[FEATURE] Prompt-length dataset pre-filtering safeguard or check `max_model_len` before forward

[vadimkantorov]

Otherwise, one can get an exception / training crash when some outlier prompt happens to be too long.

Part of the root cause:

• [Feature]: return graceful inference text input validation errors as part of output (without throwing an exception) - to enable skipping / handling bad examples after the processing of good ones vllm-project/vllm#16732

(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] File "/tmp/ray/session_2025-09-18_02-35-21_268823_22052/runtime_resources/working_dir_files/_ray_pkg_83ec426c3cf029b3/.venv/lib/python3.12/site-packages/vllm/v1/engine/async_llm.py", line 237, in add_request
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] prompt_str, request = self.processor.process_inputs(
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] File "/tmp/ray/session_2025-09-18_02-35-21_268823_22052/runtime_resources/working_dir_files/_ray_pkg_83ec426c3cf029b3/.venv/lib/python3.12/site-packages/vllm/v1/engine/processor.py", line 266, in process_inputs
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] self._validate_model_inputs(processed_inputs, lora_request)
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] File "/tmp/ray/session_2025-09-18_02-35-21_268823_22052/runtime_resources/working_dir_files/_ray_pkg_83ec426c3cf029b3/.venv/lib/python3.12/site-packages/vllm/v1/engine/processor.py", line 365, in _validate_model_inputs
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] self._validate_model_input(decoder_inputs,
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] File "/tmp/ray/session_2025-09-18_02-35-21_268823_22052/runtime_resources/working_dir_files/_ray_pkg_83ec426c3cf029b3/.venv/lib/python3.12/site-packages/vllm/v1/engine/processor.py", line 418, in _validate_model_input
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] raise ValueError(
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163] ValueError: The decoder prompt (length 1282) is longer than the maximum model length of 1280. Make sure that `max_model_len` is no smaller than the number of text tokens.
(WorkflowRunner pid=42768) ERROR 09-18 02:43:09 [workflow_runner.py:163]

[pan-x-c]

It is recommended to manually check the prompt length in the workflow and provide corresponding exception handling methods. Because the prompt length can only be obtained when the workflow is running, directly filtering out long prompts that exceed the context length will result in insufficient output data, and block the trainer.

[vadimkantorov]

In principle I agree, but in practice these vllm exceptions can be nasty - also because of multi-turn can get prompts longer than expected - so maybe some exception handling of vllm failing on too long prompts should be added...

Maybe at least a filtering script using the same Trinity machinery for formatting a prompt + iterating over a dataset and printing out number of prompt tokens for every input prompt should be good.

Even if the prompts are not exceeding the limit, (unexpectedly) long prompts are reducing the number of tokens available for model answer...