lang: support direct video inference (#9936)

Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>

python/sglang/lang/api.py

@@ -229,7 +229,7 @@ def image(expr: SglExpr):
     return SglImage(expr)
 
 
-def video(path: str, num_frames: int):
+def video(path: str, num_frames: int = -1):
     return SglVideo(path, num_frames)
 
 

python/sglang/lang/backend/runtime_endpoint.py

@@ -104,6 +104,7 @@ class RuntimeEndpoint(BaseBackend):
     def commit_lazy_operations(self, s: StreamExecutor):
         data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
         self._add_images(s, data)
+        self._add_videos(s, data)
         res = http_request(
             self.base_url + "/generate",
             json=data,
@@ -115,6 +116,7 @@ class RuntimeEndpoint(BaseBackend):
     def fill_image(self, s: StreamExecutor):
         data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
         self._add_images(s, data)
+
         res = http_request(
             self.base_url + "/generate",
             json=data,
@@ -181,6 +183,7 @@ class RuntimeEndpoint(BaseBackend):
                 data[item] = value
 
         self._add_images(s, data)
+        self._add_videos(s, data)
 
         res = http_request(
             self.base_url + "/generate",
@@ -222,6 +225,7 @@ class RuntimeEndpoint(BaseBackend):
 
         data["stream"] = True
         self._add_images(s, data)
+        self._add_videos(s, data)
 
         res = http_request(
             self.base_url + "/generate",
@@ -324,6 +328,8 @@ class RuntimeEndpoint(BaseBackend):
 
     def _generate_http_request(self, s: StreamExecutor, data):
         self._add_images(s, data)
+        self._add_videos(s, data)
+
         res = http_request(
             self.base_url + "/generate",
             json=data,
@@ -338,6 +344,11 @@ class RuntimeEndpoint(BaseBackend):
             assert len(s.images_) == 1, "Only support one image."
             data["image_data"] = s.images_[0][1]
 
+    def _add_videos(self, s: StreamExecutor, data):
+        if s.videos_:
+            assert len(s.videos_) == 1, "Only support one video."
+            data["video_data"] = s.videos_
+
     def _assert_success(self, res):
         if res.status_code != 200:
             try:

python/sglang/lang/chat_template.py

@@ -16,7 +16,9 @@ class ChatTemplate:
     role_prefix_and_suffix: Dict[str, Tuple[str, str]]
     stop_str: List[str] = ()
     image_token: str = "<image>"
+    video_token: str = "<video>"
     audio_token: str = "<audio>"
+
     style: ChatTemplateStyle = ChatTemplateStyle.PLAIN
 
     def get_prefix_and_suffix(
@@ -161,6 +163,7 @@ register_chat_template(
         style=ChatTemplateStyle.PLAIN,
         stop_str=("<|im_end|>",),
         image_token="<|vision_start|><|image_pad|><|vision_end|>",
+        video_token="<|vision_start|><|video_pad|><|vision_end|>",
     )
 )
 

python/sglang/lang/interpreter.py

@@ -32,11 +32,7 @@ from sglang.lang.ir import (
     SglVarScopeEnd,
     SglVideo,
 )
-from sglang.utils import (
-    encode_image_base64,
-    encode_video_base64,
-    get_exception_traceback,
-)
+from sglang.utils import encode_image_base64, get_exception_traceback
 
 
 def run_internal(state, program, func_args, func_kwargs, sync):
@@ -286,6 +282,7 @@ class StreamExecutor:
         # For vision
         self.images_ = []
         self.cur_images = []
+        self.videos_ = []
 
         # For fork/join
         self.fork_start_text_pos = None
@@ -372,6 +369,7 @@ class StreamExecutor:
             exes[i].cur_role_begin_pos = self.cur_role_begin_pos
             exes[i].fork_start_text_pos = len(self.text_)
             exes[i].images_ = list(self.images_)
+            exes[i].videos_ = list(self.videos_)
 
             # TODO(ying): handle API speculative execution
 
@@ -508,13 +506,9 @@ class StreamExecutor:
 
     def _execute_video(self, expr: SglVideo):
         path = expr.path
-        num_frames = expr.num_frames
 
-        base64_data = encode_video_base64(path, num_frames)
-
-        self.images_.append((path, base64_data))
-        self.cur_images.append((path, base64_data))
-        self.text_ += self.chat_template.image_token
+        self.videos_.append(path)
+        self.text_ += self.chat_template.video_token
 
     def _spec_gen(self, sampling_params):
         stop = sampling_params.stop

python/sglang/lang/ir.py

@@ -445,7 +445,7 @@ class SglImage(SglExpr):
 
 
 class SglVideo(SglExpr):
-    def __init__(self, path: str, num_frames: int):
+    def __init__(self, path: str, num_frames: int = -1):
         self.path = path
         self.num_frames = num_frames