Consolidate MIC usage, slim weights and reviewed framework expectations

4b1f143 verified 3 days ago

1.98 kB

---
tags:
  - mic
  - test-artifact
base_model: llava-hf/llava-1.5-7b-hf
---
# Slim fixture: mic-llava-slim

A 4.3M-parameter fixture for the [llava MIC](https://huggingface.co/hf-internal-testing/llava-integration), covering [`llava-hf/llava-1.5-7b-hf`](https://huggingface.co/llava-hf/llava-1.5-7b-hf). It is a test artifact, not a usable model, tagged `test-artifact` to keep it out of candidate-model search.

- **Fixture key `232603d3d5e9`.** It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
- **Same architecture, smaller config:** `LlavaForConditionalGeneration`, shrunk by the generic rule in [`sweep.py`](https://github.com/huggingface/model-integration-contracts/blob/main/experiments/sweep.py) (vision num_hidden_layers 2, text num_hidden_layers 2).
- **Processor files copied byte for byte** from `llava-hf/llava-1.5-7b-hf`: `added_tokens.json`, `chat_template.jinja`, `chat_template.json`, `generation_config.json`, `preprocessor_config.json`, `processor_config.json`, `special_tokens_map.json`, `tokenizer.json`, `tokenizer.model`, `tokenizer_config.json`.
- **Random weights (seed 17).** They passed the input-sensitivity gate: every perturbation of the contract's inputs moved the logits by at least 0.001 (weakest: `pixel_values:halve`, 0.035), so no training was needed.
- **Same tensor names as the real checkpoint.** The real checkpoint has every parameter, and so does the fixture.
- **Saved in float16**, the dtype `llava-hf/llava-1.5-7b-hf` loads in by default.
- **Finite in float32, bfloat16, and float16** (forward, loss, gradients).

Built by [`build_fixtures.py`](https://github.com/huggingface/transformers-mic/blob/main/example/build_fixtures.py); [`build.json`](build.json) records the pinned sources, gate results, and file hashes. The source repository commits only `README.md` and `build.json`; rebuild the other files with `python example/build_fixtures.py --fixture mic-llava-slim`.

	---
	tags:
	- mic
	- test-artifact
	base_model: llava-hf/llava-1.5-7b-hf
	---

	# Slim fixture: mic-llava-slim

	A 4.3M-parameter fixture for the [llava MIC](https://huggingface.co/hf-internal-testing/llava-integration), covering [`llava-hf/llava-1.5-7b-hf`](https://huggingface.co/llava-hf/llava-1.5-7b-hf). It is a test artifact, not a usable model, tagged `test-artifact` to keep it out of candidate-model search.

	- Fixture key `232603d3d5e9`. It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
	- Same architecture, smaller config: `LlavaForConditionalGeneration`, shrunk by the generic rule in [`sweep.py`](https://github.com/huggingface/model-integration-contracts/blob/main/experiments/sweep.py) (vision num_hidden_layers 2, text num_hidden_layers 2).
	- Processor files copied byte for byte from `llava-hf/llava-1.5-7b-hf`: `added_tokens.json`, `chat_template.jinja`, `chat_template.json`, `generation_config.json`, `preprocessor_config.json`, `processor_config.json`, `special_tokens_map.json`, `tokenizer.json`, `tokenizer.model`, `tokenizer_config.json`.
	- Random weights (seed 17). They passed the input-sensitivity gate: every perturbation of the contract's inputs moved the logits by at least 0.001 (weakest: `pixel_values:halve`, 0.035), so no training was needed.
	- Same tensor names as the real checkpoint. The real checkpoint has every parameter, and so does the fixture.
	- Saved in float16, the dtype `llava-hf/llava-1.5-7b-hf` loads in by default.
	- Finite in float32, bfloat16, and float16 (forward, loss, gradients).

	Built by [`build_fixtures.py`](https://github.com/huggingface/transformers-mic/blob/main/example/build_fixtures.py); [`build.json`](build.json) records the pinned sources, gate results, and file hashes. The source repository commits only `README.md` and `build.json`; rebuild the other files with `python example/build_fixtures.py --fixture mic-llava-slim`.