Correct FPM runtime dtype annotations

#6
by nv-simonec - opened
NVIDIA org

Summary

  • Correct reviewed FPM dtype annotations for DSV4-Flash, DSV4-Pro, Kimi-K3, and GLM-5.2-NVFP4.
  • Rewrite current and historical Parquets plus sidecar, manifest, and catalog hashes.
  • Keep latency, workload coordinates, GEMM identity, row order, and row count unchanged.
  • Apply and validate the corrections in the dataset publisher, not FPM Gym.
  • Fix DSV4-Pro producer-config classification to use recorded source paths during clean rebuilds.

Validation

  • Dataset import from pinned AISimPlayground revision 1a5edbd5505e1505230517fa116774cb335cc1df
  • uv run --with pyarrow python scripts/manage_dataset.py validate
  • Ruff format/check and git diff --check
  • Column-level comparison of all seven DSV4-Pro Parquets: 36,669 rows, only moe_quant_mode changed
nv-simonec changed pull request status to open
NVIDIA org

Base alignment: main had already deleted the Qwen3-8B files. This PR removes the stale Qwen catalog and publisher entries so the resulting dataset validates and does not resurrect those files.

nv-simonec changed pull request status to merged

Sign up or log in to comment