ms

Diagnose a training failure, accuracy problem, or performance bottleneck by routing to the right specialist skill in diagnose mode

Diagnose and fix a training failure, accuracy problem, or performance bottleneck by routing to the right specialist skill in fix mode

Integrate algorithm features or operators into an existing model codebase

Route to MindSpore migration tools (HF models, third-party repos)

Check whether the local workspace is ready to train or run inference

Diagnose accuracy regressions, numerical drift, wrong-result issues, and cross-platform mismatch after successful execution by analyzing the symptom, validating consistency across data, config, model, checkpoint, and runtime, preserving a reusable snapshot, and emitting an actionable report.

Adapt a paper feature, released reference implementation, or user-described algorithm change such as manifold-constrained hyper-connections (mHC) or Attention Residuals (AttnRes) into an existing model codebase, generate the minimal patch, and hand the updated workspace to readiness validation.

Auto-invoked when users ask mindspore api questions. such as mint.*, tensor.*, forward, backward, cpu/gpu/npu.

Diagnose training and runtime failures across MindSpore and PTA (PyTorch + torch_npu) by analyzing failure evidence, validating the most likely root causes, preserving a reusable diagnosis snapshot, and emitting an actionable report.

Migrate model implementations into the MindSpore ecosystem by first analyzing the source model or repo, then selecting the correct migration route, building the migration, and verifying the result. Use this as the top-level migration entry instead of asking users to choose `hf-transformers`, `hf-diffusers`, or generic PyTorch migration paths up front.

Build framework operators through one of two implementation methods: custom-access integration that avoids changing framework source, or native-framework integration that routes into the concrete framework workflow, verifies the result, and delivers the expected artifact.

Diagnose throughput, latency, memory, utilization, dataloader, and communication bottlenecks after a MindSpore or torch_npu workload already runs by analyzing performance evidence, validating the most likely bottlenecks, preserving a reusable snapshot, and emitting an actionable report.

Check whether a local single-machine workspace is ready to train or run inference, explain what is missing, emit a readiness report, and optionally apply safe user-space fixes. Use for pre-run workspace readiness checks, training or inference preflight, missing-item analysis, runtime_smoke-based readiness certification, or safe environment remediation before execution.

1 hook across 1 event