Fix some minor issues and provide tests for Pipeline

[windreamer]

Based on the commit history and code changes, I'll write a professional PR description in English.

---

PR Description

Motivation

This PR addresses several minor issues in the LMDeploy inference pipeline and engine to improve robustness, fix edge cases in token generation, and establish comprehensive test coverage for the pipeline functionality.

Modification

1. Fix kwargs forwarding in pipeline chat method
The chat method in Pipeline class was not properly forwarding keyword arguments to stream_infer, which could cause unexpected behavior when users passed custom generation parameters. This fix ensures all kwargs are correctly propagated through the call chain.

2. Fix off-by-one error when max_new_tokens=0
Previously, when max_new_tokens was set to 0, the engine would still generate 1 extra token due to incorrect boundary checking logic. This PR refactors the token limit validation in AsyncEngine to handle the zero-token case correctly, ensuring immediate termination without producing any output tokens when explicitly requested.

3. Provide default engine outputs for empty generators
Added defensive initialization of EngineOutput with INTERNAL_ENGINE_ERROR status as a fallback when async_stream_infer yields an empty generator. This prevents potential unbound variable errors and provides clearer error signaling in edge cases where the engine fails to produce outputs.

4. Add comprehensive pipeline test suite
Introduced a new test file tests/test_lmdeploy/test_pipeline.py with extensive coverage of:

• Single and batch inference across both PyTorch and Turbomind backends
• OpenAI-style message format handling
• Streaming and non-streaming generation modes
• Multi-turn conversation with session management
• Perplexity calculation (get_ppl)
• Edge cases including max_new_tokens=0 and varying generation configurations

The test suite uses pytest parametrization to ensure consistent behavior across both backend implementations.

[Copilot]

Pull request overview

This PR improves robustness and API consistency in LMDeploy’s inference pipeline/engine, and adds an end-to-end test suite intended to cover key Pipeline behaviors and edge cases (notably max_new_tokens=0).

Changes:

• Update AsyncEngine.generate() to early-exit when max_new_tokens resolves to 0, and add a default EngineOutput fallback when the engine yields no outputs.
• Forward **kwargs from Pipeline.chat() into stream_infer() for better parameter propagation.
• Add a new tests/test_lmdeploy/test_pipeline.py integration test suite for Pipeline infer/stream/chat/session/ppl behaviors, including max_new_tokens=0.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 9 comments.

File	Description
tests/test_lmdeploy/test_pipeline.py	Adds Pipeline integration tests across backends, streaming/chat flows, and max_new_tokens=0 edge case.
lmdeploy/serve/core/async_engine.py	Refactors gen-config determination/early-exit and initializes default EngineOutput to handle empty engine generators.
lmdeploy/pipeline.py	Fixes Pipeline.chat() kwarg propagation to stream_infer().

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Logging run out of tokens at error level is misleading when the user explicitly sets max_new_tokens=0 (a valid request). Consider lowering the level (info/debug) and/or changing the message to reflect an intentional zero-token generation.

Suggested change

	logger.error(f'run out of tokens. session={session_id}.')
	logger.info(f'no tokens requested (max_new_tokens=0). session={session_id}.')

[lvhan028]

May let us know the case that yields empty generator

[windreamer]

lmdeploy/lmdeploy/turbomind/turbomind.py

Lines 784 to 786 in 59e8944

	except (GeneratorExit, asyncio.CancelledError) as e:
	logger.info(f'[async_stream_infer] {type(e).__name__}')
	self.model_inst.cancel()

If we cancel a request before the generation of the first token, I think we will trigger this case.