Rebase 2026/02/23

Update Dockerfile for vllm wheel version and enhance input processing in Omni models. Added decoding for byte messages in async requests, included new metrics in benchmarks, and improved pause state handling in the scheduler. Refined input and output processing methods for better multimodal support and error handling.

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
This commit is contained in:
tzhouam committed 2026-02-23 03:36:01 +00:00
1 parent 292da4b0d8
commit bf69cd94bb
9 files changed
+90 -77

No files matched your search

+5 -3
View File
@@ -92,6 +92,8 @@ class OmniInputPreprocessor(InputPreprocessor):
inputs["prompt_embeds"] = prompt_embeds
if additional_information is not None:
inputs["additional_information"] = additional_information
if prompt_text := parsed_content.get("prompt"):
inputs["prompt"] = prompt_text
if cache_salt := parsed_content.get("cache_salt"):
inputs["cache_salt"] = cache_salt
@@ -132,14 +134,14 @@ class OmniInputPreprocessor(InputPreprocessor):
* [`SingletonInputs`][vllm.inputs.data.SingletonInputs] instance
"""
if "prompt_embeds" in prompt:
return self._process_embeds(prompt) # type: ignore[arg-type]
if "prompt_token_ids" in prompt:
return self._process_tokens(
prompt, # type: ignore[arg-type]
)
if "prompt_embeds" in prompt:
return self._process_embeds(prompt) # type: ignore[arg-type]
if "prompt" in prompt:
return self._process_text(
prompt, # type: ignore[arg-type]