,5 @@
 /// Payload validation logic
-use crate::config::Config;
+use crate::config::{Config, Idefics2};
 use crate::validation::ValidationError::{BestOfSampling, BestOfSeed, EmptyInput};
 use crate::{
     GenerateParameters, GenerateRequest, GrammarType, HubPreprocessorConfig, Idefics2Preprocessor,
@@ -605,7 +605,9 @@ fn image_tokens(
 fn image_id(config: &Config) -> u32 {
     use Config::*;
     match config {
-        Paligemma(pali_gemma) => pali_gemma.image_token_index,
+        Idefics2(idefics) => idefics.image_token_id,
+        LlavaNext(llava) => llava.image_token_index,
+        Paligemma(paligemma) => paligemma.image_token_index,
         _ => unimplemented!("Images tokens are not supported for this model configuration"),
     }
 }

server/text_generation_server/models/vlm_causal_lm.py

@@ -256,8 +256,6 @@ class VlmCausalLM(FlashCausalLM):
         trust_remote_code: bool,
         **kwargs,
     ):
-        if PREFIX_CACHING:
-            raise NotImplementedError("Vlm do not work with prefix caching yet")
         if processor_kwargs is None:
             processor_kwargs = {}
         self.processor = processor_class.from_pretrained(
@@ -347,7 +345,6 @@ class VlmCausalLM(FlashCausalLM):
             # This makes sure the max_s for the decode pass is correct.
             max_s = min(self.max_past(), max_s)
 
-        bs = input_ids.shape[0]
         # Try to find an associated cuda graph
         bs = input_ids.shape[0]
         sorted_padded_bs = sorted([k for k in self.cuda_graphs.keys() if k >= bs])
@@ -358,7 +355,7 @@ class VlmCausalLM(FlashCausalLM):
             cuda_graph = None
         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,