feat: expose media_io_kwargs at runtime (#34778)

Signed-off-by: Alexandre Milesi <milesial@users.noreply.github.com>
This commit is contained in:
milesial
2026-03-07 04:27:04 +00:00
committed by GitHub
parent 58928475e4
commit 755356b3d1
20 changed files with 298 additions and 16 deletions
+2
View File
@@ -49,6 +49,7 @@ class DeepseekV32Renderer(BaseRenderer[DeepseekV32Tokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = tokenizer.apply_chat_template(
@@ -75,6 +76,7 @@ class DeepseekV32Renderer(BaseRenderer[DeepseekV32Tokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = tokenizer.apply_chat_template(
+2
View File
@@ -49,6 +49,7 @@ class Grok2Renderer(BaseRenderer[Grok2Tokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = tokenizer.apply_chat_template(
@@ -75,6 +76,7 @@ class Grok2Renderer(BaseRenderer[Grok2Tokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = tokenizer.apply_chat_template(
+2
View File
@@ -635,6 +635,7 @@ class HfRenderer(BaseRenderer[HfTokenizer]):
tokenizer=tokenizer,
model_config=model_config,
),
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = safe_apply_chat_template(
@@ -689,6 +690,7 @@ class HfRenderer(BaseRenderer[HfTokenizer]):
tokenizer=tokenizer,
model_config=model_config,
),
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = safe_apply_chat_template(
+2
View File
@@ -90,6 +90,7 @@ class MistralRenderer(BaseRenderer[MistralTokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = safe_apply_chat_template(
@@ -116,6 +117,7 @@ class MistralRenderer(BaseRenderer[MistralTokenizer]):
messages,
self.model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt_raw = await self._apply_chat_template_async(
+14 -2
View File
@@ -6,6 +6,7 @@ from typing import TYPE_CHECKING, Any, TypeVar
from vllm.exceptions import VLLMValidationError
from vllm.inputs import EmbedsPrompt, TextPrompt, TokensPrompt
from vllm.logger import init_logger
from vllm.multimodal.media.connector import merge_media_io_kwargs
from vllm.tokenizers import TokenizerLike
from vllm.utils.import_utils import LazyLoader
@@ -52,8 +53,15 @@ class ChatParams:
chat_template_kwargs: dict[str, Any] = field(default_factory=dict)
"""The kwargs to pass to the chat template."""
def with_defaults(self, default_chat_template_kwargs: dict[str, Any] | None):
if not default_chat_template_kwargs:
media_io_kwargs: dict[str, dict[str, Any]] | None = None
"""Per-modality kwargs for media I/O (loading/decoding images, videos, etc.)."""
def with_defaults(
self,
default_chat_template_kwargs: dict[str, Any] | None = None,
default_media_io_kwargs: dict[str, dict[str, Any]] | None = None,
):
if not default_chat_template_kwargs and not default_media_io_kwargs:
return self
return ChatParams(
@@ -63,6 +71,10 @@ class ChatParams:
default_chat_template_kwargs,
self.chat_template_kwargs,
),
media_io_kwargs=merge_media_io_kwargs(
default_media_io_kwargs,
self.media_io_kwargs,
),
)
def get_apply_chat_template_kwargs(self) -> dict[str, Any]:
+2
View File
@@ -43,6 +43,7 @@ class TerratorchRenderer(BaseRenderer):
messages,
model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt = parse_dec_only_prompt([1]) # Dummy token IDs
@@ -64,6 +65,7 @@ class TerratorchRenderer(BaseRenderer):
messages,
model_config,
content_format="string",
media_io_kwargs=params.media_io_kwargs,
)
prompt = parse_dec_only_prompt([1]) # Dummy token IDs