Merge branch 'main' into warmup_gaudi_backend

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

clients/python/pyproject.toml

@@ -11,15 +11,15 @@ repository = "https://github.com/huggingface/text-generation-inference"
 
 
 [tool.poetry.dependencies]
-python = "^3.7"
+python = "^3.9"
 pydantic = "> 2, < 3"
-aiohttp = "^3.8"
+aiohttp = "^3.11"
 huggingface-hub = ">= 0.12, < 1.0"
 
-[tool.poetry.dev-dependencies]
-pytest = "^6.2.5"
-pytest-asyncio = "^0.17.2"
-pytest-cov = "^3.0.0"
+[tool.poetry.group.dev.dependencies]
+pytest = "^8"
+pytest-asyncio = "^0.26"
+pytest-cov = "^6.0.0"
 
 [tool.pytest.ini_options]
 asyncio_mode = "auto"

router/src/config.rs

@@ -229,10 +229,13 @@ impl Llama4 {
     pub fn pixel_shuffle_ratio(&self) -> f64 {
         self.vision_config.pixel_shuffle_ratio
     }
-    pub fn get_aspect_ratios(&self, height: usize, width: usize) -> (usize, usize) {
+    pub fn get_aspect_ratios(
+        &self,
+        height: usize,
+        width: usize,
+        max_chunks: usize,
+    ) -> (usize, usize) {
         let patch_size = self.vision_config.image_size;
-        // How to avoid hardcoding this?
-        let max_chunks = 15;
         let supported = find_supported_resolutions(max_chunks, patch_size);
         let (target_h, target_w) = get_best_fit(height, width, &supported, false);
         (target_h / patch_size, target_w / patch_size)

router/src/lib.rs

@@ -204,7 +204,7 @@ pub struct Gemma3Processor {
 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub struct Llama4Processor {
     #[serde(default)]
-    do_image_splitting: bool,
+    max_patches: usize,
 }
 
 #[derive(Debug, Clone, Deserialize, Default)]

router/src/validation.rs

@@ -698,10 +698,14 @@ fn image_tokens(
             let image_height = config.image_size();
             let patch_size = config.patch_size();
             let pixel_shuffle_ratio = config.pixel_shuffle_ratio();
+            let max_patches = match preprocessor_config {
+                Some(HubPreprocessorConfig::Llama4Processor(cfg)) => cfg.max_patches,
+                _ => panic!("Expected Llama4Processor in preprocessor_config"),
+            };
             let downsample_ratio =
                 (1.0 / (pixel_shuffle_ratio * pixel_shuffle_ratio)).round() as usize;
 
-            let (ratio_h, ratio_w) = config.get_aspect_ratios(height, width);
+            let (ratio_h, ratio_w) = config.get_aspect_ratios(height, width, max_patches);
             let image_width = image_height; // Assuming pixel shape: [H][W][C]
 
             let num_patches_per_chunk =

server/text_generation_server/models/__init__.py

@@ -1041,7 +1041,6 @@ def get_model(
                 trust_remote_code=trust_remote_code,
                 processor_kwargs={
                     "use_fast": True,
-                    "size": {"height": 336, "width": 336},
                 },
             )
     elif model_type == BAICHUAN: