Add LMMSEvaluator (integrate lmms-eval toolkit) for multimodal (vision+audio) evaluation

[DelwinKim]

Describe your changes

This PR adds multimodal (vision + audio) evaluation to Olive by integrating the lmms-eval harness as an Olive evaluator, plus the supporting pass needed to make quantized multi-component ORT-GenAI packages evaluable.

What's added

1. LMMSEvaluator (olive/evaluator/olive_evaluator.py)
A new evaluator ("type": "LMMSEvaluator") that runs lmms-eval benchmarks from within a single olive run (can also run eval standalone). It supports two model handler types:

• ONNXModelHandler pointing at an ORT-GenAI multimodal package (genai_config.json + quantized ONNX, e.g. from MobiusBuilder + OnnxKQuantQuantization) → dispatches to the new ortgenai_mm adapter.
• HfModelHandler for HuggingFace PyTorch multimodal models → dispatches to lmms-eval's native wrapper, auto-detected from the HF model_type (overridable via model_class).

Benchmarks are selected purely by config ("tasks": ["ocrbench", "docvqa_val_lite", "librispeech_test_clean", "fleurs_en", ...]), and the harness's official scorers are used (OCRBench scoring, DocVQA ANLS, WER, etc.), so results can be compared directly against published literature.

2. ortgenai_mm adapter (olive/evaluator/lmms_ort.py)
A lmms-eval model adapter that drives ORT-GenAI multimodal generation. Implements generate_until and loglikelihood, handles image and audio inputs (og.Images / og.Audios), per-model-type processor argument shapes (Phi-4-MM vs Whisper), and EOS/stop handling (including Whisper's BOS==EOS collision).

3. CompositeToOnnxPackage pass (olive/passes/onnx/composite_to_onnx_package.py)
MobiusBuilder emits a multi-component CompositeModelHandler (decoder / vision / audio / embedding in subdirectories). A composite can't be evaluated directly — it has no single session, and LocalSystem.evaluate_model rejects composite handlers. This pass converts the fanned-out components into a single runnable ONNXModelHandler (an ORT-GenAI package), so the evaluator can load it.

4. Registration

• olive/olive_config.json: registers the new evaluator and pass.

Example recipe

A complete example recipe — build + quantize Gemma‑4 E2B, package to an ORT‑GenAI package, then evaluate vision and audio benchmarks:

{
  "input_model": { "type": "HfModel", "model_path": "google/gemma-4-E2B-it" },
  "systems": {
    "local_system": {
      "type": "LocalSystem",
      "accelerators": [
        { "device": "gpu", "execution_providers": ["CUDAExecutionProvider"] }
      ]
    }
  },
  "passes": {
    "mobius_build":  { "type": "MobiusBuilder", "precision": "fp16" },
    "int4_quantize": { "type": "OnnxKQuantQuantization", "bits": 4, "block_size": 32, "save_as_external_data": true },
    "flatten":       { "type": "CompositeToOnnxPackage" }
  },
  "evaluators": {
    "evaluator": {
      "type": "LMMSEvaluator",
      "tasks": ["ai2d_lite"],
      "batch_size": 1,
      "max_length": 16384,
      "limit": 10,
      "log_samples": false,
      "ignore_stop_strings": ["\n\n"],
      "output_path": "results/gemma4_vision_audio.json"
    }
  },
  "evaluator": "evaluator",
  "evaluate_input_model": false,
  "target": "local_system",
  "output_dir": "models/gemma4_vision_audio",
  "cache_dir": "cache/gemma4_vision_audio",
  "no_artifacts": true
}

A complete example recipe, evaluate the HF Pytorch model:

{
  "input_model": { "type": "HfModel", "model_path": "google/gemma-4-E2B-it" },
  "systems": {
    "local_system": {
      "type": "LocalSystem",
      "accelerators": [
        { "device": "gpu", "execution_providers": ["CUDAExecutionProvider"] }
      ]
    }
  },
  "evaluators": {
    "evaluator": {
      "type": "LMMSEvaluator",
      "model_class": "gemma4",
      "tasks": ["ai2d_lite"],
      "batch_size": 1,
      "max_length": 16384,
      "limit": 10,
      "log_samples": false,
      "output_path": "results/gemma4_vision_audio.json"
    }
  },
  "evaluator": "evaluator",
  "evaluate_input_model": true,
  "target": "local_system",
  "output_dir": "models/gemma4_vision_audio",
  "cache_dir": "cache/gemma4_vision_audio",
  "no_artifacts": true
}

Run it with:

olive run --config gemma4_vision_audio.json

LMMSEvaluator config reference

Field	Default	Applies to	Description
tasks	(required)	both	List of lmms-eval task names, e.g. ocrbench, docvqa_val_lite, chartqa_lite, ai2d_lite, mmmu_val, textvqa_val, librispeech_test_clean, librispeech_test_other, fleurs_en.
limit	None (full set)	both	Cap the number of samples per task.
batch_size	1	both	Currently only 1 is supported (see limitations).
max_new_tokens	256	both	Max tokens generated per generate_until request.
max_length	32768	ONNX/genai	Total sequence budget (prompt + multimodal embeds + completion) for the ORT-GenAI adapter. Image/audio embeds can be 1000+ tokens, so keep this generous; lower it (e.g. 4096) to bound memory.
system_prompt	None (auto)	both	System prompt. When unset, the adapter resolves a per-model-family default (matching lmms-eval's per-model wrappers).
prompt_template	None	ONNX/genai	Optional full override of the prompt scaffolding (format string). Escape hatch for non-standard templates; normally unnecessary since media tokens are auto-handled.
ignore_stop_strings	None	ONNX/genai	Stop strings to drop from each task's lmms-eval until list (e.g. ["\n\n"]). Use with a larger max_new_tokens for step-by-step reasoning tasks that lmms-eval would otherwise truncate at the first blank line.
log_samples	false	both	Write per-sample predictions alongside the summary.
output_path	None	both	Where to write the results JSON.

Vision-only or audio-only runs are the same recipe with the corresponding subset of tasks. Audio tasks (librispeech, fleurs, covost) additionally require FFmpeg available at runtime (HF datasets decodes audio via torchcodec).

Tests

• test/evaluator/test_lmms_ort.py — adapter unit tests (generation, loglikelihood, image/audio handling, processor branching, EOS logic, prompt building / media-token auto-detection).
• test/passes/onnx/test_composite_to_onnx_package.py — composite→ONNX pass (entry-point component selection, genai_config discovery).

Notes / limitations

• batch_size is currently fixed at 1: ORT-GenAI's per-request Generator + set_inputs flow is single-sequence, and batching ragged, per-request multimodal (image/audio) inputs isn't currently supported by the runtime.
• lmms-eval is imported lazily, so non-multimodal Olive usage is unaffected if it isn't installed. Running the LMMSEvaluator (and the full adapter test suite) requires pip install lmms-eval.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass. (Adapter tests require lmms-eval installed; mocked tests pass without it, but the tests that instantiate the real adapter need the dependency.)
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.
  • Release note: Add LMMSEvaluator for multimodal (vision + audio) evaluation via the lmms-eval harness, and CompositeToOnnxPackage to convert multi-component ORT-GenAI packages for evaluation.

(Optional) Issue link

[DelwinKim]

Sample Benchmark Results — Gemma-4 E2B-it, per-component INT4 (KQuant) sensitivity

To show the evaluator end-to-end (and that it produces literature-comparable numbers), here is a small quantization sweep (weight only) on google/gemma-4-E2B-it, evaluated with LMMSEvaluator (new evaluator that integrates lmms-eval) on CUDA using each benchmark's official scorer

Pipeline: MobiusBuilder(fp16) → OnnxKQuantQuantization(bits=4, block_size=32, nodes_to_exclude=…) → CompositeToOnnxPackage → LMMSEvaluator. The variants differ only in which components are excluded from INT4 (kept FP16):

Variant	Components at INT4	Components kept FP16	Decoder size	Package size (includes embedding)
FP16 (baseline Mobius build, no quant)	—	all	4.5 GB	~11 GB
INT4 (all)	decoder + lm_head + embedding + vision + audio	—	1.5 GB	~6.9 GB
INT4, FP16 encoders (exclude vision_encoder/*, audio_encoder/*)	decoder + lm_head + embedding	vision + audio encoders	1.5 GB	~7.5 GB
INT4, decoder-only (exclude vision_encoder/*, audio_encoder/*, embedding/*)	decoder + lm_head	embedding + vision + audio encoders	1.5 GB	~7.5 GB

Vision (higher is better; ai2d/ocrbench = accuracy, chartqa = relaxed-overall, docvqa = ANLS):

Variant	ai2d_lite	chartqa_lite	docvqa_val_lite	ocrbench
FP16 (baseline)	45.0	42.6	74.5	64.0
INT4 (all)	35.4	43.4	70.7	59.4
INT4, FP16 encoders	36.4	42.6	71.0	61.5
INT4, decoder-only	36.0	42.6	71.0	61.1

Audio (WER, lower is better):

Variant	fleurs_en	librispeech_test_clean	librispeech_test_other
FP16 (baseline)	8.78	5.54	12.98
INT4 (all)	12.86	7.00	15.96
INT4, FP16 encoders	9.90	6.70	14.76
INT4, decoder-only	9.83	6.89	14.69

Observations:

• INT4 on the decoder gives essentially all the size savings (decoder 4.5 GB → 1.5 GB, ~3×) at a modest vision cost.
• Keeping the vision/audio encoders at FP16 recovers most of the audio degradation from full INT4 — e.g. fleurs_en WER 12.86 → 9.90 and ocrbench 59.4 → 61.5 — for only ~0.6 GB more than full INT4.
• The embedding table makes almost no difference: INT4-decoder-only vs INT4-with-FP16-encoders are within noise on every task, so quantizing the embedding to INT4 is effectively free.
• chartqa_lite is flat (even slightly up) across all variants — it's not a discriminating benchmark for this model at this precision.

_{Setup: CUDA EP, batch_size=1, max_new_tokens=256, limit=2000 (lite splits are smaller), official lmms-eval scorers. FP16 = MobiusBuilder fp16 export with no quantization pass.}

[microsoft-github-policy-service]

@DelwinKim please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.

@microsoft-github-policy-service agree [company="{your company}"]

Options:

• (default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.

@microsoft-github-policy-service agree

• (when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.

@microsoft-github-policy-service agree company="Microsoft"

Contributor License Agreement

Contribution License Agreement

This Contribution License Agreement (“Agreement”) is agreed to by the party signing below (“You”),
and conveys certain license rights to Microsoft Corporation and its affiliates (“Microsoft”) for Your
contributions to Microsoft open source projects. This Agreement is effective as of the latest signature
date below.

1. Definitions.
   “Code” means the computer software code, whether in human-readable or machine-executable form,
   that is delivered by You to Microsoft under this Agreement.
   “Project” means any of the projects owned or managed by Microsoft and offered under a license
   approved by the Open Source Initiative (www.opensource.org).
   “Submit” is the act of uploading, submitting, transmitting, or distributing code or other content to any
   Project, including but not limited to communication on electronic mailing lists, source code control
   systems, and issue tracking systems that are managed by, or on behalf of, the Project for the purpose of
   discussing and improving that Project, but excluding communication that is conspicuously marked or
   otherwise designated in writing by You as “Not a Submission.”
   “Submission” means the Code and any other copyrightable material Submitted by You, including any
   associated comments and documentation.
2. Your Submission. You must agree to the terms of this Agreement before making a Submission to any
   Project. This Agreement covers any and all Submissions that You, now or in the future (except as
   described in Section 4 below), Submit to any Project.
3. Originality of Work. You represent that each of Your Submissions is entirely Your original work.
   Should You wish to Submit materials that are not Your original work, You may Submit them separately
   to the Project if You (a) retain all copyright and license information that was in the materials as You
   received them, (b) in the description accompanying Your Submission, include the phrase “Submission
   containing materials of a third party:” followed by the names of the third party and any licenses or other
   restrictions of which You are aware, and (c) follow any other instructions in the Project’s written
   guidelines concerning Submissions.
4. Your Employer. References to “employer” in this Agreement include Your employer or anyone else
   for whom You are acting in making Your Submission, e.g. as a contractor, vendor, or agent. If Your
   Submission is made in the course of Your work for an employer or Your employer has intellectual
   property rights in Your Submission by contract or applicable law, You must secure permission from Your
   employer to make the Submission before signing this Agreement. In that case, the term “You” in this
   Agreement will refer to You and the employer collectively. If You change employers in the future and
   desire to Submit additional Submissions for the new employer, then You agree to sign a new Agreement
   and secure permission from the new employer before Submitting those Submissions.
5. Licenses.

• Copyright License. You grant Microsoft, and those who receive the Submission directly or
  indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license in the
  Submission to reproduce, prepare derivative works of, publicly display, publicly perform, and distribute
  the Submission and such derivative works, and to sublicense any or all of the foregoing rights to third
  parties.
• Patent License. You grant Microsoft, and those who receive the Submission directly or
  indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license under
  Your patent claims that are necessarily infringed by the Submission or the combination of the
  Submission with the Project to which it was Submitted to make, have made, use, offer to sell, sell and
  import or otherwise dispose of the Submission alone or with the Project.
• Other Rights Reserved. Each party reserves all rights not expressly granted in this Agreement.
  No additional licenses or rights whatsoever (including, without limitation, any implied licenses) are
  granted by implication, exhaustion, estoppel or otherwise.

6. Representations and Warranties. You represent that You are legally entitled to grant the above
   licenses. You represent that each of Your Submissions is entirely Your original work (except as You may
   have disclosed under Section 3). You represent that You have secured permission from Your employer to
   make the Submission in cases where Your Submission is made in the course of Your work for Your
   employer or Your employer has intellectual property rights in Your Submission by contract or applicable
   law. If You are signing this Agreement on behalf of Your employer, You represent and warrant that You
   have the necessary authority to bind the listed employer to the obligations contained in this Agreement.
   You are not expected to provide support for Your Submission, unless You choose to do so. UNLESS
   REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING, AND EXCEPT FOR THE WARRANTIES
   EXPRESSLY STATED IN SECTIONS 3, 4, AND 6, THE SUBMISSION PROVIDED UNDER THIS AGREEMENT IS
   PROVIDED WITHOUT WARRANTY OF ANY KIND, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTY OF
   NONINFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7. Notice to Microsoft. You agree to notify Microsoft in writing of any facts or circumstances of which
   You later become aware that would make Your representations in this Agreement inaccurate in any
   respect.
8. Information about Submissions. You agree that contributions to Projects and information about
   contributions may be maintained indefinitely and disclosed publicly, including Your name and other
   information that You submit with Your Submission.
9. Governing Law/Jurisdiction. This Agreement is governed by the laws of the State of Washington, and
   the parties consent to exclusive jurisdiction and venue in the federal courts sitting in King County,
   Washington, unless no federal subject matter jurisdiction exists, in which case the parties consent to
   exclusive jurisdiction and venue in the Superior Court of King County, Washington. The parties waive all
   defenses of lack of personal jurisdiction and forum non-conveniens.
10. Entire Agreement/Assignment. This Agreement is the entire agreement between the parties, and
    supersedes any and all prior agreements, understandings or communications, written or oral, between
    the parties relating to the subject matter hereof. This Agreement may be assigned by Microsoft.