#3577 [Engine] MoE fused layout buffer too small during prefill — inference fails on short prompts (Strata 0.2.1)

Descripción

In Strata 0.2.1, inference consistently fails during the prefill stage when running a MoE model with IQ3_S (GSQ-RCO) quantization.

Even very short prompts such as "Hello", "Hi Strata", or "What are your features?" trigger an error indicating that the fused layout's Mixture-of-Experts (MoE) buffers are too small.

The model loads and the UI indicates that the available memory should be sufficient, but the engine fails to generate a response.

## Environment

- **Strata version:** 0.2.1
- **GPU 1:** NVIDIA RTX 5080
- **GPU 2:** NVIDIA RTX 3060
- **Memory configuration:** RAM + VRAM
- **Quantization:** IQ3_S (GSQ-RCO)
- **Model memory requirement shown:** 51.3 GB (RAM + VRAM)
- **Available memory reported by UI:** 78 GiB
- **Multi-GPU configuration available:** RTX 5080 + RTX 3060 + RAM (layer split)

## Steps to Reproduce

1. Launch Strata 0.2.1.
2. Load the MoE model with IQ3_S (GSQ-RCO) quantization.
3. Start a chat session.
4. Send a simple prompt such as "Hello".
5. Observe the error during the prefill stage.
6. Retry with another short prompt.

## Actual Behavior

Inference fails with errors such as:

```text
Prefill: the fused layout's MoE buffers are too small for layer 30's expert path at a chunk of 8192 tokens
```

The same error occurs again on subsequent attempts.

Another attempt produces:

```text
Prefill: the fused layout's MoE buffers are too small for layer 34's expert path at a chunk of 7168 tokens
```

The application displays:

```text
Try sending the message again.
```

No response is generated.

## Expected Behavior

- The model should successfully process short prompts and generate responses.
- MoE expert buffers should be allocated correctly for the selected prefill chunk size.
- If the requested chunk size exceeds available buffer capacity, the engine should automatically reduce the chunk size or use a safe fallback instead of aborting inference.

## Additional Observations

- The issue occurs even with extremely short input prompts.
- The failure references large internal prefill chunks (8192 and 7168 tokens), despite the actual prompts containing only a few tokens.
- Different failures reference different MoE layers (30 and 34).
- The UI reports sufficient available memory to load the model.
- Repeated attempts do not resolve the problem.

## Possible Cause

This may indicate a buffer allocation or prefill chunk-sizing issue in the fused MoE execution path, potentially related to memory allocation across GPU and system RAM.

It would be useful to verify whether the engine can dynamically resize MoE buffers or reduce the prefill chunk size when the expert path exceeds the allocated buffer capacity.

## Attachments

Screenshots showing:

1. Repeated MoE buffer errors during chat inference.
2. The selected IQ3_S quantization and reported memory requirements.
3. The RTX 5080 + RTX 3060 + RAM layer-split configuration option.

<img width="681" height="621" alt="Image" src="https://github.com/user-attachments/assets/c0956b74-dc3d-43c1-9fb2-d8741d504d49" />

<img width="428" height="38" alt="Image" src="https://github.com/user-attachments/assets/93f6abc6-53e5-4f0b-9bd0-a533e2250eaa" />

<img width="420" height="66" alt="Image" src="https://github.com/user-attachments/assets/4981d1ce-4444-41af-998d-909818aa2f5a" />