Skip to main content

← All model recipe tasks

qwen_vl model family

Tasks

Supported architectures and task heads

The Hugging Face values below are copied from checkpoint metadata at the recorded revision. The TRTMC task contract comes from the exact E2E recipe. Architecture identity selects or describes a source graph; it does not imply that TRTMC reproduces every Hugging Face head with the same model_type. For example, an encoder recipe may consume the base model and intentionally return hidden states instead of the checkpoint's pretraining or classification logits.

HF model_typeHF architecture / pipeline classTRTMC task contractExact recipe profiles
qwen2_5_vlQwen2_5_VLForConditionalGeneration
Source: config.architectures
Metadata: config.json at 66285546d2b8
vision_language_generation
qwen25vl-3b
qwen25vl-3b-tp2
qwen3_vlQwen3VLForConditionalGeneration
Source: config.architectures
Metadata: config.json at 89644892e4d8
vision_language_generation
qwen3-vl-2b
qwen3-vl-2b-tp4

Declared recipes

Each row comes from a model-owned E2E manifest. It declares a test recipe; it is not a live pass receipt for every hardware target.

RecipeExact Hugging Face checkpointTaskBuild configurationDeclared E2E casesManifest
qwen25vl-3bQwen/Qwen2.5-VL-3B-Instructfp32; single deviceqwen25vl-3btests/e2e/models/qwen_vl/manifests/qwen25vl-3b.json
qwen25vl-3b-tp2Qwen/Qwen2.5-VL-3B-Instructfamily default; TP2qwen25vl-3b-tp2tests/e2e/models/qwen_vl/manifests/qwen25vl-3b-tp2.json
qwen3-vl-2bQwen/Qwen3-VL-2B-Instructbf16; single deviceqwen3-vl-2btests/e2e/models/qwen_vl/manifests/qwen3-vl-2b.json
qwen3-vl-2b-tp4Qwen/Qwen3-VL-2B-Instructfamily default; TP4qwen3-vl-2b-tp4tests/e2e/models/qwen_vl/manifests/qwen3-vl-2b-tp4.json

Family-owned configuration

These keys are extracted from the family's registered config schema. Pass one key per repeatable --set namespace.field=value argument. Build-time keys belong on trtmc build; session keys belong on the family's inference command.

qwen_vl_decoder

Schema source: python/tensorrt_model_connect/families/qwen_vl/runtime_config_schema.py

--set keyTypeDefaultCLI surfaceAllowed configuration layers
qwen_vl_decoder.decode_attentionstring"native"buildBuild Time, Bundle Default, Session Request
qwen_vl_decoder.max_prefill_lengthint320buildBuild Time, Bundle Default, Session Request
qwen_vl_decoder.opt_prefill_lengthint3264buildBuild Time, Bundle Default, Session Request
qwen_vl_decoder.builder_workspace_gibint321buildBuild Time, Bundle Default, Session Request
trtmc build <hf-id> --output <bundle.bundle> --set qwen_vl_decoder.decode_attention=<value>

qwen_vl_lora

Schema source: python/tensorrt_model_connect/families/qwen_vl/runtime_config_schema.py

--set keyTypeDefaultCLI surfaceAllowed configuration layers
qwen_vl_lora.enabledboolfalsebuildBuild Time, Bundle Default, Session Request
qwen_vl_lora.max_rankint320buildBuild Time, Bundle Default, Session Request
qwen_vl_lora.target_modulesstring"q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj"buildBuild Time, Bundle Default, Session Request
trtmc build <hf-id> --output <bundle.bundle> --set qwen_vl_lora.enabled=<value>

qwen_vl_vision

Schema source: python/tensorrt_model_connect/families/qwen_vl/runtime_config_schema.py

--set keyTypeDefaultCLI surfaceAllowed configuration layers
qwen_vl_vision.image_heightint32448buildBuild Time, Bundle Default, Session Request
qwen_vl_vision.image_widthint32448buildBuild Time, Bundle Default, Session Request
qwen_vl_vision.dynamic_resolutionboolfalsebuildBuild Time, Bundle Default, Session Request
qwen_vl_vision.min_pixelsint320buildBuild Time, Bundle Default, Session Request
qwen_vl_vision.opt_pixelsint32200704buildBuild Time, Bundle Default, Session Request
qwen_vl_vision.max_pixelsint320buildBuild Time, Bundle Default, Session Request
trtmc build <hf-id> --output <bundle.bundle> --set qwen_vl_vision.image_height=<value>

The registry rejects unknown namespaces and fields. A --config file can set the same keys in JSON or YAML form.

Family-specific CLI contracts

Inputs and options below are filtered by the declared E2E task and the methods and configuration fields used by that family's native runtime implementation. The global CLI parser accepts a wider union of flags; flags absent here are not declared for this family.

trtmc run

Generate text from an image and text prompt.

Declared recipes: qwen25vl-3b, qwen25vl-3b-tp2, qwen3-vl-2b, qwen3-vl-2b-tp4

trtmc run <bundle.bundle> --prompt "<text>" --image <input.png> [generation options]
Supported input or optionRequirementRuntime behavior
--prompt <TEXT>RequiredText prompt for the vision-language recipe.
--image <PATH>Required by declared image recipesImage input consumed by the runtime image-aware generate overload.
--max-new-tokens <N>OptionalLimit the number of generated tokens or audio frames.
--temperature <F>OptionalSet sampling temperature.
--top-k <N>OptionalRestrict sampling to the top K tokens.
--top-p <F>OptionalEnable nucleus sampling at probability P.
--min-p <F>OptionalFilter tokens below min-p times the maximum probability.
--seed <N>OptionalSet the sampling or diffusion seed.
--lora-adapter <DIR> --lora-adapter-id <ID>OptionalLoad and select a dynamic LoRA adapter.
--greedyOptionalSelect deterministic greedy decoding by setting temperature to zero.
Code basis

Runtime provider: qwen_vl

  • include/trtmc/pipeline.h
  • src/cli/main.cpp
  • src/runtime/models/qwen_vl/pipeline.cpp
  • src/runtime/models/qwen_vl/pipeline.h
  • src/runtime/models/qwen_vl/sampler.cpp
  • tests/e2e/models/qwen_vl/manifests/qwen25vl-3b-tp2.json
  • tests/e2e/models/qwen_vl/manifests/qwen25vl-3b.json
  • tests/e2e/models/qwen_vl/manifests/qwen3-vl-2b-tp4.json
  • tests/e2e/models/qwen_vl/manifests/qwen3-vl-2b.json

Bundle lifecycle commands

These commands apply to every declared recipe in this family; they do not add task inputs or model capabilities.

trtmc build

Build one exact checkpoint into a TensorRT-Model-Connect bundle.

trtmc build <hf-id> --output <bundle.bundle>

trtmc inspect

Inspect bundle metadata, runtime identity, and packaged sections.

trtmc inspect <bundle.bundle>

See the CLI Reference for all options and limitations.