huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-04-19 22:02:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into warmup_gaudi_backend

Browse Source 
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
This commit is contained in:
Wang, Yi A 2025-04-14 18:24:34 -07:00
commit 6b21985c95
6 changed files with 1059 additions and 843 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1872
clients/python/poetry.lock generated
View File

File diff suppressed because it is too large Load Diff

12
clients/python/pyproject.toml
View File

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

9
router/src/config.rs
View File

@ -229,10 +229,13 @@ impl Llama4 {
pub fn pixel_shuffle_ratio(&self) -> f64 { pub fn pixel_shuffle_ratio(&self) -> f64 {
self.vision_config.pixel_shuffle_ratio 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; 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 supported = find_supported_resolutions(max_chunks, patch_size);
let (target_h, target_w) = get_best_fit(height, width, &supported, false); let (target_h, target_w) = get_best_fit(height, width, &supported, false);
(target_h / patch_size, target_w / patch_size) (target_h / patch_size, target_w / patch_size)

2
router/src/lib.rs
View File

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

6
router/src/validation.rs
View File

@ -698,10 +698,14 @@ fn image_tokens(
let image_height = config.image_size(); let image_height = config.image_size();
let patch_size = config.patch_size(); let patch_size = config.patch_size();
let pixel_shuffle_ratio = config.pixel_shuffle_ratio(); 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 = let downsample_ratio =
(1.0 / (pixel_shuffle_ratio * pixel_shuffle_ratio)).round() as usize; (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 image_width = image_height; // Assuming pixel shape: [H][W][C]
let num_patches_per_chunk = let num_patches_per_chunk =

1
server/text_generation_server/models/__init__.py
View File

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