[osabnis]

Overview

SmolLM3 skips RoPE on every 4th layer. On these layers, Q reaches FLASH_ATTN_EXT through the following chain:

PERMUTE -> VIEW -> RESHAPE

This is the stream view created in build_attn_mha over the reshape_3d output of the Q matmul.

is_supported_flash_attn_pattern did not recognize this pattern, so with -fa on/auto, these 9 layers fell back to CPU. This caused the graph to split and the GPU portion to fail with 4 different errors across stateful and stateless execution.

This PR extends the support check to accept:

PERMUTE -> VIEW -> RESHAPE -> base

The existing translation already handles this Q path: VIEW is a no-op on the dynamic path, and -fa off uses the same Q representation.

Additional information

Tested on Arc Pro B70 with OpenVINO 2026.4.1 using dev_backend_openvino commit b1d8ce791 and the ticket command (1024-token prompt, -c 1382 -n 128).

Decode performance

Execution	-fa on	-fa off	-fa auto
Stateful	fail -> 130.0	131.4 -> 134.9	fail -> 128.1
Stateless	fail -> 106.0	98.1 -> 98.5	fail -> 106.4

Values are decode tok/s, shown as before -> after.

Accuracy

KLD against the CPU backend, measured on 4 WikiText-2 slices with context length 1024:

• Stateless: -fa on 0.0435, same as -fa off 0.0435
• Stateful: -fa on 0.0491 vs -fa off 0.0441
• Unpatched control: stateful -fa off 0.0438, stateless -fa off 0.0435

Additional validation

• Panther Lake (Arc B390, OpenVINO dev build, on 25e3135cc + ggml-openvino: keep stateful execution with -fa off #361 + ggml-openvino: support gemma-2 soft-capped attention #362): -fa on/auto now pass in both stateful and stateless modes.
• Regression on Panther Lake (25e3135cc + ggml-openvino: keep stateful execution with -fa off #361 + ggml-openvino: support gemma-2 soft-capped attention #362, this change applied together with the Ling-mini fixes): 7 other models × 6 configurations, identical greedy output in all 42 pairs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI assistance was used for analysis, editing the PR description, and running tests.