feat: add initial qwen2.5-vl model and test

docs/source/supported_models.md

@@ -27,6 +27,7 @@ Text Generation Inference enables serving optimized models. The following sectio
 - [StarCoder 2](https://huggingface.co/bigcode/starcoder2-15b-instruct-v0.1)
 - [Qwen 2](https://huggingface.co/collections/Qwen/qwen2-6659360b33528ced941e557f)
 - [Qwen 2 VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [Qwen 2.5 VL](https://huggingface.co/collections/Qwen/qwen25-66e81a666513e518adb90d9e)
 - [Opt](https://huggingface.co/facebook/opt-6.7b)
 - [T5](https://huggingface.co/google/flan-t5-xxl)
 - [Galactica](https://huggingface.co/facebook/galactica-120b)

integration-tests/models/__snapshots__/test_flash_qwen2_5_vl/test_flash_qwen2_5_vl_bay.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "The image showcases the Statue of Liberty, a colossal bronze statue located in New York Harbor, a heritage building in the United States. The statue has a majestic presence, with one arm raised towards the sun and the other hitched on her hip. It sits atop a keeper's walkway, observed from the water. Surrounding the statue is a lush green meadow, where picnic spots, walkways, and a visitor desk can be found. In front of the statue, a large marina can accommodate fourteen different kinds of boats. In the backdrop stands the Empire State Building, marking the crowded skyscrapers of New York City.",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1738342753,
+  "id": "",
+  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
+  "object": "chat.completion",
+  "system_fingerprint": "3.0.2-dev0-native",
+  "usage": {
+    "completion_tokens": 128,
+    "prompt_tokens": 8736,
+    "total_tokens": 8864
+  }
+}

integration-tests/models/__snapshots__/test_flash_qwen2_5_vl/test_flash_qwen2_5_vl_inpaint.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "The image shows a whimsical scene set in what appears to be a fast-food restaurant. Dominating the foreground is a large, green, inflatable dinosaur with realistic textures, giving it a Jurassic Park-like appearance. The dinosaur is wearing a red Adult Swim logo hat, adding a humorous touch to its appearance.\n\nSurrounding the dinosaur are various food items typically found in a fast-food restaurant, including French fries in a plastic cup, a hamburger on a plate, and a beverage in another cup. The hamburger is detailed with lettuce, tomato, and other typical fast-food ingredients.\n\nAccompanying the dinosaur is a realistic-looking owl perched on the table, which adds to the surreal and playful atmosphere of the scene. The background features the interior of the restaurant with neon signs and other typical decor elements, enhancing the overall theme of a fun and fantastical fast-food experience.\n\nOverall, the image is a playful and imaginative blend of a standard fast-food setting with an unexpected and amusing twist provided by the dinosaur and owl characters.",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1738343775,
+  "id": "",
+  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
+  "object": "chat.completion",
+  "system_fingerprint": "3.0.2-dev0-native",
+  "usage": {
+    "completion_tokens": 206,
+    "prompt_tokens": 5375,
+    "total_tokens": 5581
+  }
+}

integration-tests/models/__snapshots__/test_flash_qwen2_5_vl/test_flash_qwen2_5_vl_simple.json

@@ -0,0 +1,26 @@
+{
+  "choices": [
+    {
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null,
+      "message": {
+        "content": "The image depicts an anthropomorphic rabbit character wearing an intricate space suit, which includes a helmet with a starry face pattern and multiple suitors. The rabbit's ears are significantly large and upright, and it has a hitchhiker-like star antennas on its chest. The background is a reddish-orange, rocky landscape, suggesting a Martian environment. The suit has various buttons, a red button on the chest, and a reflective or illuminated dome on the head. The overall color scheme is dominated by shades of red, orange, and gray, giving a sense of a rugged, otherworldly setting.",
+        "name": null,
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "usage": null
+    }
+  ],
+  "created": 1738342872,
+  "id": "",
+  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
+  "object": "chat.completion",
+  "system_fingerprint": "3.0.2-dev0-native",
+  "usage": {
+    "completion_tokens": 121,
+    "prompt_tokens": 1363,
+    "total_tokens": 1484
+  }
+}

integration-tests/models/__snapshots__/test_flash_qwen2_5_vl/test_flash_qwen2_5_vl_simple_streaming.json

@@ -0,0 +1,20 @@
+{
+  "choices": [
+    {
+      "delta": {
+        "content": "",
+        "role": "assistant",
+        "tool_calls": null
+      },
+      "finish_reason": "stop",
+      "index": 0,
+      "logprobs": null
+    }
+  ],
+  "created": 1738343559,
+  "id": "",
+  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
+  "object": "chat.completion.chunk",
+  "system_fingerprint": "3.0.2-dev0-native",
+  "usage": null
+}

integration-tests/models/test_flash_qwen2_5_vl.py

@@ -0,0 +1,122 @@
+import pytest
+
+
+@pytest.fixture(scope="module")
+def flash_qwen2_5_vl_handle(launcher):
+    with launcher("Qwen/Qwen2.5-VL-3B-Instruct") as handle:
+        yield handle
+
+
+@pytest.fixture(scope="module")
+async def flash_qwen2_5(flash_qwen2_5_vl_handle):
+    await flash_qwen2_5_vl_handle.health(300)
+    return flash_qwen2_5_vl_handle.client
+
+
+@pytest.mark.private
+async def test_flash_qwen2_5_vl_simple(flash_qwen2_5, response_snapshot):
+    response = await flash_qwen2_5.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
+                        },
+                    },
+                    {"type": "text", "text": "Describe the image"},
+                ],
+            },
+        ],
+    )
+
+    assert (
+        response.choices[0].message.content
+        == "The image depicts an anthropomorphic rabbit character wearing an intricate space suit, which includes a helmet with a starry face pattern and multiple suitors. The rabbit's ears are significantly large and upright, and it has a hitchhiker-like star antennas on its chest. The background is a reddish-orange, rocky landscape, suggesting a Martian environment. The suit has various buttons, a red button on the chest, and a reflective or illuminated dome on the head. The overall color scheme is dominated by shades of red, orange, and gray, giving a sense of a rugged, otherworldly setting."
+    )
+
+    assert response == response_snapshot
+
+
+@pytest.mark.private
+async def test_flash_qwen2_5_vl_simple_streaming(flash_qwen2_5, response_snapshot):
+    responses = await flash_qwen2_5.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png"
+                        },
+                    },
+                    {"type": "text", "text": "Describe the image"},
+                ],
+            },
+        ],
+        stream=True,
+    )
+
+    count = 0
+    generated = ""
+    last_response = None
+    async for response in responses:
+        count += 1
+        generated += response.choices[0].delta.content
+        last_response = response
+
+    assert (
+        generated
+        == "The image depicts an anthropomorphic rabbit character wearing an intricate space suit, which includes a helmet with a starry face pattern and multiple suitors. The rabbit's ears are significantly large and upright, and it has a hitchhiker-like star antennas on its chest. The background is a reddish-orange, rocky landscape, suggesting a Martian environment. The suit has various buttons, a red button on the chest, and a reflective or illuminated dome on the head. The overall color scheme is dominated by shades of red, orange, and gray, giving a sense of a rugged, otherworldly setting."
+    )
+    assert count == 121
+    assert last_response == response_snapshot
+
+
+@pytest.mark.private
+async def test_flash_qwen2_5_vl_bay(flash_qwen2_5, response_snapshot):
+    response = await flash_qwen2_5.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                        },
+                    },
+                    {"type": "text", "text": "Describe the image"},
+                ],
+            },
+        ],
+    )
+    assert response == response_snapshot
+
+
+@pytest.mark.private
+async def test_flash_qwen2_5_vl_inpaint(flash_qwen2_5, response_snapshot):
+    response = await flash_qwen2_5.chat(
+        seed=42,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-inpaint.png"
+                        },
+                    },
+                    {"type": "text", "text": "Describe the image"},
+                ],
+            },
+        ],
+    )
+    assert response == response_snapshot

launcher/src/main.rs

@@ -2053,7 +2053,7 @@ fn main() -> Result<(), LauncherError> {
                 // this is a short term temporary fix to enable vlms to avoid rejecting images
                 let default_optimal = match config {
                     Some(ref config) => match config.model_type.as_deref() {
-                        Some("qwen2_vl") => 10_000,
+                        Some("qwen2_vl") | Some("qwen2_5_vl") => 10_000,
                         _ => 4096,
                     },
                     None => 4096,

nix/overlay.nix

@@ -13,16 +13,16 @@ final: prev: {
     (
       python-self: python-super: with python-self; {
         # Python package override example:
-        #   transformers = python-super.transformers.overrideAttrs (
-        #     _: _: {
-        #       src = final.fetchFromGitHub {
-        #         owner = "huggingface";
-        #         repo = "transformers";
-        #         rev = "2bd4d5897dc73e8b172832070a6f9e567a0df017";
-        #         hash = "sha256-JOIpKH9ssDEfI2Tf15e0iPKtThJwQ9GxMvRAnm+M2Pg=";
-        #       };
-        #    }
-        # );
+        transformers = python-super.transformers.overrideAttrs (
+          _: _: {
+            src = final.fetchFromGitHub {
+              owner = "huggingface";
+              repo = "transformers";
+              rev = "8d73a38606bc342b370afe1f42718b4828d95aaa";
+              hash = "sha256-MxroG6CWqrcmRS+eFt7Ej87TDOInN15aRPBUcaycKTI=";
+            };
+          }
+        );
       }
     )
   ];

router/src/config.rs

@@ -184,10 +184,43 @@ impl Qwen2Vl {
     }
 }
 
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub struct Qwen2_5VlVisionConfig {
+    pub(crate) depth: usize,
+    pub(crate) hidden_act: String,
+    pub(crate) hidden_size: usize,
+    pub(crate) intermediate_size: usize,
+    pub(crate) num_heads: usize,
+    pub(crate) in_chans: usize,
+    pub(crate) out_hidden_size: usize,
+    pub(crate) patch_size: usize,
+    pub(crate) spatial_merge_size: usize,
+    pub(crate) spatial_patch_size: usize,
+    pub(crate) window_size: usize,
+    pub(crate) fullatt_block_indexes: Vec<usize>,
+    pub(crate) tokens_per_second: usize,
+    pub(crate) temporal_patch_size: usize,
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub struct Qwen2_5Vl {
+    pub(crate) vision_config: Qwen2_5VlVisionConfig,
+}
+
+impl Qwen2_5Vl {
+    pub fn get_number_of_features(&self, height: usize, width: usize) -> usize {
+        let num_pixels = height * width;
+        num_pixels / self.vision_config.patch_size.pow(2)
+    }
+}
+
 #[derive(Clone, Debug, Serialize, Deserialize)]
 #[serde(tag = "model_type")]
 #[serde(rename_all = "snake_case")]
 pub enum Config {
+    Qwen2_5Vl(Qwen2_5Vl),
     Qwen2Vl(Qwen2Vl),
     LlavaNext(LlavaNext),
     ClipVisionModel(ClipVisionModel),

router/src/validation.rs

@@ -684,6 +684,10 @@ fn image_tokens(
             "<|vision_start|>{:?}<|vision_end|>",
             "<|image_pad|>".repeat(config.get_number_of_features(height, width))
         ),
+        Qwen2_5Vl(config) => format!(
+            "<|vision_start|>{:?}<|vision_end|>",
+            "<|image_pad|>".repeat(config.get_number_of_features(height, width))
+        ),
         _ => unimplemented!("Images tokens are not supported for this model configuration"),
     }
 }
@@ -712,7 +716,7 @@ fn prepare_input<T: TokenizerTrait>(
     let (tokenizer_query, input_chunks) = match config {
         Some(
             config @ (Idefics | Mllama | Idefics2(_) | Idefics3(_) | Paligemma(_) | LlavaNext(_)
-            | Qwen2Vl(_)),
+            | Qwen2Vl(_) | Qwen2_5Vl(_)),
         ) => {
             let mut input_chunks = Vec::new();
             let mut tokenizer_query = String::with_capacity(inputs.len());

server/text_generation_server/layers/rotary.py

@@ -86,7 +86,6 @@ def static(cls, config, dim, base, device):
             # `rope_type` is now standard in transformers, but some existing models
             # have `type` instead.
             rope_type = rope_scaling.get("rope_type", rope_scaling.get("type", None))
-            mrope_section = rope_scaling.get("mrope_section", None)
 
             if rope_type == "linear":
                 pass

server/text_generation_server/models/__init__.py

@@ -164,6 +164,11 @@
     from text_generation_server.models.custom_modeling.qwen2_vl import (
         Qwen2VLForConditionalGeneration,
     )
+    from text_generation_server.models.custom_modeling.qwen2_5_vl import (
+        Qwen2_5VLForConditionalGeneration,
+        Qwen2_5_VLConfig,
+        Qwen2_5_VLProcessor,
+    )
     from text_generation_server.layers.attention import SUPPORTS_WINDOWING
 except ImportError as e:
     log_master(logger.warning, f"Could not import Flash Attention enabled models: {e}")
@@ -317,6 +322,11 @@ class ModelType(enum.Enum):
         "name": "Qwen 2 VL",
         "url": "https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d",
     }
+    QWEN2_5_VL = {
+        "type": "qwen2_5_vl",
+        "name": "Qwen 2.5 VL",
+        "url": "https://huggingface.co/collections/Qwen/qwen25-66e81a666513e518adb90d9e",
+    }
     OPT = {
         "type": "opt",
         "name": "Opt",
@@ -1368,6 +1378,21 @@ def get_model(
             trust_remote_code=trust_remote_code,
             lora_adapter_ids=lora_adapter_ids,
         )
+    if model_type == QWEN2_5_VL:
+        return VlmCausalLM(
+            model_id=model_id,
+            model_class=Qwen2_5VLForConditionalGeneration,
+            revision=revision,
+            quantize=quantize,
+            speculator=speculator,
+            dtype=dtype,
+            default_dtype=torch.bfloat16,
+            kv_cache_dtype=kv_cache_dtype,
+            trust_remote_code=trust_remote_code,
+            lora_adapter_ids=lora_adapter_ids,
+            config_class=Qwen2_5_VLConfig,
+            processor_class=Qwen2_5_VLProcessor,
+        )
     if model_type == MLLAMA:
         if FLASH_ATTENTION:
             return MllamaCausalLM(