fix: run tokenizer.encode in executor to avoid blocking event loop

[sufubao]

Summary

• tokenizer.encode in httpserver/manager.py::_encode() is CPU-bound and was called synchronously on the asyncio event loop, blocking all concurrent I/O (streaming output, new request acceptance) during tokenization
• Moved to run_in_executor so the event loop stays responsive under concurrent long-text requests

Test results (Qwen3.5-35B-A3B, 22k tokens)

A heartbeat coroutine ticks every 1ms to detect event loop blocking:

Mode	Encode time	Heartbeat ticks	Max gap
sync (before)	41.7ms	0 (loop fully blocked)	N/A
run_in_executor (after)	40.2ms	36	2.1ms

Test plan

• python test/test_tokenizer_blocking.py — verifies event loop is not blocked during tokenization
• End-to-end server test with concurrent streaming requests

[gemini-code-assist]

Code Review

This pull request updates the HTTP server manager to execute tokenizer encoding within a thread pool executor, preventing synchronous operations from blocking the asyncio event loop. It also introduces a new test script to verify the non-blocking behavior. Feedback includes a recommendation to use asyncio.get_running_loop() for better consistency and a suggestion to avoid hardcoded local paths in the test suite to ensure portability.

[gemini-code-assist]

Using asyncio.get_running_loop() is preferred over asyncio.get_event_loop() inside coroutines. get_running_loop() is more explicit and ensures that an error is raised if no loop is currently running, which helps in catching bugs during development. Additionally, it avoids potential deprecation warnings in newer Python versions.

Suggested change

	prompt_ids = await asyncio.get_event_loop().run_in_executor(
	None,
	lambda: self.tokenizer.encode(
	prompt, multimodal_params, add_special_tokens=sampling_params.add_special_tokens
	),
	)
	else:
	prompt_ids = self.tokenizer.encode(prompt, add_special_tokens=sampling_params.add_special_tokens)
	prompt_ids = await asyncio.get_event_loop().run_in_executor(
	None,
	lambda: self.tokenizer.encode(prompt, add_special_tokens=sampling_params.add_special_tokens),
	)
	prompt_ids = await asyncio.get_running_loop().run_in_executor(
	None,
	lambda: self.tokenizer.encode(
	prompt, multimodal_params, add_special_tokens=sampling_params.add_special_tokens
	),
	)
	else:
	prompt_ids = await asyncio.get_running_loop().run_in_executor(
	None,
	lambda: self.tokenizer.encode(prompt, add_special_tokens=sampling_params.add_special_tokens),
	)

[gemini-code-assist]

The MODEL_DIR is hardcoded to a specific local path (/nvme/models/Qwen3.5-35B-A3B). This makes the test non-portable and likely to fail in CI environments or on other developers' machines. Consider using a small, publicly available model identifier (e.g., "Qwen/Qwen2.5-0.5B") or allowing the path to be set via an environment variable.