Hugging Face's logo

HanzoHuang
/
Qwen2.5-VL-3B-Instruct-RKLLM

Image-Text-to-Text
RKLLM
rockchip
rk3576
rk3588
qwen2.5-vl
Model card Files
xet
 Community
HanzoHuang commited on
Commit
376e844
·
verified ·
1 Parent(s): 6ce781a

Publish validated RKLLM 1.3.1 artifact

Browse files
Files changed (1) hide show
  1. README.md +101 -0
README.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen2.5-VL-3B-Instruct
4
+ library_name: rkllm
5
+ pipeline_tag: image-text-to-text
6
+ tags: [rkllm, rockchip, rk3576, rk3588, qwen2.5-vl]
7
+ ---
8
+
9
+ # Qwen2.5-VL-3B-Instruct · RKLLM
10
+
11
+ Text and image chat on Rockchip RK3576 and RK3588, converted from [Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct).
12
+
13
+ * Max context length: 8192
14
+
15
+ - **Conversion toolkit:** RKLLM Toolkit 1.3.1 and RKNN Toolkit2 2.5.0.
16
+ - **Inputs / output:** text and one image → text. Video and audio input are unavailable in this runtime.
17
+ - **License:** Apache 2.0; see the [original model](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) and [license](LICENSE).
18
+
19
+ ## Choose your files
20
+
21
+ Use files for your board's exact SoC. RK3576 offers compact W4A16-g128 and W8A8; RK3588 uses W8A8. Download size is not runtime RAM.
22
+
23
+ | Board | Quantization | Size | Language model |
24
+ | --- | --- | --- | --- |
25
+ | RK3576 | W4A16-g128 | 2.60 GiB | [Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm](v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm) |
26
+ | RK3576 | w8a8 | 3.70 GiB | [Qwen2.5-VL-3B-Instruct_RK3576_w8a8.rkllm](v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w8a8.rkllm) |
27
+ | RK3588 | w8a8 | 3.51 GiB | [Qwen2.5-VL-3B-Instruct_RK3588_w8a8.rkllm](v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_RK3588_w8a8.rkllm) |
28
+
29
+ Every language model requires its board's matching vision encoder below. Both RK3576 language variants share the RK3576 encoder.
30
+
31
+ | Board | Size | Vision encoder |
32
+ | --- | --- | --- |
33
+ | RK3576 | 1.52 GiB | [Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn](v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn) |
34
+ | RK3588 | 1.38 GiB | [Qwen2.5-VL-3B-Instruct_vision_RK3588.rknn](v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_vision_RK3588.rknn) |
35
+
36
+ ## Run on your board
37
+
38
+ The vision encoders require **RKNN Runtime 2.5.0** and the language models require **RKLLM Runtime 1.3.1**. Set `RKLLM_RUNTIME_IMAGE` to an ARM64 runtime image with those versions. The validation used a locally built image; a model-specific Docker image is not included in this release.
39
+
40
+ Download the RK3576 W4A16-g128 language model and matching encoder:
41
+
42
+ ```bash
43
+ hf download HanzoHuang/Qwen2.5-VL-3B-Instruct-RKLLM \
44
+ v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm \
45
+ v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn \
46
+ --local-dir ./models
47
+
48
+ sudo docker run --rm -it --privileged \
49
+ -p 8001:8001 -v /dev:/dev \
50
+ -v "$PWD/models/v1.3.1/RK3576:/app/models:ro" \
51
+ -e MODEL_KIND=vlm -e TARGET_PLATFORM=rk3576 -e RKLLM_EMBED_FLASH=1 \
52
+ -e MODEL_FILE=Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm \
53
+ -e VISION_MODEL_FILE=Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn \
54
+ ${RKLLM_RUNTIME_IMAGE:?Set RKLLM_RUNTIME_IMAGE to an ARM64 RKLLM 1.3.1 / RKNN 2.5.0 image} \
55
+ serve --max_context_len 4096
56
+ ```
57
+
58
+ For RK3588, download both RK3588 files, mount that directory, set `TARGET_PLATFORM=rk3588`, and select its W8A8 language filename. Use the [runtime documentation](https://github.com/Hanzo-Huang/rkllm-docker) for health checks and OpenAI-compatible image requests.
59
+
60
+ ## Release notes
61
+
62
+ - Source acquired from ModelScope `Qwen/Qwen2.5-VL-3B-Instruct`, pinned revision `1b5a7674d40a0ac9acef541a46fed12d710bd36b`. The HF link above identifies the canonical upstream model; no equivalence to an HF Git commit is assumed.
63
+ - Export context: **8192 tokens**. Functional API tests used **4096**; the speed/memory benchmark used **2048**. Longer runtime context has not been validated on these boards.
64
+ - Vision export follows the SDK placement rule, including a CPU-targeted patch-embedding convolution.
65
+ - Fixed vision input: RGB, 448 × 448, batch 1, 256 image tokens, CLIP mean [0.48145466, 0.4578275, 0.40821073] and std [0.26862954, 0.26130258, 0.27577711]. The runtime pads to a square using RGB 127, then resizes bilinearly. The encoder returns one pooler output with 2048 channels per image token.
66
+ - Language calibration: the SDK's 20 image-and-text examples, captured using the source processor and language-model input hook. W4A16-g128 uses GRQ; W8A8 uses normal quantization. Float32 CUDA loading; two NPU cores on RK3576 and three on RK3588.
67
+ - Export compatibility uses the pooler output under Transformers 5, imports only the requested calibration model family, and supplies a temporary nested text_config view for the SDK language loader. It restores the source mrope_section rotary configuration after SDK model loading. Original checkpoint weights and configuration files are unchanged.
68
+ - All three variants passed seven API smoke checks on their target board: arithmetic, independent requests, SSE, exact conversation recall, image description, image color over SSE, and text after an image. Tests used the baked RKLLM 1.3.1 application with RKNN Runtime 2.5.0 and `embed_flash=1`.
69
+ - RK3576 vision encoder cosine against original float32 embeddings: 0.99772957.
70
+ - RK3588 vision encoder cosine against original float32 embeddings: 0.99708773.
71
+
72
+ These are functional and numerical smoke checks, not a broad quality benchmark. Quantization and fixed image resolution may reduce accuracy.
73
+
74
+ Measured on the target boards:
75
+
76
+ | Board | Quantization | Decode tok/s | Startup RAM (GB) | Stable RAM (GB) | Date |
77
+ | --- | --- | ---: | ---: | ---: | --- |
78
+ | RK3576 | w4a16-g128 | 6.17 | 4.19 | 4.14 | 2026-10-23 |
79
+ | RK3576 | w8a8 | 5.13 | 5.32 | 5.32 | 2026-10-23 |
80
+ | RK3588 | w8a8 | 8.47 | 4.97 | 5.00 | 2026-10-23 |
81
+
82
+ Speed is the median native decode rate over three 128-token runs after a 32-token warmup, greedy decoding, EOS ignored, context 2048. RAM is additional board usage above a five-second idle baseline: sampled loading peak and ten-second ready-state median, in decimal GB. The matched encoder is loaded; `embed_flash=1`. First-token timing stays in the private report. Measurements depend on the board, clocks, cooling, and background load.
83
+
84
+
85
+
86
+ <details>
87
+ <summary>SHA-256 checksums</summary>
88
+
89
+ ```text
90
+ de7152bfb6c2f532a048ac0e50a353a7a74dc8f10009af8f99e80fb273eda830 v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w4a16_g128.rkllm
91
+ 41f1488381f16666ce89545028f9ec8457052be591c036cd68a309af7d671612 v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_RK3576_w8a8.rkllm
92
+ e5bb20a18703232c14870c0abb157950e5b98f3c78b9c275a1c602c17efd85c3 v1.3.1/RK3576/Qwen2.5-VL-3B-Instruct_vision_RK3576.rknn
93
+ 3621ae4e255f08fba6b7b4a46ccfda3775bd4b8da6cdf4336aeff37947b8eba3 v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_RK3588_w8a8.rkllm
94
+ bc6620d506b583b2c4bbd51fea9aecdf07d109a5cbbb5e7e35906de2041e009a v1.3.1/RK3588/Qwen2.5-VL-3B-Instruct_vision_RK3588.rknn
95
+ ```
96
+
97
+ </details>
98
+
99
+ ## Credits
100
+
101
+ [Qwen](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) provides the original model. Conversion uses [Rockchip RKLLM](https://github.com/airockchip/rknn-llm/tree/release-v1.3.1), [RKNN Toolkit2](https://github.com/airockchip/rknn-toolkit2), and the [conversion wrapper](https://github.com/Hanzo-Huang/rknn-llm).