[Bug]: chunked-prefill fallback does not synchronize llm_args.enable_chunked_prefill

System Info

• CPU architecture: N/A (code inspection issue)
• GPU: N/A (issue identified through source analysis)
• TensorRT-LLM branch: main
• TensorRT-LLM commit: current main branch at time of investigation
• OS: N/A

Additional information:

• This issue was identified through source-code inspection and call-chain analysis.
• No specific hardware is required to observe the behavior.
• The report concerns state synchronization in the PyTorch executor initialization path.

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Summary

While reviewing tensorrt_llm/_torch/pyexecutor/py_executor_creator.py, I noticed that chunked prefill can be disabled through fallback logic, but llm_args.enable_chunked_prefill is never updated to reflect the effective runtime state.

The relevant flow is:

enable_chunked_context = llm_args.enable_chunked_prefill

Later, chunked prefill may be disabled through:

• FLASHINFER_STAR_ATTENTION fallback
• MLA unsupported-SM fallback

For example:

enable_chunked_context = False

and:

model_engine.attn_runtime_features.chunked_prefill = False

However:

llm_args.enable_chunked_prefill

is never updated.

As a result, the runtime state and user-facing configuration can diverge.

Steps to reproduce the behavior

1. Enable chunked prefill.

2. Trigger a fallback path that disables chunked prefill at runtime (for example, an unsupported MLA SM configuration or FLASHINFER_STAR_ATTENTION).

3. Observe that:

   • runtime chunked prefill is disabled
   • llm_args.enable_chunked_prefill remains True

Minimal example

Relevant pattern:

enable_chunked_context = llm_args.enable_chunked_prefill

...

enable_chunked_context = False
model_engine.attn_runtime_features.chunked_prefill = False

# llm_args.enable_chunked_prefill remains unchanged

Expected behavior

When chunked prefill is disabled through fallback logic, the effective runtime state and llm_args.enable_chunked_prefill should remain synchronized.

After initialization:

llm_args.enable_chunked_prefill

should accurately reflect whether chunked prefill is actually enabled.

actual behavior

The runtime disables chunked prefill through fallback logic, but:

llm_args.enable_chunked_prefill

remains True.

This creates a state mismatch where:

• runtime chunked prefill is disabled
• user-facing configuration still reports chunked prefill as enabled

Downstream validation and feature-status reporting may therefore observe stale state.

additional notes

This appears to be a synchronization issue rather than an intentional design choice.

Notably, create_py_executor() is already expected to mutate portions of llm_args, and the nearby kv_cache_config.enable_block_reuse logic keeps runtime state and configuration synchronized.

I was unable to find an existing issue or PR tracking this specific enable_chunked_prefill synchronization problem.

Potential areas to investigate:

• FLASHINFER_STAR_ATTENTION fallback path
• MLA unsupported-SM fallback path
• downstream validation that reads llm_args.enable_chunked_prefill
• feature-status reporting based on llm_args

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

Start in tensorrt_llm/_torch/pyexecutor/py_executor_creator.py, tracing create_py_executor() through the FLASHINFER_STAR_ATTENTION and MLA unsupported-SM fallback paths. Check downstream validation and feature-status reporting that read llm_args.enable_chunked_prefill. Done means the configuration value and effective runtime chunked-prefill state remain synchronized after initialization.