feat: configure cudagraph capture batch sizes

[CUHKSZzxy]

Summary

• Add PyTorchEngineConfig support for explicit CUDA graph capture batch sizes and route them through CacheConfig and graph runners.
• Normalize configured sizes by validating positive integers, deduplicating, sorting, filtering to max_batch_size, and always including max_batch_size so valid scheduler batches stay covered without eager fallback.
• Add focused unit coverage for normalization and CUDA graph capture-size selection.

Validation

• Focused unit tests for cudagraph capture batch sizes.
• Syntax checks for touched modules.
• Real PyTorch API server smoke test with custom capture sizes confirmed normalization, graph capture, and graph replay under concurrent decode.

Assistance

Assisted with Codex + GPT-5.5 High

[Copilot]

Pull request overview

This PR adds an explicit configuration path for CUDA graph capture batch sizes in the PyTorch engine, threading the configured sizes from PytorchEngineConfig through CacheConfig into graph runners.

Changes:

• Add cudagraph_capture_batch_sizes to PytorchEngineConfig and CacheConfig, and plumb it through config building and spec-decode config cloning.
• Update graph runners to prefer the configured capture sizes over the inferred (power-of-two) defaults.
• Expose the new option via the lmdeploy serve api_server CLI.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
lmdeploy/pytorch/engine/config_builder.py	Validates/normalizes configured capture sizes and passes them into CacheConfig.
lmdeploy/pytorch/config.py	Adds cudagraph_capture_batch_sizes to CacheConfig and propagates it in SpecDecodeConfig.from_config.
lmdeploy/pytorch/backends/graph_runner.py	Uses configured capture sizes when present.
lmdeploy/pytorch/backends/cuda/graph_runner.py	Uses configured capture sizes when present (CUDA backend).
lmdeploy/messages.py	Adds cudagraph_capture_batch_sizes to PytorchEngineConfig and documents it.
lmdeploy/cli/utils.py	Adds --cudagraph-capture-batch-sizes argument.
lmdeploy/cli/serve.py	Wires the CLI argument into PytorchEngineConfig for the API server.

---

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