vision-language
Use LMDeploy vision-language and multimodal APIs for media loading, VLM prompt formats, OpenAI-style multimodal payloads, image token placement, and no-download media validation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- vision-language
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use LMDeploy vision-language and multimodal APIs for media loading, VLM prompt formats, OpenAI-style multimodal payloads, image token placement, and no-download media validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/input-formats.md6.9 KB
- 参考文档references/troubleshooting.md6.2 KB
- 参考文档references/workflows.md6.5 KB
- 脚本scripts/check_vl_media_inputs.py7.4 KB
03 SKILL.md 正文
Vision Language
Use this sub-skill when the task involves LMDeploy VLM or multimodal inputs: images, videos, audio, time series, OpenAI content lists, tuple prompts such as (prompt, image), or VisionConfig.
Read First
references/input-formats.mdfor tuple prompts, OpenAI-style content blocks, media URL/data URL formats, batch shapes, multi-turn image chat,IMAGE_TOKEN, and modality support.references/workflows.mdfor copy-ready offline VLM pipeline patterns, API payload construction, base64/local media handling,VisionConfig,mm_processor_kwargs, and payload preflight.references/troubleshooting.mdfor optional dependency failures, blocked URLs, fetch timeouts, image token placement, session length, batch media mismatches, chat-template issues, and chemistry/media gaps.scripts/check_vl_media_inputs.pyfor a no-download smoke check that creates a tiny local image, validates base64/data URLs, and rejects unsafe URL examples.
Scope
Own these tasks:
- Build VLM prompts for
lmdeploy.pipeline, including(prompt, image),(prompt, [image1, image2]), batch prompt lists, and OpenAI-style message lists. - Load or encode multimodal media with
lmdeploy.vl.load_image,load_video,load_audio,load_time_series, and theencode_*_base64helpers. - Construct OpenAI-compatible multimodal request bodies using
text,image_url,video_url,audio_url, andtime_series_urlcontent blocks. - Tune VLM-specific inputs with
VisionConfig(max_batch_size=...),media_io_kwargs,mm_processor_kwargs, and manualIMAGE_TOKENplacement where model docs require it. - Preflight payloads for safe media URL schemes, local/base64 media syntax, multiple images, and server-side multimodal parser expectations without downloading model weights.
Route elsewhere:
- Text-only offline generation,
GenerationConfigbasics, streaming, logits, PPL, andlmdeploy chat:pipeline-inference. - Launching API servers/proxies, auth, API clients, endpoint routing, and transport operations:
serving-apis. - Creating AWQ/GPTQ/SmoothQuant or other quantized VLM artifacts:
quantization. - Adding or debugging VLM model implementation internals, processor classes, backend kernels, or architecture registration:
backend-extension.
Fast Start
from lmdeploy import TurbomindEngineConfig, VisionConfig, pipeline
from lmdeploy.vl import load_image
backend_config = TurbomindEngineConfig(session_len=8192)
vision_config = VisionConfig(max_batch_size=4)
with pipeline("org/vlm-or-local-folder", backend_config=backend_config, vision_config=vision_config) as pipe:
image = load_image("/path/to/image.jpg")
response = pipe(("Describe this image.", image))
print(response.text if hasattr(response, "text") else response)
For custom local VLM folder names that LMDeploy cannot map to a built-in template, pass chat_template_config=ChatTemplateConfig(model_name="template-name") when creating the pipeline.
Validation
Run the bundled check before recommending a VLM media workflow:
python sub-skills/vision-language/scripts/check_vl_media_inputs.py --print-examples
This check does not download models or fetch remote media. Model execution still requires the user’s selected VLM weights, backend dependencies, accelerator memory, and any model-specific optional packages.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).