Nanbeige4.2-3B please

[sgmihai]

I do not see it mentioned, and maybe it is easy to do since 4.1 is already supported

[Atomic-Germ]

4.1 was apparently a llama, 4.2 has all kinds differences and it looks like it needs a new xclbin which we can't make but the AMD devs can.
Anyway, I have converted it, but it's not going to run yet (not just the hidden size):

$ flm run nanbeige4.2:3b
[FLM]  Using custom model list path: /home/atomic-germ/.config/flm/model_list.json
230 (140304904493376): PID(32024): Created pcidev (0000:c2:00.1)
495509 (140304904493376): PID(32024): Opened /dev/accel/accel0 as 3, sysfs: /sys/bus/pci/devices/0000:c2:00.1
525135 (140304904493376): PID(32024): Expanding BO from 0 to 67108864
10052009 (140304904493376): PID(32024): Created expandable AMDXDNA_BO_DEV_HEAP: hdl=1 sz=0x4000000 paddr=0x4000000 vaddr=0x7f9b14000000 uptr=0x0
10069362 (140304904493376): PID(32024): Created device (0000:c2:00.1) ...
[FLM]  Loading model: /home/atomic-germ/.config/flm/models/Nanbeige4.2-3B-NPU2
[ERROR]  Failed to load model: Unsupported hidden size: 3072

---

Actually, let's make it easier so they might actually do it. Why it doesn't run:

1. lm_config.hpp

• LM_Config::from_pretrained has no loop keys. The 4.2 config.json carries:
  num_loops: 2, skip_loop_final_norm: false. Add JSON_GET for both
  num_loops u32 default 1, skip_loop_final_norm bool default false

2. nanbeige_npu

• hardcodes hidden_size == 0xA00 (2560) as cmp $0xa00 in nanbeige_npu::Impl, throws Unsupported hidden size: if it isn't.
  4.2 uses 3072 - They could make it pull that from config like the qwen3.5 variants
• RoPE inv_freq is a hardcoded inv_freq_3b - could build that from rope_theta and head_dim (70M & 128 respectively instead to be flexible
• 4.2 is looped twice per token, num_loops = 2, skip_loop_final_norm = false & RMS norm applied after,
• cache index = loop_idx * num_hidden_layers + layer_idx 2x KV memory. Each pass attention is only in its own loop's cache

stuff we know:

key	value
model_type	nanbeige
hidden_size	3072
num_hidden_layers	22
num_attention_heads	48
num_key_value_heads	8
head_dim / kv_channels	128
intermediate_size	10752
rms_norm_eps	1e-05
rope_theta	70000000
vocab_size	166144
max_position_embeddings	262144
num_loops	2
skip_loop_final_norm	false

• Standard llama-style layout, 201 tensors (22 blocks × 9 + embed/lm_head/
  final_norm). special flags are all off, so no weight changes are needed for
  loop support.
• Tokenizer is same as 4.1

I don't think there's much more needed. Hopefully this helps.