FLM 1.0.2 on Linux: Qwen3.5 4B and 9B hang during prefill, while Llama 3.2 and Qwen3.6 work

Environment:

• Fedora 44
• Kernel: 7.1.8-200.fc44.x86_64
• Hardware: AMD Ryzen AI 7 PRO 350, RyzenAI-npu6 (6x8)
• FLM: 1.0.2
• XRT: 2.26.0
• amdxdna: 0.8
• NPU firmware: 1.1.2.64

Validation:
flm validate succeeds:

• /dev/accel/accel0 with 8 columns
• NPU firmware 1.1.2.64
• amdxdna 0.8
• memlock infinity

Repro:
flm run qwen3.5:4b
or
flm run qwen3.5:9b

Prompt:
hi

Actual result:
Both models load successfully, then print:
[FLM] Prefill chunk 1/1 with 13 tokens
and hang indefinitely. They never reach:
[FLM] Creating checkpoint at context length 13

The Qwen3.5 model directories were removed and downloaded again after FLM reported old Qwen models as incompatible.

Working controls on the exact same system:

• llama3.2:3b processes a 41-token prefill and replies correctly.
• qwen3.6-moe:35b-a3b passes prefill, creates a checkpoint, and replies correctly.

RAM is not exhausted:

• Qwen3.5 4B/9B uses roughly 45% RAM when hung.
• qwen3.6-moe:35b-a3b works despite RAM usage above 90%.

This appears specific to the Qwen3.5 prefill/runtime path in FLM 1.0.2 on Linux.

Start by reproducing flm run qwen3.5:4b and flm run qwen3.5:9b with the listed Fedora, XRT, amdxdna, and firmware versions, then compare their prefill behavior with the working Llama 3.2 and Qwen3.6 controls. Trace the Qwen3.5 prefill/runtime path around the last logged step; done means both models complete prefill, create the checkpoint, and answer the prompt.