mvp-engine

Use and extend the MVP-Engine MLLM data kit with explicit specs, handlers, and sample objects. Use for recipe data wiring, MLLMDataSpec construction, source resample/resolve_refs policy, schema/media/tokenization handlers, packing, guards, collation, dataloader setup, Qwen VL data support, and multimodal data extensions.

Add, review, update, and validate recipe-local sequence-parallel plans that reuse the tensor-parallel mesh in mvp-engine models.

Add, review, update, and validate recipe-local tensor-parallel plans, TP metadata postprocessors, and compatible mesh config for mvp-engine models.

Add, review, update, and validate VLM/MLLM data pipelines using the current spec/handler-based MLLMDataKit design. Use for raw schema normalization, MLLMDataSpec wiring, source resample/resolve_refs policy, media extensions, processor setup, packing, guards, collation, step estimation, and recipe integration.

Add, review, update, and validate VLM/MLLM packing behavior around the current MLLMDataKit design, including MLLMPackingSpec, MLLMPackingAssembler, MLLMPack metadata, block-causal masks, model-specific packed input preparation, token accounting, and step estimation.

Use LigerKernelKit for reusable Liger Kernel integration before model construction, covering official model-family dispatch, custom-model symbol patching via LigerPatch, module selection validation, and loss-kernel guards.

Decide where and how to wire Liger Kernel into an MVP-Engine recipe using LigerKernelKit, including official family dispatch, authoring a custom or composite model's LigerPatch map, module selection, and loss-kernel safety.

Use TokenNormedLossKit for unreduced per-token loss patching, accumulation-window global token normalization, gradient rescaling, and token loss logging in MVP-Engine recipes.

Add, review, update, and validate loss spike guards for scalar-loss or per-token-loss training, including config thresholds, micro-batch skip wiring, distributed loss reduction.

Use MFUKit for global model FLOPs utilization logging, including microbatch FLOPs accumulation, device peak resolution, distributed reduction, and perf/mfu log payload construction.

Use and extend MLLMModelKit for multimodal model loading, recipe patches, freeze policy, gradient checkpointing, and torch.compile wiring in MVP-Engine recipes.

Use OptimKit to build optimizers over trainable parameters and Transformers learning-rate schedulers in MVP-Engine recipes.

Add, review, update, and validate gradient checkpointing through MLLMModelKit when possible, with recipe-local fallback only for model-specific activation-checkpoint wiring.

Add, review, update, and validate torch model.compile support using MLLMModelKit when possible, including compile config, scope, placement before distributed wrapping, and recipe-local fallback for special graph boundaries.

Add, review, update, and validate VLM freeze policy using MLLMModelKit when possible, including vision/projector/language prefix groups, optimizer interaction, trainable dtype handling, and MFU impacts.

Add, review, update, and validate model FLOPs utilization using MFUKit, including recipe-provided calculate_model_flops, microbatch accumulation, device peak resolution, distributed reduction, and perf/mfu logging.

Add, review, update, and validate token-normalized training loss using TokenNormedLossKit, including unreduced per-token loss patching, accumulation-window global supervised-token normalization, gradient rescaling, and token-level logging.

Analyze mvp-engine experiment outputs and produce a structured Markdown report grounded in run artifacts, logs, checkpoints, optional eval files, and baseline comparisons.

Triage reviewer feedback on an open PR, implement scoped fixes, rerun validation, and prepare reviewer-ready responses tied to comments, changed files, and validation evidence.

Run a pre-PR or pre-push quality gate by inspecting the branch diff, tightening touched public API docstrings/types, running validation, and summarizing residual risk.

Merge an upstream branch into a recipe development branch by comparing branch deltas, mapping shared-contract breakages, resolving conflicts, adapting the recipe, validating affected paths, and completing the merge safely.

Add, review, update, and validate migration of external model/configuration code into mvp-engine recipes with strict behavior parity, checkpoint compatibility, and optional NPU variant support.

Add, review, update, and validate recipe/model-local FSDP2 custom prefetching for models whose wrapped module execution order needs explicit forward or backward prefetch edges.

Create a new recipe scaffold under recipes/ with standard config, engine, dataset, model, README, and recipe-local test structure while keeping task-specific implementation intentionally minimal.