fix(write_io): capture complete generate() I/O traces for LLM and VLM models

[talnufai]

Summary

Fix several issues in the write_io_files capture flow when using:

model.generate(..., write_io=True)

This PR resolves QRANIUMSW-63509 by:

• Capturing the correct chunk_inputs during chunked prefill instead of the full padded inputs
• Preserving all prefill and decode iterations using indexed file names
• Capturing missing vision encoder inputs and outputs in VLM paths
• Preventing collisions between vision and language prefill captures
• Generating a timestamped io_dir for each generate() invocation
• Fixing Dual-QPC language ONNX path selection
• Adding runtime and virtual environment artifacts to .gitignore

Changes

Correct prefill input capture

Use the same chunk_inputs passed to session.run() when writing captured I/O.

Preserve all iterations

Store each generation step in a unique file:

prefill_0
prefill_1
...
decode_1
decode_2
...

Also remove logic that disabled capture after the first decode step.

Capture vision encoder I/O

Add write_io_files support for VLM vision encoder execution paths.

Vision captures are stored separately from language captures:

vision_prefill_0
vision_prefill_1
...

This avoids overwrite/collision with language prefill outputs.

Timestamp output directories

Each generate() call now writes into a unique timestamped directory:

io_dir/<timestamp>/

preventing collisions across multiple runs.

Dual-QPC fix

Replace reliance on self.onnx_path[1] with the explicit language model ONNX path.

Example Output

Text-to-Text Model

~/.cache/qeff_models/Qwen2ForCausalLM/
└── Qwen2ForCausalLM-1657fb5c3133a3d4/
    └── io_dir/
        └── 20260907_033641/
            ├── aic_batch_io.json
            ├── prefill_0
            ├── decode_1
            ├── decode_2
            ├── ...
            └── decode_99

Vision-Language Model

~/.cache/qeff_models/LlavaForConditionalGeneration/
└── LlavaDecoderWrapper-84c9b297d91aeb74/
    └── io_dir/
        └── 20260907_033803/
            ├── aic_batch_io.json
            ├── vision_prefill_0
            ├── prefill_0
            ├── prefill_1
            ├── prefill_2
            ├── prefill_3
            ├── prefill_4
            ├── decode_1
            ├── decode_2
            ├── ...
            └── decode_127

This demonstrates that vision prefill, language prefill, and all decode iterations are now preserved instead of being overwritten.

Result

With write_io=True, all inputs and outputs used throughout generation are captured and preserved for both text and vision-language models, including:

• Vision prefill
• Language prefill
• All decode iterations

Validation

Verified on:

• Qwen2ForCausalLM
• GPT2LMHeadModel
• LlavaForConditionalGeneration

Confirmed:

• Correct chunked input capture
• Vision encoder capture
• No I/O overwrites
• No vision/lang file collisions
• Unique output directories per generate() call

[talnufai]

Hi @quic-rishinr, the Python formatting issue is fixed as you requested