runanywhere

/

qwen3_vl_HNPU

like 1

Follow

RunAnywhere, Inc. 8

Adds the Hexagon v75 Qwen3-VL-2B vision (VLM) bundle under v75/ (source Qwen/Qwen3-VL-2B-Instruct).

Files (~2.8G):

• qwen3vl-2b-vlm-512.json — VLM manifest (family=vlm, host-op qwen3vl_generate)
• vis_native_512_f16.bin — native-resolution vision tower (fp16)
• llm_shared_512_w8.bin — shared 28-layer Qwen3 decoder (W8A16), M-RoPE
• lmhead_wqo_w8.bin — tied lm-head (W8, weight-only)
• embed_f16.bin, tokenizer.json

Params: hidden 2048, 28 layers, vocab 151936, head_dim 128, kv_dim 1024, rope_theta 5e6, max_ctx 1024, eos 151645.
Compile: QAIRT 2.47.0, dsp_arch v75, soc_model 57 (O3/VTCM). Vision fp16 + W8 weight-only decode/lm-head.

Runtime note (important): the VLM manifest drives inference through the qwen3vl_generate host-op (vision -> M-RoPE splice -> shared decode). That op must be registered in the QHexRT runtime; the current public qhx_generate registers internvl/gemma/qwen3_5/etc. but not qwen3vl_generate yet, so this bundle needs a runtime build that includes it. The text-path v75 build (v81-style generic-op plan, no custom host-op) is separately in progress and its decode export gates greedy first-token 12095 (" Paris") == HF gold. Published for v75 arch parity with the v79/v81 dirs.