brandonmusic

/

GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78

Like 9

Summary

Follow up the BF16-shared checkpoint revision with reproducible lineage,
independent runtime evidence, focused converter tests, and two metadata
corrections. This PR does not replace or upload any model shard.

Metadata corrections

• Set safetensors metadata.total_size to 346,196,508,248, the sum of all
  186,905 tensor payloads parsed from the 157 shard headers. The current value
  is 22,130,880 bytes too high.
• Preserve the old R7 assembly manifest as SOURCE_R7_MANIFEST.json and remove
  its stale seals. It describes the parent layout (187,580 tensors and 75
  rewritten carriers through layer 77), not the current BF16-shared layout
  (186,905 tensors and 76 carriers through layer 78).

Reproducibility and validation

• Add the source/donor revisions, hashes, central-shard identity, byte
  accounting, validation totals, and an exact 228-tensor comparison of the
  canonical central shard against the pinned BF16 donor.
• Add a converter matching the canonical central-shard layout and focused
  tests for carrier rewriting, metadata ownership, and manifest lifecycle.
• Record the independent TP4 MTP0/MTP3, prefill, CUDA-graph, memory, and paired
  KLD validation that motivated the canonical revision.

Checks: 3 passed; Ruff check and format passed; all JSON parsed; the
corrected total_size was independently recomputed from every safetensors
header.

This supersedes PR #2, which was uploaded from the old parent concurrently
with the canonical BF16-shared commit and therefore conflicts with it.