Merge upstream/main into moriio branch

[knitcapcat-amd]

Summary

Fix for vllm-project#1742:

• Merge upstream/main into moriio branch
• Resolve conflicts in 3 files:
  • vllm_omni/engine/async_omni_engine.py: keep both the mori receiver_connectors attachment and upstream's pd_config = self._detect_pd_config(); pass both connectors= and pd_config= to Orchestrator
  • vllm_omni/engine/orchestrator.py: keep both mori connector field/parameter/PUT logic and upstream's PD disaggregation state init / mm_features _ms == "thinker" filtering
  • vllm_omni/model_executor/stage_configs/qwen2_5_omni_multiconnector.yaml: drop (qwen2_5_omni configs were moved to vllm_omni/deploy/ by upstream refactor [Config Refactor 2.5/N] Centralize pipeline registry vllm-project/vllm-omni#2915; the mori_connector registration is preserved in qwen2_5_omni_mori_intranode.yaml)
• Re-tested intra-node Qwen2.5-Omni-7B with mori XGMI on 6×MI300X (TP=2 per stage):
  • All 4 mori connectors initialized (sender + receiver for both 0→1 and 1→2)
  • End-to-end inference works: text + audio (base64 WAV) returned
  • Mori PUT 0→1 confirmed via Orchestrator log

[gemini-code-assist]

Code Review

This pull request overhauls the CI/CD infrastructure, documentation, and benchmarking suite. It introduces Intel XPU test pipelines, refactors Buildkite configurations to support dynamic uploads and docs-only skip logic, and adds a nightly wheel release process. The ComfyUI extension is updated with video generation and LoRA support, and new accuracy benchmarks are added for diffusion and VoxCPM models. Documentation is extensively updated to reflect new features and design patterns. Review feedback identifies several critical improvements, including fixing shell command substitution bugs in bootstrap scripts, ensuring proper file handle closure in benchmark clients, and implementing cleanup for temporary synthetic images. Additionally, suggestions were made to avoid in-place file modifications and runtime package installations within CI scripts to improve reliability and maintainability.

[gemini-code-assist]

The command substitution $(git add .) is a bug. It executes git add ., which returns nothing, and then the shell attempts to execute an empty string as a command. It should be a standalone command. Additionally, echo $(git diff ...) is redundant as git diff already outputs to stdout. If the intention is to include all changes in the diff, consider using git add . as a standalone command, but generally CI should only diff against committed changes.

Suggested change

	$(git add .)
	echo $(git diff --name-only --diff-filter=ACMDR $(git merge-base origin/main HEAD))
	git add .
	git diff --name-only --diff-filter=ACMDR $(git merge-base origin/main HEAD)

[gemini-code-assist]

Same as above, $(git add .) is a bug and echo $(...) is redundant.

Suggested change

	$(git add .)
	echo $(git diff --name-only --diff-filter=ACMDR HEAD~1)
	git add .
	git diff --name-only --diff-filter=ACMDR HEAD~1

[gemini-code-assist]

Modifying repository files in-place with sed -i during CI is generally discouraged as it can lead to side effects in shared environments or if the script is reused. It is better to update generate-nightly-index.py to handle the missing regex dependency gracefully using a try-except import block.

[gemini-code-assist]

Installing packages via pip at runtime in a CI script makes the tests dependent on external network availability and can lead to flakiness. This dependency should be baked into the Docker image (Dockerfile.xpu).

[gemini-code-assist]

The file opened with open(audio_path, "rb") is not explicitly closed. It is recommended to use a context manager to ensure the file handle is released after the request is completed.

Suggested change

	data.add_field(
	"audio_sample",
	open(audio_path, "rb"),
	filename=os.path.basename(audio_path),
	content_type="audio/wav",
	)
	async with aiohttp.ClientSession() as session:
	async with session.post(url, data=data) as resp:
	result = await resp.json()
	print(f" Upload response ({resp.status}): {json.dumps(result, indent=2)}")
	return result
	if ref_text:
	data.add_field("ref_text", ref_text)
	with open(audio_path, "rb") as audio_file:
	data.add_field(
	"audio_sample",
	audio_file,
	filename=os.path.basename(audio_path),
	content_type="audio/wav",
	)
	async with aiohttp.ClientSession() as session:
	async with session.post(url, data=data) as resp:
	result = await resp.json()
	print(f" Upload response ({resp.status}): {json.dumps(result, indent=2)}")
	return result

[gemini-code-assist]

Synthetic image files created in the system temp directory are not cleaned up. This can lead to disk space issues if the benchmark is run repeatedly. Consider using a temporary directory that is automatically cleaned up or explicitly deleting the files after the benchmark finishes.